The RecSys Challenge 2025 focuses on developing robust recommendation systems capable of generalizing across multiple tasks in a highly sparse and dynamic user-behavior dataset. Designing unified user representations that generalize across multiple recommendation tasks under extreme sparsity remains a key challenge. In this paper, we present TFT-SR (Triple-Feature Transformer with Sparsity Regularization), a unified framework for behavioral modeling in the RecSys Challenge 2025. Our method fuses three complementary types of features-statistical descriptors, quantized temporal patterns, and hashed high-cardinality IDs-into a unified user vector. A dual-path neural encoder is used to separately extract dense and sparse representations, with the sparse branch regularized by an L1 penalty to promote interpretability and efficiency. Multi-task optimization [13, 15, 18] is performed through loss weighting, ensuring balanced learning across tasks. we participated in the competition under the team name 'xunzhou,' achieving 11th place on the final leaderboard and 5th place on the academic leaderboard.
Universal user representations hold the promise of reducing the need for task-specific modelling in personalized systems, enabling a single embedding to support a wide range of predictive tasks such as churn prediction, product propensity, and category-level recommendations. In this paper, we introduce BEHAV-E (Behavioural Embedding via Hybrid Action Variational Encoder), a self-supervised architecture that learns rich, 2048-dimensional user embeddings from multi-event behavioural data. BEHAV-E integrates both temporal and semantic signals by combining kernel density estimation (KDE) of user activity timelines, product and category embeddings via shared embedding bags, and an LSTM-based autoencoder for modelling search query semantics. We evaluate our approach within the RecSys Universal Behavioural modelling Challenge. Our results demonstrate that BEHAV-E effectively captures complex, multi-modal user behaviour in a compact and transferable embedding format. This paper provides an overview of the approach we used as team DArgk1 for the ACM RecSys Challenge 2025. Our submission achieved the 13..h rank and sum of scores of 4.5713 (borda count 483) in the competition academia-track final results. We release our source code at: https://github.com/knife982000/ RecSys2025Challenge
Traditional recommender systems are usually engineered for one of several objectives; predicting clicks, conversions, or churn depend on bespoke feature engineering that limits reuse and overlooks broader behavioral patterns. The RecSys Challenge 20251, organized by Synerise, tackles this shortcoming by introducing Universal Behavioral Profiles, a unified user representation framework meant to generalize across many downstream tasks. This paper presents the winning approach of the “rec2” team for this challenge. Our approach combines two Transformer-based embeddings with a compact statistical vector that summarizes coarse behavioral frequencies. This ensemble achieved consistently high accuracy on both the open tasks and the hidden tasks, highlighting its robust generalization ability. Our findings highlight two takeaways: (i) training Transformers with contrastive learning forms a particularly effective backbone for universal user embeddings; (ii) combining embeddings produced by diverse models yields an additional performance gains.
This work presents a sequence-aware framework for learning Universal Behavioral Profiles (UBPs) in the ACM RecSys Challenge 2025. Our method combines State-Space model-based sequential encoders (SS4Rec and dual Mamba4Rec models) with explicit churn tokens, action-specific loss weighting, temporal segmentation, and a rich set of 471 user-level aggregate features. Two complementary vocabularies balance high-frequency and low-frequency events, while a three-window scheme captures both short- and long-term dynamics. Evaluated on six downstream tasks (three open and three hidden) our team, Teletubbies, ranked 6th overall with sum of six scores of 4.73, achieving top-tier performance on five tasks but underperforming on the New Product Propensity (hidden2) cold-start task. An extensive ablation study confirms the unique contributions of each component to cross-task generalization. We analyze limitations in handling novel items and outline future work integrating side-information to address cold-start challenges.
Modern enterprises require robust predictive analytics, yet tasks like churn prediction and propensity modeling are often treated as isolated problems. The ACM RecSys Challenge 2025 addressed this by proposing a unified task to create generalizable user profiles from a large-scale behavioral dataset provided by Synerise. This paper presents the 4th-place solution in the academic track from team "ririka". Our approach centers on creating a powerful, hybrid Universal Behavioral Profile by fusing an extensive, feature-engineered vector with multiple specialized embeddings from a Universal Behavioral Transformer (UBT). We introduce a comprehensive feature engineering strategy incorporating statistical, temporal, and semantic features, and develop a multi-task UBT framework to generate distinct embeddings for objectives like churn and propensity. Through a meticulous fusion and normalization methodology, we combine these components into a single, high-dimensional user representation. Our final model proved highly effective, demonstrating robust and generalizable performance across both open and hidden evaluation tasks. Our analysis confirms that synthesizing handcrafted features with deep, sequential representations is critical for building a powerful and versatile user profile.
This paper presents the solution developed by the EmbedNBreakfast team for the ACM RecSys Challenge 2025, for the construction of Universal Behavioral Profiles: general-purpose user representations derived from historical interactions. We propose a representation-learning framework that combines Recurrent Neural Networks, attention mechanisms, and collaborative filtering to jointly optimize embeddings across several predictive objectives. Our method achieved 2nd place on the Academic Leaderboard and 5th Overall, demonstrating the effectiveness of unified, representation-based modeling for diverse behavior prediction tasks.
Research in the area of recommender systems heavily relies on offline experimentation with historical data. The validity of such research efforts may however be limited by the quality and representativeness of publicly available datasets. To address these limitations, we introduce the Synerise dataset as a new, large-scale e-commerce dataset derived from real-world logs. This dataset provides rich, time-stamped user-item interactions alongside detailed item metadata—including categories, descriptions, and prices—and incorporates user search and navigation behavior for a more holistic understanding of user intent. In the paper, we provide a description of the dataset and how it can be used for model evaluation in different research questions. Furthermore, we provide an overview of the ACM RecSys 2025 Challenge, which introduced the novel task of Universal Behavioral Modeling, and which was based on the Synerise dataset. The dataset can be downloaded at https://recsys.synerise.com.
In this paper, we propose the solution developed by the UEC_ bootcamp_2025 team, which was awarded the 2nd Academic Prize in the RecSys Challenge 2025. The challenge required participants to construct universal user representations that generalize across multiple predictive tasks. To this end, we propose an approach that integrates heterogeneous user features derived from diverse modeling approaches. Each feature type captures distinct aspects of user behavior, and we apply normalization and weighting to ensure compatibility across representations. By combining these diverse embeddings, our method constructs robust user profiles that support consistent performance across open and hidden tasks. Experiments show our representation outperforms individual features and the baseline. The code is available at https://github.com/1012sk/RecSys_Challenge_2025_academic_2nd_prize.
For the RecSys Challenge 2025, our team, teitlax, presents a resource-efficient approach that lifts a 1-billion-parameter language model to the top 8% of the final leaderboard (rank 33/424). Our core thesis is to improve the input representation rather than increasing model size. Our pipeline first transforms 19M raw interaction logs into dense, multi-section user documents. It then employs a second small LLM to distil each profile into a concise behavioral narrative, a step we term “narrative distillation”. Finally, the base model is fine-tuned via parameter-efficient contrastive learning before a lightweight ensemble. This recipe reaches an average AUROC of 0.760 across the six open and hidden tasks—within 5% of the winning score (0.799), demonstrating that intelligent feature engineering can unlock state-of-the-art performance with modest computational and reproducibility budgets.
Modern digital platforms generate vast amounts of user interaction data every day. Although this data is closely related, predictive tasks such as recommendation and churn prediction are often modeled separately, resulting in inconsistent user representations. The RecSys Challenge 2025 explores this issue by promoting the creation of Universal Behavioral Profiles: single, versatile embeddings that generalize across multiple tasks. To exploit the relational nature of interaction data stored in tables, our team “sisinflabpoliba” proposed a solution rooted in the Relational Deep Learning (RDL) paradigm, which treats the user’s multi-faceted journey as a unified heterogeneous graph. Our approach enhances the LightGCN architecture to perform relation-aware message passing across diverse entities like users, products, and queries. We then create a hybrid profile by fusing these learned deep embeddings with a robust, feature-engineered baseline. This strategy effectively balances the complex relational patterns captured by the Graph Neural Network model with the stability of handcrafted features, yielding powerful, general-purpose user profiles that demonstrate strong performance across the challenge’s open and hidden tasks.
Playlist curation is a key factor in media content discovery services, yet finding diverse, relevant content is challenging for curators due to time-consuming manual query crafting. We propose a method where a large language model (LLM) expands a playlist theme into multiple diverse queries. Vectors from these expanded queries, along with the original theme vector, retrieve candidates via vector search. Experiments on Japanese TV programs show our method significantly improves precision over a theme-vector baseline, boosting Precision@10 from 0.79 to 0.98 and increasing P@50 by 22 percentage points. This approach enhances curator efficiency and improves playlist quality by delivering more accurate and diverse recommendations.
Already before the enactment of the EU AI Act, candidate or job recommendation for algorithmic hiring -- semi-automatically matching CVs to job postings -- was used as an example of a high-risk application where unfair treatment could result in serious harms to job seekers. Recommending candidates to jobs or jobs to candidates, however, is also a fitting example of a multi-stakeholder recommendation problem. In such multi-stakeholder systems, the end user is not the only party whose interests should be considered when generating recommendations. In addition to job seekers, other stakeholders -- such as recruiters, organizations behind the job postings, and the recruitment agency itself -- are also stakeholders in this and deserve to have their perspectives included in the design of relevant fairness metrics. Nevertheless, past analyses of fairness in algorithmic hiring have been restricted to single-side fairness, ignoring the perspectives of the other stakeholders. In this paper, we address this gap and present a multi-stakeholder approach to fairness in a candidate recommender system that recommends relevant candidate CVs to human recruiters in a human-in-the-loop algorithmic hiring scenario. We conducted semi-structured interviews with 40 different stakeholders (job seekers, companies, recruiters, and other job portal employees). We used these interviews to explore their lived experiences of unfairness in hiring, co-design definitions of fairness as well as metrics that might capture these experiences. Finally, we attempt to reconcile and map these different (and sometimes conflicting) perspectives and definitions to existing (categories of) fairness metrics that are relevant for our candidate recommendation scenario.
Most food recommender systems aim to boost user engagement by analyzing recipe ingredients and users' past choices. Even though consumers are paying more attention to sustainability, such as carbon and water footprints, there remains a notable lack of public corpora that combine detailed user-recipe interactions with reliable environmental impact data. This gap makes it hard to build recommendation tools that both match people's tastes and help reduce ecological damage. To this end, we present GreenFoodLens, a resource that enriches HUMMUS, one of the largest corpora for food recommendation, with environmental impact estimates derived from the hierarchical taxonomy of the SU-EATABLE-LIFE project. We achieved this result through a multi-step process involving human annotations, iterative labeling assessments, knowledge refinement, and constrained generation techniques with large language models. Finally, we evaluate recommendation baselines on HUMMUS augmented with GreenFoodLens labels and find that models are driven by popularity signals, which may exacerbate the environmental impact of users' recipe choices. These experiments demonstrate the practical benefit of GreenFoodLens for benchmarking and advancing sustainability-aware recommendation research. The resource is available at https://github.com/tail-unica/GreenFoodLens.
This paper introduces a multi-objective ranking framework deployed on a large-scale entertainment platform to jointly optimize user engagement, revenue, and content pricing. Unlike prior work, our system addresses a critical real-world challenge: extreme label imbalance across objectives, with monetization signals being over 100 times sparser than engagement. To overcome this, we adopt an output aggregation strategy that supports runtime tuning of objective weights, enabling fast iteration and dynamic prioritization without retraining. We further introduce a robust offline evaluation pipeline based on Pareto analysis and distribution-aware test datasets, exposing trade-offs that would otherwise remain hidden. Beyond engagement and revenue, we incorporate a third price-based objective optimized via constrained Bayesian search over a high-dimensional simplex by demonstrating how monetization goals can be achieved without degrading user experience. Our approach is validated through both offline evaluations and online A/B tests, showing measurable revenue improvements with minimal impact on engagement. This work provides a novel, end-to-end blueprint for scalable multi-objective optimization under production constraints, where business trade-offs must be explicit, tunable, and validated.
The last few years have seen an increasing interest of the RecSys community in the multimodal recommendation research field, as shown by the numerous contributions proposed in the literature. Our paper falls in this research line, as we released a multimodal extension of three state-of-the-art datasets (MovieLens-1M, DBbook, Last.fm-2K) in the movie, book, music recommendation domains, respectively. Although these datasets have been widely adopted for classical recommendation tasks (e.g., collaborative filtering), their use in multimodal recommendation has been hindered by the absence of multimodal information. To fill this gap, we have manually collected multimodal item raw files from different modalities (text, images, audio, and video, when available) for each dataset. Specifically, we have collected, for MovieLens-1M, movie plots (textual information), movie posters (images) and movie trailers (audio and video); for Last.fm-2K, we have collected, for each artist, the tags provided by users (textual information), the most popular album covers (images), and the most popular songs (audio); finally, for DBbook we have collected book abstracts (textual information) and book covers (image). We encoded all this information using state-of-the-art feature encoders, and we released the extended datasets, which include the mappings to the raw multimodal information and the encoded features. Finally, we conduct a benchmark analysis of various recommendation models using MMRec as a multimodal recommendation framework. Our results show that multimodal information can further enhance the quality of recommendations in these domains compared to single collaborative filtering. We release the multimodal version of such datasets to foster this research line, including links to download the raw multimodal files and the encoded item features.
Accurately capturing diverse user preferences at scale is a core challenge for large-scale recommender systems like Spotify’s, given the complexity and variability of user behavior. To address this, we propose a two-stage framework that combines representation learning and transfer learning to produce generalized user embeddings. In the first stage, an autoencoder compresses rich user features into a compact latent space. In the second, task-specific models consume these embeddings via transfer learning, removing the need for manual feature engineering. This approach enhances flexibility by allowing dynamic updates to input features, enabling near-real-time responsiveness. The framework has been deployed in production at Spotify with an efficient infrastructure that allows downstream models to operate independently. Extensive online experiments in a live setting show significant improvements in metrics such as consumption share, content discovery, and search success. Additionally, our method achieves these gains while substantially reducing infrastructure costs.
Recommendation models often encode users’ sensitive attributes (e.g., gender or age) in their learned representations during training, leading to biased (e.g., stereotypical) recommendations and potential privacy risks. To address this, previous research has predominantly focused on adversarial training to make user representations invariant to sensitive attributes. However, adversarial methods can be unstable and computationally expensive due to additional network parameters. An alternative approach is the use of regularization losses that minimize distributional discrepancies between different demographic groups during training. In particular, the Sliced Wasserstein Distance (SWD) provides a computationally efficient and stable solution for mitigating bias by directly aligning the distributions of user representations across groups. We follow this alternative strategy and propose an in-processing approach to mitigate encoded biases in user representations of implicit feedback-based recommender systems by using SWD-based regularization. We perform extensive experiments targeting the debiasing of the users’ gender on three datasets ML-1M, LFM2b-DB, and EB-NeRD from the movie, music, and news domains, respectively. Our results indicate that SWD-based regularization is an effective approach for mitigating encoded biases in user representations while keeping competitive recommendation accuracy.
In online job search platforms, autocomplete plays a crucial role in providing immediate, structured suggestions that guide users through their query process. However, inconsistencies in job title expressions, such as ’sr data scientist’ versus ’data scientist senior’, or embellished forms such as ’superstar software engineer’, can undermine the quality of autocomplete suggestions and diminish user satisfaction. Traditional normalization methods rely on manually curated vocabularies, which are labor intensive and often insufficient to capture the diverse variations in raw job titles. We present an automated and scalable framework for canonical title generation that leverages large language models (LLMs) alongside embedding-based similarity measures to derive normalized job titles directly from raw data. Our approach generalizes to domains with unstructured or inconsistently formatted titles (e.g. product catalogs or course titles): we systematically remove irrelevant information, enforce a consistent format, and eliminate overly generic or redundant titles by combining LLM-generated normalization with a two-stage deduplication process. Our method demonstrates significant improvements in normalization quality, with offline accuracy gains of 18.6% over baseline methods and online A/B tests showing an improvement of 160% in user engagement metrics.
Sedentary behavior, dubbed the disease of the 21st century, is a ubiquitous force driving chronic illness. Yet, traditional itinerary and Point-of-Interest (POI) Recommender Systems (RSs) lack engaging elements that motivate routine urban walking. This research proposes a novel framework combining narrative-driven storytelling with location-based RSs to promote physical activity and immersive urban exploration. This approach introduces a bidirectional alignment between POI and itinerary recommendations and LLM-generated narratives, transforming routine urban walks into dynamic journeys where contextually relevant stories unfold across city locations. Unlike sequential POI recommendations, this framework embeds location suggestions within contextually relevant narratives of various genres, simultaneously promoting health benefits and deeper city exploration. The research addresses three research questions using a method that builds a structured knowledge base by extracting entities (e.g., POIs, and characters) and semantic links from narrative corpora, enabling semantic alignment between recommended physical locations and story elements. The core aspects of this work are: (i) context-aware itinerary recommendations and personalized story generation, (ii) bidirectional mapping between RSs and story generation, and (iii) systems design bridging user’s needs to promote urban walking as a health activity. Evaluation employs comparative user studies measuring quality and engagement, route-narrative semantic alignment, and narrative analysis to validate the integrated proposed approach.
