Wednesday Poster & Coffee Break Session

Date: Wednesday 15:30 – 16:00 CET
Chair: To be announced

  • DMDataHunter: A System for Finding Datasets Based on Scientific Problem Descriptions
    by Michael Färber (Karlsruhe Institute of Technology (KIT), Germany) and Ann-Kathrin Leisinger (Karlsruhe Institute of Technology (KIT), Germany)

    The number of datasets is steadily rising, making it increasingly difficult for researchers and practitioners in the various scientific disciplines to be aware of all datasets, particularly of the most relevant datasets for a given research problem. To this end, dataset search engines have been proposed. However, they are based on the users’ keywords and thus have difficulties in determining precisely fitting datasets for complex research problems. In this paper, we propose the system at that recommends suitable datasets to users based on given research problem descriptions. It is based on fastText for the text representation and text classification, the Data Set Knowledge Graph (DSKG) with metadata about almost 1,700 unique datasets, as well as 88,000 paper abstracts as research problem descriptions for training the model. Overall, our system demonstrates that recommending datasets facilitates data provisioning and reuse according to the FAIR principles and that dataset recommendation is a promising future research direction.

    Full text in ACM Digital Library

  • DMTransformers4Rec: Bridging the Gap between NLP and Sequential / Session-Based Recommendation
    by Gabriel de Souza Pereira Moreira (NVIDIA, Brazil), Sara Rabhi (NVIDIA, Canada), Jeong Min Lee (Facebook AI, United States), Ronay Ak (NVIDIA, United States), and Even Oldridge (NVIDIA, Canada)

    Much of the recent progress in sequential and session-based recommendation has been driven by improvements in model architecture and pretraining techniques originating in the field of Natural Language Processing. Transformer architectures in particular have facilitated building higher-capacity models and provided data augmentation and training techniques which demonstrably improve the effectiveness of sequential recommendation. But with a thousandfold more research going on in NLP, the application of transformers for recommendation understandably lags behind. To remedy this we introduce Transformers4Rec, an open-source library built upon HuggingFace’s Transformers library with a similar goal of opening up the advances of NLP based Transformers to the recommender system community and making these advancements immediately accessible for the tasks of sequential and session-based recommendation. Like its core dependency, Transformers4Rec is designed to be extensible by researchers, simple for practitioners, and fast and robust in industrial deployments.
    In order to demonstrate the usefulness of the library and the applicability of Transformer architectures in next-click prediction for user sessions, where sequence lengths are much shorter than those commonly found in NLP, we have leveraged Transformers4Rec to win two recent session-based recommendation competitions. In addition, we present in this paper the first comprehensive empirical analysis comparing many Transformer architectures and training approaches for the task of session-based recommendation. We demonstrate that the best Transformer architectures have superior performance across two e-commerce datasets while performing similarly to the baselines on two news datasets. We further evaluate in isolation the effectiveness of the different training techniques used in causal language modeling, masked language modeling, permutation language modeling and replacement token detection for a single Transformer architecture, XLNet. We establish that training XLNet with replacement token detection performs well across all datasets. Finally, we explore techniques to include side information such as item and user context features in order to establish best practices and show that the inclusion of side information uniformly improves recommendation performance. Transformers4Rec library is available at

    Full text in ACM Digital Library

  • LBRHorizontal Cross-Silo Federated Recommender Systems
    by Saikishore Kalloori (ETH Zürich, Switzerland) and Severin Klingler (Media Technology Center, ETH Zürich, Switzerland)

    Recommender systems (RSs) completely rely on the knowledge of training information to generate recommendations. However, due to privacy, ownership, and protection of users’ information, such training information is not easily accessible or shared with an RS. Moreover, with recent regulations in privacy laws (e.g, GDPR), collecting user preferences and perform centralized training may not be feasible. Federated Learning (FL) is a form of machine learning technique where the goal is to learn a high-quality recommendation model without never directly accessing raw training data. In this work, we specifically focus on situations where multiple stakeholders (referred to as corporate companies like e-commerce business partners, hospitals, banks, news media publishers) participate in federated learning to build a shared recommendation model. We performed offline experiments by simulating a real federated learning setup and investigated the benefits federated learning brings to stakeholders in terms of ranking compared to an RS model trained without participating in federated learning. Our experimental results reveal that stakeholders can significantly benefit from federated learning to generate accurate recommendations. Moreover, we also study the use and benefits of federated learning in situations when there are not enough preferences available for users.

    Full text in ACM Digital Library

  • LBRAutomatic Collection Creation and Recommendation
    by Sanidhya Singal (Airtel, India), Piyush Singh (Airtel, India), and Manjeet Dahiya (Airtel, India)

    We present a collection recommender system that can automatically create and recommend collections of items at a user level. Unlike regular recommender systems, which output top-N relevant items, a collection recommender system outputs collections of items such that the items in the collections are relevant to a user, and the items within a collection follow a specific theme. Our system builds on top of the user-item representations learnt by item recommender systems. We employ dimensionality reduction and clustering techniques along with intuitive heuristics to create collections with their ratings and titles. We test these ideas in a real-world setting of music recommendation, within a popular music streaming service. We find that there is a 2.3x increase in recommendation-driven consumption when recommending collections over items. Further, it results in effective utilization of real estate and leads to recommending a more and diverse set of items. To our knowledge, these are first of its kind experiments at such a large scale.

    Full text in ACM Digital Library

  • LBRDynamic Graph Construction for Improving Diversity of Recommendation
    by Rui Ye (Meituan, China), Yuqing Hou (Meituan, China), Te Lei (Meituan, China), Yunxing Zhang (Meituan, China), Qing Zhang (Meituan, China), Jiale Guo (Meituan, China), Huaiwen Wu (Meituan, China), and Hengliang Luo (Meituan, China)

    The diversity of recommendation has attracted a lot of attention in recommender systems due to its ability to improve user experience. Most of the diversified recommendation tasks usually exploit user-item interaction records for mining user explicit preferences, while rarely explore the user-item non-interaction records. For diversified recommendations, however, the neglected non-interaction records are especially important for capturing users’ potential interests to improve the diversity of recommendation. Moreover, the majority of diversified recommendation methods run in two stages: first optimizing the users and items embeddings by relevance, then generating the diversified items list by post-processing methods. These methods are not end-to-end thus can hardly reach global optimum. To solve above limitations, we propose an end-to-end Dynamic Diversified Graph framework (DDGraph) which constructs the user-item graph dynamically based on the users and items embeddings. Technically, we initialize a user-item interaction graph and dynamically update the graph by selecting a set of diverse items for each user and building links between the items and user. The selection of diverse items can be achieved by different candidate selection operators. Specifically, we design a Quantile Progressive Candidate Selection (QPCS) operator based on the latent space division. To the best of our knowledge, our method is the first to diversify recommendation results by dynamic end-to-end graph construction and the QPCS has a higher computational efficiency than other operators. Extensive experiments on the benchmark dataset illustrate the effectiveness and superiority of the DDGraph framework.

    Full text in ACM Digital Library

  • LBRSequence Adaptation via Reinforcement Learning in Recommender Systems
    by Stefanos Antaris (KTH Royal Institute of Technology, Sweden) and Dimitrios Rafailidis (University of Thessaly, Greece)

    Accounting for the fact that users have different sequential patterns, the main drawback of state-of-the-art recommendation strategies is that a fixed sequence length of user-item interactions is required as input to train the models. This might limit the recommendation accuracy, as in practice users follow different trends on the sequential recommendations. Hence, baseline strategies might ignore important sequential interactions or add noise to the models with redundant interactions, depending on the variety of users’ sequential behaviours. To overcome this problem, in this study we propose the SAR model, which not only learns the sequential patterns but also adjusts the sequence length of user-item interactions in a personalized manner. We first design an actor-critic framework, where the RL agent tries to compute the optimal sequence length as an action, given the user’s state representation at a certain time step. In addition, we optimize a joint loss function to align the accuracy of the sequential recommendations with the expected cumulative rewards of the critic network, while at the same time we adapt the sequence length with the actor network in a personalized manner. Our experimental evaluation on four real-world datasets demonstrates the superiority of our proposed model over several baseline approaches. Finally, we make our implementation publicly available at

    Full text in ACM Digital Library

  • DSLeveraging Multi-Faceted User Preferences for Improving Click-Through Rate Predictions
    by Pan Li (New York University, United States)

    Recommender systems have been widely adopted by many firms across various industries over the past decade, as they could provide numerous economic benefits to the industry, such as influencing consumer choices, generating the lift in sales, and enhancing consumer trust. Many currently-deployed recommender systems primiarily optimize the similarity measure following the Collaborative Filtering paradigm during the recommendation performance, which focuses on the matching between targeted users and items of their interests. Although effective and useful in many cases, these still face several important challenges to fully address consumers’ concerns. In particular, they usually provide recommendations of the most similar items related to the consumption records, without taking into account the multi-faceted consumer preferences. For example, some consumers would be satisfied to receive recommendations that are novel and unexpected to broaden their horizons, while others might be interested to receive recommendations from other product categories to expand their interests. Falling to do so would lead to the filter bubble and exploration-exploitation trade-off phenomenon, which is undesirable because it would significantly reduce sales diversity and deteriorate user satisfaction with recommender systems.
    In my thesis proposal, I will describe new methods to design recommender systems following two research streams to tackle the aforementioned challenges, namely unexpected recommender system and cross-domain system. In particular, unexpected recommender systems address consumers’ desire for variety, and aim at providing novel and useful recommendations simultaneously, while cross-domain recommender systems learn consumer preferences from their behaviors in other domains to better predict their behaviors in the target domain. The proposed models are capable of significantly improving recommendation performance and user satisfaction, as demonstrated through extensive offline and online experiments discussed in this proposal.

    Full text in ACM Digital Library

  • DSMeasuring and Mitigating Bias and Harm in Personalized Advertising
    by Muhammad Ali (Khoury College of Computer Sciences Northeastern University, United States)

    Online personalized advertising is often very effective in identifying relevant audiences for each piece of content, which has led to its widespread adoption. In today’s internet, however, these advertising systems are used not only to market products, but also consequential life opportunities such as employment or housing, as well as socially important political messaging. This has led to increasing concerns about the presence of algorithmic bias and possible discrimination in these important domains — with results showing problematic biases along gender, race, and political affiliation, even when the advertiser might have targeted broadly.

    A growing body of work focuses on measuring and characterizing these biases, as well as finding ways to mitigate these effects and building responsible systems. However, these results often emerge from different scientific communities and are often disconnected in the literature. In this paper, I attempt at bridging the gap between isolated efforts to either measure these biases, or to mitigate them. I discuss how the need to measure bias in advertising, and the efforts to mitigate it, despite being distant in the literature, are complementary problems that need to center their methodolgy around user studies.

    This paper presents a research agenda that focuses on the need for user-centric measurements of bias, by collecting real ads from users, and using surveys to understand user perceptions for these ads. My approach also calls for incorporating user sentiments into the mitigation efforts, by constraining optimization on user values that emerge from surveys. Finally, I also emphasize the need for involving users in the evaluation of responsible advertising systems; efforts to mitigate bias eventually need to be contextualized in terms of benefits to users instead of simple performance tradeoffs. My focus on the users is motivated by the fact that they are stakeholders in personalized advertising, vulnerable at the hand of algorithmic bias and harm, and therefore crucial in both efforts to measure and mitigate these effects.

    Full text in ACM Digital Library

  • INPersonalizing Peloton: Combining Rankers and Filters To Balance Engagement and Business Goals
    by Shayak Banerjee (Personalization Team Peloton, Inc., United States), Arnab Bhadury (Personalization Team Peloton Interactive, Inc., United States), Nilothpal Talukder (Personalization Team Peloton Interactive, Inc., United States), and Santosh Thammana (Personalization Team Peloton Interactive, Inc., United States)

    Peloton is at the forefront of the at-home fitness market, with two business pillars: (a) a line of connected fitness equipment, and (b) a subscription-based service that offers access to a rich catalog of high quality fitness classes. As of May 2021, the total member base for Peloton stood at over 5.4 million, who took more than 170 million workouts. Peloton classes cover a diversity of instructors, languages, fitness disciplines, durations, intensity and muscle groups. On the other side, each user has their own specific fitness goals, time available to work out, fitness equipment and level of skill or strength. This diversity of content and individuality of user needs creates the need for a recommender system capable of personalizing the Peloton experience.
    Most recommendation engines optimize for user lifetime value or time of engagement. However, Peloton users have very different usage habits when compared to other industry recommendation problems. Users arrive on the platform with a clear intent to workout, which allows our recommendation algorithms to not just focus on the short-term classic metrics such as click-through-rates and optimizing session lengths for exploration. Instead, fitness content recommendations at Peloton also help solve longer term problems such as: 
    It helps balance engagement and business goals. A classic example of this is the introduction of a new instructor. For existing users, who already have developed affinities to certain instructors, how can we ensure that enough of them see and try some classes from the new instructor so that they can build their own following?
    It helps guide users to explore the vast library of content. Peloton users quickly develop set routines with our fitness content, with high repeat plays of the same instructor/duration/class type. Recommendations serve as a mechanism to encourage them to try something outside this comfort zone, which both increases the breadth of a user’s engagement with the platform and leads to a more holistic workout routine.
    We achieve these two goals by utilizing a combination of rankers and filters. Ranking models help order the universe of content for each user according to their preferences. Filters take a slice of this ordered content to generate a shelf of content with a reason for suggesting it. Explainability is heavily linked to business goals, while ranking is linked to engagement goals. For instance, ranking and filtering can work in tandem to populate a shelf that helps promote a new music partnership, e.g. Workouts Featuring The Beatles, where we highlight classes that contain music by The Beatles (filter), ordered by the user’s class preferences (ranker). With rankers and filters, we empower other teams to curate personalized experiences. The creation process is simplified to picking a ranker and a filter to create a shelf, and then giving it a title to then have it displayed to users.
    Further, we have context-based models that order the shelves for each user depending on both their preferences and context, such as platform and time. For ranking our various filters, we take a multi-armed bandit approach, due to the need to handle cold starts on users and balance exploration (keep putting new shelves in front of the user) with exploitation (keep showing them shelves they already interact with). To start with, we implemented a simple Thompson Sampling based bandits algorithm, which accumulates rewards (for shelves interacted with) and penalties (for shelves ignored), which in turn constantly adapts the shelf ordering for a user, making the experience more personalized over time. We are able to perform all calculations offline in batch, using Spark, and cache the Thompson Sampling parameters in DynamoDB. When a user requests their shelves, a random sampling is performed using these parameters to serve up a contextually ordered list of shelves.
    A unique feature of Peloton classes is that they are usually aired “live” first, which seeds the class with a set of users. This ameliorates the cold start problem for recommending classes, opening us up to using collaborative filtering approaches. Also, users typically take one session in a given day, and most even just take one class in a given day. This means we are able to compute class recommendations for each user offline, cache it and serve it up when requested. In the offline world, we use Spark to pre-process our user-class interaction data, and then train a deep learning model using PyTorch. Our ranking model is a sequential recommender using long-short-term memory (LSTMs). From our ranked list of classes, we apply our various filters and generate the shelves of recommendations. These are cached in AWS DynamoDB, and served up via Kubernetes-driven APIs.
    Recommending fitness content at Peloton has the potential to go beyond simply guiding users to their next class. With a holistic overview of both what classes a user is taking and a user’s feedback on their performance (explicit or implicit), there is an opportunity to tailor workout routines that optimize for long-term metrics such as health, strength or flexibility.
    Shayak Banerjee is currently a Staff Machine Learning Engineer on the Personalization team at Peloton, where he works on personalizing the Discovery and Browse experiences for users. His contributions relate to building out the infrastructure for running data pipelines and tools for balancing content production and consumption needs. Prior to Peloton, he worked on connecting users to communities at Meetup. Shayak graduated from The University of Texas at Austin with a PhD in Electrical & Computer Engineering.

    Full text in ACM Digital Library

Platinum Supporters
Gold Supporters
Silver Supporters
Special Supporter