Next generation improvements in recommendation systems

A method to create personalized product recommendations in an electronic format for a specific shopper from a group of shoppers and a target product from a group of products, where each product has one or more attributes from a group of attributes, and where the method provides personalized product recommendations in which different product recommendations are provided for different shoppers for the same target product, comprising the steps of:

• a. identifying the specific shopper's preference value for an attribute, where the preference value is calculated using the specific shopper's behavior with other products with the same attribute, and calculating preference values for several other shoppers within the group of shoppers and the attributes, resulting in a different preference value for each shopper and each attribute,
• b. caching recommendation data for each shopper, the recommendation data comprising a subset of the group of attributes for each shopper, wherein the subset contains the N attributes with the largest preference values for each shopper,
• c. creating product recommendations based on correlation values between product pairs, where each correlation value represents a likelihood of purchase, and calculating the likelihood of purchase for several product pairs, wherein the product recommendations for each product are the products with the largest likelihood of purchase,
• d. personalizing the product recommendations for the target product for the specific shopper by obtaining the product recommendations for the target product, adjusting the likelihood of purchase for each product recommendation based on at least one of the attributes of the recommended product matching an attribute in the subset of the group of attributes for the specific shopper, and adjusting the likelihood of purchase based on the specific shopper's preference value for the attribute, and
• e. automatically displaying in real-time via a web service one or more of the personalized product recommendations in order of the adjusted likelihood of purchase, wherein steps (a) and (c) are computationally complex and performed by an offline process that transforms behavioral data to product recommendation tables, and steps (b), (d), and (e) are performed by a real-time process that uses the product recommendation tables.