crab：
yum install python-dev python-numpy python-numpy-dev python-setuptools python-numpy-dev python-scipy libatlas-dev g++
pip install -U scikits.learn -i https://pypi.tuna.tsinghua.edu.cn/simple/
git clone https://github.com/muricoca/crab.git
python setup.py install
Model: user item preference matrix (sparse representation, no semantic information, you can consider using word2vec)
uid:{itemid:preference}
Similarity: Collaborative filtering based on user or item
recommender: filtering method, such as knn
Basically an unreliable implementation (too simple)
word2vec-recommender：
A user is described by the context of the item, so that the vocabulary is a collection of items, and the doc is the representation of the user. After training with word2vec, semantic similarity can be expressed.
word2vec: The essence is to reduce the dimensionality of the word, together with the context to which it belongs, so that the dimensionality-reduced vector not only stores the data in multiple dimensions of the vector, but also expresses the semantic similarity of the word. .
Recommendation engine: The essence is to use the similarity of items or user preferences to make appropriate recommendations for users.
The application of word2vec in recommendation is envisaged: it is necessary to find the counterpart of the word, as well as a collection of semantic words that conform to the facts (corpus).
Example: 1. item embedding:
1" to record a collection of items browsed by a user every day,
2" process this collection as a statement
3" After a period of time, train for all sentences of a single user
4" The similarity of items for a user is reflected in the similarity of word2vec.
The conditions for the establishment of the above assumptions: a person's preferences are stable in a certain period of time; the user has preferences but may not even be able to tell himself; the user's browsing truly expresses his preferences, that is, what he chooses to watch is himself like.
2、user embeding
1" When an item is on sale, record the user sequence purchased within one day
2" Count the purchase user sequence of all items
3" Recommend items through their vector similarity
3. After there are many user choices, new users can be clustered