Recommending for Bentham
One of my favorite people-and-preference papers is Kahneman, Wakker, and Sarin’s 1997 article ‘Back to Bentham? Explorations of Experienced Utility’. They report on a series of experiments that examine how people experience and remember utility, pain, or pleasure, and the ways in which it affects their decision-making processes.
This is the paper that introduced me to the peak-end rule, the principle that the pleasure (or pain) people remember an experience having is effectively the average of the high point and the end. Total utility and duration are both irrelevant (within reason, at least).
Lesson 1, if you want to make a movie people remember as funny: have a fantastic joke and a solid ending.
But one of the things I found particularly profitable about this paper is the questions it examines about the difference between how users experience something during the event and the way that they remember the experience (and how it factors into their future decision-making). These are not necessarily the same thing! In a prolonged medical procedure, the person is actually experiencing the pain throughout, but their memory is based on the peak and end. One of the manipulations that Kahneman and his collaborators tried was to modify a colonoscopy procedure so that the exam took longer but ended with less pain. The result was that patients remembered experiencing less pain, even though they didn’t really decrease the pain experienced at any given time during the procedure.
This difference could affect our understanding of how users interact with recommender systems, though I have yet to see a recommender or evaluation directly leverage it.
When we recommend, we are usually trying to maximize the user’s utility. However, we should ask what utility we are trying to optimize. Do we want to maximize the user’s immediate utility while watching the movie? Do we want to optimize their remembered utility? The utility that arises from its long-term affect on their life, which is built in large part on remembered utility but may include additional components? Are we optimizing their expected utility, as reflected in their decision of whether or not to click ‘play’?
Last week in my recommender systems class, we drew a chart that looked something like this:
|‘Explore more like this’
If we are evaluating a recommender on its ability to recommend items that a user will purchase or watch, in most cases (barring re-purchases of previously-consumed items), we are measuring its ability to produce expected utility: to recommend items that the user expects to enjoy.
This may be highly correlated with remembered or long-term utility, which may well be good enough in many applications. But if we want to carefully understand the recommender’s impact on its users, we are not measuring either experienced or remembered utility.
We can get at experienced utility with consumption ratings: looking at completion (if a user finishes reading an article or watching a movie, they at least found enough utility not to turn it off) and explicit consumption ratings (which combine experienced utility with an expression of desire to experience more such utility in the future).
Remembered utility basically requires memory ratings. They’re relatively hard to get compared to implicit data, and are quite noisy (although that is partially reflective of the noise of memory itself, so that isn’t all bad).
If we want to distinguish my memory of a movie and that movie’s impact on my life, things get very difficult, and it might not even be possible.
I haven’t yet seen recommender systems research tackle this issue head-on. I think Bart Knijnenberg’s self-actualization agenda will likely benefit from some insight into it, though.
I look forward to someday having insight into what difference it makes if we recommend movies users expect to enjoy, will enjoy, or will enjoy having watched.