Recommender systems normalization incase of new user

Course Q&A Machine Learning Specialization Unsupervised Learning, Recommenders, Reinforcement
1

Hi,

In the recommender systems lab, it is shown that when a new user is added to the system, and this new user gave some preferences in the form of ratings, then this new user’s ratings were added to the original ratings. After that, normalization of the entire ratings was performed using mean normalization.

My question is, how it works in a real-world scenario?

So, whenever a new user is added to the system, then every time, the system loads the entire ratings again and performs normalization for every new user being added to the system?

Can someone shed light on this, please?

Best Regards,
Bhavana Malla

2

I believe that procedure isn’t intended to show you how to treat a new user, because usually a new user doesn’t have any rating.

I think you are quite concerned about how to deal with new user, but I am afraid you can’t look at that procedure from that angle.

Let’s assume a new user doesn’t have any rating. If we only have a rating-based CF recommender, then there is really nothing much we can do except for the Mean normalization thing that you have explained in your other post.

However, I think in a real-world business, they don’t rely on just a rating-based CF recommender, instead, I believe they should have a hybrid system that takes both rating and other information into account. “Other information” may include your location (based on IP), your browser, your local time, sometimes language, and any other information it can get. Certainly, those information won’t be used to train a rating-based CF, but can be used to train a separate Content-based recommender (second part of Course 3 Week 2).

Search “youtube recommendation algorithm” and you should find many interesting stuff, or replace Youtube with other platform :wink:

Cheers,
Raymond

PS: if you find some interesting stuff from the search, feel free to share in the MLS Resources Section if you would like to :wink:

1 Like
3

You have mentioned an aspect of collaborative-based recommender systems that also made me feel like WHAT???

Yes: In a Collaborative-based recommender system, if you have a new user, you basically have to retrain your model to include that user in the next predictions. That is, if you want to include that user in the collaborative-based recommender system.

Every time you have a new user you face, with that user, the Cold Start problem, specially if you don’t have any information about that user.

Now, there are “work-arounds”. For example:

Example 1. You gather some information about the user on on-boarding, like Location, Age, Gender, and may be you directly ask the user for some of their preferences. In this case, you can leverage on the profiles created with your current users, and based on that produce recommendations to this new user, by using the current model. At some point, for example every night, or every week, you’ll want to re-train your model and include all your new users.

Example 2: For your new user(s), where you don’t have any information about them to begin with, your can start by using a Content-Based recommender system, which would be based on your items (products, movies, songs, or any other type of item). You start with recommendations from your content based system, and as the user rates (or sees, or listens to, or selects) the recommendations, you start learning about this new user’s preferences. At some point, you’ll have enough information about this new user to add them in the next retraining of your Collaborative-Based Recommender System.

Hope this sheds light on this topic!

1 Like
4

Makes sense after watching Content based recommender! Yeah sure… will do that… Thanks

5

Yes now that I have watched Content-based recommenders, I got how leveraging these extra features about users and movies helps to mitigate the cold start problem up to to some extent. So, in practice, it is both Collaborative Filtering and Content-based Filtering together will give recommendations to the users. Everything makes sense. Thank you!