A probabilistic model for item-based recommender systems

A probabilistic calculus for the calculation of conditionals based on item rather than basket counts is proposed, which has the consequence that items bough together as part of small baskets are more predictive of each other than if they co-occur in large baskets.

Abstract

Recommender systems estimate the conditional probability P(χj|χi) of item χj being bought, given that a customer has already purchased item χi. While there are different ways of approximating this conditional probability, the expression is generally taken to refer to the frequency of co-occurrence of items in the same basket, or other user-specific item lists, rather than being seen as the co-occurrence of χj with χi as a proportion of all other items bought alongside χi. This paper proposes a probabilistic calculus for the calculation of conditionals based on item rather than basket counts. The proposed method has the consequence that items bough together as part of small baskets are more predictive of each other than if they co-occur in large baskets. Empirical results suggests that this may result in better take-up of personalised recommendations.