This research propose a hybrid recommender system, leveraging the benefits of content-based filtering recommender system and collaborative filtering recommender system. The dataset in this research is acquired through web scrapping from TripAdvisor. Both structured numeric data and unstructured textural data are used in building the prediction models. In the data preparation phase, numeric data (user ratings) are converted into the format of user-item pair matrix, and the sparsity problem is resolved with the matrix factorization technique; textual data are converted to vectors by using word embedding and further reduced to five specific dimensions by using semantic analysis. Minmax and mean normalization techniques are applied to the dataset before modelling. Five different types of prediction models are constructed in this research: (1). Matrix factorization model to impute missing ratings; (2). Statistic based collaborative filtering computing; (3). K-nearest neighbour classifier; (4). Content-based filtering with Multiple Linear Regression; (5). Logistic regression with Keras; and (6). Binary classification model with Neural Network. The best prediction mode is loaded and applied with an English conversational chatbot. The chatbot is built on the Facebook platform and can be accessed by users through Facebook Messenger.
Volume 12 | 07-Special Issue
Pages: 1574-1593
DOI: 10.5373/JARDCS/V12SP7/20202261