0
$\begingroup$

As a beginner in ml, I am watching a video on YouTube about designing a model of song recommendations on Spotify. For each user, there to be predictions about which songs to be recommended.

One of the features was number of historical engagements, such as views/likes.

The point I don't understand is when the interviewer asks what happens if some user is just clicking more than others in general because they are just more active in general they are clickers who click a lot. and explains that this is why we would need to normalize the clicks (for example, by dividing by some total of a user's activity).

If a model is predicting per group as here, then the prediction is for songs per user. So do you really have to normalize? What is the difference if User A has 900 clicks for Song A and 100 clicks for Song B, while User B has 20 clicks for Song A, 5 clicks for Song B, and 5 clicks for Song C. Would it matter? At the end of the day, we are deciding between songs per user we don't care about different users.

And even if you were to normalize, you could simulate and get User A say 0.9 for Song A and 0.1 for Song B. But User B watched 3 songs, so maybe his distribution is 0.6, 0.2, and 0.2.

How does the normalization help you? You still can't compare 0.9 and 0.6. They are the best scores per group, but still are incomparable across groups. How would raw click numbers differ from ratios?

But per user it would work to get the greatest percentage just as in raw numbers.

So my point is would you need to normalize when predictions are per groups. Do you have to care about a comparison across the groups?

Improve this question
$\endgroup$
2
  • $\begingroup$ Normalization is undertaken to solve a specific problem in a specific context. A description of the problem to be solved and the context in which the problem arises is missing from this question, so it's hard to say why normalization might or might not matter. $\endgroup$ Commented Mar 7, 2025 at 23:23
  • $\begingroup$ @Sycorax I attached the video for reference. Basically it's a recommendation system for songs to users. They worried about one user being more active in general and theirfore having bigger numbers of clicks per the specific Song recommendation in the festure of historical engagement with a given song $\endgroup$ Commented Mar 7, 2025 at 23:34

1 Answer 1

Reset to default
0
$\begingroup$

By the context of your question and the typical way people use "normalization" in machine learning online, I assume you mean translating the variables to be bounded between the range $[0,1]$. If this is the case, then it should be obvious that this doesn't do anything with respect to changing the distribution given it is a linear transformation. So it is not a necessary procedure and wouldn't fix that issue.

The benefit of normalizing in this context is that it puts the values on a roughly similar scale, but any other utility that may serve would depend on the actual purpose of the model.

Improve this answer
$\endgroup$

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.