Collaborative based Filtering in Recommender Systems & Content Based Recommender Systems

So, continuing on the?previous topic. We have multiple types of Recommender Systems. Among them let us now discuss about the following types.

They are :

• Collaborative based Filtering in Recommender Systems &
• Content Based Recommender Systems.

1. Collaborative Based Filtering in Recommender Systems:

Let us look this from the practical stand point. Imagine a user who liked movies U1 = M1, M2 & M3 and some other user liked U2 = M1, M3 & M4. And now there is this third user who likes the movie M1 and we need to recommend him movies and this should be likely to be in his watch list and this is our job.

U1 = M1, M2 & M3

U2 = M1, M3 & M4

U3 = M1 & ……?

Now, from the above data, which is a data matrix, of the users we need make use of this matrix data which is A, in which we have the ratings for the users, Ui on Movies Mj.

• So, from the data that is available or given to us, U1= M1, M2 & M3 ; U2 = M1, M3 & M4 ; U3 = M1 & ……?, we can see that there is a one mutually liked movie M1 among the three users when compared. So, the task is to find the next movie which is likely to be watched by U3.

Coming to the comparison we can see that M1 is rated and liked by the three users. And we can also see that M3 is rated and liked by the U1 & U2. Now, there is a high chance that this can be rated by the U3 excluding the other interest of the U1 & U2.

And this is the core idea and flow of logic, which is, the users who agreed in the past tend to also agree in the future.

***Core-idea: Users who agreed in the past tend to also agree in the future (Assumption).

Here, in the past U1 liked and rated M1 which is also liked and rated by U2 & U3 later which means that these users agreed this movie M1 is great and worthwhile in the past. Hence in the future user U3 may like M3 which is something that U1 & U2 agreed that as a great movie in the future.

In Simple words, We need to believe this assumption as the core idea if not we cannot build a collaborative based filtering in recommender Systems.

The above is a little focus on Collaborative based filtering in this article and now we will get to see content based which is similar to the Classification/regression techniques. Content based recommender systems fits into the framework of Classification problem.

2. Content Based Recommender Systems:

The algorithm strictly says that it does not use the rating information ‘Aij’ info anymore as the previous one. On the other hand, what it needs is a way to represent the information of item ‘Ij’ and the user ‘Ui’.

Now, let us focus on the example a practical stand point, imagine we have opted an item ‘Ij’ which might be an adventurous film.

Features here are Item Ij/Mj (Movie) : action, drama, 90’s hits, action-theme, beach-theme, actors & directorial specials.

So, if we have this representation of an item ‘Ij’ (or) Movie ‘Mj’ which is as described and similarly the user ‘Ui’ that this guy likes action (or) adventure-themed movies and the person’s gender or any other related information.

Features here: Likes Adventure-themed movies, Action movies, Male, San Fransisco , aged 25.

• Now consider the above as features or imagine them as the features for our given problem and we need to somehow represent them. In the above we have features related to both User ‘Ui’ & Item/Movie ‘Ij’.
• So, what we can say is that in the content based recommender systems we do not use ‘Aij’ info which is user ‘Ui’ on an item ‘Ij’ and rather we use other meta information. And this is the reason we call it as Content-Based Recommender Systems. So here we are seeing what type of content is there in Item ‘Ij’, User ‘Ui’ likes.

• And the same that is collaborative filtering based recommender systems. Because it is the collaboration based, we are using the data from the collaboration between all of the users to share their data. And if the user ‘Ui’ does not share the data we will not have all this matrix data. We came up with this information because of collaboration from all these users shared.

So, in the content based we need data like earlier. As soon as we have the data like this we can represent an Item ‘Ij’ as a vector where the first component of vector could be ‘is an adventure-themed movie’.

Once we get a feature representation as we have seen above we can pose this as a Classification/regression problem.

If we don’t have the representation, the only way we have is Collaborative Filtering based Recommender Systems.

And if we have this data in addition to ‘Aij’. (It does not use ‘Aij’ as actively as collaborative based using).

*****If we exclude the part of feature engineering, content based is similar to Classification/regression problem.*****