What algorithms should I choose for a recommender system and why?

Question

To my knowledge, recommender systems are broadly classified into collaborative and content. Collaborative in turn is divided into 1) Memory (uses similarity metrics) and 2) Model (well known Matrix/Tensor factorization). Content based involves constructing a user profile and then an algorithm like SVM to classify and recommend items. Now here are my questions:

1. What other algorithms can I use for recommendation and why?
2. Can I use neural networks? (understanding them has been a bit difficult for me)
3. Is it true that neural networks (NN) are only suited for text and image processing and numerical data doesn't need complex algorithms like NN?

Answer 1

Short answer:

It depends on your data. What do you want to do ?

Longer answer:

1. Use content based approaches if you have data on your items. Use collaborative approaches is you have data on your users. Use both if you have both. I would say content based approaches are general machine learning problems (how do I extract meaningful information from data) whereas collaborative filtering is really a recommender system specific work (how user's behavior can suggest users/items similarity/connections).

2. Well you can. Neural nets are just a kind of algorithm, you surely can use them for content based analysis, and it might be possible to use them to enhance your collaborative algorithm.

3. NN use texts and images as numerical data, so I don't understand your question.

If you want a good insight in today's recommender systems, take a look at this article.