Bike Sharing Demand （Kaggle）

Forecast use of a city bikeshare system

Bike sharing systems are a means of renting bicycles where the process of obtaining membership, rental, and bike return is automated via a network of kiosk locations throughout a city. Using these systems, people are able rent a bike from a one location and return it to a different place on an as-needed basis. Currently, there are over 500 bike-sharing programs around the world.

The data generated by these systems makes them attractive for researchers because the duration of travel, departure location, arrival location, and time elapsed is explicitly recorded. Bike sharing systems therefore function as a sensor network, which can be used for studying mobility in a city. In this competition, participants are asked to combine historical usage patterns with weather data in order to forecast bike rental demand in the Capital Bikeshare program in Washington, D.C.

给出一个城市的自行车租借系统的历史租借数据，要求预测自行车租借数量。

Data Fields

datetime - hourly date + timestamp  
season -  1 = spring, 2 = summer, 3 = fall, 4 = winter 
holiday - whether the day is considered a holiday
workingday - whether the day is neither a weekend nor holiday
weather - 1: Clear, Few clouds, Partly cloudy, Partly cloudy 
2: Mist + Cloudy, Mist + Broken clouds, Mist + Few clouds, Mist 
3: Light Snow, Light Rain + Thunderstorm + Scattered clouds, Light Rain + Scattered clouds 
4: Heavy Rain + Ice Pallets + Thunderstorm + Mist, Snow + Fog 
temp - temperature in Celsius
atemp - "feels like" temperature in Celsius
humidity - relative humidity
windspeed - wind speed
casual - number of non-registered user rentals initiated
registered - number of registered user rentals initiated
count - number of total rentals

主要的特征工程在于datetime上，很容易想到凌晨2点和下午两点的租借数是很不一样的。用的是RandomForestRegressor，很容易的得到了一个排名在1/3位置的结果。

import csv
import pandas as pd
from math import log, sqrt
from sklearn.ensemble import RandomForestRegressor
from sklearn.cross_validation import train_test_split

data_file = pd.read_csv("train.csv", header=0)
test_file = pd.read_csv("test.csv", header=0)

data_file["hour"] = pd.factorize([i.split()[1][:2] for i in data_file["datetime"]])[0]
data_file = data_file.drop("casual", axis=1)
data_file = data_file.drop("registered", axis=1)
data_file["rent"] = data_file["count"]
data_file = data_file.drop("count", axis=1)
test_file["hour"] = pd.factorize([i.split()[1][:2] for i in test_file["datetime"]])[0]

train_x, test_x, train_y, test_y = train_test_split(
    data_file.values[:,1:-1], data_file.values[:, -1], test_size=.01, random_state=0)
test_data = test_file.values[:, 1:]
ids = test_file.values[:, 0]

def rmsle(pred, real):
    n = len(pred)
    return sqrt(sum([(log(pred[i]+1)-log(real[i]+1))**2 for i in range(n)])/n)

reg = RandomForestRegressor(n_estimators=100, random_state=0)
reg.fit(train_x, train_y)
print reg.feature_importances_
print rmsle(reg.predict(test_x), test_y)
print rmsle(reg.predict(train_x), train_y)
output = reg.predict(test_data).astype(int)

predictions_file = open("RandomForestRegssor.csv", "wb")
open_file_object = csv.writer(predictions_file)
open_file_object.writerow(["datetime","count"])
open_file_object.writerows(zip(ids, output))
predictions_file.close()
print "done."