A framework for unbiased explainable pairwise ranking for recommendation

Recent research has demonstrated the effectiveness of utilizing contrastive learning for training Transformer-based sequence encoders in sequential recommendation tasks. Items are represented using vectors and the relations between items are measured by the dot product self-attention, the feature representation in sequential recommendation can be enhanced. However, in real-world scenarios, user behavior sequences are unpredictable, and the limitations of dot product-based approaches hinder the complete capture of collaborative transferability. Moreover, the Bayesian personalized ranking (BPR) loss function, commonly utilized in recommendation systems, lacks constraints when considering positive and negative sampled items, potentially leading to suboptimal optimization outcomes. This presents a complex challenge that needs to be addressed. To tackle these issues, this article proposes a novel method involving stochastic self-attention. This article introduces uncertainty into the proposed model by utilizing elliptical Gaussian distribution controlled by mean and covariance vector to explain the unpredictability of items. At the same time, the proposed model combines a Wasserstein self-attention module to compute the positional relationships between items within a sequence in order to effectively incorporate uncertainty into the training process. The Wasserstein self-attention mechanism satisfies the triangular inequality and can not only addresses uncertainty but also promote collaborative transfer learning. Furthermore, embedding a stochastic Gaussian distribution into each item will bring additional uncertainty into the proposed model. Multi-pair contrastive learning relies on high-quality positive samples, and the proposed model combines the cloze task mask and dropout mask mechanisms to generate high-quality positive samples. It demonstrates superior performance and adaptability compared to traditional single-pair contrastive learning methods. Additionally, a dynamic loss reweighting strategy is introduced to balance the cloze task loss and the contrastive loss effectively. We conduct experiments and the results show that the proposed model outperforms the state-of-the-art models, especially on cold start items. For each metric, the hit ratio (HR) and normalized discounted cumulative gain (NDCG) on the Beauty dataset improved by an average of 1.3% and 10.27%, respectively; on the Toys dataset improved by an average of 8.24% and 5.89%, respectively; on the ML-1M dataset improved by an average of 68.62% and 8.22%, respectively; and on the ML-100M dataset improved by an average of 93.57% and 44.87% Our code is available at DOI: 10.5281/zenodo.13634624.

Sequential recommendation is a popular task in academic research and close to real-world application scenarios, where the goal is to predict the next action(s) of the user based on his/her previous sequence of actions. In the training process of recommender systems, the loss function plays an essential role in guiding the optimization of recommendation models to generate accurate suggestions for users. However, most existing sequential recommendation techniques focus on designing algorithms or neural network architectures, and few efforts have been made to tailor loss functions that fit naturally into the practical application scenario of sequential recommender systems. Ranking-based losses, such as cross-entropy and Bayesian Personalized Ranking (BPR) are widely used in the sequential recommendation area. We argue that such objective functions suffer from two inherent drawbacks: i) the dependencies among elements of a sequence are overlooked in these loss formulations; ii) instead of balancing accuracy (quality) and diversity, only generating accurate results has been over emphasized. We therefore propose two new loss functions based on the Determinantal Point Process (DPP) likelihood, that can be adaptively applied to estimate the subsequent item or items. The DPP-distributed item set captures natural dependencies among temporal actions, and a quality vs. diversity decomposition of the DPP kernel pushes us to go beyond accuracy-oriented loss functions. Experimental results using the proposed loss functions on three real-world datasets show marked improvements over state-of-the-art sequential recommendation methods in both quality and diversity metrics.

Background: A recommender algorithm’s main goal is to learn user preferences from the user-system interactions and provide a list of relevant items to the user. In information retrieval literature this problem is formulated as learning to rank (LtR) problem. Bayesian Personalized Ranking (BPR) [1] is one of the popular LtR approaches based on pair-wise comparison using single source implicit information. Aim: In this work, we aim to design a recommender system algorithm that generates accurate recommendations. The system should only use a single source implicit user preference information. This is possible through a good approximation of the posterior probability in BPR optimization function. Method: We proposed a Similarity based Monte Carlo approximate solution for the posterior probability in BPR. We used four datasets from different recommendation application domains to evaluate the performance of our proposed algorithm. The input data was pre-processed to match with the requirements of the algorithm. Result: The result of the analysis shows a significant improvement in terms of mean average precision (MAP) for our proposed algorithm compared with the BPR and another alternative extension to BPR. Conclusion: We conclude that the proposed approximate solution is successful in providing the most informative samples to approximate BPR posterior probability. This is confirmed by the significant improvement of the accuracy of the provided ranked list of items for the users. i

Location-based social networks (LBSNs), such as Gowalla, Facebook, Foursquare, Brightkite, and so on, have attracted millions of users to share their social friendship and their locations via check-ins in the past few years. Plenty of valuable information is accumulated based on the check-in behaviors, which makes it possible to learn users’ moving patterns as well as their preferences. In LBSNs, point-of-interest (POI) recommendation is one of the most significant tasks because it can help targeted users explore their surroundings as well as help third-party developers provide personalized services. Matrix factorization is a promising method for this task because it can capture users’ preferences to locations and is widely adopted in traditional recommender systems such as movie recommendation. However, the sparsity of the check-in data makes it difficult to capture users’ preferences accurately. Geographical influence can help alleviate this problem and have a large impact on the final recommendation result. By studying users’ moving patterns, we find that users tend to check in around several centers and different users have different numbers of centers. Based on this, we propose a Multi-center Gaussian Model (MGM) to capture this pattern via modeling the probability of a user’s check-in on a location. Moreover, users are usually more interested in the top 20 or even top 10 recommended POIs, which makes personalized ranking important in this task. From previous work, directly optimizing for pairwise ranking like Bayesian Personalized Ranking (BPR) achieves better performance in the top- k recommendation than directly using matrix matrix factorization that aims to minimize the point-wise rating error. To consider users’ preferences, geographical influence and personalized ranking, we propose a unified POI recommendation framework, which unifies all of them together. Specifically, we first fuse MGM with matrix factorization methods and further with BPR using two different approaches. We conduct experiments on Gowalla and Foursquare datasets, which are two large-scale real-world LBSN datasets publicly available online. The results on both datasets show that our unified POI recommendation framework can produce better performance.

This paper shows a method for building and publishing datasets in commercial services. Datasets contribute to the development of research in machine learning and recommender systems. In particular, because recommender systems play a central role in many commercial services, publishing datasets from the services are in great demand from the recommender system community. However, the publication of datasets by commercial services may have some business risks to those companies. To publish a dataset, this must be approved by a business manager of the service. Because many business managers are not specialists in machine learning or recommender systems, the researchers are responsible for explaining to them the risks and benefits. We first summarize three challenges in building datasets from commercial services: (1) anonymize the business metrics, (2) maintain fairness, and (3) reduce the popularity bias. Then, we formulate the problem of building and publishing datasets as an optimization problem that seeks the sampling weight of users, where the challenges are encoded as appropriate loss functions. We applied our method to build datasets from the raw data of our real-world mobile news delivery service. The raw data has more than 1,000,000 users with 100,000,000 interactions. Each dataset was built in less than 10 minutes. We discussed the properties of our method by checking the statistics of the datasets and the performances of typical recommender system algorithms.

In many application domains of recommender systems, explicit rating information is sparse or non-existent. The preferences of the current user have therefore to be approximated by interpreting his or her behavior, i.e., the implicit user feedback. In the literature, a number of algorithm proposals have been made that rely solely on such implicit feedback, among them Bayesian Personalized Ranking (BPR).In the BPR approach, pairwise comparisons between the items are made in the training phase and an item i is considered to be preferred over item j if the user interacted in some form with i but not with j. In real-world applications, however, implicit feedback is not necessarily limited to such binary decisions as there are, e.g., different types of user actions like item views, cart or purchase actions and there can exist several actions for an item over time.In this paper we show how BPR can be extended to deal with such more fine-granular, graded preference relations. An empirical analysis shows that this extension can help to measurably increase the predictive accuracy of BPR on realistic e-commerce datasets.

Research in recommender systems has traditionally focused on improving the predictive accuracy of recommendations by developing new algorithms or by incorporating new sources of data. However, several studies have shown that accuracy does not always correlate with a better user experience, leading to recent research that puts emphasis on Human-Computer Interaction in order to investigate aspects of the interface and user characteristics that influence the user experience on recommender systems. Following this new research this paper presents SetFusion, a visual user-controllable interface for hybrid recommender system. Our approach enables users to manually fuse and control the importance of recommender strategies and to inspect the fusion results using an interactive Venn diagram visualization. We analyze the results of two field studies in the context of a conference talk recommendation system, performed to investigate the effect of user controllability in a hybrid recommender. Behavioral analysis and subjective evaluation indicate that the proposed controllable interface had a positive effect on the user experience.

There have been significant advances made in the research of recommender systems over the past decades and have been implemented in both industry and academia. Recently, multi-criteria ratings are being incorporated into traditional recommender systems to further improve their quality, especially to handle the data sparsity and cold start issues. However, incorporation of multi-criteria ratings have improved the performance of the recommendation, but at the same time, multi-dimensionality issue also arises. This paper presents a clustering-based recommendation approach which is used for dealing with the multi-dimensionality issue in multi-criteria recommender systems. Here, we cluster the users based on their individual criteria ratings using K-means clustering and the intra-cluster similarity is computed using Mahalanobis distance measure for neighbourhood set generation. This improves the recommendations quality and predictive accuracy of both traditional and clusteringbased collaborative recommendations. The Yahoo! Movies dataset was used for testing the approach and the experiment conducted shows promising results.

Over the past several years, research in recommender systems has emphasized the importance of serendipity, but there is still no consensus on the definition of this concept and whether serendipitous items should be recommended is still not a well-addressed question. According to the most common definition, serendipity consists of three components: relevance, novelty and unexpectedness, where each component has multiple variations. In this paper, we looked at eight different definitions of serendipity and asked users how they perceived them in the context of movie recommendations. We surveyed 475 users of the movie recommender system, MovieLens regarding 2146 movies in total and compared those definitions of serendipity based on user responses. We found that most kinds of serendipity and all the variations of serendipity components broaden user preferences, but one variation of unexpectedness hurts user satisfaction. We found effective features for detecting serendipitous movies according to definitions that do not include this variation of unexpectedness. We also found that different variations of unexpectedness and different kinds of serendipity have different effects on preference broadening and user satisfaction. Among movies users rate in our system, up to 8.5% are serendipitous according to at least one definition of serendipity, while among recommendations that users receive and follow in our system, this ratio is up to 69%.

Predicting process behavior meets factorization machines

While much recommender system research has been driven by the rating prediction task, there is an emphasis in recent research on exploring new methods to evaluate the effectiveness of a recommendation. The Recommender Systems Challenge 2014 takes up this theme by challenging researchers to explore engagement as an evaluation criterion. In this paper we discuss how predicting engagement differs from the traditional rating prediction task and motivate the rationale behind our approach to the challenge. We show that standard matrix factorization recommender algorithms do not perform well on the task. Our solution depends on clustering items according to their time-dependent profile to distinguish topical movies from other movies. Our prediction engine also exploits the observation that extreme ratings are more likely to attract engagement.

With the advent of the era of big data, in order to actively cope with the massive and unstructured "information overload", recommendation systems are widely used in network products and have a profound impact on people's lives. At present, the research of recommendation system has been developed for more than ten years, but many of them are still in the incomplete stage, there are problems such as sparse data, insufficient mining of user implicit preferences, and inaccurate semantic description of item features. These pictures, videos and other content are full of users' emotions, which contain a lot of user emotions, which need further Excavation and application, resulting in huge social and economic benefits. However, there are few researches on sentiment analysis of this kind of information, and the multi-modal sentiment analysis of text, images and other information is particularly weak. Multimodal information has the same emotional characteristics and can describe the user's emotional changes from different angles, so as to obtain more accurate emotional identification.

A large body of research in recommender systems focuses on optimizing prediction and ranking. However, recent work has highlighted the importance of other aspects of the recommendations, including transparency, control and user experience in general. Building on these aspects, we introduce MoodPlay, a hybrid recommender system music which integrates content and mood-based filtering in an interactive interface. We show how MoodPlay allows the user to explore a music collection by latent affective dimensions, and we explain how to integrate user input at recommendation time with predictions based on a pre-existing user profile. Results of a user study (N=240) are discussed, with four conditions being evaluated with varying degrees of visualization, interaction and control. Results show that visualization and interaction in a latent space improve acceptance and understanding of both metadata and item recommendations. However, too much of either can result in cognitive overload and a negative impact on user experience.

Sedentary lifestyles and bad eating habits influence the onset of many serious health problems. Healthy behavior change is an arduous task, and requires a careful planning. In this thesis, we propose that behavior recommenders can help their users achieve healthy behavior change. Such a system should inspire its users with small, incremental and achievable goals. For this, it must resolve a trade-off between two opposing objectives: help the user achieve a steady improvement in target behavior, and avoid extreme goals that may injure or discourage the user. This is an unprecedented challenge in the recommender systems research. If the system understands the impacts of past interventions for behavior change, it can determine its usersâ behavioral responses to its own recommendations. This implies a specific data curation, in which we not only measure people's behavior but also deliberately introduce an intervention to monitor its effect on people's patterns. In turn, the system can use these existing users' information to derive the right procedure for effective recommendations. In this study we capitalize on this insight and develop InspiRE - our behavior recommender framework. Through InspiRE we propose the following contributions: 1) We design the data curation. 2) We develop the novel approaches for behavior profiling 3) We develop an evaluation process for this novel type of recommender system, and also compare it with traditional, similarity-based recommendation approach. We curate a dataset that contains information of daily step counts and social intervention for 83 people. InspiRE successfully uses the observations from this dataset, and proposes recommendations that are both effective and feasible. We also show that InspiRE can generalize to other dimensions of well being: we demonstrate this through a dataset that contains the snacking patterns of 73 people, who receive message-based interventions. We observe that InspiRE's recommendation strategy is in line with theories of behavior change.

Learning to rank has recently emerged as an attractive technique to train deep convolutional neural networks for various computer vision tasks. Pairwise ranking, in particular, has been successful in multi-label image classification, achieving state-of-the-art results on various benchmarks. However, most existing approaches use the hinge loss to train their models, which is non-smooth and thus is difficult to optimize especially with deep networks. Furthermore, they employ simple heuristics, such as top-k or thresholding, to determine which labels to include in the output from a ranked list of labels, which limits their use in the real-world setting. In this work, we propose two techniques to improve pairwise ranking based multi-label image classification: (1) we propose a novel loss function for pairwise ranking, which is smooth everywhere and thus is easier to optimize; and (2) we incorporate a label decision module into the model, estimating the optimal confidence thresholds for each visual concept. We provide theoretical analyses of our loss function in the Bayes consistency and risk minimization framework, and show its benefit over existing pairwise ranking formulations. We demonstrate the effectiveness of our approach on three large-scale datasets, VOC2007, NUS-WIDE and MS-COCO, achieving the best reported results in the literature.