The last decade has seen the rise of personalisation as a driving force for consumer markets. Spearheaded by companies such as Amazon.com, traditional outlets such as book-selling have been lent a new meaning by integrating search engines with personalised recommendations and this, in turn, has spawned a new generation of algorithmic developments of considerable theoretical as well as practical importance.

One of the most successful recommender system technologies is collaborative filtering, where products are recommended on the basis of the expressed opinions or observed behaviour of other customers, based on fundamental measures of similarity and association. The former identifies neighbourhoods or clusters of representative observations from which to derive a product recommendation usually with Bayesian conditional probabilities. However, the value of the massive datasets accumulated at an increasing rate at point-of-sale and even by radio-frequency tagging is massively higher than its use for associative rule mining. Beyond predictive inference, machine-learning algorithms serve to provide explanations of value chains and to explore for latent drivers that underpin observed consumer behaviour. Value chain optimisation has previously delivered Bayesian networks and been used in practice in product development and marketing, but without exploring the potential for the discovery of latent variables.

This project will develop and evaluate graphical models for collaborative filtering. Probabilistic recommender systems capable of hidden variables underpinning preference, choice and purchase behaviour will be developed within the Bayesian belief networks framework.

Project staff and support

Terence Etchells (Postdoctoral associate, Liverpool JMU)
Paulo Lisboa (Academic supervisor, Liverpool JMU)
Ben Dias (Industrial supervisor, Unilever)
Tim Boxer (Technology Translator, Industrial Mathematics KTN)

This project is being carried out at Liverpool JMU, in conjunction with Unilever. It is supported by an EPSRC industrial CASE award, made available through the Knowledge Transfer Network for Industrial Mathematics. Start date: October 2006; duration: 3.5 years.