Saturday, 4 August 2018

Recommendation System II


Recommendation System

We have always seen the famous product sites recommending products to users. There must be some set of rules which drives them to right product, and if they recommend right product, probability of acquiring or using the product increases. It works best with Netflix, Amazon, YouTube and so on.

Every business is ultimately trying to maximize their revenue or traffic and lead user to the things which they may not search but are available. Firms use their own ‘set of rules’ (tweaked algorithms) to identify what works best based on their data, in following article we will try to understand one of such technique, how it works and how we can use this in our system.

In data Science there are broadly 2 ways recommendation system works are we are going to explain with People and Name of Movies they watch, and how does the recommendation works for both of those:

  • Item/Content based recommendation system.
  • User based recommendation System

       Item/Content based recommendation system

Content based recommendation system use their knowledge to predict the user rating for that product based on the history or the similar products user have liked, recommending the top one to user, after figuring out that this is the user will like the most.

Lets start to dig down the problem with a very simple example, now we have 4 users who have seen one movie and rated them specific, we want to recommend the next movie but it should be relevant.



In content based recommendation system , we create another table, which provides the movies details with certain attributes. We want to map these attributes to the previous table and if someone like a movie on one genre he will have better probability of liking the movie in same genre, we can have so many of such attributes.



For above set, we know for Brian 'Chak De' will be next movie to recommend because he gave good rating to a sports movie and bad to a drama one. Similarly, we can easily figure out in this smaller set what to recommend, Imagine data base with 1m users and 100k movies with 200 attributes.

Key Takeaways:

Good for new users, always have some recommendation.
Table II in above pic is difficult to maintain consistently, some people will say 'Wanted' has less drama than 3 idiots and we give them equal so recommendation may look like a generalization.
Also several attributes to consider about movie, requires a lot of work on product set.

      User Based Recommendation Systems

In this case system does not take considerations about product at all , it just takes into light that how other person has attributed/rated the product




To recommend the movie to Scott (highlighted column) , we will find the people who have rated similarly to Scott in past and who are opposite to taste of Scott , quite clearly he shares rating traits with Brian and opposite to Raj , he will rate similar to Brian this movie.

Key Takeaways:

No need to maintain product feature inventory, a lot of time and space saved.
This can’t work until we have some reviews about products, can’t recommend the product straightaway to new user.
Most likely favors product with more reviews rather than one with less reviews.

Next articles with have ways to implement content based recommendations.


3 comments: