In [1]:
from __future__ import (absolute_import, division, print_function,
unicode_literals)
import pickle
import os
import pandas as pd
from surprise import SVD
from surprise import KNNBasic
from surprise import Dataset
from surprise import Reader
from surprise import dump
from surprise.accuracy import rmse
In [2]:
# We will train and test on the u1.base and u1.test files of the movielens-100k dataset.
# if you haven't already, you need to download the movielens-100k dataset
# You can do it manually, or by running:
#Dataset.load_builtin('ml-100k')
# Now, let's load the dataset
train_file = os.path.expanduser('~') + '/.surprise_data/ml-100k/ml-100k/u1.base'
test_file = os.path.expanduser('~') + '/.surprise_data/ml-100k/ml-100k/u1.test'
data = Dataset.load_from_folds([(train_file, test_file)], Reader('ml-100k'))
# We'll use the well-known SVD algorithm and a basic nearest neighbors approach.
algo_svd = SVD()
algo_knn = KNNBasic()
for trainset, testset in data.folds():
algo_svd.fit(trainset)
predictions_svd = algo_svd.test(testset)
algo_knn.fit(trainset)
predictions_knn = algo_knn.test(testset)
rmse(predictions_svd)
rmse(predictions_knn)
dump.dump('./dump_SVD', predictions_svd, algo_svd)
dump.dump('./dump_KNN', predictions_knn, algo_knn)
In [4]:
# The dumps have been saved and we can now use them whenever we want.
predictions_svd, algo_svd = dump.load('./dump_SVD')
predictions_knn, algo_knn = dump.load('./dump_KNN')
df_svd = pd.DataFrame(predictions_svd, columns=['uid', 'iid', 'rui', 'est', 'details'])
df_knn = pd.DataFrame(predictions_knn, columns=['uid', 'iid', 'rui', 'est', 'details'])
df_svd['err'] = abs(df_svd.est - df_svd.rui)
df_knn['err'] = abs(df_knn.est - df_knn.rui)
We now have two dataframes with the all the predictions for each algorithm. The cool thing is that, as both algorithm have been tested on the same testset, the indexes of the two dataframes are the same!
In [5]:
df_svd.head()
Out[5]:
In [6]:
df_knn.head()
Out[6]:
In [7]:
# Let's check how good are the KNN predictions when the SVD has a huge error:
df_knn[df_svd.err >= 3.5]
Out[7]:
In [8]:
# Well... Not much better.
# Now, let's look at the predictions of SVD on the 10 worst predictions for KNN
df_svd.iloc[df_knn.sort_values(by='err')[-10:].index]
Out[8]:
In [9]:
# How different are the predictions from both algorithms ?
# Let's count the number of predictions for each rating value
import matplotlib.pyplot as plt
import matplotlib
%matplotlib notebook
matplotlib.style.use('ggplot')
figure, (ax1, ax2) = plt.subplots(1, 2)
df_svd.est.plot(kind='hist', title='SVD', ax=ax1)
df_knn.est.plot(kind='hist', title='KNN', ax=ax2)
# As expected, one of the drawbacks of the NN algorithms is that their predictions are often
# quite concentrated around the mean. The SVD algorithm seems more confortable predicting extreme rating values.
Out[9]:
In [11]:
# Question: when a user has rated only a small number of items (less than 10), which algorithm
# gives the best predictions on average?
def get_Iu(uid):
"""Return the number of items rated by given user
Args:
uid: The raw id of the user.
Returns:
The number of items rated by the user.
"""
try:
return len(trainset.ur[trainset.to_inner_uid(uid)])
except ValueError: # user was not part of the trainset
return 0
df_knn['Iu'] = df_knn.uid.apply(get_Iu)
df_svd['Iu'] = df_svd.uid.apply(get_Iu)
df_knn[df_knn.Iu < 10].err.mean(), df_svd[df_svd.Iu < 10].err.mean()
Out[11]: