Abstract
Topic modelling and trend analysis of information are increasingly important in today’s world, especially with the information overload as a result of the increased accessibility of the Internet. As such, there is an urgent need to make good sense of the information – without which, may land humans in a state whereby it becomes counter-intuitively more difficult to find our desired resources online. Topic modelling is used to discover topics within datasets. These datasets can come in all forms - texts, images, or sounds etcetera. Over the years, topic modelling techniques have been continually improved on, evolving from more primitive, to more sophisticated methods, such Corex and FastText. Coupled with topic modelling comes a temporal factor, whereby popularity of topics can be analysed over time. This helps in better understanding the potential popularity of topics, regardless of their popularity level at any point in time. This paper proposes novel topic modelling techniques – LDA2X, LDA2Char, Doc2X, LDA2XPand, and LDA2Word, which combines traditional topic modelling techniques with recent advances in word embeddings. These new topic modelling techniques are evolved versions of older topic modelling techniques, such as LDA, Corex, FastText, Word2Vec, and Doc2Vec. The former newer topic modelling techniques are crafted through ‘hybridisation’ of the latter older topic modelling techniques, through strategic sequential combination of these older topic modelling techniques. After which, temporal factors for trend analysis are integrated with newer topic modelling techniques to dissect how popularity trends of each topic can be understood and compared to that of other topics, over time. These trend analysis results, combined with evaluations based on topic coherence scores and topic umbrella assigned to each group of keywords describing each topic, will be able to give conclusive results on which will be the best topic modelling algorithm to use depending on topic modelling purposes and circumstances of applications. These topic modelling and trend analysis techniques are crafted to inspire actual applications in the provision of business services. Some beneficiaries of these proposed topic modelling and trend analysis algorithms include researchers, entrepreneurs, and investors, through accelerating the identification of promising business ideas and markets for their joint creation of successful businesses. These business services can be part of business models that aim to provide contextual understanding of datasets, regardless of dataset size. Such a business model would be of high interest to audiences such as data-centric companies and data scientists as well, for the need to understand consumer actions and behaviours has been increasing at an exponential rate in the recent years. By implementing state-of-the-art techniques proposed in this paper, a significant number of parties stands to benefit - through possessing newfound capabilities of discovering of the quickest path to executing the right decisions, and through benefitting from accelerated knowledge gain and transfer.