Following the ACM TechTalk, Joe Konstan was kind enough to answer some additional questions we were not able to get to during the live event. These questions and answers are presented below:
Q: I’ve been working with salesforce/cloudsense, where does that fit into the galaxy?
A: Recommender systems are a underlying technology. Platforms such as salesforce and cloudsense maintain customer history data and can use recommendation algorithms to customize interaction with customers.
Q: Can we group recommendation systems under the field of AI?
A: Absolutely, the algorithms in recommender systems certainly come from AI. At the same time, the idea of collaborative filtering comes out of the CSCW (computer-supported cooperative work) and CHI (human-computer interaction) communities. Like any application, recommender systems draws on many specialty areas in computing to provide solutions.
Q: Are there special recommendation Algorithms for a conversation context, e.g. For cross-/Up-selling in a c-commerce scenario. One of the challenges is, that the recommended products should be different than the base product, e.g. Select the right tie fitting a shirt, but don’t recommend another shirt. Also the conversation should provide a convincing explanation to the user.
A: Conversational recommender systems has been an important area of research and practice from the early days of the field. A number of systems have demonstrated how to bring together dialogue management and product databases to help users find the right items.
Q: Haven’t film/music/arts critics been doing recommendations for a long while? However they don’t generally say “if you like that then you’ll like this”. Critics usually try to comment on the quality of things (e,g,“great cinematography”) and leave subjective questions to the consumer.
A: Absolutely, critics have been recommending for hundreds or thousands of years. Sometimes they very much do recommend through comparisons (if you didn’t like movie X, you won’t like Y). In fact in the book field there’s a whole database for librarians based on “read-alike” lists to help people find other books that “read similarly to” each other. Critics have a challenge: they’re trying to be experts but also to help non-experts, which is why they often highlight the features of a movie (e.g., it’s violent, or funny) to help us figure out if we’ll like it.
Q: Wouldn’t it be important to recommend right thing at the right time even if the customer would have found it anyway?
A: That depends on the application and goal. In a retail transaction, it might make sense to make a “would have found it anyway” recommendation at the very last moment (to avoid missing a sale) or very early (to get the customer engaged in a shopping session), but less so along the way when the alternative may be a greater value opportunity. My point is to get the system designer to think about that decision and not simply make the obvious recommendation because it is easy.
Q: Does diversity/serendipity preference change if the recommendations are explicitly labeled as “old favorites” vs “discover something new” (or the like)?
It can. But then we still need to decide how many “old favorites” or “something new” items should be recommended in a particular circumstance. As your question makes clear, you can’t separate the algorithmic recommendation from the design of the interface.
Q: What is an example of a successful recommender system?
Netflix, Amazon, Google, almost all the news aggregators.