Yet more recommendations
The paper from Macedo et al isn’t a million miles away from my idea. Essentially, it appears that their WebMemex system (apologies to Mr V. Bush? He’s in the references!) is made up of a tool to track a users web activity (ie. a record of page requests), a tool to link documents together by commonality of terminology (using Latent Semantic Indexing theory), and a system to recommend documents via the web activity of other people in an identified group.
There are a few things I found interesting and useful in this approach (other than Latent Semantic Indexing, which I read about, but didn’t understand, in another paper on a different issue):
- no recommendations are made per se – this is a tool that shares browsing activity;
- the yahoo chat buddies system is used to identify the social networks;
- trust is considered to be a reciprocated buddy, and you must have two or more reciprocated buddies for the system to work in order to protect the privacy of an individuals browsing path (with recommendations from only one buddy, you would know exactly which pages your buddy had visited).
I really like the fact that the social network is established outside of the application itself. The ties of trust should be able to work for other non-overlapping applications (in the same way that Sainsbury’s, BP, Barclaycard and Debenhams and share the Nectar Loyalty card): say for instance that Google, Amazon and Lastminute all used the same outsourced “trust” recommender system. This would reduce the burden on the user of duplicating information (single sign-on and then opting in to which companies can use your information), make it easier to maintain and much more likely that people would sign up to it (I reckon).
Equally, I appreciate the touch of protecting privacy through the use of a trust network only being valuable with anonymised input. I trust my friends’ opinions, but I don’t necessarily want to know which friend has made the recommendation (online trust being different in many ways to face-to-face trust).
It’s also made me think a little bit more about the trust network: where does the value come into play? I don’t really agree that reciprocation is necessary for trust – it’s more a case of mutual trust, so I don’t think it should affect any “trust values”. Additionally, I feel that mutual trust networks are more likely to be closed: closed networks in being less expansionist are less likely to be of use in covering a large dataset. I think that stronger and weaker trust links may overcome this: if I trust person A who in turn trusts person B, I may have some implicit trust in person B’s opinion, albeit at a lower level than that which I have with person A. As a rule of thumb, I would have thought that the small world six degrees of separation could express the outer boundary of trust value.
On another point, the principle of “it can be assumed that friends have common understandings and interests” expressed in the paper is an important one for trust. We cannot be 100% sure that I always trust A’s opinion on every matter, but it will be too time-consuming a task to filter out the unnecessary information. However, if the above assumption is true, value should come from the fact that A makes his recommendations in generally similar spheres to your interests, and that you are likely to be searching for recommendations within these spheres – which is the whole principle behind the pattern-matching side of recommendation engines anyway.
Overall, I have returned to my original idea, that, rather than seek to augment the quality of overall search results with a “trust weighting”, which I have been informed may be mathematically muddled in principle, a “search trusted sources” advanced option is the best bet. Essentially this could mean that a search engine could use keyword matching and popularity for a general trawl search where you’re just looking for as much information as you can find on a particular subject, and offer a “trusted sources” search where you really need to believe in the quality of the found artefacts (and then other searches for particular, known items etc.)
Anyway, this all bears a lot more structured thought than 15 minutes bashing it all out on a keyboard – I might return to this subject later. All comments appreciated, particularly those that may point out the errors in my logic!