Personalized Check-in Prediction Model Based on User’s Dissimilarity and Regression

Abstract

To solve the problem that the user check-in prediction model is difficult to provide personalized check-in services, this paper proposes a novel hybrid model, called personalized check-in prediction model based on user's dissimilarity and regression (UDR). The UDR is mainly composed of two sub-models: user's regression location prediction model (UR) and user's dissimilarity location prediction model (UD). In UR, considering the personalization of user check-ins, we propose a hybrid weighted Markov model, which combines the whole check-in data and individual check-ins. Different from other methods, for the prediction of individual check-ins, we not only consider the preference of individual users, but also the influence of friend relationships. Meanwhile, the Hidden Markov model(HMM) is used to determine users' next check-in location by using time series feature (week-hour) and location sequence. In addition, by improving the kernel density estimation, we propose a multi-level hybrid kernel density estimation model, which is built based on the individual, city and region layers, and smoothes the over-fitting phenomenon caused by few check-ins. In UD, we take into account the weather factors that most existing methods did not consider. By defining the “cold and hot spot transference” and weather similarity features, we explore the influence of weather on user's check-ins and also propose a method used to calculate the similarity between user check-in weather preferences and location weathers. At the same time, the influence of social, time, and space factors are also considered. The experiments on two LBSN datasets demonstrate that the performance of UDR is superior to the state-of-the-art check-in prediction methods.