Article Preview
Top1. Introduction
With the rapid growth of mobile users and business in the world, short message service (SMS) has become not only an important way in daily work and life but also a massive commercial industry due to its convenience and inexpensive price (Wu et al., 2008; Almeida et al., 2011; Sohn et al., 2009). At the same time, criminals use SMS to defraud and earn economic or political benefits. This situation has become more serious over the years, especially in India, Pakistan, and China where SMS lacks of effective supervisions (Delany et al., 2012). Reports from the Chinese National Spam Report Center show that Chinese mobile phone users received 12 spam messages every week in average, and the total number of spam messages reached 120 million in 2014 (CNSR, 2014; Qihoo 360, 2014). In the USA, more than 69% of the surveyed mobile users claimed to have received text spam in 2012 (Jiang et al., 2013).
Spam messages are unsolicited and unwanted messages, including advertisements, frauds, erotic services, etc. They disturb normal life, consume the resource of mobile communication equipment, and lead to congestion of communication network (Wu et al., 2008; Sohn et al., 2012; Wadhawan & Negi, 2014).
Most of the current SMS spam-filtering technologies are transplanted from email spam filtering, such as black and white or sensitive keywords lists, high frequency blocking. Besides of above, the technologies of test mining and machine learning are widely used for blocking spam message by their content, for example, SVM, K-NN, Bayes (Delany et al., 2012; Wadhawan & Negi, 2014). But they usually face the following challenges (Almeida et al., 2011; Liu & Wang, 2010; Sohn et al., 2012).
- 1.
The traditional classifiers usually use vector space model (VSM) to represent text which is apt to long text. It is easily to cause similarity-drift when it is used for SMS messages. Moreover, VSM also lacks of contextual and semantic information.
- 2.
Many of existing studies about SMS spam filtering often ignore some important characteristics in the content of SMS spam. They are usually injected with many kinds of abbreviations, symbols, websites, variant words, and distort or deform sentences, which are used deliberately by the spammers to avoid anti-spam system filtering
- 3.
The definition of SMS spam is different from country to country. For example, business advertisements are permitted in many countries, even as one of the important value-added services for the mobile operators (Delany et al., 2012). But it is consider as violating personal rights in most advanced countries. Nevertheless the existing SMS spam filtering methods usually deal with the SMS spam as binary-class problem: ham and spam, which couldn’t provide for multi-grain categories.
In our study, we present a Message Topic Model (MTM) which is based on the probability theory of latent semantic analysis, and more suitable for the task of SMS spam filtering. Compared with the existing SMS spam filtering technologies and standard topic model—Latent Dirichlet Allocation (LDA), MTM is fitter for SMS filtering due to the following main factors: