Tags
“Recommendation systems control majority of what you see but did not ask for on the Internet”
- When you search for a product on Amazon apart from that amazon suggests to you various other products.
- When you are on Linkedin browsing through updates from people you are connected with, at one section a person you may know is suggested to you.
- When you are logged into Netflix without any specific aim and search “good action movies” it suggests to a list of movies that you may be interested in.
Among these three scenarios you may see two broad distinctions. You may feel the particular scenario of Netflix may not be exactly in accordance with the opening statement – “Recommendation systems control majority of what you see but did not ask for …”, because you specifically asked for “good action movies” right? Wrong. A good action movie like everything that people typically desire are very subjective and is bound to have different perceptions to different users. An action movie that your friend liked may not be exactly what you have in mind. The action movies Iron Man and Saving Private Ryan are not exactly similar and suggesting one because another person watched another isn’t going to make him happy.
So the search results that appear are not exactly the consolidated best action movie, rather they are what the recommendation systems suggests you to watch among all action movies. You can consider this type to be like an on-demand setup and the former two types to be a passive setup where the recommendations are made continuously learning from each new user activity more like an advertisement.
This way across the web in every possible business recommendation systems engage with you and control what you see. These recommendations systems are used for a business purpose in each case. Typically these business purposes are to increase revenue or increase customer satisfaction and at times to drive down cost. We will concentrate on the majority two:
1. Increasing User Engagement for Revenue and Satisfaction
2. E-Commerce Systems – Purchase Recommendations for more Revenue
1. Increasing User Engagement
For our specific case and in relation to recommendation systems we define user engagement as the idea where our recommendation systems learn from an user and keeps suggesting items, both on-demand and passively, towards increasing the likelihood of him returning to our website and being more engaged in our website. These recommendation systems these days have become crucial and clearly distinguish online services.
We will see two industry pioneers of such a setup:
1.1 Netflix – Beyond the 5-stars
Netflix is an early pioneer of exploiting recommendation systems towards increasing user engagement and making them watch more movies through Netflix than any other platform. In 2006 Netflix announced the Netflix Prize, a machine learning and data mining competition for movie rating prediction. They offered $1 million to whoever improved the accuracy of their existing system, that recommended movies to user called Cinematch, by 10%. After a year, a team from AT&T Research Labs had won the first part of the progress prize.
What was the justification for $1 million in prize money?
Early on Netflix users had a DVD rental business model. So when users went through the list of movies available online and chose to rent a DVD, those titles weren’t readily available. So users typically would plan ahead and add titles to their queues. The selection was distant in time from actual viewing, so people would carefully select titles, and exchanging a DVD for another takes more than a day and meant no movie for the time period. Thus for a great customer satisfaction and making them come back and rent more from Netflix, it had to recommend the absolute best movie that the user would like. Its entire business model relied on customers being happy with selecting those movies that were recommended to them.
On the other side of the coin it had to predict which movie the user is most likely to enjoy based on his previous rentals and feedback. This was crucial because popular titles that a user wants weren’t always readily available. During those times if the user has no preference, and trusts the recommendations made to him he is more likely to rent a movie thereby bringing in revenue where there was none initially due to supply constraints ensuring customer engagement.
Since then much has changed thanks to faster Internet, we now have live streaming service and Netflix’s range of service delivery like integration with Xbox, IPhone Live Streaming, and Integration with Roku Player. This is also brought in real time data about users and their preferences.
Current Scenario
Today Netflix’s business is centered on providing personalized service. It starts at their homepage where a group of videos are arranged horizontally that a user is most likely to rate higher. They suggest titles in various genres that you may like. The collection of genre rows range from familiar high-level categories like Comedies and Dramas to highly tailored slices such as “Imaginative Time Travel Movies from the 1980s”[2].
While recommendation systems suggest content to you, they also run the risk of hiding content from you that you may have liked. Netflix approaches this with two broad principles in their recommendation system that addresses both their business needs and customer’s preferences ensuring continued engagement.
* It recommends a movie to you not because it suits their business needs but because it matches the information they have on you.
* It also keeps in mind the diversity that a user/household may have, and thus always suggests across genres to cater needs of different age groups in a household.
Having the second principle alongside the objective of increasing customer engagement is of course subjective to each business and depends on how they are setup and what segment they are serving.
For example in the case of an online Radio service like Pandora, which is very much individual user specific, when a user “dislikes” a song it deemphasizes certain attributes and refines station results corresponding to that attributes. It has less concern on running the risk of clouding users with narrow services. This is the case because unlike movies users will listen to a song they like more any number of times and will prefer a radio that plays that song which they like. Thus Netflix with limited supply of movies has to engage users by providing him with content he likes while occasionally exposing him to new contents he may like.
Recalling the previous blog we can see that emphasis is on the user in this case. Such a setup is called Collaborative Filtering. We will delve into technical specifics in coming weeks. Towards the emphasis on user they are vocal and explicit in communicating that Netflix adapt to their taste. Whether this helps is beyond the scope of this blog. They gather wide range of details about users like their connections on Facebook and use social proof to suggest users newer flavors.
1.2 Social Networks
Today social networking sites like Facebook and Twitter take away a significant portion of our time. And the companies wish to engage us even more. Their success lies in being able to take away the majority of our time and learning as much as possible about us and start being able to predict our likes, actions and dislikes.
What if this is all driven by recommendation systems? You can be quick in rebutting that we spend time on it not because they present us something we like, but only to connect with our friends in real world. This is partially true.
We would like to focus on two major social networking sites, Facebook and LinkedIn. The story of connecting with friends and family on a real time basis applies to Facebook and justifies the fact that we spend a lot of time there to an extent. On the other hand LinkedIn is very successful as a networking site even though majority of the people don’t look for jobs everyday.
We see two distinct ways how recommendation systems are exploited in each of the social networks.
* Facebook capitalized on our initial engagement, learnt about us and setup recommendation systems to market targeted ads by predicting the likelihood of us clicking an ad.
* LinkedIn on the other hand, early on recommended people to who we may be interested in connecting with and ensured our engagement with the website.
Online social networks, such as Facebook have access to a myriad of information. This can be used to recommend a wide variety of artifacts ranging from news articles, books, movies etc. This helps them gain more revenue as well. At the same time people are more likely to discover new things on social network and this will encourage them to engage more if recommendations are inline with their expectations.
For example their app center is becoming a major traffic channel for developers and 220 million people visit per month [3]. Using user data the end result is to learn user’s preferences to produce app recommendations that are socially relevant as well as timely.
In the case of LinkedIn they used recommendation systems to predict social graphs and suggested users. This made it easier for people to build their social network. Once users were connected LinkedIn suggests jobs.
2. E-Commerce Systems – Purchase Recommendations for more Revenue
This setup of recommendation systems is primarily motivated to increase direct revenue from users. In this case we mainly see the emphasis is put on products by trying to identify people who is more likely to buy this product. Recalling previous week this is an example of item based filtering. A prominent example for this is Amazon.
At Amazon.com, they use recommendation algorithms to personalize the online store for each customer. The online store changes radically based on customer interests, showing programming titles to a software engineers and baby toys to a new mother.The primary problem that they address using recommendation system is increasing sales through traditional cross-sell and up-sell techniques.
In coming weeks we will delve deeper into how these organizations actually do what we explained so far. Going deeper into the technical specifications
References:
[1]http://www.netflixprize.com/assets/ProgressPrize2007_KorBell.pdf
[2]http://techblog.netflix.com/2012/04/netflix-recommendations-beyond-5-stars.html