Research On Ad Recommendation System Based On Classification Algorithms

With the explosive growth of the Internet data, users are facing more and more kinds of information, and data mining technology used in personalized recommendation systems has become a research hotspot.This paper summarizes several kinds of common methods in recommendation system:collaborative filtering method, content-based method, hybrid methods, and data mining technologies related to them. Besides, we analyze the advantages and disadvantages of these methods. Then, this paper focuses on ad recommendation problems, and introduces two kinds of content-based methods in ad recommendation: single-label classification model and multi-label classification model.We develop a tool for visualization of statistics and analysis of ad log data provided by one internet company with the Hadoop platform, and we use this tool to analyze the data and find the dependency problems among features and ads. This paper proposes an improved single-label classification model using dependencies between ad features and other features, and an improved multi-label classification model using dependencies between ads. The former one selects and adds some combined features, which containing dependent information between features, according to mutual information between features. The latter one discovers the dependencies among ads based on a heuristic method which is used for building the classifier chain. Finally, we implement these algorithms and do some experiments on the Hadoop platform. Two improved methods proposed can show better results than other classic methods. Therefore, both dependencies between features and dependencies between ads are important factors in ad recommendation systems.