Initial commit

README.md

@@ -0,0 +1,19 @@
+# 基于用户画像的商品推荐
+## Usage
+### Requirments
+
+- Tensorflow-GPU 2.4.1
+- CUDA 1.10
+
+### 代码说明
+#### 数据处理
+- 根据tagid是否缺失把train和test(复赛数据)分出两部分数据集
+- 将复赛数据集的train和test的tagid未缺失用户的tagid序列用来做Word2Vector
+
+#### 模型说明
+- 两层GRU
+- 五折交叉验证
+
+#### 结果输出
+- test中tagid缺失的用户label直接预测为1
+- 线下train_tagidNotnull_F1Score为0.6773461

all_code.py

@@ -0,0 +1,149 @@
+# _*_coding: utf-8-*_
+# @Project : 基于用户画像的商品推荐挑战赛
+# @FileNAme: all_code.py
+# @Author  : Rocket,Qian
+# @Time    : 2021/9/19 18:28
+import pandas as pd
+import numpy as np
+from sklearn.feature_extraction.text import TfidfVectorizer
+from sklearn.model_selection import StratifiedKFold
+from sklearn.metrics import accuracy_score, f1_score, roc_auc_score, recall_score, precision_score
+from tensorflow.keras.preprocessing import text, sequence
+# from tensorflow.keras.utils import to_categorical
+from tensorflow.keras.models import *
+from tensorflow.keras.layers import *
+from tensorflow.keras.callbacks import *
+from tensorflow.keras.layers import GRU
+import tensorflow as tf
+from gensim.models import Word2Vec
+from config import parser
+import os
+import warnings
+
+warnings.filterwarnings('ignore')
+
+args = parser.parse_args()
+# 读取数据，简单处理list数据
+train_all = pd.read_csv(args.train_file, header=None)
+test_all = pd.read_csv(args.test_file, header=None)
+# train_first = pd.read_csv(r'E:\Competition\基于用户画像的商品推荐挑战赛\dataset\train.txt', header=None)
+train_all.columns = ['pid', 'label', 'gender', 'age', 'tagid', 'time', 'province', 'city', 'model', 'make']
+test_all.columns = ['pid', 'gender', 'age', 'tagid', 'time', 'province', 'city', 'model', 'make']
+print('导入数据成功')
+train = train_all[train_all['tagid'].notnull()]
+test = test_all[test_all['tagid'].notnull()]
+
+flag = 0
+if flag == 1:
+    train = pd.concat([train_first, train])
+    # train.to_csv(r'E:\Competition\基于用户画像的商品推荐挑战赛\train_sum.csv', index=False)
+
+train['label'] = train['label'].astype(int)
+
+data = pd.concat([train, test])
+data['label'] = data['label'].fillna(-1)
+data['tagid'] = data['tagid'].apply(lambda x: eval(x))
+data['tagid'] = data['tagid'].apply(lambda x: [str(i) for i in x])
+
+embed_size = args.embed_size
+MAX_WORDS_NUM = args.MAX_WORDS_NUM
+MAX_SEQUENCE_LENGTH = args.MAX_SEQUENCE_LENGTH
+w2v_model = Word2Vec(sentences=data['tagid'].tolist(), vector_size=embed_size, window=args.window, min_count=1,
+                     epochs=args.epochs, hs=1)
+
+X_train = data[:train.shape[0]]['tagid']
+X_test = data[train.shape[0]:]['tagid']
+
+tokenizer = text.Tokenizer(num_words=MAX_WORDS_NUM)
+tokenizer.fit_on_texts(list(X_train) + list(X_test))
+X_train = tokenizer.texts_to_sequences(X_train)
+X_test = tokenizer.texts_to_sequences(X_test)
+X_train = sequence.pad_sequences(X_train, maxlen=MAX_SEQUENCE_LENGTH, padding='pre', truncating='pre')
+X_test = sequence.pad_sequences(X_test, maxlen=MAX_SEQUENCE_LENGTH, padding='pre', truncating='pre')
+word_index = tokenizer.word_index
+nb_words = len(word_index) + 1
+
+embedding_matrix = np.zeros((nb_words, embed_size))
+for word, i in word_index.items():
+    try:
+        embedding_vector = w2v_model.wv.get_vector(word)
+    except KeyError:
+        continue
+    if embedding_vector is not None:
+        embedding_matrix[i] = embedding_vector
+y_cat = train['label'].values
+
+# GPU设置
+os.environ['TF_XLA_FLAGS'] = '--tf_xla_enable_xla_devices'
+os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
+
+physical_devices = tf.config.list_physical_devices('GPU')
+tf.config.experimental.set_memory_growth(physical_devices[0], enable=True)
+
+
+# 定义模型
+def my_model():
+    embedding_input = Input(shape=(MAX_SEQUENCE_LENGTH,), dtype='int32')
+    # 词嵌入（使用预训练的词向量）
+    embedder = Embedding(nb_words,
+                         embed_size,
+                         input_length=MAX_SEQUENCE_LENGTH,
+                         weights=[embedding_matrix],
+                         trainable=False
+                         )
+    embed = embedder(embedding_input)
+    l = GRU(args.GRU1_hidden_size, return_sequences=True)(embed)
+    flat = BatchNormalization()(l)
+    drop = Dropout(args.dropout)(flat)
+    l2 = GRU(args.GRU2_hidden_size)(drop)
+    output = Dense(1, activation='sigmoid')(l2)
+    model = Model(inputs=embedding_input, outputs=output)
+    model.compile(loss=tf.keras.losses.BinaryCrossentropy(), optimizer='adam', metrics=['accuracy'])
+    return model
+
+
+# 五折交叉验证
+folds = StratifiedKFold(n_splits=5, shuffle=True, random_state=2021)
+triain_pre = np.zeros([len(train), 1])
+test_predictions = np.zeros([len(test), 1])
+
+for fold_, (trn_idx, val_idx) in enumerate(folds.split(train, train['label'])):
+    print("fold n{}".format(fold_ + 1))
+    model = my_model()
+    early_stopping = EarlyStopping(monitor='val_accuracy', patience=5)
+    bst_model_path = "./{}.h5".format(fold_)
+    model_checkpoint = ModelCheckpoint(bst_model_path, save_best_only=True, save_weights_only=True)
+
+    X_tra, X_val = X_train[trn_idx], X_train[val_idx]
+    y_tra, y_val = y_cat[trn_idx], y_cat[val_idx]
+
+    model.fit(X_tra, y_tra,
+              validation_data=(X_val, y_val),
+              epochs=args.nn_epochs, batch_size=args.batch_size, shuffle=True,
+              callbacks=[early_stopping, model_checkpoint])
+
+    model.load_weights(bst_model_path)
+    triain_pre[val_idx] = model.predict(X_val)
+    test_predictions += model.predict(X_test) / folds.n_splits
+    del model
+
+
+submit = test[['pid']]
+submit['tmp'] = test_predictions
+submit.columns = ['user_id', 'tmp']
+
+submit['rank'] = submit['tmp'].rank()
+submit['category_id'] = 1
+submit.loc[submit['rank'] <= int(submit.shape[0] * 0.859), 'category_id'] = 0
+
+submit_null = test_all[test_all['tagid'].isna()][['pid']]
+submit_null['category_id'] = 1
+
+submit_notnull = submit[['user_id', 'category_id']]
+submit_notnull.columns = ['pid', 'category_id']
+
+sub = pd.concat([submit_null, submit_notnull])
+sub.sort_values(by='pid', ascending=True, inplace=True)
+sub.to_csv(
+    r'E:\Competition\基于用户画像的商品推荐挑战赛\result\0920GRUemd64b400win1_0859.csv',
+    index=False)

config.py

@@ -0,0 +1,32 @@
+# _*_coding: utf-8-*_
+# @Project : 基于用户画像的商品推荐挑战赛
+# @FileNAme: config.py
+# @Author  : Rocket,Qian
+# @Time    : 2021/9/10 20:25
+import argparse
+path = r'E:\Competition\基于用户画像的商品推荐挑战赛\dataset'
+parser = argparse.ArgumentParser(description="基于用户画像的商品推荐挑战赛")
+
+# ========================= Dataset Configs ==========================
+parser.add_argument('--train_file', type=str, default=path + r'\data2\train.txt')
+parser.add_argument('--test_file', type=str, default=path + r'\data2\test.txt')
+
+# ========================= Word2Vec Configs ==========================
+parser.add_argument('--embed_size', type=int, default=64, help='embedding_size of every tagid')
+parser.add_argument('--MAX_WORDS_NUM', type=int, default=224253, help='all word of fusai data')
+parser.add_argument('--MAX_SEQUENCE_LENGTH', type=int, default=256)
+parser.add_argument('--window', type=int, default=1)
+parser.add_argument('--epochs', type=int, default=10, help='Num epochs of word2vec')
+
+# ========================= Model Configs ==========================
+parser.add_argument('--GRU1_hidden_size', type=int, default=128, help='GRU1 hidden_size ')
+parser.add_argument('--GRU2_hidden_size', type=int, default=256, help='GRU2 hidden_size ')
+parser.add_argument('--dropout', type=float, default=0.2, help='dropout ratio')
+parser.add_argument('--nn_epochs', type=int, default=128, help='number of total epochs to train')
+parser.add_argument('--batch_size', type=int, default=400, help='number of total epochs to train')
+
+
+args = parser.parse_args()
+print(args.dropout)
+print(args.train_file)
+print(args.batch_size)

requirements.txt

@@ -0,0 +1,5 @@
+numpy==1.19.5
+tensorflow_gpu==2.4.1
+gensim==4.0.1
+pandas==0.25.3
+scikit_learn==0.24.2