Restc Sbr Eval

restc-sbr-eval

Spatio-Temporal Contrastive Learning Enhanced GNNs for Session-based Recommendation — Wan et al. (2022) (arXiv:2209.11461, 2022)

What this evaluates

Evaluates a model's ability to perform session-based next-item recommendation by capturing both spatial graph structures and temporal dynamics. It probes how well the model aggregates collaborative filtering signals and session-specific sequences to predict the subsequent item in a user's browsing session.

Datasets

• Tmall — total 377166; splits: train (351268), test (25898)
• Diginetica — total 780328; splits: train (719470), test (60858)
• Gowalla — total 574532; splits: train (419200), test (155332)
• RetailRocket — total 448780; splits: train (433648), test (15132)
• Nowplaying — total 915128; splits: train (825304), test (89824)
• LastFM — total 3510163; splits: train (2837330), test (672833)

Metrics

• cross-entropy (primary) — range: [0, 1]
  • Cross-entropy loss between the predicted probability distribution over candidate items and the ground truth next item: $\mathcal{L}{main} = -\sum{i=1}^{N} y_i \log(\hat{y}_i) + (1-y_i)\log(1-\hat{y}_i)$.

Input / output format

Input: A session sequence consisting of ordered item IDs and their corresponding timestamps.

Output: A probability distribution vector of size N (number of items) representing the likelihood of each candidate item being the next click.

Scoring recipe

def compute_loss(predictions, targets):
    loss = -torch.sum(targets * torch.log(predictions + 1e-9)) / targets.shape[0]
    return loss

Common pitfalls

• Data leakage when constructing the global Collaborative Filtering Graph (CFG) if future session interactions are inadvertently included in the training graph.
• Handling of directed cycles and high in-degree items, which can distort message passing in the graph encoder if not properly regularized.
• Session sparsity: short sessions (avg length ~3-7) may lack sufficient temporal signals, making spatial graph aggregation critical but prone to noise.

Evidence (verbatim from paper)

Then, we apply cross-entropy as our objective function of the main task with the ground truth ${\mathbf{y}{1},\mathbf{y}{2},\mathbf{y}{3},\ldots,\mathbf{y}{N}}:

Citation

@misc{wan2022restc,
  title={Spatio-Temporal Contrastive Learning Enhanced GNNs for Session-based Recommendation},
  author={Wan et al. (2022)},
  year={2022},
  note={arXiv:2209.11461}
}

• arXiv: 2209.11461