Content-based mobile spam classification using stylistically motivated features

Abstract

The feature of brevity in mobile phone messages makes it difficult to distinguish lexical patterns to identify spam. This paper proposes a novel approach to spam classification of extremely short messages using not only lexical features that reflect the content of a message but new stylistic features that indicate the manner in which the message is written. Experiments on two mobile phone message collections in two different languages show that the approach outperforms previous content-based approaches significantly, regardless of language.

Introduction

Mobile spam, which most often refers to the abuse of the Short Message Service (SMS) to indiscriminately send out unsolicited (and unwanted) bulk messages, has progressively become a major issue from the early 2000s with the increasing popularity of mobile phones. Spam messages may be particularly irritating for many recipients because of the inconvenience that they cause, and also because of the fees that may apply for every message received according to the market; for example, fees apply for recipients in the United States. In order to reduce mobile spam, governments and many mobile service providers have taken various countermeasures (e.g. by imposing substantial fines on spammers, blocking specific phone numbers, creating an alias address, etc.). Nevertheless, the mobile spam rate is still on the rise.

Recently, a content-based approach for spam classification, which had previously brought enormous success in detecting e-mail spam, has started to gain attention among mobile spam researchers. Gómez Hidalgo et al. (2006) explored the use of statistical learning-based classifiers that are trained with lexical features, such as character and word n-grams, for mobile spam classification; the approach explored classification similar to e-mail spam classification (Sahami et al., 1998, Cormack and Lynam, 2007). Unlike for e-mails, however, mobile phone messages are often much shorter, because of the limited size of mobile devices. The feature of brevity causes each message to have fare less information for content-based spam classifiers, which makes the task very challenging. Therefore, succeeding studies on mobile spam classification focused on expanding the feature set for learning the classifiers with features additionally engineered from the message content, in order to consider the contextual information with relative word positions (Cormack et al., 2007a, Cormack et al., 2007b).

Collectively, the features used in earlier content-based approaches for mobile spam classification were topical terms and phrases that statistically indicated the spamness of a message, such as “% off sale,” “poker,” and “no pay”. However, there is no guarantee that legitimate messages would not contain such expressions. For example, an average person may send messages such as:

“Danny, ABC mall is having a 56% off sale tomorrow!!”

“Jane, we are planning to play poker with no pay tomorrow night, got it?”

Current content-based mobile spam filters may label such legitimate messages as spam messages. Such a misclassification can cause serious problems for instant mobile communication. We have thus decided to not only depend on the message content itself, but also to incorporate new features that indicate how the messages are written from a linguistic point of view. We propose the use of new features that reflect this “style”, or the manner in which the content is expressed, namely stylistic features.

In this paper, we introduce a number of stylistic features (Sohn et al., 2009)² that originated mostly from authorship classification, which involves identifying the author of a given text, for learning mobile spam classifiers. The features include readily extractable and countable information from message texts, such as word and sentence lengths (Mendenhall, 1887), function word counts (Mosteller and Wallace, 1984), part-of-speech tags (Argamon-Engelson et al., 1998), and syntactic information (Stamatatos et al., 2000). We then use a machine learning-based classifier to learn such features. We conduct an empirical evaluation that uses two real-world SMS test collections.

The remainder of this paper is organized as follows. Our mobile spam classification algorithm is briefly described in Section 2. In Section 3 we introduce the proposed stylistic features. In Section 4 Experimental design, 5 Results and discussion, we report the setting of our experiments and discuss the results. Finally, we conclude the paper in Section 6.