I’m building a recommendation system and need interaction data plus a benchmark set. What should I look at?

If you need interaction data and a benchmark set for a recommendation system, you’ll usually want to look at datasets that include one or more of these signals:

• Explicit feedback: ratings, likes/dislikes, reviews
• Implicit feedback: clicks, views, purchases, watch time, listen history
• Context: timestamps, session info, device, location
• Item/user metadata: categories, text, images, demographics

Good benchmark datasets to consider

1. MovieLens

Best for: quick prototyping, collaborative filtering, ranking baselines
What it has: user-item ratings, timestamps, some movie metadata
Why use it: clean, widely used, many standard benchmark splits
Sizes: from small to very large variants

2. Amazon Reviews / Amazon Product Data

Best for: implicit/explicit feedback, large-scale item recommendation
What it has: reviews, ratings, timestamps, product categories, text
Why use it: huge and diverse; good for cold-start and text-enhanced models

3. Yelp Open Dataset

Best for: business recommendation, review-based modeling
What it has: user reviews, ratings, check-ins, business metadata
Why use it: strong benchmark for recommendation and review sentiment tasks

4. Last.fm

Best for: music recommendation
What it has: user-artist interaction logs, tags, social links in some versions
Why use it: classic implicit-feedback dataset

5. Netflix Prize data

Best for: rating prediction benchmarks
What it has: user ratings on movies
Why use it: historically important, though less commonly used now due to age and availability constraints

6. Retail / e-commerce datasets

Examples:

• Instacart Market Basket Analysis
• Alibaba/Taobao session logs
• RetailRocket
• Diginetica
• Criteo click logs

Best for: session-based, next-item prediction, purchase recommendation
What they have: clicks, carts, purchases, session sequences, timestamps

7. Open Graph / social recommendation datasets

Examples:

• Epinions
• Douban
• Foursquare
• Pinterest-style datasets

Best for: social, location, or content discovery recommendations

---

What to look for in a benchmark set

Interaction type

Decide whether you need:

• Ratings prediction → MovieLens, Yelp, Amazon
• Top-N ranking / implicit feedback → Amazon, Last.fm, RetailRocket
• Session-based recommendation → Diginetica, Yoochoose, RetailRocket
• Sequential recommendation → Amazon, Steam, Netflix-like logs

Dataset quality

Check for:

• Sparse or dense interactions
• Cold-start users/items
• Missing values
• Noise/spam reviews
• Temporal coverage
• Availability of metadata

Standard evaluation splits

Prefer datasets with:

• Predefined train/validation/test splits, or
• A clear temporal split protocol

For recommendation, temporal splits are often more realistic than random splits.

---

A practical shortlist

If you want just a few strong starting points:

• MovieLens 1M / 20M for a clean classic benchmark
• Amazon Reviews for large-scale implicit + content-rich recommendation
• Yelp Open Dataset for real-world review/business recommendation
• RetailRocket or Diginetica for session-based recommendation
• Last.fm for music recommendation

---

If you want your own benchmark

A strong benchmark set usually includes:

1. Raw interaction logs
2. User/item metadata
3. A fixed train/val/test split
4. Negative sampling protocol
5. Evaluation metrics like:
   • Recall@K
   • NDCG@K
   • MAP@K
   • Hit Rate@K
   • AUC

If you want, I can also give you:

• a dataset recommendation based on your domain,
• a comparison table of the best benchmark datasets,
• or a download/use checklist for building your evaluation pipeline.