Two-Sided State-Space Models for Sequential Recommendation with Non-Random Multimodal Review Feedback Open Access

Abstract

Two-sided digital platforms are dynamic: user preferences evolve and item popularity shifts over time. User reviews play a central role in these dynamics. Yet most existing methods treat reviews as passive information for updating user representations. This leaves two important aspects underexplored. First, how are reviews generated? The generation is often not random, depending on the evolving latent states of users and items. Second, how do reviews propagate through item-side dynamics? Realized reviews can induce complex spillovers across related items and influence later decisions. To address these challenges, this thesis develops a two-sided state-space model with non-random feedback. It consists of: (1) a modality-missing-not-at-random review fusion module that leverages informative signals whether rating, text, or image modalities are present, and their expression patterns; (2) a weighted message passing algorithm that adjusts for the endogeneity in review presence and propagates across a local user-item-item graph that allows for cross-item review spillovers; and (3) a two-sided temporal state evolution module that models global temporal shocks as well as asymmetric carryover of negative versus positive review feedback. The framework is studied in the setting of event-conditioned sequential recommendation, where observed reviews are used as inputs for ranking instead of being predicted. Across six Amazon categories, the proposed method consistently outperforms strong sequential, multimodal, and debiasing baselines, improving Recall@20 by 14.8%-18.8% over BSARec, the strongest baseline.

Table of Contents

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . 1

2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2.1 Sequential Recommendation . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2.2 Multimodal Recommendation and Missing Modalities . . . . . . . . . . . . . 5

2.3 Missing Data, Debiasing, and Causal Perspectives . . . . . . . . . . . . . . . 6

2.4 Temporal Dynamics, User-Generated Content, and Personalization . . . . . . 7

3 Problem Setup and Notation . . . . . . . . . . . . . . . . . . . . . . . . . . 8

4 Data and Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . 12

4.1 Data Source and Category Coverage . . . . . . . . . . . . . . . . . . . . . . 12

4.2 Preprocessing and Temporal Discretization . . . . . . . . . . . . . . . . . . . 12

4.3 Dataset Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

4.4 Motivating Empirical Patterns . . . . . . . . . . . . . . . . . . . . . . . . . . 14

5 Method . . . . . . . . . . . . . . . . . . . . . . . . . . 16

5.1 Neural Instantiation of the State-Space Formulation . . . . . . . . . . . . . . 17

5.2 MNAR Multimodal Review Encoding . . . . . . . . . . . . . . . . . . . . . . 19

5.3 User-Side Latent-State Evolution . . . . . . . . . . . . . . . . . . . . . . . . 21

5.4 Item-Side Latent-State Evolution . . . . . . . . . . . . . . . . . . . . . . . . 23

5.5 MNAR-Aware Weighted Message Passing and Local Propagation . . . . . . . 25

5.6 Dynamic Item Scoring and Training . . . . . . . . . . . . . . . . . . . . . . . 26

6 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . . 29

6.1 Datasets and Evaluation Scope . . . . . . . . . . . . . . . . . . . . . . . . . 29

6.2 Multimodal Feature Extraction and TS-SSM Configuration . . . . . . . . . . 30

6.3 Baselines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

6.4 Evaluation Protocol . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

7 Results . . . . . . . . . . . . . . . . . . . . . . . . . . 33

7.1 Main Results on the Focal Categories . . . . . . . . . . . . . . . . . . . . . . 33

7.2 Cross-Dataset Generalization . . . . . . . . . . . . . . . . . . . . . . . . . . 34

7.3 Statistical Robustness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

7.4 Information Boundary Controls . . . . . . . . . . . . . . . . . . . . . . . . . 36

8 Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 39

8.1 Component Contribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

8.2 Limitations and Future Directions . . . . . . . . . . . . . . . . . . . . . . . . 41

9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . 42

About this Honors Thesis

Rights statement

• Permission granted by the author to include this thesis or dissertation in this repository. All rights reserved by the author. Please contact the author for information regarding the reproduction and use of this thesis or dissertation.

School	Emory College
Department	Data and Decision Sciences
Degree	B.S.
Submission	Honors Thesis
Language	English
Research Field	Computer Science Statistics Applied Mathematics
Keyword	Multimodal Learning Missing Not At Random (MNAR) Message Passing State-Space Models Recommender Systems Sequential Recommendation
Committee Chair / Thesis Advisor	Ruoxuan Xiong, Emory University
Committee Members	Manuela Manetta, Emory University Musa Ayar, Emory University

Last modified

Primary PDF

Thumbnail	Title	Date Uploaded	Actions
	Two-Sided State-Space Models for Sequential Recommendation with Non-Random Multimodal Review Feedback ()	2026-04-06 22:05:18 -0400	Press to Select an action Download

Supplemental Files

Thumbnail	Title	Date Uploaded	Actions