Robust heterogeneous graph representation learning for enhanced recommender systems

نوع: Type: Thesis

مقطع: Segment: PHD

عنوان: Title: Robust heterogeneous graph representation learning for enhanced recommender systems

ارائه دهنده: Provider: Amin Nazari

اساتید راهنما: Supervisors: Dr. Muharram Mansoorizadeh

اساتید مشاور: Advisory Professors: Dr. Hassan Khotanlou, Dr. MirHossein Dezfoulian

اساتید ممتحن یا داور: Examining professors or referees: Dr. Behrouz Minaei, Dr. Mohammad Aliannejadi

زمان و تاریخ ارائه: Time and date of presentation: 2026

مکان ارائه: Place of presentation: اتاق سمینار گروه کامپیوتر

چکیده: Abstract: Heterogeneous information networks are widely used to model real-world data such as social networks, citation networks, and commercial platforms, as these data encompass diverse types of entities and rich semantic relationships. However, these networks often have extremely large dimensions, which can severely impact the efficiency of analytical methods and machine learning techniques. To address this issue, heterogeneous graph representation learning methods have been developed with the aim of providing a new representation in a lower-dimensional space. An optimal representation should simultaneously preserve the structural properties of the graph, the latent semantics embedded in the relationships between nodes, and the attributes of individual nodes, so that it can be utilized in applications such as similarity computation, classification, clustering, link prediction, graph visualization, and recommender systems. Despite the remarkable successes of graph neural networks in learning node representations, these methods still face challenges in heterogeneous, noisy, or dynamic environments due to limitations in semantic modeling and vulnerability to structural imperfections. To tackle these challenges, we first propose a noise-resistant multi-view representation learning framework called RoGAT. RoGAT integrates attention mechanisms at three levels: node, edge, and meta-path. Each view within this framework learns an embedding based on semantics or relation types; subsequently, contrastive learning is employed to pull together embeddings derived from different views of the same node while pushing apart embeddings of other nodes, thereby enhancing the richness of the learned representations. Our approach dynamically combines rich semantic dependencies from multiple perspectives through multi-level representations and improves them via contrastive learning and information integration. This design enables reliable representation learning even under noisy conditions. However, learning effective representations from heterogeneous graphs remains challenging due to severe data sparsity. To overcome these limitations, we introduce our second framework, NR-GAT, a multimodal, node-type-aware and relation-type-aware graph neural network. NR-GAT integrates node-level and relation-level attention mechanisms into a unified architecture, allowing the model to capture rich semantic dependencies across multiple levels of abstraction and from heterogeneous data sources. To support this framework, we construct a semantically enriched heterogeneous information network by integrating interaction signals and deep semantic representations extracted from textual sources. Topic modeling is employed to distill latent semantic topics from subtitle data, which are then added to the graph as additional node types. Furthermore, to demonstrate the impact of integrating heterogeneous sources, we introduce another model, MHG-Rec, which leverages a multi-task learning paradigm to jointly optimize node classification and recommendation objectives. This design acts as an implicit regularizer, encouraging the learning of shared and generalizable representations while enhancing model robustness under noisy and data-scarce conditions.

فایل: ّFile: Download فایل