pyrecmd

Pure python implementation of collaborative filtering algorithm, one of popular methods when to build recommendation system

Installation

1. Download python package file: pyrecmd-0.1.0.tar.gz
2. pip install pyrecmd-0.1.0.tar.gz

Example

import pyrecmd
import os

# assume that you already have the movie-lens 100k dataset
with open(os.path.join("ml-100k","u.data"),"r") as fin:
    ratings = np.asarray([ [int(token) for token in line.split("\t")] for line in fin])
    ratings[:,[0,1]] = ratings[:,[1,0]] # switch user and item columns

with open(os.path.join("ml-100k","u.item"),"r",encoding="latin1") as fin:
    itemnames = [ line.split("|")[1] for line in fin]
    itemnames.insert(0, None)

# initialize with 100 features and ratings each record in form of (item, user, rating)
# parameter 'itemnames' and 'usernames' are optional and required to output more readable names
movierecmd = pyrecmd.PyRecmd(100, ratings, itemnames=itemnames)

# parameter 'regularization' is optional but it's worth tuning this parameter for better validation error
# parameter 'learningrate' is optional but it's worth tuning this parameter to shorten learning time
# parameter 'validationsplit' is optional and usually needed when you're tyring to tune learning parameters like 'regularization' and 'learningrate'
movierecmd.fit(regularization=0.8, iterations=10000, learningrate=3e-4, validationsplit=0.1, verbose=True)

  # learning-rate=0.0003, regularization-lambda=0.8, iterations=10000, validationsplit=0.1
  # 90700 for train, 9300 for test, total 100000 ratings
  # iteration #1: loss=0.04257519975299331, valloss=0.04123926882097799
  #iteration #1000: loss=0.003767156102278313, valloss=0.02942087829419617
  #iteration #2000: loss=0.0019754647266246525, valloss=0.029991390881000023
  #iteration #3000: loss=0.0017366852642300109, valloss=0.029905459133939213
  #iteration #4000: loss=0.001660970523927064, valloss=0.029934834364596683
  #iteration #5000: loss=0.0016250544223948558, valloss=0.03011373533301576
  #iteration #6000: loss=0.0016085053613471238, valloss=0.030277677577385687
  #iteration #7000: loss=0.001600799510030617, valloss=0.030360609212595906
  #iteration #8000: loss=0.0015958981491327761, valloss=0.03042506479935819
  #iteration #9000: loss=0.0015925062894997233, valloss=0.030413337290199106
  #iteration #10000: loss=0.0015910422433985266, valloss=0.030346999656762637

# predict all the ratings given users 1,2,3,4,5
userresult = movierecmd.predict(movierecmd.USER,[1,2,3,4,5])

  #{1: [('Toy Story (1995)', 4.85, 5.0),
  #    ('GoldenEye (1995)', 3.02, 3.0),
  #    ('Four Rooms (1995)', 3.92, 4.0),
  #    ('Get Shorty (1995)', 4.16, 3.0),
  #    ('Copycat (1995)', 2.83, 3.0),
  #         ...
  #2: [('Toy Story (1995)', 4.03, 4.0),
  #   ('GoldenEye (1995)', 2.78, None),
  #   ('Four Rooms (1995)', 2.32, None),
  #   ('Get Shorty (1995)', 2.95, None),
  #   ('Copycat (1995)', 3.08, None),
  # ...

sample_ratings = [(r[0],r[1]) for r in ratings if r[1] == 5]
# instant_fit learns a single user or item instantly and return prediction results given the sample
instant_fit_result = movierecmd.instant_fit(movierecmd.USER, sample_ratings, regularization=0.8, iterations=5000, learningrate=3e-4, verbose=True)

  #iteration #1: loss=0.16384997286426806
  #iteration #500: loss=0.04543407281702286
  #iteration #1000: loss=0.0346454254970892
  #iteration #1500: loss=0.02911853749379627
  #iteration #2000: loss=0.026296828685820364
  #iteration #2500: loss=0.02471713670331908
  #iteration #3000: loss=0.023771358973847783
  #iteration #3500: loss=0.023201403212681094
  #iteration #4000: loss=0.02287483390299474
  #iteration #4500: loss=0.02270596322264621
  #iteration #5000: loss=0.02263077845384043

print(instant_fit_result)

  #{'prediction': [('Toy Story (1995)', 5, 5.0),
  #                ('GoldenEye (1995)', 4.99, 5.0),
  #                ('Four Rooms (1995)', 5, None),
  #                ('Get Shorty (1995)', 5, None),
  #                ('Copycat (1995)', 5, None),
  # ...
  # 'weight': array([-0.0012171 ,  0.49456538,  0.4138671 ,  0.02527371,  0.22232359,
  # ...

Note

The input "Ratings" is a list of tuples, each has (item, user, rating)

Recommended steps to follow

1. (PyRecmd() construct) Initialize PyRecmd object with rating information
2. (PyRecmd.fit()) Try several times varying with 'regularization' and 'learningrate' parameters also given 'validationsplit' parameter, 0.1 for an example
   • 'validationsplit > 0' splits input ratings into train and test set
   • 'validationsplit > 0' outputs validation loss as well, which is important to check the model is overfitted
   • 'regularization' is weight regularization and validation loss is quite depedent to this parameter
   • 'learningrate' plays a role in convergence speed as usual machine learning method does
3. (PyRecmd.fit()) Given the best 'learningrate' and 'regularization' parameters found, learn again the whole dataset without 'validationsplit' in order to materialize the final weights
4. (PyRecmd.predict()) Predict all the ratings given users or items
5. (PyRecmd.instant_fit) Fit any external user or item not learned yet to the existing model and returns prediction results

Limitation

• Too slow to learn a large dataset in a feasible time which size is more than 100K

Future items(if they are in need)

• Speed up by OpenCL
• Incremental model building

Reference

• Wikipedia - Collaborative Filtering
• Coursera - Machine Learning - Recommender System