So, I have been forced to think about recommendation on the web in recent weeks. My former employers laid me off from work and I had to interview and answer the question: “So what kind of recommendation algorithms did you work on?”
There have been many categorizations of recommendation systems, and here I will list just a few that I have had the chance to describe to my potential employers. The description is non-technical because often it is difficult to get into details of the algorithm due to numerous NDA’s I have signed at many companies. However high-level description should be okay, and please let me know informally and discreetly if you object to my listing them. [[This is to say, more explicitly, I do not believe any of this information is secret. And if they are your secret, it is my believe that either you have made the information public it or that I found it independently.]]
In the most grossly general way, a recommendation is the indication of availability and desirability of a specific object to be acted upon by the internet user so as to benefit the recommender and the internet user. Typically, one can think of a product (balloons, diapers, beer, games), media (movie/music/picture), or textual content (news story, FAQ articles). Typically also, the indicator of availability is either a specifically designated area with a creative which represents the product and text describing it. The indication could also be a popup window, a link, a pre-roll, post-roll, inlaid text or video. The indication of availability could also be sounds of speech, music, or other human audible sounds.
The benefit to the recommender typically materializes when the internet user purchases a product, video, music. But there could be intangible benefit that are quantified by surrogate measurements: clicks, hovers, add-to-carts, etc.
The User Experience
Ultimately tho, if I had to explain to the internet user what it is that I am doing, then these are the typical categories of explanation out there:
- Relevance: This is your run-of-the-mill Collaborative Filtering algorithm making (item->item) recommendation, Items relevant to current item. (creative->blog page) (the blog page’s text creates the context for which to calculate relevance, AKA Contextual Recommendation/Advertisement), etc. Content Based Filtering may also apply.
- Personalization: Typical collaborative filtering algorithm for making (item->user) recommendations. Typically, personalization is more personal than behavior targeting, but behavior results should fall under this category.
- Performance: Ideally, every single recommendation made should be performance based, however, this category, for me, contains non-contextual, non-personalized recommendations. Popular among DVD & music stores is Movers & Shakers, Bill Board top 100 music charts, Oscar Winners). “Most Viewed Video Today”, notice there can be some restrictions to Performance based system–in this case “Today”, but it should be minimal and easily comprehended. Also, secretly, these could be high converters, although you’d never say that to your customer…
- Referral: This is a very cool area of recommendation. This is your typical eHarmony match making (user->user). There is also (site->user), (blog->blog aka blogroll); Referral service is not very popular because it is very hard to put a value on the referral.
- Discovery/Navigational: And finally, if your site lacks hierarchical/facet based browsing capability, and it does not have a good search engine, then recommendations can help to orchestrate the user experience by leading them into purchasing the product.
Btw, I got laid off from work doing this for a living, so this information may not be the best thing to repeat in your interview or your next paper, Caveat emptor. And if you’re a recommendations vendor or a retailer or a media content provider, Caveat venditor. [[And to say this explicitly, I am not certain of the value of this approach to making a universal recommender system. Writing about it may bring you bad luck. Using the system for commercial purposes may not generate the desired revenue]]
Interesting Future Work:
Explore versus Exploit: The precise meaning and nature of this style of recommendation is unclear to me. This must have something similar to active learning algorithms that selects some examples for exploration to improve it’s performance…
User Experience versus Time-on-site versus Spend-maximization versus cost-minimization versus total user value.
User retention versus sales volume: If you sell lots of users a bad product, it can be made so that they don’t return the product, but they may never come back again.
AOV versus revenue: This is a hard call. I think I’d go for revenue, but there’re a lot of gotcha’s…
volume versus margin: This one is pretty complicated too.
CTR versus conversion rate: very difficult to determine what to use. It’s almost impossible to be sure form day to day and page to page and product to product and user to user which is more valuable.
Another thing not entailed in the above listing is the recently popular game theoretic approach to advertising. In the presence of many possible recommendations, Because the user are aware of the above five aspects of recommendation, they may second guess the value of the products being recommended based on the assumed underlying reason for those products being recommended. (More specifically, the user will think recommendations are being made to sell them something more expensive than something else that is of the same quality but lower price, etc.) This thought process obviously takes place inside most intelligent shoppers, but how do we leverage it so as to provide the best user experience?