ml-recsys-tools 0.5.3

Tools for recommendation systems development

ml-recsys-tools

Open source repo for various tools for recommender systems development work. Work in progress.

Main purpose is to provide a single wrapper for various recommender packages to train, tune, evaluate and get data in and recommendations / similarities out.

Installation:

Pip: includes only basic dependecies + lightfm/implicit for now: pip install ml_recsys_tools

Docker (build and run interactively):

docker build -t ml_recsys_tools:local .
docker run -it --rm ml_recsys_tools:local python

Basic example:

# dataset: download and prepare dataframes
from ml_recsys_tools.datasets.prep_movielense_data import get_and_prep_data
rating_csv_path, users_csv_path, movies_csv_path = get_and_prep_data()

# read the interactions dataframe and create a data handler object and  split to train and test
import pandas as pd

ratings_df = pd.read_csv(rating_csv_path)
from ml_recsys_tools.data_handlers.interaction_handlers_base import ObservationsDF    
obs = ObservationsDF(ratings_df, uid_col='userid', iid_col='itemid')
train_obs, test_obs = obs.split_train_test(ratio=0.2)

# train and test LightFM recommender
from ml_recsys_tools.recommenders.lightfm_recommender import LightFMRecommender    
lfm_rec = LightFMRecommender()
lfm_rec.fit(train_obs, epochs=10)

# print summary evaluation report:
print(lfm_rec.eval_on_test_by_ranking(test_obs.df_obs, prefix='lfm ', n_rec=100))

# get all recommendations and print a sample (training interactions are filtered out by default)
recs = lfm_rec.get_recommendations(lfm_rec.all_users, n_rec=5)
print(recs.sample(5))

# get all similarities and print a sample
simils = lfm_rec.get_similar_items(lfm_rec.all_items, n_simil=5)
print(simils.sample(10))

Recommender models and tools:

• LightFM package based recommender.

• Spotlight package based implicit recommender.

• Implicit package based ALS recommender.

• Serving / Tuning / Evaluation features added for most recommenders:

  • Dataframes for all inputs and outputs
    • adding external features (for LightFM hybrid mode)
    • early stopping fit (for iterative models: LightFM, ALS, Spotlight)
    • hyperparameter search
    • fast batched methods for:
      • user recommendation sampling
      • similar items samplilng with different similarity measures
      • similar users sampling
      • evaluation by sampling and ranking

• Additional recommender models:

  • Similarity based:

    • cooccurence (items, users)
    • generic similarity based (can be used with external features)

• Ensembles:

  • subdivision based (multiple recommenders each on subset of data - e.g. geographical region):
    • geo based: simple grid, equidense grid, geo clustering
    • LightFM and cooccurrence based
  • combination based - combining recommendations from multiple recommenders
  • similarity combination based - similarity based recommender on similarities from multiple recommenders
  • cascade ensemble

• Interaction dataframe and sparse matrix handlers / builders:

  • sampling, data splitting,
  • external features matrix creation (additional item features), with feature engineering: binning / one*hot encoding (via pandas_sklearn)
  • evaluation and ranking helpers
  • handlers for observations coupled with external features and features with geo coordinates
  • mappers for geo features, observations, recommendations, similarities etc.

• Evaluation utils:

  • score reports on lightfm metrics (AUC, precision, recall, reciprocal)
  • n-DCG, and n-MRR metrics, n-precision / recall
  • references: best possible ranking and chance ranking

• Utilities:

  • hyperparameter tuning utils (by skopt)
  • similarity calculation helpers (similarities, dot, top N, top N on sparse)
  • parallelism utils
  • sklearn transformer extenstions (for feature engineering)
  • google maps util for displaying geographical data
  • logging, debug printouts decorators and other isntrumentation and inspection tools
  • pandas utils
  • data helpers: redis, s3
  • basic rankings server that updates models from S3

• Examples:

  • a basic example on movielens 1M demonstrating:
    • basic data ingestion without any item/user features
    • LightFM recommender: fit, evaluation, early stopping, hyper-param search, recommendations, similarities
    • Cooccurrence recommender
    • Two combination ensembles (Ranks and Simils)

• Still to add:

  • add and reorganize examples
  • much more comments and docstrings
  • more tests