What occupies our geeky minds

From Good To Great Recommendations

Short URL 0 Comments

Frozen Brainstorm by Steve Jurvetson

In 2006 Netflix, the US online movie rental company, launched a competition to see who could improve their recommendation engine for movies by at least 10%. They offered a prize of $1 million, and within days they’d already had entries with improvements.

The competition lasted for some years, and I remember reading a fantastic article in Wired magazine titled “This Psychologist Might Outsmart the Math Brains Competing for the Netflix Prize.”

The crux of the article is about one of the entrants, Gavin Potter, who had an undergraduate degree in psychology and a master’s in operations research rather than computers or maths, who intended to use human psychology rather than maths to improve the algorithm.

Anyone who’s done a little sales will know that it involves understanding the customer, and so Potter was obviously on the right track.  A great salesperson builds a relationship with their customers, understands their desires, and tries to solve their problems. How in the world, can a computer do that. The answer is, it’s not simple, but with today’s technology, it’s a lot more feasible.

Part of Potter’s solution included taking into account the psychology of the ratings of previous movies. “If a customer watches three movies in a row that merit four stars — say, the Star Wars trilogy — and then sees one that’s a bit better — say, Blade Runner — they’ll likely give the last movie five stars. But if they started the week with one-star stinkers like the Star Wars prequels, Blade Runner might get only a 4 or even a 3.” Knowing a little human psychology meant Potter could improve his algorithm quickly.

So where does that leave companies who want to implement a recommendation engine, who don’t have a million dollar prize offering, or a degree in psychology. It took 3 years for the Netflix prize to be won, but they started with a simple approach and built upon it. We’ve been involved in the recommendation space for 5 years and have found that a phased approach works best, starting from something simple and working up to a very tailored solution.

In this time we’ve found that every client’s recommendation needs are different. Like Netflix, a company can implement a simple recommendation algorithm and improve it over time. We usually suggest a phased approach for adding recommendations to any companies web or mobile site; start with a simple algorithm that provides recommendations based on matching content, then add extra information based on what clients have purchased or “liked,” and then tweak the algorithm to include other smarter ways of determining matches. We do all of this by working closely with clients to ensure the psychological smarts that they understand are included in the end result.

To highlight the approach it’s worth looking a few of the Internet’s popular recommendation systems. The Echo Nest provides a music recommendation system, and to provide the results they not only “listen” to each track to determine characteristics including tempo, song structure, timbre, they also trawl the web to analyse conversations about  artists, albums and songs. Combining these results makes a much smarter recommendation.

Amazon’s approach to recommendations is also complex, with a lot of secret sauce. Part of it’s trick is to use collaborative filtering; looking at what each customer rates or buys, and recommending the results to similar people. However, they tweak this with a range of tricks, like reducing the importance of very popular items that results in making less well-known items much more relevant. They use this to “trick” the recommendation system into not just showing everyone the most popular books, but ones that are more obscure but just as likely to be relevant.

These tweaks take years to perfect. That’s why we like to work with clients in partnership to perfect the recommendations. It might be we start the process with very little information, creating recommendations based on keywords. Then we aim to collect information about what the companies clients love, and then further tweak these results — just like The Echo Nest and Amazon — based on how clients make decisions around the companies products. Over time we end up with highly relevant recommendations specific to the product our client is selling.

It took 3 years and $1 million to improve their results by 10.9%. Many might question the value of the result, but when a good algorithm can increase sales or click-throughs by more than 30%, it seems to make a lot of sense. This is exactly why we offer a couple of services for clients where we analyse the business and data, and design a path to go from a cold start through to a tailored recommendation system. That way you don’t have to wait 3 years or spend a million to get an idea of what you need to do.

Potter didn’t win the Netflix Prize. But surprisingly most of the competitors collaborated in some way, sharing their findings to help each other improve their results. The winner was a combined team (from several of the teams) known as BellKor’s Pragmatic Chaos, which improved the Netflix’s original by 10.09%.

 

Posted by Richard in Recommendations