Linear item-item models have recently garnered significant attention in recommender systems due to their high efficiency and competitive performance compared to complex non-linear models. This high efficiency makes them effective and highly suitable f...
Linear item-item models have recently garnered significant attention in recommender systems due to their high efficiency and competitive performance compared to complex non-linear models. This high efficiency makes them effective and highly suitable for rapid adoption in real-world applications. However, these models face several limitations stemming from the nature of recommendation data. Key challenges include popularity bias, which leads to the over-recommendation of popular items, and neighborhood bias, which overemphasizes local relationships and fails to capture global patterns. Furthermore, their reliance on high sparsity creates a semantic gap, and their foundational design is often static, failing to capture the dynamic non-temporal nature of user preferences. This dissertation systematically addresses these challenges through a three-part investigation.
First, we address intrinsic issues with recommendation data. (1) To combat both popularity bias and neighborhood bias, we propose a structural normalization method called Data-Adaptive Normalization (DAN). DAN is a model-agnostic technique that dynamically adjusts item and user influence based on dataset-specific characteristics, such as distribution skewness and a novel weighted homophily metric, effectively mitigating both biases. (2) To tackle high sparsity and the resulting semantic gap, we propose semantic enrichment with LLMs in a model named L³AE. This is the first framework to integrate Large Language Models (LLMs) into linear autoencoders, using a two-phase, closed-form solution. It distills rich semantic knowledge from LLMs to enhance the collaborative signal, especially for sparse interactions.
Second, we extend the model beyond its static nature. (3) To address the non-temporal nature of classic linear models, we introduce a temporal-aware extension called the TemporAl LinEar item-item model (TALE). TALE is an efficient linear sequential recommendation model that successfully incorporates temporal context. It introduces three key components: single-target augmentation to preserve temporal correlations, time interval-aware weighting to adjust item significance based on actual timestamps, and trend-aware normalization to capture dynamic item popularity.
In conclusion, the methods presented in this dissertation demonstrate that linear item-item models, when enhanced with sophisticated normalization, rich semantic understanding from LLMs, and dynamic temporal awareness, can overcome their traditional weaknesses. These contributions establish advanced linear models as a highly efficient, accurate, and robust alternative to more complex neural architectures in modern recommender systems.