Efficient sequential recommendation has emerged as a promising paradigm for providing real-time personalized service. To deploy models under resource-constrained environment, model compression becomes a necessity, yet existing methods typically apply uniform compression ratios, overlooking the heterogeneity in user intent complexity. This "one-size-fits-all" approach leads to a fundamental Capacity-Complexity Mismatch, where computational resources are either wasted on simple instances or insufficient for complex ones. To address this, we propose RACE (Ranking-uncertainty Aware Capacity-Adaptive Customization for Efficient Sequential Recommendation), a framework that dynamically adapts both model structures and input sequences with uncertainty-aware compression ratios. To ensure capacity-adaptive inference, RACE captures the model's intrinsic uncertainty of its predictions and transforms it into a proactive resource allocator. Specifically, we derive a list-wise ranking-uncertainty metric from the distributional stability of top-K predictions, which serves as a measurable proxy for instance complexity. A lightweight predictor is then trained via self-supervision to anticipate this uncertainty from early-layer features. Guided by this signal, RACE performs instance-specific slimming in two dimensions: a hypernetwork-style channel sensitivity extractor identifies parameter redundancy, while importance-aware progressive sequence filtering mitigates data redundancy. Extensive experiments on four popular datasets demonstrate that RACE outperforms state-of-the-art baselines on accuracy, efficiency, and adaptability.
Figures & tables
Figure 1. Comparison of on-device recommendation paradigms. (a) On-device Finetuning incurs computational overhead due to backpropagation, and often falls short for capturing shifting interests. (b) On-device Pruning applies a unified-sparsity mask and relies on additional extractors, unable to adapt to the complexity of user intents. (c) Our proposed RACE framework introduces a Ranking-Uncertainty Driven mechanism to dynamically adapt model size to instance complexity. In comparison, RACE achieves sequence and structure slimming and customization in an intrinsic feature reuse and retraining-free manner.
Figure 2. Correlation analysis between model performance and two input-side heuristics: (a) history popularity; (b) interest stability. Neither reliably indicates instance complexity.
Figure 3. Overview of the RACE framework. Our approach consists of three key components: (a) Ranking-Uncertainty Guided Resource Allocation : RACE designs a list-wise ranking-uncertainty metric and transforms it into a proactive instance-complexity proxy to guide resource allocation, utilizing a self-supervised threshold predictor to anticipate ratios early in inference. (b) Interest-Compatible Subnets Identification : Using the predicted ratio, a lightweight hypernetwork-style sensitivity extractor generates channel-wise masks to customize each feed-forward layer. (c) Progressive Sequence Filtering : Using the predicted ratio, RACE calculates relative importance scores to retain semantically relevant items while suppressing noise block by block.
Model
Method
MovieLens-100k
Douban
Amazon-CD
Amazon-Game
NDCG ↑
Hit ↑
FLOPs ↓
NDCG ↑
Hit ↑
FLOPs ↓
NDCG ↑
Hit ↑
FLOPs ↓
NDCG ↑
Hit ↑
FLOPs ↓
SASRec
Normal
0.0299
0.0471
1.8432
0.0141
0.0193
1.8432
0.0250
0.0295
1.8432
0.0334
0.0418
1.8432
Finetune
0.0268
0.0442
1.8432
0.0131
0.0174
1.8432
0.0251
0.0297
1.8432
0.0338
0.0424
1.8432
SNIP
0.0285
0.0464
0.8352
0.0109
0.0153
0.8352
0.0240
0.0280
0.8352
0.0331
0.0401
0.8352
GEM
0.0330
0.0516
0.8352
0.0144
0.0199
0.8352
0.0263
0.0312
0.8352
0.0343
0.0422
0.8352
APG
0.0338
0.0536
0.8947
0.0224
0.0305
0.8947
0.0267
0.0312
0.8947
0.0385
0.0468
0.8947
Table 1. Overall performance comparison of RACE and competitive baselines across four datasets and two representative backbones. For metrics NDCG@K and Hit@K , K=5 is used by default. ↑ and ↓ denote that larger and smaller values are preferred, respectively. Bold font indicates the best performance in each column.
Model
Ablated Model
MovieLens-100k
Douban
NDCG ↑
Hit ↑
NDCG ↑
Hit ↑
SASRec
w/o uncertainty
0.0299
0.0471
0.0141
0.0193
w/o customization
0.0263
0.0447
0.0139
0.0193
w/o sequence filtering
0.0388
0.0592
0.0229
0.0296
w/o subnet identification
0.0296
0.0492
0.0104
0.0171
w/o dynamic ratio
0.0376
0.0584
0.0250
0.0321
Table 2. Ablation experiment.
Uncertainty Design
SASRec
BSARec
NDCG ↑
Hit ↑
NDCG ↑
Hit ↑
MC Dropout
0.0216
0.0292
0.0284
0.0366
Ensemble
0.0197
0.0266
0.0261
0.0344
Mean Variance
0.0248
0.0317
0.0310
0.0415
Entropy
0.0227
0.0311
0.0296
0.0389
Pairwise Margin
0.0228
0.0289
0.0323
0.0407
Table 3. Comparison of alternative uncertainty designs for guiding dynamic thresholds on Douban.
Model
Method
Douban
Amazon-Game
NDCG ↑
Hit ↑
NDCG ↑
Hit ↑
SASRec
Mean
0.0238
0.0312
0.0393
0.0471
Masked Mean
0.0241
0.0318
0.0390
0.0474
RACE
0.0261
0.0337
0.0402
0.0490
BSARec
Mean
0.0310
0.0395
0.0472
0.0567
Masked Mean
0.0307
0.0397
0.0477
0.0568
Table 4. Comparison of different sequence-filtering methods on Douban and Amazon-Game datasets.
Figure 4. Uncertainty Analysis: (a) Performance (before or after customization) vs. Uncertainty (b) Predicted Channel Pruning Ratio vs. Uncertainty.
Figure 5. Training Curve of RACE over Five Seeds.
Figure 6. Sensitivity Analysis of Sequence Pruning Ratio.
Figure 7. Sensitivity Analysis of Channel Pruning Ratio.
Method
Backbone FLOPs ↓
Other FLOPs ↓
Latency ↓
Finetune
1.8432
3.6864
9.732
DIET
0.9010
3.0597
10.518
RACE
0.7549
0.0490
5.780
Table 5. Efficiency analysis on the SASRec backbone.
Dataset
Test-time Constraint
NDCG ↑
Hit ↑
FLOPs ↓
Douban
Stricter than training
0.0259
0.0337
0.6266
Same as training
0.0261
0.0337
0.7928
Looser than training
0.0262
0.0341
1.0408
CD
Stricter than training
0.0287
0.0328
0.6718
Same as training
0.0288
0.0329
0.7549
Looser than training
0.0288
0.0330
0.8567
Table 6. Adaptivity analysis under varying test-time constraints on the SASRec backbone.
Figure S1 . Sensitivity of NDCG@5 and Hit@5 to loss coefficients λ1 and λ2 on BSARec with Amazon-CD. Stars mark the default setting used in the main experiments.
Transformers have redefined sequential recommendation by effectively modeling dynamic user behaviors and long-range dependencies. However, they remain inherently inefficient: standard architectures operate at a fixed rate, allocating comparable computation to every item in a user's history regardless of its information content. This leads to prohibitive computational overhead on long sequences and increased sensitivity to behavioral noise. To address this, practitioners often resort to lossy sequence compression, staged modeling, or truncation. This limits the model's ability to leverage the full context of long histories during inference. Inspired by the recent success of Byte Latent Transformer, we propose DP-Rec, a dynamic latent patching architecture for recommendation. DP-Rec shifts from item-level modeling to patch-level modeling by segmenting interaction sequences using contrastive entropy surprise to identify informative behavioral boundaries. A lightweight patch encoder compresses these temporally contextualized segments into a reduced set of dynamic latent behavior vectors, which are then processed by a larger latent transformer and decoded for next-item prediction. Extensive experiments show that, under constrained computational budgets, DP-Rec scales effectively to long sequences and achieves a superior efficiency-accuracy trade-off over both non-compressed and fixed-size compression baselines.
Dwipam Katariya, Thomas Caputo, Akshat Shreemali +6
Sequential recommendation models, while powerful, have limited flexibility in responding to immediate user requests, making it difficult to adapt their recommendations to the user's timely interests. Unfortunately, existing user request adaptation methods often incur high computational overhead due to either 1) retraining the entire backbone network or 2) leveraging the inference ability of large language models (a.k.a. prompt engineering), limiting their applicability in large-scale recommendation services. This paper presents Control via Request-Aware Masking for Editing Recommenders (CRAMER), a framework that takes users' natural-language requests to immediately change sequential recommendation models' behavior. Specifically, inspired by the model control theory, CRAMER treats user requests as control signals to modulate frozen backbone parameters through masking, achieving instant adaptation to diverse requests while avoiding costly retraining. Experiments on multiple large-scale benchmark datasets show that CRAMER outperforms four state-of-the-art request-aware baselines across multiple recommendation metrics while achieving minimal overhead. Moreover, the proposed framework exhibits enhanced controllability and cross-domain adaptability, establishing a new paradigm for request-aware sequential recommendation.
Zhiyuan Julian Su, Naihe Feng, Zhen Luther Qin +1
Gaoling School of Artificial Intelligence, Renmin University of China · Faculty of Computer Science, Dalhousie University
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