Why is deep learning used in recommender systems?

Question

I am currently reading a lot about recommender systems (RS) and came across that many RS are based on deep learning. However, I never find a good scientific article why deep learning is used in RS and why it is more successful compared to other methods.

Now my question is:

• Why is deep learning used in RS?

• Why would it be better to use deep learning in RS?

• How does deep learning work in RS?

I would be very happy if you could help me to answer my unanswered questions. I would be very happy if you would add a source (that is scientific, no medium etc.) that I can read more about it.

Answer 1

"Recommender Systems" is a very broad area and can be approached from different optics: latent variable models, graph models, etc.

"Deep learning" is an umbrella term for gradient-based optimization of deep differentiable models, and has been used to model all sorts of supervised learning problems, including graphs and latent variable models. Therefore, it has been applied to many different areas, with a great degree of success.

Multiple deep learning approaches have been applied to recommender systems because deep learning performed well in many other areas, so why not expect it to perform well for recommender systems.

Neural approaches have proven to be effective also for recommender systems. That is why they are used nowadays.

Also, different inductive biases have been applied to neural networks in other areas and, in some cases, such inductive biases are also useful for recommendation, like the sequential bias for recurrent networks, or the structured bias for graph neural networks.

If you want to explore the academic publications related to neural recommender systems, I would suggest that you start with the article Neural collaborative filtering and the survey Deep learning based recommender system: A survey and new perspectives.

Also, in order to get a broad view of the state of the art in recommender systems, you may have a look at papers with code, where you can see the best-performing systems for different benchmark datasets, with links to the associated articles and source code repositories.

Answer 2

Pure hype for neural-everything, supported by parties with an interest in such methods (e.g. hardware vendors and cloud providers).

For recommender systems this is partly to blame on badly design experiments in earlier research too (see e.g. Are We Really Making Much Progress? A Worrying Analysis of Recent Neural Recommendation Approaches or On the Difficulty of Evaluating Baselines), which made them look better than alternatives until evaluated more carefully.

Answer 3

I understood once that one reasons is because of the loss function.

Classical ML (non DL) can handle correctly point wise and pair wise ranking systems. But for listwise there are a lot of problems.

One of the advantages that DL methods have over non DL methods is teh loss function. With NN you can have a much more flexible loss function and you can tackle listwise problems.