4 Recommended Benchmark Datasets Compared in 2026: Why Tenrec's Scale and True Negative Feedback Design Stand Out

I've seen models produce impressive numbers in papers only to fall short when deployed in real products—more than once.

The problem usually isn't the model; it's the dataset you're using. MovieLens gives you ratings, Amazon gives you reviews, MIND gives you news clicks—most of them have only one type of user feedback.

Tenrec was built to address this. It is real behavior data from two of Tencent's feed recommendation apps, covering four recommendation scenarios, with data and code publicly released, published on arXiv in 2022 (ID 2210.10629).

Let's be clear: this article is not a benchmark run comparing the four datasets; it's a side-by-side comparison of their specifications. When choosing a dataset, looking at specs is usually more time-efficient than looking at benchmark scores.

Tenrec vs. Common Recommender System Datasets: Quick Overview

Tenrec is a collection of real user behaviors from Tencent's feed products, covering four recommendation scenarios in one dataset.

MovieLens is the most common rating dataset in academia, Amazon Reviews is e-commerce review data, and MIND is news recommendation click data released by Microsoft.

Here's a simple table summarizing the differences:

Item	Tenrec	MovieLens	Amazon Reviews	MIND
Core focus	Real behaviors with multi-scenario, multi-type feedback	Rating prediction and collaborative filtering	E-commerce reviews and category recommendation	News recommendation and click prediction
Number of scenarios	4	Mainly single movie rating	Multiple categories but single task	Single news scenario
Positive feedback types	Clicks, likes, shares, follows, etc.	Ratings	Ratings and reviews	Clicks
True negative feedback	Yes	No	No	No
Scale	~5M users, 140M interactions	Relatively small	Depends on version	Medium

Next, I'll compare them across five dimensions, each addressing the same question: can this dataset test what you really want to solve?

Dimension 1: Scale and Scenario Coverage—Small Datasets Can't Reveal Real System Issues

Rating datasets like MovieLens are clean and easy to process; running an experiment takes only minutes. The trade-off is their small scale and single rating behavior.

Models that perform well on such datasets often lose relevance when applied to real large-scale systems—a problem the Tenrec paper points out right at the beginning.

Tenrec's scale is about 5 million users and 140 million interactions, split across four different recommendation scenarios, including two feed types: video and article.

The authors state that Tenrec is one of the largest datasets in the recommender systems field, covering most recommendation scenarios and tasks.

→ This dimension: Tenrec wins. To validate model performance at real scale, it is more convincing than small rating datasets.

Dimension 2: True Negative Feedback—Evaluating with Only Positive Samples Is Like Tuning Ranking Blindfolded

Most recommendation datasets contain only positive signals. You get a bunch of "what users liked" but no "what users explicitly didn't want."

This forces models to train under one-class settings, and evaluation can only compare who ranks higher, without knowing whether the model truly learned to distinguish preferences.

Tenrec explicitly includes true negative feedback. This is one of its most substantial differences from MovieLens and MIND.

For those working on ranking models, this difference is practical: with both positive and negative samples in the same dataset, the model's discriminative power becomes meaningful.

→ This dimension: Tenrec wins, and by a clear margin.

Dimension 3: Overlapping Users and Items Across Scenarios—Can Cross-Domain Recommendation Really Be Tested?

Cross-domain recommendation has been popular in recent years, but most datasets have no "cross-domain" at all—only one scenario. Where would you cross to?

Amazon Reviews has many categories, but it's closer to multi-task rather than behavior transfer of the same users across different product scenarios.

Tenrec's four scenarios share overlapping users and items. That is, the same person has behavior records in different scenarios, which is exactly the structure needed for cross-domain recommendation, multi-task learning, and lifelong learning.

→ This dimension: Tenrec wins. For cross-scenario research, it is almost the most direct option available.

Dimension 4: Types of Positive Feedback—Click, Like, Share, Follow Are Not the Same

Many datasets treat "clicked" as "liked." But in real products, clicking, liking, and sharing represent completely different levels of intensity.

Click datasets like MIND are at a disadvantage here: they only have impression and click signals.

Tenrec records multiple types of positive feedback, including clicks, likes, shares, follows, etc. This allows you to do more: treat different feedback as multi-task objectives, weight feedback intensity, or study the conversion from weak to strong signals.

Note that the data volume for different feedback types is inherently uneven—clicks far outnumber shares—so you'll need to handle sample imbalance during modeling.

→ This dimension: Tenrec wins, but remember to check the distribution of each feedback type first.

Dimension 5: Feature Fields and Reproducibility—Without Extra Features, Models Can Only Compare Embeddings

MovieLens basically gives you user ID, item ID, rating, and timestamp. Models can only play with embedding dimensions and network architecture.

In addition to user ID and item ID, Tenrec includes extra feature fields, which is friendlier for feature engineering and feature interaction research.

Regarding reproducibility, the Tenrec paper states that all datasets and code have been released, with standard citation formats provided, and multiple community implementations are available for reference.

It's worth noting that the paper mentions the names of the two production systems have been anonymized; if you need to know the names for research, you can contact the authors as explained in the paper.

→ This dimension: Tenrec wins.

What Unique Advantages Does Each Have?

Tenrec's advantages are mainly concentrated in these points:

First, it is a rare public dataset that simultaneously possesses large scale, multiple scenarios, multiple feedback types, and true negative samples—a combination that is not easy to find.

Second, overlapping users and items across scenarios allow it to directly support cross-domain recommendation and multi-task learning research without having to piece together data yourself.

Third, data and code are released together, with accumulated community implementations, lowering the barrier for those wanting to reproduce experiments.

As for the advantages of other datasets, we should be honest:

MovieLens is clean and lightweight, suitable for quick idea validation and teaching, with the lowest experiment cost.

Amazon Reviews and MIND have mature evaluation conventions in their respective domains (e-commerce reviews, news clicks); if your topic is closely tied to these scenarios, using them is more appropriate.

Conclusion: Which Dataset Should You Choose?

If your research requires an experimental environment for cross-scenario, cross-domain, or lifelong learning, Tenrec is currently the most complete option.

If you're dealing with real ranking problems and need negative samples to evaluate model discriminative power, Tenrec's true negative feedback design is far more practical than one-class datasets.

If you want to study user feedback of different intensities (click vs. share vs. follow), Tenrec's multi-type signals can directly support that.

If you just want to quickly validate a collaborative filtering idea or do a classroom teaching demo, lightweight datasets like MovieLens remain the easiest.

In one sentence: Tenrec fills exactly the gaps that existing benchmark datasets have long had—scale, number of scenarios, and negative samples. For researching real-world recommendation problems, it deserves a top spot on your list.

FAQ

When was Tenrec released?

The paper was published on arXiv in 2022, ID 2210.10629, by Guanghu Yuan and 11 other authors.

Where does Tenrec's data come from?

It comes from real user behaviors of two of Tencent's feed recommendation apps, covering four recommendation scenarios. The paper mentions that the names of the two production systems have been anonymized.

Does Tenrec only have click data?

No. In addition to clicks, it includes multiple positive feedback types such as likes, shares, and follows, and also contains true negative feedback.

How large is Tenrec?

Official description: about 5 million users and 140 million interactions; the paper calls it one of the largest datasets in the recommender systems field.

What are the license and citation requirements?

The paper states that the dataset and code are publicly released and should be cited when used. The standard citation is Yuan, Guanghu et al., "Tenrec: A Large-scale Multipurpose Benchmark Dataset for Recommender Systems," arXiv preprint arXiv:2210.10629, 2022.

What research topics is Tenrec suitable for?

According to the paper and official description, it can support collaborative filtering, cross-domain recommendation, multi-task learning, lifelong learning, and most popular recommendation tasks; however, you should first confirm whether the data volume for the corresponding scenario is sufficient for your specific topic.

References

• 2210.10629 Tenrec: A Large-scale Multipurpose Benchmark Dataset for Recommender Systems
• Tenrec: A Large-scale Multipurpose Benchmark Dataset for Recommender Systems
• Tenrec
• Tenrec: A Large-scale Multipurpose Benchmark Dataset for Recommender Systems | OpenReview
• Tenrec: A Large-scale Multipurpose Benchmark Dataset for Recommender Systems | alphaXiv