I wrote here last year about the marvellous Fishscale of academicness, as a great way to teach students information literacy skills by starting with how evaluate what they’ve found. I’m currently teaching information ethics to Masters students at Humboldt Uni, and this week’s theme is “Trust”: it touches on all sorts of interesting topics in this area, including recommendation systems, also known as recommendation engines.
An example of such a recommendation system in action would be the customer star ratings for products on Amazon, which are averaged out and may be used as a way to suggest further purchases to customers, amongst other information. Or reviews for hotels/cafes on Tripadvisor, film suggestions on Netflix, etc. Recommendations are everywhere these days: Facebook recommends apps you might like, and will suggest “people you may know” : LinkedIn and Twitter work in similar ways.
For me, these recommendations beg certain questions, which also turn up in debates about privacy and about altmetrics, such as:
How much information do you have to give them about yourself, do you trust them with it, and how good are their recommendations anyway? Are you happy to be influenced by what others have done/said online?
Recommendation systems use “relevance” algorithms, which are similar to those used when you perform a search. They might combine a number of factors, including:
- Items you’ve already interacted with (i.e. suggesting similar items, called an item-to-item approach)
- User-to-user: it finds people who are similar to you, eg they have displayed similar choices to you already, and suggests things based on their choices
- Popularity of items (eg Facebook recommends apps to you depending on how much use they’ve had) Note that this may have to be balanced against novelty: new items will necessarily not have achieved high popularity.
- Ratings from other users/customers (here, they might weight certain users’ scores more heavily, or average star ratings, or just preference items with a review)
- Information that they already have about you, against a profile of what such a person might like (eg information gleaned from tracking you online through your browser or on your user profile on their site, or that you have given them in some way)
The sophistication of the algorithm used and the size of the data pool drawn on (or lack thereof) might also depend on the need for speed of the system.
Naturally, those working on recommendation engines have given quite a bit of consideration to how they might evaluate the recommendations given, as this paper from Microsoft discusses, in a relatively accessible way. It introduces many relevant concepts, such as the notion that recommending things that it knows you’ve already seen will increase your trust in the recommendations, although it is very difficult to measure trust in a test situation.
We see that human evaluation of these recommendation systems is important as “click through rate (CTR)” is so easily manipulated and inadequate as a measure of the usefulness of recommendations, as described and illustrated in this blog post by Edwin Chen.
Which recommendations do you value, and why? I also came across a review of movie recommendation sites from 2009, which explains why certain sites were preferred, which gives plenty of food for thought. From my reading and experience, I’d start my list of the kind of things that I’d like from recommendation systems with:
- It doesn’t take information about me without asking me first (lots of sites now have to tell you about cookies, as the Cookie collective explain)
- It uses a minimal amount of information that I’ve given it (and doesn’t link with other sites/services I’ve used, to either pull in or push out data about me, unless I tell it that it can!)
- Suggestions are relevant to my original interest, but with the odd curveball thrown in, to support a more serendipitous discovery and to help me break out of the “filter bubble“
- Suggestions feature a review that was written by a person (in a language that I speak), so more than just a star rating
- Suggestions are linked in a way that allows me to surf and explore further, eg filtering for items that match one particular characteristic that I like from the recommendation
- I don’t want the suggestions to be too creepily accurate: I like to think I’ve made a discovery for myself, and I doubt the trustworthiness of a company that knows too much about me!
I’m sure there’s more, but I’m equally sure that we all want something slightly different from recommendation systems! My correspondence with Alke Groeppel-Wegener suggests that her students are very keen on relevance and not so interested in serendipity. For me, if that relevance comes at the expense of my privacy, so that I have to give the system lots of information about myself, then I definitely don’t want it. What about you?