qhjqhj00 rec-splitting-eval
Exploring Data Splitting Strategies for the Evaluation of Recommendation Models — Meng et al. (2020) (arXiv:2007.13237, 2020)
What this evaluates
Evaluates how different data splitting strategies (leave-one-last-item, leave-one-last-basket, temporal global) impact the performance ranking of recommendation models on e-commerce datasets. It probes whether evaluation protocols introduce temporal leakage or distribution shifts that confound model comparisons and invalidate cross-paper rankings.
Datasets
- Tafeng Dataset — total ?; splits: train (-1), test (-1)
- Dunnhumby Dataset — total ?; splits: train (-1), test (-1)
Metrics
NDCG@10(primary) — range: [0, 1]- Normalized Discounted Cumulative Gain at rank 10. Measures the quality of the ranked recommendation list by applying a logarithmic discount to gains based on position, normalized by the ideal DCG.
Recall@10— range: [0, 1]- Fraction of relevant (ground truth) items successfully retrieved within the top 10 recommended items for each user.
Input / output format
Input: User-item interaction histories (often structured as baskets or sessions) with implicit timestamps, used to train collaborative filtering or neural recommendation models.
Output: A ranked list of candidate items for each user, typically truncated to top-K for evaluation against held-out ground truth interactions.
Scoring recipe
def ndcg_at_k(relevant, pred, k=10):
dcg = sum(1.0 / math.log2(i + 2) for i, item in enumerate(pred[:k]) if item in relevant)
idcg = sum(1.0 / math.log2(i + 2) for i in range(min(len(relevant), k)))
return dcg / idcg if idcg > 0 else 0.0
def recall_at_k(relevant, pred, k=10):
hits = sum(1 for item in pred[:k] if item in relevant)
return hits / len(relevant) if relevant else 0.0
Common pitfalls
- Temporal leakage occurs when future interactions are inadvertently included in training due to non-temporal splits, artificially inflating performance.
- Ranking swaps between models are highly sensitive to the splitting strategy, making cross-paper comparisons invalid without standardized protocols.
- Leave-one-last-basket splits group interactions by session/basket, fundamentally changing the evaluation distribution compared to item-level splits.
Evidence (verbatim from paper)
Figure 2 plots the NDCG@10 performance of these models for pairs of splitting strategies across each of the two datasets, as well as reporting Kendall's $ au$ correlation between the score distributions for each.
Citation
@misc{meng2020exploring,
title={Exploring Data Splitting Strategies for the Evaluation of Recommendation Models},
author={Meng et al. (2020)},
year={2020},
note={arXiv:2007.13237}
}
- arXiv: 2007.13237