Extracting visual features for personalized recommendation using autoencoder

In recent years, product images have garnered increasing interest within recommendation systems, as the visual appearance of products strongly influences users’ decision. The primary goal of personalized recommendation systems is to provide suggestions that reflect with the individual preferences of each user. Recently, deep learning models have demonstrated remarkable performance, revealing a strong potential in analyzing visual features. This paper present an innovative approach that utilizes item images to construct a personalized model for users. The approach involves employing Autoencoder for feature extraction of visual features in item images. Subsequently, these latent features are correlated with user preferences to establish the personalized model, which is then integrated into a collaborative filtering (CF) algorithm for recommendation purposes. The effectiveness of the proposed approach is assessed through experiments on two substantial real datasets, encompassing fashion items from Amazon.com and movies from the MovieLens dataset. A comparative analysis with alternative collaborative filtering-based methods is conducted to validate our results.