SessionRec: Predict the next session instead of the next item

简体中文

Fidelity: core mechanism reproduction. Runs within-session pooling, a cross-session Transformer, multi-positive retrieval and same-session hard negatives. Private features and ANN serving are not reproduced.

Paper information

Field	Details
Paper	Paper or original source
Institution	Meituan
First public release	2025-02-14 (arXiv v1)
Author code	No official author code identified (checked 2026-08-09).
Local adapter or method key	sessionrec
Local reproduction code	src/auto_research/reproductions/sessionrec/

Paper method and evidence

Item-by-item training forces an order onto a single consumption intent. SessionRec uses an item-based session encoder (ISE) to pool a visit into a token, then a session-based sequence encoder (SSE) to learn changes across visits. One query retrieves every positive from the next session. Mean pooling beats other paper ISE variants; SSE can be Transformer or HSTU. The negative loss uses genuine next-session exposures without positive feedback.

flowchart LR
  A[Exposure sessions] --> B[Mean item pooling]
  B[Mean item pooling] --> C[Session sequence encoder]
  C[Session sequence encoder] --> D[Next-session multi-positive retrieval]
  D[Next-session multi-positive retrieval] --> E[Hard-negative loss]

Original paper figure

Figure 2 shows the architecture and how its modules connect. From the original paper; copyright remains with the authors.

Core equation

\[ s_t=\frac{1}{|S_t|}\sum_{i\in S_t}e_i,\qquad q_t=\operatorname{SSE}(s_1,\ldots,s_t)_{t}. \]

\[ \mathcal L=\mathcal L_{multi\text{-}positive\ retrieval} +\alpha\frac{1}{|P||N|}\sum_{p\in P,n\in N}\log(1+e^{q^\top e_n-q^\top e_p}). \]

\(P_{t+1}\) contains every next-session positive and \(N_{t+1}\) the next-session unengaged exposures.

Reported results

In KuaiSAR leave-one-session-out testing, SessionRec-HSTU Recall@500/NDCG@500 is 0.3318/0.1287, vs 0.2009/0.0701 for HSTU and 0.2212/0.0785 for HSTU+. A seven-day Meituan homepage A/B replacing SASRec reports Pay PV +0.603% and PVCTCVR +0.564%.

Local reproduction

KuaiRand-Pure standard-traffic exposures are sessionized at 30 minutes. is_click=1 or long_view=1 is positive; other same-session exposures are hard negatives. Each user's last two valid sessions are validation/test. Dimension, layers, steps and sampling match the flat item Transformer. Training cycles through individual positive CE targets and evaluation retains all positives. MovieLens lacks session/exposure information, and DIN is not a baseline here. Seed-42 validation searches \(\alpha\in\{0,0.01,0.1\}\); 0 and 0.01 tie at NDCG@20 0.002856, so 0.01 retains the hard-negative objective without selecting on test.

Method	Recall@20 mean ± SD	NDCG@20 mean ± SD	Train + eval seconds
Item Transformer	0.011894 ± 0.002272	0.004718 ± 0.000924	11.32
SessionRec Transformer	0.009302 ± 0.001201	0.003678 ± 0.000640	10.53

NDCG@20 drops 22.05% with only 1/3 seeds improving, so the paper's gain is not reproduced. Few usable sessions per user, mixed interests within short-video sessions and lost order through mean pooling are plausible explanations, not demonstrated causes. Stable metrics. Archives, extracted data, runs and checkpoints remain ignored.

pip install -e '.[neural-recs]'
for seed in 42 43 44; do
  AUTO_RESEARCH_SESSIONREC_ROWS=350000 \
  AUTO_RESEARCH_SESSIONREC_USERS=2000 \
  AUTO_RESEARCH_SESSIONREC_STEPS=180 \
  auto-research reproduce --paper sessionrec --dataset-dir data --seed "$seed"
done