Update benchmark results after fixing top_k/remove_seen in recommend_k_* helpers

[miguelgfierro]

Summary

Follow-up to #2322, which fixed top_k and remove_seen being ignored in several recommend_k_* helpers in examples/06_benchmarks/benchmark_utils.py.

This PR reruns movielens.ipynb with the corrected code and updates the Algorithm Comparison table in README.md accordingly. It also cleans up several issues found during the rerun:

• Remove TensorFlow import block — none of the active algorithms (ALS, SAR, NCF, embdotbias, BPR, BiVAE, LightGCN) depend on TF.
• Remove all surprise/SVD dead code from benchmark_utils.py (import + 4 functions) and movielens.ipynb (commented-out blocks in 7 cells), since SVD is not in core dependencies.
• Split the single GPU try/except import block in benchmark_utils.py into three independent blocks (LightGCN, NCF, embdotbias) so a failure in one model's imports does not silently prevent the others from loading — this was the root cause of a NameError: name 'RecoDataLoader' is not defined during the rerun.
• Update the machine description (24 CPUs, GeForce RTX 5090 Laptop GPU with 24Gb).
• Minor notebook cleanup: add colons to markdown section headers, remove orphaned commented-out stubs.
• Add RATING_THRESHOLD = 3.5 for BPR: BPR is an implicit feedback model and should not treat all ratings equally. Ratings >= 4 are binarized to 1 (positive) and ratings <= 3 are excluded. This threshold is applied consistently to both training data (via prepare_training_bpr) and test data (via prepare_metrics_bpr). The remove_seen masking still uses the full unfiltered training set so that all previously seen items are suppressed from recommendations.
• Fix BPR.recommend_k_items: seen-item masking (via -inf) is now applied to the score matrix before argpartition, guaranteeing exactly top_k valid recommendations per user. The old implementation filtered after selection, which could return fewer than top_k results when seen items appeared in the top-k window.
• Move RATING_THRESHOLD constant to module level (top of benchmark_utils.py) alongside other module constants.

Metric changes

Algo	Old MAP	New MAP	Old nDCG	New nDCG	Change
ALS	0.0047	0.0104	0.0442	0.0328	within noise
BiVAE	0.1461	0.3300	0.4751	0.4717	large MAP increase — old run used different params
BPR	0.1325	0.2292	0.4420	0.3623	see note below
embdotbias	0.0190	0.0549	0.1178	0.1190	within noise
LightGCN	0.0885	0.2772	0.4198	0.4191	within noise
NCF	0.1077	0.2603	0.3961	0.3931	within noise
SAR	0.1106	0.2585	0.3825	0.3938	within noise
SVD	removed	—	removed	—	dropped with surprise

BPR note: Two independent corrections affect the BPR result:

1. The old nDCG of 0.44 was artificially inflated because recommend_k_items was called without top_k, so the evaluator received a differently-ordered full candidate set rather than a top-10 list. Fixing this alone brought nDCG to ~0.197.
2. Applying RATING_THRESHOLD = 3.5 (ratings >= 4 as positive) is methodologically correct for BPR, which is an implicit feedback model — treating low ratings as positive inflates metrics because BPR ignores rating magnitude entirely. With the threshold, nDCG rises to 0.3623 because the model now trains on a cleaner positive signal and is evaluated against the same filtered ground truth.

SAR rating metrics (RMSE/MAE/R²) are now shown as N/A — SAR does not produce rating predictions and never should have had those columns populated.

Test plan

• python -m py_compile examples/06_benchmarks/benchmark_utils.py
• Full rerun of examples/06_benchmarks/movielens.ipynb on Movielens 100k — output committed.

FYI @he-yufeng

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[miguelgfierro]

ready to review @anargyri @SimonYansenZhao

[anargyri]

Why does SAR have no rating metrics now but before it had them?