This week we looked at Recommender Systems. We all see Recommender Systems everywhere on the Internet, the big ones being Amazon.com recommending “Other products you may like”, and Netflix recommending other movies to watch.
Recommender systems recommend things we should concentrate our attention on. Recommender systems have been around for ages, when we look at the articles in the newspaper, do we ever stop to wonder, why has the newspaper recommended this article for me? Well, obviously the newspaper can’t make a perfect newspaper for everybody personally, so it uses editorial recommendations, or hand selected recommendations.
However, now that we have the Internet, mass-recommendation is possible, and preferable in the age of The Long Tail.
The basic concepts of recommendation engines is:
- Users want recommendations
- Sites have a list of items for recommendation (movies, books, people)
- We take some inputs like ratings, demographics, and content data
- We output a prediction of what people might like, the top choice being the recommendation
We could recommend what people are likely to hate, but there’s not much point in that, so recommender engines are tuned towards positive recommendations.
There’s two types of recommender systems:
- Content-based, which looks at content (like movies), and recommends similar content to that user. So if you’ve watched a cowboy movie and rated it highly on Netflix, Netflix will recommend other cowboy movies that are similar (same actors, same time period, same genre etc).
- Collaborative filtering, which looks at customers or items, and says, if you like the same book as other users who purchased this book, then here are the items that those customers purchased as well. For example, if you purchased a camera, and other people who purchased that camera also bought a case, Amazon will recommend that case to you.
Sites that recommend similar to Netflix include:
- Hulu.com (popular TV shows)
- Pandora.com (music)
- Crunchyroll.com (Japanese Anime shows)
Sites that recommend differently to Netflix include:
- iPredict. iPredict recommends outcomes based on the concept of shares. A share pays out $1 if an outcome becomes true. Therefore, $1 = 100% likelihood that an outcome is true. So if a share is trading at $0.04, then there is a 4% likelihood according to the free market than that outcome is true. This is a bit like having 100,000 people bet on recommendations for what movie you would enjoy based on you liking Harry Potter, and then taking the average of their bets, and saying “that’s what the market thinks you’ll enjoy”.
We also were to look at item profiles for the various items, and describe attributes used to describe those items:
- Ant Man Movie – Year of production, age classification, length of movie, genre, release date, critic scores, director, writers, actors, award nominations,tropes contained, etc.
- A document (like the Wikipedia page on Recommender systems) – Links to other Wikipedia pages, categories of the page, Links to External pages, number of words in the article, bounce rate of people visiting the page who then left the site, number of people who arrive at Wikipedia at this article first.
Finally, we were to combine all the previous topics we discussed (search engines, advertisers, and recommenders) into a start up business model that delivered groceries to the home.
For me, the model would be shipping American candies and snacks automatically to people’s home, as a Snack Subscription.
First, explaining the business model – the value proposition is delivering delicious American candies and snacks regularly to people who enjoy and value them. The revenue model is Subscription based, where every month people exchange $30 to get a box of snacks delivered to them. The target customer is anyone in the world who values American snacks so much they’re willing to pay $30 a month every month ($360 a year) on American snacks. The distribution channel is via the web for the ordering service, and physical courier for the delivery service.
The second step would be to create a website and get that crawled by a search engine like Google. It would be important to know the keywords that customers are searching for using something like Google Trends for Snack Subscription. Google knows that people searching for Snack Box also want to search for Snack Subscription. We’d have to get links to our websites from other popular websites, to increase our rankings with the Pagerank algorithm.
Of course, it’s not always easy to get links on popular websites, so instead we can advertise instead. Advertising is all getting attention to our links, so we can take our above research for Snack Subscriptions, look at our competitors, and start bidding on keywords. To get better value for money, we can increase the information quality of our advertising by having a site optimised for mobile, including video, and describing our site in a format that the Google Knowledge Graph understands. This in turn will decrease the amount of money we have to pay per click, and increase the likelihood of our ad being displayed on the limited number of advertising slots available on a search query.
Finally, once we have customers arriving at our site, we can present a range of snack subscriptions, and give them the choice of ranking which types of snacks they prefer. These explicit rankings will provide a motivated customer with a more accurate selection of products they’d enjoy. But if they didn’t want to rank anything, we could provide options based on what other customers on the site prefer.