A1612
Title: Topic Model for multiple supervised information based on non-linear functions
Authors: Kotono Waki - Doshisha university (Japan) [presenting]
Shintaro Yuki - Doshisha University (Japan)
Hiroshi Yadohisa - Doshisha University (Japan)
Abstract: In marketing, purchase history comprises the number of purchases made by each consumer for each product and is characterised by high dimensionality and extreme sparseness. In many cases, supervisory information, such as consumer attribute data and service evaluations, corresponds to purchase history. Supervised LDA (SLDA) is a method for learning consumer requirements from purchase history and interpreting them based on supervisory information. Supervisory information is crucial for understanding consumer requirements, and SLDA can accommodate a single piece of supervisory information. However, in cases involving multiple pieces of supervisory information, SLDA cannot capture their correlated structure. Additionally, SLDA assumes a linear relationship between consumers and topics during learning, which limits the expressive ability of the model. By contrast, an embedded topic model can estimate topics using nonlinear functions. The aim is to develop a nonlinear topic-learning method that considers the correlated structure of multiple pieces of supervisory information based on the abovementioned methods. This approach enables a more precise understanding of consumer requirements.
