Effective Product Recommender System using Hybrid FastTex, Attention and Probabilistic Matrix Factorization

Recommender system (RS) is a machine learning model to support consumer in online shopping. RS responsible to show essential product/item to customer. A variant of RS that adopted in many e-commerce businesses called Collaborative Filtering (CF). CS calculated connection between user and item using previous activity of user. RS framework that leverages past customer action, such as ratings, to provide recommendations. Unfortunately, we have collected a relatively small number of customer ratings that are below 4%. An implicit factor model is a type of matrix factorization-based CF that generates rating predictions. However, depending exclusively on matrix factorization would result in a recommendation that lacks accuracy. Product review records are utilized in various models to enhance the accuracy of rating predictions. Most individuals utilized the TF-IDF method, which is an LDA methodology, to analyze texts pertaining to product reviews. However, the standard models such as TF-IDF and LDA had limitations in terms of their ability to accurately understand the content of the document. This work employed probabilistic matrix factorization (PMF) to generate rating predictions by incorporating word sequential contextual representation through the use of LSTM (Long Short-Term Memory) and Attention. The contextual portrayal utilized Fasttext to convert word vector representation, each in their own way. The average values for Hybridization PMF, Attention, and Fasttext were 0.880691. Furthermore, our Hybridization PMF, Attention, and Fasttext achieved an average score of 0.7246844. Our model demonstrated a superior performance over the earlier studies.