Click Analysis And Prediction For Product Search On C2C Web Sites

Millions of dollars turnover are generated every day on popular ecommerce web sites. In China, more than 30 billion dollars transactions were generated from online C2C (Customer to Customer) market in 2009. With the booming of this market, predicting the click probability for search results is crucial for user experience, as well as conversion probability. The objective of this paper is to propose a click prediction framework for product search on C2C web sites.We segment this paper into three main parts. In the first part, we provide a briefly description on the characteristics of product search and an analysis on the page view and click of search result for product search is described. In the second part, we present the overview of the system framework for click prediction, as well as the mathematics models used in our system. Finally, we deeply analyze features on each dimension of product search and the experiments with analysis are presented.Click prediction is deeply researched for sponsored search and web search. However, to the best of our knowledge, few studies were reported referred to the domain of online product search. We validate the performance of state-of-the-art techniques used in sponsored search for predicting click probability on C2C web sites. Besides, based on the characteristics of product search, we segment features of prediction model into four different groups according to different roles in the process of product search: search, buyer, seller and product itself. Features of each group are deeply researched and prediction model is built on them. Plenty of experiments are performed on these models, and the results demonstrate that the combined model improves both precision and recall significantly.We discuss the problems existing in click models commonly. Position bias is a common problem in click logs. From statistical data, we can see users usually prefer to click items which are ranked in the higher position on the result page. Plenty studies reported are focusing on this problem. Based on these research results, we provide the solution which is fit to product search. Unbalanced data is another problem in the real click logs, which leads serious performance reduction in prediction model. In our paper, we try to give a solution and discuss on the experiment results.