Tailoring online shopping recommendations
Recommendation systems apply in online shopping contexts. Typically, the system will suggest other items based on a user's previous choices. Yet such 'collaborative filtering' systems do not learn very well, and may never realise that a consumer might be interested in a not-yet-popular product. Otherwise, a system may employ content filtering to recommend additional products based on the content of a consumer's previous purchases. This is also inefficient because the offered products would be very similar, and thus redundant, to purchased products. The EU-funded ASSORTMENTCUSTOMIZE project aims to develop software methods for tailoring recommendations given to users according to their individual characteristics. It uses purchase data to determine consumer class and situational factors based on interest, purpose or age. The system uses these classes and product categories, along with product price and cost, to compile the best list of suggestions for the particular consumer. Each decision the consumer makes in the future will refine the algorithm. Such computations are normally time consuming; however, the ASSORTMENTCUSTOMIZE tool will be able to do it quickly using innovative programming. This will enable retailers to offer sets of product suggestions to customers that match their characteristics and are also therefore profit-maximising. The project will test the algorithms in the cable television marketplace, but the technology will be applicable to other choice-rich contexts. During the project's first phase, the original parameters of the model were redesigned to allow the software to learn any number of customer characteristics simultaneously. Previously, it had been limited to one at a time. The project also obtained a large set of choice data from an American video retailer, which establishes the customer histories necessary for the project. Simulations using these data and an optimised algorithm led to a working web version prototype. Numerous other software optimisations have also been achieved. The next phase was a test of the system in June 2013 involving over 2 000 users. This makes the results more realistic. The resulting paper will be submitted to Management Science, a prestigious journal.
