Buying Pattern Prediction

Client:Wing Yung Chan, The Hut Group <wingyungchan@gmail.com>

Online retailers face the interesting challenge of trying to understand a potentially large and unknown customer base. Once a customer has bought a product from a site, we'd like to know what they'll likely buy next and importantly, WHEN. This project is to build a Machine Learning system that can use the historic behaviour of hundreds of thousands of customers (anonymised) to predict how new customers to the site will behave.

The dataset will be taken from a collection of popular and diverse websites that belong to The Hut Group. Proposed sites include MyProtein, LookFantastic and IWantOneOfThose.

feedback

We do get quite a lot of suggestions for generic applications of machine learning to business (including two or three stock market prediction systems in a typical year). I don't have any objection to using machine learning techniques as a core of the project, but in their second year, students won't have the skills to develop a very sophisticated approach - they may do something like simply applying a standard regression toolkit. This doesn't really provide a great deal of challenge.

The students work in teams of six, so while one or two of them are doing this, we'd need to expand the project idea with functionality that the others might be working on. How about this as an idea …

There are many online shopping sites where users become frustrated - they can't find what they want, get bored, or simply fail to complete a purchase because of a usability problem. Retailers would be happy to help out such customers, if only they which ones were most likely to benefit. If you could predict which shoppers are about to abandon their shopping basket based on click stream analysis, then it would be possible for an intelligent agent (ideally assisted by a live human operator whose contributions to the conversation are seamlessly interleaved with robot responses about routine matters) to offer the customer real-time hints and assistance via a pop-up window.