Recommender System

Classification for a recommender system

There are two types: 1. Using Meta Data - where we know someone likes something from a survey or other data - and we have grouped the choices to present them to the user 2. Collaborative Filtering - Find other people like you (rated the same thing in similar ways) and recommend the same. Asking for behavior -- is worse then looking for behavior

Embeddings

1. Embeddings are simply word vectors - which is nothing but matrix multiplication of one hot encoded vector with Embedding matrix. Thus simply it is a look up of the embedding entry for a particular index. With Word2Vec and Glove - they are simple embeddings which can be used for a dataset - which are precalculated.
2. Using the above idea, we can convert the text into embeddings - i.e.
   • Index all the unique words
   • create embeddings (with random data) or we can use Glove vectors for embeddings
   • for the given text - simply convert by a lookup into Embedding matrix
3. How many latent factors - depends on architecture decision - look at Word2Vec paper, experiment with sizes.
4. Error range - for good model - model specific

DataSet :

1. Movie Lens Benchmark