Deep Hotel Recommender System Using Aspect-based Sentiment Analysis of Users’ Reviews

Abstract

Listen

Translate article

The value of search engines has increased with the digitalization of industries such as e-commerce and entertainment. There are numerous areas where recommendations are used and many different methods and technologies have been used to develop recommender systems. Deep learning-based methods are most effective as they improve the system’s overall performance by obtaining state-of-the-art predictive accuracy. However, collaborative filtering is the most famous technique used in the recommneder systems that rely on the user-item interaction. Besides, Aspect-Based Sentiment Analysis (ABSA) is a text analysis method that divides opinions into aspects and determines the sentiment associated with each. Unlike previous works, in this paper, we combine aspect-based sentiment analysis with multi-criteria collaborative filtering using deep learning. In this study, we use polarities obtained by aspect-based sentiment analysis from user reviews combined with the overall rating and sub-ratings in a deep neural network model with attention mechanism to predict user-item interactions. The model is composed of three parts: The first part includes Aspect-based Sentiment Analysis (ABSA) which is applied to the reviews, and polarities are extracted, then the final dataset is generated. In the second part, the user and item embeddings are concatenated and an attention layer is added after user and item embeddings. This layers provides the insights to user preferred words by the model by learning aspect weights. Then the obtained features are given as input to the Polarities and Sub-ratings Deep Neural Network (DNN) to predict the sub-ratings and polarities. Finally, in the third part, we give obtained criteria array as input to the Overall Score DNN to predict the overall score. Based on polarities the proposed approach leverages aspect-based user’s opinion. The obtained results indicate that proposed approach performs significantly better on the TripAdvisor dataset.