Global temporal and leave-last-out splitting for next-item evaluation

[hieuddo]

Description

The current NextItemEvaluation only supports pre-split data via from_splits with three datasets: Diginetica, RetailRocket, and Cosmetics; already preprocessed and uploaded in bundle train/val/test. We should move to more transparent data loading/splitting. Recent studies have also shown some flaws/issues in data splitting that lead to data leakage:

• [1] Meng et al. (2020). Exploring Data Splitting Strategies for the Evaluation of Recommendation Models. RecSys 2020.
• [2] Ji et al. (2023). A Critical Study on Data Leakage in Recommender System Offline Evaluation. ACM TOIS 2023.
• [3] Hidasi & Czapp (2023). Widespread Flaws in Offline Evaluation of Recommender Systems. RecSys 2023.

This PR adds:

• NextItemEvaluation.from_timestamps: a global temporal split, the leakage-free protocol recommended by [2], similar to the existing TimestampSplit for rating data.
• NextItemEvaluation.leave_last_out: the per-user leave-one-out split, commonly used in the sequential literature. We'll raise a warning that this splitting method leaks cross-user future information; we support this because it is still very common.
• A sort-order warning in SequentialDataset.build: warns when rows within a session are not chronological.
• Add three new amazon_review datasets: Beauty, Sports_and_Outdoors, and Toys_and_Games; which are widely used in Semantic ID papers. For now, we only implement load_feedback to use these three as normal rating data. Any load_text or load_item_metadata are for future implementation with some Semantic ID models (e.g., TIGER, RQ-VAE).

Related Issues

Checklist:

• I have added tests.
• I have updated the documentation accordingly.
• I have updated datasets/README.md (if you are adding a new dataset).

[hieuddo]

Quick runs (10 epochs) with SASRec and BERT4Rec show better performance than their corresponding results in TIGER paper

Beauty:

Model	Recall@10	NDCG@10
SASRec	0.0669	0.0356
BERT4Rec	0.0656	0.0360

Sports:

Model	Recall@10	NDCG@10
SASRec	0.0351	0.0182
BERT4Rec	0.0330	0.0174

Toys:

Model	Recall@10	NDCG@10
SASRec	0.0739	0.0376
BERT4Rec	0.0688	0.0378