Characterizing Users in an Online Classified Ad Network

Unlike online social networking sites (e.g. Twitter and Facebook) which are heavily used for disseminating content and sharing information between users and shopping sites (e.g. Ebay) where buyers and sellers are reviewed, the flow of information between users such as buyers and sellers in a classified ad network is very limited. Characterizing users or assigning them to some classes in one such network is challenging due to the sparsity of the data about users, the vague separation of user classes and sometimes the tendency of users to hide or misrepresent their profile information. In this paper, we study the information revealed in the ads posted to an online classified ads site and analyze the behaviour of users posting those ads; our study is conducted using data collected from Kijiji over a year. We study the problem in the context of one specific task where we seek to detect if a user posting an ad belongs to one of the two classes business and non-business, based on the ads the user has posted. We study an approach based on user profiling, where given statistics on how an ad mentions terms and features from a class profile, the affinity of an ad (and subsequently a user) to a particular class is determined. We report the effectiveness of this approach in detecting user classes solely based on the information revealed in their ads and study the impact of the profile size on the accuracy. In the absence of labeled training data, we show that a simple bootstrapping technique with only a few n-grams as a seed set can give nearly good results in terms of F-measure. We further report our experiments on characterizing the collective behavior of users in posting ads and some of the distinctive usage patterns that emerge.