Introduction

Web mining is a hot topic nowadays [1–۱۲]. Web mining algorithm can be used to find hidden information from a large number of web data, which has been widely used in e-commerce decision support applications, such as personalized recommendation. Personalized recommendation can improve e-commerce sales from three aspects: browser into buyers, increase cross selling, and build loyalty. No matter what kind of algorithm is used, the data model influences mining results greatly [13–۱۸]. The quality of a data model can be evaluated from three parts, including collecting, storing and preparing. Data collecting can deploy in client-side, proxy-side and serve-side. Client-side collecting can be forbidden by users for privacy reasons. Proxy-side collecting can only collect the behavior of clients that send requests to the proxy. Serve-side collecting can collect all server activities of all accessing clients. Files (structured or unstructured) and databases are the main mediums for web data storing. Data preparing, including cleaning, integration, selection and transformation, is to pre-process stored data and transform them according to the requirement of web mining algorithms. At present, there are several kinds of common web data source model, including web log [19], application server log [20], Client-side log, Packet sniffer [21], and 5-gram united events model [22]. Web log is the most frequently used data source in web mining system. This data is common in a variety of web servers, such as IIS, Apache, etc. Kohavi [23] points out the shortcomings of the web log as a data source for the web mining algorithm: (1) poor data collection. First, due to the impact of local cache, proxy servers and firewalls, part of the data can’t reach the server, which causes it is not complete in web log data collection. Secondly, there are many problems in the clock synchronization of multi file system. Thirdly, the web log contains a lot of redundant information, such as the JPG, BMP and other pictures downloaded due to the need of Http protocol. (2) It is not designed for analysis. First, the URL string lacks the semantics, which is logged by web log. Especially in today’s most dynamic web site, URL analysis is almost meaningless because the page content and site structure can be dynamically generated. The second aspect is the lack of storage form information,