Data Analysis and Processing for Recommendation System

Abstract

In this graduation thesis, we focused on exploring novel methods and approaches in data anal- ysis and processing for recommendation systems. We aimed to address the challenges posed by diverse and heterogeneous data to contribute to the advancement of recommendation sys- tems. By effectively analyzing and processing such data, we can unlock the true potential of recommendation systems, enabling users to make better decisions and discover new experiences. To build our recommendation system, we established models for points of interest (POIs) and users. Our solution incorporated three key factors: sentiment analysis, user preferences, and ratings, culminating in the integration of the lightGCN model. Preprocessing and filtering of the data were performed to ensure data quality, followed by modelling the POIs based on their unique characteristics. The sentiment analysis factor played a crucial role in analyzing user reviews and predicting ratings. By employing sentiment analysis techniques, we accurately represented the user’s opinion by aligning the sentiment expressed in textual reviews with the user’s rating. The user preference factor enabled us to recommend the most suitable POIs based on individual preferences and interests. The rating factor, which examined the ratings users gave to visited POIs, facilitated tracking and updating user preferences. This allowed for dynamic adjustment and refinement of the user preference profile, ensuring recommenda- tions aligned with their evolving interests. The culmination of these factors, along with the preprocessing, filtering, and modelling of the POIs, led to the integration of the lightGCN model. By combining similarity scores derived from the user preference profile, sentiment anal- ysis score, and ratings score, the lightGCN model predicted the most suitable POIs for each user, enhancing the recommendation system’s accuracy. During the experimentation phase, we utilized Yelp datasets in a Jupyter environment to preprocess, filter, and model the POIs, incorporating sentiment analysis of reviews. The rec- ommendation system, developed in the same environment, utilized the combined results of the three factors and the lightGCN model to provide improved POI recommendations for users