SASRec: Self-Attentive Sequential Recommendation¶

简体中文

Fidelity: Full core pipeline. The model, paper's pairwise BCE training objective, and sequential inference all run. Public data is scaled from MovieLens-1M to MovieLens-100K and evaluated with a stricter full-catalog protocol.

Paper information¶

Field	Value
Paper link	arXiv 1808.09781
Organization	UC San Diego
First published	2018-08-20 (arXiv v1)
Original author code	Yes: official author code.
Adapter	`sasrec`
Local reproduction code	`src/auto_research/reproductions/sasrec/`

Original paper summary¶

Background and main changes¶

Markov chains favor recent behavior; RNNs compute step by step. SASRec uses causal self-attention to select behaviors relevant to the next item, giving a more flexible receptive field across sparse and dense data. It became a common sequential recommendation baseline.

flowchart LR
    S["Chronological item IDs"] --> E["Shared item embeddings plus positions"]
    E --> A["Causal multi-head self-attention"]
    A --> F["Point-wise ReLU FFN"]
    F --> R["Residual + LayerNorm"]
    R --> T["Dot product with shared item embeddings"]
    T --> O["Full-catalog next-item scores"]

Key figure from the paper¶

Original Figure 1 (training): Training flow and key optimization steps. Image from the original paper; copyright belongs to the authors. Click the image for its source.

Core equation¶

For sequence representation \(E\), each layer computes:

\[ \operatorname{Attn}(E)=\operatorname{softmax}\left(\frac{QK^\top}{\sqrt d}+M_{causal}\right)V, \qquad F_i=\operatorname{ReLU}(S_iW_1+b_1)W_2+b_2. \]

Input item embeddings and output classifier weights are tied. Pairwise BCE over the positive item and a sampled negative at each position preserves the paper's objective.

Offline and online results in the paper¶

Offline: with the ground-truth item and 100 sampled negatives, Hit@10/NDCG@10 are 0.4854/0.3219 on Beauty, 0.7410/0.5360 on Games, 0.8729/0.6306 on Steam, and 0.8245/0.5905 on MovieLens-1M.
Online: no production A/B is reported.

SASRec is a user-approved named exception for a classic baseline, not a change to the quantified online A/B requirement for new industrial papers.

Local reproduction¶

Control: Popularity versus SASRec. NDCG@10 falls from 0.02969 to 0.02933 (-1.24%). This compares models, not a SASRec component ablation.

MovieLens-100K has 932 eligible users and 1,682 items, with chronological leave-two-out. Seeds 42/43/44 each run 320 steps on Apple MPS. Full-catalog evaluation is substantially harder than 100 sampled negatives, so absolute scores are not comparable with the paper.

Model	Hit@10	NDCG@10	Head share@10
Popularity	0.06009	0.02969	1.00000
SASRec	0.05830 ± 0.00395	0.02933 ± 0.00051	0.97775

SASRec NDCG@10 is -1.24% relative to popularity. Three seeds give similar results without beating the strong popularity baseline. The run provides a reusable neural sequence baseline with a clear protocol and exposes head bias under a short budget.

See metrics/movielens-100k-seeds42-44.json.

Reproduction command¶

AUTO_RESEARCH_SASREC_STEPS=320 auto-research reproduce --paper sasrec --seed 42