1、背景:https://tianchi.aliyun.com/competition/introduction.htm?spm=5176.100066.0.0.3f6e7d83RQgWEL&raceId=231598
分数是0.58比较弱,代码这里参考。
2、通过比赛提取的特征如下:
特征 | 类型 | 说明 |
link_ID | string | 每条路段(link)唯一标识 |
link_seq | int | 132条路段从1-132编号; |
length | int | link长度(米) |
width | int | link宽度(米) |
link_class | int | link道路等级,如1代表主干道 |
date | string | 日期,如2015-10-01 |
week | int | 星期,根据日期映射到星期,从1到7 |
time_interval | string | 时间段,如[2015-09-01 00:00:00,2015-09-01 00:02:00) |
time_slot | int | 时间片,根据时间段映射一天24小时,从1到720,每段2分 |
avg_travel_time | float | 该时间片平均旅行时间的均值,反应集中趋势 |
inlinks_atl_1 | float | 车辆在该路段上(timeslot-上游平均旅行时间)时间段的平均旅行时间,上游最多4个路段汇入,如果小于4,则大于4的为0。 测试集中,该值是通过决策树回归预测出来的。 |
inlinks_atl_2 | float | |
inlinks_atl_3 | float | |
inlinks_atl_4 | float | |
inlinks_atl_1 | float | 车辆在该路段上(timeslot+平均旅行时间)时间段的平均旅行时间,下游最多4个路段汇出,如果小于4,则大于4的为0。 测试集中,该值是通过决策树回归预测出来的。 |
inlinks_atl_2 | float | |
inlinks_atl_3 | float | |
inlinks_atl_4 | float | |
travel_time | float | 车辆在该路段上的平均旅行时间(秒) |
# -*- coding: utf-8 -*-
import pandas as pd
import time
import numpy as np
from sklearn import metrics
from sklearn import tree
from sklearn.linear_model import LinearRegression
def main():
#加载标记数据
label_ds=pd.read_csv(r"link_train_0801.txt",sep='\t',encoding='utf8',\
names=['link_id','link_seq','length','width','link_class','start_date','week','time_interval','time_slot','travel_time',\
'avg_travel_time','sd_travel_time','inlinks_num','outlinks_num','inlinks_avg_travel_time','outlinks_avg_travel_time',\
'inlinks_atl_1','inlinks_atl_2','inlinks_atl_3','inlinks_atl_4','outlinks_atl_1','outlinks_atl_2','outlinks_atl_3','outlinks_atl_4'])
label_ds["link_id"] = label_ds["link_id"].astype("string")
label_ds["link_seq"] = label_ds["link_seq"].astype("int")
label_ds["length"] = label_ds["length"].astype("int")
label_ds["width"] = label_ds["width"].astype("int")
label_ds["link_class"] = label_ds["link_class"].astype("int")
label_ds["start_date"] = label_ds["start_date"].astype("string")
label_ds["week"] = label_ds["week"].astype("int")
label_ds["time_interval"] = label_ds["time_interval"].astype("string")
label_ds["time_slot"] = label_ds["time_slot"].astype("int")
label_ds["travel_time"] = label_ds["travel_time"].astype("float")
label_ds["avg_travel_time"] = label_ds["avg_travel_time"].astype("float")
label_ds["sd_travel_time"] = label_ds["sd_travel_time"].astype("float")
label_ds["inlinks_num"] = label_ds["inlinks_num"].astype("int")
label_ds["outlinks_num"] = label_ds["outlinks_num"].astype("int")
label_ds["inlinks_avg_travel_time"] = label_ds["inlinks_avg_travel_time"].astype("float")
label_ds["outlinks_avg_travel_time"] = label_ds["outlinks_avg_travel_time"].astype("float")
label_ds["inlinks_atl_1"] = label_ds["inlinks_atl_1"].astype("float")
label_ds["inlinks_atl_2"] = label_ds["inlinks_atl_2"].astype("float")
label_ds["inlinks_atl_3"] = label_ds["inlinks_atl_3"].astype("float")
label_ds["inlinks_atl_4"] = label_ds["inlinks_atl_4"].astype("float")
label_ds["outlinks_atl_1"] = label_ds["outlinks_atl_1"].astype("float")
label_ds["outlinks_atl_2"] = label_ds["outlinks_atl_2"].astype("float")
label_ds["outlinks_atl_3"] = label_ds["outlinks_atl_3"].astype("float")
label_ds["outlinks_atl_4"] = label_ds["outlinks_atl_4"].astype("float")
#加载预测数据
unlabel_ds=pd.read_csv(r"link_test_0801.txt",sep='\t',encoding='utf8',\
names=['link_id','link_seq','length','width','link_class','start_date','week','time_interval','time_slot',\
'avg_travel_time','sd_travel_time','inlinks_num','outlinks_num','inlinks_avg_travel_time','outlinks_avg_travel_time',\
'inlinks_atl_1','inlinks_atl_2','inlinks_atl_3','inlinks_atl_4','outlinks_atl_1','outlinks_atl_2','outlinks_atl_3','outlinks_atl_4'])
unlabel_ds["link_id"] = unlabel_ds["link_id"].astype("string")
unlabel_ds["link_seq"] = unlabel_ds["link_seq"].astype("int")
unlabel_ds["length"] = unlabel_ds["length"].astype("int")
unlabel_ds["width"] = unlabel_ds["width"].astype("int")
unlabel_ds["link_class"] = unlabel_ds["link_class"].astype("int")
unlabel_ds["start_date"] = unlabel_ds["start_date"].astype("string")
unlabel_ds["week"] = unlabel_ds["week"].astype("int")
unlabel_ds["time_interval"] = unlabel_ds["time_interval"].astype("string")
unlabel_ds["time_slot"] = unlabel_ds["time_slot"].astype("int")
unlabel_ds["avg_travel_time"] = unlabel_ds["avg_travel_time"].astype("float")
unlabel_ds["sd_travel_time"] = unlabel_ds["sd_travel_time"].astype("float")
unlabel_ds["inlinks_num"] = unlabel_ds["inlinks_num"].astype("int")
unlabel_ds["outlinks_num"] = unlabel_ds["outlinks_num"].astype("int")
unlabel_ds["inlinks_avg_travel_time"] = unlabel_ds["inlinks_avg_travel_time"].astype("float")
unlabel_ds["outlinks_avg_travel_time"] = unlabel_ds["outlinks_avg_travel_time"].astype("float")
unlabel_ds["inlinks_atl_1"] = unlabel_ds["inlinks_atl_1"].astype("float")
unlabel_ds["inlinks_atl_2"] = unlabel_ds["inlinks_atl_2"].astype("float")
unlabel_ds["inlinks_atl_3"] = unlabel_ds["inlinks_atl_3"].astype("float")
unlabel_ds["inlinks_atl_4"] = unlabel_ds["inlinks_atl_4"].astype("float")
unlabel_ds["outlinks_atl_1"] = unlabel_ds["outlinks_atl_1"].astype("float")
unlabel_ds["outlinks_atl_2"] = unlabel_ds["outlinks_atl_2"].astype("float")
unlabel_ds["outlinks_atl_3"] = unlabel_ds["outlinks_atl_3"].astype("float")
unlabel_ds["outlinks_atl_4"] = unlabel_ds["outlinks_atl_4"].astype("float")
outit=pd.DataFrame()#输出结果
mr_df=pd.DataFrame()#输出link的mape和rmse
mape=0;
rmse=0;
train_df=label_ds.loc[(pd.to_datetime(label_ds["start_date"])<'2016-06-01')]#训练集
valid_df=label_ds.loc[(pd.to_datetime(label_ds["start_date"])>='2016-06-01')]#验证集train_df.sample(frac=0.2)
for linkid in range(1,133):
#提取训练集、验证集、测试集
train_df_id=train_df.loc[(train_df["link_seq"]==linkid)]
print "训练集,有", train_df_id.shape[0], "行", train_df_id.shape[1], "列"
valid_df_id=valid_df.loc[(valid_df["link_seq"]==linkid)]
print "验证集,有", valid_df_id.shape[0], "行", valid_df_id.shape[1], "列"
test_df=unlabel_ds.loc[(unlabel_ds["link_seq"]==linkid)]#测试集
print "测试集,有", test_df.shape[0], "行", test_df.shape[1], "列"
#特征选择
#模型训练
train_X=train_df_id[['link_seq','time_slot','length','avg_travel_time',\
'inlinks_atl_1','inlinks_atl_2','inlinks_atl_3','inlinks_atl_4','outlinks_atl_1','outlinks_atl_2','outlinks_atl_3','outlinks_atl_4']]
train_X=train_X.fillna(0)#空值替换为0
train_y = train_df_id['travel_time']#标记
model_it=LinearRegression()#tree.DecisionTreeRegressor()
model_it.fit(train_X, train_y)
#模型验证
valid_X=valid_df_id[['link_seq','time_slot','length','avg_travel_time',\
'inlinks_atl_1','inlinks_atl_2','inlinks_atl_3','inlinks_atl_4','outlinks_atl_1','outlinks_atl_2','outlinks_atl_3','outlinks_atl_4']]
valid_X=valid_X.fillna(0)#空值替换为0
valid_y=valid_df_id['travel_time']
pre_valid_y=model_it.predict(valid_X)
abs_y=abs(pre_valid_y-valid_y)
abs_error=abs_y.sum()#求和
mape_id=abs_error/valid_df_id.shape[0]
rmse_id=np.sqrt(metrics.mean_squared_error(valid_y, pre_valid_y))#均方差,模型评估
print "linkseq="+str(linkid)+"的mape=",mape_id
print "linkseq="+str(linkid)+"的RMSE=",rmse_id
mr_list=[[linkid,mape_id,rmse_id]]
mr_df=mr_df.append(mr_list)
mape=mape+mape_id
rmse=rmse+rmse_id
#模型预测
test_X = test_df[['link_seq','time_slot','length','avg_travel_time',\
'inlinks_atl_1','inlinks_atl_2','inlinks_atl_3','inlinks_atl_4','outlinks_atl_1','outlinks_atl_2','outlinks_atl_3','outlinks_atl_4']]
test_X=test_X.fillna(0)#空值替换为0
test_info = test_df[['link_id','start_date','time_interval']]
test_y=model_it.predict(test_X)
test_info["travel_time"]=test_y
outit=outit.append(test_info)#追加到输出结果
print "all mape:",mape/132
print "all RMSE:",rmse/132
mr_df.to_csv('linkmape.txt',sep='#',index=False,header=None)
outit.to_csv('outit.txt',sep='#',index=False,header=None)#输出预测数据
#执行
if __name__ == '__main__':
start = time.clock()
main()
end = time.clock()
print('finish all in %s' % str(end - start))