Recommender Systems using Latent-Factor Models
In this post, I describe an implementation of a recommender system based on latent-factor models and its application to the MovieLens 100K Dataset. The system is built within the Scikit-Learn framework to allow for using Scikit-Learn features such as Pipelines and GridSearchCV.
Import the relevant libraries
import collections
import numpy as np
import pandas as pd
from scipy.sparse import csr_matrix
from sklearn.base import BaseEstimator, TransformerMixin
from sklearn.pipeline import Pipeline
from sklearn.model_selection import GridSearchCV
from sklearn.model_selection import ParameterGrid
from sklearn.metrics import r2_score
from sklearn.metrics import mean_squared_error
import pickle
from time import time
import requests
import urllib
from IPython.display import Image, display
from IPython.core.display import HTML
import re
import matplotlib.pyplot as plt
%matplotlib inline
import seaborn as sns
sns.set(style='white', palette='Set2')
np.set_printoptions(precision=2, suppress=True)
np.random.seed(seed=0)
Load the data
Download the MovieLens 100K Dataset. This a classic dataset which is often used to test the performance of recommender systems. We will use files ‘movies.csv’ and ‘ratings.csv’ from this dataset.
"""!wget http://files.grouplens.org/datasets/movielens/ml-latest-small.zip
!unzip ml-latest-small.zip
!mv ml-latest-small/ ml-100k"""
'!wget http://files.grouplens.org/datasets/movielens/ml-latest-small.zip\n!unzip ml-latest-small.zip\n!mv ml-latest-small/ ml-100k'
# Read in the data into pandas dataframes
df_movies = pd.read_csv('ml-100k/movies.csv')
df_ratings = pd.read_csv('ml-100k/ratings.csv')
df_ratings.columns = ['userId', 'itemId', 'rating', 'timestamp']
df_ratings['timestamp'] = df_ratings['timestamp'].astype(str)
print (df_movies.head(2))
print (df_ratings.head(2))
movieId title genres
0 1 Toy Story (1995) Adventure|Animation|Children|Comedy|Fantasy
1 2 Jumanji (1995) Adventure|Children|Fantasy
userId itemId rating timestamp
0 1 31 2.5 1260759144
1 1 1029 3.0 1260759179
Split the dataset into train (80%) and test (20%) datasets.
First let’s explore the distribution of the number of movies versus number of ratings. From the following histogram, we see that many movies are rated only by less than 10 users.
plt.hist(df_ratings.groupby(['itemId'])['userId']
.count().values, bins=50);
plt.xlim(0,80)
plt.xlabel('Nummber of ratings')
plt.ylabel('Nummber of movies')
<matplotlib.text.Text at 0x7f752dc35828>
We partition our dataset into train and test sets by removing 10 randomly selected movies rated by each user from the original dataset and placing them in the test set. The remaining dataset is our train set.
The following function can be used to partition our dataset into train and test sets. We will use it later in this post.
def train_test_split_df(df, test_size=10):
def drop_rows(df, remove_n):#, itemIds_single_rating):
drop_indices = np.random.choice(df.index, remove_n, replace=False)
#drop_indices = [i for i in drop_indices if i not in itemIds_single_rating]
return df.drop(drop_indices)
# ItemIds' occuring more than 20 times
counts = collections.Counter(df['itemId'])
#itemIds_single_rating = [k for k,v in counts.items() if v<2]
grouped = df.groupby(by=['userId'], group_keys=False)
remove_n = test_size
df_train = df.groupby(['userId']).apply(
lambda x: drop_rows(x, remove_n))#,
#itemIds_single_rating))
df_train = df_train.reset_index(level=0, drop=True)
df_test = df[~df.index.isin(df_train.index)]
removed_item_ids = set(df['itemId']) - set(df_train['itemId'])
dummy_ratings = []
dummy_userIds = np.random.choice(df['userId'].unique(),
size=len(removed_item_ids))
for i,itemid in enumerate(removed_item_ids):
dummy_ratings.append([dummy_userIds[i],itemid,0.001,'NaN'])
df_train_dummy = pd.DataFrame(dummy_ratings, columns=['userId', 'itemId', 'rating', 'timestamp'])
df_train = pd.concat([df_train, df_train_dummy])
return df_train, df_test
"""#df_train, df_test = train_test_split_df(df_ratings.iloc[:1000], test_size=10)
df_train, df_test = train_test_split_df(df_ratings, test_size=10)
print (len(df_train), len(df_test), len(df_ratings))"""
"""ratings_extractor = RatingsExtractor()
ratings_extractor.transform(df_train)
ratings_extractor = RatingsExtractor()
ratings_extractor.transform(df_ratings)"""
93423 6710 100004
'ratings_extractor = RatingsExtractor()\nratings_extractor.transform(df_train)\n\nratings_extractor = RatingsExtractor()\nratings_extractor.transform(df_ratings)'
Transformer
The following transformer (called RatingsExtractor) converts a pandas dataframe containing itemIds, userIds, and ratings into an $m \times n$ sparse matrix, where $m$ is the number of items and $n$ is the number of users. It also implements dictionaries to convert the row and column indices into item and user ids and vice versa.
class RatingsExtractor(BaseEstimator, TransformerMixin):
def __init__(self):
self.index2user_dict = {}
self.index2item_dict = {}
self.user2index_dict = {}
self.item2index_dict = {}
self.ratings = csr_matrix([], dtype=np.float)
def fit(self,df_ratings):
return self
def transform(self, df_ratings):
for i,uid in enumerate(sorted(df_ratings['userId'].unique())):
self.index2user_dict[i] = uid
for i,mid in enumerate(sorted(df_ratings['itemId'].unique())):
self.index2item_dict[i] = mid
self.user2index_dict = {v:k for k,v in self.index2user_dict.items()}
self.item2index_dict = {v:k for k,v in self.index2item_dict.items()}
n_items = len(self.index2item_dict)
n_users = len(self.index2user_dict)
row_ind = [self.item2index_dict[mov]
for mov in df_ratings['itemId'].values]
col_ind = [self.user2index_dict[mov]
for mov in df_ratings['userId'].values]
data = df_ratings['rating'].values
self.ratings = csr_matrix((data, (row_ind, col_ind)), shape=(n_items, n_users))
return self.ratings
def get_row_col_indices(self, df_test):
row_indices = [self.item2index_dict[k]
for k in df_test['itemId'].values]
col_indices = [self.user2index_dict[k]
for k in df_test['userId'].values]
return list(zip(row_indices,col_indices))
"""ratings_extractor = RatingsExtractor()
ratings = ratings_extractor.transform(df_ratings)"""
'ratings_extractor = RatingsExtractor()\nratings = ratings_extractor.transform(df_ratings)'
Estimator
Our estimator implements the intuition shown in the figure below.
So our predicted rating for item-user pair ($i,x$) is given by:
\begin{align}
\hat{r}_{ix} = \mu + b_i + b_x + \mathbf{q_i} \cdot \mathbf{p_x}
\end{align}
where,
$\mu$ is the global mean rating,
$b_i$ is the item bias for item $i$,
$b_x$ is the user bias for user $x$,
$\mathbf{q_i}$ is the latent factor vector for item $i$, and
$\mathbf{p_x}$ is the latent factor vector for user $x$.
The objective function can be written as
\begin{align}
J =
\sum_{(i,x) \in R} \left( r_{i,x} - \left( \mu +
b_i + b_x +
\mathbf{q_i} \cdot \mathbf{p_x} \right) \right)^2
+ \lambda_{if} \sum_i \left\lVert q_i \right\rVert ^2
+ \lambda_{xf} \sum_x \left\lVert p_x \right\rVert ^2 \
+ \lambda_{ib} \sum_i b_i^2 +
\lambda_{xb} \sum_x b_x^2
\end{align}
Here, $\lambda_{if}, \lambda_{xf}, \lambda_{ib}, \lambda_{xb}$ are regularization parameters. I used Stochastic Gradient Descent (SGD) algorithm to minimize the objective function $J$. Alternatively, one can also use Alternating Least Squares (ALS) algorithm.
In the SGD algorithm, we update each parameter ($b_x$, $b_i$, $q_i$, and $p_x$) with each sample using gradient update of the form
\begin{align} \theta := \theta - \eta \frac{\partial J}{\partial \theta} \end{align}.
Our SGD parameter updates become
\begin{align}
b_i := b_i + \eta \left( e_{ix} - \lambda{ib} b_i \right)
b_x := b_x + \eta \left( e_{ix} - \lambda_{xb} b_x \right)
\mathbf{q_i} := \mathbf{q_i} + \eta \left( e_{ix} \mathbf{p_x} - \lambda_{if} \mathbf{q_i} \right)
\mathbf{p_x} := \mathbf{p_x} + \eta \left( e_{ix} \mathbf{q_i} - \lambda_{xf} \mathbf{p_x} \right)
\end{align}
where, $e_{ix} = r_{ix} - \hat{r}_{ix}$ is the prediction error.
The above algorithm is implemented in LatentFactorRecSys class below.
class LatentFactorRecSys(BaseEstimator, TransformerMixin):
def __init__(self, reg_item_fact=0.1, reg_user_fact=0.1,
reg_item_bias=0.1, reg_user_bias=0.1,
num_factors=20,
learning_rate=0.001, max_iter=10,
tolerance=0.1):
self.reg_item_fact = reg_item_fact
self.reg_user_fact = reg_user_fact
self.reg_item_bias = reg_item_bias
self.reg_user_bias = reg_user_bias
self.num_factors = num_factors
self.learning_rate = learning_rate
self.max_iter = max_iter
self.tolerance = tolerance
"""self.R = csr_matrix([], dtype=np.float)
self.mu = 0.0
self.bx = csr_matrix([], dtype=np.float)
self.bi = csr_matrix([], dtype=np.float)
self.P = csr_matrix([], dtype=np.float)
self.Q = csr_matrix([], dtype=np.float)
self.rmse_iter = []"""
def fit(self,ratings,y=None):
self.R = ratings
f = self.num_factors
eta = self.learning_rate
lambda_if = self.reg_item_fact
lambda_xf = self.reg_user_fact
lambda_ib = self.reg_item_bias
lambda_xb = self.reg_user_bias
#nonzero_indices = list(zip(self.R.nonzero()[0].tolist(),
# self.R.nonzero()[1].tolist()))
nonzero_indices = list(zip(*ratings.nonzero()))
n_items,n_users = self.R.shape
self.P = np.random.uniform(0,np.sqrt(self.R.max()/f),
size=(n_users,f)) # User factors
self.Q = np.random.uniform(0,np.sqrt(self.R.max()/f),
size=(n_items,f)) # Item factors
self.bx = np.random.uniform(-1,1,n_users) # User bias
self.bi = np.random.uniform(-1,1,n_items) # Item bias
self.mu = self.R.sum()/self.R.size # Overall mean rating
self.rmse_iter = []
for it in range(self.max_iter):
for i,x in nonzero_indices:
res_ix = self.R[i,x] - self.predict_ix(i,x)
self.bx[x] += eta*(res_ix - lambda_xb*self.bx[x])
self.bi[i] += eta*(res_ix - lambda_ib*self.bi[i])
ptemp = self.P[x,:]
self.P[x,:] += eta*(res_ix*self.Q[i,:] -
lambda_xf*self.P[x,:])
self.Q[i,:] += eta*(res_ix*ptemp -
lambda_if*self.Q[i,:])
cur_rmse = self.rmse_all()
self.rmse_iter.append((it,cur_rmse))
if (it>0 and self.rmse_iter[it-1][1]-cur_rmse < self.tolerance):
print ('converged at iter = %d, rmse = %f' % (it+1, cur_rmse))
return self
print ('not converged, number of iterations exceeded max_iter = %d ' % self.max_iter)
print ('current rmse = %f' % cur_rmse)
return self
def transform(self,ratings,y=None):
return self
def predict_ix(self,item_index,user_index):
return (self.mu + self.bx[user_index] + self.bi[item_index] +
np.dot(self.P[user_index,:],self.Q[item_index,:]))
def predict(self,itemid_userid_array,ratings_extractor):
pred = []
for itemid,userid in itemid_userid_array:
item_index = ratings_extractor.item2index_dict[itemid]
user_index = ratings_extractor.user2index_dict[userid]
#print (itemid,userid,item_index,user_index)
pred.append(self.predict_ix(item_index,user_index))
return pred
def predict_all(self):
n_items, n_users = self.R.shape
#nonzero_indices = list(zip(self.R.nonzero()[0].tolist(),
# self.R.nonzero()[1].tolist()))
nonzero_indices = list(zip(*self.R.nonzero()))
row_ind = [i for i,x in nonzero_indices]
col_ind = [x for i,x in nonzero_indices]
mask = csr_matrix(([1.]*self.R.size, (row_ind, col_ind)),
shape=(n_items, n_users))
biases = (self.mu*mask
+ mask.multiply(np.tile(self.bx,(n_items,1)))
+ mask.multiply(np.tile(self.bi[np.newaxis].transpose(),
(1,n_users))))
predictions = biases + mask.multiply(np.matmul(self.Q,self.P.T)) # elementwise multiplication
return predictions
def rmse_all(self):
Residue = self.R - self.predict_all()
return np.sqrt(Residue.multiply(Residue).sum()/self.R.size)
def score(self,ratings,y=None):
# rec.named_steps['lf_recommender'].R.data
return r2_score(self.R.data, self.predict_all().data)
def get_convergence_curve(self):
return self.rmse_iter
"""rec = LatentFactorRecSys(max_iter=2)
# print (rec.get_params())
# rec.set_params(max_iter=10)
rec.fit(ratings)
conv_curve = rec.get_convergence_curve()
iter_index,rmse = tuple(zip(*conv_curve))
plt.plot(iter_index,rmse, '-ok')
plt.xlabel('iteration')
plt.ylabel('RMSE')"""
"rec = LatentFactorRecSys(max_iter=2)\n# print (rec.get_params())\n# rec.set_params(max_iter=10)\n\nrec.fit(ratings)\n\nconv_curve = rec.get_convergence_curve()\niter_index,rmse = tuple(zip(*conv_curve))\n\nplt.plot(iter_index,rmse, '-ok')\nplt.xlabel('iteration')\nplt.ylabel('RMSE')"
Pipeline
We combine our RatingsExtractor and LatentFactorRecSys classes in a scikit-learn pipeline.
rec_pipeline = Pipeline(
[('ratings_extractor', RatingsExtractor()),
('lf_recommender', LatentFactorRecSys(max_iter=10, tolerance=0.01))
])
"""num_samples = 1000
rec_pipeline.fit(df_ratings.sample(num_samples, random_state=0))
print (rec_pipeline)
conv_curve = rec_pipeline.named_steps['lf_recommender'].get_convergence_curve()
iter_index,rmse = tuple(zip(*conv_curve))
plt.plot(iter_index,rmse, '-ok')
plt.xlabel('iteration')
plt.ylabel('RMSE')"""
"num_samples = 1000\nrec_pipeline.fit(df_ratings.sample(num_samples, random_state=0))\nprint (rec_pipeline)\n\nconv_curve = rec_pipeline.named_steps['lf_recommender'].get_convergence_curve()\niter_index,rmse = tuple(zip(*conv_curve))\n\nplt.plot(iter_index,rmse, '-ok')\nplt.xlabel('iteration')\nplt.ylabel('RMSE')"
GridSearch
Next, we perform a gridsearch over our model parameters $f$ (number of latent factors), $\lambda_{if}, \lambda_{xf}, \lambda_{ib}, \lambda_{xb}$ and the hyperparameter $\eta$ (learning rate).
In the following grid search, I have used only 1,000 samples and $max_iter = 30$. In addition, all the reguralization coefficients are set to the same value to limit the search space. This is done to keep the computations tractable.
"""num_latent_factors = [10] #[10, 20]
regularizations = [0.1] #[0.1, 1.0]
learning_rates = [0.01] #[0.01, 0.1]
iterations = [2] #[25, 50, 75, 100]"""
num_latent_factors = [40, 60, 80, 100]
regularizations = [0.001, 0.01, 0.1, 1.0]
learning_rates = [0.0001, 0.001, 0.01]
iterations = [30]
grid= []
for reg in regularizations:
grid.append({'lf_recommender__num_factors': num_latent_factors,
'lf_recommender__learning_rate': learning_rates,
'lf_recommender__max_iter': iterations,
'lf_recommender__reg_item_bias': [reg],
'lf_recommender__reg_item_fact': [reg],
'lf_recommender__reg_user_bias': [reg],
'lf_recommender__reg_user_fact': [reg],
})
#print (list(ParameterGrid(grid)))
grid_search = GridSearchCV(rec_pipeline, param_grid=grid,
cv=2, verbose=0, n_jobs=3)
print (grid_search)
num_samples = 1000
grid_search.fit(df_ratings.sample(num_samples, random_state=0))
GridSearchCV(cv=2, error_score='raise',
estimator=Pipeline(steps=[('ratings_extractor', RatingsExtractor()), ('lf_recommender', LatentFactorRecSys(learning_rate=0.001, max_iter=10, num_factors=20,
reg_item_bias=0.1, reg_item_fact=0.1, reg_user_bias=0.1,
reg_user_fact=0.1, tolerance=0.01))]),
fit_params={}, iid=True, n_jobs=3,
param_grid=[{'lf_recommender__num_factors': [40, 60, 80, 100], 'lf_recommender__learning_rate': [0.0001, 0.001, 0.01], 'lf_recommender__max_iter': [30], 'lf_recommender__reg_item_bias': [0.001], 'lf_recommender__reg_item_fact': [0.001], 'lf_recommender__reg_user_bias': [0.001], 'lf_recommender__reg_..._item_fact': [1.0], 'lf_recommender__reg_user_bias': [1.0], 'lf_recommender__reg_user_fact': [1.0]}],
pre_dispatch='2*n_jobs', refit=True, return_train_score=True,
scoring=None, verbose=0)
converged at iter = 2, rmse = 1.777826
converged at iter = 2, rmse = 1.828849
converged at iter = 2, rmse = 1.820899
converged at iter = 2, rmse = 1.886549
converged at iter = 2, rmse = 1.815157
converged at iter = 2, rmse = 1.892730
converged at iter = 2, rmse = 1.958522
converged at iter = 2, rmse = 1.842210
not converged, number of iterations exceeded max_iter = 30
current rmse = 1.495032
converged at iter = 29, rmse = 1.465206
not converged, number of iterations exceeded max_iter = 30
current rmse = 1.531471
converged at iter = 30, rmse = 1.455406
not converged, number of iterations exceeded max_iter = 30
current rmse = 1.502702
not converged, number of iterations exceeded max_iter = 30
current rmse = 1.506124
converged at iter = 30, rmse = 1.450238
not converged, number of iterations exceeded max_iter = 30
current rmse = 0.447158
not converged, number of iterations exceeded max_iter = 30
current rmse = 0.438726
converged at iter = 27, rmse = 1.419675
not converged, number of iterations exceeded max_iter = 30
current rmse = 0.443659
not converged, number of iterations exceeded max_iter = 30
current rmse = 0.447455
not converged, number of iterations exceeded max_iter = 30
current rmse = 0.462130
not converged, number of iterations exceeded max_iter = 30
current rmse = 0.441478
not converged, number of iterations exceeded max_iter = 30
current rmse = 0.435423
converged at iter = 2, rmse = 1.784997
converged at iter = 2, rmse = 1.893901
converged at iter = 2, rmse = 1.874529
converged at iter = 2, rmse = 1.833197
converged at iter = 2, rmse = 1.837853
converged at iter = 2, rmse = 1.898421
converged at iter = 2, rmse = 1.826771
converged at iter = 2, rmse = 1.874525
not converged, number of iterations exceeded max_iter = 30
current rmse = 0.458926
not converged, number of iterations exceeded max_iter = 30
current rmse = 1.498145
not converged, number of iterations exceeded max_iter = 30
current rmse = 1.472105
converged at iter = 27, rmse = 1.468265
converged at iter = 27, rmse = 1.486684
converged at iter = 30, rmse = 1.415607
not converged, number of iterations exceeded max_iter = 30
current rmse = 1.484826
converged at iter = 29, rmse = 1.453231
not converged, number of iterations exceeded max_iter = 30
current rmse = 1.470400
not converged, number of iterations exceeded max_iter = 30
current rmse = 0.469220
not converged, number of iterations exceeded max_iter = 30
current rmse = 0.436152
not converged, number of iterations exceeded max_iter = 30
current rmse = 0.457324
not converged, number of iterations exceeded max_iter = 30
current rmse = 0.459726
not converged, number of iterations exceeded max_iter = 30
current rmse = 0.438211
not converged, number of iterations exceeded max_iter = 30
current rmse = 0.440269
converged at iter = 2, rmse = 1.888414
converged at iter = 2, rmse = 1.830839
converged at iter = 2, rmse = 1.809775
converged at iter = 2, rmse = 1.950127
converged at iter = 2, rmse = 1.879681
not converged, number of iterations exceeded max_iter = 30
current rmse = 0.443645
converged at iter = 2, rmse = 1.811025
converged at iter = 2, rmse = 1.759191
converged at iter = 2, rmse = 1.803322
not converged, number of iterations exceeded max_iter = 30
current rmse = 0.450282
not converged, number of iterations exceeded max_iter = 30
current rmse = 1.431791
not converged, number of iterations exceeded max_iter = 30
current rmse = 1.521118
not converged, number of iterations exceeded max_iter = 30
current rmse = 1.517283
converged at iter = 24, rmse = 1.440038
not converged, number of iterations exceeded max_iter = 30
current rmse = 1.474165
not converged, number of iterations exceeded max_iter = 30
current rmse = 1.478043
converged at iter = 26, rmse = 1.422170
converged at iter = 30, rmse = 1.463759
not converged, number of iterations exceeded max_iter = 30
current rmse = 0.478135
not converged, number of iterations exceeded max_iter = 30
current rmse = 0.465071
not converged, number of iterations exceeded max_iter = 30
current rmse = 0.470699
not converged, number of iterations exceeded max_iter = 30
current rmse = 0.455117
not converged, number of iterations exceeded max_iter = 30
current rmse = 0.472355
not converged, number of iterations exceeded max_iter = 30
current rmse = 0.471608
converged at iter = 2, rmse = 1.823752
converged at iter = 2, rmse = 1.849292
not converged, number of iterations exceeded max_iter = 30
current rmse = 0.474583
converged at iter = 2, rmse = 1.891493
converged at iter = 2, rmse = 1.889316
converged at iter = 2, rmse = 1.846779
converged at iter = 2, rmse = 1.813008
not converged, number of iterations exceeded max_iter = 30
current rmse = 0.459500
converged at iter = 2, rmse = 1.906590
converged at iter = 2, rmse = 1.865887
not converged, number of iterations exceeded max_iter = 30
current rmse = 1.365381
not converged, number of iterations exceeded max_iter = 30
current rmse = 1.353588
not converged, number of iterations exceeded max_iter = 30
current rmse = 1.407780
not converged, number of iterations exceeded max_iter = 30
current rmse = 1.354890
not converged, number of iterations exceeded max_iter = 30
current rmse = 1.395890
not converged, number of iterations exceeded max_iter = 30
current rmse = 1.380766
converged at iter = 24, rmse = 0.681698
not converged, number of iterations exceeded max_iter = 30
current rmse = 1.418345
not converged, number of iterations exceeded max_iter = 30
current rmse = 1.364138
converged at iter = 24, rmse = 0.676327
converged at iter = 23, rmse = 0.660403
converged at iter = 24, rmse = 0.698259
converged at iter = 24, rmse = 0.690734
converged at iter = 25, rmse = 0.672983
converged at iter = 25, rmse = 0.675286
converged at iter = 25, rmse = 0.662770
not converged, number of iterations exceeded max_iter = 30
current rmse = 0.456928
GridSearchCV(cv=2, error_score='raise',
estimator=Pipeline(steps=[('ratings_extractor', RatingsExtractor()), ('lf_recommender', LatentFactorRecSys(learning_rate=0.001, max_iter=10, num_factors=20,
reg_item_bias=0.1, reg_item_fact=0.1, reg_user_bias=0.1,
reg_user_fact=0.1, tolerance=0.01))]),
fit_params={}, iid=True, n_jobs=3,
param_grid=[{'lf_recommender__num_factors': [40, 60, 80, 100], 'lf_recommender__learning_rate': [0.0001, 0.001, 0.01], 'lf_recommender__max_iter': [30], 'lf_recommender__reg_item_bias': [0.001], 'lf_recommender__reg_item_fact': [0.001], 'lf_recommender__reg_user_bias': [0.001], 'lf_recommender__reg_..._item_fact': [1.0], 'lf_recommender__reg_user_bias': [1.0], 'lf_recommender__reg_user_fact': [1.0]}],
pre_dispatch='2*n_jobs', refit=True, return_train_score=True,
scoring=None, verbose=0)
grid_search.cv_results_
{'mean_fit_time': array([ 0.18, 0.25, 0.29, 0.3 , 3.88, 4.25, 3.7 , 4.09, 4.3 ,
4.02, 4.03, 4.04, 0.3 , 0.28, 0.3 , 0.23, 3.99, 3.71,
3.81, 3.87, 3.84, 4.04, 4.54, 4.55, 0.24, 0.33, 0.35,
0.36, 4.02, 3.27, 4.11, 3.54, 3.61, 3.92, 4. , 4.12,
0.31, 0.32, 0.29, 0.32, 4.1 , 4.01, 3.93, 4.07, 3.1 ,
3.1 , 2.96, 2.54]),
'mean_score_time': array([ 0.01, 0.01, 0.01, 0.01, 0.02, 0.01, 0.01, 0.01, 0.02,
0.01, 0.02, 0.01, 0.02, 0.02, 0.01, 0.01, 0.03, 0.01,
0.01, 0.04, 0.02, 0.01, 0.02, 0.02, 0.01, 0.03, 0.02,
0.03, 0.01, 0.01, 0.02, 0.02, 0.02, 0.01, 0.02, 0.02,
0.02, 0.01, 0.02, 0.01, 0.01, 0.02, 0.01, 0.01, 0.01,
0.01, 0.03, 0.02]),
'mean_test_score': array([-1.76, -1.91, -1.91, -2.06, -0.84, -0.95, -0.87, -0.74, 0.83,
0.83, 0.83, 0.83, -1.87, -1.91, -1.84, -2.01, -0.87, -0.85,
-0.78, -0.81, 0.83, 0.82, 0.84, 0.83, -1.93, -2. , -1.88,
-1.69, -0.85, -0.85, -0.85, -0.76, 0.81, 0.82, 0.81, 0.82,
-1.86, -2.03, -1.84, -2.01, -0.63, -0.6 , -0.58, -0.64, 0.61,
0.61, 0.61, 0.62]),
'mean_train_score': array([-1.76, -1.91, -1.91, -2.06, -0.84, -0.95, -0.87, -0.74, 0.83,
0.83, 0.83, 0.83, -1.87, -1.91, -1.84, -2.01, -0.87, -0.85,
-0.78, -0.81, 0.83, 0.82, 0.84, 0.83, -1.93, -2. , -1.88,
-1.69, -0.85, -0.85, -0.85, -0.76, 0.81, 0.82, 0.81, 0.82,
-1.86, -2.03, -1.84, -2.01, -0.63, -0.6 , -0.58, -0.64, 0.61,
0.61, 0.61, 0.62]),
'param_lf_recommender__learning_rate': masked_array(data = [0.0001 0.0001 0.0001 0.0001 0.001 0.001 0.001 0.001 0.01 0.01 0.01 0.01
0.0001 0.0001 0.0001 0.0001 0.001 0.001 0.001 0.001 0.01 0.01 0.01 0.01
0.0001 0.0001 0.0001 0.0001 0.001 0.001 0.001 0.001 0.01 0.01 0.01 0.01
0.0001 0.0001 0.0001 0.0001 0.001 0.001 0.001 0.001 0.01 0.01 0.01 0.01],
mask = [False False False False False False False False False False False False
False False False False False False False False False False False False
False False False False False False False False False False False False
False False False False False False False False False False False False],
fill_value = ?),
'param_lf_recommender__max_iter': masked_array(data = [30 30 30 30 30 30 30 30 30 30 30 30 30 30 30 30 30 30 30 30 30 30 30 30 30
30 30 30 30 30 30 30 30 30 30 30 30 30 30 30 30 30 30 30 30 30 30 30],
mask = [False False False False False False False False False False False False
False False False False False False False False False False False False
False False False False False False False False False False False False
False False False False False False False False False False False False],
fill_value = ?),
'param_lf_recommender__num_factors': masked_array(data = [40 60 80 100 40 60 80 100 40 60 80 100 40 60 80 100 40 60 80 100 40 60 80
100 40 60 80 100 40 60 80 100 40 60 80 100 40 60 80 100 40 60 80 100 40 60
80 100],
mask = [False False False False False False False False False False False False
False False False False False False False False False False False False
False False False False False False False False False False False False
False False False False False False False False False False False False],
fill_value = ?),
'param_lf_recommender__reg_item_bias': masked_array(data = [0.001 0.001 0.001 0.001 0.001 0.001 0.001 0.001 0.001 0.001 0.001 0.001
0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.1 0.1 0.1
0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0
1.0 1.0 1.0],
mask = [False False False False False False False False False False False False
False False False False False False False False False False False False
False False False False False False False False False False False False
False False False False False False False False False False False False],
fill_value = ?),
'param_lf_recommender__reg_item_fact': masked_array(data = [0.001 0.001 0.001 0.001 0.001 0.001 0.001 0.001 0.001 0.001 0.001 0.001
0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.1 0.1 0.1
0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0
1.0 1.0 1.0],
mask = [False False False False False False False False False False False False
False False False False False False False False False False False False
False False False False False False False False False False False False
False False False False False False False False False False False False],
fill_value = ?),
'param_lf_recommender__reg_user_bias': masked_array(data = [0.001 0.001 0.001 0.001 0.001 0.001 0.001 0.001 0.001 0.001 0.001 0.001
0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.1 0.1 0.1
0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0
1.0 1.0 1.0],
mask = [False False False False False False False False False False False False
False False False False False False False False False False False False
False False False False False False False False False False False False
False False False False False False False False False False False False],
fill_value = ?),
'param_lf_recommender__reg_user_fact': masked_array(data = [0.001 0.001 0.001 0.001 0.001 0.001 0.001 0.001 0.001 0.001 0.001 0.001
0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.1 0.1 0.1
0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0
1.0 1.0 1.0],
mask = [False False False False False False False False False False False False
False False False False False False False False False False False False
False False False False False False False False False False False False
False False False False False False False False False False False False],
fill_value = ?),
'params': ({'lf_recommender__learning_rate': 0.0001,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 40,
'lf_recommender__reg_item_bias': 0.001,
'lf_recommender__reg_item_fact': 0.001,
'lf_recommender__reg_user_bias': 0.001,
'lf_recommender__reg_user_fact': 0.001},
{'lf_recommender__learning_rate': 0.0001,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 60,
'lf_recommender__reg_item_bias': 0.001,
'lf_recommender__reg_item_fact': 0.001,
'lf_recommender__reg_user_bias': 0.001,
'lf_recommender__reg_user_fact': 0.001},
{'lf_recommender__learning_rate': 0.0001,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 80,
'lf_recommender__reg_item_bias': 0.001,
'lf_recommender__reg_item_fact': 0.001,
'lf_recommender__reg_user_bias': 0.001,
'lf_recommender__reg_user_fact': 0.001},
{'lf_recommender__learning_rate': 0.0001,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 100,
'lf_recommender__reg_item_bias': 0.001,
'lf_recommender__reg_item_fact': 0.001,
'lf_recommender__reg_user_bias': 0.001,
'lf_recommender__reg_user_fact': 0.001},
{'lf_recommender__learning_rate': 0.001,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 40,
'lf_recommender__reg_item_bias': 0.001,
'lf_recommender__reg_item_fact': 0.001,
'lf_recommender__reg_user_bias': 0.001,
'lf_recommender__reg_user_fact': 0.001},
{'lf_recommender__learning_rate': 0.001,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 60,
'lf_recommender__reg_item_bias': 0.001,
'lf_recommender__reg_item_fact': 0.001,
'lf_recommender__reg_user_bias': 0.001,
'lf_recommender__reg_user_fact': 0.001},
{'lf_recommender__learning_rate': 0.001,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 80,
'lf_recommender__reg_item_bias': 0.001,
'lf_recommender__reg_item_fact': 0.001,
'lf_recommender__reg_user_bias': 0.001,
'lf_recommender__reg_user_fact': 0.001},
{'lf_recommender__learning_rate': 0.001,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 100,
'lf_recommender__reg_item_bias': 0.001,
'lf_recommender__reg_item_fact': 0.001,
'lf_recommender__reg_user_bias': 0.001,
'lf_recommender__reg_user_fact': 0.001},
{'lf_recommender__learning_rate': 0.01,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 40,
'lf_recommender__reg_item_bias': 0.001,
'lf_recommender__reg_item_fact': 0.001,
'lf_recommender__reg_user_bias': 0.001,
'lf_recommender__reg_user_fact': 0.001},
{'lf_recommender__learning_rate': 0.01,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 60,
'lf_recommender__reg_item_bias': 0.001,
'lf_recommender__reg_item_fact': 0.001,
'lf_recommender__reg_user_bias': 0.001,
'lf_recommender__reg_user_fact': 0.001},
{'lf_recommender__learning_rate': 0.01,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 80,
'lf_recommender__reg_item_bias': 0.001,
'lf_recommender__reg_item_fact': 0.001,
'lf_recommender__reg_user_bias': 0.001,
'lf_recommender__reg_user_fact': 0.001},
{'lf_recommender__learning_rate': 0.01,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 100,
'lf_recommender__reg_item_bias': 0.001,
'lf_recommender__reg_item_fact': 0.001,
'lf_recommender__reg_user_bias': 0.001,
'lf_recommender__reg_user_fact': 0.001},
{'lf_recommender__learning_rate': 0.0001,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 40,
'lf_recommender__reg_item_bias': 0.01,
'lf_recommender__reg_item_fact': 0.01,
'lf_recommender__reg_user_bias': 0.01,
'lf_recommender__reg_user_fact': 0.01},
{'lf_recommender__learning_rate': 0.0001,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 60,
'lf_recommender__reg_item_bias': 0.01,
'lf_recommender__reg_item_fact': 0.01,
'lf_recommender__reg_user_bias': 0.01,
'lf_recommender__reg_user_fact': 0.01},
{'lf_recommender__learning_rate': 0.0001,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 80,
'lf_recommender__reg_item_bias': 0.01,
'lf_recommender__reg_item_fact': 0.01,
'lf_recommender__reg_user_bias': 0.01,
'lf_recommender__reg_user_fact': 0.01},
{'lf_recommender__learning_rate': 0.0001,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 100,
'lf_recommender__reg_item_bias': 0.01,
'lf_recommender__reg_item_fact': 0.01,
'lf_recommender__reg_user_bias': 0.01,
'lf_recommender__reg_user_fact': 0.01},
{'lf_recommender__learning_rate': 0.001,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 40,
'lf_recommender__reg_item_bias': 0.01,
'lf_recommender__reg_item_fact': 0.01,
'lf_recommender__reg_user_bias': 0.01,
'lf_recommender__reg_user_fact': 0.01},
{'lf_recommender__learning_rate': 0.001,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 60,
'lf_recommender__reg_item_bias': 0.01,
'lf_recommender__reg_item_fact': 0.01,
'lf_recommender__reg_user_bias': 0.01,
'lf_recommender__reg_user_fact': 0.01},
{'lf_recommender__learning_rate': 0.001,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 80,
'lf_recommender__reg_item_bias': 0.01,
'lf_recommender__reg_item_fact': 0.01,
'lf_recommender__reg_user_bias': 0.01,
'lf_recommender__reg_user_fact': 0.01},
{'lf_recommender__learning_rate': 0.001,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 100,
'lf_recommender__reg_item_bias': 0.01,
'lf_recommender__reg_item_fact': 0.01,
'lf_recommender__reg_user_bias': 0.01,
'lf_recommender__reg_user_fact': 0.01},
{'lf_recommender__learning_rate': 0.01,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 40,
'lf_recommender__reg_item_bias': 0.01,
'lf_recommender__reg_item_fact': 0.01,
'lf_recommender__reg_user_bias': 0.01,
'lf_recommender__reg_user_fact': 0.01},
{'lf_recommender__learning_rate': 0.01,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 60,
'lf_recommender__reg_item_bias': 0.01,
'lf_recommender__reg_item_fact': 0.01,
'lf_recommender__reg_user_bias': 0.01,
'lf_recommender__reg_user_fact': 0.01},
{'lf_recommender__learning_rate': 0.01,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 80,
'lf_recommender__reg_item_bias': 0.01,
'lf_recommender__reg_item_fact': 0.01,
'lf_recommender__reg_user_bias': 0.01,
'lf_recommender__reg_user_fact': 0.01},
{'lf_recommender__learning_rate': 0.01,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 100,
'lf_recommender__reg_item_bias': 0.01,
'lf_recommender__reg_item_fact': 0.01,
'lf_recommender__reg_user_bias': 0.01,
'lf_recommender__reg_user_fact': 0.01},
{'lf_recommender__learning_rate': 0.0001,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 40,
'lf_recommender__reg_item_bias': 0.1,
'lf_recommender__reg_item_fact': 0.1,
'lf_recommender__reg_user_bias': 0.1,
'lf_recommender__reg_user_fact': 0.1},
{'lf_recommender__learning_rate': 0.0001,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 60,
'lf_recommender__reg_item_bias': 0.1,
'lf_recommender__reg_item_fact': 0.1,
'lf_recommender__reg_user_bias': 0.1,
'lf_recommender__reg_user_fact': 0.1},
{'lf_recommender__learning_rate': 0.0001,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 80,
'lf_recommender__reg_item_bias': 0.1,
'lf_recommender__reg_item_fact': 0.1,
'lf_recommender__reg_user_bias': 0.1,
'lf_recommender__reg_user_fact': 0.1},
{'lf_recommender__learning_rate': 0.0001,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 100,
'lf_recommender__reg_item_bias': 0.1,
'lf_recommender__reg_item_fact': 0.1,
'lf_recommender__reg_user_bias': 0.1,
'lf_recommender__reg_user_fact': 0.1},
{'lf_recommender__learning_rate': 0.001,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 40,
'lf_recommender__reg_item_bias': 0.1,
'lf_recommender__reg_item_fact': 0.1,
'lf_recommender__reg_user_bias': 0.1,
'lf_recommender__reg_user_fact': 0.1},
{'lf_recommender__learning_rate': 0.001,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 60,
'lf_recommender__reg_item_bias': 0.1,
'lf_recommender__reg_item_fact': 0.1,
'lf_recommender__reg_user_bias': 0.1,
'lf_recommender__reg_user_fact': 0.1},
{'lf_recommender__learning_rate': 0.001,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 80,
'lf_recommender__reg_item_bias': 0.1,
'lf_recommender__reg_item_fact': 0.1,
'lf_recommender__reg_user_bias': 0.1,
'lf_recommender__reg_user_fact': 0.1},
{'lf_recommender__learning_rate': 0.001,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 100,
'lf_recommender__reg_item_bias': 0.1,
'lf_recommender__reg_item_fact': 0.1,
'lf_recommender__reg_user_bias': 0.1,
'lf_recommender__reg_user_fact': 0.1},
{'lf_recommender__learning_rate': 0.01,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 40,
'lf_recommender__reg_item_bias': 0.1,
'lf_recommender__reg_item_fact': 0.1,
'lf_recommender__reg_user_bias': 0.1,
'lf_recommender__reg_user_fact': 0.1},
{'lf_recommender__learning_rate': 0.01,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 60,
'lf_recommender__reg_item_bias': 0.1,
'lf_recommender__reg_item_fact': 0.1,
'lf_recommender__reg_user_bias': 0.1,
'lf_recommender__reg_user_fact': 0.1},
{'lf_recommender__learning_rate': 0.01,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 80,
'lf_recommender__reg_item_bias': 0.1,
'lf_recommender__reg_item_fact': 0.1,
'lf_recommender__reg_user_bias': 0.1,
'lf_recommender__reg_user_fact': 0.1},
{'lf_recommender__learning_rate': 0.01,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 100,
'lf_recommender__reg_item_bias': 0.1,
'lf_recommender__reg_item_fact': 0.1,
'lf_recommender__reg_user_bias': 0.1,
'lf_recommender__reg_user_fact': 0.1},
{'lf_recommender__learning_rate': 0.0001,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 40,
'lf_recommender__reg_item_bias': 1.0,
'lf_recommender__reg_item_fact': 1.0,
'lf_recommender__reg_user_bias': 1.0,
'lf_recommender__reg_user_fact': 1.0},
{'lf_recommender__learning_rate': 0.0001,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 60,
'lf_recommender__reg_item_bias': 1.0,
'lf_recommender__reg_item_fact': 1.0,
'lf_recommender__reg_user_bias': 1.0,
'lf_recommender__reg_user_fact': 1.0},
{'lf_recommender__learning_rate': 0.0001,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 80,
'lf_recommender__reg_item_bias': 1.0,
'lf_recommender__reg_item_fact': 1.0,
'lf_recommender__reg_user_bias': 1.0,
'lf_recommender__reg_user_fact': 1.0},
{'lf_recommender__learning_rate': 0.0001,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 100,
'lf_recommender__reg_item_bias': 1.0,
'lf_recommender__reg_item_fact': 1.0,
'lf_recommender__reg_user_bias': 1.0,
'lf_recommender__reg_user_fact': 1.0},
{'lf_recommender__learning_rate': 0.001,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 40,
'lf_recommender__reg_item_bias': 1.0,
'lf_recommender__reg_item_fact': 1.0,
'lf_recommender__reg_user_bias': 1.0,
'lf_recommender__reg_user_fact': 1.0},
{'lf_recommender__learning_rate': 0.001,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 60,
'lf_recommender__reg_item_bias': 1.0,
'lf_recommender__reg_item_fact': 1.0,
'lf_recommender__reg_user_bias': 1.0,
'lf_recommender__reg_user_fact': 1.0},
{'lf_recommender__learning_rate': 0.001,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 80,
'lf_recommender__reg_item_bias': 1.0,
'lf_recommender__reg_item_fact': 1.0,
'lf_recommender__reg_user_bias': 1.0,
'lf_recommender__reg_user_fact': 1.0},
{'lf_recommender__learning_rate': 0.001,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 100,
'lf_recommender__reg_item_bias': 1.0,
'lf_recommender__reg_item_fact': 1.0,
'lf_recommender__reg_user_bias': 1.0,
'lf_recommender__reg_user_fact': 1.0},
{'lf_recommender__learning_rate': 0.01,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 40,
'lf_recommender__reg_item_bias': 1.0,
'lf_recommender__reg_item_fact': 1.0,
'lf_recommender__reg_user_bias': 1.0,
'lf_recommender__reg_user_fact': 1.0},
{'lf_recommender__learning_rate': 0.01,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 60,
'lf_recommender__reg_item_bias': 1.0,
'lf_recommender__reg_item_fact': 1.0,
'lf_recommender__reg_user_bias': 1.0,
'lf_recommender__reg_user_fact': 1.0},
{'lf_recommender__learning_rate': 0.01,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 80,
'lf_recommender__reg_item_bias': 1.0,
'lf_recommender__reg_item_fact': 1.0,
'lf_recommender__reg_user_bias': 1.0,
'lf_recommender__reg_user_fact': 1.0},
{'lf_recommender__learning_rate': 0.01,
'lf_recommender__max_iter': 30,
'lf_recommender__num_factors': 100,
'lf_recommender__reg_item_bias': 1.0,
'lf_recommender__reg_item_fact': 1.0,
'lf_recommender__reg_user_bias': 1.0,
'lf_recommender__reg_user_fact': 1.0}),
'rank_test_score': array([34, 41, 42, 48, 25, 32, 30, 21, 3, 2, 5, 6, 38, 40, 36, 46, 31,
28, 23, 24, 7, 8, 1, 4, 43, 44, 39, 33, 27, 29, 26, 22, 11, 9,
12, 10, 37, 47, 35, 45, 19, 18, 17, 20, 14, 15, 16, 13], dtype=int32),
'split0_test_score': array([-1.65, -1.78, -1.76, -2.22, -0.87, -0.97, -0.8 , -0.76, 0.83,
0.84, 0.82, 0.84, -1.67, -1.95, -1.83, -2.02, -0.88, -0.81,
-0.68, -0.77, 0.82, 0.82, 0.84, 0.83, -1.99, -1.75, -1.96,
-1.73, -0.94, -0.93, -0.82, -0.7 , 0.81, 0.81, 0.81, 0.81,
-1.79, -2. , -1.76, -2.05, -0.56, -0.54, -0.54, -0.56, 0.61,
0.59, 0.6 , 0.62]),
'split0_train_score': array([-1.65, -1.78, -1.76, -2.22, -0.87, -0.97, -0.8 , -0.76, 0.83,
0.84, 0.82, 0.84, -1.67, -1.95, -1.83, -2.02, -0.88, -0.81,
-0.68, -0.77, 0.82, 0.82, 0.84, 0.83, -1.99, -1.75, -1.96,
-1.73, -0.94, -0.93, -0.82, -0.7 , 0.81, 0.81, 0.81, 0.81,
-1.79, -2. , -1.76, -2.05, -0.56, -0.54, -0.54, -0.56, 0.61,
0.59, 0.6 , 0.62]),
'split1_test_score': array([-1.86, -2.04, -2.06, -1.9 , -0.81, -0.94, -0.93, -0.72, 0.83,
0.83, 0.84, 0.82, -2.07, -1.87, -1.85, -2. , -0.85, -0.89,
-0.89, -0.85, 0.84, 0.82, 0.83, 0.83, -1.87, -2.25, -1.8 ,
-1.65, -0.75, -0.77, -0.87, -0.83, 0.82, 0.82, 0.81, 0.82,
-1.92, -2.05, -1.92, -1.98, -0.69, -0.67, -0.63, -0.72, 0.61,
0.63, 0.61, 0.62]),
'split1_train_score': array([-1.86, -2.04, -2.06, -1.9 , -0.81, -0.94, -0.93, -0.72, 0.83,
0.83, 0.84, 0.82, -2.07, -1.87, -1.85, -2. , -0.85, -0.89,
-0.89, -0.85, 0.84, 0.82, 0.83, 0.83, -1.87, -2.25, -1.8 ,
-1.65, -0.75, -0.77, -0.87, -0.83, 0.82, 0.82, 0.81, 0.82,
-1.92, -2.05, -1.92, -1.98, -0.69, -0.67, -0.63, -0.72, 0.61,
0.63, 0.61, 0.62]),
'std_fit_time': array([ 0.02, 0. , 0.04, 0.02, 0.34, 0.12, 0.08, 0.06, 0. ,
0. , 0.05, 0.15, 0.04, 0.02, 0.04, 0.05, 0.07, 0.11,
0.1 , 0.05, 0.04, 0.13, 0.02, 0.15, 0.04, 0.03, 0.01,
0.05, 0.18, 0.7 , 0.05, 0.42, 0.03, 0.2 , 0.06, 0.09,
0. , 0.01, 0.03, 0.03, 0.12, 0.19, 0.12, 0.03, 0.3 ,
0.04, 0.11, 0.68]),
'std_score_time': array([ 0.01, 0. , 0. , 0. , 0.01, 0.01, 0. , 0. , 0.01,
0.01, 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. ,
0. , 0.01, 0. , 0. , 0.01, 0. , 0. , 0. , 0. ,
0. , 0. , 0.01, 0.01, 0. , 0. , 0.01, 0. , 0. ,
0. , 0.01, 0. , 0. , 0. , 0. , 0. , 0.01, 0. ,
0. , 0.01, 0. ]),
'std_test_score': array([ 0.11, 0.13, 0.15, 0.16, 0.03, 0.01, 0.07, 0.02, 0. ,
0. , 0.01, 0.01, 0.2 , 0.04, 0.01, 0.01, 0.01, 0.04,
0.1 , 0.04, 0.01, 0. , 0. , 0. , 0.06, 0.25, 0.08,
0.04, 0.09, 0.08, 0.02, 0.07, 0. , 0. , 0. , 0. ,
0.07, 0.03, 0.08, 0.04, 0.07, 0.07, 0.05, 0.08, 0. ,
0.02, 0.01, 0. ]),
'std_train_score': array([ 0.11, 0.13, 0.15, 0.16, 0.03, 0.01, 0.07, 0.02, 0. ,
0. , 0.01, 0.01, 0.2 , 0.04, 0.01, 0.01, 0.01, 0.04,
0.1 , 0.04, 0.01, 0. , 0. , 0. , 0.06, 0.25, 0.08,
0.04, 0.09, 0.08, 0.02, 0.07, 0. , 0. , 0. , 0. ,
0.07, 0.03, 0.08, 0.04, 0.07, 0.07, 0.05, 0.08, 0. ,
0.02, 0.01, 0. ])}
Retrain optimal model on test set
Next, we retrain the best estimator (estimator with optimal parameters) found by our grid search. Here, we increase $max_iter$ to 100 and decrease the convergence tolerance of our SGD to 0.001.
grid_search.best_estimator_
#grid_search.best_estimator_.get_params()
grid_search.best_estimator_.set_params(lf_recommender__max_iter=100)
grid_search.best_estimator_.set_params(lf_recommender__tolerance=0.001)
Pipeline(steps=[('ratings_extractor', RatingsExtractor()), ('lf_recommender', LatentFactorRecSys(learning_rate=0.01, max_iter=100, num_factors=80,
reg_item_bias=0.01, reg_item_fact=0.01, reg_user_bias=0.01,
reg_user_fact=0.01, tolerance=0.001))])
rec = grid_search.best_estimator_
print (rec)
df_train, df_test = train_test_split_df(df_ratings, test_size=10)
print ('Retraining optimal model ...')
t0 = time()
rec.fit_transform(df_train)
print("done in %0.3fs." % (time() - t0))
Pipeline(steps=[('ratings_extractor', RatingsExtractor()), ('lf_recommender', LatentFactorRecSys(learning_rate=0.01, max_iter=100, num_factors=80,
reg_item_bias=0.01, reg_item_fact=0.01, reg_user_bias=0.01,
reg_user_fact=0.01, tolerance=0.001))])
Retraining optimal model ...
converged at iter = 77, rmse = 0.142825
done in 397.656s.
#conv_curve = rec.steps[1][1].get_convergence_curve()
conv_curve = rec.named_steps['lf_recommender'].get_convergence_curve()
iter_index,rmse = tuple(zip(*conv_curve))
plt.plot(iter_index,rmse, '-ok')
plt.xlabel('iteration')
plt.ylabel('RMSE')
<matplotlib.text.Text at 0x7fb36a66f438>
RMSE on test set
Next, we calculate the root mean square error (RMSE) on our test set.
"""ratings_extractor = RatingsExtractor()
test_ratings_matrix = ratings_extractor.transform(df_test)
row_col_indices = ratings_extractor.get_row_col_indices(df_test)
test_ratings = [test_ratings_matrix[r,c] for r,c in row_col_indices]
pred = rec.named_steps['lf_recommender'].predict(row_col_indices)
rmse = np.sqrt(mean_squared_error(test_ratings, pred))
print (rmse)"""
"ratings_extractor = RatingsExtractor()\ntest_ratings_matrix = ratings_extractor.transform(df_test)\nrow_col_indices = ratings_extractor.get_row_col_indices(df_test)\ntest_ratings = [test_ratings_matrix[r,c] for r,c in row_col_indices]\n\npred = rec.named_steps['lf_recommender'].predict(row_col_indices)\n\nrmse = np.sqrt(mean_squared_error(test_ratings, pred))\nprint (rmse)"
test_ratings = df_test['rating'].values
itemid_userid_array = [tuple(x) for x in df_test[['itemId', 'userId']]
.to_records(index=False)]
rec_pipeline.named_steps
{'lf_recommender': LatentFactorRecSys(learning_rate=0.001, max_iter=10, num_factors=20,
reg_item_bias=0.1, reg_item_fact=0.1, reg_user_bias=0.1,
reg_user_fact=0.1, tolerance=0.01),
'ratings_extractor': RatingsExtractor()}
lf_recommender = rec.named_steps['lf_recommender']
ratings_extractor = rec.named_steps['ratings_extractor']
#ratings_extractor = RatingsExtractor()
#ratings_extractor.transform(df_ratings)
pred = lf_recommender.predict(itemid_userid_array, ratings_extractor)
rmse = np.sqrt(mean_squared_error(test_ratings, pred))
print (rmse)
0.968712576811
The RMSE of 0.97 looks pretty good, given that we didn’t perform exhaustive search for the best model parameters.
# Save trained model
with open('lf_movie_recommender.pkl', 'wb') as handle:
pickle.dump(rec, handle, protocol=pickle.HIGHEST_PROTOCOL)
# Load trained model
#with open('lf_movie_recommender.pkl', 'rb') as handle:
# rec = pickle.load(handle)
Compute Recommendations
Helper functions: Filter similar movies based on cosine distance as a similarity metric
def cosine_similarity(item_factors):
cos_sim = np.matmul(item_factors, item_factors.transpose())
norm = np.diagonal(cos_sim)
norm_mat = np.sqrt(np.outer(norm, norm))
cos_sim = np.divide(cos_sim, norm_mat)
return cos_sim
def cosine_distance(item_factors):
return 1 - cosine_similarity(item_factors)
cos_dist = cosine_distance(lf_recommender.Q)
array([[ 0. , 0.9 , 0.89, ..., 0.71, 0.69, 0.83],
[ 0.9 , 0. , 1.09, ..., 1. , 0.89, 0.99],
[ 0.89, 1.09, 0. , ..., 0.92, 0.82, 0.89],
...,
[ 0.71, 1. , 0.92, ..., 0. , 0.46, 0.36],
[ 0.69, 0.89, 0.82, ..., 0.46, 0. , 0.31],
[ 0.83, 0.99, 0.89, ..., 0.36, 0.31, 0. ]])
def get_movie_name(df_movies, itemid):
movie_name = df_movies[df_movies['movieId']==itemid]['title'].values[0]
return movie_name
def get_similar_items(itemid, num_recs):
itemindex = ratings_extractor.item2index_dict[itemid]
dists = cos_dist[itemindex,:]
indices = dists.argsort()[1:num_recs+1]
similar_item_ids = [ratings_extractor.index2item_dict[i] for i in indices]
similar_item_names = []
for itemid in similar_item_ids:
similar_item_names.append(get_movie_name(df_movies, itemid))
return similar_item_names
"""num_recs = 5
itemid = 1
movie_name = get_movie_name(df_movies, itemid)
print ('%s ==>' %movie_name)
get_similar_items(itemid, num_recs)"""
"num_recs = 5 \nitemid = 1\nmovie_name = get_movie_name(df_movies, itemid)\nprint ('%s ==>' %movie_name)\nget_similar_items(itemid, num_recs)"
Helper functions: Retrieve IMDB ID given a movie title
IMDB ID is used to get movie poster from themoviedb.org API.
def strip_year(movie_name):
movie_name = re.sub(r'\([^)]*\)', '', movie_name).strip()
return movie_name
def imdb_id_from_title(title):
title = strip_year(title)
pattern = 'http://www.imdb.com/xml/find?json=1&nr=1&tt=on&q={movie_title}'
url = pattern.format(movie_title=urllib.parse.quote_plus(title))
r = requests.get(url)
res = r.json()
# sections in descending order or preference
for section in ['popular','exact','substring']:
key = 'title_' + section
if key in res:
return res[key][0]['id']
"""strip_year('Toy Story (1995)')"""
"""imdb_id_from_title('Toy Story (1995)')"""
'Toy Story'
Helper functions: Display movie posters
CONFIG_PATTERN = 'http://api.themoviedb.org/3/configuration?api_key={key}'
KEY = 'cd3fbcc9cb70affd9a68b0951b3d0997'
def size_str_to_int(x):
return float("inf") if x == 'original' else int(x[1:])
url = CONFIG_PATTERN.format(key=KEY)
r = requests.get(url)
config = r.json()
base_url = config['images']['base_url']
sizes = config['images']['poster_sizes']
max_size = max(sizes, key=size_str_to_int)
def get_poster_url(movie_name):
imdb_id = imdb_id_from_title(movie_name)
IMG_PATTERN = 'http://api.themoviedb.org/3/movie/{imdbid}/images?api_key={key}'
r = requests.get(IMG_PATTERN.format(key=KEY,imdbid=imdb_id))
api_response = r.json()
poster = api_response['posters'][0] # Get the best rated poster
rel_path = poster['file_path']
poster_url = "{0}{1}{2}".format(base_url, max_size, rel_path)
return poster_url
"""movie_name = 'Toy Story'
get_poster(movie_name)"""
"movie_name = 'Toy Story' \nget_poster(movie_name)"
Display Recommendations
# see https://stackoverflow.com/questions/19471814/display-multiple-images-in-one-ipython-notebook-cell
def make_html(image_url):
return "<img style='width: 120px; margin: 0px; float: left; \
border: 1px solid black;' src='%s' />" % image_url
def display_rec_movies(movieid, num_recs):
movie_name = get_movie_name(df_movies, movieid)
poster_url = get_poster_url(movie_name)
print ('Input movie:')
display(HTML(make_html(poster_url)))
recommended_item_names = get_similar_items(movieid, num_recs)
print ('Recommended movies: ')
imagesList = ''
for movie_name in recommended_item_names:
poster_url = get_poster_url(movie_name)
imagesList += make_html(poster_url)
display(HTML(imagesList))
"""movieid = 1
display_rec_movies(movieid)"""
'movieid = 1\ndisplay_rec_movies(movieid)'
movieids = [1, 10, 50]
num_recs = 5
for movieid in movieids:
display_rec_movies(movieid, num_recs)
print ('\n\n')
Input movie:
Recommended movies:
Input movie:
Recommended movies:
Input movie:
Recommended movies:
