From Agnostic to Specific: Latent Preference Diffusion for Multi-Behavior Sequential Recommendation

Ruochen Yang; Xiaodong Li; Jiawei Sheng; Jiangxia Cao; Xinkui Lin; Shen Wang; Shuang Yang; Zhaojie Liu

cs.IRFeb 26, 2026

From Agnostic to Specific: Latent Preference Diffusion for Multi-Behavior Sequential Recommendation

Authors: Ruochen Yang, Xiaodong Li, Jiawei Sheng, Jiangxia Cao, Xinkui Lin, Shen Wang, Shuang Yang, Zhaojie Liu,

Organizations: Institute of Information Engineering, Chinese Academy of Sciences · School of Cyber Security, UCAS · Kuaishou Technology

Abstract

Multi-behavior sequential recommendation (MBSR) aims to learn the dynamic and heterogeneous interactions of users' multi-behavior sequences, so as to capture user preferences under target behavior for the next interacted item prediction. Unlike previous methods that adopt unidirectional modeling by mapping auxiliary behaviors to target behavior, recent concerns are shifting from behavior-fixed to behavior-specific recommendation. However, these methods still ignore the user's latent preference that underlying decision-making, leading to suboptimal solutions. Meanwhile, due to the asymmetric deterministic between items and behaviors, discriminative paradigm based on preference scoring is unsuitable to capture the uncertainty from low-entropy behaviors to high-entropy items, failing to provide efficient and diverse recommendation. To address these challenges, we propose FatsMB, a framework based diffusion model that guides preference generation From Behavior-Agnostic To Behavior-Specific in latent spaces, enabling diverse and accurate Multi-Behavior Sequential Recommendation. Specifically, we design a Multi-Behavior AutoEncoder (MBAE) to construct a unified user latent preference space, facilitating interaction and collaboration across Behaviors, within Behavior-aware RoPE (BaRoPE) employed for multiple information fusion. Subsequently, we conduct target behavior-specific preference transfer in the latent space, enriching with informative priors. A Multi-Condition Guided Layer Normalization (MCGLN) is introduced for the denoising. Extensive experiments on real-world datasets demonstrate the effectiveness of our model.

Figures & tables

Figure 1. Motivation of our work. Discriminative paradigm is constrained by candidate set and interaction history, while generative paradigm incorporating user’s latent preference achieves accuracy and diversity.

Dataset	$\text{H}(I)$	$\text{H}(B)$	$\text{H}(B\|I)$	$\text{H}(I\|B)$	$\text{I}(I,B)$
Yelp	9.6591	1.2540	1.1569	9.5620	0.0971
Retail	10.8756	0.7227	0.6702	10.8231	0.0525
IJCAI	12.1743	0.5691	0.5041	12.1093	0.0651

Table 1. The entropy

\text{H}(\cdot)

, conditional entropy

\text{H}(\cdot|\cdot)

and mutual information

\text{I}(\cdot,\cdot)

between items I and behaviors B of the three datasets. Details refer to Appendix. B .

Figure 2. The overall framework of our model. Stage 1 trains MBAE to construct the latent space for unifying preferences. Stage 2 achieves the transfer from behavior-agnostic to behavior-specific in the latent space based on DM.

Figure 4. The architecture of MCGLN. The behavior-agnostic preference $z_{\varnothing}$ serves as the transfer subject. The target behavior $b_{t}$ guides the direction as hard router for MoE. Condition representation $e_{b}$ and $e_{t}$ generate adaptive weights.

Dataset	User	Item	Interactions	Behavior Type
Yelp	19,800	22,734	$1.4\times 10^{6}$	Tip, Dislike, Neutral, Like
Retail	147,894	99,037	$7.6\times 10^{6}$	Click, Favorite, Cart, Purchase
IJCAI	423,423	874,328	$3.6\times 10^{7}$	Click, Favorite, Cart, Purchase

Table 2. The statistics of datasets

Models	Yelp				Retail				IJCAI
Models	R@10	N@10	R@20	N@20	R@10	N@10	R@20	N@20	R@10	N@10	R@20	N@20
BERT4Rec	0.0501	0.0245	0.0828	0.0331	0.1414	0.0811	0.1853	0.0921	0.2386	0.1473	0.3009	0.1632
DiffuRec	0.0526	0.0280	0.0840	0.0362	0.1476	0.0860	0.1914	0.0971	0.2811	0.1702	0.3405	0.1816
MBGCN	0.0497	0.0229	0.0796	0.0323	0.1398	0.0801	0.1832	0.0910	0.2274	0.1401	0.2937	0.1590
MB-GMN	0.0517	0.0257	0.0859	0.0343	0.1434	0.0829	0.1859	0.0914	0.2408	0.1488	0.3072	0.1654
MB-STR	0.0545	0.0275	0.0938	0.0372	0.1532	0.0887	0.1990	0.1002	0.2904	0.1999	0.3430	0.2132

Table 3. The overall performance comparison of different models on three datasets. The best results are boldfaced and the second-best results are underlined , with the relative improvements denoted as Improve

\uparrow

. All baseline results are derived from reproduced experiments referencing the original papaers and released code.

Variants		Yelp		Retail		IJCAI
Variants		R@20	N@20	R@20	N@20	R@20	N@20
MBAE	w APE	0.0936	0.0381	0.2183	0.1103	0.3701	0.2105
	w RoPE	0.0981	0.0430	0.2252	0.1170	0.3712	0.2096
	w BaRoPE	0.1003	0.0438	0.2280	0.1189	0.3880	0.2227
FatsMB	w MLP	0.0980	0.0432	0.2279	0.1197	0.3798	0.2211
	w AdaLN	0.1019	0.0447	0.2286	0.1195	0.3857	0.2253

Table 4. The performance comparison of ablation for different modules with different components. Comparison can be divided into inter-module and intra-module. The best results of diverse module variants are boldfaced.

Panel of Figure 5 — Figure 5. The comparison of attention matrics of APE and BaRoPE on user_1 in Retail dataset. The sequential arrangement of Item_Behavior means the user interacted with item through behavior .

Panel of Figure 6 — Figure 6. The preference distribution and similarity experiment in Retail dataset.

Panel of Figure 7 — Figure 7. The comparison of FatsMB and MB-STR in the few-shot experiment in Yelp dataset.

Panel of Figure 8 — Figure 8. The performance comparison of different hyperparameters on Yelp and Retail datasets. Y-R and R-R indicates Recall on Yelp and Retail datasets. Y-N and R-N indicate NDGC on Yelp and Retail datasets. Among these, (a)(b) are the results on MBAE and (c)(d)(e) are the results on FatsMB.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 9 — Figure 9. Comparison of computational time.

Panel of Figure 10 — Figure 10. Comparison of computational space.

Recommender Systems

Recommender systems predict and rank items for people using interaction histories, content, context, and feedback.

See all 508 papers

MCLMR: A Model-Agnostic Causal Learning Framework for Multi-Behavior Recommendation
Mar 26, 2026Ranxu Zhang, Junjie Meng, Ying Sun +5Recommender Systems
School of Artificial Intelligence and Data Science, University of Science and Technology of China, Hefei, China · Artificial Intelligence Thrust, The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China · RMIT University, Melbourne, Australia +2
Modeling Behavioral Intensity and Transitions for Generative Recommendation
Apr 27, 2026Wenxuan Yang, Xiaoyang Xu, Hanyu Zhang +3Sequential Recommendation Recommender Systems
Ant Group · Fudan University
Beyond Observed Auxiliary Relations: Environment-Conditioned Modeling for Multi-Behavior Recommendation
Aug 24, 2026Seunghan Lee, Hyunsik Yoo, Jian Kang +2Recommender Systems User Preference Modeling
Korea University Seoul, Republic of Korea · University of Illinois Urbana-Champaign Champaign, IL, USA · MBZUAI Abu Dhabi, United Arab Emirates

Mar 26, 2026J/K move · Enter open · S save

MCLMR: A Model-Agnostic Causal Learning Framework for Multi-Behavior Recommendation

Ranxu Zhang, Junjie Meng, Ying Sun, Ziqi Xu+4

School of Artificial Intelligence and Data Science, University of Science and Technology of China, Hefei, China · Artificial Intelligence Thrust, The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China · RMIT University, Melbourne, Australia+2

Recommender Systems

Open paper AskPDF arXiv

Multi-Behavior Recommendation (MBR) leverages multiple user interaction types (e.g., views, clicks, purchases) to enrich preference modeling and alleviate data sparsity issues in traditional single-behavior approaches. However, existing MBR methods face fundamental challenges: they lack principled frameworks to model complex confounding effects from user behavioral habits and item multi-behavior distributions, struggle with effective aggregation of heterogeneous auxiliary behaviors, and fail to align behavioral representations across semantic gaps while accounting for bias distortions. To address these limitations, we propose MCLMR, a novel model-agnostic causal learning framework that can be seamlessly integrated into various MBR architectures. MCLMR first constructs a causal graph to model confounding effects and performs interventions for unbiased preference estimation. Under this causal framework, it employs an Adaptive Aggregation module based on Mixture-of-Experts to dynamically fuse auxiliary behavior information and a Bias-aware Contrastive Learning module to align cross-behavior representations in a bias-aware manner. Extensive experiments on three real-world datasets demonstrate that MCLMR achieves significant performance improvements across various baseline models, validating its effectiveness and generality. All data and code will be made publicly available. For anonymous review, our code is available at the following the link: https://github.com/gitrxh/MCLMR.

Abstract

Recommender Systems

Explore similar work

MCLMR: A Model-Agnostic Causal Learning Framework for Multi-Behavior Recommendation

Modeling Behavioral Intensity and Transitions for Generative Recommendation

Beyond Observed Auxiliary Relations: Environment-Conditioned Modeling for Multi-Behavior Recommendation