IMPUTATION OF MISSING MODALITIES IN MULTIMODAL RECOMMENDATION SYSTEMS

Multimodal recommendation systems, leveraging heterogeneous information such as text, images, and audio to provide personalized and accurate recommendations, represent one of the most advanced frontiers in machine learning. However, one of the main obstacles in implementing these systems is the management of missing modalities, a frequent problem in real-world scenarios where data is often incomplete or absent. As a standardized solution for multimodal feature extraction in recommendation systems, the literature proposes a framework named Ducho, which however presents a significant characteristic that limits its applicability in practical contexts: the absence of dedicated mechanisms for handling missing modalities. This thesis presents the development of a new module for Ducho, designed to effectively handle missing modalities through the implementation of state-of-the-art imputation techniques; the module integrates traditional statistical imputation approaches, graph-based techniques and deep learning-based methods. The experimental evaluation, conducted on a multimodal dataset with missing modalities, demonstrates the effectiveness of the developed module in imputing missing modalities. The main contribution of this work consists in extending the capabilities of Ducho, making it applicable to real-world scenarios characterized by incomplete multi modal data, thus providing the scientific community with a standardized tool for the comparative evaluation of extraction and imputation techniques in the multimodal domain.