Datawhale干货

作者：周远哲，斯特拉斯堡大学硕士

我的本科与硕士都是非科班，几乎没有相关的基础。 接触机器学习竞赛的半年左右的时候参加了去年kaggle规模最大的jane street量化大赛，比赛前期屠榜了几次，最终也拿到了冠军。与此同时也拿了一些其他比赛的top名次。本次比赛的主要目的还是学习，通过比赛学习一些CTR建模方面的基础知识。这一次比赛的队友很强，有幸再次拿到了好名次。

本次参加了微信大数据挑战赛，比赛中我只用到了一个模型（DCNv1），其他的时间更多是在探索和验证一些奇特的想法，初赛的方案没有构造任何手工特征，因此上分的速度相对于别的队伍要慢上很多。读者最关注的应该是模型的效果。这个方案的分数在复赛a榜，做bagging后可达到0.724+的成绩，在比赛中拿到亚军。下面具体介绍一下赛题及我的方案。后台回复 211209 可获取完整代码。

代码地址：

https://aistudio.baidu.com/aistudio/projectdetail/2536106

实践背景

赛题背景

本次比赛基于脱敏和采样后的数据信息，对于给定的一定数量到访过微信视频号“热门推荐”的用户， 根据这些用户在视频号内的历史n天的行为数据，通过算法在测试集上预测出这些用户对于不同视频内容的互动行为（包括点赞、点击头像、收藏、转发等）的发生概率。本次比赛以多个行为预测结果的加权uAUC值进行评分

比赛提供训练集用于训练模型，测试集用于评估模型效果，提交结果demo文件用于展示提交结果的格式。所有数据文件格式都是带表头的.csv格式，不同字段列之间用英文逗号分隔。初赛与复赛的数据分布一致，数据规模不同。初赛提供百万级训练数据，复赛提供千万级训练数据。

赛题地址：https://algo.weixin.qq.com/

评价指标

本次比赛采用uAUC作为单个行为预测结果的评估指标，uAUC定义为不同用户下AUC的平均值，计算公式如下：

其中，n为测试集中的有效用户数，有效用户指的是对于某个待预测的行为，过滤掉测试集中全是正样本或全是负样本的用户后剩下的用户。AUCi为第i个有效用户的预测结果的AUC（Area Under Curve）。AUC的定义和计算方法可参考维基百科。

初赛的最终分数为4个行为（查看评论、点赞、点击头像、转发）的uAUC值的加权平均。复赛的最终分数为7个行为（查看评论、点赞、点击头像、转发、收藏、评论和关注）的uAUC值的加权平均。分数越高，排名越靠前。

性能要求

出于性能评估的可操作性考虑，本次比赛只要求晋级决赛的Top6队伍需满足最低性能标准，晋级空缺名额后补。组委会届时将对队伍提交的代码和模型进行性能评估。关于性能评估的具体说明如下：性能评估方法：

• 在组委会指定的机器上（2卡P40 48G显存 14核CPU 112G内存），预测给定测试数据集7个目标行为的概率，记录预测时长（只统计模型推理部分耗时，不包含数据处理、特征提取等部分耗时），并计算单个目标行为2000条样本的平均预测时长（单位：毫秒）。

• 最低性能要求：单个目标行为2000条样本的预测时长不超过200毫秒。

赛题难点

1. 如何利用多模态向量。

2. 如何加速训练，提高迭代效率，同时节省内存。

3. 如何做采样？

4. 预训练向量存在信息泄漏，导致训练过程中过拟合训练集怎么办？

赛事总结

以往的大数据挑战赛只允许在校学生参与，这一次的大数据挑战赛在职人员也可以参加，竞争更加激烈，参与数达到了惊人的6768支队伍，比赛第一天的有效提交队伍数便突破了1000，国内很多推荐算法相关的从业人员都或多或少会了解或参与，可能是今年参与人数最多的比赛了，同时也是数据挖掘爱好者证明自己实力最好的舞台。

本次大数据挑战赛的赛事方对于比赛的公平公正做了许多努力，同时提供了大量充足的算力支持（有了112G内存，写代码不需要那么抠抠搜搜了）。赛事方针对部分队伍可能利用b榜测试集数据的泄漏信息进行特征工程的情况，采取许多必要的措施。最终的比赛b榜，选手只允许使用训练好的数据进行预测，测试集的时间信息被去除，同时ab榜采用了完全不同的用户集合，既考验了选手的工程能力，也考验了选手的的建模能力。

代码实践

读取数据

比赛时能用到的有效数据越多越好，多多益善。这里我们把初赛数据也用上，用于预训练词向量。

这次分享不提供比赛数据集，使用生成的dummy数据代替。代码按照复赛的格式实现。

!pip install gensimimport pandas as pd
import numpy as np
from gensim.models import Word2Vec
from gensim.models import KeyedVectors# notebook掉线的话会无法监测进度，因此把训练进度打印进一个文件。当然这里也可以用nohup挂起运行。
# !pip install loguru -i http://mirrors.tencentyun.com/pypi/simple
# from loguru import loggerimport os
import gcimport time
import traceback
from collections import defaultdict
from sklearn.metrics import roc_auc_scorefrom paddle.io import Dataset
from paddle.static import InputSpecimport paddle
import paddle.fluid as fluidfrom paddle.fluid.dygraph import Linear
from paddle.fluid.dygraph import Layer, to_variable
import paddle.fluid.dygraph as dygraph# paddle.enable_static()# 读取测试集
test_a = pd.read_csv('./test_a.csv')# 读取初赛数据
test_pre_a = pd.read_csv('wbdc2021/data/wedata/wechat_algo_data1/test_a.csv')
test_pre_b = pd.read_csv('wbdc2021/data/wedata/wechat_algo_data1/test_b.csv')
test_pre_a = test_pre_a.append(test_pre_b)# 读取多模态向量
feed_embedding = pd.read_csv('./feed_embeddings.csv')# 读取初赛、复赛训练集
user_action = pd.read_csv('./user_action.csv')
user_action_prelimi = pd.read_csv('wbdc2021/data/wedata/wechat_algo_data1/user_action.csv')
user_action = user_action.append(user_action_prelimi)# 读取feed信息
feed_info = pd.read_csv('./feed_info.csv')N_test = test_a.shape[0]# 读取多模态词向量为词典
feed_id_embedding_dict = {}# 使用词典存储多模态向量
for id, emb in feed_embedding.values:feed_id_embedding_dict[id] = np.array(emb.split(' ')[:512]).astype('float32')
del feed_embedding# uAUC评测函数
def uAUC(labels, preds, user_id_list):"""Calculate user AUC"""user_pred = defaultdict(lambda: [])user_truth = defaultdict(lambda: [])for idx, truth in enumerate(labels):user_id = user_id_list[idx]pred = preds[idx]truth = labels[idx]user_pred[user_id].append(pred)user_truth[user_id].append(truth)user_flag = defaultdict(lambda: False)for user_id in set(user_id_list):truths = user_truth[user_id]flag = False# 若全是正样本或全是负样本，则flag为Falsefor i in range(len(truths) - 1):if truths[i] != truths[i + 1]:flag = Truebreakuser_flag[user_id] = flagtotal_auc = 0.0size = 0.0for user_id in user_flag:if user_flag[user_id]:auc = roc_auc_score(np.asarray(user_truth[user_id]), np.asarray(user_pred[user_id]))total_auc += auc size += 1.0user_auc = float(total_auc)/sizereturn user_auc# 统一设置一些超参数N_USERID = 250250
N_FEEDID = 112871  # feedid个数
N_AUTHORSONGSINGER = 30000EMBEDDING_SIZE_FEEDID_PRETRAINED = 256  # 预训练feedid的维度N_duomo = 64   # 多模态向量降维
PCA_DIM = 64   # 文本类向量降维
EMBEDDING_SIZE_AUTHORID = 128
EMBEDDING_SIZE_SONGSINGERID = 32
EMBEDDING_SIZE_KEYWORDTAG = 20BATCH_SIZE = 8000
EPOCHS = 2

预训练词向量

# 预训练feedid词向量
# 这里没有对id的顺序进行处理，按默认的顺序训练词向量。if os.path.exists('./feedmodel_cbow.model'):model_sg = Word2Vec.load('./feedmodel_cbow.model')
else:feedid_seq_list = np.concatenate([user_action.groupby('userid').feedid.apply(lambda x: [str(id) for id in x] ).values,test_a.groupby('userid').feedid.apply(lambda x: [str(id) for id in x] ).values, test_pre_a.groupby('userid').feedid.apply(lambda x: [str(id) for id in x] ).values])# 比赛中使用的原始参数model_sg = Word2Vec(feedid_seq_list, vector_size=EMBEDDING_SIZE_FEEDID_PRETRAINED, window=32, min_count=1, sg=0, sample=1e-3, negative=15, workers=32, seed=1, epochs=10)model_sg.save('./feedmodel_cbow.model')# 处理一下缺失值，这里的处理比较粗糙
user_id_a = set(list(test_a.userid.values))
user_id_train = set(list(user_action.userid.values))
feed_id_a = set(list(test_a.feedid.values))
feed_id_train = set(list(user_action.feedid.values))feed_info['authorid'].fillna(feed_info['authorid'].max()+1, inplace=True)
feed_info['bgm_song_id'].fillna(feed_info['bgm_song_id'].max()+1, inplace=True)
feed_info['bgm_singer_id'].fillna(feed_info['bgm_singer_id'].max()+1, inplace=True)feedid_authorid_dict = dict(zip(list(feed_info['feedid'].values), list(feed_info['authorid'].values)))
feedid_bgm_song_id_dict = dict(zip(list(feed_info['feedid'].values), list(feed_info['bgm_song_id'].values)))
feedid_bgm_singer_id_dict = dict(zip(list(feed_info['feedid'].values), list(feed_info['bgm_singer_id'].values)))
feedid_vps_dict = dict(zip(list(feed_info['feedid'].values), list(feed_info['videoplayseconds'].values)))# 将其他的id类特征拼接入训练数据。这里没有使用pandas自带的merge函数，而是使用词典映射，避免了OOM的问题。
# 这里实际上可以用embedding layer来做映射处理，但是比赛中遇到了一个bug没有调试通，因此保留了这些id特征作为输入。user_action['authorid'] = user_action['feedid'].map(dict(zip(feed_info.feedid.values, feed_info.authorid.values)))
user_action['bgm_song_id'] = user_action['feedid'].map(dict(zip(feed_info.feedid.values, feed_info.bgm_song_id.values)))
user_action['bgm_singer_id'] = user_action['feedid'].map(dict(zip(feed_info.feedid.values, feed_info.bgm_singer_id.values)))
user_action['vps'] = user_action['feedid'].map(feedid_vps_dict).astype('float32')
user_action['stay/vps'] = user_action['stay']/1000/user_action['vps']# 对测试集数据用相同的方式处理
test_a['authorid'] = test_a['feedid'].map(dict(zip(feed_info.feedid.values, feed_info.authorid.values)))
test_a['bgm_song_id'] = test_a['feedid'].map(dict(zip(feed_info.feedid.values, feed_info.bgm_song_id.values)))
test_a['bgm_singer_id'] = test_a['feedid'].map(dict(zip(feed_info.feedid.values, feed_info.bgm_singer_id.values)))test_pre_a['authorid'] = test_pre_a['feedid'].map(dict(zip(feed_info.feedid.values, feed_info.authorid.values)))
test_pre_a['bgm_song_id'] = test_pre_a['feedid'].map(dict(zip(feed_info.feedid.values, feed_info.bgm_song_id.values)))
test_pre_a['bgm_singer_id'] = test_pre_a['feedid'].map(dict(zip(feed_info.feedid.values, feed_info.bgm_singer_id.values)))# 文本类向量使用PCA进行处理from sklearn import feature_extraction
from sklearn.feature_extraction.text import TfidfTransformer, TfidfVectorizer
from sklearn.feature_extraction.text import CountVectorizer
import os
from sklearn.decomposition import PCAfeed_info['description'].fillna('0', inplace=True)
feed_info['ocr'].fillna('0', inplace=True)
feed_info['asr'].fillna('0', inplace=True)if os.path.exists('./tfidf_pca.npy'):tfidf_pca = np.load('tfidf_pca.npy')
else:vectorizer = CountVectorizer()transformer = TfidfVectorizer(min_df=3, max_df=0.9, use_idf=1, smooth_idf=1, sublinear_tf=1)corpus= feed_info[~feed_info.description.isna()].description.valuestfidf = transformer.fit_transform(corpus)Nrow = tfidf.shape[0]from sklearn.decomposition import TruncatedSVDclf = TruncatedSVD(PCA_DIM)tfidf_pca = clf.fit_transform(tfidf)np.save('tfidf_pca.npy', tfidf_pca)# 将降低维度后的文本类向量保存入array
embedding_matrix5 = np.zeros((N_FEEDID + 2, PCA_DIM))
for i, id in enumerate(feed_info.feedid.values):try:embedding_matrix5[id+1] = tfidf_pca[i].astype('float32')except:pass

构建embedding层

# 将降低多模态向量保存入array
embedding_matrix3 = np.zeros((N_FEEDID + 2, 512))
for i in range(N_FEEDID):try:embedding_matrix3[i+1] = feed_id_embedding_dict[i].astype('float32')except:pass# 使用PCA对多模态向量降维。
if N_duomo!=512:pca = PCA(n_components=N_duomo)embedding_matrix3 = pca.fit_transform(embedding_matrix3)# 拼接两组向量
embedding_matrix3 = np.hstack([embedding_matrix3, embedding_matrix5])# 构建embedding layer
embedding_layer_fid_multimodal = paddle.nn.Embedding(num_embeddings=N_FEEDID+2,embedding_dim=N_duomo + PCA_DIM,padding_idx=0,)
embedding_layer_fid_multimodal.weight.set_value(embedding_matrix3.astype('float32'))# 将预训练的feedid向量保存入array
embedding_matrix = np.zeros((N_FEEDID + 2, EMBEDDING_SIZE_FEEDID_PRETRAINED))
for i in range(N_FEEDID):try:embedding_matrix[i+1] = model_sg.wv[str(i)].astype('float32')except:pass# 构建embedding layer
embedding_layer_fid_pretrained = paddle.nn.Embedding(num_embeddings=N_FEEDID+2,embedding_dim=EMBEDDING_SIZE_FEEDID_PRETRAINED,padding_idx=0,)
embedding_layer_fid_pretrained.weight.set_value(embedding_matrix.astype('float32'))

构建模型输入

# 训练集数据
X_train = user_action.sort_values(by=['date_', 'userid'])[['userid', 'feedid']].copy()
X_train['authorid'] = X_train['feedid'].map(feedid_authorid_dict).astype('float32')
X_train['songid'] = X_train['feedid'].map(feedid_bgm_song_id_dict).astype('float32')
X_train['singerid'] = X_train['feedid'].map(feedid_bgm_singer_id_dict).astype('float32')
X_train['vps'] = X_train['feedid'].map(feedid_vps_dict).astype('float32')
X_train['device'] = user_action.sort_values(by=['date_', 'userid'])['device'].astype('float32')y_train = user_action.sort_values(by=['date_', 'userid'])[['read_comment', 'like', 'click_avatar', 'forward', 'favorite', 'comment', 'follow', ]].copy()# 测试集X_test = test_a[['userid', 'feedid']].copy()
X_test['authorid'] = X_test['feedid'].map(feedid_authorid_dict).astype('float32')
X_test['songid'] = X_test['feedid'].map(feedid_bgm_song_id_dict).astype('float32')
X_test['singerid'] = X_test['feedid'].map(feedid_bgm_singer_id_dict).astype('float32')
X_test['vps'] = X_test['feedid'].map(feedid_vps_dict).astype('float32')
X_test['device'] = test_a['device'].astype('float32')X_train = X_train.values
X_test = X_test.values

构建feedid与tag/keyword的映射

# 构建keyword的embedding层def operate(X):xl = X.split(';')xl = [int(x.split(' ')[0]) for x in xl if float(x.split(' ')[1])>0.5 ]return xl# 将tag key word进行拼接
manual_tag_list = feed_info.manual_tag_list.fillna('0;0').apply(lambda x: [int(i) for i in (x).split(';')]).values
machine_tag_list_ = feed_info.machine_tag_list.fillna('0 0;0 0').apply(lambda x: operate(x)).valuesmachine_keyword_list = feed_info.machine_keyword_list.fillna('0;0').apply(lambda x: [int(i) for i in (x).split(';')]).values
manual_keyword_list = feed_info.manual_keyword_list.fillna('0;0').apply(lambda x: [int(i) for i in (x).split(';')]).valuesdef padding(x, N=EMBEDDING_SIZE_KEYWORDTAG):if len(x)>N:return x[:N]else:return x+[0 for i in range(N-len(x))]key_list = []for l1, l2, l3, l4 in zip(machine_keyword_list, manual_keyword_list, manual_tag_list, machine_tag_list_):tmp = l2tmp2 = l3key_list.append(padding(tmp, EMBEDDING_SIZE_KEYWORDTAG)+padding(tmp2, EMBEDDING_SIZE_KEYWORDTAG))feedid_keyword_list_dict = dict(zip(list(feed_info.feedid.values), key_list))embedding_matrix_exex = np.zeros((N_FEEDID + 2, EMBEDDING_SIZE_KEYWORDTAG*2))
for i in range(N_FEEDID):try:embedding_matrix_exex[i+1] = np.array(feedid_keyword_list_dict[i]).astype('int32')except:pass# 使用embedding层来映射id
embedding_layer_fid_keytag = paddle.nn.Embedding(num_embeddings=N_FEEDID+2,embedding_dim=EMBEDDING_SIZE_KEYWORDTAG*2,padding_idx=0,)
embedding_layer_fid_keytag.weight.set_value(embedding_matrix_exex.astype('float32'))import gc
gc.collect()

author、song、singer的预训练embedding

# 以 userid 维度聚类，训练authour、song、singer id 的词向量。
# 这里给予authorid的词向量维度更高一些，是实验得出的结论。 if os.path.exists('./author_cbow.model'):model_author_cbow  = Word2Vec.load('./author_cbow.model')
else:tmp = user_action.groupby('userid').authorid.apply(lambda x: [str(id) for id in x] ).valuestmp_ = test_a.groupby('userid').authorid.apply(lambda x: [str(id) for id in x] ).valuestmp = np.concatenate([tmp, tmp_])tmp_ = test_pre_a.groupby('userid').authorid.apply(lambda x: [str(id) for id in x] ).valuestmp = np.concatenate([tmp, tmp_])model_author_cbow = Word2Vec(tmp, vector_size=EMBEDDING_SIZE_AUTHORID, window=32, min_count=1, sg=0, sample=1e-3, negative=15, workers=32, seed=1, epochs=10)model_author_cbow.save('./author_cbow.model')# 映射词表
embedding_matrix_author = np.zeros((N_AUTHORSONGSINGER + 2, EMBEDDING_SIZE_AUTHORID))
for i in range(N_AUTHORSONGSINGER):try:embedding_matrix_author[i+1] =  model_author_cbow.wv[str(i)].astype('float32') except:pass# 构建embedding层
embedding_layer_author = paddle.nn.Embedding(num_embeddings=N_AUTHORSONGSINGER + 2,embedding_dim=EMBEDDING_SIZE_AUTHORID,padding_idx=0,weight_attr=paddle.ParamAttr(initializer=paddle.nn.initializer.Assign(embedding_matrix_author),trainable=False))
embedding_layer_author.weight.set_value(embedding_matrix_author.astype('float32'))if os.path.exists('./bgm_song_id_cbow.model'):model_bgm_song_id_cbow = Word2Vec.load('./bgm_song_id_cbow.model')
else:tmp = user_action.groupby('userid').bgm_song_id.apply(lambda x: [str(id) for id in x] ).valuestmp_ = test_a.groupby('userid').bgm_song_id.apply(lambda x: [str(id) for id in x] ).valuestmp = np.concatenate([tmp, tmp_])tmp_ = test_pre_a.groupby('userid').bgm_song_id.apply(lambda x: [str(id) for id in x] ).valuestmp = np.concatenate([tmp, tmp_])model_bgm_song_id_cbow = Word2Vec(tmp, vector_size=EMBEDDING_SIZE_SONGSINGERID, window=32, min_count=1, sg=0, sample=1e-3, negative=15, workers=32, seed=1, epochs=10)model_bgm_song_id_cbow.save('./bgm_song_id_cbow.model')# 映射词表
embedding_matrix_bgm_song_id = np.zeros((N_AUTHORSONGSINGER + 2, EMBEDDING_SIZE_SONGSINGERID))
for i in range(N_AUTHORSONGSINGER):try:embedding_matrix_bgm_song_id[i+1] =  model_bgm_song_id_cbow.wv[str(float(i))].astype('float32')except:pass# 构建embedding层
embedding_layer_song = paddle.nn.Embedding(num_embeddings=N_AUTHORSONGSINGER + 2,embedding_dim=EMBEDDING_SIZE_SONGSINGERID,padding_idx=0,)
embedding_layer_song.weight.set_value(embedding_matrix_bgm_song_id.astype('float32'))if os.path.exists('./bgm_singer_id_cbow.model'):model_bgm_singer_id_cbow = Word2Vec.load('./bgm_singer_id_cbow.model')
else:tmp = user_action.groupby('userid').bgm_singer_id.apply(lambda x: [str(id) for id in x] ).valuestmp_ = test_a.groupby('userid').bgm_singer_id.apply(lambda x: [str(id) for id in x] ).valuestmp = np.concatenate([tmp, tmp_])tmp_ = test_pre_a.groupby('userid').bgm_singer_id.apply(lambda x: [str(id) for id in x] ).valuestmp = np.concatenate([tmp, tmp_])model_bgm_singer_id_cbow = Word2Vec(tmp, vector_size=EMBEDDING_SIZE_SONGSINGERID, window=32, min_count=1, sg=0, sample=1e-3, negative=15, workers=32, seed=1, epochs=10)model_bgm_singer_id_cbow.save('./bgm_singer_id_cbow.model')# 映射词表
embedding_matrix_bgm_singer_id = np.zeros((N_AUTHORSONGSINGER + 2, EMBEDDING_SIZE_SONGSINGERID))
for i in range(N_AUTHORSONGSINGER):try:embedding_matrix_bgm_singer_id[i+1] = model_bgm_singer_id_cbow.wv[str(float(i))].astype('float32')except:pass
# 构建embedding层
embedding_layer_singer = paddle.nn.Embedding(num_embeddings=N_AUTHORSONGSINGER + 2,embedding_dim=EMBEDDING_SIZE_SONGSINGERID,padding_idx=0,)
embedding_layer_singer.weight.set_value(embedding_matrix_bgm_singer_id.astype('float32'))

TF-IDF

之前我们构筑了很多feedid的词向量特征，现在我们为userid构建tf-idf向量作为特征。

# 构建tf-idf向量作为特征。
# 词向量作为特征偏向于提升模糊泛化的效果，而tfidf则属于更加精准的匹配。
# 这里同样给予authorid更高的向量维度import osif os.path.exists('./tfidf_user.npy'):tfidf_userid_feat = np.load('tfidf_user.npy')
else:from tqdm import tqdmtfidf_userid_feat = []for group in tqdm(['authorid', 'bgm_song_id', 'bgm_singer_id',]):tmp = user_action[['userid', group]].sort_values(by='userid').groupby('userid')[group].apply(lambda x: ' '.join(map(str, x)))vectorizer = CountVectorizer()transformer = TfidfTransformer(sublinear_tf=True)from sklearn.decomposition import TruncatedSVDtfidf = transformer.fit_transform(vectorizer.fit_transform(tmp))if group =='authorid':clf = TruncatedSVD(EMBEDDING_SIZE_AUTHORID)else:clf = TruncatedSVD(EMBEDDING_SIZE_SONGSINGERID)tfidf_pca_author = clf.fit_transform(tfidf)tfidf_userid_feat.append(tfidf_pca_author)tfidf_userid_feat = np.hstack(tfidf_userid_feat)np.save('tfidf_user.npy', tfidf_userid_feat)# 映射词表
embedding_matrix_userid_tfidf = np.zeros((N_USERID+2, EMBEDDING_SIZE_AUTHORID+EMBEDDING_SIZE_SONGSINGERID*2))
for index, i in enumerate(sorted(user_action.userid.unique())):try:embedding_matrix_userid_tfidf[i+1] = tfidf_userid_feat[index].astype('float32')except:pass# 构建embedding层
embedding_layer_userid_tfidf = paddle.nn.Embedding(num_embeddings=N_USERID + 2,embedding_dim=EMBEDDING_SIZE_AUTHORID+EMBEDDING_SIZE_SONGSINGERID*2,padding_idx=0,)
embedding_layer_userid_tfidf.weight.set_value(embedding_matrix_userid_tfidf.astype('float32'))

模型结构与训练

模型结构_DCN模型

https://arxiv.org/abs/1708.05123

DCN 的全称是Deep & Cross Network，网络架构如下：

使用DCNv1的弊端也十分明显，模型本身17年就发表了，调参成本低，但是模型性能上限偏低。

模型结构_词向量

模型中使用的词向量种类包括了：

1. 按照userid聚合的feedid的预训练词向量。

2. feedid的文本类、多模态特征向量。

3. 按照userid聚合的authorid预训练词向量。

4. 按照userid聚合的songid预训练词向量。

5. 按照userid聚合的singerid预训练词向量。

6. 使用authorid、songid、singerid的tfidf表示的userid词向量。

7. 可训练的feedid向量。

8. 可训练的userid向量。

9. 可训练的tag、keyword向量。

10. 可训练的authorid、songid、singerid向量。

模型中没有使用点击率等统计类型的特征。

import numpy as np
import pandas as pd
from numba import njit
from scipy.stats import rankdata
from joblib import Parallel, delayed
import warnings
warnings.filterwarnings("ignore", category=FutureWarning)paddle.disable_static()@njit
def _auc(actual, pred_ranks):actual = np.asarray(actual)pred_ranks = np.asarray(pred_ranks)n_pos = np.sum(actual)n_neg = len(actual) - n_posreturn (np.sum(pred_ranks[actual == 1]) - n_pos*(n_pos+1)/2) / (n_pos*n_neg)def auc(actual, predicted):pred_ranks = rankdata(predicted)return _auc(actual, pred_ranks)class network(fluid.dygraph.Layer):def __init__(self, name_scope='baseline'):super(network, self).__init__(name_scope)name_scope = self.full_name()# 定义三层全连接层，输入维度是最终选取的特征数量，输出维度是1，激活函数为reluself.embedding_layer_feedid_trainable = paddle.nn.Embedding(N_FEEDID + 2, 64)self.embedding_layer_userid_trainable = paddle.nn.Embedding(N_USERID + 2, 128)self.embedding_layer_author_trainable = paddle.nn.Embedding(N_AUTHORSONGSINGER+2, 64)self.embedding_layer_song_trainable = paddle.nn.Embedding(N_AUTHORSONGSINGER+2, 16)self.embedding_layer_singer_trainable = paddle.nn.Embedding(N_AUTHORSONGSINGER+2, 16)self.embedding_layer_author = embedding_layer_author # 128self.embedding_layer_song = embedding_layer_song # 32self.embedding_layer_singer = embedding_layer_singer # 32self.embedding_layer_userid_tfidf = embedding_layer_userid_tfidf # 128 + 32 + 32self.embedding_layer_fid_pretrained = embedding_layer_fid_pretrained # 256self.embedding_layer_fid_multimodal = embedding_layer_fid_multimodal # 128self.embedding_layer_fid_keytag = embedding_layer_fid_keytag # ...self.embedding_layer_tag = paddle.nn.Embedding(N_AUTHORSONGSINGER+2, 32)self.embedding_layer_keyword = paddle.nn.Embedding(N_AUTHORSONGSINGER+2, 32)self.x0_size = 128+64+64+16+16+128+32+32+128+32+32+256+128+32+32+2self.mlp1 = paddle.fluid.dygraph.Linear(self.x0_size, 256, act='relu')self.mlp_bn1 = paddle.nn.BatchNorm(256)self.mlp2 = paddle.fluid.dygraph.Linear(256, 256, act='relu')self.mlp_bn2 = paddle.nn.BatchNorm(256)self.cross1 = paddle.fluid.dygraph.Linear(self.x0_size, self.x0_size, )self.cross2 = paddle.fluid.dygraph.Linear(self.x0_size, self.x0_size, )self.cross3 = paddle.fluid.dygraph.Linear(self.x0_size, self.x0_size, )self.cross4 = paddle.fluid.dygraph.Linear(self.x0_size, self.x0_size, )self.cross5 = paddle.fluid.dygraph.Linear(self.x0_size, self.x0_size, )self.bn_output = paddle.nn.BatchNorm(256+self.x0_size)self.output = paddle.fluid.dygraph.Linear(self.x0_size+256, 7, )self.sigmoid = paddle.nn.Sigmoid()# 网络的前向计算函数def forward(self, inputs):        input_userid = paddle.cast(inputs[:,0]+1, dtype='int64')input_feedid = paddle.cast(inputs[:,1]+1, dtype='int64')input_authorid = paddle.cast(inputs[:,2]+1, dtype='int64')input_songid = paddle.cast(inputs[:,3]+1, dtype='int64')input_singerid = paddle.cast(inputs[:,4]+1, dtype='int64')embedding_userid_trainable = self.embedding_layer_userid_trainable(input_userid)embedding_userid_tfidf = self.embedding_layer_userid_tfidf(input_userid)embedding_userid_tfidf.stop_gradient=Falseembedding_feedid_trainable = self.embedding_layer_feedid_trainable(input_feedid)embedding_fid_multimodal = self.embedding_layer_fid_multimodal(input_feedid)embedding_fid_pretrained = self.embedding_layer_fid_pretrained(input_feedid)embedding_fid_multimodal.stop_gradient=Falseembedding_fid_pretrained.stop_gradient=Falseembedding_author = self.embedding_layer_author(input_authorid+1)embedding_song = self.embedding_layer_song(input_songid+1)embedding_singer = self.embedding_layer_singer(input_singerid+1)embedding_author_trainable = self.embedding_layer_author_trainable(input_authorid+1)embedding_song_trainable = self.embedding_layer_song_trainable(input_songid+1)embedding_singer_trainable = self.embedding_layer_singer_trainable(input_singerid+1)input_tag_keyword = paddle.cast(self.embedding_layer_fid_keytag(input_feedid), dtype='int64')tag_embedding = paddle.sum(self.embedding_layer_tag(input_tag_keyword[:,20:]), axis=1)/20.keyword_embedding = paddle.sum(self.embedding_layer_keyword(input_tag_keyword[:,:20]), axis=1)/20.embedding_author.stop_gradient=Falseembedding_song.stop_gradient=Falseembedding_singer.stop_gradient=Falseinput_tag_keyword.stop_gradient=Falsedense_input = inputs[:,5:]x0 = paddle.fluid.layers.concat([embedding_userid_trainable,embedding_userid_tfidf,embedding_feedid_trainable,embedding_fid_multimodal,embedding_fid_pretrained,embedding_author,embedding_song,embedding_singer,embedding_author_trainable,embedding_song_trainable,embedding_singer_trainable,tag_embedding,keyword_embedding,dense_input], axis=-1)x0 = paddle.cast(x0, dtype='float32')x_mlp = self.mlp1(paddle.fluid.layers.dropout(x0, dropout_prob=0.4))x_mlp = self.mlp_bn1(x_mlp)x_mlp = self.mlp2(paddle.fluid.layers.dropout(x_mlp, dropout_prob=0.4))x_mlp = self.mlp_bn2(x_mlp)cross = x0cross = cross + x0 * self.cross1( paddle.fluid.layers.dropout(cross, dropout_prob=0.15) )cross = cross + x0 * self.cross2( paddle.fluid.layers.dropout(cross, dropout_prob=0.15) )cross = cross + x0 * self.cross3( paddle.fluid.layers.dropout(cross, dropout_prob=0.15) )cross = cross + x0 * self.cross4( paddle.fluid.layers.dropout(cross, dropout_prob=0.15) )cross = cross + x0 * self.cross5( paddle.fluid.layers.dropout(cross, dropout_prob=0.15) )x = paddle.fluid.layers.concat([cross, x_mlp],axis=-1)x = self.bn_output(x)x = self.output(paddle.fluid.layers.dropout(x, dropout_prob=0.5))output = self.sigmoid(x) output = paddle.cast(output, dtype='float32')return outputclass TrainDataset(Dataset):def __init__(self, x_train_array, y_train_array):# 样本数量self.training_data, self.training_label = x_train_array.astype('float32'), y_train_array.astype('float32')self.num_samples = self.training_data.shape[0] def __getitem__(self, idx):data = self.training_data[idx]label = self.training_label[idx]return data,  labeldef __len__(self):# 返回样本总数量return self.num_samplesfor i in range(14):try:import gcdel x_train, x_validgc.collect()del y_train_, y_valid_gc.collect()except:pass# 这里没有使用验证集，只是将第i天的负样本去掉了index_val = (user_action.date_==i).valuesindex_train = (~index_val)x_train, x_valid = X_train[index_train], X_train[index_val]y_train_, y_valid_ = y_train.values[index_train].astype('float32'), y_train.values[index_val].astype('float32')train_dataset = TrainDataset(x_train, y_train_)train_loader = paddle.io.DataLoader(train_dataset, batch_size=1024*8, shuffle=True)valid_dataset = TrainDataset(x_valid, y_valid_)valid_loader = paddle.io.DataLoader(valid_dataset, batch_size=1024*8, shuffle=False)input = [InputSpec(shape=[None, 7], dtype='float32',name='inputs')]model = paddle.Model(network(), input)model.prepare(optimizer=paddle.optimizer.Adam(learning_rate=0.002, parameters=model.parameters()),loss=paddle.nn.BCELoss(),)model.fit(train_data=train_loader, eval_data=valid_loader,epochs=2, verbose=1,)model.save('./model_weight_{}'.format(i), False)del train_datasetdel train_loaderdel modelgc.collect()y_test_preds = []for i in range(14):model = paddle.jit.load('./model_weight_{}'.format(i)).eval()preds = model(X_test.astype('float32'))y_test_preds.append(preds)test_prediction = np.mean(y_test_preds, axis=0)

整理不易，点赞三连↓