s406917/workshops_recommender_systems
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity
45e88a01aa
workshops_recommender_systems/P5. Graph-based.ipynb
Robert 45e88a01aa Changes in slides
2020-05-21 22:53:34 +02:00

554 KiB
Raw Blame History

Self made RP3-beta

import helpers
import pandas as pd
import numpy as np
import scipy.sparse as sparse
from collections import defaultdict
from itertools import chain
import random
import time
import matplotlib.pyplot as plt

train_read=pd.read_csv('./Datasets/ml-100k/train.csv', sep='\t', header=None)
test_read=pd.read_csv('./Datasets/ml-100k/test.csv', sep='\t', header=None)
train_ui, test_ui, user_code_id, user_id_code, item_code_id, item_id_code = helpers.data_to_csr(train_read, test_read)
class RP3Beta():
    def fit(self, train_ui, alpha, beta):
        """We weight our edges by user's explicit ratings so if user rated movie high we'll follow that path
        with higher probability."""
        self.train_ui=train_ui
        self.train_iu=train_ui.transpose()
        
        self.alpha = alpha
        self.beta = beta
        
        # Define Pui 
        Pui=sparse.csr_matrix(self.train_ui/self.train_ui.sum(axis=1))
        
        # Define Piu
        to_divide=np.vectorize(lambda x: x if x>0 else 1)(self.train_iu.sum(axis=1)) # to avoid dividing by zero
        Piu=sparse.csr_matrix(self.train_iu/to_divide)
        item_orders=(self.train_ui>0).sum(axis=0)
        
        Pui = Pui.power(self.alpha)
        Piu = Piu.power(self.alpha)

        P3=Pui*Piu*Pui
        
        P3/=np.power(np.vectorize(lambda x: x if x>0 else 1)(item_orders), self.beta)
        
        self.estimations=np.array(P3)
    
    def recommend(self, user_code_id, item_code_id, topK=10):
        
        top_k = defaultdict(list)
        for nb_user, user in enumerate(self.estimations):
            
            user_rated=self.train_ui.indices[self.train_ui.indptr[nb_user]:self.train_ui.indptr[nb_user+1]]
            for item, score in enumerate(user):
                if item not in user_rated and not np.isnan(score):
                    top_k[user_code_id[nb_user]].append((item_code_id[item], score))
        result=[]
        # Let's choose k best items in the format: (user, item1, score1, item2, score2, ...)
        for uid, item_scores in top_k.items():
            item_scores.sort(key=lambda x: x[1], reverse=True)
            result.append([uid]+list(chain(*item_scores[:topK])))
        return result
    
    def estimate(self, user_code_id, item_code_id, test_ui):
        result=[]
        for user, item in zip(*test_ui.nonzero()):
            result.append([user_code_id[user], item_code_id[item], 
                           self.estimations[user,item] if not np.isnan(self.estimations[user,item]) else 1])
        return result
model=RP3Beta()
model.fit(train_ui, alpha=1, beta=0)
top_n=pd.DataFrame(model.recommend(user_code_id, item_code_id, topK=10))

top_n.to_csv('Recommendations generated/ml-100k/Self_P3_reco.csv', index=False, header=False)

estimations=pd.DataFrame(model.estimate(user_code_id, item_code_id, test_ui))
estimations.to_csv('Recommendations generated/ml-100k/Self_P3_estimations.csv', index=False, header=False)
import evaluation_measures as ev
estimations_df=pd.read_csv('Recommendations generated/ml-100k/Self_P3_estimations.csv', header=None)
reco=np.loadtxt('Recommendations generated/ml-100k/Self_P3_reco.csv', delimiter=',')

ev.evaluate(test=pd.read_csv('./Datasets/ml-100k/test.csv', sep='\t', header=None),
            estimations_df=estimations_df, 
            reco=reco,
            super_reactions=[4,5])
943it [00:00, 6131.26it/s]
RMSE MAE precision recall F_1 F_05 precision_super recall_super NDCG mAP MRR LAUC HR Reco in test Test coverage Shannon Gini
0 3.702446 3.527273 0.282185 0.192092 0.186749 0.21698 0.204185 0.240096 0.339114 0.204905 0.572157 0.593544 0.875928 1.0 0.077201 3.875892 0.974947

Let's check hiperparameters

Alpha
from tqdm import tqdm
result=[]
for alpha in tqdm([round(i,1) for i in np.arange(0.2,1.6001,0.2)]):
    model=RP3Beta()
    model.fit(train_ui, alpha=alpha, beta=0)
    reco=pd.DataFrame(model.recommend(user_code_id, item_code_id, topK=10))
    estimations_df=pd.DataFrame(model.estimate(user_code_id, item_code_id, test_ui))
    to_append=ev.evaluate(test=pd.read_csv('./Datasets/ml-100k/test.csv', sep='\t', header=None),
            estimations_df=estimations_df, 
            reco=np.array(reco),
            super_reactions=[4,5])
    to_append.insert(0, "Alpha", alpha)
    result.append(to_append)
    
result=pd.concat(result)
result
  0%|          | 0/8 [00:00<?, ?it/s]
0it [00:00, ?it/s]
943it [00:00, 5033.65it/s]
 12%|█▎        | 1/8 [00:15<01:47, 15.30s/it]
0it [00:00, ?it/s]
943it [00:00, 5085.50it/s]
 25%|██▌       | 2/8 [00:31<01:33, 15.53s/it]
0it [00:00, ?it/s]
943it [00:00, 5274.96it/s]
 38%|███▊      | 3/8 [00:46<01:17, 15.56s/it]
0it [00:00, ?it/s]
943it [00:00, 4951.74it/s]
 50%|█████     | 4/8 [01:02<01:02, 15.64s/it]
0it [00:00, ?it/s]
943it [00:00, 5223.55it/s]
 62%|██████▎   | 5/8 [01:18<00:47, 15.72s/it]
0it [00:00, ?it/s]
943it [00:00, 5301.89it/s]
 75%|███████▌  | 6/8 [01:34<00:31, 15.64s/it]
0it [00:00, ?it/s]
943it [00:00, 5440.58it/s]
 88%|████████▊ | 7/8 [01:49<00:15, 15.65s/it]
0it [00:00, ?it/s]
943it [00:00, 5136.56it/s]
100%|██████████| 8/8 [02:05<00:00, 15.65s/it]
Alpha RMSE MAE precision recall F_1 F_05 precision_super recall_super NDCG mAP MRR LAUC HR Reco in test Test coverage Shannon Gini
0 0.2 268.177832 211.732649 0.262672 0.166858 0.166277 0.197184 0.187661 0.203252 0.320910 0.196132 0.563378 0.580866 0.850477 1.000000 0.060606 3.669627 0.979636
0 0.4 10.546689 7.792373 0.268505 0.172669 0.171569 0.202643 0.192489 0.212653 0.326760 0.200172 0.565148 0.583801 0.854719 1.000000 0.064214 3.726996 0.978426
0 0.6 3.143988 2.948790 0.274655 0.180502 0.177820 0.208730 0.198176 0.222746 0.332872 0.203290 0.568872 0.587738 0.870626 1.000000 0.065657 3.785282 0.977090
0 0.8 3.670728 3.495735 0.281972 0.189868 0.185300 0.216071 0.203541 0.236751 0.339867 0.206688 0.573729 0.592432 0.874867 1.000000 0.070707 3.832415 0.975998
0 1.0 3.702446 3.527273 0.282185 0.192092 0.186749 0.216980 0.204185 0.240096 0.339114 0.204905 0.572157 0.593544 0.875928 1.000000 0.077201 3.875892 0.974947
0 1.2 3.704441 3.529251 0.280912 0.193633 0.187311 0.216872 0.203004 0.240588 0.338049 0.203453 0.571830 0.594313 0.883351 1.000000 0.085859 3.910718 0.974073
0 1.4 3.704580 3.529388 0.273595 0.190651 0.183874 0.212183 0.199464 0.239118 0.329550 0.195433 0.566171 0.592793 0.871686 1.000000 0.107504 3.961915 0.972674
0 1.6 3.704591 3.529399 0.263097 0.186255 0.178709 0.205170 0.191094 0.232920 0.317439 0.184917 0.552349 0.590545 0.868505 0.999576 0.156566 4.060156 0.969203 
metrics=list(result.columns[[i not in ['Alpha'] for i in result.columns]])

charts_per_row=6
charts_per_column=3

fig, axes = plt.subplots(nrows=charts_per_row, ncols=charts_per_column,figsize=(18, 7*charts_per_row ))
import itertools
to_iter=[i for i in itertools.product(range(charts_per_row), range(charts_per_column))]

for i in range(len(metrics)):
    df=result[['Alpha', metrics[i]]]
    df.plot(ax=axes[to_iter[i]], title=metrics[i], x=0, y=1)
Beta
from tqdm import tqdm
result=[]
for beta in tqdm([round(i,1) for i in np.arange(0,1,0.1)]):
    model=RP3Beta()
    model.fit(train_ui, alpha=1, beta=beta)
    reco=pd.DataFrame(model.recommend(user_code_id, item_code_id, topK=10))
    estimations_df=pd.DataFrame(model.estimate(user_code_id, item_code_id, test_ui))
    to_append=ev.evaluate(test=pd.read_csv('./Datasets/ml-100k/test.csv', sep='\t', header=None),
            estimations_df=estimations_df, 
            reco=np.array(reco),
            super_reactions=[4,5])
    to_append.insert(0, "Beta", beta)
    result.append(to_append)
    
result=pd.concat(result)
result
  0%|          | 0/10 [00:00<?, ?it/s]
0it [00:00, ?it/s]
943it [00:00, 5006.82it/s]
 10%|█         | 1/10 [00:15<02:18, 15.43s/it]
0it [00:00, ?it/s]
943it [00:00, 5039.36it/s]
 20%|██        | 2/10 [00:30<02:03, 15.45s/it]
0it [00:00, ?it/s]
943it [00:00, 5447.56it/s]
 30%|███       | 3/10 [00:46<01:48, 15.47s/it]
0it [00:00, ?it/s]
943it [00:00, 5024.04it/s]
 40%|████      | 4/10 [01:01<01:32, 15.42s/it]
0it [00:00, ?it/s]
943it [00:00, 4743.56it/s]
 50%|█████     | 5/10 [01:17<01:17, 15.55s/it]
0it [00:00, ?it/s]
943it [00:00, 5014.86it/s]
 60%|██████    | 6/10 [01:33<01:02, 15.57s/it]
0it [00:00, ?it/s]
943it [00:00, 4970.05it/s]
 70%|███████   | 7/10 [01:48<00:46, 15.49s/it]
0it [00:00, ?it/s]
943it [00:00, 5221.08it/s]
 80%|████████  | 8/10 [02:04<00:31, 15.55s/it]
0it [00:00, ?it/s]
943it [00:00, 5267.88it/s]
 90%|█████████ | 9/10 [02:19<00:15, 15.55s/it]
0it [00:00, ?it/s]
943it [00:00, 5219.38it/s]
100%|██████████| 10/10 [02:35<00:00, 15.54s/it]
Beta RMSE MAE precision recall F_1 F_05 precision_super recall_super NDCG mAP MRR LAUC HR Reco in test Test coverage Shannon Gini
0 0.0 3.702446 3.527273 0.282185 0.192092 0.186749 0.216980 0.204185 0.240096 0.339114 0.204905 0.572157 0.593544 0.875928 1.000000 0.077201 3.875892 0.974947
0 0.1 3.703312 3.528128 0.290138 0.197597 0.192259 0.223336 0.210944 0.246153 0.347768 0.212034 0.581038 0.596328 0.884411 1.000000 0.085137 3.957416 0.972784
0 0.2 3.703825 3.528636 0.297137 0.201202 0.196067 0.228169 0.218026 0.252767 0.355655 0.219909 0.588904 0.598160 0.886532 1.000000 0.094517 4.053212 0.969980
0 0.3 3.704130 3.528939 0.303499 0.204749 0.199901 0.232829 0.225107 0.260797 0.363757 0.226825 0.599969 0.599964 0.888653 1.000000 0.105339 4.147779 0.966948
0 0.4 3.704313 3.529120 0.308908 0.208811 0.203854 0.237241 0.229614 0.266918 0.370758 0.232673 0.609385 0.602014 0.895016 0.999894 0.132035 4.259682 0.962989
0 0.5 3.704422 3.529229 0.314316 0.211411 0.206768 0.240986 0.237124 0.273416 0.378307 0.239297 0.622792 0.603327 0.903499 0.999046 0.168831 4.411281 0.956648
0 0.6 3.704488 3.529295 0.314634 0.206209 0.204818 0.240159 0.242489 0.273850 0.376438 0.238428 0.622042 0.600721 0.897137 0.996394 0.212843 4.621938 0.945932
0 0.7 3.704528 3.529335 0.304136 0.187298 0.191990 0.228749 0.238305 0.256201 0.358807 0.226808 0.593897 0.591207 0.868505 0.983033 0.256854 4.898568 0.928065
0 0.8 3.704552 3.529360 0.266384 0.147571 0.158660 0.194838 0.214485 0.209336 0.299850 0.184356 0.492852 0.571152 0.803818 0.936373 0.341270 5.257397 0.895882
0 0.9 3.704567 3.529375 0.162354 0.076967 0.089233 0.114583 0.134657 0.113253 0.160868 0.085486 0.243590 0.535405 0.580064 0.800106 0.415584 5.563910 0.857396 
### import matplotlib.pyplot as plt

metrics=list(result.columns[[i not in ['Beta'] for i in result.columns]])

charts_per_row=6
charts_per_column=3

fig, axes = plt.subplots(nrows=charts_per_row, ncols=charts_per_column,figsize=(18, 7*charts_per_row ))
import itertools
to_iter=[i for i in itertools.product(range(charts_per_row), range(charts_per_column))]

for i in range(len(metrics)):
    df=result[['Beta', metrics[i]]]
    df.plot(ax=axes[to_iter[i]], title=metrics[i], x=0, y=1)

Check sample recommendations

train=pd.read_csv('./Datasets/ml-100k/train.csv', sep='\t', header=None, names=['user', 'item', 'rating', 'timestamp'])
items=pd.read_csv('./Datasets/ml-100k/movies.csv')

user=random.choice(list(set(train['user'])))

train_content=pd.merge(train, items, left_on='item', right_on='id')
display(train_content[train_content['user']==user][['user', 'rating', 'title', 'genres']]\
        .sort_values(by='rating', ascending=False)[:15])

reco = np.loadtxt('Recommendations generated/ml-100k/Self_P3_reco.csv', delimiter=',')
items=pd.read_csv('./Datasets/ml-100k/movies.csv')

# Let's ignore scores - they are not used in evaluation: 
reco_users=reco[:,:1]
reco_items=reco[:,1::2]
# Let's put them into one array
reco=np.concatenate((reco_users, reco_items), axis=1)

# Let's rebuild it user-item dataframe
recommended=[]
for row in reco:
    for rec_nb, entry in enumerate(row[1:]):
        recommended.append((row[0], rec_nb+1, entry))
recommended=pd.DataFrame(recommended, columns=['user','rec_nb', 'item'])

recommended_content=pd.merge(recommended, items, left_on='item', right_on='id')
recommended_content[recommended_content['user']==user][['user', 'rec_nb', 'title', 'genres']].sort_values(by='rec_nb')
user rating title genres
24762 813 5 Gattaca (1997) Drama, Sci-Fi, Thriller
45157 813 5 Tomorrow Never Dies (1997) Action, Romance, Thriller
57294 813 4 Evita (1996) Drama, Musical
7820 813 4 Devil's Advocate, The (1997) Crime, Horror, Mystery, Thriller
74585 813 4 Mortal Kombat: Annihilation (1997) Action, Adventure
26242 813 4 Air Force One (1997) Action, Thriller
43721 813 4 Amistad (1997) Drama
55467 813 4 Seven Years in Tibet (1997) Drama, War
7505 813 4 Starship Troopers (1997) Action, Adventure, Sci-Fi, War
3996 813 3 G.I. Jane (1997) Action, Drama, War
77862 813 3 Steel (1997) Action
63929 813 3 Anastasia (1997) Animation, Children's, Musical
56993 813 3 Jungle2Jungle (1997) Children's, Comedy
25979 813 3 Dead Man Walking (1995) Drama
78546 813 3 For Richer or Poorer (1997) Comedy
user rec_nb title genres
5999 813.0 1 Contact (1997) Drama, Sci-Fi
4994 813.0 2 Titanic (1997) Action, Drama, Romance
1953 813.0 3 Scream (1996) Horror, Thriller
1282 813.0 4 English Patient, The (1996) Drama, Romance, War
3705 813.0 5 Conspiracy Theory (1997) Action, Mystery, Romance, Thriller
4051 813.0 6 Game, The (1997) Mystery, Thriller
6230 813.0 7 L.A. Confidential (1997) Crime, Film-Noir, Mystery, Thriller
5213 813.0 8 Full Monty, The (1997) Comedy
3922 813.0 9 Saint, The (1997) Action, Romance, Thriller
6331 813.0 10 Good Will Hunting (1997) Drama

project task 6: generate recommendations of RP3Beta for hiperparameters found to optimize recall

# use better values than (1,0) for alpha and beta
# if you want you can also modify the model to consider different weights (we took as weights user ratings, maybe take ones or squares of ratings instead)
# save the outptut in 'Recommendations generated/ml-100k/Self_RP3Beta_estimations.csv'
# and 'Recommendations generated/ml-100k/Self_RP3Beta_reco.csv'

project task 7 (optional): implement graph-based model of your choice

# for example change length of paths in RP3beta
# save the outptut in 'Recommendations generated/ml-100k/Self_GraphTask_estimations.csv'
# and 'Recommendations generated/ml-100k/Self_GraphTask_reco.csv'