def global_feature():
cutoffLine('-')
print 'Generate global feature'
# 统计每种商品每天销量，为统计每种商品在同类商品种排名服务， 为了避免使用未来信息
global ci_sale
if os.path.exists('data/ci_sale.pkl'):
ci_sale_file = open('data/ci_sale.pkl', 'rb')
ci_sale = pickle.load(ci_sale_file)
# for c in ci_rank: print ci_rank[c]
ci_sale_file.close()
else:
u_file = file('data/nuser.csv', 'r')
u_reader = csv.reader(u_file)
ci_sale = {}
for line in u_reader:
doneCount(u_reader.line_num)
item = int(line[1])
behavior = int(line[2])
category = int(line[4])
date = int(line[5])
if not ci_sale.has_key(category): ci_sale[category] = {}
if behavior == 4:
if not ci_sale[category].has_key(item): ci_sale[category][item] = [0]*(TOTAL_DAY+1)
ci_sale[category][item][date] += 1
ci_sale_file = open('data/ci_sale.pkl', 'wb')
pickle.dump(ci_sale, ci_sale_file)
ci_sale_file.close()
u_file.close()
def drop_no_buy_user():
cutoffLine('-')
rfile = file('data/nuser.csv','r')
reader = csv.reader(rfile)
buyed_user = set()
print 'user behavior stat'
for line in reader:
doneCount(reader.line_num)
if int(line[2]) == 4: buyed_user.add(int(line[0]))
rfile.close()
print '\ndrop...'
rfile = file('data/nuser.csv','r')
wfile = file('data/nuser_cleaned','w')
reader = csv.reader(rfile)
writer = csv.writer(wfile)
count = 0
for line in reader:
doneCount(reader.line_num)
if int(line[0]) in buyed_user:
writer.writerow(line)
count += 1
cutoffLine('-')
print count
rfile.close()
wfile.close()
def predict(window, model, item_subset, proportion, algo, confidence):
cutoffLine('-')
print 'Generate result set with confidence %f' % confidence
feature_file = file('splited_data_%d/set_for_prediction.csv'%window, 'r')
result_file = file('data/tianchi_mobile_recommendation_predict_%d_%s_%d_%s.csv'%\
(window, algo, proportion, str(confidence)), 'w')
f_reader = csv.reader(feature_file)
r_writer = csv.writer(result_file)
r_writer.writerow(['user_id','item_id'])
predict_set = set()
UI = []
X = []
each_time = 500000
for line in f_reader:
doneCount(f_reader.line_num)
line = map(int, line)
UI.append(tuple(line[0:2]))
X.append(line[3:])
if f_reader.line_num % each_time == 0:
if algo == 'lr' or algo == 'svm': X = preprocessing.scale(X)
if algo == 'lr' or algo == 'rf':
y_pred = model.predict_proba(X)
print y_pred
for index, y in enumerate(y_pred):
if y[1] > confidence: predict_set.add(UI[index])
if algo == 'svm':
y_pred = model.predict(X)
for index, y in enumerate(y_pred):
if y == 1: predict_set.add(UI[index])
UI = []
X = []
if len(UI) > 0:
if algo == 'lr' or algo == 'svm': X = preprocessing.scale(X)
if algo == 'lr' or algo == 'rf':
y_pred = model.predict_proba(X)
for index, y in enumerate(y_pred):
if y[1] > confidence: predict_set.add(UI[index])
if algo == 'svm':
y_pred = model.predict(X)
for index, y in enumerate(y_pred):
if y == 1: predict_set.add(UI[index])
UI = []
X = []
cutoffLine('-')
print "Prediction set size before drop: %d" % len(predict_set)
predict_set = dropItemsNotInSet(predict_set, item_subset)
r_writer.writerows(predict_set)
print "Prediction set size after drop: %d" % len(predict_set)
feature_file.close()
result_file.close()
return len(predict_set)
def merge_training_set():
cutoffLine('*')
print 'Merge training set'
start_time = time.time()
positive_count = 0
negative_count = 0
total_count = 0
total_file = file(PRE_DIR + '/' + 'train_set.csv', 'w')
pos_file = file(PRE_DIR + '/' + 'positive_set.csv', 'w')
neg_file = file(PRE_DIR + '/' + 'negative_set.csv', 'w')
total_writer = csv.writer(total_file)
pos_writer = csv.writer(pos_file)
neg_writer = csv.writer(neg_file)
for i in range(1, FILES-1):
cutoffLine('-')
print 'load train set %d' % i
r_file = file(PRE_DIR + '/' + 'set_%d.csv' % i)
reader = csv.reader(r_file)
for line in reader:
doneCount(reader.line_num)
line = map(int, line)
if line[-1] == 1:
positive_count += 1
pos_writer.writerow(line)
if line[-1] == 0:
negative_count += 1
neg_writer.writerow(line)
total_count += 1
total_writer.writerow(line)
r_file.close()
total_file.close()
pos_file.close()
neg_file.close()
cutoffLine('-')
# 44114
print 'Positive Example: %d' % positive_count
# 59373295
print 'Negative Example: %d' % (total_count - positive_count)
# 59417409
print 'Total Example: %d' % total_count
# 一致性判断
print 'Is right? %s'%('Yes' if positive_count + negative_count == total_count else 'No')
end_time = time.time()
duration = timekeeper(start_time, end_time)
cutoffLine('*')
print 'It takes %s to merge training set and backup negative and positive set' % duration
def predict(model, index):
cutoffLine('-')
print 'Generate result set %d' % index
feature_file = file('splited_data/set_for_prediction.csv', 'r')
result_file = file(TRAIN_SET_DIR + '/' + 'lr_result_%d.csv' % index, 'w')
f_reader = csv.reader(feature_file)
r_writer = csv.writer(result_file)
r_writer.writerow(['user_id','item_id'])
for line in f_reader:
doneCount(f_reader.line_num)
line = map(int, line)
if model.predict([line[2:]])[0] == 1: r_writer.writerow(line[0:2])
feature_file.close()
result_file.close()
def stat():
cutoffLine('-')
print 'stat some information...'
user_file = file('data/nuser.csv','r')
item_file = file('data/item.csv','r')
stat_file = open('data/stat.txt','w')
row_count = 0
user_set = set()
sub_item_set = set()
all_item_set = set()
category_set = set()
user_geo_count = 0
item_geo_count = 0
reader = csv.reader(item_file)
for line in reader:
doneCount(reader.line_num)
if reader.line_num == 1: continue
if line[1]: item_geo_count += 1
category_set.add(line[2])
sub_item_set.add(line[0])
reader = csv.reader(user_file)
for line in reader:
doneCount(reader.line_num)
row_count += 1
user_set.add(line[0])
all_item_set.add(line[1])
if line[3]: user_geo_count += 1
interact_item_set = all_item_set & sub_item_set
stat_file.write('%s : %s\n'%(u'Total Count',row_count))
stat_file.write('%s : %s\n'%(u'User Count',len(user_set)))
stat_file.write('%s : %s\n'%(u'All Item Count',len(all_item_set)))
stat_file.write('%s : %s\n'%(u'Sub Item Count',len(sub_item_set)))
stat_file.write('%s : %s %f\n'%(u'Interact Item Count',
len(interact_item_set),
float(len(interact_item_set))/len(sub_item_set)))
stat_file.write('%s : %s\n'%(u'Category Count',len(category_set)))
stat_file.write('%s : %s\n'%(u'User Geo Count',user_geo_count))
stat_file.write('%s : %s\n'%(u'Item Geo Count',item_geo_count))
stat_file.close()
user_file.close()
item_file.close()
def evaluate_model(model, index):
cutoffLine('-')
print 'offline evaluate RF model %d' % index
test_file = file('splited_data/set_test.csv', 'r')
test_reader = csv.reader(test_file)
predict_set = set()
real_set = set()
for line in test_reader:
doneCount(test_file.line_num)
line = map(int, line)
if line[-1] == 1 : real_set.add((line[0],line[1]))
if model.predict([line[2:-1]])[0] == 1: predict_set.add((line[0],line[1]))
import evaluate
P, R, F = evaluate.evaluate(predict_set, real_set)
test_file.close()
return P, R, F
def stat():
cutoffLine('-')
print 'stat some information...'
user_file = file('data/nuser.csv', 'r')
item_file = file('data/item.csv', 'r')
stat_file = open('data/stat.txt', 'w')
row_count = 0
user_set = set()
sub_item_set = set()
all_item_set = set()
category_set = set()
user_geo_count = 0
item_geo_count = 0
reader = csv.reader(item_file)
for line in reader:
doneCount(reader.line_num)
if reader.line_num == 1: continue
if line[1]: item_geo_count += 1
category_set.add(line[2])
sub_item_set.add(line[0])
reader = csv.reader(user_file)
for line in reader:
doneCount(reader.line_num)
row_count += 1
user_set.add(line[0])
all_item_set.add(line[1])
if line[3]: user_geo_count += 1
interact_item_set = all_item_set & sub_item_set
stat_file.write('%s : %s\n' % (u'Total Count', row_count))
stat_file.write('%s : %s\n' % (u'User Count', len(user_set)))
stat_file.write('%s : %s\n' % (u'All Item Count', len(all_item_set)))
stat_file.write('%s : %s\n' % (u'Sub Item Count', len(sub_item_set)))
stat_file.write('%s : %s %f\n' %
(u'Interact Item Count', len(interact_item_set),
float(len(interact_item_set)) / len(sub_item_set)))
stat_file.write('%s : %s\n' % (u'Category Count', len(category_set)))
stat_file.write('%s : %s\n' % (u'User Geo Count', user_geo_count))
stat_file.write('%s : %s\n' % (u'Item Geo Count', item_geo_count))
stat_file.close()
user_file.close()
item_file.close()
def evaluate_model(algo, window, model, item_subset, confidence):
cutoffLine('-')
print 'offline evaluate model with confidence %f' % confidence
test_file = file('splited_data_%d/set_test.csv'%window, 'r')
test_reader = csv.reader(test_file)
predict_set = set()
real_set = set()
UI = []
X = []
each_time = 500000
for line in test_reader:
doneCount(test_reader.line_num)
line = map(int, line)
UI.append(tuple(line[0:2]))
X.append(line[3:-1])
if line[-1] == 1 : real_set.add((line[0],line[1]))
if test_reader.line_num % each_time == 0:
if algo == 'lr' or algo == 'svm': X = preprocessing.scale(X)
if algo == 'lr' or algo == 'rf':
y_pred = model.predict_proba(X)
for index, y in enumerate(y_pred):
if y[1] > confidence: predict_set.add(UI[index])
if algo == 'svm':
y_pred = model.predict(X)
for index, y in enumerate(y_pred):
if y == 1: predict_set.add(UI[index])
UI = []
X = []
if len(UI) > 0:
if algo == 'lr' or algo == 'svm': X = preprocessing.scale(X)
if algo == 'lr' or algo == 'rf':
y_pred = model.predict_proba(X)
for index, y in enumerate(y_pred):
if y[1] > confidence: predict_set.add(UI[index])
if algo == 'svm':
y_pred = model.predict(X)
for index, y in enumerate(y_pred):
if y == 1: predict_set.add(UI[index])
UI = []
X = []
predict_set = dropItemsNotInSet(predict_set, item_subset)
real_set = dropItemsNotInSet(real_set, item_subset)
import evaluate
P, R, F = evaluate.evaluate(predict_set, real_set)
test_file.close()
return P, R, F
def train(window, proportion, algo, confidence):
start_time = time.time()
cutoffLine('*')
print '%s model training with sample proportion 1:%d...' %(algo, proportion)
t_file = file('data/training_set_%d_%d.csv' % (window, proportion), 'r')
t_reader = csv.reader(t_file)
X = []
y = []
for line in t_reader:
doneCount(t_reader.line_num)
line = map(int, line)
X.append(line[3:-1])
y.append(line[-1])
model_name = 'data/model/%s_%d_%d.model'%(algo, window, proportion)
if os.path.exists(model_name): model = joblib.load(model_name)
else:
if algo == 'lr': model = LR(X, y)
if algo == 'rf': model = RF(X, y)
if algo == 'svm': model = SVM(X, y)
joblib.dump(model, model_name)
cutoffLine('-')
print model.classes_
item_subset = loadItemSubset()
record_file = open('data/model_evaluate_record.txt','a')
P, R, F= evaluate_model(algo, window, model, item_subset, confidence)
predict_set_size = predict(window, model, item_subset, proportion, algo, confidence)
record_file.write('window %d '%window + algo+' %d'%proportion + ' %.2f\n'%confidence)
record_file.write('\tP: %f\n'%P)
record_file.write('\tR: %f\n'%R)
record_file.write('\tF1: %f\n'%F)
record_file.write('Predict Set Size: %d\n'%predict_set_size)
record_file.write('-'*30+'\n')
record_file.close()
t_file.close()
cutoffLine('*')
end_time = time.time()
duration = timekeeper(start_time, end_time)
print 'I takes %s to train , evaluate model and generate result' % duration
def predict(window, model, item_subset, proportion, algo, confidence):
cutoffLine('-')
print 'Generate result set with confidence %f' % confidence
feature_file = file('splited_data_%d/set_for_prediction.csv' % window, 'r')
result_file = file('data/tianchi_mobile_recommendation_predict_%d_%s_%d_%s.csv'%\
(window, algo, proportion, str(confidence)), 'w')
f_reader = csv.reader(feature_file)
r_writer = csv.writer(result_file)
r_writer.writerow(['user_id', 'item_id'])
predict_set = set()
UI = []
X = []
each_time = 500000
for line in f_reader:
doneCount(f_reader.line_num)
line = map(int, line)
UI.append(tuple(line[0:2]))
X.append(line[3:])
if f_reader.line_num % each_time == 0:
if algo == 'lr': X = preprocessing.scale(X)
y_pred = model.predict_proba(X)
for index, y in enumerate(y_pred):
if y[1] > confidence: predict_set.add(UI[index])
UI = []
X = []
if len(UI) > 0:
if algo == 'lr': X = preprocessing.scale(X)
y_pred = model.predict_proba(X)
for index, y in enumerate(y_pred):
if y[1] > confidence: predict_set.add(UI[index])
UI = []
X = []
cutoffLine('-')
print "Prediction set size before drop: %d" % len(predict_set)
predict_set = dropItemsNotInSet(predict_set, item_subset)
r_writer.writerows(predict_set)
print "Prediction set size after drop: %d" % len(predict_set)
feature_file.close()
result_file.close()
return len(predict_set)
def cartBuy():
user_file = file('data/nuser.csv','r')
reader = csv.reader(user_file)
cart_30 = set()
buy_31 = set()
cart_31 = set()
for line in reader:
doneCount(reader.line_num)
if int(line[5]) == 30 and int(line[6]) > 15:
if int(line[2]) == 3:cart_30.add((int(line[0]),int(line[1])))
if int(line[2]) == 4:
if (line[0],line[1]) in cart_30:cart_30.remove((int(line[0]),int(line[1])))
if int(line[5]) == 31 and int(line[6]) > 15:
if int(line[2]) == 3:cart_31.add((int(line[0]),int(line[1])))
if int(line[2]) == 4:
if (line[0],line[1]) in cart_31:cart_31.remove((int(line[0]),int(line[1])))
if int(line[5]) == 31 and int(line[2]) == 4:
buy_31.add((int(line[0]),int(line[1])))
user_file.close()
return cart_30 , buy_31 , cart_31
def predict(model, item_subset):
cutoffLine('-')
print 'Generate result set'
feature_file = file('splited_data/set_for_prediction.csv', 'r')
result_file = file('data/prediction_lr.csv', 'w')
f_reader = csv.reader(feature_file)
r_writer = csv.writer(result_file)
r_writer.writerow(['user_id','item_id'])
predict_set = set()
for line in f_reader:
doneCount(f_reader.line_num)
line = map(int, line)
if model.predict([line[2:]])[0] == 1: predict_set.add((line[0], line[1]))
cutoffLine('-')
print "Prediction set size before drop: %d" % len(predict_set)
predict_set = dropItemsNotInSet(predict_set, item_subset)
r_writer.writerows(predict_set)
print "Prediction set size after drop: %d" % len(predict_set)
feature_file.close()
result_file.close()
def lineCount():
stat_file = file(PRE_DIR + '/stat.csv','w')
writer = csv.writer(stat_file)
for i in range(1,FILES+1):
print '\n' + '-'*50
if i == FILES: file_name = 'for_prediction.csv'
else: file_name = '%d.csv' % i
file_path = PRE_DIR + '/' + file_name
print 'processing %s' % file_name
rfile = file(file_path,'r')
reader = csv.reader(rfile)
count = 1
for line in reader:
doneCount(reader.line_num)
count += 1
writer.writerow([file_name, count])
rfile.close()
stat_file.close()