Abstract:

The explosion of e-commerce platforms in recent years has brought about an enormous volume of product data. Recommender systems have become increasingly important in the e-commerce industry to help users navigate through this vast amount of data and to provide personalized recommendations based on user preferences. Traditionally, recommendation systems have relied on either image-based or text-based features to make recommendations. However, combining these two features could potentially lead to more accurate and effective recommendations. In this paper, we propose an approach that combines image and text-based features to provide more accurate and personalized product recommendations. We use ResNet-50 to extract image embeddings, and the Sentence- Transformers model with the BERT-base-NLI-mean-tokens architecture to generate text embeddings. Cosine similarity is then used to measure the similarity between the embeddings, which serves as the basis for product recommendations. The main contribution of this paper is to navigate the effectiveness of combining image and text-based features for product recommendations. Specifically, we evaluate the proposed approach on a large dataset of product images and descriptions to get recommendations. We also focus on computability to make our approach run on commodity-level hardware with a single GPU

Description:

BERT, E-commerce, Embeddings, Image-based Recommendation, Text-based Recommendation, Product Recommendation, Recommender Systems, ResNet-50, Sentence- Transformer, Cosine Similarity, Personalized Recommendations