workshops_recommender_systems/P2. Evaluation.ipynb

Prepare test set

import pandas as pd
import numpy as np
import scipy.sparse as sparse
from collections import defaultdict
from itertools import chain
import random
from tqdm import tqdm

# In evaluation we do not load train set - it is not needed
test=pd.read_csv('./Datasets/ml-100k/test.csv', sep='\t', header=None)
test.columns=['user', 'item', 'rating', 'timestamp']

test['user_code'] = test['user'].astype("category").cat.codes
test['item_code'] = test['item'].astype("category").cat.codes

user_code_id = dict(enumerate(test['user'].astype("category").cat.categories))
user_id_code = dict((v, k) for k, v in user_code_id.items())
item_code_id = dict(enumerate(test['item'].astype("category").cat.categories))
item_id_code = dict((v, k) for k, v in item_code_id.items())

test_ui = sparse.csr_matrix((test['rating'], (test['user_code'], test['item_code'])))

Estimations metrics

estimations_df=pd.read_csv('Recommendations generated/ml-100k/Ready_Baseline_estimations.csv', header=None)
estimations_df.columns=['user', 'item' ,'score']

estimations_df['user_code']=[user_id_code[user] for user in estimations_df['user']]
estimations_df['item_code']=[item_id_code[item] for item in estimations_df['item']]
estimations=sparse.csr_matrix((estimations_df['score'], (estimations_df['user_code'], estimations_df['item_code'])), shape=test_ui.shape)

def estimations_metrics(test_ui, estimations):
    result=[]

    RMSE=(np.sum((estimations.data-test_ui.data)**2)/estimations.nnz)**(1/2)
    result.append(['RMSE', RMSE])

    MAE=np.sum(abs(estimations.data-test_ui.data))/estimations.nnz
    result.append(['MAE', MAE])
    
    df_result=(pd.DataFrame(list(zip(*result))[1])).T
    df_result.columns=list(zip(*result))[0]
    return df_result

# in case of error (in the laboratories) you might have to switch to the other version of pandas
# try !pip3 install pandas=='1.0.3' (or pip if you use python 2) and restart the kernel

estimations_metrics(test_ui, estimations)

	RMSE	MAE
0	0.949459	0.752487

Ranking metrics

import numpy as np
reco = np.loadtxt('Recommendations generated/ml-100k/Ready_Baseline_reco.csv', delimiter=',')
# Let's ignore scores - they are not used in evaluation: 
users=reco[:,:1]
items=reco[:,1::2]
# Let's use inner ids instead of real ones
users=np.vectorize(lambda x: user_id_code.setdefault(x, -1))(users)
items=np.vectorize(lambda x: item_id_code.setdefault(x, -1))(items) # maybe items we recommend are not in test set
# Let's put them into one array
reco=np.concatenate((users, items), axis=1)
reco

array([[663, 475,  62, ..., 472, 269, 503],
       [ 48, 313, 475, ..., 591, 175, 466],
       [351, 313, 475, ..., 591, 175, 466],
       ...,
       [259, 313, 475, ...,  11, 591, 175],
       [ 33, 313, 475, ...,  11, 591, 175],
       [ 77, 313, 475, ...,  11, 591, 175]])

def ranking_metrics(test_ui, reco, super_reactions=[], topK=10):
    
    nb_items=test_ui.shape[1]
    relevant_users, super_relevant_users, prec, rec, F_1, F_05, prec_super, rec_super, ndcg, mAP, MRR, LAUC, HR=\
    0,0,0,0,0,0,0,0,0,0,0,0,0
    
    cg = (1.0 / np.log2(np.arange(2, topK + 2)))
    cg_sum = np.cumsum(cg)
    
    for (nb_user, user) in tqdm(enumerate(reco[:,0])):
        u_rated_items=test_ui.indices[test_ui.indptr[user]:test_ui.indptr[user+1]]
        nb_u_rated_items=len(u_rated_items)
        if nb_u_rated_items>0: # skip users with no items in test set (still possible that there will be no super items)
            relevant_users+=1
            
            u_super_items=u_rated_items[np.vectorize(lambda x: x in super_reactions)\
            (test_ui.data[test_ui.indptr[user]:test_ui.indptr[user+1]])]
            # more natural seems u_super_items=[item for item in u_rated_items if test_ui[user,item] in super_reactions]
            # but accesing test_ui[user,item] is expensive -we should avoid doing it
            if len(u_super_items)>0:
                super_relevant_users+=1
            
            user_successes=np.zeros(topK)
            nb_user_successes=0
            user_super_successes=np.zeros(topK)
            nb_user_super_successes=0
            
            # evaluation
            for (item_position,item) in enumerate(reco[nb_user,1:topK+1]):
                if item in u_rated_items:
                    user_successes[item_position]=1
                    nb_user_successes+=1
                    if item in u_super_items:
                        user_super_successes[item_position]=1
                        nb_user_super_successes+=1
                        
            prec_u=nb_user_successes/topK 
            prec+=prec_u
            
            rec_u=nb_user_successes/nb_u_rated_items
            rec+=rec_u
            
            F_1+=2*(prec_u*rec_u)/(prec_u+rec_u) if prec_u+rec_u>0 else 0
            F_05+=(0.5**2+1)*(prec_u*rec_u)/(0.5**2*prec_u+rec_u) if prec_u+rec_u>0 else 0
            
            prec_super+=nb_user_super_successes/topK
            rec_super+=nb_user_super_successes/max(len(u_super_items),1) # to set 0 if no super items
            ndcg+=np.dot(user_successes,cg)/cg_sum[min(topK, nb_u_rated_items)-1]
            
            cumsum_successes=np.cumsum(user_successes)
            mAP+=np.dot(cumsum_successes/np.arange(1,topK+1), user_successes)/min(topK, nb_u_rated_items)
            MRR+=1/(user_successes.nonzero()[0][0]+1) if user_successes.nonzero()[0].size>0 else 0
            LAUC+=(np.dot(cumsum_successes, 1-user_successes)+\
            (nb_user_successes+nb_u_rated_items)/2*((nb_items-nb_u_rated_items)-(topK-nb_user_successes)))/\
            ((nb_items-nb_u_rated_items)*nb_u_rated_items)
            
            HR+=nb_user_successes>0
            
            
    result=[]
    result.append(('precision', prec/relevant_users))
    result.append(('recall', rec/relevant_users))
    result.append(('F_1', F_1/relevant_users))
    result.append(('F_05', F_05/relevant_users))
    result.append(('precision_super', prec_super/super_relevant_users))
    result.append(('recall_super', rec_super/super_relevant_users))
    result.append(('NDCG', ndcg/relevant_users))
    result.append(('mAP', mAP/relevant_users))
    result.append(('MRR', MRR/relevant_users))
    result.append(('LAUC', LAUC/relevant_users))
    result.append(('HR', HR/relevant_users))

    df_result=(pd.DataFrame(list(zip(*result))[1])).T
    df_result.columns=list(zip(*result))[0]
    return df_result

ranking_metrics(test_ui, reco, super_reactions=[4,5], topK=10)

943it [00:00, 9409.37it/s]

	precision	recall	F_1	F_05	precision_super	recall_super	NDCG	mAP	MRR	LAUC	HR
0	0.09141	0.037652	0.04603	0.061286	0.079614	0.056463	0.095957	0.043178	0.198193	0.515501	0.437964

Diversity metrics

def diversity_metrics(test_ui, reco, topK=10):
    
    frequencies=defaultdict(int)
    
    # let's assign 0 to all items in test set
    for item in list(set(test_ui.indices)):
        frequencies[item]=0
        
    # counting frequencies
    for item in reco[:,1:].flat:
        frequencies[item]+=1
        
    nb_reco_outside_test=frequencies[-1]
    del frequencies[-1]
    
    frequencies=np.array(list(frequencies.values()))
                         
    nb_rec_items=len(frequencies[frequencies>0])
    nb_reco_inside_test=np.sum(frequencies)
                         
    frequencies=frequencies/np.sum(frequencies)
    frequencies=np.sort(frequencies)
    
    with np.errstate(divide='ignore'): # let's put zeros put items with 0 frequency and ignore division warning
        log_frequencies=np.nan_to_num(np.log(frequencies), posinf=0, neginf=0)
                         
    result=[]
    result.append(('Reco in test', nb_reco_inside_test/(nb_reco_inside_test+nb_reco_outside_test)))
    result.append(('Test coverage', nb_rec_items/test_ui.shape[1]))
    result.append(('Shannon', -np.dot(frequencies, log_frequencies)))
    result.append(('Gini', np.dot(frequencies, np.arange(1-len(frequencies), len(frequencies), 2))/(len(frequencies)-1)))
    
    df_result=(pd.DataFrame(list(zip(*result))[1])).T
    df_result.columns=list(zip(*result))[0]
    return df_result

# in case of errors try !pip3 install numpy==1.18.4 (or pip if you use python 2) and restart the kernel

import evaluation_measures as ev
import imp
imp.reload(ev)

x=diversity_metrics(test_ui, reco, topK=10)
x

	Reco in test	Test coverage	Shannon	Gini
0	1.0	0.033911	2.836513	0.991139

To be used in other notebooks

import evaluation_measures as ev
import imp
imp.reload(ev)

estimations_df=pd.read_csv('Recommendations generated/ml-100k/Ready_Baseline_estimations.csv', header=None)
reco=np.loadtxt('Recommendations generated/ml-100k/Ready_Baseline_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])
#also you can just type ev.evaluate_all(estimations_df, reco) - I put above values as default

943it [00:00, 8290.29it/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	0.949459	0.752487	0.09141	0.037652	0.04603	0.061286	0.079614	0.056463	0.095957	0.043178	0.198193	0.515501	0.437964	1.0	0.033911	2.836513	0.991139

import evaluation_measures as ev
import imp
imp.reload(ev)

dir_path="Recommendations generated/ml-100k/"
super_reactions=[4,5]
test=pd.read_csv('./Datasets/ml-100k/test.csv', sep='\t', header=None)

df=ev.evaluate_all(test, dir_path, super_reactions)
#also you can just type ev.evaluate_all() - I put above values as default

943it [00:00, 9608.49it/s]
943it [00:00, 9837.12it/s]
943it [00:00, 10292.46it/s]
943it [00:00, 9906.94it/s]
943it [00:00, 9162.09it/s]

df.iloc[:,:9]

	Model	RMSE	MAE	precision	recall	F_1	F_05	precision_super	recall_super
0	Self_TopPop	2.508258	2.217909	0.188865	0.116919	0.118732	0.141584	0.130472	0.137473
0	Ready_Baseline	0.949459	0.752487	0.091410	0.037652	0.046030	0.061286	0.079614	0.056463
0	Self_GlobalAvg	1.125760	0.943534	0.061188	0.025968	0.031383	0.041343	0.040558	0.032107
0	Ready_Random	1.518964	1.222159	0.046554	0.020603	0.023679	0.031216	0.028970	0.021179
0	Self_BaselineUI	0.967585	0.762740	0.000954	0.000170	0.000278	0.000463	0.000644	0.000189

df.iloc[:,np.append(0,np.arange(9, df.shape[1]))]

	Model	NDCG	mAP	MRR	LAUC	HR	Reco in test	Test coverage	Shannon	Gini
0	Self_TopPop	0.214651	0.111707	0.400939	0.555546	0.765642	1.000000	0.038961	3.159079	0.987317
0	Ready_Baseline	0.095957	0.043178	0.198193	0.515501	0.437964	1.000000	0.033911	2.836513	0.991139
0	Self_GlobalAvg	0.067695	0.027470	0.171187	0.509546	0.384942	1.000000	0.025974	2.711772	0.992003
0	Ready_Random	0.050489	0.019185	0.123856	0.506812	0.322375	0.987805	0.184704	5.103172	0.906873
0	Self_BaselineUI	0.000752	0.000168	0.001677	0.496424	0.009544	0.600530	0.005051	1.803126	0.996380

Check metrics on toy dataset

import evaluation_measures as ev
import imp
import helpers
imp.reload(ev)

dir_path="Recommendations generated/toy-example/"
super_reactions=[4,5]
test=pd.read_csv('./Datasets/toy-example/test.csv', sep='\t', header=None)

display(ev.evaluate_all(test, dir_path, super_reactions, topK=3))
#also you can just type ev.evaluate_all() - I put above values as default

toy_train_read=pd.read_csv('./Datasets/toy-example/train.csv', sep='\t', header=None, names=['user', 'item', 'rating', 'timestamp'])
toy_test_read=pd.read_csv('./Datasets/toy-example/test.csv', sep='\t', header=None, names=['user', 'item', 'rating', 'timestamp'])
reco=pd.read_csv('Recommendations generated/toy-example/Self_BaselineUI_reco.csv', header=None)
estimations=pd.read_csv('Recommendations generated/toy-example/Self_BaselineUI_estimations.csv', names=['user', 'item', 'est_score'])
toy_train_ui, toy_test_ui, toy_user_code_id, toy_user_id_code, \
toy_item_code_id, toy_item_id_code = helpers.data_to_csr(toy_train_read, toy_test_read)

print('Training data:')
display(toy_train_ui.todense())

print('Test data:')
display(toy_test_ui.todense())

print('Recommendations:')
display(reco)

print('Estimations:')
display(estimations)

3it [00:00, 1651.52it/s]

	Model	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	Self_BaselineUI	1.612452	1.4	0.444444	0.888889	0.555556	0.478632	0.333333	0.75	0.676907	0.574074	0.611111	0.638889	1.0	0.888889	0.8	1.386294	0.25

Training data:

matrix([[3, 4, 0, 0, 5, 0, 0, 4],
        [0, 1, 2, 3, 0, 0, 0, 0],
        [0, 0, 0, 5, 0, 3, 4, 0]], dtype=int64)

Test data:

matrix([[0, 0, 0, 0, 0, 0, 3, 0],
        [0, 0, 0, 0, 5, 0, 0, 0],
        [5, 0, 4, 0, 0, 0, 0, 2]], dtype=int64)

Recommendations:

	0	1	2	3	4	5	6
0	0	30	5.0	20	4.0	60	4.0
1	10	40	3.0	60	2.0	70	2.0
2	20	40	5.0	20	4.0	70	4.0

Estimations:

	user	item	est_score
0	0	60	4.0
1	10	40	3.0
2	20	0	3.0
3	20	20	4.0
4	20	70	4.0

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')

print('Here is what user rated high:')
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_BaselineUI_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')

print('Here is what we recommend:')
recommended_content[recommended_content['user']==user][['user', 'rec_nb', 'title', 'genres']].sort_values(by='rec_nb')

Here is what user rated high:

	user	rating	title	genres
24754	105	5	Gattaca (1997)	Drama, Sci-Fi, Thriller
37073	105	5	Contact (1997)	Drama, Sci-Fi
40592	105	5	Titanic (1997)	Action, Drama, Romance
46032	105	5	L.A. Confidential (1997)	Crime, Film-Noir, Mystery, Thriller
17916	105	4	English Patient, The (1996)	Drama, Romance, War
23767	105	4	Chasing Amy (1997)	Drama, Romance
47003	105	4	Cop Land (1997)	Crime, Drama, Mystery
48891	105	4	Lost Highway (1997)	Mystery
67355	105	4	Good Will Hunting (1997)	Drama
54121	105	3	Boogie Nights (1997)	Drama
55439	105	3	Seven Years in Tibet (1997)	Drama, War
69687	105	3	Wag the Dog (1997)	Comedy, Drama
7321	105	2	Saint, The (1997)	Action, Romance, Thriller
45056	105	2	Tomorrow Never Dies (1997)	Action, Romance, Thriller
64052	105	2	Alien: Resurrection (1997)	Action, Horror, Sci-Fi

Here is what we recommend:

	user	rec_nb	title	genres
103	105.0	1	Great Day in Harlem, A (1994)	Documentary
1046	105.0	2	Tough and Deadly (1995)	Action, Drama, Thriller
1988	105.0	3	Aiqing wansui (1994)	Drama
2930	105.0	4	Delta of Venus (1994)	Drama
3872	105.0	5	Someone Else's America (1995)	Drama
4814	105.0	6	Saint of Fort Washington, The (1993)	Drama
5755	105.0	7	Celestial Clockwork (1994)	Comedy
6698	105.0	8	Some Mother's Son (1996)	Drama
8592	105.0	9	Maya Lin: A Strong Clear Vision (1994)	Documentary
7638	105.0	10	Prefontaine (1997)	Drama

project task 3: implement some other evaluation measure

# it may be your idea, modification of what we have already implemented 
# (for example Hit2 rate which would count as a success users whoreceived at least 2 relevant recommendations) 
# or something well-known
# expected output: modification of evaluation_measures.py such that evaluate_all will also display your measure

dir_path="Recommendations generated/ml-100k/"
super_reactions=[4,5]
test=pd.read_csv('./Datasets/ml-100k/test.csv', sep='\t', header=None)

ev.evaluate_all(test, dir_path, super_reactions)

943it [00:00, 9023.32it/s]
943it [00:00, 9004.98it/s]
943it [00:00, 8532.40it/s]
943it [00:00, 8974.06it/s]
943it [00:00, 8818.01it/s]

	Model	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	Self_TopPop	2.508258	2.217909	0.188865	0.116919	0.118732	0.141584	0.130472	0.137473	0.214651	0.111707	0.400939	0.555546	0.765642	1.000000	0.038961	3.159079	0.987317
0	Ready_Baseline	0.949459	0.752487	0.091410	0.037652	0.046030	0.061286	0.079614	0.056463	0.095957	0.043178	0.198193	0.515501	0.437964	1.000000	0.033911	2.836513	0.991139
0	Self_GlobalAvg	1.125760	0.943534	0.061188	0.025968	0.031383	0.041343	0.040558	0.032107	0.067695	0.027470	0.171187	0.509546	0.384942	1.000000	0.025974	2.711772	0.992003
0	Ready_Random	1.518964	1.222159	0.046554	0.020603	0.023679	0.031216	0.028970	0.021179	0.050489	0.019185	0.123856	0.506812	0.322375	0.987805	0.184704	5.103172	0.906873
0	Self_BaselineUI	0.967585	0.762740	0.000954	0.000170	0.000278	0.000463	0.000644	0.000189	0.000752	0.000168	0.001677	0.496424	0.009544	0.600530	0.005051	1.803126	0.996380