No full-text available
To read the full-text of this research,
you can request a copy directly from the authors.
A Graph-Based Approach for Mitigating Multi-Sided Exposure Bias in Recommender Systems
Abstract
Fairness is a critical system-level objective in recommender systems that has been the subject of extensive recent research. A specific form of fairness is supplier exposure fairness, where the objective is to ensure equitable coverage of items across all suppliers in recommendations provided to users. This is especially important in multistakeholder recommendation scenarios where it may be important to optimize utilities not just for the end user but also for other stakeholders such as item sellers or producers who desire a fair representation of their items. This type of supplier fairness is sometimes accomplished by attempting to increase aggregate diversity to mitigate popularity bias and to improve the coverage of long-tail items in recommendations. In this article, we introduce FairMatch, a general graph-based algorithm that works as a post-processing approach after recommendation generation to improve exposure fairness for items and suppliers. The algorithm iteratively adds high-quality items that have low visibility or items from suppliers with low exposure to the users’ final recommendation lists. A comprehensive set of experiments on two datasets and comparison with state-of-the-art baselines show that FairMatch, although it significantly improves exposure fairness and aggregate diversity, maintains an acceptable level of relevance of the recommendations.
... As an important tool to alleviate information overload, the recommender system makes a significant contribution to improving personalized experiences such as e-commerce shopping, movie recommendations, and travel recommendations by using historical interaction data between users and items to generate recommendation predictions [1,2]. However, the recommender systems that only aims to improve the recommendation utility can easily lead to the Matthew effect [3], Filter bubble [4], Process Fairness [5], Outcome Fairness [5], and other problems, and the bias of the recommender systems is one of the reasons for increasing the unfairness of process and unfairness of outcome [6]. ...
... Popularity bias refers to the fact that items with high popularity are recommended more often than their popularity [9]. Popularity bias has an important impact on the data, model, and results of the recommender systems, and is one of the reasons why most items are not fairly recommended [2]. ...
... The feedback loop ecology of the recommender system further intensifies this Matthew effect. Mansoury et al. [2] pointed out that one of the main reasons why different items do not receive fair exposure in recommendations is the influence of popularity bias, that is, a few popular items are over-recommended, while most other items do not receive due attention. Abdollahpouri et al. [23] also showed that this bias towards popular items will have a negative impact on less popular items and new items in the system. ...
To reduce the impact of rating bias and popularity bias in recommender system, and make the recommender system reach a balance between recommendation utility and debias effect at the same time, we propose a bi-process debiasing recommendation model based on matrix factorization. Firstly, considering the problem that the user’s ratings are affected by the herd mentality, which leads to a consistency between the rating and the selection of rating items, resulting in the power-law distribution, the k-times parabolic fuzzy distribution was used to fuse the user’s age to redistribute the ratings. Secondly, the loss function is optimized by the continuously increasing flow and popularity of items. Finally, user emotion and item popularity are combined to construct user psychological tendency, which is divided into three levels: strong, medium and weak, and different levels are given different weights. To verify the performance of the model, the experimental results on real datasets show that the model proposed in this paper not only effectively reduces the recommendation bias but also ensures the recommendation utility.
... The long-term impact of such decision systems has recently been explored through explicit modeling of the dynamics and feedback effects in the interactions between the decision-makers and the targeted populations [16,28,35,24,14]. In particular, the recent work of [16] has demonstrated, with concrete simulation examples, how long-term fairness can not be analyzed in closed forms, but requires the use of more computational analysis tools based on simulations. ...
... For example, [21,26] find that providing a direct subsidy for a disadvantaged group with the purpose of improving some institutional utility actually widens the gap between advantaged and disadvantaged groups over time, which further shows that long-term fairness is difficult to achieve. There have been a growing number of studies on fairness in the long-term with various algorithmic approaches [28,35,24,14]. [24] proposes a graph-based algorithm to improve fairness in recommendations for items and suppliers. ...
... There have been a growing number of studies on fairness in the long-term with various algorithmic approaches [28,35,24,14]. [24] proposes a graph-based algorithm to improve fairness in recommendations for items and suppliers. They relate fairness to breaking the perpetuation of bias in the interactions between users and items. ...
Long-term fairness is an important factor of consideration in designing and deploying learning-based decision systems in high-stake decision-making contexts. Recent work has proposed the use of Markov Decision Processes (MDPs) to formulate decision-making with long-term fairness requirements in dynamically changing environments, and demonstrated major challenges in directly deploying heuristic and rule-based policies that worked well in static environments. We show that policy optimization methods from deep reinforcement learning can be used to find strictly better decision policies that can often achieve both higher overall utility and less violation of the fairness requirements, compared to previously-known strategies. In particular, we propose new methods for imposing fairness requirements in policy optimization by regularizing the advantage evaluation of different actions. Our proposed methods make it easy to impose fairness constraints without reward engineering or sacrificing training efficiency. We perform detailed analyses in three established case studies, including attention allocation in incident monitoring, bank loan approval, and vaccine distribution in population networks.
... This phenomenon may result in an unfair representation of items from different suppliers in recommendations. Various notions of exposure bias (or its counterpart, exposure fairness) are introduced in the literature: 1) aggregate diversity which is defined as the fraction of items that appeared at least once or times in the recommendation lists [3,5,25,27,29], 2) equality of exposure which is defined as how equally the total exposure is distributed among all items [25,29,30,34], and 3) equity of exposure which is defined as how equally the total exposure is distributed among all items in the system with respect to their qualification/relevance, meaning that high-quality items (based on users' interaction) are expected to receive higher exposure than the low-quality ones [36,37,41]. In this paper, we consider the second definition and introduce several evaluation metrics for evaluating the performance of the recommendation system in long run. ...
... This phenomenon may result in an unfair representation of items from different suppliers in recommendations. Various notions of exposure bias (or its counterpart, exposure fairness) are introduced in the literature: 1) aggregate diversity which is defined as the fraction of items that appeared at least once or times in the recommendation lists [3,5,25,27,29], 2) equality of exposure which is defined as how equally the total exposure is distributed among all items [25,29,30,34], and 3) equity of exposure which is defined as how equally the total exposure is distributed among all items in the system with respect to their qualification/relevance, meaning that high-quality items (based on users' interaction) are expected to receive higher exposure than the low-quality ones [36,37,41]. In this paper, we consider the second definition and introduce several evaluation metrics for evaluating the performance of the recommendation system in long run. ...
... Most existing research to study exposure bias are conducted in static settings where a single round of recommendation results is analyzed [25,33,39]. While these studies unveiled important aspects of exposure bias and proposed solutions for tackling it, the long-term impact of this bias is yet a significant research gap which we seek to investigate in this paper. ...
Exposure bias is a well-known issue in recommender systems where items and suppliers are not equally represented in the recommendation results. This is especially problematic when bias is amplified over time as a few items (e.g., popular ones) are repeatedly over-represented in recommendation lists and users' interactions with those items will amplify bias towards those items over time resulting in a feedback loop. This issue has been extensively studied in the literature on model-based or neighborhood-based recommendation algorithms, but less work has been done on online recommendation models, such as those based on top-K contextual bandits, where recommendation models are dynamically updated with ongoing user feedback. In this paper, we study exposure bias in a class of well-known contextual bandit algorithms known as Linear Cascading Bandits. We analyze these algorithms on their ability to handle exposure bias and provide a fair representation for items in the recommendation results. Our analysis reveals that these algorithms tend to amplify exposure disparity among items over time. In particular, we observe that these algorithms do not properly adapt to the feedback provided by the users and frequently recommend certain items even when those items are not selected by users. To mitigate this bias, we propose an Exposure-Aware (EA) reward model that updates the model parameters based on two factors: 1) user feedback (i.e., clicked or not), and 2) position of the item in the recommendation list. This way, the proposed model controls the utility assigned to items based on their exposure in the recommendation list. Extensive experiments on two real-world datasets using three contextual bandit algorithms show that the proposed reward model reduces exposure bias amplification in long run while maintaining the recommendation accuracy.
... Gini coefficient. Gini coefficient is widely used in sociology and economics to measure the degree of social unfairness [28,30,52,60,61]. To our knowledge, it is also the most commonly used metric for consistent individual fairness. ...
... To our knowledge, it is also the most commonly used metric for consistent individual fairness. The utility can be the predicted relevance for a user [28,52] or the exposure for an item [30,60,61]. The lower the value, the fairer the recommendations. ...
... Entropy is often used to measure the uncertainty of a system. In recommendation, it is used to measure the inconsistency of item exposure [60,61,71]. The lower the value, the fairer the recommendations. ...
Recommender systems are an essential tool to relieve the information overload challenge and play an important role in people's daily lives. Since recommendations involve allocations of social resources (e.g., job recommendation), an important issue is whether recommendations are fair. Unfair recommendations are not only unethical but also harm the long-term interests of the recommender system itself. As a result, fairness issues in recommender systems have recently attracted increasing attention. However, due to multiple complex resource allocation processes and various fairness definitions, the research on fairness in recommendation is scattered. To fill this gap, we review over 60 papers published in top conferences/journals, including TOIS, SIGIR, and WWW. First, we summarize fairness definitions in the recommendation and provide several views to classify fairness issues. Then, we review recommendation datasets and measurements in fairness studies and provide an elaborate taxonomy of fairness methods in the recommendation. Finally, we conclude this survey by outlining some promising future directions.
... • I particularly study the multi-sided exposure bias in recommender systems and its impact on different sides in the system. I perform sets of experiments using existing recommendation algorithms on different datasets and investigate the fairness and equality of exposure for items and suppliers (Chapter 4) [116,120]. ...
... The proposed technique works as a post-processing recommendation and is able to mitigate exposure bias for items and suppliers with negligible loss in recommendation accuracy. A comprehensive set of experiments on different datasets and comparison with various state-of-the-art baselines show the effectiveness of FairMatch algorithm on improving exposure fairness for items and suppliers in recommender systems (Chapter 7) [118,120]. ...
... I show that existing metrics cannot properly measure the exposure bias and hide important aspects of measuring exposure bias. I propose several metrics that can better measure the effectiveness of an algorithm in mitigating the exposure bias for items and suppliers (Chapter 5) [120]. ...
Fairness is a critical system-level objective in recommender systems that has been the subject of extensive recent research. It is especially important in multi-sided recommendation platforms where it may be crucial to optimize utilities not just for the end user, but also for other actors such as item sellers or producers who desire a fair representation of their items. Existing solutions do not properly address various aspects of multi-sided fairness in recommendations as they may either solely have one-sided view (i.e. improving the fairness only for one side), or do not appropriately measure the fairness for each actor involved in the system. In this thesis, I aim at first investigating the impact of unfair recommendations on the system and how these unfair recommendations can negatively affect major actors in the system. Then, I seek to propose solutions to tackle the unfairness of recommendations. I propose a rating transformation technique that works as a pre-processing step before building the recommendation model to alleviate the inherent popularity bias in the input data and consequently to mitigate the exposure unfairness for items and suppliers in the recommendation lists. Also, as another solution, I propose a general graph-based solution that works as a post-processing approach after recommendation generation for mitigating the multi-sided exposure bias in the recommendation results. For evaluation, I introduce several metrics for measuring the exposure fairness for items and suppliers, and show that these metrics better capture the fairness properties in the recommendation results. I perform extensive experiments to evaluate the effectiveness of the proposed solutions. The experiments on different publicly-available datasets and comparison with various baselines confirm the superiority of the proposed solutions in improving the exposure fairness for items and suppliers.
... At each of these stages, new biases can be introduced [5,52]. This circular behavior can complicate the process of bias recognition and mitigation [5,53]. ...
... Bias and fairness Understanding, detection, and evaluation of bias and/or fairness in ML [26,[29][30][31][32][33][34][35]38,40] in ML Fairness in information networks [27] Bias management in ML [10] Bias and fairness Understanding, detection, and evaluation of bias and/or fairness in GNNs [25,41,43,45] in GNNs Bias and/or fairness mitigation in GNNs [46][47][48][49] Bias and fairness Understanding, detection, and evaluation of bias and/or fairness in RS [5,16,50,52] in RS Bias and/or fairness mitigation in RS [4,7,51,54,55,[60][61][62][63][64][65][66]69] Bias and fairness Understanding, detection, and evaluation of bias and/or fairness in GNN-based RS [18,20,21,23,56,57] in GNN-based RS Bias and/or fairness mitigation in GNN-based RS [22,53,58,67,68,[70][71][72][73] ...
In today’s technology-driven society, many decisions are made based on the results provided by machine learning algorithms. It is widely known that the models generated by such algorithms may present biases that lead to unfair decisions for some segments of the population, such as minority or marginalized groups. Hence, there is concern about the detection and mitigation of these biases, which may increase the discriminatory treatments of some demographic groups. Recommender systems, used today by millions of users, are not exempt from this drawback. The influence of these systems on so many user decisions, which in turn are taken as the basis for future recommendations, contributes to exacerbating this problem. Furthermore, there is evidence that some of the most recent and successful recommendation methods, such as those based on graphical neural networks (GNNs), are more sensitive to bias. The evaluation approaches of some of these biases, as those involving protected demographic groups, may not be suitable for recommender systems since their results are the preferences of the users and these do not necessarily have to be the same for the different groups. Other assessment metrics are aimed at evaluating biases that have no impact on the user. In this work, the suitability of different user-centered bias metrics in the context of GNN-based recommender systems are analyzed, as well as the response of recommendation methods with respect to the different types of biases to which these measures are addressed.
... This circular behavior of biases increases the complexity to recognize where actions are needed. Exposure bias happens due to the exposure of specific parts of items for users achieved from implicit feedback, and it can also be caused by popularity bias due to the recommendation of the most popular items [2,40]. In other words, bias can result in the limitation of the users' choices and contaminate users' feedback, which can amplify exposure bias [41]. ...
Recommender Systems (RSs) are used to provide users with personalized item recommendations and help them overcome the problem of information overload. Currently, recommendation methods based on deep learning are gaining ground over traditional methods such as matrix factorization due to their ability to represent the complex relationships between users and items and to incorporate additional information. The fact that these data have a graph structure and the greater capability of Graph Neural Networks (GNNs) to learn from these structures has led to their successful incorporation into recommender systems. However, the bias amplification issue needs to be investigated while using these algorithms. Bias results in unfair decisions, which can negatively affect the company reputation and financial status due to societal disappointment and environmental harm. In this paper, we aim to comprehensively study this problem through a literature review and an analysis of the behavior against biases of different GNN-based algorithms compared to state-of-the-art methods. We also intend to explore appropriate solutions to tackle this issue with the least possible impact on the model performance.
... Definition: Users tend to interact with items in higher position of the recommendation list regardless of the items' actual relevance (Collins et al., 2018). Impact: Position bias makes the evaluation of recommender systems challenging, because it also affects how the users interact with a series of recommendations (Mansoury, 2021). Tracking a user's clicks on a series of recommendations can often be used to infer the relevance of a set of recommendations. ...
With the remarkable increase in the number of scientific entities such as publications, researchers, and scientific topics, and the associated information overload in science, academic recommender systems have become increasingly important for millions of researchers and science enthusiasts. However, it is often overlooked that these systems are subject to various biases. In this article, we first break down the biases of academic recommender systems and characterize them according to their impact and prevalence. In doing so, we distinguish between biases originally caused by humans and biases induced by the recommender system. Second, we provide an overview of methods that have been used to mitigate these biases in the scholarly domain. Based on this, third, we present a framework that can be used by researchers and developers to mitigate biases in scholarly recommender systems and to evaluate recommender systems fairly. Finally, we discuss open challenges and possible research directions related to scholarly biases.
... The various graph CF strategies described above have historically centered on the enhancement of system accuracy, but, actually, never focused on the recommendation fairness dimensions. Despite some recent graph-based approaches have specifically been designed to address C-fairness [11,17,27,40,41,48] and P-fairness [6,19,20,35,51,52], there is a notable knowledge gap in the literature about the effects of the state-of-the-art graph strategies on the three objectives of C-fairness, P-fairness, and system accuracy. This work intends to complement the previous research and provide answers to pending research problems such as how different graph models perform for the three evaluation objectives. ...
To date, graph collaborative filtering (CF) strategies have been shown to outperform pure CF models in generating accurate recommendations. Nevertheless, recent works have raised concerns about fairness and potential biases in the recommendation landscape since unfair recommendations may harm the interests of Consumers and Producers (CP). Acknowledging that the literature lacks a careful evaluation of graph CF on CP-aware fairness measures, we initially evaluated the effects on CP-aware fairness measures of eight state-of-the-art graph models with four pure CF recommenders. Unexpectedly, the observed trends show that graph CF solutions do not ensure a large item exposure and user fairness. To disentangle this performance puzzle, we formalize a taxonomy for graph CF based on the mathematical foundations of the different approaches. The proposed taxonomy shows differences in node representation and neighbourhood exploration as dimensions characterizing graph CF. Under this lens, the experimental outcomes become clear and open the doors to a multi-objective CP-fairness analysis.
... This circular behavior of biases increases the complexity to recognize where actions are needed. Exposure bias happens due to the exposure of specific parts of items for users achieved from implicit feedback, and it can also be caused by popularity bias due to the recommendation of the most popular items [2,40]. In other words, bias can result in the limitation of the users' choices and contaminate users' feedback, which can amplify exposure bias [41]. ...
Recommender Systems (RSs) are used to provide users with personalized item recommendations and help them overcome the problem of information overload. Currently, recommendation methods based on deep learning are gaining ground over traditional methods such as matrix factorization due to their ability to represent the complex relationships between users and items and to incorporate additional information. The fact that these data have a graph structure and the greater capability of Graph Neural Networks (GNNs) to learn from these structures has led to their successful incorporation into recommender systems. However, the bias amplification issue needs to be investigated while using these algorithms. Bias results in unfair decisions, which can negatively affect the company’s reputation and financial status due to societal disappointment and environmental harm. In this paper, we aim to comprehensively study this problem through a literature review and an analysis of the behavior against biases of different GNN-based algorithms compared to state-of-the-art methods. We also intend to explore appropriate solutions to tackle this issue with the least possible impact on the model’s performance.
... Specifically, with this popularity bias, users receive more recommendations for popular items, and non-popular items receive less exposure than popular items. This disparate, i.e., unfair, treatment of popular and non-popular items has been thoroughly discussed in recent research [49,57,58,62]. In general, privacy and fairness are seen as two fundamental challenges of recommender systems research [22,63]. ...
User-based KNN recommender systems (UserKNN ) utilize the rating data of a target user's k nearest neighbors in the recommendation process. This, however, increases the privacy risk of the neighbors since their rating data might be exposed to other users or malicious parties. To reduce this risk, existing work applies differential privacy by adding randomness to the neighbors' ratings, which reduces the accuracy of UserKNN. In this work, we introduce ReuseKNN, a novel differentially-private KNN-based recommender system. The main idea is to identify small but highly reusable neighborhoods so that (i) only a minimal set of users requires protection with differential privacy, and (ii) most users do not need to be protected with differential privacy, since they are only rarely exploited as neighbors. In our experiments on five diverse datasets, we make two key observations: Firstly, ReuseKNN requires significantly smaller neighborhoods, and thus, fewer neighbors need to be protected with differential privacy compared to traditional UserKNN. Secondly, despite the small neighborhoods, ReuseKNN outperforms UserKNN and a fully differentially private approach in terms of accuracy. Overall, ReuseKNN leads to significantly less privacy risk for users than in the case of UserKNN
... Alternatively, due to the unavailability of a directly suitable public data set for MSRS research, the researchers have modified the real-world data set to incorporate the stakeholders' information. For example, the recent MSRS has crawled the IMDB to append the ML-100 K and ML-1 M (Harper and Konstan, 2015) data set with producer information (Abdollahpouri et al., 2020;Mansoury et al., 2021). The recent research indicates the popularity of ML-100 K and ML-1 M data set in MSRS research. ...
Purpose
In a multi-stakeholder recommender system (MSRS), stakeholders are the multiple entities (consumer, producer, system, etc.) benefited by the generated recommendations. Traditionally, the exclusive focus on only a single stakeholders' (for example, only consumer or end-user) preferences obscured the welfare of the others. Two major challenges are encountered while incorporating the multiple stakeholders' perspectives in MSRS: designing a dedicated utility function for each stakeholder and optimizing their utility without hurting others. This paper proposes multiple utility functions for different stakeholders and optimizes these functions for generating balanced, personalized recommendations for each stakeholder.
Design/methodology/approach
The proposed methodology considers four valid stakeholders user, producer, cast and recommender system from the multi-stakeholder recommender setting and builds dedicated utility functions. The utility function for users incorporates enhanced side-information-based similarity computation for utility count. Similarly, to improve the utility gain, the authors design new utility functions for producer, star-cast and system to incorporate long-tail and diverse items in the recommendation list. Next, to balance the utility gain and generate the trade-off recommendation solution, the authors perform the evolutionary optimization of the conflicting utility functions using NSGA-II. Experimental evaluation and comparison are conducted over three benchmark data sets.
Findings
The authors observed 19.70% of average enhancement in utility gain with improved mean precision, diversity and novelty. Exposure, hit, reach and target reach metrics are substantially improved.
Originality/value
A new approach considers four stakeholders simultaneously with their respective utility functions and establishes the trade-off recommendation solution between conflicting utilities of the stakeholders.
... In this way, recommendation system and users form a feedback loop [16,27]. Various research efforts have studied the exposure bias in offline recommendation algorithms [22,23,28], but less work has been done on interactive recommendation algorithms such as online learning to rank algorithms where the system iteratively learns from users' interaction feedback by balancing exploration versus exploitation [7]. ...
Exposure bias is a well-known issue in recommender systems where items and suppliers are not equally represented in the recommendation results. This is especially problematic when bias is amplified over time as a few popular items are repeatedly over-represented in recommendation lists. This phenomenon can be viewed as a recommendation feedback loop: the system repeatedly recommends certain items at different time points and interactions of users with those items will amplify bias towards those items over time. This issue has been extensively studied in the literature on model-based or neighborhood-based recommendation algorithms, but less work has been done on online recommendation models such as those based on multi-armed Bandit algorithms. In this paper, we study exposure bias in a class of well-known bandit algorithms known as Linear Cascade Bandits. We analyze these algorithms on their ability to handle exposure bias and provide a fair representation for items and suppliers in the recommendation results. Our analysis reveals that these algorithms fail to treat items and suppliers fairly and do not sufficiently explore the item space for each user. To mitigate this bias, we propose a discounting factor and incorporate it into these algorithms that controls the exposure of items at each time step. To show the effectiveness of the proposed discounting factor on mitigating exposure bias, we perform experiments on two datasets using three cascading bandit algorithms and our experimental results show that the proposed method improves the exposure fairness for items and suppliers.
To date, graph collaborative filtering (CF) strategies have been shown to outperform pure CF models in generating accurate recommendations. Nevertheless, recent works have raised concerns about fairness and potential biases in the recommendation landscape since unfair recommendations may harm the interests of Consumers and Producers (CP). Acknowledging that the literature lacks a careful evaluation of graph CF on CP-aware fairness measures, we initially evaluated the effects on CP-aware fairness measures of eight state-of-the-art graph models with four pure CF recommenders. Unexpectedly, the observed trends show that graph CF solutions do not ensure a large item exposure and user fairness. To disentangle this performance puzzle, we formalize a taxonomy for graph CF based on the mathematical foundations of the different approaches. The proposed taxonomy shows differences in node representation and neighbourhood exploration as dimensions characterizing graph CF. Under this lens, the experimental outcomes become clear and open the doors to a multi-objective CP-fairness analysis (Codes are available at: https://github.com/sisinflab/ECIR2023-Graph-CF.).KeywordsGraph collaborative filteringFairnessMulti-objective analysis
Conversational recommendation system (CRS) aims at recommending appropriate items to the user through a multi-turn conversation. The end-to-end CRS is a type of CRS that models the recommendation task and the conversation task simultaneously which has attracted more and more attention in recent years. At the same time, knowledge graph and Transformer are incorporated into the end-to-end CRS to generate better recommendations and better responses to the user, which makes the CRS have state-of-the-art performance. It is known that there exist semantic relations in a conversation. However, we observe that existing end-to-end CRSs in general ignore the semantic relations in the conversation and therefore would likely hinder the performance of CRSs. Motivated by this, we propose a gated cross- and self-attention based CRS utilizing semantic relation information (ASR) model, which can explicitly model and utilize the semantic relations in a conversation. To the best of our knowledge, we are the first to advocate for modeling and utilizing the semantic relations in the end-to-end CRS, which could help to improve the performance of the CRS. Furthermore, to mitigate the class-imbalance problem that most end-to-end CRSs face, we propose a new negative sampling method which could make the proposed CRS learn better. Moreover, we design a Transformer-based dialogue module integrating the semantic relations in a conversation to generate more diversified and precise responses. Extensive experiments on widely used benchmark datasets demonstrate that the proposed ASR model achieves state-of-the-art results in both recommendation and conversation tasks.
Masoud Mansoury is a postdoctoral researcher at Amsterdam Machine Learning Lab at University of Amsterdam, Netherlands. He is also a member of Discovery Lab collaborating with Data Science team at Elsevier Company in the area of recommender systems. Masoud received his PhD in Computer and Information Science from Eindhoven University of Technology, Netherlands, in 2021. He has published his research works in top conferences such as FAccT, RecSys, and CIKM. His research interests include recommender systems, algorithmic bias, and contextual bandits.
This research conducted by Masoud Mansoury investigated the impact of unfair recommendations on different actors in the system and proposed solutions to tackle the unfairness of recommendations. The solutions were a rating transformation technique that works as a pre-processing step before recommendation generation and a general graph-based solution that works as a post-processing approach after recommendation generation for mitigating the multi-sided exposure bias in the recommendation results. For evaluation, he introduced several metrics for measuring the exposure fairness for items and suppliers, and showed that the proposed metrics better capture the fairness properties in the recommendation results. Extensive experiments on different publicly-available datasets confirmed the superiority of the proposed solutions in improving the exposure fairness for items and suppliers.
Social norms govern collective behaviour by guiding individual behaviour in the absence of a central enforcing authority, which makes them powerful self-regulating mechanisms for societies. This is in stark contrast to policy or legislative norms - also targeted at governing behaviour in society - which are issued by a central authority who also then needs to enforce compliance. Enforcing compliance is expensive. Also, these norms might come into conflict with existing social norms, which causes issues. It is therefore not surprising that much research is aimed at understanding existing social norms around behaviours connected to important issues like health or climate change. Designing policy that piggybacks on existing norms to promote behaviour is faster and cheaper than using the classic carrot-and-stick approach of most policy design. The modelling community has invested quite a bit of effort into developing normative frameworks, models and simulations. Yet, very little of this effort has been directed towards the study of the norm life-cycle. Besides, these research efforts have omitted explicit representation of norms and the assessment of norm stability and reactivity in the face of some environmental changes. Values as a stabilizing factor, must be considered while studying the reactivity and stability of social norms. Without such stabilizing elements, modeled norms react swiftly to any change in the environment and are mere behavioural patterns rather than social norms. In this thesis, I use values as drivers of behavioural choices, and Schwartz’s theory of abstract values as a basis. As these values are very abstract, there is a need to translate them to more concrete values and assign behavioural choices to them. A theory or methodology for this step has not been developed in a way that is widely applicable. Thus, a precise way of such a translation is necessary for practical purposes. I designed a practical but formal framework that can be used to study the value-driven behaviour of agents in social simulations. I showed how this formal design can be used in practice to implement multi-agent simulations. Then, I continued with proposing a social norm framework that is focused on finding an explanation for norm dynamics - their emergence, perpetuation, and eventual disappearance. I operationalized the framework by way of a multi-agent simulation in the context of environmental change and absence of sanctions for deviant behaviour. I showed that the values are an intuitive stabilizing factor that allow norms to persist through changes in the agents’ environment and perpetuate and spread even in the absence of punishment. A norm will, however, change, evolve or disappear altogether if it becomes impossible to perform or if the value priorities of the agents change. I explained the norm dynamic and its strong connection to values by implementing various multi-agent simulation scenarios.
As one of the most successful AI-powered applications, recommender systems aim to help people make appropriate decisions in an effective and efficient way, by providing personalized suggestions in many aspects of our lives, especially for various human-oriented online services such as e-commerce platforms and social media sites. In the past few decades, the rapid developments of recommender systems have significantly benefited human by creating economic value, saving time and effort, and promoting social good. However, recent studies have found that data-driven recommender systems can pose serious threats to users and society, such as spreading fake news to manipulate public opinion in social media sites, amplifying unfairness toward under-represented groups or individuals in job matching services, or inferring privacy information from recommendation results. Therefore, systems' trustworthiness has been attracting increasing attention from various aspects for mitigating negative impacts caused by recommender systems, so as to enhance the public's trust towards recommender systems techniques. In this survey, we provide a comprehensive overview of Trustworthy Recommender systems (TRec) with a specific focus on six of the most important aspects; namely, Safety & Robustness, Nondiscrimination & Fairness, Explainability, Privacy, Environmental Well-being, and Accountability & Auditability. For each aspect, we summarize the recent related technologies and discuss potential research directions to help achieve trustworthy recommender systems in the future.
Recommender systems are an essential tool to relieve the information overload challenge and play an important role in people’s daily lives. Since recommendations involve allocations of social resources (e.g., job recommendation), an important issue is whether recommendations are fair. Unfair recommendations are not only unethical but also harm the long-term interests of the recommender system itself. As a result, fairness issues in recommender systems have recently attracted increasing attention. However, due to multiple complex resource allocation processes and various fairness definitions, the research on fairness in recommendation is scattered. To fill this gap, we review over 60 papers published in top conferences/journals, including TOIS, SIGIR, and WWW. First, we summarize fairness definitions in the recommendation and provide several views to classify fairness issues. Then, we review recommendation datasets and measurements in fairness studies and provide an elaborate taxonomy of fairness methods in the recommendation. Finally, we conclude this survey by outlining some promising future directions.
Academic research in recommender systems has been greatly focusing on the accuracy-related measures of recommendations. Even when non-accuracy measures such as popularity bias, diversity, and novelty are studied, it is often solely from the users' perspective. However, many real-world recommenders are often multi-stakeholder environments in which the needs and interests of several stakeholders should be addressed in the recommendation process. In this paper, we focus on the popularity bias problem which is a well-known property of many recommendation algorithms where few popular items are over-recommended while the majority of other items do not get proportional attention and address its impact on different stakeholders. Using several recommendation algorithms and two publicly available datasets in music and movie domains, we empirically show the inherent popularity bias of the algorithms and how this bias impacts different stakeholders such as users and suppliers of the items. We also propose metrics to measure the exposure bias of recommendation algorithms from the perspective of different stakeholders.
Research has shown that recommender systems are typically biased towards popular items, which leads to less popular items being underrepresented in recommendations. The recent work of Abdollahpouri et al. in the context of movie recommendations has shown that this popularity bias leads to unfair treatment of both long-tail items as well as users with little interest in popular items. In this paper, we reproduce the analyses of Abdollahpouri et al. in the context of music recommendation. Specifically, we investigate three user groups from the Last.fm music platform that are categorized based on how much their listening preferences deviate from the most popular music among all Last.fm users in the dataset: (i) low-mainstream users, (ii) medium-mainstream users, and (iii) high-mainstream users. In line with Abdollahpouri et al., we find that state-of-the-art recommendation algorithms favor popular items also in the music domain. However, their proposed Group Average Popularity metric yields different results for Last.fm than for the movie domain, presumably due to the larger number of available items (i.e., music artists) in the Last.fm dataset we use. Finally, we compare the accuracy results of the recommendation algorithms for the three user groups and find that the low-mainstreaminess group significantly receives the worst recommendations.
Matching algorithms are used routinely to match donors to recipients for solid organs transplantation, for the assignment of medical residents to hospitals, record linkage in databases, scheduling jobs on machines, network switching, online advertising, and image recognition, among others. Although many optimal solutions may exist to a given matching problem, when the elements that shall or not be included in a solution correspond to individuals, it becomes of paramount importance that the solution is selected fairly. In this paper we study individual fairness in matching problems. Given that many maximum matchings may exist, each one satisfying a different set of individuals, the only way to guarantee fairness is through randomization. Hence we introduce the distributional maxmin fairness framework which provides, for any given input instance, the strongest guarantee possible simultaneously for all individuals in terms of satisfaction probability (the probability of being matched in the solution). Specifically, a probability distribution over feasible solutions is maxmin-fair if it is not possible to improve the satisfaction probability of any individual without decreasing it for some other individual which is no better off. Our main contribution is a polynomial-time algorithm building on techniques from minimum cuts, and edge-coloring algorithms for regular bipartite graphs, and transversal theory. In the special case of bipartite matching, our algorithm runs in \(O((|V|^2 + |E| |V|^{2/3})\cdot (\log |V|)^2)\) expected time. An experimental evaluation of our fair-matching algorithm shows its ability to scale to graphs with tens of millions of vertices and hundreds of millions of edges, taking only a few minutes on a simple architecture. To the best of our knowledge, this yields the first large-scale implementation of the egalitarian mechanism of Bogomolnaia and Moulin (Econometrica 72(1):257–279, 2004). Our analysis confirms that our method provides stronger satisfaction probability guarantees than non-trivial baselines.
Recommender systems provide personalized information access to users of Internet services from social networks to e-commerce to media and entertainment. As is appropriate for research in a field with a focus on personalization, academic studies of recommender systems have largely concentrated on optimizing for user experience when designing, implementing and evaluating their algorithms and systems. However, this concentration on the user has meant that the field has lacked a systematic exploration of other aspects of recommender system outcomes. A user-centric approach limits the ability to incorporate system objectives, such as fairness, balance, and profitability, and obscures concerns that might come from other stakeholders, such as the providers or sellers of items being recommended. Multistakeholder recommendation has emerged as a unifying framework for describing and understanding recommendation settings where the end user is not the sole focus. This article outlines the multistakeholder perspective on recommendation, highlighting example research areas and discussing important issues, open questions, and prospective research directions.
Recommender Systems (RSs) are widely used to help online users discover products, books, news, music, movies, courses, restaurants, etc. Because a traditional recommendation strategy always shows the most relevant items (thus with highest predicted rating), traditional RS’s are expected to make popular items become even more popular and non-popular items become even less popular which in turn further divides the haves (popular) from the have-nots (unpopular). Therefore, a major problem with RSs is that they may introduce biases affecting the exposure of items, thus creating a popularity divide of items during the feedback loop that occurs with users, and this may lead the RS to make increasingly biased recommendations over time. In this paper, we view the RS environment as a chain of events that are the result of interactions between users and the RS. Based on that, we propose several debiasing algorithms during this chain of events, and evaluate how these algorithms impact the predictive behavior of the RS, as well as trends in the popularity distribution of items over time. We also propose a novel blind-spot-aware matrix factorization (MF) algorithm to debias the RS. Results show that propensity matrix factorization achieved a certain level of debiasing of the RS while active learning combined with the propensity MF achieved a higher debiasing effect on recommendations.
Due to the complexity of recommendation algorithms, experimentation on recommender systems has become a challenging task. Current recommendation algorithms, while powerful, involve large numbers of hyperparameters. Tuning hyperparameters for finding the best recommendation outcome often requires execution of large numbers of algorithmic experiments particularly when multiples evaluation metrics are considered. Existing recommender systems platforms fail to provide a basis for systematic experimentation of this type. In this paper, we describe librec-auto, a wrapper for the well-known LibRec library, which provides an environment that supports automated experimentation.
Many recommender systems suffer from popularity bias: popular items are recommended frequently while less popular , niche products, are recommended rarely or not at all. However, recommending the ignored products in the "long tail" is critical for businesses as they are less likely to be discovered. In this paper, we introduce a personalized diversification re-ranking approach to increase the representation of less popular items in recommendations while maintaining acceptable recommendation accuracy. Our approach is a post-processing step that can be applied to the output of any recommender system. We show that our approach is capable of managing popularity bias more effectively, compared with an existing method based on regularization. We also examine both new and existing metrics to measure the coverage of long-tail items in the recommendation.
Recommender systems are typically designed to optimize the utility of the end user. In many settings, however, the end user is not the only stakeholder and this exclusive focus may produce unsatisfactory results for other stakeholders. One such setting is found in multisided platforms, which bring together buyers and sellers. In such platforms, it may be necessary to jointly optimize the value for both buyers and sellers. This paper proposes a constraint-based integer programming optimization model, in which different sets of constraints are used to reflect the goals of the different stakeholders. This model is applied as a post-processing step, so it can easily be added onto an existing recommendation system to make it multi-stakeholder aware. For computational tractability with larger data sets, we reformulate the integer problem using the Lagrangian dual and use subgradient optimization. In experiments with two data sets, we evaluate empirically the interaction between the utilities of buyers and sellers and show that our approximation can achieve good upper and lower bounds in practical situations.
Rankings are ubiquitous in the online world today. As we have transitioned from finding books in libraries to ranking products, jobs, job applicants, opinions and potential romantic partners, there is a substantial precedent that ranking systems have a responsibility not only to their users but also to the items being ranked. To address these often conflicting responsibilities, we propose a conceptual and computational framework that allows the formulation of fairness constraints on rankings in terms of exposure allocation. As part of this framework, we develop efficient algorithms for finding rankings that maximize the utility for the user while provably satisfying a specifiable notion of fairness. Since fairness goals can be application specific, we show how a broad range of fairness constraints can be implemented using our framework, including forms of demographic parity, disparate treatment, and disparate impact constraints. We illustrate the effect of these constraints by providing empirical results on two ranking problems.
In the debate about filter bubbles caused by algorithmic news recommendation, the conceptualization of the two core concepts in this debate, diversity and algorithms, has received little attention in social scientific research. This paper examines the effect of multiple recommender systems on different diversity dimensions. To this end, it maps different values that diversity can serve, and a respective set of criteria that characterizes a diverse information offer in this particular conception of diversity. We make use of a data set of simulated article recommendations based on actual content of one of the major Dutch broadsheet newspapers and its users (N=21,973 articles, N=500 users). We find that all of the recommendation logics under study proved to lead to a rather diverse set of recommendations that are on par with human editors and that basing recommendations on user histories can substantially increase topic diversity within a recommendation set.
In this work, we define and solve the Fair Top-k Ranking problem, in which we want to determine a subset of k candidates from a large pool of n » k candidates, maximizing utility (i.e., select the "best" candidates) subject to group fairness criteria.
Our ranked group fairness definition extends group fairness using the standard notion of protected groups and is based on ensuring that the proportion of protected candidates in every prefix of the top-k ranking remains statistically above or indistinguishable from a given minimum. Utility is operationalized in two ways: (i) every candidate included in the top-k should be more qualified than every candidate not included; and (ii) for every pair of candidates in the top-k, the more qualified candidate should be ranked above.
An efficient algorithm is presented for producing the Fair Top-k Ranking, and tested experimentally on existing datasets as well as new datasets released with this paper, showing that our approach yields small distortions with respect to rankings that maximize utility without considering fairness criteria. To the best of our knowledge, this is the first algorithm grounded in statistical tests that can mitigate biases in the representation of an under-represented group along a ranked list.
Many recommendation algorithms suffer from popularity bias in their output: popular items are recommended frequently and less popular ones rarely, if at all. However, less popular, long-tail items are precisely those that are often desirable recommendations. In this paper, we introduce a flexible regularization-based framework to enhance the long-tail coverage of recommendation lists in a learning-to-rank algorithm. We show that regularization provides a tunable mechanism for controlling the trade-off between accuracy and coverage. Moreover, the experimental results using two data sets show that it is possible to improve coverage of long tail items without substantial loss of ranking performance.
Recommender systems are becoming increasingly important and prevalent because of the ability of solving information overload. In recent years, researchers are paying increasing attention to aggregate diversity as a key metric beyond accuracy, because improving aggregate recommendation diversity may increase long tails and sales diversity. Trust is often used to improve recommendation accuracy. However, how to utilize trust to improve aggregate recommendation diversity is unexplored. In this paper, we focus on solving this problem and propose a novel trust-aware recommendation method by incorporating time factor into similarity computation. The rationale underlying the proposed method is that, trustees with later creation time of trust relation can bring more diverse items to recommend to their trustors than other trustees with earlier creation time of trust relation. Through relevant experiments on publicly available dataset, we demonstrate that the proposed method outperforms the baseline method in terms of aggregate diversity while maintaining almost the same recall.
Recommender system aim at providing a personalized list of items ranked according to the preferences of the user, as such ranking methods are at the core of many recommendation algorithms. The topic of this tutorial focuses on the cutting-edge algorithmic development in the area of recommender systems. This tutorial will provide an in depth picture of the progress of ranking models in the field, summarizing the strengths and weaknesses of existing methods, and discussing open issues that could be promising for future research in the community. A qualitative and quantitative comparison between different models will be provided while we will also highlight recent developments in the areas of Reinforcement Learning.
Fairness in recommendation has attracted increasing attention due to bias and discrimination possibly caused by traditional recommenders. In Interactive Recommender Systems (IRS), user preferences and the system’s fairness status are constantly changing over time. Existing fairness-aware recommenders mainly consider fairness in static settings. Directly applying existing methods to IRS will result in poor recommendation. To resolve this problem, we propose a reinforcement learning based framework, FairRec, to dynamically maintain a long-term balance between accuracy and fairness in IRS. User preferences and the system’s fairness status are jointly compressed into the state representation to generate recommendations. FairRec aims at maximizing our designed cumulative reward that combines accuracy and fairness. Extensive experiments validate that FairRec can improve fairness, while preserving good recommendation quality.
We present a framework for quantifying and mitigating algorithmic bias in mechanisms designed for ranking individuals, typically used as part of web-scale search and recommendation systems. We first propose complementary measures to quantify bias with respect to protected attributes such as gender and age. We then present algorithms for computing fairness-aware re-ranking of results. For a given search or recommendation task, our algorithms seek to achieve a desired distribution of top ranked results with respect to one or more protected attributes. We show that such a framework can be tailored to achieve fairness criteria such as equality of opportunity and demographic parity depending on the choice of the desired distribution. We evaluate the proposed algorithms via extensive simulations over different parameter choices, and study the effect of fairness-aware ranking on both bias and utility measures. We finally present the online A/B testing results from applying our framework towards representative ranking in LinkedIn Talent Search, and discuss the lessons learned in practice. Our approach resulted in tremendous improvement in the fairness metrics (nearly three fold increase in the number of search queries with representative results) without affecting the business metrics, which paved the way for deployment to 100% of LinkedIn Recruiter users worldwide. Ours is the first large-scale deployed framework for ensuring fairness in the hiring domain, with the potential positive impact for more than 630M LinkedIn members.
Two-sided marketplaces are platforms that have customers not only on the demand side (e.g. users), but also on the supply side (e.g. retailer, artists). While traditional recommender systems focused specifically towards increasing consumer satisfaction by providing relevant content to consumers, two-sided marketplaces face the problem of additionally optimizing for supplier preferences, and visibility. Indeed, the suppliers would want afair opportunity to be presented to users. Blindly optimizing for consumer relevance may have a detrimental impact on supplier fairness. Motivated by this problem, we focus on the trade-off between objectives of consumers and suppliers in the case of music streaming services, and consider the trade-off betweenrelevance of recommendations to the consumer (i.e. user) andfairness of representation of suppliers (i.e. artists) and measure their impact on consumersatisfaction.
We propose a conceptual and computational framework using counterfactual estimation techniques to understand, and evaluate different recommendation policies, specifically around the trade-off between relevance and fairness, without the need for running many costly A/B tests. We propose a number of recommendation policies which jointly optimize relevance and fairness, thereby achieving substantial improvement in supplier fairness without noticeable decline in user satisfaction. Additionally, we consider user disposition towards fair content, and propose a personalized recommendation policy which takes into account consumer's tolerance towards fair content. Our findings could guide the design of algorithms powering two-sided marketplaces, as well as guide future research on sophisticated algorithms for joint optimization of user relevance, satisfaction and fairness.
Recommendation systems are ubiquitous and impact many domains; they have the potential to influence product consumption, individuals' perceptions of the world, and life-altering decisions. These systems are often evaluated or trained with data from users already exposed to algorithmic recommendations; this creates a pernicious feedback loop. Using simulations, we demonstrate how using data confounded in this way homogenizes user behavior without increasing utility.
Rankings of people and items are at the heart of selection-making, match-making, and recommender systems, ranging from employment sites to sharing economy platforms. As ranking positions influence the amount of attention the ranked subjects receive, biases in rankings can lead to unfair distribution of opportunities and resources, such as jobs or income. This paper proposes new measures and mechanisms to quantify and mitigate unfairness from a bias inherent to all rankings, namely, the position bias, which leads to disproportionately less attention being paid to low-ranked subjects. Our approach differs from recent fair ranking approaches in two important ways. First, existing works measure unfairness at the level of subject groups while our measures capture unfairness at the level of individual subjects, and as such subsume group unfairness. Second, as no single ranking can achieve individual attention fairness, we propose a novel mechanism that achieves amortized fairness, where attention accumulated across a series of rankings is proportional to accumulated relevance. We formulate the challenge of achieving amortized individual fairness subject to constraints on ranking quality as an online optimization problem and show that it can be solved as an integer linear program. Our experimental evaluation reveals that unfair attention distribution in rankings can be substantial, and demonstrates that our method can improve individual fairness while retaining high ranking quality.
Collaborative filtering is a broad and powerful framework for building recommendation systems that has seen widespread adoption. Over the past decade, the propensity of such systems for favoring popular products and thus creating echo chambers have been observed. This has given rise to an active area of research that seeks to diversify recommendations generated by such algorithms. We address the problem of increasing diversity in recom- mendation systems that are based on collaborative filtering that use past ratings to predict a rating quality for potential recommendations. Following our earlier work, we formulate recommendation system design as a subgraph selection problem from a candidate super-graph of potential recommendations where both diversity and rating quality are explicitly optimized: (1) On the modeling side, we define a new flexible notion of diversity that allows a system designer to prescribe the number of recommendations each item should receive, and smoothly penalizes deviations from this distribution. (2) On the algorithmic side, we show that minimum-cost network flow methods yield fast algorithms in theory and practice for designing recommendation subgraphs that optimize this notion of diversity. (3) On the empirical side, we show the effectiveness of our new model and method to increase diversity while maintaining high rating quality in standard rating data sets from Netflix and MovieLens.
In recent years, deep neural networks have yielded immense success on speech recognition, computer vision and natural language processing. However, the exploration of deep neural networks on recommender systems has received relatively less scrutiny. In this work, we strive to develop techniques based on neural networks to tackle the key problem in recommendation --- collaborative filtering --- on the basis of implicit feedback.
Although some recent work has employed deep learning for recommendation, they primarily used it to model auxiliary information, such as textual descriptions of items and acoustic features of musics. When it comes to model the key factor in collaborative filtering --- the interaction between user and item features, they still resorted to matrix factorization and applied an inner product on the latent features of users and items.
By replacing the inner product with a neural architecture that can learn an arbitrary function from data, we present a general framework named NCF, short for Neural network-based Collaborative Filtering. NCF is generic and can express and generalize matrix factorization under its framework. To supercharge NCF modelling with non-linearities, we propose to leverage a multi-layer perceptron to learn the user-item interaction function. Extensive experiments on two real-world datasets show significant improvements of our proposed NCF framework over the state-of-the-art methods. Empirical evidence shows that using deeper layers of neural networks offers better recommendation performance.
What makes a good recommendation or good list of recommendations?
Research into recommender systems has traditionally focused on accuracy, in particular how closely the recommender’s predicted ratings are to the users’ true ratings. However, it has been recognized that other recommendation qualities—such as whether the list of recommendations is diverse and whether it contains novel items—may have a significant impact on the overall quality of a recommender system. Consequently, in recent years, the focus of recommender systems research has shifted to include a wider range of “beyond accuracy” objectives.
In this article, we present a survey of the most discussed beyond-accuracy objectives in recommender systems research: diversity, serendipity, novelty, and coverage. We review the definitions of these objectives and corresponding metrics found in the literature. We also review works that propose optimization strategies for these beyond-accuracy objectives. Since the majority of works focus on one specific objective, we find that it is not clear how the different objectives relate to each other.
Hence, we conduct a set of offline experiments aimed at comparing the performance of different optimization approaches with a view to seeing how they affect objectives other than the ones they are optimizing. We use a set of state-of-the-art recommendation algorithms optimized for recall along with a number of reranking strategies for optimizing the diversity, novelty, and serendipity of the generated recommendations. For each reranking strategy, we measure the effects on the other beyond-accuracy objectives and demonstrate important insights into the correlations between the discussed objectives. For instance, we find that rating-based diversity is positively correlated with novelty, and we demonstrate the positive influence of novelty on recommendation coverage.
We present the LFM-1b dataset of more than one billion music listening events created by more than 120,000 users of Last.fm. Each listening event is characterized by artist, album, and track name, and further includes a timestamp. On the (anonymous) user level, basic demographics and a selection of more elaborate user descriptors are included.
The dataset is foremost intended for benchmarking in music information retrieval and recommendation. To facilitate experimentation in a straightforward manner, it also includes a precomputed user-item-playcount matrix. In addition, sample Python scripts showing how to load the data and perform efficient computations are provided. An implementation of a simple collaborative filtering recommender rounds off the code package.
We discuss in detail the LFM-1b dataset's acquisition, availability, statistics, and content, and place it in the context of existing datasets. We also showcase its usage in a simple artist recommendation task, whose results are intended to serve as baseline against which more elaborate techniques can be assessed. The two unique features of the dataset in comparison to existing ones are (i) its substantial size and (ii) a wide range of additional user descriptors that reflect their music taste and consumption behavior.
Novelty and diversity have been identified, along with accuracy, as foremost properties of useful recommendations. Considerable progress has been made in the field in terms of the definition of methods to enhance such properties, as well as methodologies and metrics to assess how well such methods work. In this chapter we give an overview of the main contributions to this area in the field of recommender systems, and seek to relate them together in a unified view, analyzing the common elements underneath the different forms under which novelty and diversity have been addressed, and identifying connections to closely related work on diversity in other fields.
Recommender systems are being used to help users find relevant items from a large set of alternatives in many online applications. Most existing recommendation techniques have focused on improving recommendation accuracy; however, diversity of recommendations has also been increasingly recognized in research literature as an important aspect of recommendation quality. This paper proposes a graph-theoretic approach for maximizing aggregate recommendation diversity based on maximum flow or maximum bipartite matching computations. The proposed approach is evaluated using real-world movie rating datasets and demonstrates substantial improvements in both diversity and accuracy, as compared to the recommendation re-ranking approaches, which have been introduced in prior literature for the purpose of diversity improvement.
Sales diversity is considered a key feature of Recommender Systems from a business perspective. Sales diversity is also linked with the long-tail novelty of recommendations, a quality dimension from the user perspective. We explore the inversion of the recommendation task as a means to enhance sales diversity - and indirectly novelty - by selecting which users an item should be recommended to instead of the other way around. We address the inverted task by two approaches: a) inverting the rating matrix, and b) defining a probabilistic reformulation which isolates the popularity component of arbitrary recommendation algorithms. We find that the first approach gives rise to interesting reformulations of nearest-neighbor algorithms, which essentially introduce a new neighbor selection policy. The second approach, as well as the first, ultimately result in substantial sales diversity enhancements, and improved trade-offs with recommendation precision and novelty. Two experiments on movie and music recommendation datasets show the effectiveness of the resulting approach, even when compared to direct optimization approaches of the target metrics proposed in prior work.
The MovieLens datasets are widely used in education, research, and industry. They are downloaded hundreds of thousands of times each year, reflecting their use in popular press programming books, traditional and online courses, and software. These datasets are a product of member activity in the MovieLens movie recommendation system, an active research platform that has hosted many experiments since its launch in 1997. This article documents the history of MovieLens and the MovieLens datasets. We include a discussion of lessons learned from running a long-standing, live research platform from the perspective of a research organization. We document best practices and limitations of using the MovieLens datasets in new research.
Focusing on the problems of over-specialization and concen- tration bias, this paper presents a novel probabilistic method for recommending items in the neighborhood-based collab- orative filtering framework. For the probabilistic neighbor- hood selection phase, we use an efficient method for weighted sampling of k neighbors that takes into consideration the similarity levels between the target user (or item) and the candidate neighbors. We conduct an empirical study show- ing that the proposed method increases the coverage, dis- persion, and diversity reinforcement of recommendations by selecting diverse sets of representative neighbors. We also demonstrate that the proposed approach outperforms pop- ular methods in terms of item prediction accuracy, utility- based ranking, and other popular measures, across various experimental settings. This performance improvement is in accordance with ensemble learning theory and the phe- nomenon of \hubness" in recommender systems.
Introduction. The problem discussed in this paper was formulated by T. Harris as follows:
“Consider a rail network connecting two cities by way of a number of intermediate cities, where each link of the network has a number assigned to it representing its capacity. Assuming a steady state condition, find a maximal flow from one given city to the other.”
We present an approach to control information flow in object-oriented systems. The decision of whether an informatin flow is permitted or denied depends on both the authorizations specified on the objects and the process by which information is obtained ...
Recommender systems are becoming increasingly important to individual users and businesses for providing personalized recommendations. However, while the majority of algorithms proposed in recommender systems literature have focused on improving recommendation accuracy (as exemplified by the recent Netflix Prize competition), other important aspects of recommendation quality, such as the diversity of recommendations, have often been overlooked. In this paper, we introduce and explore a number of item ranking techniques that can generate recommendations that have substantially higher aggregate diversity across all users while maintaining comparable levels of recommendation accuracy. Comprehensive empirical evaluation consistently shows the diversity gains of the proposed techniques using several real-world rating datasets and different rating prediction algorithms.
The Recommender Systems community is paying increasing attention to novelty and diversity as key qualities beyond accuracy in real recommendation scenarios. Despite the raise of interest and work on the topic in recent years, we find that a clear common methodological and conceptual ground for the evaluation of these dimensions is still to be consolidated. Different evaluation metrics have been reported in the literature but the precise relation, distinction or equivalence between them has not been explicitly studied. Furthermore, the metrics reported so far miss important properties such as taking into consideration the ranking of recommended items, or whether items are relevant or not, when assessing the novelty and diversity of recommendations.
We present a formal framework for the definition of novelty and diversity metrics that unifies and generalizes several state of the art metrics. We identify three essential ground concepts at the roots of novelty and diversity: choice, discovery and relevance, upon which the framework is built. Item rank and relevance are introduced through a probabilistic recommendation browsing model, building upon the same three basic concepts. Based on the combination of ground elements, and the assumptions of the browsing model, different metrics and variants unfold. We report experimental observations which validate and illustrate the properties of the proposed metrics.
Recommendations from the long tail of the popularity distribution of items are generally considered to be particularly valuable. On the other hand, recommendation accuracy tends to decrease towards the long tail. In this paper, we quantitatively examine this trade-off between item popularity and recommendation accuracy. To this end, we assume that there is a selection bias towards popular items in the available data. This allows us to define a new accuracy measure that can be gradually tuned towards the long tail. We show that, under this assumption, this measure has the desirable property of providing nearly unbiased estimates concerning recommendation accuracy. In turn, this also motivates a refinement for training collaborative-filtering approaches. In various experiments with real-world data, including a user study, empirical evidence suggests that only a small, if any, bias of the recommendations towards less popular items is appreciated by users.
When a Web user's underlying information need is not clearly specified from the initial query, an effective approach is to diversify the results retrieved for this query. In this paper, we introduce a novel probabilistic framework for Web search result diversification, which explicitly accounts for the various aspects associated to an underspecified query. In particular, we diversify a document ranking by estimating how well a given document satisfies each uncovered aspect and the extent to which different aspects are satisfied by the ranking as a whole. We thoroughly evaluate our framework in the context of the diversity task of the TREC 2009 Web track. Moreover, we exploit query reformulations provided by three major Web search engines (WSEs) as a means to uncover different query aspects. The results attest the effectiveness of our framework when compared to state-of-the-art diversification approaches in the literature. Additionally, by simulating an upper-bound query reformulation mechanism from official TREC data, we draw useful insights regarding the effectiveness of the query reformulations generated by the different WSEs in promoting diversity.
All previously known efficient maximum-flow algorithms work by finding augmenting paths, either one path at a time (as in the original Ford and Fulkerson algorithm) or all shortest-length augmenting paths at once (using the layered network approach of Dinic). An alternative method based on the preflow concept of Karzanov is introduced. A preflow is like a flow, except that the total amount flowing into a vertex is allowed to exceed the total amount flowing out. The method maintains a preflow in the original network and pushes local flow excess toward the sink along what are estimated to be shortest paths. The algorithm and its analysis are simple and intuitive, yet the algorithm runs as fast as any other known method on dense graphs, achieving an O(n3) time bound on an n-vertex graph. By incorporating the dynamic tree data structure of Sleator and Tarjan, we obtain a version of the algorithm running in O(nm log(n2/m)) time on an n-vertex, m-edge graph. This is as fast as any known method for any graph density and faster on graphs of moderate density. The algorithm also admits efficient distributed and parallel implementations. A parallel implementation running in O(n2log n) time using n processors and O(m) space is obtained. This time bound matches that of the Shiloach-Vishkin algorithm, which also uses n processors but requires O(n2) space.
Collaborative filers help people make choices based on the opinions of other people. GroupLens is a system for collaborative filtering of netnews, to help people find articles they will like in the huge stream of available articles. News reader clients display predicted scores and make it easy for users to rate articles after they read them. Rating servers, called Better Bit Bureaus, gather and disseminate the ratings. The rating servers predict scores based on the heuristic that people who agreed in the past will probably agree again. Users can protect their privacy by entering ratings under a pseudonym, without reducing the effectiveness of the score prediction. The entire architecture is open: alternative software for news clients and Better Bit Bureaus can be developed independently and can interoperate with the components we have developed.
Deconvolving feedback loops in recommender systems
- David F Ayan Sinha
- Karthik Gleich
- Ramani
LibRec: A Java library for recommender systems
- Guibing Guo
- Jie Zhang
- Zhu Sun
- Neil Yorke-Smith
- Guo Guibing
Modeling and Counteracting Exposure Bias in Recommender Systems
- Sami Khenissi
- Khenissi Sami
