聚类分析|基于RFM模型对用户价值进行分析

#导入基础包import pandas as pdimport numpy as npimport seaborn as snsimport matplotlib.pyplot as pltplt.rcParams["font.sans-serif"]=["SimHei"] #用来正常显示中文标签plt.rcParams["axes.unicode_minus"]=False #用来正常显示负号%matplotlib inline#读取数据df=pd.read_csv("C:/Users/dwhyx/Downloads/data/userlostprob_train.txt",sep="	")df.describe()

# 转为日期型格式df["arrival"]=pd.to_datetime(df["arrival"])df["d"]=pd.to_datetime(df["d"])# 相减得到"提前预定天数"列df["day_advanced"]=(df["arrival"]-df["d"]).dt.days# 删除原有列df=df.drop(["d","arrival"],axis=1)

for col in ["delta_price1","delta_price2","lowestprice"]:    df.loc[df[col]<0,col]=df[col].median()  # 填充中位数for col in ["customer_value_profit","ctrip_profits"]:    df.loc[df[col]<0,col]=0  # 填充0#极值处理for i in df.columns:    df.loc[df[i]<np.percentile(df[i],1),i]=np.percentile(df[i],1)    df.loc[df[i]>np.percentile(df[i],99),i]=np.percentile(df[i],99)

print("删除空值前数据维度是:{}".format(df.shape))df.dropna(axis=0,thresh=df.shape[1]*0.2,inplace=True)  df.dropna(axis=1,thresh=df.shape[0]*0.2,inplace=True)print("删除空值后数据维度是:{}".format(df.shape))

df.hist(figsize=(20,20))plt.savefig("hist.png")

filter_mean=["businessrate_pre","businessrate_pre2","cancelrate_pre","customereval_pre2 "]for i in df.columns:    if i in filter_mean:        df[i].fillna(df[i].mean(),inplace=True)    else:        df[i].fillna(df[i].median(),inplace=True)

for i in df.columns:    df.loc[df[i]<np.percentile(df[i],1),i]=np.percentile(df[i],1)   #小于1%分位数的用1%分位数填充    df.loc[df[i]>np.percentile(df[i],99),i]=np.percentile(df[i],99)  # 大于99%分位数的用99%分位数填充

consume_level=["avgprice","consuming_capacity"]    # 合并字段作为消费水平from sklearn.decomposition import PCA   #利用PCA主成分分析无监督的降维方法pca=PCA(n_components=1)  #n_components设置为1df["consume_level"]=pca.fit_transform(df[consume_level])  #返回降维后的数据df.drop(consume_level,axis=1,inplace=True)  #删除冗余列

#字段重名rfm = df[["lasthtlordergap","ordernum_oneyear","consume_level"]]  rfm.rename(columns={"lasthtlordergap":"recency","ordernum_oneyear":"frequency","consume_level":"monetary"},inplace=True)#利用MinMaxScaler进行归一化处理from sklearn.preprocessing import MinMaxScalerscaler = MinMaxScaler()scaler.fit(rfm)rfm = pd.DataFrame(scaler.transform(rfm),columns=["recency","frequency","monetary"])#分箱rfm["R"]=pd.qcut(rfm["recency"], 2)rfm["F"]=pd.qcut(rfm["frequency"], 2)rfm["M"]=pd.qcut(rfm["monetary"], 2)# 根据分箱情况进行编码，二分类可以直接用标签编码方式from sklearn.preprocessing import LabelEncoderrfm["R"]=LabelEncoder().fit(rfm["R"]).transform(rfm["R"])  rfm["F"]=LabelEncoder().fit(rfm["F"]).transform(rfm["F"])rfm["M"]=LabelEncoder().fit(rfm["M"]).transform(rfm["M"])#定义RFM模型，需要特别注意的是，R值代表距离上次消费时间间隔，值越小客户价值越高，与F和M值正好相反。def get_label(r,f,m):    if (r==0)&(f==1)&(m==1):        return "高价值客户"    if (r==1)&(f==1)&(m==1):        return "重点保持客户"    if((r==0)&(f==0)&(m==1)):        return "重点发展客户"    if (r==1)&(f==0)&(m==1):        return "重点挽留客户"    if (r==0)&(f==1)&(m==0):        return "一般价值客户"    if (r==1)&(f==1)&(m==0):        return "一般保持客户"    if (r==0)&(f==0)&(m==0):        return "一般发展客户"    if (r==1)&(f==0)&(m==0):        return "潜在客户"def RFM_convert(df):    df["Label of Customer"]=df.apply(lambda x:get_label(x["R"],x["F"],x["M"]),axis=1)        df["R"]=np.where(df["R"]==0,"高","低")    df["F"]=np.where(df["F"]==1,"高","低")    df["M"]=np.where(df["M"]==1,"高","低")        return df[["R","F","M","Label of Customer"]]rfm1=RFM_convert(rfm)rfm1.head(5)

df = df.join(rfm1)df.to_csv("label.csv")

value_counts=rfm1["Label of Customer"].value_counts().valueslabels=rfm1["Label of Customer"].value_counts().indexexplode=[0.1,0.1,0.1,0,0,0,0,0]color=["deepskyblue","steelblue","lightskyblue","aliceblue","skyblue","cadetblue","cornflowerblue","dodgerblue"]plt.figure(figsize=(10, 7))plt.pie(x=value_counts,labels=labels,autopct="%.2f%%",explode=explode,colors=color,wedgeprops={"linewidth":0.5,"edgecolor":"black"},textprops={"fontsize":12,"color":"black"})plt.legend(labels,bbox_to_anchor=(1, 1), loc="best", borderaxespad=0.7)plt.title("客户类别细分情况")plt.show()

# 用户特征提取user_features=["visitnum_oneyear","starprefer","sid","price_sensitive","ordernum_oneyear","ordercanncelednum","ordercanceledprecent","lastpvgap",               "lasthtlordergap","landhalfhours","iforderpv_24h","historyvisit_totalordernum","historyvisit_avghotelnum","h",               "delta_price2","delta_price1","decisionhabit_user","customer_value_profit","ctrip_profits","cr","consume_level"]#生成用户特征的相关性矩阵user_corr=df[user_features].corr()#绘制用户特征的相关性矩阵热度图fig,ax = plt.subplots(figsize=(18, 12))sns.heatmap(user_corr, xticklabels=True, yticklabels=True, square=False, linewidths=.5, annot=True, cmap="YlGnBu")

#删除historyvisit_totalordernum列delete_columns = ["historyvisit_totalordernum"]df.drop(delete_columns,axis=1,inplace= True)#PCA主成分分析from sklearn.decomposition import PCA#定义降维函数def PCA_transform(df,col,new_col,n=1):    pca=PCA(n_components=n)    pca.fit(df[col])    df[new_col]=pca.transform(df[col])    # 添加新生成列    df.drop(col,axis=1,inplace= True)   # 删除原来的特征列#选择特征price_prefer=["delta_price1","delta_price2"]                       # 用户价格偏好visit_num=["decisionhabit_user","historyvisit_avghotelnum"]        #用户访问数c_value=["customer_value_profit","ctrip_profits"]                  # 用户价值#应用函数PCA_transform(df,price_prefer,"price_prefer")PCA_transform(df,visit_num,"visit_num")PCA_transform(df,c_value,"c_value"

# 选取刻画用户的重要指标user_feature = ["consume_level","c_value","day_advanced","h","lasthtlordergap","lastpvgap",                "landhalfhours","ordernum_oneyear","ordercanceledprecent","price_sensitive","price_prefer","sid","starprefer","visit_num"]user_attributes = df[user_feature]# 数据标准化from sklearn.preprocessing import StandardScalerscaler = StandardScaler()scaler.fit(user_attributes)user_attributes = scaler.transform(user_attributes)#K-Means聚类from sklearn.cluster import KMeansKmeans=KMeans(n_clusters=3,random_state=13)                                     # 建立KMean模型Kmeans.fit(user_attributes)                                                     # 训练模型k_char=Kmeans.cluster_centers_                                                  # 得到每个分类的质心personas=pd.DataFrame(k_char.T,index=user_feature,columns=["0类","1类","2类"])  # 用户画像表personas

#绘制热力图fig,ax = plt.subplots(figsize=(5, 10))sns.heatmap(personas, xticklabels=True, yticklabels=True, square=False, linewidths=.5, annot=True, cmap="YlGnBu")

#绘制饼形图class_k=list(Kmeans.labels_)  # 每个类别的用户个数percent=[class_k.count(0)/len(user_attributes),class_k.count(1)/len(user_attributes),class_k.count(2)/len(user_attributes)]   # 每个类别用户个数占比fig, ax = plt.subplots(figsize=(11,11))colors=["green","darkblue","lightyellow"]types=["中价值用户","高价值用户","低价值用户"]ax.pie(percent,radius=1,autopct="%.2f%%",pctdistance=0.75,colors=colors,labels=types)ax.pie([1], radius=0.6,colors="w")plt.show()