Dynamic Adaptation of Recommendation Models: Structures, Representations, and Data Utility

Abstract

Personalized recommender systems have become a cornerstone of online platforms and information services, aiming to capture user preference dynamics and deliver timely, tailored experiences. With the explosive growth of mobile interactions and online activities, user behavior exhibits strong temporal patterns and non-stationarity, including the continuous arrival of new users and items as well as the gradual drift and sudden shift of existing user interests. In such streaming environments, traditional static recommendation models often rely on batch training with fully sampled historical data. These models lack sensitivity to temporal evolution and struggle with challenges such as cold-start cases, interest shifts, and storage limitations. Meanwhile, advanced modeling approaches such as graph convolution and user/item embeddings have achieved remarkable success, yet their adaptability to dynamic environments remains limited. In particular, the problems of how to expand structures, adjust representation capacity, and exploit data utility under continuous updates remain pressing. To systematically solve these issues, this thesis takes dynamic adaptation as its central principle and conducts research around challenges at levels of structure, representation, and data utility.

Concretely, these challenges manifest as follows. First, at the structural level, graph convolutional models effectively capture high-order collaborative signals but risk becoming over-stationary in streaming contexts, overly dependent on long-term preferences while lagging behind rapid shifts. The intermittent activity of users and the continuous expansion of the item space further limit the scalability of fixed-size architectures. Second, at the representational level, most approaches adopt a uniform and static embedding dimension for all users and items, overlooking vast differences in interaction frequency and interest diversity. This issue is exacerbated in dynamic and streaming scenarios: highly active users demand evolving representation capacities, new items require efficient capacity allocation, while cold-start cases suffer from over-allocated fixed embedding sizes that are largely wasted and rarely utilized. Such static allocation not only causes significant memory redundancy but also prevents efficient resource distribution under limited budgets, undermining overall recommendation quality. Third, at the data utility level, pairwise ranking methods such as Bayesian Personalized Ranking (BPR) widely rely on uniform sampling or heuristic weighting, lacking principled schemes for evaluating sample contribution. As a result, informative samples are underutilized, the true value of training data is diluted, and performance degrades particularly in long-tail and cold-start situations.

To address the above challenges, this thesis proposes three contributions that systematically advance the dynamic adaptation of recommendation models across structure, representation, and data utility. At the structural level, we introduce Dynamically Expandable Graph Convolution (DEGC), which evolves graph convolutional filters through pruning, refinement, and expansion. A Kalman-inspired temporal attention mechanism enables initialization of user embeddings without historical replay, balancing efficiency with adaptability, while adaptive pruning prevents uncontrolled growth. At the representational level, we propose Dynamic Embedding Size Search (DESS), framing embedding dimension allocation as a non-stationary bandit problem. Built upon the LinUCB framework, DESS integrates user INterest Diversity (IND) and item PrOperty Diversity (POD) as feedback signals, along with a weighted forgetting mechanism, to achieve provable sublinear regret. An embedding-adaptive neural collaborative filtering framework further supports efficient online updates, achieving both precision and resource efficiency. At the data utility level, we present Interpretable Triplet Importance for Personalized Ranking (ITIPR), which defines (user, positive item, negative item) triplet contribution via triplet Shapley values. Efficient approximation is achieved through gradient-based truncated Monte Carlo with variance reduction, enabling stable estimation. ITIPR supports importance-aware resampling and reweighting, leading to improved ranking quality while offering interpretable insights into which user–item interactions matter most. Unlike DEGC and DESS, which primarily target streaming recommendation dynamics, ITIPR has general applicability to various recommendation scenarios, providing a transparent and fair mechanism for data utilization.

Extensive experiments on large-scale public datasets and real-world industrial benchmarks validate the proposed approaches. DEGC consistently outperforms strong graph convolution neural network baselines such as LightGCN, NGCF, and MGCCF across four datasets, mitigating over-stationarity while maintaining stable efficiency without requiring historical replay. DESS demonstrates superior performance in both Top-k recommendation and rating prediction across four benchmarks, achieving significant gains in accuracy while substantially reducing memory and training cost, thus enabling practical online alternation training. ITIPR achieves robust improvements across six datasets and two backbone families, including matrix factorization and graph neural networks, delivering consistent gains in Recall and NDCG while providing interpretable sample-level contributions that uncover the most valuable interactions. Collectively, this thesis advances the dynamic adaptation of recommendation models from structural, representational, and data utility perspectives, offering both theoretical foundations and empirical evidence for building the next generation of effective, efficient, scalable, and trustworthy recommender systems.

Date of Award	1 Dec 2025
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Chen MA (Supervisor)