import os import sys sys.path.append(os.getcwd()) import numpy as np from tqdm import tqdm import faiss import warnings import pickle import collections import random from sklearn.preprocessing import MinMaxScaler from sklearn.preprocessing import LabelEncoder import tensorflow as tf from tensorflow.python.keras import backend as K from tensorflow.python.keras.models import Model from tensorflow.keras.preprocessing.sequence import pad_sequences from deepmatch.models import YoutubeDNN from deepmatch.utils import sampledsoftmaxloss, NegativeSampler from deepctr.feature_column import SparseFeat, VarLenSparseFeat, DenseFeat from collections import Counter from config import logger, offline_mode from config import need_metric_recall from config import embedding_dim, ef_construction, M from config import subject_youtubednn_recall_dict from config import subject_youtube_user_emb_dict from config import subject_youtube_item_emb_dict from config import subject_youtube_item_faiss_model_path from config import subject_youtube_user_faiss_model_path from config import subject_youtube_user_embedding_index_dict from config import subject_youtube_item_embedding_index_dict from config import subject_youtubednn_train_input_data from config import subject_youtubednn_train_label_data from config import subject_youtubednn_test_input_data from config import subject_youtubednn_test_label_data from config import subject_youtubednn_train_set_data from config import subject_youtubednn_test_set_data from config import subject_youtube_user_embedding_data from config import subject_youtube_item_embedding_data from config import samples_mode from config import subject_max_seq_len from matching.subject.recall_comm import get_all_select_df from matching.subject.recall_comm import get_user_info_df, get_item_info_df from matching.subject.recall_comm import get_all_hist_and_last_select, metrics_recall K.set_learning_phase(True) if tf.__version__ >= '2.0.0': tf.compat.v1.disable_eager_execution() tqdm.pandas() warnings.filterwarnings('ignore') def gen_data_set(data, negsample=0): """ 获取YoutubeDNN召回时的训练验证数据 negsample指的是通过滑窗构建样本的时候，负采集样样本的数量 """ data.sort_values("created_timestamp", inplace=True) item_ids = data['subject_id'].unique() train_set = [] test_set = [] if os.path.exists(subject_youtubednn_train_set_data) and \ os.path.exists(subject_youtubednn_test_set_data): train_set = pickle.load(open(subject_youtubednn_train_set_data, 'rb')) test_set = pickle.load(open(subject_youtubednn_test_set_data, 'rb')) return train_set, test_set for reviewerID, hist_select in tqdm(data.groupby('user_id')): pos_list = hist_select['subject_id'].tolist() if negsample > 0: # 用户没选择的物品里面选择负样本 candidate_set = list(set(item_ids) - set(pos_list)) # 对于每个正样本，选择n个负样本 neg_list = np.random.choice(candidate_set,size=len(pos_list)*negsample,replace=True) # 长度只有一个的时候，需要把这条数据也放到训练集中，不然最终学到的embedding就会有缺失 if len(pos_list) == 1: train_set.append((reviewerID, [pos_list[0]], pos_list[0], 1, len(pos_list), hist_select.iloc[0]['gender'], hist_select.iloc[0]['identity'], hist_select.iloc[0]['edu_background'],hist_select.iloc[0]['logins'], hist_select.iloc[0]['grade'],hist_select.iloc[0]['experience'])) test_set.append((reviewerID, [pos_list[0]], pos_list[0], 1, len(pos_list), hist_select.iloc[0]['gender'], hist_select.iloc[0]['identity'], hist_select.iloc[0]['edu_background'],hist_select.iloc[0]['logins'], hist_select.iloc[0]['grade'],hist_select.iloc[0]['experience'])) # 滑窗构造正负样本 for i in range(1, len(pos_list)): hist_sel = pos_list[:i] if i != len(pos_list) - 1: # 正样本 [user_id, his_item, pos_item, label, len(his_item)] train_set.append((reviewerID, hist_sel[::-1], pos_list[i], 1, len(hist_sel[::-1]), hist_select.iloc[0]['gender'], hist_select.iloc[0]['identity'], hist_select.iloc[0]['edu_background'],hist_select.iloc[0]['logins'], hist_select.iloc[0]['grade'],hist_select.iloc[0]['experience'])) for negi in range(negsample): # 负样本 [user_id, his_item, neg_item, label, len(his_item)] train_set.append((reviewerID, hist_sel[::-1], neg_list[i*negsample+negi], 0, len(hist_sel[::-1]), hist_select.iloc[0]['gender'], hist_select.iloc[0]['identity'], hist_select.iloc[0]['edu_background'],hist_select.iloc[0]['logins'], hist_select.iloc[0]['grade'],hist_select.iloc[0]['experience'])) else: # 将最长的那一个序列长度作为测试数据 test_set.append((reviewerID, hist_sel[::-1], pos_list[i], 1, len(hist_sel[::-1]), hist_select.iloc[0]['gender'], hist_select.iloc[0]['identity'], hist_select.iloc[0]['edu_background'],hist_select.iloc[0]['logins'], hist_select.iloc[0]['grade'],hist_select.iloc[0]['experience'])) random.shuffle(train_set) random.shuffle(test_set) pickle.dump(train_set, open(subject_youtubednn_train_set_data, 'wb')) pickle.dump(test_set, open(subject_youtubednn_test_set_data, 'wb')) return train_set, test_set def gen_model_input(train_set, user_profile, seq_max_len): """ 将输入的数据进行padding，使得序列特征的长度都一致 """ train_uid = np.array([line[0] for line in train_set]) train_seq = [line[1] for line in train_set] train_iid = np.array([line[2] for line in train_set]) train_label = np.array([line[3] for line in train_set]) train_hist_len = np.array([line[4] for line in train_set]) train_gender = np.array([line[5] for line in train_set]) train_identity = np.array([line[6] for line in train_set]) train_edu_background = np.array([line[7] for line in train_set]) train_logins = np.array([line[8] for line in train_set]) train_grade = np.array([line[9] for line in train_set]) train_experience = np.array([line[10] for line in train_set]) train_seq_pad = pad_sequences(train_seq, maxlen=seq_max_len, padding='post', truncating='post', value=0) train_model_input = {"user_id": train_uid, "subject_id": train_iid, "hist_subject_id": train_seq_pad, "hist_len": train_hist_len, "gender":train_gender, "identity":train_identity, "edu_background":train_edu_background, "logins":train_logins, "grade":train_grade, "experience":train_experience} return train_model_input, train_label def youtubednn_recall(data, topk=20): # 定义用户和物品特征 user_sparse_features = ['user_id'] #用户连续性特征：年龄，身份，教育背景，登录次数，金币数，经验值 user_dense_features = ['gender','identity','edu_background','logins', 'grade', 'experience'] item_sparse_features = ['subject_id'] # 用户选择序列的长度，短的填充，长的截断 SEQ_LEN = subject_max_seq_len user_profile_dup = data[["user_id"]].drop_duplicates('user_id') item_profile_dup = data[["subject_id"]].drop_duplicates('subject_id') # 定义特征 sparse_features = user_sparse_features + item_sparse_features feature_max_idx = {} for feature in sparse_features: lbe = LabelEncoder() data[feature] = lbe.fit_transform(data[feature]) feature_max_idx[feature] = data[feature].max() + 1 for feature in user_dense_features: min_max_scaler = MinMaxScaler() data[feature] = min_max_scaler.fit_transform(data[[feature]]) # 提取user和item的特征，具体选择哪些特征需要进一步分析 user_profile = data[user_sparse_features + user_dense_features].drop_duplicates('user_id') item_profile = data[item_sparse_features].drop_duplicates('subject_id') user_index_2_rawid = dict(zip(user_profile['user_id'], user_profile_dup['user_id'])) item_index_2_rawid = dict(zip(item_profile['subject_id'], item_profile_dup['subject_id'])) # 生成训练集和测试集 # 由于深度学习需要的数据量非常大 # 为了保证召回的效果，通过滑窗的形式扩充训练样本 train_set, test_set = gen_data_set(data, 1) if os.path.exists(subject_youtubednn_train_input_data) and \ os.path.exists(subject_youtubednn_train_label_data): # 加载YoutubeDNN模型的训练集数据 train_model_input = pickle.load(open(subject_youtubednn_train_input_data, 'rb')) train_label = pickle.load(open(subject_youtubednn_train_label_data, 'rb')) else: # 生成并保存YoutubeDNN模型的训练集数据 train_model_input, train_label = gen_model_input(train_set, user_profile, SEQ_LEN) pickle.dump(train_model_input, open(subject_youtubednn_train_input_data, 'wb')) pickle.dump(train_label, open(subject_youtubednn_train_label_data, 'wb')) if os.path.exists(subject_youtubednn_test_input_data) and \ os.path.exists(subject_youtubednn_test_label_data): # 加载YoutubeDNN模型的测试集数据 test_model_input = pickle.load(open(subject_youtubednn_test_input_data, 'rb')) test_label = pickle.load(open(subject_youtubednn_test_label_data, 'rb')) else: # 生成并保存YoutubeDNN模型的测试集数据 test_model_input, test_label = gen_model_input(test_set, user_profile, SEQ_LEN) pickle.dump(test_model_input, open(subject_youtubednn_test_input_data, 'wb')) pickle.dump(test_label, open(subject_youtubednn_test_label_data, 'wb')) # 将数据整理成模型可以直接输入的形式 # 用户特征由稀疏特征user_id和可变长度特征hist_subject_id组成 user_feature_columns = [SparseFeat('user_id', feature_max_idx['user_id'], embedding_dim), DenseFeat('gender', dimension=1), DenseFeat('identity', dimension=1), DenseFeat('edu_background', dimension=1), DenseFeat('logins', dimension=1), DenseFeat('grade', dimension=1), DenseFeat('experience', dimension=1), VarLenSparseFeat(SparseFeat('hist_subject_id', feature_max_idx['subject_id'], embedding_dim, embedding_name="subject_id"), SEQ_LEN, 'mean', 'hist_len'),] # 物品循环由稀疏特征subject_id组成 item_feature_columns = [SparseFeat('subject_id', feature_max_idx['subject_id'], embedding_dim)] # 进行负采样 train_counter = Counter(train_model_input['subject_id']) item_count = [train_counter.get(i, 0) for i in range(item_feature_columns[0].vocabulary_size)] sampler_config = NegativeSampler('frequency', num_sampled=5, item_name='subject_id', item_count=item_count) # 模型的定义 # num_sampled: 负采样时的样本数量 model = YoutubeDNN(user_feature_columns, item_feature_columns, sampler_config=sampler_config, user_dnn_hidden_units=(128, embedding_dim)) # 模型编译 model.compile(optimizer="adam", loss=sampledsoftmaxloss) print("开始训练") # 模型训练，这里可以定义验证集的比例，如果设置为0的话就是全量数据直接进行训练 history = model.fit(train_model_input, train_label, batch_size=256, epochs=10, verbose=1, validation_split=0.0) # 训练完模型之后,提取训练的Embedding，包括user端和item端 test_user_model_input = test_model_input all_item_model_input = {"subject_id": item_profile['subject_id'].values} user_embedding_model = Model(inputs=model.user_input, outputs=model.user_embedding) item_embedding_model = Model(inputs=model.item_input, outputs=model.item_embedding) # 保存输出的item_embedding和user_embedding # 排序的时候用到，保存的时候需要和原始的id对应 user_embs = user_embedding_model.predict(test_user_model_input, batch_size=2 ** 12) item_embs = item_embedding_model.predict(all_item_model_input, batch_size=2 ** 12) # embedding保存之前归一化一下 user_embs = user_embs / np.linalg.norm(user_embs, axis=1, keepdims=True) item_embs = item_embs / np.linalg.norm(item_embs, axis=1, keepdims=True) pickle.dump(user_embs, open(subject_youtube_user_embedding_data, 'wb')) pickle.dump(item_embs, open(subject_youtube_item_embedding_data, 'wb')) logger.info('构建youtube user embedding HNSW索引') build_hnsw(user_embs, subject_youtube_user_faiss_model_path, ef_construction, M, embedding_dim) logger.info('构建youtube item embedding HNSW索引') build_hnsw(item_embs, subject_youtube_item_faiss_model_path, ef_construction, M, embedding_dim) # 将embedding转换成字典的形式方便查询 raw_user_id_emb_dict = {user_index_2_rawid[k]: v for k, v in zip(user_profile['user_id'], user_embs)} raw_item_id_emb_dict = {item_index_2_rawid[k]: v for k, v in zip(item_profile['subject_id'], item_embs)} pickle.dump(raw_user_id_emb_dict, open(subject_youtube_user_emb_dict, 'wb')) pickle.dump(raw_item_id_emb_dict, open(subject_youtube_item_emb_dict, 'wb')) # 生成user embedding和item embedding索引字典 youtube_user_embedding_index_dict = {k: user_index_2_rawid[v] for k, v in enumerate(user_profile['user_id'])} youtube_item_embedding_index_dict = {k: item_index_2_rawid[v] for k, v in enumerate(item_profile['subject_id'])} pickle.dump(youtube_user_embedding_index_dict, open(subject_youtube_user_embedding_index_dict, 'wb')) pickle.dump(youtube_item_embedding_index_dict, open(subject_youtube_item_embedding_index_dict, 'wb')) # 定义用户召回字典 user_recall_items_dict = collections.defaultdict(dict) if not os.path.exists(subject_youtubednn_recall_dict): # 加载保存的faiss索引 index = faiss.read_index(subject_youtube_item_faiss_model_path) # 通过user_embedding去查询最相似的topk+1个item_embedding sim, idx = index.search(np.ascontiguousarray(user_embs), topk+1) logger.info('生成YoutubeDNN所有用户的召回列表') for target_idx, sim_value_list, rele_idx_list in tqdm(zip(test_user_model_input['user_id'], sim, idx)): target_raw_id = user_index_2_rawid[target_idx] # 从1开始是为了去掉物品本身 for rele_idx, sim_value in zip(rele_idx_list[1:], sim_value_list[1:]): rele_raw_id = item_index_2_rawid[rele_idx] user_recall_items_dict[target_raw_id][rele_raw_id] = \ user_recall_items_dict.get(target_raw_id, {}).get(rele_raw_id, 0) + sim_value # 将召回的结果进行排序，直接得到召回的结果 user_recall_items_dict = {k: sorted(v.items(), key=lambda x: x[1], reverse=True) \ for k, v in user_recall_items_dict.items()} pickle.dump(user_recall_items_dict, open(subject_youtubednn_recall_dict, 'wb')) return user_recall_items_dict def build_hnsw(vecs, to_file, ef=2000, m=64, dim=100): """ 训练hnsw模型 """ vecs = vecs.astype('float32') # 构建索引 index = faiss.IndexFlatIP(dim) index.add(vecs) # 保存hnsw模型 faiss.write_index(index, to_file) return index def youtubednn_recall_train(): """ youtubednn召回训练和评估 """ # 需要召回的数量 recall_item_num = 100 logger.info("加载物品行为数据") all_select_df = get_all_select_df(offline=offline_mode) logger.info("加载物品信息数据") item_info = get_item_info_df() logger.info("获取用户信息数据") users_info = get_user_info_df() all_select_df = all_select_df.merge(users_info, on='user_id') all_select_df = all_select_df.merge(item_info, on='subject_id') # 为了召回评估，提取最后一次选择作为召回评估 # 如果不需要做召回评估直接使用全量的训练集进行召回 if need_metric_recall: logger.info('获取物品行为数据历史和最后一次选择') train_hist_select_df, train_last_select_df =get_all_hist_and_last_select(all_select_df) else: train_hist_select_df = all_select_df # YoutubeDNN模型训练 user_recall_items_dict = youtubednn_recall(train_hist_select_df, topk=recall_item_num) # 召回效果评估 if samples_mode == True and need_metric_recall: logger.info('YoutebeDNN召回效果评估') metrics_recall(user_recall_items_dict, train_last_select_df, topk=recall_item_num) if __name__ == '__main__': youtubednn_recall_train()