notebook.community

Edit and run

New song recommendation baselines





In [ ]:



    


%matplotlib inline
%load_ext autoreload
%autoreload 2

import os, sys, time, gzip
import pickle as pkl
import numpy as np
import pandas as pd
from scipy.sparse import lil_matrix, issparse, hstack, vstack

import matplotlib.pyplot as plt
import seaborn as sns



    











In [ ]:



    


from models import MTC
from sklearn.linear_model import LogisticRegression
# from tools import calc_RPrecision_HitRate
from tools import calc_metrics



    











In [ ]:



    


TOPs = [5, 10, 20, 30, 50, 100, 200, 300, 500, 1000]



    











In [ ]:



    


datasets = ['aotm2011', '30music']



    











In [ ]:



    


dix = 1
dataset_name = datasets[dix]
dataset_name



    











In [ ]:



    


data_dir = 'data/%s/setting1' % dataset_name
Y_trndev = pkl.load(gzip.open(os.path.join(data_dir, 'Y_train_dev.pkl.gz'), 'rb'))
Y_test = pkl.load(gzip.open(os.path.join(data_dir, 'Y_test.pkl.gz'), 'rb'))
song2pop = pkl.load(gzip.open('data/%s/setting2/song2pop.pkl.gz' % dataset_name, 'rb'))
songsets = pkl.load(gzip.open(os.path.join(data_dir, 'songs_train_dev_test_s1.pkl.gz'), 'rb'))



    











In [ ]:



    


songset_trndev = songsets['train_song_set'] + songsets['dev_song_set']
songset_test = songsets['test_song_set']

Given a new song, recommend to the longest playlist

Random tie breaking if there are more than one longest playlist.





In [ ]:



    


pl_indices = np.where(Y_test.sum(axis=0).A.reshape(-1) > 0)[0]
lengths = Y_trndev.sum(axis=0).A.reshape(-1)[pl_indices]

Y_pred = lil_matrix(Y_test.shape, dtype=np.float)

np.random.seed(1234567890)
for ix in range(len(songset_test)):
    sort_ix = np.argsort(-lengths)
    long_ix = [sort_ix[0]]
    longest = lengths[sort_ix[0]]
    for i in range(1, sort_ix.shape[0]):
        if lengths[sort_ix[i]] < longest:
            break
        else:
            short_ix.append(sort_ix[i])
    long_ix = np.random.permutation(long_ix)
    rec_ix = long_ix[0]
    Y_pred[ix, pl_indices[rec_ix]] = 1
    lengths[rec_ix] += 1
    
Y_pred = Y_pred.tocsc()



    











In [ ]:



    


rps_longest = []
hitrates_longest = {top: [] for top in TOPs}
aucs_longest = []

for j in range(Y_test.shape[1]):
    if (j+1) % 100 == 0:
        sys.stdout.write('\r%d / %d' % (j+1, Y_test.shape[1]))
        sys.stdout.flush()
    y_true = Y_test[:, j].toarray().reshape(-1)
    if y_true.sum() < 1:
        continue
    y_pred = Y_pred[:, j].A.reshape(-1)
    
    rp, hr_dict, auc = calc_metrics(y_true, y_pred, tops=TOPs)
    rps_longest.append(rp)
    for top in TOPs:
        hitrates_longest[top].append(hr_dict[top])
    aucs_longest.append(auc)
    
print('\n%d / %d' % (len(rps_longest), Y_test.shape[1]))



    











In [ ]:



    


longest_perf = {dataset_name: 
               {'Test': {'R-Precision': np.mean(rps_longest), 
                         'Hit-Rate': {top: np.mean(hitrates_longest[top]) for top in hitrates_longest},
                         'AUC': np.mean(aucs_longest)}}}
longest_perf



    











In [ ]:



    


fperf_longest = os.path.join(data_dir, 'perf-longest.pkl')
print(fperf_longest)
pkl.dump(longest_perf, open(fperf_longest, 'wb'))
pkl.load(open(fperf_longest, 'rb'))

Given a new song, recommend to the shortest playlist

Random tie breaking if there are more than one shortest playlist.





In [ ]:



    


pl_indices = np.where(Y_test.sum(axis=0).A.reshape(-1) > 0)[0]
lengths = Y_trndev.sum(axis=0).A.reshape(-1)[pl_indices]

Y_pred = lil_matrix(Y_test.shape, dtype=np.float)

np.random.seed(1234567890)
for ix in range(len(songset_test)):
    sort_ix = np.argsort(lengths)
    short_ix = [sort_ix[0]]
    shortest = lengths[sort_ix[0]]
    for i in range(1, sort_ix.shape[0]):
        if lengths[sort_ix[i]] > shortest:
            break
        else:
            short_ix.append(sort_ix[i])
    short_ix = np.random.permutation(short_ix)
    rec_ix = short_ix[0]
    Y_pred[ix, pl_indices[rec_ix]] = 1
    lengths[rec_ix] += 1
    
Y_pred = Y_pred.tocsc()



    











In [ ]:



    


rps_shortest = []
hitrates_shortest = {top: [] for top in TOPs}
aucs_shortest = []
ndcgs_shortest = []

for j in range(Y_test.shape[1]):
    if (j+1) % 100 == 0:
        sys.stdout.write('\r%d / %d' % (j+1, Y_test.shape[1]))
        sys.stdout.flush()
    y_true = Y_test[:, j].toarray().reshape(-1)
    if y_true.sum() < 1:
        continue
    y_pred = Y_pred[:, j].A.reshape(-1)
    
    rp, hr_dict, auc, ndcg = calc_metrics(y_true, y_pred, tops=TOPs)
    rps_shortest.append(rp)
    for top in TOPs:
        hitrates_shortest[top].append(hr_dict[top])
    aucs_shortest.append(auc)
    ndcgs_shortest.append(ndcg)
    
print('\n%d / %d' % (len(rps_shortest), Y_test.shape[1]))



    











In [ ]:



    


shortest_perf = {dataset_name: 
                 {'Test': {'R-Precision': np.mean(rps_shortest), 
                           'Hit-Rate': {top: np.mean(hitrates_shortest[top]) for top in hitrates_shortest},
                           'AUC': np.mean(aucs_shortest),
                           'NDCG': np.mean(ndcgs_shortest)}}}
shortest_perf



    











In [ ]:



    


fperf_shortest = os.path.join(data_dir, 'perf-shortest.pkl')
print(fperf_shortest)
pkl.dump(shortest_perf, open(fperf_shortest, 'wb'))
pkl.load(open(fperf_shortest, 'rb'))

Popularity (in test set) as song score





In [ ]:



    


rps_poptest = []
hitrates_poptest = {top: [] for top in TOPs}
aucs_poptest = []
ndcgs_poptest = []

for j in range(Y_test.shape[1]):
    if (j+1) % 100 == 0:
        sys.stdout.write('\r%d / %d' % (j+1, Y_test.shape[1]))
        sys.stdout.flush()
    y_true = Y_test[:, j].toarray().reshape(-1)
    if y_true.sum() < 1:
        continue    
    y_pred = np.asarray([song2pop[sid] for sid, _ in songset_test])
    
    rp, hr_dict, auc, ndcg = calc_metrics(y_true, y_pred, tops=TOPs)
    rps_poptest.append(rp)
    for top in TOPs:
        hitrates_poptest[top].append(hr_dict[top])
    aucs_poptest.append(auc)
    ndcgs_poptest.append(ndcg)
    
print('\n%d / %d' % (len(rps_poptest), Y_test.shape[1]))



    











In [ ]:



    


fig = plt.figure(figsize=[20, 5])
ax1 = plt.subplot(131)
ax1.hist(rps_poptest, bins=100)
ax1.set_yscale('log')
ax1.set_title('R-Precision')
#ax.set_xlim(0, xmax)
ax2 = plt.subplot(132)
ax2.hist(aucs_poptest, bins=100)
ax2.set_yscale('log')
ax2.set_title('AUC')
ax3 = plt.subplot(133)
ax3.hist(ndcgs_poptest, bins=100)
ax3.set_yscale('log')
ax3.set_title('NDCG')
pass



    











In [ ]:



    


poptest_perf = {dataset_name: {'Test': {'R-Precision': np.mean(rps_poptest), 
                                     'Hit-Rate': {top: np.mean(hitrates_poptest[top]) for top in hitrates_poptest},
                                     'AUC': np.mean(aucs_poptest),
                                     'NDCG': np.mean(ndcgs_poptest)}}}
poptest_perf



    











In [ ]:



    


fperf_poptest = os.path.join(data_dir, 'perf-poptest.pkl')
print(fperf_poptest)
pkl.dump(poptest_perf, open(fperf_poptest, 'wb'))
pkl.load(open(fperf_poptest, 'rb'))

Logistic Regression with only song popularity as feature





In [ ]:



    


rps_lrpop = []
hitrates_lrpop = {top: [] for top in TOPs}
aucs_lrpop = []
ndcgs_lrpop = []

nsong_trndev = len(songset_trndev)
nsong_test = len(songset_test)
for j in range(Y_test.shape[1]):
    if (j+1) % 10 == 0:
        sys.stdout.write('\r%d / %d' % (j+1, Y_test.shape[1]))
        sys.stdout.flush()
    y_true = Y_test[:, j].toarray().reshape(-1)
    if y_true.sum() < 1:
        continue
    
    X_train = np.asarray([song2pop[sid] for sid, _ in songset_trndev]).reshape(nsong_trndev, 1)
    Y_train = Y_trndev[:, j].A.reshape(-1)
    clf = LogisticRegression()
    clf.fit(X_train, Y_train)
    X_test = np.asarray([song2pop[sid] for sid, _ in songset_test]).reshape(nsong_test, 1)
    
    y_pred = clf.decision_function(X_test).reshape(-1)
    
    rp, hr_dict, auc, ndcg = calc_metrics(y_true, y_pred, tops=TOPs)
    rps_lrpop.append(rp)
    for top in TOPs:
        hitrates_lrpop[top].append(hr_dict[top])
    aucs_lrpop.append(auc)
    ndcgs_lrpop.append(ndcg)
    
print('\n%d / %d' % (len(rps_lrpop), Y_test.shape[1]))



    











In [ ]:



    


lrpop_perf = {dataset_name: {'Test': {'R-Precision': np.mean(rps_lrpop), 
                                      'Hit-Rate': {top: np.mean(hitrates_lrpop[top]) for top in hitrates_lrpop},
                                      'AUC': np.mean(aucs_lrpop),
                                      'NDCG': np.mean(ndcgs_lrpop)}}}
lrpop_perf



    











In [ ]:



    


fperf_lrpop = os.path.join(data_dir, 'perf-lrpop.pkl')
print(fperf_lrpop)
pkl.dump(lrpop_perf, open(fperf_lrpop, 'wb'))
pkl.load(open(fperf_lrpop, 'rb'))

Content source: chengsoonong/digbeta

Similar notebooks:

notebook.community | gallery | about