Get startedGet started for free

Uses for recommendation engines

1. Uses for recommendation engines

So far, we've only considered recommendations as the use case for the ALS algorithm, but there are other applications that are also useful. These include latent feature discovery, item grouping, dimensionality reduction and image compression. In this course we'll only talk about some of these. First let's talk about latent features. As mentioned earlier, people will go to great lengths to effectively categorize items. But some products span various categories making them difficult to organize. Movies are often like this. Horror movies can be comedies. Dramas can be satires. Documentaries can be romances or even mysteries. Because of this, they can sometimes be difficult to market. If we had a better understanding of how consumers categorize movies based on their experience watching them, we could add more power to marketing strategies. ALS can help with this.

2. Basic factorization of ALS

When we have a matrix that contains users and movie ratings, ALS

3. Basic factorization of ALS (cont.)

will factor that matrix into two matrices, one containing

4. User matrix

user information and the other containing

5. Product matrix

product information, or in this case, movie information. Each matrix takes the respective

6. Factor matrix axes

labeled axis from the original matrix, and is given another axis that is unlabeled. The unlabeled axes contain what’s called

7. Latent feature axes

latent features. The number of latent features is referred to as the "rank" of these matrices. In this case, the rank chosen is 3. You, as a data scientist get to choose how many of these ALS will create. These latent features represent groups that are created from patterns in the original ratings matrix and the values in these columns represent how much each item falls into these groups. For example,

8. Horror vs drama

in the original ratings matrix, there might be a lot of people who like horror movies and don’t like dramas. They would rate

9. Horror vs Drama Part II

horror movies high

10. Horror vs Drama Part III

and dramas low. Likewise other people might like dramas and not like horror movies, and would rate

11. Horror vs Drama Part IV

dramas high

12. Horror vs Romance Part V

and horror films low. ALS can see this and determine that these are different types of movies.

13. Horror vs Drama Part VI

And if we were to look at the movie factor matrix, we would likely see that in one of the latent feature rows the dramas would score high, and the horror movies would score low, while in another latent feature column we might see the

14. Horror vs Drama Part VII

opposite. Knowing a little about these movies, we could determine that those latent features reflect those two genres. This allows us to mathematically see how users experience these movies and to what degree users feel each movie falls into each respective category. This concept goes a bit deeper though. For example, if we were to look at a

15. Uncovering Features

movie matrix, we might see in one latent-feature column,

16. Uncovering Features Part II

that several movies have scored very high, but they

17. Uncovering Features Part III

don’t seem to have anything in common. If they are all popular movies,

18. Uncovering Features Part IV

we might want to research what’s going on here to see if there's a business opportunity. Digging deeper,

19. Shakespeare Adaptations

we find that these movies are all adaptations of Shakespeare plays, and that there seems to be a strong customer group that likes these types of movies. Now that we know this, we can use this information to inform how we choose what movies to make and hopefully give our customers more of what they want. It’s worth reiterating that in the original data set there was no column anywhere called “Shakespeare Adaptations”. It’s also worth noting that many or all of these customers may not even know that this is something that draws them to these movies. This is the type of powerful information that ALS can help us uncover.

20. Let's practice!

Now lets try and actually use this.

Create Your Free Account

or

By continuing, you accept our Terms of Use, our Privacy Policy and that your data is stored in the USA.