Athens.
Up next at WebSci '09 is Patricia Victor, who begins by noting the growth in recommendation systems, including, for example, the advanced functionality on Amazon and in other e-commerce applications. Some 60% of Netflix users, for example, base their viewing on recommendations, and Netflix has offered a US$10m prize for an algorithm that improves its recommendation system by 10%.
There are two classes of recommendation systems: systems which are content-based and systems which are collaborative filtering-based. The latter focusses on similarities in the rating behaviour of users, and trust-based systems are often based on such algorithms. Epinions offers such a social trust network, and also allows users to evaluate other users by placing them in their network of trust, thereby conferring particular importance on these users' trust ratings. This also alleviates the 'cold start' problem with new users; it provides more reliable and accurate recommendations and leads to a kind of trust propagation through the network.
Such algorithms variously understand trust as a kind of weight, as a collaborative filtering neighbour selector, or work with a combination of trust and correlation. (Patricia now moves through a number of implementations of such algorithms, by Golbeck and Massa for the first type, by O'Donovan for the second, and by Patricia herself for the third.) The problem of such algorithms is that they tend to have been tested on different datasets, so that no direct head-to-head comparison of these algorithms has been possible to date. Patricia has undertaken such a comparison using Epinions data.
But what about distrust? In the first place, how does propagation work here - if person A trusts person B, and person B distrusts person C, does this mean that C snould also be considered to be distrusted by A? Results appear to be mixed; distrust remains a difficult problem. There is a need for more applications and more datasets to test this in more detail.
