Trust Propagation and Personalization on eBay by Vesile Evrim

Introduction:

The easy access of devices such as mobile phones, computers and PDAs make it possible for almost anybody to be an information provider. As the number of uncontrolled information providers increases, choosing trusted sources for communication become an essential issue. Today, the internet is one of the best examples of the environments with major trust issues in many subareas, such as e-commerce, P2P systems, gaming, and virtual reality.

As one of the most popular e-commerce sites, eBay is the environment that is having millions of hits every day. The structure (buyer, seller) and the availability of the information (ratings, comments) provide us the base to analyze trust propagation and personalization on eBay where can be extended to the other applications in the future work.

Approach:

According to our analysis, comments on eBay are over 90% biased towards positive which represent unnaturally positive trust scores in the system. We are proposing that trust values can be propagated throughout an e-commerce application between buyers and sellers, and that we can harness this information to compute a tailored trust value for a previously unseen user.

The main approach in this research is to use comments instead of ratings in the trust calculations. We do believe that the comments provided by users have more information about the trustworthiness of the other sources, products than the provided binary ratings. In order to analyze and understand user comments on eBay, we have developed a technique for approximating the goodness of a user comment for the purposes of building our trust graph.

AuctionRules Algorithm

AuctionRules operates under the assumption that people will generally use the same set of terms to express some form of dissatisfaction in their online auction comments. The algorithm captures negativity in comments where users have complained but still marked the comment as positive. AuctionRules is a machine learning algorithm. As with most machine learning techniques, training examples were required for the algorithm to learn. The algorithm works only with words and phrases which explicitly express negativity. Many of the words, expressions and characters in the raw comments were of no value to the learning process, so before training examples were compiled, preprocessing was done to reduce complexity (Figure 1)

Figure 1 Graphical overview of the trust-modeling process.

Results

Initially we crawled over 10,000 comments from the eBay site. We used comments in two experiments; firstly Comparing Classifier Accuracy for Computing Trust and then Comparison of distributions between current eBay trust values and AuctionRules generated values. We used only the set of comments which were rated by real people in our user evaluations. This is a set of 1000 classified user comments (Figure 2).

Results show a more realistic distribution using the AuctionRules values, and consistent improvements of up to 21% over seven popular classification algorithms. AuctionRules also produces a false negative rating of 0% compared with 8.1% from other tested algorithms.

For more details please refer to papers Personalizing Trust in Online Auction” and “Extracting and “Visualizing Trust Relationships from Online Auction Feedback Comments”.