A Hybrid Approach for Topic Discovery and Recommendations based on Topic Modeling and Deep Learning

Abstract

The elevated growth rate of internet users and boom in the applications development, mainly e-business, has familiarized users to write their comments and reviews about the product they received. These reviews help customers in decision making. In this thesis, we aim to propose a recommender system that is primarily based on user comments or reviews. We aim to utilize the user-generated data to perform topic discovery and recommendations based on the topics discovered. We take benefit of the language modeling practices and attempt to join them with neural networks (NN) to identify some latent discussion patterns in the feedbacks of users. Our goal is to design a probabilistic language modeling based neural network architecture, where RNN's special type Long-Short Term Memory (LSTM) model is used to predict discussion trends or topics. Probabilistic topic models utilize word co-occurrences across documents to identify topically related words. Due to their complementary nature, these models define different notions of word similarity, which, when combined, can produce better topical representations. We have trained our Latent Dirichlet Allocation (LDA) model with word embeddings to improve the topic quality. The hybrid of LDA-LSTM takes multiple inputs including words and topic embeddings, the temporal details, and probability distributions from LDA. To capture the long contextual dependencies and limited vocabulary challenge, LSTM model takes both direct word embeddings and temporal topic distributions. Latent topic distributions are used to feed LSTM layers that are learnt based on the LDA model which is pre trained. This proposed LDA-LSTM model, unlike previous studies, is capable of capturing both long range contextual and temporal dependencies. To enhance the recommendations results, we us objective function that takes into account the identified topic categories and user preferences. We have formulated an equation model to access the credibility or authenticity of a recommendation. The proposed hybrid approach is experimented on crowdfunding campaigns and is used for reliable project recommendations. We collected data for scam and non-scam crowdfunding projects and applied our proposed approach to suggest secure and optimized recommendations to investors by using their comments. For the proposed approach following metrics are considered for the performance analysis and evaluation of the proposed approach: i) prediction accuracy, ii) an optimal number of identified topics, and iii) the number of epochs. We compared our results with NN and NN-LDA based on these performance metrics. The strengths of both integrated models offer that the proposed model can play a significant part in an improved and effective understanding of user generated data such as crowdfunding comments.