train (on local machine)

• run from train.py

  make train
  

• (recommend) run from notebook as /research_phase/train.ipynb

serve

• on local machine

make serve

• on docker

make docker_run_serve

• after start serve (local or docker)
  • endpoint is http://localhost:8080/recommendations
  • api spec is on http://localhost:8080/redoc and http://localhost:8080/docs

test

make test_serve

• user based collabortation filtering with SVD

• main.py
  • this module is wrap recommender and serve as RestAPI by FastAPI
• modules/db.py
  • this module is use for query movielen data
  • now it just load movielen data as dataframe
• modules/feature_store.py
  • this module is use for query feature to use for model prediction
  • now return features are "histories" (watched movie id) and "unwatched_movie_ids"
  • now it just query from DB, but in future if model is more complex we can get benefit from this module
• modules/recommender.py
  • this module is wrap model to use for recommend data from user_id
  • given user_id and use SVD model to predict rating of unwatch movies
  • then get top 10 movie and return to user

• clean data
  • use pandasprofiling to see static of data
  • merge movie in movie_df that have duplicate title
  • remove rating data that have rating outside range[0, 5] and user_id is Nan
• train model
  • user surprise lib to train SVD model
  • get RMSE of algorithm on train test data split
  • get Hitrate (precision) of this algorithm on leave one out data split
  • get final model by use all data as trainset (to memory it)
  • save artifacts

• after serve access this url http://localhost:8080/docs
• you can use interaction web UI to try call API

• you can call API follow specification by your prefer tools ex. curl, python requests, postman