A Multimodal Deep Learning Framework for Personalized Event Recommendation with Hyperparameter Optimization via Genetic Algorithm

Abstract

Personalized event recommendation poses a fusion challenge: textual descriptions, social graph connectivity, and user interaction history each carry information that no single-modality model can fully exploit. This paper presents a multimodal deep learning framework that combines DistilBERT text embeddings, Heterogeneous Graph Transformer (HGT) graph embeddings, and interaction features through a Gated Multimodal Unit (GMU) trained with Bayesian Personalized Ranking (BPR) loss. A Genetic Algorithm optimizes six hyperparameters including learning rate, batch size, embedding dimensions, and graph encoder depth. An order-preserving, leakage-free strategy is proposed for handling out-of-vocabulary tokens: rather than mapping unknown tokens to a shared index, they are removed from the input sequence while surviving tokens retain their original positional order, preventing representational collapse at small vocabulary sizes. Experiments are conducted on a controlled synthetic dataset comprising 10,000 users, 3,000 events, and 80,000 interactions, with random and popularity baselines included as reference points. Results confirm correct gradient flow and optimizer behavior; the framework is positioned as an architectural proof of concept intended for subsequent validation on datasets with genuine preference structure.