A Location-Based Indoor Recommendation System Using A Hidden Markov Model

Abstract

Recommendation systems play an important role in both the industrial and the academic worlds because they can combine theoretical study and market use together perfectly. Through improving user experience, recommendation systems bring considerable profits for businesses. However, most recommendation system applications are on the Internet. Due to the limitation of data-collecting techniques, economic costs, customer privacy protections and other factors, in online retail stores, usually the same advertisement will be pushed to every customer with no differentiation. Advertisements are usually chosen only based on the selling rates of goods, which cannot deliver advertisements to customers precisely. With the development of RFID technology, a way of tracking customers' movements and purchasing behaviours with low costs and privacy protections is provided. Hidden Markov model (HMM) is a statistical Markov model in which the system being modeled is assumed to be a Markov process with unobserved (hidden) states. Therefore, it performs well when predicting results for things which patterns are hard to find. With only sequences of observations vations, HMM is able to predict the next emission. In this thesis, a recommendation system that recommends goods for customers in a grocery store is proposed and implemented. To simplify the problem, our research considers zones in a grocery store that are detected and strictly separated by RFID readers and every shopping cart's movement stands for a customer's movement. Our system takes location information as well as customer movements into consideration as a parameter (state) for training the HMM model. Every one of our designed algorithms is compared with probability selection method, and we prove that by adding location and customer movements parameters, our solution has a higher hit rate for recommending the right goods for customers.