0 概论

NFL以Game Pass产品作为主要营销方法（Pat Even，2019），在欧洲的观众人数增加。但是，上一季的Game Pass广告策略仍存在问题。在市场方面，一些国家获得了更多投资，但投资回报率非常有限，而其他国家则相反。为了解决这个问题，我选择英国市场进行分析并给出相应的解决方案，以达到扩大市场和提高投资回报率的效果。

0.1 构成

为了解决这个问题，本篇文章分为6个部分。

第1部分是问题的提出以及GamePass的业务概述。

第2部分是MySQL的数据导入与连接工作。

第3部分是Notebook上数据清洗工作。

第4部分是在Tableau上对GamePass的广告投放表现做分析，以及指出英国市场的问题。

第5部分是针对英国市场的广告投放所影响的用户消费具体数据，进行特征工程处理，并通过处理过的特征，进行简单的机器学习预测。

第6部分则是对用户进行细分，并针对不同群体的购买金额、购买次数对比，提出相应的简单营销策略。

0.2 注明

（本次项目数据来自NFL和TwoCircles。由于NFL规定，本次项目的数据不公开，望谅解）

本次项目在maxaishaojun的Github上可以浏览（注意Github移动端浏览可能效果不佳）。

"4 问题分析"部分可移步到Tableau Dashboard上浏览 NFL GamePass 2018-19赛季广告投放问题分析。

0.3 引用

Pat, E. (2019) ‘NFL Viewership Growth Throughout Europe Exposes Opportunities in the US’, Front Office Sport, 7 March. Available at: frntofficesport.com/two-circles… (Accessed: 6 April 2019)

1 项目背景

（注：本部分文档来源 - Two Circles）

1.1 问题提出

假设现在是2018/19 NFL赛季的结束，你是NFL Game Pass Europe营销团队的成员。 你的任务是分析广告投放在获取新的NFL Game Pass订阅者方面的有效性，通过部署的广告渠道挖掘购买者的客户档案，并为下一季的数字广告活动提供建议，涵盖广告素材，受众群体，地理位置定位 ，渠道组合和预算。

1.2 项目概述

NFL Game Pass是国外NFL球迷的首选OTT订阅产品，也是NFL国际业务的重中之重。与“NFL的Netflix”相似，粉丝可以访问每个实时游戏，游戏回放和精彩集锦，NFL RedZone，NFL网络24/7直播电视频道，原创内容，节目和游戏档案，下载等等。

近年来，NFL在欧洲的人气越来越高，这已反映在非洲大陆的NFL Game Pass的用户数量上。保持这种用户增长水平需要采取策略来最小化用户流失，优化用户获取并增加尚未准备好转换的粉丝群体的考虑。

NFL Game Pass Europe的订户收购营销活动中最大的组成部分是数字广告，每季占收购订户的约46％。该活动通常在NFL季前赛（8月）开始，并在NFL赛季（9月）的前4周达到峰值强度，最终在超级碗（2月初）之后结束。该活动中部署了各种付费营销渠道，主要是Google广告（搜索/ PPC），Facebook和YouTube（社交）以及各种程序化展示广告合作伙伴（展示广告）。此外，展示广告在www.nfl.com和NFL免费拥有的其他数字资产（O＆O）上发布。预算，广告创意和受众群体定位策略因渠道而异，在整个广告系列中获得平衡对其成功至关重要。

NFL Game Pass Europe刚刚完成了2018/19赛季的数字广告活动，并正在进入2019/20赛季的规划过程。通过访问丰富的数据集，可以分析广告效果以及购买者人口统计数据，NFL Game Pass Europe正在寻求挖掘数据集，以获取有助于未来数字广告策略的见解。

1.3 附录 2018/19 NFL Game Pass Europe 产品定义

2018/19 Audience Strategy and Definitions

2018/19 Market Strategy and Definitions 2018/19 Example Ad Units 2018/19 Marketing Plan and Promotions

2 数据导入

-- 1 在MediaPerformanceData中增加numweek与revenueselect a.date, a.nflweek, a.numweek, a.platform, a.market, a.audience, round(a.`Spend (GBP)`*1.3, 2) as spend_usd, a.impressions, a.clicks, a.transactions, b.revenue_usd
from (select *,@number := casewhen @weeks = nflweek then @numberelse @number+1end as numweek,@weeks := nflweek as weeksfrom mediaperformancedatajoin (select @weeks := null, @number := 0) as variableorder by date) a
join (select mt.date, mt.nflweek, mt.platform, mt.market, mt.audience, round(sum(`Revenue (USD)`),2) as revenue_usdfrom MediaTransactionsdata mtjoin Subscriptionsdata s on mt.transactionID = s.transactionidgroup bymt.date, mt.nflweek, mt.platform, mt.market, mt.audienceorder by mt.date, mt.platform, mt.market, mt.audience) b
on (a.date = b.date and a.nflweek = b.nflweek and a.platform = b.platform and a.market = b.market and a.audience = b.audience)
;-- 2 连接剩下三表select mt.transactionid, s.customerid, mt.date, mt.nflweek, mt.platform, mt.market, mt.audience,c.`NFL Game Pass Segment`, c.gender, c.age, c.`NFL Tickets`, c.`NFL Shop`, c.`NFL Fantasy`, c.`New To NFL Database`, c.`Email Opt-In`, c.`Favourite Team`, s.SKU, s.`Buy Type`, s.`Converted Free Trial`, s.`Revenue (USD)`
fromSubscriptionsdata s
left joinmediatransactionsdata mt on s.transactionid = mt.transactionid
left join customersdata c on s.customerid = c.customerid
order bymt.date
;
复制代码

3 数据清洗

import pandas as pd
import numpy as npmp_file_path = '/Users/apple/Downloads/nfl/mediaperformance.csv'
nfl_file_path = '/Users/apple/Downloads/nfl/nfladverts.csv'mp_data = pd.read_csv(mp_file_path)
nfl_data = pd.read_csv(nfl_file_path)# 重新命名列nfl_data.columns = ['transaction_id', 'customer_id', 'date', 'nflweek', 'platform', 'market','audience', 'segment', 'gender', 'age_group', 'tickets','shop', 'fantasy', 'new_to_database', 'email_opt_in','favourite_team', 'sku', 'buy_type', 'converted_free_trial','revenue_usd']nfl_data['revenue_usd'].describe()
# mp_data.columns# 重复值mp_unique = mp_data.groupby(['date', 'nflweek', 'numweek', 'platform', 'market', 'audience']).size().reset_index(name='Freq')
mp_unique = mp_unique.sort_values(by=['Freq'], ascending=False)
print(mp_unique)nfl_data.head()
nfl_unique = nfl_data.groupby(['transaction_id']).size().reset_index(name='Freq')
nfl_unique = nfl_unique.sort_values(by=['Freq'], ascending=False)
print(nfl_unique)
# data.drop_duplicates(subset ="columns Name", keep = False, inplace = True) # 缺失值mp_null_total = mp_data.isnull().sum(axis=0).sort_values(ascending=False)
mp_null_percent = (mp_data.isnull().sum()/len(mp_data.index)).sort_values(ascending=False).round(3)
mp_missing_data = pd.concat([mp_null_total, mp_null_percent], axis=1, keys=['Total', 'Percent'])nfl_null_total = nfl_data.isnull().sum(axis=0).sort_values(ascending=False)
nfl_null_percent = (nfl_data.isnull().sum()/len(nfl_data.index)).sort_values(ascending=False).round(3)
nfl_missing_data = pd.concat([nfl_null_total, nfl_null_percent], axis=1, keys=['Total', 'Percent'])# 去掉没有transaction_id的行
nfl_data = nfl_data.dropna(subset=['transaction_id'])# 没有填喜好球队的编为‘No Team’
nfl_data['favourite_team'] = nfl_data['favourite_team'].fillna('No Team')# 没填性别编为'U'
nfl_data['gender'] = nfl_data['gender'].fillna('U')# 由于gender缺失太多，和其他列也没有存在明显的逻辑关系，尝试看有没有男女id差异，结果没有
# https://pandas.pydata.org/pandas-docs/stable/reference/api/pandas.pivot_table.html
gender_predict = nfl_data[nfl_data.gender.notnull()]
gender_pivot = pd.pivot_table(gender_predict, values='customer_id', index=['gender'], aggfunc=np.mean)# 按照逻辑填充buy_type
# nfl_data['buy_type'].unique()
nfl_data['buy_type'] = nfl_data.apply(lambda row: 'Buy Now' if row['sku'] =='Free' else ('Buy Now' if (row['converted_free_trial'] == 0 and row['revenue_usd'] > 0) else ('Free Trial' if (row['converted_free_trial'] == 0 and row['revenue_usd'] == 0) or (row['converted_free_trial'] == 1 and row['revenue_usd'] > 0) else row['buy_type'])), axis=1)# nfl_data.isnull().sum(axis=0).sort_values(ascending=False)# 异常值# binary都变成0&1，方便计算；并去掉异常值
# nfl_data['market'].unique()
nfl_data['tickets'] = np.where(nfl_data['tickets'] == 'N', 0, 1)
nfl_data['shop'] = np.where(nfl_data['shop'] == 'N', 0, 1)
nfl_data['fantasy'] = np.where(nfl_data['fantasy'] == False, 0, 1)
nfl_data['new_to_database'] = np.where(nfl_data['new_to_database'] == 'N', 0, 1)
nfl_data['email_opt_in'] = np.where(nfl_data['email_opt_in'] == 'False', 0, np.where(nfl_data['email_opt_in'] == '0', 0, 1))
nfl_data['buy_type'] = np.where(nfl_data['buy_type'] == 'Free Trial', 0, 1)
nfl_data['favourite_team'] = np.where(nfl_data['favourite_team'] == 'NFL', 'No Team', nfl_data['favourite_team'])
nfl_data['favourite_team'] = np.where(nfl_data['favourite_team'] == 'No Team', 0, 1)# 纠正sku、buy_type、converted_free_trial的逻辑错误
# https://stackoverflow.com/questions/36921951/truth-value-of-a-series-is-ambiguous-use-a-empty-a-bool-a-item-a-any-o
# https://discuss.analyticsvidhya.com/t/how-to-resolve-python-error-cannot-compare-a-dtyped-int64-array-with-a-scalar-of-type-bool/73065
nfl_data['sku'] = np.where((nfl_data['revenue_usd'] == 0) & (nfl_data['sku'] != 'Pro'), 'Free', nfl_data['sku'])
nfl_data['sku'] = np.where((nfl_data['revenue_usd'] > 0) & (nfl_data['sku'] == 'Free'), 'Pro', nfl_data['sku'])
nfl_data['buy_type'] = np.where((nfl_data['sku'] == 'Pro') & (nfl_data['revenue_usd'] == 0), 0, nfl_data['buy_type'])
nfl_data['converted_free_trial'] = np.where((nfl_data['sku'] == 'Pro') & (nfl_data['buy_type'] == 0) & (nfl_data['revenue_usd'] > 0), 1, nfl_data['converted_free_trial'])
nfl_data['converted_free_trial'] = np.where((nfl_data['sku'] == 'Pro') & (nfl_data['buy_type'] == 1) & (nfl_data['revenue_usd'] > 0), 0, nfl_data['converted_free_trial'])mp_data.to_excel(r'/Users/apple/Downloads/nfl/mp_data.xlsx', index=True)
nfl_data.to_csv(r'/Users/apple/Downloads/nfl/nfl_data.csv', index=True)
复制代码

4 问题分析

问题分析部分可移步到Tableau Dashboard上浏览 NFL GamePass 2018-19赛季广告投放问题分析。

5 特征工程

import pandas as pd
from pandas import DataFrame
import numpy as np
import matplotlib.pyplot as plt
import seaborn as snsuser_file_path = '/Users/apple/Downloads/nfl/nfl_data.csv'
user_data = pd.read_csv(user_file_path)# extract UK market
uk_data = user_data.loc[user_data['market'] == 'UK']
# uk_data['revenue_usd'].describe()# user_data.customer_id = user_data.customer_id.astype(str)
uk_data = uk_data.drop(['Unnamed: 0', 'date', 'customer_id', 'audience', 'market', 'nflweek', 'platform'], axis=1)
# uk_data.head()# 看消费的分布情况
uk_data.revenue_usd.hist(bins=20, alpha=0.5)
plt.title("Game Pass Europe Revenue Distribution")
plt.xlabel("Revenue($)")
plt.ylabel("Frequency")
复制代码

5.1 数据预处理

# 量纲的问题# 这里的data都是 yes/no 或者是category的数据，除了revenue_usd之外，没有continuous data
# 规范化处理(norm)对数据异常值很敏感，处理之后数据中的异常值会消失，因此如果数据集中存在异常值，则这是一种不好的做法。
# 而标准化不受数据限制，所以一般我们采用标准化来处理数据。def normalize(data, column):for col in column:data['normalize_'+col] = (data[col] - np.min(data[col])) / (np.max(data[col]) - np.min(data[col]))return datadef standardize(data, column):for col in column:data['standardize_'+col] = (data[col] - np.mean(data[col])) / (np.std(data[col]))return datacolumns = ['revenue_usd']
uk_data = standardize(uk_data, columns)# 再介绍几种激活函数，其本意都是把数值压缩在某个区间，其中有的区间敏感，有的不敏感def tanh(data, column):  for col in column:data['tanh_'+col] = np.tanh(data[col])return data  def sigmoid(data, column):for col in column:data['sigmoid_'+col] = 1.0 / (1.0 + (np.exp(data[col])*(-1)))return datadef leakyrelu(data, column, a=1): for col in column:data['leakyrelu_'+col] = np.array([x if x > 0 else a * x for x in data[col]])return datadef softplus(data, column):for col in column:data['softplus_'+col] = np.log(np.exp(data[col]) + 1)return datauk_data.standardize_revenue_usd.hist(bins=20, alpha=0.5)
plt.title("Game Pass Europe Revenue Distribution")
plt.xlabel("Revenue($)")
plt.ylabel("Frequency")# 特征值类型的问题# category编码 - dummy coding
# 这种虚拟变量的做法容易增加数据特征的维度dummy_data = pd.get_dummies(uk_data,columns=['segment', 'gender', 'age_group', 'sku'],prefix=['segment', 'gender', 'age_group', 'sku'],prefix_sep="_")# uk_data = uk_data.drop(['gender_U'], axis=1)# 连续值转为categoryconditions = [(uk_data['revenue_usd'] == 0),(uk_data['revenue_usd'] <= 13),(uk_data['revenue_usd'] <= 98)]
choices = [0,1,2]
dummy_data['revenue_category'] = np.select(conditions, choices, default= 3)
uk_data['revenue_category'] = np.select(conditions, choices, default= 3)fig, ax = plt.subplots()
dummy_data['revenue_category'].value_counts().plot(ax=ax, kind='bar')复制代码

5.2 特征选择

# dummy_data.columnsselect_feature = ['tickets', 'shop', 'fantasy', 'new_to_database', 'email_opt_in','favourite_team', 'buy_type', 'converted_free_trial', 'segment_Acq','segment_Ret', 'segment_iOS', 'gender_F', 'gender_M', 'gender_U','age_group_18-21', 'age_group_22-25', 'age_group_26-30','age_group_31-35', 'age_group_36-40', 'age_group_41-50','age_group_51-60', 'age_group_60+', 'age_group_Under 18','age_group_Unknown', 'sku_Essential', 'sku_Free', 'sku_Playoffs','sku_Pro', 'sku_Super Bowl', 'sku_Weekly']# 方差选择法from sklearn.feature_selection import VarianceThresholdvarianceThreshold = VarianceThreshold(threshold = 0.2)
varianceThreshold.fit_transform(dummy_data[select_feature])
var_result = varianceThreshold.get_support()# 相关系数法
# 选择基本的feature，然后匹配其他
# 注意逻辑关系from sklearn.feature_selection import SelectKBest
from sklearn.feature_selection import f_regression
selectKBest = SelectKBest(f_regression, k=10)
feature = dummy_data[select_feature]
bestFeature = selectKBest.fit_transform(feature, dummy_data[['revenue_usd']])
feature_result = selectKBest.get_support()def feature_results(list_feature, list_result):dic = {}for i in range(len(list_feature)):feature = list_feature[i]result = list_result[i]dic[feature] = resultresult_tuple = sorted(dic.items(), key=lambda kv: kv[1])return result_tuplevar_results = feature_results(select_feature, var_result)
weights_results = feature_results(select_feature, feature_result)print(var_results[-5:-1])
print(weights_results[-10:-1])复制代码

5.3 维度下降

# 逻辑降维
# 根据发散性和关联性的结果，对愿特征进行处理
# 合并tickets、shop、fantasy、new_to_database、email_opt_in，以提高相关性
# 如果可以的话应该加上权重
# uk_data.sku.unique()uk_data['user_behaviour'] = uk_data['tickets'] + uk_data['shop'] + uk_data['fantasy'] + uk_data['new_to_database'] + uk_data['email_opt_in']    # age_group 分成老中青 0-21 22-40 40+
uk_data['age'] = np.where((uk_data['age_group']=='Under 18') | (uk_data['age_group']=='18-21'), 'young', np.where((uk_data['age_group']=='22-25') | (uk_data['age_group']=='26-30') | (uk_data['age_group']=='31-35') | (uk_data['age_group']=='36-40'), 'adult','old'))# sku, buy_ype & converted_free_trial 是属于购买行为了，应该归在target_variable中
uk_data['sku_category'] = np.where((uk_data['sku']=='Pro') & (uk_data['buy_type']== 1), 'Pro-BuyNow', np.where((uk_data['sku']=='Pro') & (uk_data['buy_type']== 0) & (uk_data['converted_free_trial']== 0), 'Pro-FreeTrial-NoConvert',np.where((uk_data['sku']=='Pro') & (uk_data['buy_type']== 0) & (uk_data['converted_free_trial']== 1), 'Pro-FreeTrial-Convert',uk_data['sku'])))
uk_data = uk_data.drop(['tickets', 'shop', 'fantasy', 'new_to_database', 'email_opt_in', 'sku', 'buy_type', 'converted_free_trial', 'age_group', 'standardize_revenue_usd'], axis=1)uk_data = uk_data[['transaction_id', 'age','gender', 'favourite_team', 'user_behaviour', 'segment', 'revenue_usd', 'sku_category', 'revenue_category']]
uk_data.head(10)# 哑编码，以适应某些模型不接受categorical data
dummy_data = pd.get_dummies(uk_data,columns=['age', 'gender', 'segment'],prefix=['age', 'gender', 'segment'],prefix_sep="_")# 将哑编码与原数据结合
feature_data = pd.merge(dummy_data, uk_data[['transaction_id', 'age', 'gender', 'segment']], on='transaction_id', how='inner')
feature_data.head()# 透视表"""
feature_data['transaction_count'] = 1
pd.pivot_table(feature_data,columns=["age"],index = ['favourite_team'],values=['revenue_usd', 'transaction_count'],aggfunc=[np.mean,np.sum])
"""# 相关性可视化"""
variables = ['favourite_team', 'user_behaviour', 'revenue_usd', 'age_adult', 'age_old', 'age_young','gender_F', 'gender_M', 'segment_Acq', 'segment_Ret', 'segment_iOS']sns.set()
sns.pairplot(feature_data[variables], size = 2.5)
plt.show()
"""# 多重共线性检验"""
x = feature_data[['favourite_team', 'user_behaviour', 'age_adult', 'age_old', 'age_young','gender_F', 'gender_M', 'segment_Acq', 'segment_Ret', 'segment_iOS']]from statsmodels.stats.outliers_influence import variance_inflation_factorvif = pd.DataFrame()
vif["VIF Factor"] = [variance_inflation_factor(x.values, i) for i in range(x.shape[1])]
vif["features"] = x.columns
vif.round(1)feature_data.to_csv(r'/Users/apple/Downloads/nfl/feature_data.csv', index=True)
"""# 算法试验
# 检验方法-train-validation
from sklearn.model_selection import train_test_splitx = feature_data[['favourite_team', 'user_behaviour', 'age_adult', 'age_old', 'age_young','gender_F', 'gender_M', 'segment_Acq', 'segment_Ret', 'segment_iOS']]
y1 = feature_data[['revenue_usd']] #regression
y2 = feature_data[['revenue_category']] #classificationx_train1, x_val1, y_train1, y_val1 = train_test_split(x, y1, test_size=0.2, random_state=1)
x_train2, x_val2, y_train2, y_val2 = train_test_split(x, y2, test_size=0.2, random_state=1)print("the number of data for training:")
print(y_train1.count())
print("the number of data for validation:")
print(y_val1.count())#衡量方法-accuracy与RMSEfrom sklearn.metrics import mean_squared_errordef rmse_model(model, x, y):predictions = model.predict(x)rmse = np.sqrt(mean_squared_error(predictions, y))return rmse"""
from sklearn import metricsdef confusion_matrix(model, x, y):model_confusion_test = metrics.confusion_matrix(y, model.predict(x))matrix = pd.DataFrame(data = model_confusion_test, columns = ['Predicted 0', 'Predicted 1', 'Predicted 2', 'Predicted 3'],index = ['Predicted 0', 'Predicted 1', 'Predicted 2', 'Predicted 3'])return matrix
"""# regression
# 由于运算量和时间问题，就只放上代码了"""
from sklearn.linear_model import LinearRegressionlinear_regression = LinearRegression()
linear_regression.fit(x_train1, y_train1)print(rmse_model(linear_regression, x_val1, y_val1))
"""# bias-variance trade-off"""
from sklearn.preprocessing import PolynomialFeaturestrain_rmses = []
val_rmses = []
degrees = range(1,8)for i in degrees:poly = PolynomialFeatures(degree=i, include_bias=False)x_train_poly = poly.fit_transform(x_train1)poly_reg = LinearRegression()poly_reg.fit(x_train_poly, y_train1)# training RMSEy_train_pred = poly_reg.predict(x_train_poly)train_poly_rmse = np.sqrt(mean_squared_error(y_train1, y_train_pred))train_rmses.append(train_poly_rmse)# validation RMSEx_val_poly = poly.fit_transform(x_val1)y_val_pred = poly_reg.predict(x_val_poly)val_poly_rmse = np.sqrt(mean_squared_error(y_val1, y_val_pred))val_rmses.append(val_poly_rmse)print('degree = %s, training RMSE = %.2f, validation RMSE = %.2f' % (i, train_poly_rmse, val_poly_rmse))fig = plt.figure()
ax = fig.add_subplot(111)
ax.plot(degrees, train_rmses,label= 'training set')
ax.plot(degrees, val_rmses,label= 'validation set')
ax.set_yscale('log')
ax.set_xlabel('Degree')
ax.set_ylabel('RMSE')
ax.set_title('Bias/Variance Trade-off')
plt.legend()
plt.show()
"""# regularization in order to reduce the effect of overfitting
# ridge (lasso, elasticnet代码相似，lasso和elasticnet对RMSE压缩得更狠一些，不过不会去掉有collinearity的feature，ridge相反)"""
from sklearn.linear_model import Ridge
from sklearn.pipeline import make_pipelinermse=[]
alpha=[1, 2, 5, 10, 20, 30, 40, 50, 75, 100]for a in alpha:ridge = make_pipeline(PolynomialFeatures(4), Ridge(alpha=a))ridge.fit(x_train1, y_train1)predict=ridge.predict(x_val1)rmse.append(np.sqrt(mean_squared_error(predict, y_val1)))
print(rmse)
plt.scatter(alpha, rmse)alpha=np.arange(20, 60, 2)
rmse=[]
for a in alpha:#ridge=Ridge(alpha=a, copy_X=True, fit_intercept=True)#ridge.fit(x_train1, y_train1)ridge = make_pipeline(PolynomialFeatures(4), Ridge(alpha=a))ridge.fit(x_train1, y_train1)predict=ridge.predict(x_val1)rmse.append(np.sqrt(mean_squared_error(predict, y_val1)))
print(rmse)
plt.scatter(alpha, rmse)ridge = make_pipeline(PolynomialFeatures(4), Ridge(alpha=24.6))
ridge_model = ridge.fit(x_train1, y_train1)
predictions = ridge_model.predict(x_val1)
print("Ridge RMSE is: " + str(rmse_model(ridge_model, x_val1, y_val1)))
"""# classficationlist(y_train2['revenue_category'].unique())# decision treefrom sklearn import tree
from sklearn.tree import DecisionTreeClassifierdecision_tree_model = DecisionTreeClassifier(criterion='entropy')
decision_tree_model.fit(x_train2, y_train2)
print(decision_tree_model.score(x_train2,y_train2))
print(decision_tree_model.score(x_val2,y_val2))# tuning
train_score = []
val_score = []
for depth in np.arange(1,20):decision_tree = tree.DecisionTreeClassifier(max_depth = depth,min_samples_leaf = 5)decision_tree.fit(x_train2, y_train2)train_score.append(decision_tree.score(x_train2, y_train2))val_score.append(decision_tree.score(x_val2, y_val2))plt.plot(np.arange(1,20),train_score)
plt.plot(np.arange(1,20),val_score)
plt.legend(['Training Accuracy','Validation Accuracy'])
plt.title('Decision Tree Tuning')
plt.xlabel('Depth')
plt.ylabel('Accuracy')train_score = []
val_score = []
for leaf in np.arange(20,100):decision_tree = tree.DecisionTreeClassifier(max_depth = 10, min_samples_leaf = leaf)decision_tree.fit(x_train2, y_train2)train_score.append(decision_tree.score(x_train2, y_train2))val_score.append(decision_tree.score(x_val2, y_val2))plt.plot(np.arange(20,100),train_score)
plt.plot(np.arange(20,100),val_score)
plt.legend(['Training Accuracy','Validation Accuracy'])
plt.title('Decision Tree Tuning')
plt.xlabel('Minimum Samples Leaf')
plt.ylabel('Accuracy')my_decision_tree_model = DecisionTreeClassifier(max_depth = 10, min_samples_leaf = 20)
my_decision_tree_model.fit(x_train2, y_train2)
print(my_decision_tree_model.score(x_train2,y_train2))
print(my_decision_tree_model.score(x_val2,y_val2))# confusion matrix
from sklearn.metrics import accuracy_score, confusion_matrix, precision_recall_fscore_supporty_predict = my_decision_tree_model.predict(x_val2)
cm = confusion_matrix(y_val2, y_predict) # Transform to df for easier plotting
cm_df = pd.DataFrame(cm,index = ['free', 'median', 'high', 'low'], columns = ['free', 'median', 'high', 'low'])plt.figure(figsize=(5.5,4))
sns.heatmap(cm_df, annot=True)
plt.title('Decision Tree \nAccuracy:{0:.3f}'.format(accuracy_score(y_val2, y_predict)))
plt.ylabel('True label')
plt.xlabel('Predicted label')
plt.show()# learning curvefrom sklearn.model_selection import learning_curvetrain_sizes, train_scores, val_scores = learning_curve(DecisionTreeClassifier(max_depth = 10, min_samples_leaf = 20), x, y2,# Number of folds in cross-validationcv=5,# Evaluation metricscoring='accuracy',# Use all computer coresn_jobs=-1, # 50 different sizes of the training settrain_sizes=np.linspace(0.1, 1.0, 5))# Create means and standard deviations of training set scores
train_mean = np.mean(train_scores, axis=1)
train_std = np.std(train_scores, axis=1)# Create means and standard deviations of validation set scores
val_mean = np.mean(val_scores, axis=1)
val_std = np.std(val_scores, axis=1)# Draw lines
plt.plot(train_sizes, train_mean, '--', color="#ff8040",  label="Training score")
plt.plot(train_sizes, val_mean, color="#40bfff", label="Cross-validation score")# Draw bands
plt.fill_between(train_sizes, train_mean - train_std, train_mean + train_std, color="#DDDDDD")
plt.fill_between(train_sizes, val_mean - val_std, val_mean + val_std, color="#DDDDDD")# Create plot
plt.title("Learning Curve \n k-fold=5, number of neighbours=5")
plt.xlabel("Training Set Size"), plt.ylabel("Accuracy Score"), plt.legend(loc="best")
plt.tight_layout()
plt.show()# Curse of Dimensionalityd_train = []
d_val = []for i in range(1,9):X_train_index = x_train2.iloc[: , 0:i]X_val_index = x_val2.iloc[: , 0:i]classifier = DecisionTreeClassifier(max_depth = 10, min_samples_leaf = 20)dt_model = classifier.fit(X_train_index, y_train2.values.ravel())d_train.append(dt_model.score(X_train_index, y_train2))d_val.append(dt_model.score(X_val_index, y_val2))plt.title('Decision Tree Curse of Dimensionality')
plt.plot(range(1,9),d_val,label="Validation")
plt.plot(range(1,9),d_train,label="Train")
plt.xlabel('Number of Features')
plt.ylabel('Score (Accuracy)')
plt.legend()
plt.xticks(range(1,9))
plt.show()# 预测结果很差，可能和feature选择有关系，也有可能和算法不适合有关系。
复制代码

6 用户细分及建议

6.1 用户细分

通过特征工程后，我们这一步的工作就是将我们的用户细分，并根据不同的用户群体提出不同的策略。在通过Excel PivotTable & 可视化操作后，以喜欢球队、NFL消费行为、用户segment、年龄、性别为比较维度，购买金额（平均）、购买次数为目标维度，构成矩阵。

6.2 Marketing建议

基于英国在Game Pass欧洲市场战略的地位，其目的应该是让越来越多的人参与到NFL中的同时，提高ROI。因此，在不考虑Budget的情况下，应该想办法改进四个群体的消费额与用户体验。

¶ 对于明星用户（中老年男性，有喜欢的球队），应该保持和提高用户的产品体验，并组织球迷活动，增进关系，提高消费意愿。

¶ 对于便宜套餐用户（男性新用户，没有喜欢的球队），其对GamePass还处在观望的状态，或者归属感不强影响消费意愿（喜欢看比赛，但是更多的是寻找免费套餐）。这个时候可以在线上线下举行增进球迷关系的活动，或者在赛季期间搞价格促销活动，提高购买意愿。另一方面，其购买能力也有可能影响消费金额，不过可以通过广告从而为GamePass间接获取流量收入。

¶ 对于“只花一次大价钱”用户，首先明确他们的消费能力应该是没问题的。应该举行调研，为什么愿意花大价钱，但是为什么只消费了一次？如何提高这部分用户对产品的体验从而促使消费次数增加？

¶ 而对于瘦狗用户（中老年女性or青年，没有喜欢球队），首先要做的是NFL对于这部分群体的普及营销工作（如果不喜欢NFL，怎么会用GamePass呢），包括推出关于女性、青少年的NFL宣传广告（提高印象）、举行多次推广活动（提高认知，并鼓励用户参与）、以及公关活动，以提升这部分群体对NFL的功能认知和道德认知。