SM-RS — Single- and Multi-Objective Recommendations

SM-RS 2.0 — the extended version of SM-RS (SIGIR'24), and a superset of v1. To our knowledge the only public RecSys dataset linking users' self-declared propensities toward beyond-accuracy objectives (relevance, diversity, novelty, exploration) with contextual impressions, item selections, and explicit perceived-quality judgments — across two domains (movies, books).

If you use this dataset, please cite BOTH papers (SM-RS 2.0, TORS 2026; and SM-RS, SIGIR 2024) — see Citation.

• 📦 Benchmark code (loaders + canonical evaluators): https://github.com/pdokoupil/SM-RS
• 🧊 Reproducibility archive: OSF — https://osf.io/wsakx (v2), https://osf.io/hkzje (v1)

Quick start

from datasets import load_dataset
behaviors = load_dataset("pdokoupil/SM-RS", "behaviors")
# or use the benchmark package: pip install sm-rs

What's inside

Core tables

Config	Rows	Description
behaviors	~10k	impression logs: userId, timestamp, history, impressions (itemId/selected/timestamp), mors, iteration
propensities	—	per-user self-declared relevance / diversity / novelty / exploration propensities (+ userId, timestamp)
objective_perceptions	—	per-impression perceived relevance/diversity/novelty/exploration (+ impressionsIds)
criteria_values	—	computed objective values per impression
comparative_diversity	—	triplet judgments (list1/list2/list3, perceived_diversities, domain)
users	—	demographics + domain (movies/books)
movies, books	—	item id maps

Derived matrices — recomputed locally, not hosted here

Scoring recommendation-list tasks (2, 3) needs item-item (relevance), distance (diversity), and popularity (novelty) matrices. These are not distributed — neither here nor on this dataset — because they derive from third-party catalogs (MovieLens, goodbooks-10k) whose terms restrict redistribution, and because they'd add gigabytes. Instead the benchmark package recomputes them from a rating matrix you build from your own download of those public datasets (smrs.derived: EASE^R item-item, 1−cosine distance, frequency popularity — all deterministic). Bit-exact originals are archived on OSF.

Tasks (benchmark)

Six tasks share this dataset; each has a canonical evaluate() in the benchmark repo (smrs.tasks.*):

#	Task	Module	Metrics
1	Propensity estimation	task1_propensity	MAE · MSE · KL
2	Results proportionality	task2_proportionality	MAE · KL · wSUM · Pearson ρ
3	Selections-aware reranking	task3_reranking	nDCG@10 · Precision@5
4	Diversity-metric definition	task4_diversity	MAE · MSE · KL
5	Perceived quality (5.1 rel / 5.2 div / 5.3 nov / 5.4 ser)	task5_perceived	MAE · MSE · Kendall τ
6	Satisfaction (6.1 / 6.2)	task6_satisfaction	MAE · MSE · Kendall τ

Leaderboard

Self-service: run the canonical evaluate() and open a PR adding your row (with a repro link). The paper baselines below are the rows to beat (from SM-RS 2.0, TORS 2026, Tables 2–4).

Task 1 — Propensity estimation · MAE / MSE / KL (↓)

Method	MAE	MSE	KL
Locally normalized	0.150	0.040	0.439
Globally normalized	0.149	0.039	0.271

Task 2 — Results proportionality · MAE↓ / KL↓ / wSUM↑ / ρ↑

Method	MAE	KL	wSUM	ρ
FAI	0.181	0.374	0.568	−0.033
Probabilistic FAI	0.141	0.242	0.568	0.318
Incremental Weighted Average	0.145	0.245	0.580	0.297
RLProp	0.114	0.181	0.549	0.356
GreedyLM	0.154	0.278	0.582	0.256
DQNMORS	0.230	0.869	0.447	0.013

Task 3 — Selections-aware reranking · nDCG@10 / Precision@5 (↑)

Method	nDCG@10	Precision@5
Popularity based	0.406	0.244
EASE^R	0.368	0.225
VAE	0.292	0.174
MultVAE	0.328	0.189
ImplicitMF	0.337	0.210
UserKNN	0.367	0.224
Identity (no reranking)	0.215	0.131

Task 4 — Diversity-metric definition · MAE / MSE / KL (↓)

Method	MAE	MSE	KL
CF-ILD	0.470	0.316	0.279
CB-ILD	0.292	0.116	0.282
CF-ILMD	0.409	0.246	0.272
CB-ILMD	0.287	0.114	0.278
BIN-DIV	0.294	0.127	0.288

Task 5 — Perceived quality · MAE↓ / MSE↓ / Kendall τ↑

5.1 relevance

Method	MAE	MSE	τ
Popularity	0.205	0.079	0.228
Relevance-based	0.191	0.063	0.263
Linear Regression	0.235	0.085	0.080
LSTM	0.189	0.061	0.022

5.2 diversity

Method	MAE	MSE	τ
CF-ILD	0.270	0.116	0.171
CB-ILD	0.255	0.081	−0.004
CF-ILMD	0.227	0.074	0.146
CB-ILMD	0.346	0.170	−0.025
BIN-DIV	0.421	0.242	0.069
Linear Regression	0.222	0.061	0.196
LSTM	0.218	0.066	0.071

5.3 novelty

Method	MAE	MSE	τ
Popularity complement	0.369	0.200	0.233
Linear Regression	0.259	0.104	0.143
LSTM	0.261	0.089	−0.008

5.4 serendipity

Method	MAE	MSE	τ
CF-Unexpectedness	0.303	0.142	0.128
CB-Unexpectedness	0.250	0.077	0.054
CF-EASE	0.283	0.117	0.068
CB-EASE	0.348	0.179	0.040
1−PMI	0.336	0.170	0.068
Linear Regression	0.270	0.103	0.039
LSTM	0.261	0.093	−0.031

Task 6 — Satisfaction · MAE↓ / MSE↓ / Kendall τ↑

Method	MAE	MSE	τ	Subtask
Popularity based	0.233	0.100	0.156	6.1
Relevance based	0.223	0.083	0.145	6.1
Linear Regression	0.255	0.102	0.045	6.1
LSTM	0.210	0.072	−0.023	6.1
LSTM Combined	0.138	0.035	0.515	6.2

The Linear Regression and popularity/relevance baselines are reproduced by the benchmark package (examples/reproduce_perceived.py) — MAE matches the paper.

What's new vs. SM-RS 1.0 / getting the v1 subset

SM-RS 2.0 is a superset of the SIGIR'24 release (adds perceived-quality judgments and more tasks). The v1 subset is recoverable from the same tables — see the benchmark repo for the column mapping.

Licensing

Collected study data (this dataset): CC-BY-4.0 — free to use with attribution (cite both papers). The dataset references catalog items by ID only and does not redistribute MovieLens or goodbooks-10k. Derived matrices are recomputed by the benchmark from your own copies of those datasets, obtained under their licenses (MovieLens: no redistribution; goodbooks-10k: per its repository license).

Citation

@article{dokoupil2026smrs2,
  author  = {Dokoupil, Patrik and Peska, Ladislav},
  title   = {SM-RS 2.0: User-perceived Qualities of Single- and Multi-Objective Recommender Systems},
  journal = {ACM Transactions on Recommender Systems}, volume = {4}, number = {3}, year = {2026},
  doi = {10.1145/3754459}
}
@inproceedings{dokoupil2024smrs,
  author    = {Dokoupil, Patrik and Peska, Ladislav and Boratto, Ludovico},
  title     = {SM-RS: Single- and Multi-Objective Recommendations with Contextual Impressions and Beyond-Accuracy Propensity Scores},
  booktitle = {Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval},
  series    = {SIGIR '24}, pages = {988--995}, year = {2024}, doi = {10.1145/3626772.3657863}
}