The study, titled “Datasets for recommender systems: A survey and comparative overview,” is open access and builds on a structured search of academic databases including Google Scholar, DBLP, IEEE Xplore, the ACM Digital Library and SpringerLink. The authors selected datasets according to strict criteria: they had to be publicly available where possible, contain genuine user-item interaction information such as ratings, reviews, clicks or purchase behavior, and have been used frequently in the experimental evaluation of recommendation algorithms. The goal, the team emphasizes, was not to produce an exhaustive inventory but to assemble a representative set spanning as many domains and usage scenarios as possible, from the earliest benchmark collections dating back to 1998 to resources released as recently as 2025.

To bring order to this sprawling landscape, the survey proposes a taxonomy built on four primary dimensions: application domain, feedback type, dataset size and temporal characteristics. Domain-based categories range from media datasets, the largest and historically most important group, through e-commerce, news recommendation, academic recommendation and location-based social networks. Feedback type distinguishes explicit feedback, such as the familiar one-to-five star ratings that define the MovieLens collections, from implicit feedback, the behavioral residue of clicks, purchases, listening events and browsing activity that modern platforms generate in enormous volumes. Size and temporal structure complete the classification, reflecting the reality that datasets have grown from a few thousand ratings in the field’s infancy to hundreds of millions of interaction records today, and that timestamps and session information are now essential for modeling how preferences shift over time.

The authors add two further analytical dimensions that they argue are too often ignored in dataset descriptions. The first is sparsity, the technical term for the fact that users typically interact with only a tiny fraction of the items available to them. In a movie-rating dataset with thousands of titles, a single user may have rated only a few dozen, leaving the vast majority of entries in the user-item matrix empty. This sparsity directly undermines collaborative filtering methods, which rely on overlapping rating histories to identify similar users or similar items; when the overlap is thin, similarity estimates become unreliable and accuracy suffers, particularly for new users and obscure titles. The second added dimension is bias, most notoriously popularity bias, in which blockbuster items accumulate vastly more interactions than niche content, teaching algorithms to favor what is already famous and starving long-tail items of exposure. Activity bias, demographic skew and exposure artifacts can similarly distort experimental results and reduce the fairness of the resulting recommendations.

No account of the field’s data heritage can skip the two benchmarks that defined an era. The Netflix Prize dataset, released for the famous competition that began in 2006, contains more than 100 million movie ratings and triggered a worldwide surge of research into matrix factorization, model ensembles and hybrid methods that pushed prediction accuracy toward its practical limits. MovieLens, maintained by the GroupLens research group at the University of Minnesota, has served as a long-standing standard of experimentation, with successive variants sized at 100K, 1M, 10M, 20M and 25M ratings that allowed the community to scale its experiments in step with algorithmic ambitions. Both remain explicit-feedback collections, which is precisely why the survey notes that they capture only part of the modern recommendation problem. Beyond movies, the authors highlight the Amazon Product Reviews corpus, spanning millions of reviews from books to electronics and pairing ratings with rich textual feedback, alongside datasets covering music streaming sessions, fashion from Alibaba’s iFashion platform, food recommendation via food.com, news via Microsoft’s MIND dataset and the more recent EB-NeRD collection, academic citation networks such as DBLP and CiteULike, location-based social networks, and game recommendation on Steam.

A distinctive feature of the survey is its treatment of the ACM RecSys Challenge, the prestigious annual competition attached to the ACM Conference on Recommender Systems. Each year, an industrial sponsor releases a large-scale dataset alongside defined evaluation metrics, effectively setting the community’s research agenda for the following year. Recent editions illustrate how the focus has shifted: from session-based hotel and music playlist recommendations, through offline and online evaluation and Twitter’s privacy-aware home timeline, to fairness in fashion sessions, privacy in online advertising and, most recently, the construction of universal user representations suitable for multiple predictive tasks rather than a single recommendation objective. The authors distill the trends across these challenges into four observations: datasets have become substantially larger; the community has moved beyond pure prediction accuracy toward fairness and privacy; modern datasets cover interactions from just the last few weeks or months rather than years; and the latest challenges target general-purpose user modeling.

That shift toward recency and behavioral data reflects a deeper truth the survey makes explicit: the economics of recommendation have changed. The newest datasets tend to originate in domains where user interaction directly generates revenue, such as music with Spotify, fashion with Alibaba, food delivery, news with Microsoft, and gaming with Steam. What these domains share is a constant stream of fresh interactions, a reliance on contextual information rather than static ratings alone, and recommender systems demonstrably successful at converting personalization into income. Older datasets, by contrast, often include an onboarding phase in which users supplied a handful of ratings and then never returned, yet those ratings still populate research benchmarks today, a mismatch between training data and deployment reality that the authors flag as a recurring problem.

The comparative analysis at the heart of the paper connects dataset properties directly to research consequences. Movie-oriented collections remain the benchmarks of choice for rating prediction and collaborative filtering because of their clean, explicitly structured user-item matrices. E-commerce corpora such as Amazon and RetailRocket are better suited to hybrid methods that combine collaborative signals with product attributes, textual reviews and implicit purchase behavior. News datasets such as MIND and EB-NeRD excel at evaluating time-sensitive methods in environments where user interests and item relevance change within hours. Social and trust-based datasets such as Epinions and FilmTrust support relational approaches, while knowledge-linked resources enable graph-based and knowledge-aware models. The authors stress that although some algorithmic ideas transfer across domains, dataset choice should be aligned with the intended scenario, because differences in interaction type, sparsity, user behavior and temporal structure can materially change experimental outcomes and the validity of the conclusions drawn from them.

The survey closes with a candid assessment of what is missing. Many datasets are static snapshots incapable of modeling preference drift, dynamic catalogs or delayed feedback. Contextual information such as time of day, location, device or social setting is frequently absent, limiting the development of genuinely context-aware recommenders. Domain coverage is lopsided: movies, music and retail are abundant while education, sustainability and mental well-being, domains of obvious social importance, remain underexplored, biasing research toward problems where data happen to be accessible rather than where they matter most. On fairness, the authors call for datasets that document their collection processes transparently, include long-tail items and less active users, and make it possible to distinguish genuine preference from biased exposure. Future collections, they argue, should embrace real-time or streaming data, richer context and representativeness by design. The recommendations are ultimately practical as well as programmatic: because dataset selection shapes algorithm behavior, reproducibility and the generalizability of every published result, the survey is intended as a roadmap for newcomers and experienced researchers alike, a guide to choosing the right data for the right question in a field whose influence over daily digital life shows no sign of diminishing.