| Title | : | A Multi-view Approach to Clickstream Mining |
| Speaker | : | Chandra Mohan T N (IITM) |
| Details | : | Tue, 2 Aug, 2016 1:00 PM @ BSB 361 |
| Abstract: | : | E-commerce has seen tremendous growth over the past few years, so much so that only those companies which analyze browsing behaviour of users, can hope to survive the stiff competition in the market. In order to model complex behaviours of a wide range of users, one must understand their online activities which are embodied in their clickstream data. Hence, it is now almost a necessity to study clickstreams. In this paper, we aim to develop a Multi-view learning model that uses clickstreams to understand the purchasing behaviour of users to predict whether a user would make a purchase or not. Our work is based on the belief that using all of the features mined from clickstreams at one go for building a prediction model may not capture all the purchase behavioural dynamics. A set of users might follow the same path almost all the time when purchasing a product. Similarly users might spend similar durations on pages. Furthermore, a lot of noise is generated in the data when both time and path are factored in. To avoid this, the feature space is partitioned into multiple views and separate models known as experts are trained on each of those views. The experts are then carefully combined using Mixture of Experts model to make the final prediction. Experimental results show that the feature partitioning works and that it indeed works better than the natural way of using all the features at one go. Moreover, our mixture model using LSTMs involves less of feature engineering and more of natural modeling of the click events. |
