关键词: 非负矩阵分解, 垃圾短信, 伪基站, 时空规律, 北京

Abstract:

The rampant pseudo base stations have become a major public hazard. They undermine the normal telecommunications order, endanger public safety, seriously infringe the property rights of the masses, and violate citizen privacy. How to dig out the spatio-temporal patterns of the pseudo base stations’ activities from massive spam messages, design effective prevention and control programs, and fight against the crime from the source, has become the focus of government agencies and researchers. The traditional methods for identifying pseudo base stations through the user terminal, however, face great challenges in terms of accuracy, comprehensiveness, and analytical ability, which no longer meet the requirements of identifying small-scale and mobile pseudo base stations. Utilizing data on the spam messages from February 23^rd, 2017 to April 26^th, 2017 in Beijing, this paper analyzes the spatio-temporal distribution of pseudo base stations through non-negative matrix factorization. We also constructed a classification model through TF-IDF (Term Frequency-Inverse Document Frequency) which compares types from different classifiers (k-Nearest Neighbors / K-Support Vector Machine /Random Forest/ Single-Layer Neural Network) and selects the most accurate random forest classification method. Combined with the land use data, we analyzed the spatio-temporal distribution of pseudo base stations that send different types of spam messages. The results of non-negative matrix factorization and spam message classification were analyzed in detail. The results show that most of the spam messages in Beijing are sent along the road network and in the central city. The number of spam messages during the day is much more than that during the evening. As time goes by in the day, the distribution of spam messages along the road network gradually shrinks inward. The pseudo base stations that send different types of spam messages differ in the spatio-temporal distribution, but all of them favor the traffic facilities and residential area within the Fourth Ring. The non-negative matrix factorization, which provides reliable results that match with traditional spam message classification, has shown simplicity in performing the analysis and interpretability in the form and result of the decomposition. It can help understand the spatio-temporal patterns of different types of spam messages and provide evident-based suggestions for government agencies to fight against the pseudo base stations effectively. By targeting the source of the spam messages, it is also beneficial for governments to combat the illegal behaviors based on pseudo base stations.

Key words: visualization analysis, non-negative matrix factorization, spam messages, spatio-temporal patterns, Beijing