IJCAI 2018 阿里妈妈广告预测算法大赛

背景

本项目是天池的一个比赛，由阿里妈妈和天池大数据众智平台举办的广告预测算法大赛，本次参赛人数5200多个队伍，而我们只取得了731的成绩，最遗憾的是当我们写好CNN预测结果准备上传，队伍却意外解散，哈哈。虽然没-能进入复赛，但是过程还是很有意思，特此记录一下。

目标

本次比赛以阿里电商广告为研究对象，提供了淘宝平台的海量真实交易数据，参赛选手通过人工智能技术构建预测模型预估用户的购买意向，即给定广告点击相关的用户（user）、广告商品（ad）、检索词（query）、上下文内容（context）、商店（shop）等信息的条件下预测广告产生购买行为的概率（pCVR），形式化定义为：pCVR=P(conversion=1/query, user, ad, context, shop)。

结合淘宝平台的业务场景和不同的流量特点，我们定义了以下两类挑战：
- （1）日常的转化率预估
- （2）特殊日期的转化率预估

评估指标，公式如下：

通过logarithmic loss（记为logloss）评估模型效果（越小越好）
其中N表示测试集样本数量，yi表示测试集中第i个样本的真实标签，pi表示第i个样本的预估转化率。

数据说明

本次比赛为参赛选手提供了5类数据（基础数据、广告商品信息、用户信息、上下文信息和店铺信息）。基础数据表提供了搜索广告最基本的信息，以及“是否交易”的标记。广告商品信息、用户信息、上下文信息和店铺信息等4类数据，提供了对转化率预估可能有帮助的辅助信息。

1 . 基础数据

字段	解释
instance_id	样本编号，Long
is_trade	是否交易的标记位，Int类型；取值是0或者1，其中1 表示这条样本最终产生交易，0 表示没有交易
item_id	广告商品编号，Long类型
user_id	用户的编号，Long类型
context_id	上下文信息的编号，Long类型
shop_id	店铺的编号，Long类型

2. 广告商品信息

字段	解释
item_id	广告商品编号，Long类型
item_category_list	广告商品的的类目列表，String类型；从根类目（最粗略的一级类目）向叶子类目（最精细的类目）依次排列，数据拼接格式为 “category_0;category_1;category_2”，其中 category_1 是 category_0 的子类目，category_2 是 category_1 的子类目
item_property_list	广告商品的属性列表，String类型；数据拼接格式为 “property_0;property_1;property_2”，各个属性没有从属关系
item_brand_id	广告商品的品牌编号，Long类型
item_city_id	广告商品的城市编号，Long类型
item_price_level	广告商品的价格等级，Int类型；取值从0开始，数值越大表示价格越高
item_sales_level	广告商品的销量等级，Int类型；取值从0开始，数值越大表示销量越大
item_collected_level	广告商品被收藏次数的等级，Int类型；取值从0开始，数值越大表示被收藏次数越大
item_pv_level	广告商品被展示次数的等级，Int类型；取值从0开始，数值越大表示被展示次数越大

3. 用户信息

字段	解释
user_id	用户的编号，Long类型
user_gender_id	用户的预测性别编号，Int类型；0表示女性用户，1表示男性用户，2表示家庭用户
user_age_level	用户的预测年龄等级，Int类型；数值越大表示年龄越大
user_occupation_id	用户的预测职业编号，Int类型
user_star_level	用户的星级编号，Int类型；数值越大表示用户的星级越高

4. 上下文信息

字段	解释
context_id	上下文信息的编号，Long类型
context_timestamp	广告商品的展示时间，Long类型；取值是以秒为单位的Unix时间戳，以1天为单位对时间戳进行了偏移
context_page_id	广告商品的展示页面编号，Int类型；取值从1开始，依次增加；在一次搜索的展示结果中第一屏的编号为1，第二屏的编号为2
predict_category_property	根据查询词预测的类目属性列表，String类型；数据拼接格式为 “category_A:property_A_1,property_A_2,property_A_3;category_B:-1;category_C:property_C_1,property_C_2” ，其中 category_A、category_B、category_C 是预测的三个类目；property_B 取值为-1，表示预测的第二个类目 category_B 没有对应的预测属性

5. 店铺信息

字段	解释
shop_id	店铺的编号，Long类型
shop_review_num_level	店铺的评价数量等级，Int类型；取值从0开始，数值越大表示评价数量越多
shop_review_positive_rate	店铺的好评率，Double类型；取值在0到1之间，数值越大表示好评率越高
shop_star_level	店铺的星级编号，Int类型；取值从0开始，数值越大表示店铺的星级越高
shop_score_service	店铺的服务态度评分，Double类型；取值在0到1之间，数值越大表示评分越高
shop_score_delivery	店铺的物流服务评分，Double类型；取值在0到1之间，数值越大表示评分越高
shop_score_description	店铺的描述相符评分，Double类型；取值在0到1之间，数值越大表示评分越高

思路

我们的实验思路如下：

统计分析 -> 数据预处理 -> 特征抽取 -> 特征表示 -> 模型拟合和预测 -> 模型选择

其实从实验思路我们可以明显看出特征工程在这次比赛尤为重要，只有刻画好特征，才能利用模型得到好的预测结果，接下来我将按照实验思路进行总结。

实验

1. 统计分析

目的：看清数据分布，了解广告、商品、店铺、用户与购买概率的关系

基础数据的统计分析（饼图、柱状图和折线图结合) ，将数据按照is_trade属性分为两张子表，分别进行对比统计分析

购买的用户分析：
单变量：性别分布、年龄分布、职业分布、星级分布
交叉变量：（重点）性别-年龄、性别-星级、年龄-星级、职业-星级,（参考）年龄-职业，性别-职业
购买的商品分布对比
（重点）标签分布、属性分布、品牌分布、价格分布、销量分布，展示次数（后四项需考虑粒度的粗细）
（参考）城市分布、收藏次数分布
上下文信息对比
（重点）时间戳分布
（参考）页面分布（看能否精确到类别）、预测类目的准确度（？）
购买的店铺分布对比
（重点）评论数分布，好评率分布、星级分布
（参考）服务评分、物流评分、描述评分

实施：利用R对数据分布进行了统计

结果如下

以在不同属性上is_trade=0/1为例, 简要分析

1.转化分布

从上图可以明显看出在给定情景下转化率很低，也就是说，我们的训练数据存在了极度平衡的现象，甚至是可以把购买理解成异常值，我们的算法要能够极好的检测出异常实例。

2.年龄,性别,星级分布

从图1可以明显看出，年龄越大，转化率先增加后减少（-1表示未知年龄），这个结果与我们常识一致，中间年龄段更具有消费能力，性别转化分布没有贴出来，结果跟我们常识也是一致的，女性转化率高于男性。从图2中可以看出，星级越高购买率相对要更高一些，但是差距不太明显。

3.价格,收藏,展示分布

图1，可以看出价格越高转化率先增加后降低，这与我们对电商平台的认知有关，价格太低必然会让人觉得物品质量不佳，但是随着价格增加，购买会带来更高的风险，转化率自然会降低。图2收藏次数越高，购买的可能性越大，收藏在电商市场的本质，就是商品入选了用户的购买集，对相关商品综合排序后，收藏的商品更有可能转化。图3，总体趋势是展示次数（广告效应）越多，购买率越高。

4.商店星级,评论数量分布

图1.商店星级差异不明显。图2.评论数量居中的购买率更高

5 城市 | 商品标签分布

这两幅图是仅仅选择了高频的城市和商标分布，可以看出城市和商品图，都有集中表现类，而商品更为明显。

总结：数据统计分析的目的是分析变量之间的关系，观察具体特征对转化率的影响，从而用于模型中初始化权重

2. 数据预处理

处理缺失值
主要处理缺失值，以及属性值为-1的值，因为后期特征表示时，我们调用的sklearn借口进行one-hot表征，而接口要求输入数据不包括负数
特征映射
由于城市和商品的值字段太长，在表征时会出现错误，因此将他们分别映射，并更新原始数据，代码如下：


def map_field(train_data, test_data, path):train_data = set(train_data.unique())test_data = set(test_data.unique())all_property = list(train_data.union(test_data))print(len(all_property))map_data = {}for i in range(len(all_property)):map_data[str(all_property[i])] = iwith open(path, "w", encoding="utf-8") as dump:json.dump(map_data, dump)return map_datadef update_data(raw_data, field_1, field_2, path_1, path_2):with open(path_1, "r", encoding="utf-8") as dump:update_field_1 = json.load(dump)with open(path_2, "r", encoding="utf-8") as dump:update_field_2 = json.load(dump)raw_data[field_1] = raw_data[field_1].apply(lambda x: update_field_1[str(x)])raw_data[field_2] = raw_data[field_2].apply(lambda x: update_field_2[str(x)])return raw_data

3. 特征抽取

本小节主要涉及对特征的转化和抽取，比如在上下文中的时间轴数据，考虑到节假日流量问题（比赛提出的挑战解决方法），我节假日和周末前后时间戳进行映射，代码见下：

def convert_interval_time(hour, size=3):"""方法：把一天24小时按照力粒度为size大小进行分割:param hour::param size::return:"""interval_time = [list(range(i, i + size)) for i in range(0, 24, size)]interval_time_factor = {}for i in range(len(interval_time)):interval_time_factor[i] = interval_time[i]for factor, h in interval_time_factor.items():if hour in h:return factorelse:passdef time_to_factor(self):"""方法：将时间戳转为区间因子tips: add two features, is_weekend and is_holiday:param data: 输入带时间戳的数据,:param size:将原始时间戳转化为粒度为size个小时一个因子,默认为size=3:return: 对于时间戳转为化为区间因子"""data_time = self.raw_data[self.modify_features["time"]]data_time["format_time"] = data_time["context_timestamp"].apply(lambda x: time.localtime(x))convert_data_time = pd.DataFrame(columns=["interval_time", "is_weekends", "is_holidays"])convert_data_time["is_weekends"] = \data_time["format_time"].apply(lambda x: 1 if x[6] in self.weekends else 0)convert_data_time["is_holidays"] = \data_time["format_time"].apply(lambda x: 1 if str(x[1]) + "." + str(x[2]) in self.holidays else 0)convert_data_time["interval_time"] = \data_time["format_time"].apply(lambda x: self.convert_interval_time(hour=x[3]))return convert_data_time

属性值除了离散的，还包括的连续属性值，比如店铺的好评率、服务态度评分等等，连续属性离散化的代码如下：

def continuous_var_to_factor(self, size=0.01):"""方法：将连续数据按照粒度为size进行转化:param data_continuous: 输入带有连续变量的数据:param properity: 指定属性:param size： 粒度默认为0.01:return:"""data_continuous = self.raw_data[self.modify_features['continuous']]data_continuous = (data_continuous / size).round()data_continuous = data_continuous.astype('int64')return data_continuous

4. 特征表示

在这一部分，我们主要是是通过One-hot对所有数据特征进行表征，然后用One-hot的最大问题，尤其是在电商环境下特征表征，我们可以想象，这个数据维度非常巨大，所以，在处理这个高维数据时，我们先将其分为五个大的领域进行表征，再对其使用SVD进行降维处理。
One-hot表征代码如下：

def one_hot_represent(self, new_data, data_type='value'):""":param new_data: pd.DataFrame. The data used to transform to the one-hot form.:param data_type: str, default 'value', alternative 'lst'. Aimed at different forms, we can use this parameter tospecify the method of handling.:return:"""if data_type == 'value':df_one_hot = pd.DataFrame()for col in new_data.columns:# print(col)one_hot_matrix = self.one_hot_model.fit_transform(new_data[[col]])feature_names = [col + str(attr) for attr in self.one_hot_model.active_features_]one_hot_matrix = pd.DataFrame(one_hot_matrix.toarray(), columns=feature_names)df_one_hot = pd.concat([df_one_hot, one_hot_matrix], axis=1)return df_one_hot# return pd.DataFrame(df_one_hot.toarray(), columns=feature_names)elif data_type == 'lst':cols = new_data.columnsdf_one_hot = pd.DataFrame()for col in cols:# print(col)data = new_data[col]all_values = list(reduce(lambda x, y: x | y, data, set()))one_hot_dim = len(all_values)one_hot_matrix = []for line in data:one_hot_vec = np.zeros(one_hot_dim)for value in line:one_hot_vec[all_values.index(value)] = 1one_hot_matrix.append(one_hot_vec)one_hot_matrix = pd.DataFrame(one_hot_matrix, columns=all_values)df_one_hot = pd.concat([df_one_hot, one_hot_matrix], axis=1)return df_one_hotelse:raise ValueError('Can\'t recognize the type, please enter \'value\' or \'lst\'')

SVD降维代码如下：


class SVDReduce(object):"""The class is used to reduce the dimension of the data outputed from the class DataPreprocess with SVD method."""def __init__(self, data, dimension=500):"""Initialize the class with the parameters.:param data: pd.DataFrame, the output data from the class DataPreprocess.:param dimension: int, default 500. To specify the output dimension."""self.data = dataself.target_dim = dimensionself.format_data_path = '../../data/format_2/'self.field = ['user', 'product', 'context', 'shop']# self.field = ['product']def judge(self, data):"""Abandon方法：判读大领域的维度标准维度,判断：不足补零,大于转为svd():return:"""logger.info("judge the dimension...")field_matrix_shape = data.shapedimension = field_matrix_shape[1]if dimension > self.target_dim:return Trueelse:return Falsedef svd(self, field_matrix):"""方法：对大的领域数据进行降维:param field_matrix: list(2d) or np.array, 每一行(list)表示一条record:return: 返回领域的降维矩阵"""logger.info("use svd to reduce the dimension")indices = field_matrix.indexfm = field_matrixfield_matrix = np.array(field_matrix)field_matrix_dim = field_matrix.shapeprint(field_matrix_dim)# 对维度进行判断是否需要降维if field_matrix_dim[1] <= self.target_dim:logger.info('Filed_matrix_dim if smaller than the target, no need to perform reduction, thus we''only add extra zero element to make up the dimension.')dim_make_up = self.target_dim - field_matrix_dim[1]matrix_make_up = np.zeros([field_matrix_dim[0], dim_make_up])matrix_make_up = pd.DataFrame(matrix_make_up, index=indices)return pd.concat([fm, matrix_make_up], axis=1)else:svd = TruncatedSVD(n_components=self.target_dim)return pd.DataFrame(svd.fit_transform(field_matrix), index=indices)def run(self):"""1. Extract the one-hot-form data from the self.new_data_one_hot according to the field-instruction.2. Based on the given self.target_dimension, judge the field matrix whether satisfy the dimension requirement.3. If so, do the svd method, else add extra zero element to achieve the self.target_dimension."""output_matrix = []for i, field_data in enumerate(self.data):# field_data = self.split_field(field=item)svd_matrix = self.svd(field_matrix=field_data)svd_matrix.to_csv(self.format_data_path + 'svd_' + self.field[i] + '.csv')output_matrix.append(svd_matrix)return output_matrix

除了通过分领域和SVD降维以外，商品的属性上高达10多万类，所以我们还对商品属性计算了信息增益从而筛选了部分重要的商品属性。

def feature_selection_with_info_gain(self, data, num_feature=500, feature='item_property_list'):print(os.path.exists('../../data/format_2/infomation_gain.txt'))if os.path.exists('../../data/format_2/infomation_gain.txt'):with open('../../data/format_2/infomation_gain.txt', 'r', encoding='utf-8') as r:selected_feature = []for i in range(num_feature):line = r.readline().replace('\ufeff', '').strip().split(',')selected_feature.append(line[0])else:fea_s = list(data.columns)fea_s.remove(feature)property = []for lst in data[feature]:for pro in lst:if pro not in property:property.append(pro)info_gain = pd.Series()for pro in property:series = pd.Series([1 if pro in lst else 0 for lst in data[feature]], index=data.index, name=pro)concat_data = pd.concat([series, self.raw_data['is_trade']], axis=1)info_gain[pro] = self.cal_info_gain(data=concat_data, independent_variable=pro,dependent_variable='is_trade')info_gain = info_gain.sort_values(ascending=False)info_gain.to_csv('../../data/format_2/infomation_gain.txt', encoding='utf-8')selected_feature = list(info_gain.index[: num_feature])new_feature = []for lst in data[feature]:new_fea = []for pro in lst:if pro in selected_feature:new_fea.append(pro)new_feature.append(set(new_fea))data[feature] = new_featurereturn data

总结：这几小节主要是对特征的映射、筛选、表征，遇到的最大困难就是数据维度太高以致于服务器多次出现memory error，所以我们对原始表征数据时按照它给定的基础数据、广告商品信息、用户信息、上下文信息和店铺信息5大块分别one-hot，并通过信息增益和SVD进行降维处理，所有代码均在data_helper.py中

5.模型拟合和预测

在模型过程中，我们考虑了很多个模型，首先是在广告预测领域用得最多且效果还可以的逻辑回归、Field-aware Factorization Machines、卷积神经网络，以及非常常用的分类方法：随机森林、提升数、简单的感知器等等，最终表现效果最好的是卷积神经网络。

实验框架如下：

实验过程：

我们首先将原始训练数据和测试集，将比赛提供的loss评估指标作为我们的损失函数，通过卷积神经网络进行训练，事实上因为CNN就自带降维效果，所以，输入CNN的数据是没有用SVD进行降维的。

NN方法代码如下：


class CNN1(object):def __init__(self, n_input, n_output, x_shape, batch_size, load=0):if load == 1:saver = tf.train.import_meta_graph("../../data/model/cnn_model.ckpt.meta")self.sess = tf.Session()saver.restore(self.sess, "../../data/model/cnn_model.ckpt")else:logger.info('building the graph...')self.kb = 0.8self.batch_size = batch_sizeself.x = tf.placeholder(tf.float32, [None, n_input], name='input')self.y = tf.placeholder(tf.float32, [None, n_output], name='true_label')self.x_ = tf.reshape(self.x, shape=x_shape)# define the first convolution layerself.W_conv1 = self.weight_variable([2, 2, 1, 16])self.b_conv1 = self.bias_variable([16])self.h_conv1 = tf.nn.relu(self.conv2d(self.x_, self.W_conv1) + self.b_conv1)self.h_pool1 = self.max_pool_2x2(self.h_conv1)# define the second convolution layerself.W_conv2 = self.weight_variable([2, 2, 16, 32])self.b_conv2 = self.bias_variable([32])self.h_conv2 = tf.nn.relu(self.conv2d(self.h_pool1, self.W_conv2) + self.b_conv2)self.h_pool2 = self.max_pool_2x2(self.h_conv2)# transform the result of h_pool2 into 1D-formself.h_pool2_ = tf.reshape(self.h_pool2, [-1, (x_shape[1] // 4) * (x_shape[2] // 4) *32])h_pool2_shape = self.h_pool2_.get_shape()self.W_fc1 = self.weight_variable([h_pool2_shape[1].value, 500])self.b_fc1 = self.bias_variable([500])self.h_fc1 = tf.nn.relu(tf.matmul(self.h_pool2_, self.W_fc1) + self.b_fc1)# add a dropout layerself.keep_prob = tf.placeholder(tf.float32, name='keep')self.h_fc1_drop = tf.nn.dropout(self.h_fc1, self.keep_prob)# add a softmax layer, and get the final probabilityself.W_fc2 = self.weight_variable([500, n_output])self.b_fc2 = self.bias_variable([n_output])self.pred = tf.nn.softmax(tf.matmul(self.h_fc1_drop, self.W_fc2) + self.b_fc2, name='pred')# self.loss_func = tf.reduce_mean(- self.y * tf.log(self.pred), name='loss_func')self.loss_func = tf.reduce_mean(- self.y * tf.log(self.pred), name='loss_func') + \0.001 * tf.nn.l2_loss(self.W_conv1) + \0.001 * tf.nn.l2_loss(self.W_conv2) + \0.001 * tf.nn.l2_loss(self.W_fc1) + \0.001 * tf.nn.l2_loss(self.W_fc2)self.optm = tf.train.AdadeltaOptimizer(0.005).minimize(self.loss_func)self.init_op = tf.global_variables_initializer()self.sess = tf.Session()self.sess.run(self.init_op)@staticmethoddef weight_variable(shape):"""the method used to define the weight variables of the convolution layers:param shape:tuple or list, 该权重的形状:return:"""initial = tf.truncated_normal(shape, stddev=0.1)return tf.Variable(initial)@staticmethoddef bias_variable(shape):"""the method used to define the weight variables of the bias of each convolution layer:param shape::return:"""initial = tf.constant(0.1, shape=shape)return tf.Variable(initial)@staticmethoddef conv2d(x, W):return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')@staticmethoddef max_pool_2x2(x):return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')@staticmethoddef get_batch_data(batch_size, train_data, label_data):total = len(train_data)if batch_size > 1:chose_samples = random.sample(range(total), batch_size)else:chose_samples = random.randint(0, total)return train_data[chose_samples], label_data[chose_samples]def train(self, train_data, label_data, epoches):train_data = np.array(train_data)label_data = np.array(label_data)# for i in range(epoches):#     logger.info('running the {}-th round of training process...'.format(i))#     attr_data, labe_data = self.get_batch_data(self.batch_size, train_data, label_data)#     _, loss = self.sess.run([self.optm, self.loss_func],#                             feed_dict={self.x: attr_data, self.y:labe_data, self.keep_prob: self.kb})#     if i + 1 == epoches:#         logger.info('finish training process and the loss is {}'.format(loss))#     elif (i + 10) % 10 == 0:#         logger.info('running the {}-th epoch and the loss is {}.'.format(i, loss))with tf.device("/gpu:0"):for i in range(epoches):logger.info('running the {}-th round of training process...'.format(i))attr_data, labe_data = self.get_batch_data(self.batch_size, train_data, label_data)_, loss = self.sess.run([self.optm, self.loss_func],feed_dict={self.x: attr_data, self.y:labe_data, self.keep_prob: self.kb})if i + 1 == epoches:logger.info('finish training process and the loss is '.format(loss))elif (i + 100) % 100 == 0:logger.info('running the {}-th epoch and the loss is {}.'.format(i, loss))def predict(self, test_data, test_label, mode='test'):logger.info('predicting the result...')if mode == 'test':pred, loss = self.sess.run([self.pred, self.loss_func], feed_dict={self.x: test_data, self.y: test_label, self.keep_prob: self.kb})return pred, losselif mode == 'predict':result = self.sess.run(self.pred, feed_dict={self.x: test_data, self.keep_prob: self.kb})return resultdef load_model_predict(self, test_data, test_label, mode='test'):if mode == 'test':result = self.sess.run(['pred: 0', 'loss_func: 0'],  feed_dict={'input: 0': test_data, 'true_label: 0': test_label, 'keep: 0': 0.8})return  resultelif mode == 'predict':result = self.sess.run('pred: 0', feed_dict={'input: 0': test_data, 'keep: 0': 0.8})return resultdef save_cnn_model(self, path):saver = tf.train.Saver()saver.save(self.sess, path)def data_matrix(all_field_path, user_file, product_file, context_file, shop_file, mode=0):user_field = list(pd.read_csv(user_file, sep=',', nrows=3).columns)product_field = list(pd.read_csv(product_file, sep=',', nrows=3).columns)context_field = list(pd.read_csv(context_file, sep=',', nrows=3).columns)shop_field = list(pd.read_csv(shop_file, sep=',', nrows=3).columns)all_field_data = pd.read_csv(all_field_path, sep=',')all_field_attrs = list(all_field_data.columns)# exclude_attrs = ['user_id', 'item_id', 'context_id', 'shop_id']attrs = [user_field, product_field, context_field, shop_field]for field in attrs:for attr in field:if attr not in all_field_attrs:field.remove(attr)max_length = max([len(attr) for attr in attrs]) + 1label = 0for field in attrs:diff = max_length - len(field)if diff > 0:for i in range(label, label + diff):field.append('x' + str(i))all_field_data['x' + str(i)] = 0label += diffelse:passattrs_orders = reduce(lambda x, y: x + y, attrs, [])if mode == 0:return all_field_data[attrs_orders], max_lengthelif mode == 1:return all_field_data[attrs_orders], all_field_data.indexdef split_train_test(data, label_data, ratio):data, label_data = np.array(data), np.array(label_data)total = len(data)train_chosen_samples = random.sample(range(total), int(ratio * total))test_chosen_samples = []for ind in range(total):if ind not in train_chosen_samples:test_chosen_samples.append(ind)train_set_attrs, train_set_target = data[train_chosen_samples], label_data[train_chosen_samples]test_set_attrs, test_set_target = data[test_chosen_samples], label_data[test_chosen_samples]return train_set_attrs, train_set_target, test_set_attrs, test_set_target

除了上述的神经网络的方法，我也通过Sklean调取相关API计算了LR、贝叶斯分类器、随机森林、提升树、感知器与上诉方法进行对比


class Data_Preprocess(object):def __init__(self, train_path, test_path, raw_train_path, raw_test_path):"""Read the data including the train_data/test_data of one hot, raw_train_data/test_data, and the label ofraw_train_data.:param train_path::param test_path::param raw_train_path::param raw_test_path:"""self.raw_train_data = self.read_data(raw_train_path, data_type="raw")  # 获取is_trade# 需要把她它分为测试集和训练集self.X_data = self.read_data(train_path, data_type="one-hot").drop("instance_id", axis=1)self.Y_label = self.raw_train_data["is_trade"]self.predict_data = self.read_data(test_path, data_type="one-hot")self.predict_x = self.alignment_data()self.predict_index = self.read_data(raw_test_path, data_type="raw")["instance_id"]# 交叉验证数据集self.X_train, self.X_test, self.Y_train, self.Y_test = self.cross_data()@staticmethoddef read_data(path, data_type):"""Read data according to the path of data:param data_type::param path::return:"""if data_type == "raw":return pd.read_csv(path, sep=" ")elif data_type == "one-hot":return pd.read_csv(path, sep=",")def alignment_data(self):logger.info("数据对齐...")return self.predict_data.reindex(columns=self.X_data.columns, fill_value=0)@staticmethoddef save_model(obj, path):pickle.dump(obj, open(path, "wb"))logger.info('The model has been saved to ' + path + '...')def cross_data(self):X_train, X_test, Y_train, Y_test = train_test_split(self.X_data, self.Y_label, test_size=0.1, random_state=0)return X_train, X_test, Y_train, Y_testclass LR_Model(object):def __init__(self):self.train_x, self.train_y, self.test_x, self.test_y = X_train, Y_train, X_test, Y_testself.predict_x = predict_xself.predict_index = predict_indexdef lr_model(self):"""Method: logisticRegression:return: return the probability of test data with list format"""logger.info('LR_model beginning ...')classifier = LogisticRegression(solver="sag", class_weight="balanced")classifier.fit(self.train_x, self.train_y)index = list(classifier.classes_).index(1)test_y_predict = pd.DataFrame(classifier.predict_proba(self.test_x), columns=list(classifier.classes_))test_y_predict[index] = test_y_predict[index].apply(lambda x: 0 if x <= 0.01 else x)predict_y = list(map(lambda x: x[index], classifier.predict_proba(self.predict_x)))data_results.save_model(obj=classifier, path="../../data/results_2/lr_model.pk")return test_y_predict, predict_y@staticmethoddef evaluate(y_true, y_pred):logger.info("LR_model evaluating...")logloss = log_loss(y_true, np.array(y_pred))logger.info("The value of logloss:" + str(logloss))return loglossdef write_result(self, predict_pro, path="../../data/results_2/lr_results.txt"):logger.info('Write_result finishing ...')with open(path, "w", encoding="utf-8") as f:f.write("instance_id" + " " + "predicted_score" + "\r")for i in range(len(predict_pro)):if predict_pro[i] > 0.01:f.write(str(self.predict_index[i]) + " " + str(predict_pro[i]) + "\r")else:f.write(str(self.predict_index[i]) + " " + str(0.0) + "\r")logger.info('Write_result finished ...')def run(self):test_y_predict, predict_y = self.lr_model()self.evaluate(self.test_y, test_y_predict)self.write_result(predict_pro=predict_y)logger.info('lr_model finished ...')class Bayes_Model(object):def __init__(self):self.train_x, self.train_y, self.test_x, self.test_y = X_train, Y_train, X_test, Y_testself.predict_x = predict_xself.predict_index = predict_indexdef bayes_model(self):logger.info('Bayes_model beginning ...')classifier = BernoulliNB()classifier.fit(self.train_x, self.train_y)index = list(classifier.classes_).index(1)test_y_predict = pd.DataFrame(classifier.predict_proba(self.test_x), columns=list(classifier.classes_))test_y_predict[index] = test_y_predict[index].apply(lambda x: 0 if x <= 0.01 else x)predict_y = list(map(lambda x: x[index], classifier.predict_proba(self.predict_x)))data_results.save_model(obj=classifier, path="../../data/results_2/bayes_model.pk")return test_y_predict, predict_y@staticmethoddef evaluate(y_true, y_pred):logger.info("Bayes_model evaluating...")logloss = log_loss(y_true, np.array(y_pred))logger.info("The value of logloss:" + str(logloss))return loglossdef write_result(self, predict_pro, path="../../data/results_2/bayes_results.txt"):logger.info('Write_result finishing ...')with open(path, "w", encoding="utf-8") as f:f.write("instance_id" + " " + "predicted_score" + "\r")for i in range(len(predict_pro)):if predict_pro[i] > 0.01:f.write(str(self.predict_index[i]) + " " + str(predict_pro[i]) + "\r")else:f.write(str(self.predict_index[i]) + " " + str(0.0) + "\r")logger.info('Write_result finished ...')def run(self):test_y_predict, predict_y = self.bayes_model()self.evaluate(self.test_y, test_y_predict)self.write_result(predict_pro=predict_y)logger.info('bayes_model finished ...')class RandomTree(object):def __init__(self):self.train_x, self.train_y, self.test_x, self.test_y = X_train, Y_train, X_test, Y_testself.predict_x = predict_xself.predict_index = predict_indexdef randomtree_model(self):logger.info('RandomTree_model beginning ...')classifier = RandomForestClassifier(class_weight="balanced")classifier.fit(self.train_x, self.train_y)index = list(classifier.classes_).index(1)test_y_predict = pd.DataFrame(classifier.predict_proba(self.test_x), columns=list(classifier.classes_))test_y_predict[index] = test_y_predict[index].apply(lambda x: 0 if x <= 0.01 else x)predict_y = list(map(lambda x: x[index], classifier.predict_proba(self.predict_x)))data_results.save_model(obj=classifier, path="../../data/results_2/random_tree_model.pk")return test_y_predict, predict_y@staticmethoddef evaluate(y_true, y_pred):logger.info("Random_tree_model evaluating...")logloss = log_loss(y_true,np.array(y_pred))logger.info("The value of logloss:" + str(logloss))return loglossdef write_result(self, predict_pro, path="../../data/results_2/random_tree_results.txt"):logger.info('Write_result finishing ...')with open(path, "w", encoding="utf-8") as f:f.write("instance_id" + " " + "predicted_score" + "\r")for i in range(len(predict_pro)):if predict_pro[i] > 0.01:f.write(str(self.predict_index[i]) + " " + str(predict_pro[i]) + "\r")else:f.write(str(self.predict_index[i]) + " " + str(0.0) + "\r")logger.info('Write_result finished ...')def run(self):test_y_predict, predict_y = self.randomtree_model()self.evaluate(self.test_y, test_y_predict)self.write_result(predict_pro=predict_y)logger.info('random_tree_model finished ...')class GTB(object):def __init__(self):self.train_x, self.train_y, self.test_x, self.test_y = X_train, Y_train, X_test, Y_testself.predict_x = predict_xself.predict_index = predict_indexdef gtb_model(self):logger.info('GTB_model beginning ...')classifier = GradientBoostingClassifier()classifier.fit(self.train_x, self.train_y)index = list(classifier.classes_).index(1)test_y_predict = pd.DataFrame(classifier.predict_proba(self.test_x), columns=list(classifier.classes_))test_y_predict[index] = test_y_predict[index].apply(lambda x: 0 if x <= 0.01 else x)predict_y = list(map(lambda x: x[index], classifier.predict_proba(self.predict_x)))data_results.save_model(obj=classifier, path="../../data/results_2/gtb_model.pk")return test_y_predict, predict_y@staticmethoddef evaluate(y_true, y_pred):logger.info("GTB_model evaluating...")logloss = log_loss(y_true, np.array(y_pred))logger.info("The value of logloss:" + str(logloss))return loglossdef write_result(self, predict_pro, path="../../data/results_2/gtb_results.txt"):logger.info('Write_result finishing ...')with open(path, "w", encoding="utf-8") as f:f.write("instance_id" + " " + "predicted_score" + "\r")for i in range(len(predict_pro)):if predict_pro[i] > 0.01:f.write(str(self.predict_index[i]) + " " + str(predict_pro[i]) + "\r")else:f.write(str(self.predict_index[i]) + " " + str(0.0) + "\r")logger.info('Write_result finished ...')def run(self):test_y_predict, predict_y = self.gtb_model()self.evaluate(self.test_y, test_y_predict)self.write_result(predict_pro=predict_y)logger.info('GTB_model finished ...')class NeuralNetwork(object):def __init__(self):self.train_x, self.train_y, self.test_x, self.test_y = X_train, Y_train, X_test, Y_testself.predict_x = predict_xself.predict_index = predict_indexdef nn_model(self):logger.info('NN_model beginning ...')classifier = MLPClassifier(solver="sgd", hidden_layer_sizes=(500, 3))classifier.fit(self.train_x, self.train_y)index = list(classifier.classes_).index(1)test_y_predict = pd.DataFrame(classifier.predict_proba(self.test_x), columns=list(classifier.classes_))test_y_predict[index] = test_y_predict[index].apply(lambda x: 0 if x <= 0.01 else x)predict_y = list(map(lambda x: x[index], classifier.predict_proba(self.predict_x)))data_results.save_model(obj=classifier, path="../../data/results_2/nn_model.pk")return test_y_predict, predict_y@staticmethoddef evaluate(y_true, y_pred):logger.info("NN_model evaluating...")logloss = log_loss(y_true, np.array(y_pred))logger.info("The value of logloss:" + str(logloss))return loglossdef write_result(self, predict_pro, path="../../data/results_2/nn_results.txt"):logger.info('Write_result beginning ...')with open(path, "w", encoding="utf-8") as f:f.write("instance_id" + " " + "predicted_score" + "\r")for i in range(len(predict_pro)):if predict_pro[i] > 0.01:f.write(str(self.predict_index[i]) + " " + str(predict_pro[i]) + "\r")else:f.write(str(self.predict_index[i]) + " " + str(0.0) + "\r")logger.info('Write_result finished ...')def run(self):test_y_predict, predict_y = self.nn_model()self.evaluate(self.test_y, test_y_predict)self.write_result(predict_pro=predict_y)logger.info('NN_model finished ...')

对于不同的模型，我们使用自己分的测试集对其进行预测，并计算相关的损失函数，损失函数的结果如下：

【注意：结果值越小越好】

Mehod	LR	Bayes	Random_T	GTB	NN	CNN
Loss	4.10158	1.015423	0.539459	0.09011	0.089561	0.046641

从上表可以看到逻辑回归、贝叶斯分类器、随机森林、简单感知器、CNN的结果，而我们提出的CNN算法在测试集上的效果为0.046641明显优于其他方法，然后我们却没能够得到最终的验证，哭死哭死。

总结

虽然这次比赛比较遗憾和难过，但是不得不说，真的学习了很多很多。

其实，对于很多事，除了【努力】、【机遇】、【幸运】，还是要注意细节，一直都喜欢对自己说，把每件小事做好了，那么结果一定不会太差，那么，未来，请继续努力吧！

【注意更多更完整的代码详见github】

用户名：DWJWendy
链接：https://github.com/DWJWendy/IJCAI_2018_CTR.git