LSA: A Long-Short-term Aspect Interest Transformer for Aspect-Based Recommendation
Le Liu, Junrui Liu, Yunhan Gao, Ziheng Wang, Tong Li · 2026
Read LSA: A Long-Short-term Aspect Interest Transformer for Aspect-Based Recommendation by Le Liu, Junrui Liu, Yunhan Gao, Ziheng Wang, Tong Li (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-21243.txt.
Introduction
Liu and co-authors present LSA, a transformer that models long and short term aspect interests for recommendation. Aspect based methods mine review terms, such as price, to learn fine-grained user tastes. User interests shift over time, though, and static graph methods miss brief new focuses. LSA models short term interest from recent aspect interactions and long term interest from global patterns. On four real data sets, it cuts mean squared error by about two and a half percent over the best baseline.
Catalog record
| CREATOR | Le Liu, Junrui Liu, Yunhan Gao, Ziheng Wang, Tong Li |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 24548 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.21243
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
LSA: A Long-Short-term Aspect Interest Transformer for Aspect-Based Recommendation Le Liu, Junrui Liu, Yunhan Gao, Ziheng Wang, Tong Li Abstract Aspect-based recommendation methods extract aspect terms from reviews, such as price, to model fine-grained user preferences on items, making them a critical approach in personalized recommender systems. Existing methods utilize graphs to represent the relationships among users, items, and aspect terms, modeling user preferences based on graph neural networks. However, they overlook the dynamic nature of user interests - users may temporarily focus on aspects they previously paid little attention to - making it difficult to assign accurate weights to aspect terms for each user-item interaction. In this paper, we propose a long-short-term aspect interest Transformer (LSA) for aspect-based recommendation, which effectively captures the dynamic nature of user preferences by integrating both long-term and short-term aspect interests. Specifically, the short-term interests model the temporal changes in the importance of recently interacted aspect terms, while the long-term interests consider global behavioral patterns, including aspects that users have not interacted with recently. Finally, LSA combines long- and short-term interests to evaluate the importance of aspects within the union of user and item aspect neighbors, therefore accurately assigns aspect weights for each user-item interaction. Experiments conducted on four real-world datasets demonstrate that LSA improves MSE by 2.55 % 2.55\% on average over the best baseline. Keywords Recommender Systems ⋅ \cdot Aspect-based Recommendation ⋅ \cdot Long-term Interest Modeling ⋅ \cdot Short-term Interest Modeling 1 Introduction Aspect-based recommendation methods effectively learn the fine-grained user preferences and have attracted significant attention [ 4 , 15 , 19 ] . Existing approaches primarily employ two paradigms: CNN-based and GNN-based methods. CNN-based methods [ 5 , 8 ] leverage CNN to extract aspect features from reviews. In contrast, recent graph-based methods [ 11 , 13 ] explicitly model user-item-aspect relations using graph structures to better capture contextual and sentiment-aware representations. However, existing methods only focus on modeling static user preferences from historical reviews and overlook the dynamic nature of user interests. Users may temporarily focus on aspects they previously paid little attention to, which makes themn difficult to assign accurate weights to aspect terms for each user-item interaction [ 1 , 11 , 14 ] . For example, a user expresses consistent interest in aspect terms, such as “sound quality” and “build quality” in history reviews, showing a long-term preference for high-performance audio equipment. The user then often mentioned “portability” in recent product reviews, suggesting a short-term interest in “portability” due to a possible upcoming trip. Long-term interests represent the stable interests of users, while short-term interests are the preferences of users recently. In this paper, we propose a Long-Short-term Aspect interest Transformer(LSA) for aspect-based recommendation, which effectively captures the dynamic nature of user preferences by integrating both long-term and short-term aspect interests. Specifically, to capture users’ transient needs, we propose a time-aware Transformer that models temporal changes in the importance of recently interacted aspect terms. Furthermore, LSA models the long-term interests by considering global behavioral patterns, including aspects that users have not interacted with recently. Those two interests are integrated through a gated fusion mechanism. To generate a personalized prediction, we introduce an interest-aware aspect term aggregation method, which uses user preferences to justify the importance of aspect terms, over the union of user- and item-specific aspect terms. The resulting aspect representation is concatenated with user and item embeddings and passed into a factorization machine for final scoring. The main contributions of this paper are summarized as follows:(1) We propose a Long-Short-term Aspect Interest Transformer(LSA) for aspect-based recommendation, which models the evolution of user interests through considering long- and short-term of interests. (2) We design a dual aspect selection method to capture users’ long-term and short-term interests effectively at the aspect granularity. (3) We introduce an interest-aware aspect term aggregation mechanism to align user preferences with item attributes for personalized prediction. (4) Extensive experiments conducted on four real-world datasets demonstrate that LSA improves MSE by an average of 2.55\% over the best baseline. 2 Related Work In this section, we review previous work relevant to our work, including Aspect-Based Recommendation and Modeling Long- and Short-Term User Interests. 2.1 Aspect-Based Recommendation Aspect-based recommender systems leverage fine-grained information from user reviews (i.e. aspect–sentiment pairs) to provide more personalized recommendations. Early works start to explore the sentiment of each aspects in user review instead of serving as a single source to explore the user’s preference on different aspects [ 15 , 19 ] . Building on this idea, researchers have proposed models to extract aspect terms from reviews and incorporate them into recommendation algorithms. For example, ANR [ 4 ] applies attention-based neural networks to identify aspect-level semantics and employs co-attention to capture asymmetric importance of aspects between users and items. Subsequent works integrate sentiment reasoning into aspect modeling. ARPM [ 6 ] incorporates social signals for preference modeling, while CARP [ 7 ] introduces logic-unit based capsule reasoning for fine-grained aspect-sentiment understanding. SENGR [ 17 ] integrates sentiment-aware review attention and graph-based user-item-user modeling to enhance rating prediction. Most recently, APH [ 11 ] models the aspect performance of items using a sentiment-aware hypergraph neural network, introducing a novel aspect hypergraph to aggregate conflicting user sentiments for fine-grained recommendation. 2.2 Modeling Long and Short-Term User Interests Users’ preferences are not static. Traditional models mainly consider long and short-term user interests at the item level. PLASTIC [ 21 ] proposes an adversarial framework that fuses long - term matrix - factorization profiles with short - term RNN session dynamics. Yu et al. [ 20 ] model short - term intent with an LSTM equipped with time - and content - aware controllers, then adaptively fuse it—via attention—with an attentive long - term preference vector. Most recently, LS-TGNN [ 14 ] leverages a temporal graph neural network and self-supervised contrastive learning to disentangle and model users’ evolving long- and short-term interests, significantly enhancing session-based recommendation performance. 3 Methodology To model the dynamic nature of user interests at the aspect level, this paper proposes LSA for aspect-based recommendation, with its architecture illustrated in Figure 1 . Specifically, LSA (1) extracts aspect terms from reviews and constructs a user–item–aspect graph ; (2) captures long - term preferences by a graph Transformer; (3) captures short - term preferences by forming time - ordered aspect sequences within temporal window; and (4) fuses the two views through a long–short attention mechanism to aggregate aspect terms. Finally, we describe the final prediction layer. Figure 1: Architecture of LSA. 3.1 Aspect Graph We initially extract aspect terms from reviews to construct the aspect graph. Following existing methods [ 9 , 11 ] , we choose a rule-based unsupervised method, which considers three dependency relations to extract aspect terms: amod , dobj and nsubj+acomp . After extracting aspect-sentiment from each review, we construct user-item-aspect graph, where nodes represent users, items, and aspect terms. Edges are created between nodes based on review interactions, and edge weights reflect either aspect frequency or user-item rating. In addition, the interaction timestamps are also recorded to support time-aware modeling. 3.2 Long- and Short-Term Interest Modeling Users’ interests are shaped by both stable long-term behavior and transient short-term needs. Thus, LSA models short-term interests using