Detecting Click Fraud in Online Advertising: A Data Mining Approach

摘要

Click fraud--the deliberate clicking on advertisements with noreal interest on the product or service offered--is one of themost daunting problems in online advertising. Building aneffective fraud detection method is thus pivotal for onlineadvertising businesses. We organized a Fraud Detection inMobile Advertising (FDMA) 2012 Competition, opening theopportunity for participants to work on real-world fraud datafrom BuzzCity Pte. Ltd., a global mobile advertising companybased in Singapore. In particular, the task is to identifyfraudulent publishers who generate illegitimate clicks, anddistinguish them from normal publishers. The competition washeld from September 1 to September 30, 2012, attracting 127teams from more than 15 countries. The mobile advertising dataare unique and complex, involving heterogeneous information,noisy patterns with missing values, and highly imbalanced classdistribution. The competition results provide a comprehensivestudy on the usability of data mining-based fraud detectionapproaches in practical setting. Our principal findings are thatfeatures derived from fine-grained time-series analysis arecrucial for accurate fraud detection, and that ensemble methodsoffer promising solutions to highly-imbalanced nonlinearclassification tasks with mixed variable types and noisy/missingpatterns. The competition data remain available for furtherstudies at href="http://palanteer.sis.smu.edu.sg/fdma2012/">palanteer.sis.smu.edu.sg/fdma2012. color="gray">