Distributed context discovering for predictive modeling

  • Z. Rong

Student thesis: Master


Click prediction has applications in various areas such as advertising, search and online sales. Usually user-intent information such as query terms and previous click history is used in click prediction. However, this information is not always available. For example, there are no queries from users on the webpages of content publishers, such as personal blogs. The available information for click prediction in this scenario are implicitly derived from users, such as visiting time and IP address. Thus, the existing approaches utilizing user-intent information may be inapplicable in this scenario; and the click prediction problem in this scenario remains unexplored to our knowledge. In addition, the challenges in handling skewed data streams also exist in prediction, since there is often a heavy traffic on webpages and few visitors click on them. In this thesis, we propose to use the pattern-based classification approach to tackle the click prediction problem. Attributes in webpage visits are combined by a pattern mining algorithm to enhance their power in prediction. To make the pattern-based classification handle skewed data streams, we adopt a sliding window to capture recent data, and an undersampling technique to handle the skewness. As a side problem raised by the pattern-based approach, mining patterns from large datasets is addressed by a distributed pattern sampling algorithm proposed by us. This algorithm shows its scalability in experiments. We validate our pattern-based approach in click prediction on a real-world dataset from a Dutch portal website. The experiments show our pattern-based approach can achieve an average AUC of 0.675 over a period of 36 days with a 5-day sized sliding window, which surpasses the baseline, a statically trained classification model without patterns by 0.002. Besides, the average weighted F-measure of our approach is 0.009 higher than the baseline. Therefore, our proposed approach can slightly improve classification performance; yet whether this improvement worth deployment in real scenarios remains a question.
Date of Award31 Dec 2013
Original languageEnglish
SupervisorMykola Pechenizkiy (Supervisor 1) & J. De Knijf (Supervisor 2)

Cite this