TF-IDF model was used to create a recommender. It was the only option available from Microsoft recommenders that dealt with text data. Recommenders usually work on a user-product-rating framework and our dataset is not build like that. It has users and the tags they use.

Maybe, some further investigation can show that other models can be used in this case.

TF-IDF is a standard model used for text data. It basically tokenizes all text, in a standard format and builds the frequency of each token. Building from that it calculates cosine similarity of the vector tokens of a user with another one and recommends accordingly.

Cosine similarity is a distance metric used for vector tokens because of its nature of high-dimensionality. Euclidean distance does not work in this cases and basically cosine distance measures the angle between different vector of tokens.

TF-IDF does not scale well so I took a small sample of 500 users to build a model. In my machine it crashed when I tried to do all users at once.

• delete model in /model/ and it will create a new model for the new input you have provided

• docker build --tag microsoft-recommender .

• docker run -p 3001:3001 --name backend microsoft-recommender