User Feedback Analysis for Business Intelligence: Semantics, Sentiment and Model Robustness
Type of DegreePhD Dissertation
Computer Science and Software Engineering
MetadataShow full item record
Business intelligence (BI) is a set of enterprise decision support tools designed to help managers, analysts, and executives rapidly make wise decisions. The growth of the business intelligence applications is fueled and challenged by the large amounts of data of customers arising from internet and the adoption of the technology of Artificial Intelligence for sophisticated data analysis. The corporate world shows particularly strong interests in user-feedback analysis by leveraging the state-of-the-art AI-based natural language processing technology. Automated systems supporting the design of software play a vital role in the field of BI. For example, in the mobile app market, a multi-billion-dollar industry, and online consumer feedback can provide good insights into product strengths and weaknesses. Therefore, in the first part of this dissertation, we put forward an AI-driven framework that supports application design. We utilize deep unsupervised learning models to build a framework for harnessing potential customer feedback information. The first component of the framework applies Bidirectional Encoder Representations from the transformers (BERT)-based topic modelling approach to identify topics and key themes that emerge from user reviews of mobile applications belonging to the health and fitness genre. Sentiment analytics integrates the accompanying ratings to reveal the market acceptance of various aspects of product design. Asides from the semantic information, the emotional information extracted from the nature language data of customer is also valuable for marketing and branding activities. Therefore, in the second part of this dissertation study, we delve into the emotion analysis of public opinion in the scenario of telemedicine, in which we devise a novel emotional analytics framework. We investigate several emotion models and proposed a novel framework to solve the challenge of non-polarity emotion analysis. We use BERT to extend the traditional dictionary-based to detect the emotion of all words in the context even not included in the dictionary. We compare our new method against the other baseline methods in semantic analysis task and applied the best performance method on user reviews of telemedicine applications and reveal the social acceptance of the telemedicine. As the wide use of deep learning models in the business intelligence and its applications, the robustness of deep learning models becomes a challenging issue. When deep learning models are deployed in decision support, the reliability is a big concern especially when it comes to life-critical missions. As the last part of this dissertation research, we study the well-known problem of adversarial example in the realm of image data recognition. Evidence shows that small interpretation on input images before feeding fail the AI classifiers. We argue that the small-difference transformations commonly used are the blame and; therefore, we propose a new model-agnostic defense using a large-difference transformation. Specifically, we apply the novel primitive-based transformation that rebuilds input images by primitives of colorful triangles. In terms of the distortions required to completely break the defenses, our experiments on the ImageNet subset demonstrate that significantly large distortions (0.12) are needed to break the defense compared to other state-of-the-art model-agnostic defenses (0.05-0.06) under strong attacks. This finding indicates that large difference transformations tend to improve the adversarial robustness, thereby suggesting a promising new direction towards solving the challenge of adversarial robustness.