Dataset

ml_netflix.csv: join on ml-latest-small and Netflix data

• code for generating the dataset in utils/gen_data.ipynb

Experimented on

1. LDA
   • pure Python
     • can run the notebook straight through
   • Stan
     • warning: takes ~1.67 hours to finish sampling on all 1056 movies
     • run python3 stan-lda.py in terminal with the following available flags:
       • regen_words_df (bool): True if you'd like to regenerate the dataframe mapping each word (ID) to a document/movie (ID)
         • saved to cache/words_df.csv
       • regen_data_lemmatized (bool): True if you'd like to regenerate the lemmatized movie descriptions
         • saved to cached/data_lemmatized.txt
       • num_movies (int): the first num_movies movies from the data set that you'd like to train on
         • by default, it's the number of movies in the data set (1056)
       • just_eval (bool): True if you'd like to just calculate the evaluation metrics. Assumes you already have the trained posterior values in results/theta.npy.
   • Pyro
     • can run the notebook straight through
     • modify number of topics, number of epochs run, etc. in cell 4
   • Turing
     • can run the notebook using Julia runtime
     • results are output to CSV (cache/julia_out.csv) for evaluation in Python using eval_julia.ipynb
2. PMF
   • pure Python
     • can run the notebook straight through
   • Stan
     • run python3 stan-pmf.py in terminal with the followiing available flags:
       • just_eval (bool): True if you'd like to just calculate the evaluation metrics. Assumes you already have the trained posterior values in results/Z.npy and results/W.npy.

Eval Metrics

1. Personalization
2. MAP@K
3. Mean Precision
4. RMSE