Double Deep Features for Apparel Recommendation System

This study describes a recommendation system embedded in the double features extracted by convolutional neural networks (CNNs). Several probabilistic models, such as probabilistic matrix factorization (PMF)-based approaches, have been utilized for recommendation systems based on a CNN model. Each recommendation algorithm utilizes a single CNN model to extract precise features about documents and pictures, and the systems with CNN have contributed in improving the performance in rating prediction. Meanwhile, the systems for some items should consider at least two precise features simultaneously, and the extension to embed multiple CNN models is necessary. However, methods that integrate multiple CNN-based features into existing recommendation systems, such as PMF, are not available. Thus, this study proposes a novel probabilistic model that integrates double CNNs into PMF. For apparel goods, two trained CNNs from document and image shape features are combined, and the latent variables of users and items are optimized based on the vectorized features of CNNs and rating. Extensive experiments demonstrate that our model outperforms other recommendation models. Index Terms – recommender system, deep learning, image shape feature,convolutional neural network, probabilistic matrix factorization