Advanced Recommendation Systems for Overcoming Cold Start and Data Sparsity Challenges
- Publication Type:
- Thesis
- Issue Date:
- 2026
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Recommender systems alleviate information overload by delivering personalised content, yet their effectiveness in real platforms is frequently constrained by imperfect signals, including sparse user–item interactions, item cold-start, and non-stationary user behaviour. Moreover, as modern recommender systems increasingly rely on item-side content, missing or degraded modalities further distort representation learning and fusion. This thesis studies robust recommendations under these practical imperfections through three interconnected contributions that progress from temporal modelling to structural representation enhancement, and finally to missing-modality robustness.
Non-stationary user behaviour is examined through the lens of sequential recommendation with long interaction histories. It introduces a state-space-based sequential architecture augmented with a wavelet frequency rescaling mechanism, enabling frequency-aware adaptation to dynamic user behaviours while maintaining efficient modelling of long-range dependencies.
Interaction sparsity is investigated from an item representation perspective in graph-based recommendation. It proposes a graph learning framework that injects global relational priors via fuzzy clustering and strengthens item representations beyond local neighbourhood aggregation, improving robustness when collaborative signals are limited or unevenly distributed.
Item cold-start is further studied in multimodal recommendation, where item content modalities may be incomplete, inducing a completeness shift between complete and incomplete items and entangling modality-invariant and modality-specific information. It proposes a modality-adversarial generative enhancement framework that reconstructs missing modalities from observed ones, aligns complete–incomplete distributions through domain-adversarial learning, and disentangles invariant versus specific signals via mutual-information regularisation to support more reliable multimodal fusion.
Taken together, these three studies demonstrate that robust recommendations in realistic settings benefit from jointly addressing non-stationary temporal dynamics in user behaviour, globally informed structural modelling under interaction sparsity, and distribution-aware generation and fusion under missing modalities. By connecting frequency-aware sequential modelling, fuzzy-structured graph representation learning, and completeness-shift-aware multimodal enhancement, this thesis provides a coherent framework for building recommender systems that remain dependable when data are sparse, behaviours evolve, and content is incomplete.
Please use this identifier to cite or link to this item: