A semantic-expansion approach to personalized knowledge recommendation

Abstract

The rapid propagation of the Internet and information technologies has changed the nature of many industries. Fast response and personalized recommendations have become natural trends for all businesses. This is particularly important for content-related products and services, such as consulting, news, and knowledge management in an organization. The digital nature of their products allows for more customized delivery over the Internet. To provide personalized services, however, a complete understanding of user profile and accurate recommendation are essential.

In this paper, an Internet recommendation system that allows customized content to be suggested based on the user's browsing profile is developed. The method adopts a semantic-expansion approach to build the user profile by analyzing documents previously read by the person. Once the customer profile is constructed, personalized contents can be provided by the system. An empirical study using master theses in the National Central library in Taiwan shows that the semantic-expansion approach outperforms the traditional keyword approach in catching user interests. The proper usage of this technology can increase customer satisfaction.

Introduction

The rapid growth of the Internet has changed the nature of many businesses. The large amount of transactional data collected from the use of information systems allows for better understanding of customer needs and the integration of knowledge for the customization of products and services. This is particularly important for content-based applications, such as consulting, news services, and knowledge management.

Due to the importance of product and service customization, Internet recommendation systems (also called the Internet recommender systems) have become an important research area in electronic commerce [46]. Its major purpose is to reduce irrelevant content and provide users with more pertinent information or product. A recent study indicates that the use of personalized recommendation can significantly increase user satisfaction due to its ability to offset information overload [27].

Many information filtering and recommendation methods have been developed in literature, most existing techniques fall into three categories: rule-based filtering, content-based filtering, and collaborative filtering. Rule-based filtering uses pre-specified if-then rules to select relevant information for recommendation. Content-based filtering uses keywords or other product-related attributes to make recommendations. Collaborative filtering uses preferences of similar users in the same reference group as a basis for recommendation.

Content-based filtering and collaborative filtering are more popular in practical applications. However, they both have limitations. The content-based approach can classify services based on their nature, but often have difficulties in identifying related interests of the same user. Collaborative filtering can find similarities among different users but is unable to handle new items that do not have existing usage information.

Content-based filtering is better than collaborative filtering when it is applied to digital products such as customized news services and document recommendation in knowledge management because documents like reports have certain semantic linkages that cannot be captured by collaborative filtering.

Two general directions are popular for using these methods: profile generation and maintenance, and profile exploitation. Profile generation and maintenance include user profile representation, profile generation, and relevance feedback. Profile exploration includes information filtering, user profile-item matching, and profile adaptation [34]. Profile generation explores the interests of a particular user, while profile exploration finds information relevant to a particular user query for recommendation. Since most queries use keywords for document search, information retrieval techniques can facilitate content-based recommendations. For example, if a user reads a report about knowledge management and auctions, recommending reports of interest to the user is similar to retrieving documents using knowledge management and auction as two keywords.

Typical research in information retrieval uses key word weighting to find documents relevant to a particular query [5]. Since key words in a user query often have semantic meanings and certain semantic relationships may exist in certain documents, simple key word matching may result in an underweight or overweight of certain key words due to their semantic similarities. Advanced techniques that take semantics into consideration in building and exploitation of user profiles are useful.

In this paper, we propose a semantic-expansion approach to build user profile and content recommendations. This approach uses semantic networks and the spreading activation model (SA) in cognitive psychology to build user profiles and then make recommendations accordingly [12], [13]. The method includes three modules: (1) analyzing document structures, (2) building user interest profiles, and (3) making recommendations. An experiment was conducted to compare the performance between the semantic-expansion approach and a typical keyword-based approach. The result indicates that the semantic-expansion approach significantly outperformed the key-word-only approach.

The remainder of this paper is organized as follows. In the next section, literature in information filtering and content-based recommendation is reviewed. In Section 3, the semantic-expansion approach is described. Section 4 presents a prototype implementation and findings from our experimental study. Finally, conclusions and suggestions for future research are discussed.