Sequential recommendation models typically predict a user's next action from recent session interactions, which limits personalization when short-term signals are weak or ambiguous. Rather than assuming that incorporating longer-term history and community signals is universally beneficial, we study when it actually helps. We propose BehaveRec, a three-signal framework combining a causal session transformer, a bidirectional User History Encoder over a 90-day window, and a GraphSAGE module over a user-user graph built from behavioral embedding similarity rather than item co-occurrence; a learned gate blends the three signals per user. On an industrial dataset of 137M sequences from 12.6M users, BehaveRec improves R@5 by +3.2% over SASRec, with a larger-capacity SASRec baseline (more parameters than BehaveRec) still trailing by +2.5%, ruling out capacity as the explanation. More importantly, a user-level stratification by history length reveals a clear dose-response relationship: BehaveRec's advantage is negligible below roughly 10 interactions and grows monotonically with history richness, reaching +0.017 R@5 for users with 101+ events. This pattern is corroborated across one industrial and three public datasets spanning history-rich to history-sparse regimes, where session-only models remain preferable. These headline results, and the resulting deployment threshold, are established for category-level intent prediction; a preliminary, single-run check on the full item-level vocabulary yields a directionally mixed result and is scoped as an open question rather than a claim. We use these results to derive a practical deployment rule -- roughly 20 interactions per user -- for when long-term history and behavioral-graph signals are worth their added complexity. Despite this added complexity, BehaveRec remains lightweight: it uses 5.67M parameters and trains on a single GPU in under two days.
Figures & tables
Dataset
∣C∣
Users
Avg hist
Med hist
MovieLens ML-1M
18
6,040
165.6
96.0
CDs & Vinyl
310
23,970
17.2
11.0
Electronics
673
39,991
14.0
11.0
TABLE I: Public benchmark statistics.
Dataset
Model
R@1
R@5
R@10
N@5
N@10
Industrial Dataset
Repeat-Last
0.4548
0.4548
0.4548
0.4548
0.4548
BERT4Rec [ 2 ]
0.4998
0.7032
0.7752
0.6182
0.6234
SASRec [ 1 ]
0.4976
0.6975
0.7757
0.5969
0.6154
BehaveRec V5
0.5010
0.7201
0.7772
0.6211
0.6396
MovieLens ML-1M ‡
Repeat-Last
0.3853
0.5517
0.6575
0.4854
0.5177
BERT4Rec [ 2 ]
0.4162
0.8672
0.9631
0.6611
0.6925
TABLE II: Recommendation performance. Bold = best per dataset. Amazon uses category-level prediction (TLSAN benchmark [ 18 ] ). † Amazon avg history 14–17 events; ‡ MovieLens: single run; five-seed mean and per-seed breakdown in Table IX . See Section V-G for cross-domain analysis.
90-day history
SASRec
V5
Δ R@5
1–5
0.612
0.610
−0.002
6–10
0.638
0.639
+0.001
11–20
0.662
0.666
+0.004
21–50
0.691
0.699
+0.008
51–100
0.716
0.728
+0.012
101+
0.731
0.748
+0.017
TABLE III: User-level history stratification on the industrial test set. R@5 by 90-day history length bin, same domain and vocabulary throughout (controls for dataset-level confounds). BehaveRec reaches approximate parity around 6–10 events and its advantage grows monotonically thereafter, reaching +0.017 R@5 at 101+ events. SASRec is preferable only in the sparsest bin (1–5 events).
Fig. 1 : Dose-response relationship between 90-day history length and BehaveRec’s R@5 advantage over SASRec. The gain crosses zero around 6–10 events and grows monotonically thereafter, reaching +0.017 R@5 in the 101+ bin. Data as in Table III .
Sequence type
SASRec
V5
Δ R@5
N
Continuation (45.5%)
0.7471
0.7523
+0.0052
1,356,496
Shift (54.5%)
0.6561
0.6932
+0.0371
1,625,835
TABLE IV : BehaveRec’s advantage over SASRec, stratified by whether the ground truth continues the last session cluster or shifts to a new one (Section V-D ). Single run, no confidence intervals.
Hist. 1–10
Hist. 11–50
Hist. 51+
N
Session 1–3 events
+0.0051
+0.0182
+0.0251
1,268,267
Session 4+ events
−0.0009
+0.0091
+0.0189
1,714,064
TABLE V : Δ R@5 (V5 − SASRec) jointly stratified by session length and 90-day history length. Single run, no confidence intervals.
Model
R@5
HR@10
NDCG@10
SASRec (session-only)
0.4324
0.8236
0.5914
BehaveRec V4 (UHE + session)
0.4218
0.8346
0.5934
TABLE VI : Item-level (full-vocabulary) evaluation on MovieLens ML-1M, ∣I∣=3,706 . Single run, no confidence intervals; reported as a preliminary, directionally mixed result (see discussion below).
Method
R@5
N@5
Δ R@5 vs. V4
(A) No graph (V4)
0.7163
0.6175
–
(B) Item co-occurrence (LightGCN-style)
0.7188
0.6197
+0.0025
(C) Behavioral embedding (V5, ours)
0.7201
0.6211
+0.0038
TABLE VII: Graph construction ablation. The same BehaveRec architecture is used throughout; only the edge definition differs. Evaluated on a 50-part test set (seed=99). All differences are significant at p<0.05 . Edge definitions are swapped at inference using the same trained backbone; variants are not independently retrained from scratch.
#
Variant
R@5
NDCG@5
Δ R@5
0
Full BehaveRec V5
0.7201
0.6211
–
1
w/o graph (= V4)
0.7163
0.6175
−0.0038
2
w/o UHE (= SASRec)
0.6975
0.5969
−0.0226
3
w/o InfoNCE ( λcon=0 )
0.7124
0.6148
−0.0077
4
w/o ThreeWayGate
0.7127
0.6146
−0.0074
5
w/o weight tying
0.7126
0.6145
−0.0075
TABLE VIII: Ablation study on the industrial dataset. All rows trained under the same protocol as BehaveRec V5 (Section IV-I ).
Model
s=42
s=99
s=123
s=456
s=789
Mean
Std
SASRec
0.8790
0.8770
0.8770
0.8785
0.8778
0.8778
0.0009
BehaveRec V5
0.8825
0.8843
0.8845
0.8838
0.8846
0.8839
0.0009
Δ
+0.0035
+0.0073
+0.0075
+0.0053
+0.0068
+0.0061
0.0017
TABLE IX: MovieLens ML-1M R@5 across five seeds. BehaveRec V5 outperforms SASRec on all five seeds ( +0.0061 mean, std 0.0017 ). Absolute values differ slightly from the single run reported in Table II due to run-to-run variation; the direction and approximate magnitude of the improvement are consistent between the two.
Metric
Lift
Surface
Annualized GMV
+\2.4$ M
Item + post-add-to-cart
post-add-to-cart page views
+55 bps
post-add-to-cart
Third-party add-to-cart
+300 bps
post-add-to-cart
New category adoption (10-day)
∼ 2%
Item + post-add-to-cart
Consumables GMV
+0.9%
Home (iOS)
Sponsored impressions
+0.85% / +0.32%
Home iOS/Android
TABLE X : Left: Production A/B test results (lift vs. existing carousel baselines). Right: Signal attribution across 12 illustrative test customers.
Session len
V4 R@5
V5 R@5
Δ
1 event
0.541
0.552
+0.011
2–3 events
0.612
0.619
+0.007
4–6 events
0.701
0.704
+0.003
7–10 events
0.768
0.769
+0.001
11–20 events
0.812
0.812
+0.000
TABLE XI: R@5 by session length: V5 (behavioral graph) vs. V4 (no graph). Graph contribution is largest for cold sessions and diminishes as session evidence strengthens.
Sessionized interaction histories contain behavioral patterns that can improve sequential recommendation. However, existing models process all sessions through the same parameterized blocks, regardless of their behavioral differences. Mixture of Experts (MoE) enables conditional computation, but it leaves open what should guide expert allocation. We propose RouteRec, a sequential recommender that uses observed session behavior as the routing criterion. RouteRec summarizes four types of behavioral evidence from sessionized histories: interaction tempo, item-group focus, repetition and carryover, and popularity tendency. It uses these cues to route computation at macro, mid, and micro scopes. Cue-derived scores first select expert groups; within each selected group, the current backbone state then refines expert selection. Across six public datasets and 18 dataset-metric combinations, RouteRec ranks first in 12 and second in three, yielding the best overall average rank of 1.61 compared with 4.11 for the next-best baseline. Additional analyses suggest that the behavioral cues guide expert allocation beyond added capacity and produce routing patterns aligned with observed behavior. Our code is available at https://github.com/jy1559/RouteRec
Junyeong Song, Jaemin Yoo
Korea Advanced Institute of Science and Technology Daejeon, Republic of Korea · Seoul National University Seoul, Republic of Korea
Modeling long sequences of user behaviors has emerged as a critical frontier in generative recommendation. However, existing solutions face a dilemma: linear attention mechanisms achieve efficiency at the cost of retrieval precision due to limited state capacity, while softmax attention suffers from prohibitive computational overhead. To address this challenge, we propose HyTRec, a model featuring a Hybrid Attention architecture that explicitly decouples long-term stable preferences from short-term intent spikes. By assigning massive historical sequences to a linear attention branch and reserving a specialized softmax attention branch for recent interactions, our approach restores precise retrieval capabilities within industrial-scale contexts involving ten thousand interactions. To mitigate the lag in capturing rapid interest drifts within the linear layers, we furthermore design Temporal-Aware Delta Network (TADN) to dynamically upweight fresh behavioral signals while effectively suppressing historical noise. Empirical results on industrial-scale datasets confirm the superiority that our model maintains linear inference speed and outperforms strong baselines, notably delivering over 8% improvement in Hit Rate for users with ultra-long sequences with great efficiency.
Lei Xin, Yuhao Zheng, Ke Cheng +3
Shanghai Dewu Information Group · Wuhan University · USTC +1
Modern recommender systems leverage ultra-long user behavior sequences to capture dynamic preferences, but end-to-end modeling is infeasible in production due to latency and memory constraints. While summarizing history via interest centers offers a practical alternative, existing methods struggle to (1) identify user-specific centers at appropriate granularity and (2) accurately assign behaviors, leading to quantization errors and loss of long-tail preferences. To alleviate these issues, we propose Hierarchical Sparse Activation Compression (HiSAC), an efficient framework for personalized sequence modeling. HiSAC encodes interactions into multi-level semantic IDs and constructs a global hierarchical codebook. A hierarchical voting mechanism sparsely activates personalized interest-agents as fine-grained preference centers. Guided by these agents, Soft-Routing Attention aggregates historical signals in semantic space, weighting by similarity to minimize quantization error and retain long-tail behaviors. Deployed on Taobao's "Guess What You Like" homepage, HiSAC achieves significant compression and cost reduction, with online A/B tests showing a consistent 1.65% CTR uplift -- demonstrating its scalability and real-world effectiveness.
Kun Yuan, Junyu Bi, Daixuan Cheng +5
Alibaba Group · Beijing, China · Renmin University