Study On User Attribute Inference Method For Sparse Features

With the pervasive availability of smart devices and the booming of social media,billions of users' behaviors are recorded and collected.The aggregated human behaviors reveal users' characteristics and reflect their demographic preference,i.e.,gender,age,marital status and even personality,interests and occupations.User profiling from social media data is an attractive option for precision marketing,personalized recommendation,and other commercial applications,without severe privacy concerns.However,it carries great difficulties in sparsity and vagueness.This paper proposes a novel approach,i.e.,Continuous Conditional Random Fields(C-CRF)based user profiling method.We first carefully analyze and report the correlation between different kinds of features and users' attributes based on a real-world dataset.And then we design methods to extract users' content,location,mobility pattern and network preference,especially designs a hierarchical-graph-based feature extracting method for sparse mobility pattern data.Finally,present a comprehensive C-CRF based prediction method to fuse features and predict attributes.Empirical studies confirm that the new approach outperforms other baseline methods,and it shows the discriminative power of check-in and mobility data to reveal users' preference.Main contributions of this paper are as follows:We collect,process and enrich a social media dataset,which consists of real users'posting records,check-in logs,corresponding place annotations and their social net-work,attribute information.We present a new hierarchical-graph-based feature extracting method for mobility pattern data,which alleviates the sparsity problem of mobility pattern records.We present a Continuous Conditional Random Fields(C-CRF)based user profiling method and conduct an extensive study and evaluation on a real-world dataset.It not only confirms the improvement of our proposed method but also provides new insights into the trajectory data analysis.