PhD Research Proposal: Trust-aware Decentralized Recommender Systems

14 Replies

I realised today I didn’t write yet an entry about my PhD Research Proposal “Trust-aware Decentralized Recommender Systems” (TaDRS).
So here it is the PDF file. If you have any comment or criticism, I’ll be happy to hear from you.
The PhD research proposal is a little bit outdated (29th May 2003) but I didn’t have a blog at that time. Enjoy and let me know what you think.

UPDATE:
Abstract
This PhD thesis addresses the following problem: exploiting of trust information in order to enhance the accuracy and the user acceptance of current Recommender Systems (RS). RSs suggest to users items they will probably like. Up to now, current RSs mainly generate recommendations based on users’ opinions on items. Nowadays, with the growth of online communities, e-marketplaces, weblogs and peer-to-peer networks, a new kind of information is available: rating expressed by an user on another user (trust). We analyze current RS weaknesses and show how use of trust can overcome them. We proposed a solution about exploiting of trust into RSs and underline what experiments we will run in order to test our solution.

14 thoughts on “PhD Research Proposal: Trust-aware Decentralized Recommender Systems”

  1. Conor Hayes

    The following is an initial response to the points made in P. Massa’s PhD proposal “Trust Aware Decentralized Recommender Systems”.

    27/02/04

    It is clear that there is an increasing amount of data available on the Internet on a variety of subjects in the form of simple rating data to more elaborate opinions represented on Blogger sites. We can consider such data to be “expert knowledge” because it consists of descriptions/reactions from people based on their direct experience of the items/subjects in question. Although, these descriptions are subjective they incorporate features which would be difficult or impossible to extract using automatic feature extraction techniques such as aesthetic value, social and cultural responses.

    The question is how to harness such information to allow people find new information that is useful for them? Automated Collaborative filtering (ACF) systems have been successfully used to create personalization systems where there is very little content description available for the items being filtered. ACF systems operate by aggregating the preferences of a distributed set of users in order to make recommendations to another user. Typically, a similarity measure is employed in order to determine the weight that should be allocated to each user, so that the preferences of a ‘similar’ user contributes more to the predicted rating for an item than a more distant neighbour.

    Although, ACF systems have the advantage that detailed content description is not necessary, they have many drawbacks such as the ‘cold start’ problem, the issue of recommendation transparency and the fact that the user data for each domain is stored and owned by the provider of the web service.

    P. Massa suggests that these drawbacks can be addressed and the benefits extended in the following ways:

    1. Developing a set of reputation and trust metrics to allow users to tap into consistently good sources of information.
    2. Developing a set of protocols that allows opinion data to be used to power recommendation/personalization strategies between distributed users without the need to centralize this data. Thus, each user has ownership and control of the opinion data he/she makes public. An example of how this might be achieved is to consider each web log to be a logical peer, or information source and to leverage p2p technology to aggregate such sources of information.

    Some historical background
    While Collaborative filtering systems now rely upon automatically identifying nearest neighbours using similarity measures such as the Pearson Correlation (Hence the term ‘Automated Collaborative Filtering’, the earliest collaborative filters were much closer to the ideas being suggested by P. Massa in this paper. Collaborative filtering systems emerged out of early research into the problem of information overload in a networked environment (Denning 1982, Palme 1984, Hill & Turoff 1985). Malone identified three types of intelligent information filter that might allow people to find desired information and eliminate undesirable material: cognitive, social and economic (Malone et al. 1987). The categories are based on the information resources that each filter draws upon to make a prediction on an item for a particular user. Cognitive filtering systems analyse the text within documents and match the results against a user-profile. These types of systems are generally now known as content-based or knowledge-based filters. Economic filters select documents based on the costs and benefits of producing and reading them, a criterion which will become relevant as content on the Internet increasingly requires payment. Social filtering, according to Malone, selects documents based on the evaluation of other people with whom the target user may have a relationship. A social filtering system recognises that people may choose to read material annotated by someone in authority or by a trusted source.
    In its purest sense, collaborative filtering refers to a process of filtering solely according to the opinions of other users (Resnick et al. 1994). The first collaborative filtering systems, however, were essentially content-based systems augmented by some characteristics of social filtering. The Lens system of Malone et al. was primarily a rule-based content filter for e-mail that included a social filtering component which allowed the receiver to set his/her filter to receive messages according to the characteristics of the sender (Malone et al. 1987). The first social filtering system that involved a collaborative element was Tapestry, a mail filtering system developed at the Xerox Palo Alto Research Center (Goldberg et al. 1992). Tapestry allowed members of its user base to annotate each document with details on how interesting (or uninteresting) the reader had found the document. The document contents and the annotations could be accessed by each user’s filter. This allowed users to filter the system for documents on a topic that had been annotated by a particular person, or to receive documents ‘replied to by user A AND user B’. The Tapestry system had some drawbacks. Users needed to learn a query language for the system, TQL, and were required to actively annotate their documents for the system to work effectively. Furthermore, since it was assumed that users worked in a relatively small environment it provided no means of automatically identifying users with similar interests. Maltz and Ehrlich employed a similar approach using the term active collaborative filtering to describe a process of providing ‘pointers’ to recommended documents for their friends and colleagues (Maltz & Erlich 1995). The relationship between these systems and the ‘collaborative bloggers’ proposed in this paper is evident.

    It is interesting to note that the researchers behind Grouplens, the first Automated Collaborative filtering System, identify the need to identify trusted sources a priori as a major weakness of a collaborative filtering System like Tapestry (Resnick et al. 1994). A similarity metric was introduced to automatically identify potentially useful sources of information from the pool of users in the Grouplens system. Thus the similarity metric used in ACF was originally intended as a proxy for a measure of trust or reputation. Accordingly, it is quite timely that the relationship between similarity metrics and trust metrics be investigated again.
    I would suggest that some of the older papers relating to early collaborative filtering/social filtering be examined. The idea of trust/reputation is implicit in the systems described in these papers, whereas this idea is lost in the later papers on automated collaborative filtering which consider only the concept of similarity.

    References

    Denning, P. (1982) Electronic junk. Communications of the ACM 23(3), 163–165.

    Hiltz, S.R., Turoff, M. (1985) Structuring computer-mediated communication systems to avoid information overload. Communications of the ACM 28(7), 680–689.
    Malone, T.W., Grant, K.R., Turbak, F.A., Brobst, S.A., Cohen, M.D. (1987) Intelligent information sharing systems. Communications of the ACM 30(5), 390–402, May.
    Maltz, D., Ehrlich, K. (1995) Pointing the way: Active Collaborative Filtering, in: Proceedings of ACM SIGCHI Conference, 1995.
    Palme, J. (1984) You have 134 unread mail do you want to read them now? in: IFIP Conference on Computer Based Message Services, Nottingham, UK, IFIP
    Resnick, P., Iacovou, N., Suchak, M., Bergstrom, P., Riedl, J. (1994) An open architecture for collaborative filtering of Netnews, in: ACM Conference on Computer Supported Co-operative Work, 1994, pp. 175–186.

  2. Pingback: Web Dawn - Rebirth of the Social Marketplace

  3. Pingback: Web Dawn - Rebirth of the Social Marketplace

  4. Pingback: Web Dawn - Rebirth of the Social Marketplace

  5. Pingback: Web Dawn - Rebirth of the Social Marketplace

  6. Pingback: Web Dawn - Rebirth of the Social Marketplace

  12. Pingback: Web Dawn - Rebirth of the Social Marketplace

Leave a Reply

Your email address will not be published. Required fields are marked *