Machine learning for canteen food recommendation in NTU

• Main Author: Le, Tan Khang
• Other Authors: Hui Siu Cheung
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Computer science and engineering
• Online Access: https://hdl.handle.net/10356/147948
• Institution: Nanyang Technological University

Food not only plays an essential role in our daily lives but also critically influences and shapes our culture. Thus, food reviews and recommendations have become essential in any form of society. However, reviewing and recommending food is not easy for anyone as it is likely to be overwhelmed by disparate preferences and cultures. Nevertheless, there have been many breakthroughs in the field of machine learning, and particularly Natural Language Processing, with the recent development in the architecture and training mechanisms. By using multiple deep learning methods, this project addresses the problems of sentiment analysis, adjective-noun word pair extraction, and information retrieval of food reviews. In this project, we aim to develop a web-based food review system at Nanyang Technological University (NTU), which systematically recommends the most suitable food selection for users. Our goal is to incorporate BERT and LSTM deep learning models into the system for sentiment analysis of food reviews. Furthermore, we implement a Part-of-Speech (POS) algorithm to automatically identify and extract adjective-noun pairs from any text content based on the principles of POS tagging and dependency parsing. The search system, using Apache Solr as its server, is also a part of the website that allows users to search for food reviews. Our RankNet model in the search system will match users’ queries with food reviews by their textual, semantic, and food category similarities and then return a ranked list of food reviews as a result. We conduct the performance evaluations of our approaches in the system. First, the experimental results show that the BERT model achieves the state-of-the-art accuracy of 70.52% on the Yelp fine-grained sentiment classification task and outperforms the LSTM model by 3.76%. Second, for our POS algorithm, 95% of its adjective-noun pair extraction results are indicated as valid by a group of evaluators. Last but not least, in the food review search system, our proposed approach, using the RankNet model, significantly outperforms the classical text retrieval-based methods, namely BM25 and tf-idf, by between 15% and 33% on the MAP@{1, 3, 5} and MRR metrics.