The goal of this notebook is not to do the best model for each Time series. It is just a comparison of few models when you have one Time Series. The presentation present a different approaches to forecast a Time Series.In this notebook we will be using web traffic data from kaggle.
The plan of the notebook is:
I. Importation & Data Cleaning
II. Aggregation & Visualization
III. Machine Learning Approach
IV Basic Model Approach
V. ARIMA approach (Autoregressive Integrated Moving Average)
VI. (FB) Prophet Approach
VII. Keras Starter
VIII. Comparaison & Conclusion
I. Importation & Data Cleaning
In this first part we will choose the Time Series to work in the others parts. The idea is to find a Time Serie who could be interesting to work with. So in the data we can find 145K Time Series. We will Find a good Time Series to introduce four approaches! So the first step is to import few libraries and the data. The four approaches are Basic Approach / ML Approach / GAM Approach / ARIMA Approach.
import pandas as pd
import numpy as np
import warnings
import scipy
from datetime import timedelta
# Forceasting with decompasable model
from pylab import rcParams
import statsmodels.api as sm
from statsmodels.tsa.stattools import adfuller
# For marchine Learning Approach
from statsmodels.tsa.tsatools import lagmat
from sklearn.linear_model import LinearRegression, RidgeCV
from sklearn.ensemble import RandomForestRegressor
from sklearn.metrics import r2_score
# Visualisation
import matplotlib.pyplot as plt
import seaborn as sns
plt.style.use('fivethirtyeight')
warnings.filterwarnings('ignore')
/opt/conda/lib/python3.6/site-packages/statsmodels/compat/pandas.py:56: FutureWarning: The pandas.core.datetools module is deprecated and will be removed in a future version. Please use the pandas.tseries module instead. from pandas.core import datetools
# Load the data
train = pd.read_csv("../input/train_1.csv")
train_flattened = pd.melt(train[list(train.columns[-50:])+['Page']], id_vars='Page', var_name='date', value_name='Visits')
train_flattened['date'] = train_flattened['date'].astype('datetime64[ns]')
train_flattened['weekend'] = ((train_flattened.date.dt.dayofweek) // 5 == 1).astype(float)
# Median by page
df_median = pd.DataFrame(train_flattened.groupby(['Page'])['Visits'].median())
df_median.columns = ['median']
# Average by page
df_mean = pd.DataFrame(train_flattened.groupby(['Page'])['Visits'].mean())
df_mean.columns = ['mean']
# Merging data
train_flattened = train_flattened.set_index('Page').join(df_mean).join(df_median)
train_flattened.reset_index(drop=False,inplace=True)
train_flattened['weekday'] = train_flattened['date'].apply(lambda x: x.weekday())
# Feature engineering with the date
train_flattened['year']=train_flattened.date.dt.year
train_flattened['month']=train_flattened.date.dt.month
train_flattened['day']=train_flattened.date.dt.day
train_flattened.head()
| Page | date | visits | weekend | mean | median | weekday | Year | Month | Day |
|---|---|---|---|---|---|---|---|---|---|
| !vote_en.wikipedia.org_all-access_all-agents | 2016-11-12 | 2.0 | 1.0 | 2.836735 | 2.0 | 5 | 2016 | 11 | 12 |
| !vote_en.wikipedia.org_all-access_all-agents | 2016-11-13 | 2.0 | 1.0 | 2.836735 | 2.0 | 6 | 2016 | 11 | 13 |
| !vote_en.wikipedia.org_all-access_all-agents | 2016-11-14 | 4.0 | 0.0 | 2.836735 | 2.0 | 0 | 2016 | 11 | 14 |
| !vote_en.wikipedia.org_all-access_all-agents | 2016-11-15 | 8.0 | 0.0 | 2.836735 | 2.0 | 1 | 2016 | 11 | 15 |
| !vote_en.wikipedia.org_all-access_all-agents | 2016-11-16 | 3.0 | 0.0 | 2.836735 | 2.0 | 2 | 2016 | 11 | 16 |
This part allowed us to prepare our data. We had created new features that we use in the next steps. Days, Months, Years are interesting to forecast with a Machine Learning Approach or to do an analysis. If you have another idea to improve this first part: Fork this notebook and improve it or share your idea in the comments.
II. Aggregation & Visualisation
plt.figure(figsize=(50, 8))
mean_group = train_flattened[['Page','date','Visits']].groupby(['date'])['Visits'].mean()
plt.plot(mean_group)
plt.title('Time Series - Average')
plt.show()
plt.figure(figsize=(50, 8))
median_group = train_flattened[['Page','date','Visits']].groupby(['date'])['Visits'].median()
plt.plot(median_group, color = 'r')
plt.title('Time Series - median')
plt.show()
plt.figure(figsize=(50, 8))
std_group = train_flattened[['Page','date','Visits']].groupby(['date'])['Visits'].std()
plt.plot(std_group, color = 'g')
plt.title('Time Series - std')
plt.show()
# For the next graphics
train_flattened['month_num'] = train_flattened['month']
train_flattened['month'].replace('11','11 - November',inplace=True)
train_flattened['month'].replace('12','12 - December',inplace=True)
train_flattened['weekday_num'] = train_flattened['weekday']
train_flattened['weekday'].replace(0,'01 - Monday',inplace=True)
train_flattened['weekday'].replace(1,'02 - Tuesday',inplace=True)
train_flattened['weekday'].replace(2,'03 - Wednesday',inplace=True)
train_flattened['weekday'].replace(3,'04 - Thursday',inplace=True)
train_flattened['weekday'].replace(4,'05 - Friday',inplace=True)
train_flattened['weekday'].replace(5,'06 - Saturday',inplace=True)
train_flattened['weekday'].replace(6,'07 - Sunday',inplace=True)
train_group = train_flattened.groupby(["month", "weekday"])['Visits'].mean().reset_index()
train_group = train_group.pivot('weekday','month','Visits')
train_group.sort_index(inplace=True)
sns.set(font_scale=3.5)
# Draw a heatmap with the numeric values in each cell
f, ax = plt.subplots(figsize=(50, 30))
sns.heatmap(train_group, annot=False, ax=ax, fmt="d", linewidths=2)
plt.title('Web Traffic Months cross Weekdays')
plt.show()
This heatmap show us in average the web traffic by weekdays cross the months. In our data we can see there are less activity in Friday and Saturday for December and November. And the biggest traffic is on the period Monday - Wednesday. It is possible to do Statistics Test to check if our intuition is ok. But You have a lot of works !
train_day = train_flattened.groupby(["month", "day"])['Visits'].mean().reset_index()
train_day = train_day.pivot('day','month','Visits')
train_day.sort_index(inplace=True)
train_day.dropna(inplace=True)
# Draw a heatmap with the numeric values in each cell
f, ax = plt.subplots(figsize=(50, 30))
sns.heatmap(train_day, annot=False, ax=ax, fmt="d", linewidths=2)
plt.title('Web Traffic Months cross days')
plt.show()
With this graph it is possible to see they are two periods with a bigger activity than the rest. The two periods are 25-29 December and 13-14 November. And we can see one period with little activity 15-17 December. They are maybe few outliers during these two periods. You must to investigate more. (coming soon…)
III. ML Approach
The first approach introduces is the Machine Learning Approach. We will use just a AdaBoostRegressor but you can try with other models if you want to find the best model. I tried with a linear model as like Ridge but ADA model is better. I will be interesting to check if GB or XGB can bit ADA. It is possible to do a Neural Network approach too. But this approach will be done if the kagglers want more !!
times_series_means = pd.DataFrame(mean_group).reset_index(drop=False)
times_series_means['weekday'] = times_series_means['date'].apply(lambda x: x.weekday())
times_series_means['Date_str'] = times_series_means['date'].apply(lambda x: str(x))
times_series_means[['year','month','day']] = pd.DataFrame(times_series_means['Date_str'].str.split('-',2).tolist(), columns = ['year','month','day'])
date_staging = pd.DataFrame(times_series_means['day'].str.split(' ',2).tolist(), columns = ['day','other'])
times_series_means['day'] = date_staging['day']*1
times_series_means.drop('Date_str',axis = 1, inplace =True)
times_series_means.head()
| date | Visits | weekday | year | month | day |
|---|---|---|---|---|---|
| 2016-11-12 | 1536.471757 | 5 | 2016 | 11 | 12 |
| 2016-11-13 | 1679.630589 | 6 | 2016 | 11 | 13 |
| 2016-11-14 | 1701.941831 | 0 | 2016 | 11 | 14 |
| 2016-11-15 | 1564.172228 | 1 | 2016 | 11 | 15 |
| 2016-11-16 | 1520.673959 | 2 | 2016 | 11 | 16 |
times_series_means.reset_index(drop=True,inplace=True)
def lag_func(data,lag):
lag = lag
X = lagmat(data["diff"], lag)
lagged = data.copy()
for c in range(1,lag+1):
lagged["lag%d" % c] = X[:, c-1]
return lagged
def diff_creation(data):
data["diff"] = np.nan
data.ix[1:, "diff"] = (data.iloc[1:, 1].as_matrix() - data.iloc[:len(data)-1, 1].as_matrix())
return data
df_count = diff_creation(times_series_means)
# Creation of 7 features with "diff"
lag = 7
lagged = lag_func(df_count,lag)
last_date = lagged['date'].max()
lagged.head()
| date | Visits | weekday | year | month | day | diff | lag1 | lag2 | lag3 | lag4 | lag5 | lag6 | lag7 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 2016-11-12 | 1536.471757 | 5 | 2016 | 11 | 12 | NaN | 0.000000 | 0.000000 | 0.000000 | 0.0 | 0.0 | 0.0 | 0.0 |
| 2016-11-13 | 1679.630589 | 6 | 2016 | 11 | 13 | 143.158833 | NaN | 0.000000 | 0.000000 | 0.0 | 0.0 | 0.0 | 0.0 |
| 2016-11-14 | 1701.941831 | 0 | 2016 | 11 | 14 | 22.311241 | 143.158833 | NaN | 0.000000 | 0.0 | 0.0 | 0.0 | 0.0 |
| 2016-11-15 | 1564.172228 | 1 | 2016 | 11 | 15 | -137.769602 | 22.311241 | 143.158833 | NaN | 0.0 | 0.0 | 0.0 | 0.0 |
| 2016-11-16 | 1520.673959 | 2 | 2016 | 11 | 16 | -43.498270 | -137.769602 | 22.311241 | 143.158833 | NaN | 0.0 | 0.0 | 0.0 |
# Train Test split
def train_test(data_lag):
xc = ["lag%d" % i for i in range(1,lag+1)] + ['weekday'] + ['day']
split = 0.70
xt = data_lag[(lag+1):][xc]
yt = data_lag[(lag+1):]["diff"]
isplit = int(len(xt) * split)
x_train, y_train, x_test, y_test = xt[:isplit], yt[:isplit], xt[isplit:], yt[isplit:]
return x_train, y_train, x_test, y_test, xt, yt
x_train, y_train, x_test, y_test, xt, yt = train_test(lagged)
# Linear Model
from sklearn.ensemble import ExtraTreesRegressor,GradientBoostingRegressor, BaggingRegressor, AdaBoostRegressor
from sklearn.metrics import mean_absolute_error, r2_score
def modelisation(x_tr, y_tr, x_ts, y_ts, xt, yt, model0, model1):
# Modelisation with all product
model0.fit(x_tr, y_tr)
prediction = model0.predict(x_ts)
r2 = r2_score(y_ts.as_matrix(), model0.predict(x_ts))
mae = mean_absolute_error(y_ts.as_matrix(), model0.predict(x_ts))
print ("-----------------------------------------------")
print ("mae with 70% of the data to train:", mae)
print ("-----------------------------------------------")
# Model with all data
model1.fit(xt, yt)
return model1, prediction, model0
model0 = AdaBoostRegressor(n_estimators = 5000, random_state = 42, learning_rate=0.01)
model1 = AdaBoostRegressor(n_estimators = 5000, random_state = 42, learning_rate=0.01)
clr, prediction, clr0 = modelisation(x_train, y_train, x_test, y_test, xt, yt, model0, model1)
----------------------------------------------- mae with 70% of the data to train: 77.19244965544415 -----------------------------------------------
# Performance 1
plt.style.use('ggplot')
plt.figure(figsize=(50, 12))
line_up, = plt.plot(prediction,label='Prediction')
line_down, = plt.plot(np.array(y_test),label='Reality')
plt.ylabel('Series')
plt.legend(handles=[line_up, line_down])
plt.title('Performance of predictions - Benchmark Predictions vs Reality')
plt.show()
# Prediction
def pred_df(data,number_of_days):
data_pred = pd.DataFrame(pd.Series(data["date"][data.shape[0]-1] + timedelta(days=1)),columns = ["date"])
for i in range(number_of_days):
inter = pd.DataFrame(pd.Series(data["date"][data.shape[0]-1] + timedelta(days=i+2)),columns = ["date"])
data_pred = pd.concat([data_pred,inter]).reset_index(drop=True)
return data_pred
data_to_pred = pred_df(df_count,30)
def initialisation(data_lag, data_pred, model, xtrain, ytrain, number_of_days):
# Initialisation
model.fit(xtrain, ytrain)
for i in range(number_of_days-1):
lag1 = data_lag.tail(1)["diff"].values[0]
lag2 = data_lag.tail(1)["lag1"].values[0]
lag3 = data_lag.tail(1)["lag2"].values[0]
lag4 = data_lag.tail(1)["lag3"].values[0]
lag5 = data_lag.tail(1)["lag4"].values[0]
lag6 = data_lag.tail(1)["lag5"].values[0]
lag7 = data_lag.tail(1)["lag6"].values[0]
lag8 = data_lag.tail(1)["lag7"].values[0]
data_pred['weekday'] = data_pred['date'].apply(lambda x:x.weekday())
weekday = data_pred['weekday'][0]
row = pd.Series([lag1,lag2,lag3,lag4,lag5,lag6,lag7,lag8,weekday]
,['lag1', 'lag2', 'lag3','lag4','lag5','lag6','lag7','lag8','weekday'])
to_predict = pd.DataFrame(columns = ['lag1', 'lag2', 'lag3','lag4','lag5','lag6','lag7','lag8','weekday'])
prediction = pd.DataFrame(columns = ['diff'])
to_predict = to_predict.append([row])
prediction = pd.DataFrame(model.predict(to_predict),columns = ['diff'])
# Loop
if i == 0:
last_predict = data_lag["Visits"][data_lag.shape[0]-1] + prediction.values[0][0]
if i > 0 :
last_predict = data_lag["Visits"][data_lag.shape[0]-1] + prediction.values[0][0]
data_lag = pd.concat([data_lag,prediction.join(data_pred["date"]).join(to_predict)]).reset_index(drop=True)
data_lag["Visits"][data_lag.shape[0]-1] = last_predict
# test
data_pred = data_pred[data_pred["date"]>data_pred["date"][0]].reset_index(drop=True)
return data_lag
model_fin = AdaBoostRegressor(n_estimators = 5000, random_state = 42, learning_rate=0.01)
lagged = initialisation(lagged, data_to_pred, model_fin, xt, yt, 30)
lagged[lagged['diff']<0]
lagged.ix[(lagged.Visits < 0), 'Visits'] = 0
df_lagged = lagged[['Visits','date']]
df_train = df_lagged[df_lagged['date'] <= last_date]
df_pred = df_lagged[df_lagged['date'] >= last_date]
plt.style.use('ggplot')
plt.figure(figsize=(30, 5))
plt.plot(df_train.date,df_train.Visits)
plt.plot(df_pred.date,df_pred.Visits,color='b')
plt.title('Training time series in red, Prediction on 30 days in blue -- ML Approach')
plt.show()
Finshed for the first approach ! The ML method requires a lot of work ! You need to create the features, the data to collect the prediction, optimisation etc… This method done a good results when there are a weekly pattern identified or a monthly pattern but we need more data.
IV. Basic Approach
For this model We will use a simple model with the average of the activity by weekdays. In general rules the simplest things give good results !
lagged_basic = lagged[['date','Visits','weekday']]
lagged_basic_tr = lagged_basic[lagged_basic['date'] < last_date]
lagged_basic_pred = lagged_basic[lagged_basic['date'] >= last_date]
lagged_basic_pred.drop('Visits',inplace=True,axis=1)
prediction_by_days = pd.DataFrame(lagged_basic.groupby(['weekday'])['Visits'].mean())
prediction_by_days.reset_index(drop=False,inplace=True)
prediction_by_days
| Weekday | Visits | |
|---|---|---|
| 0 | 0.0 | 1613.317878 |
| 1 | 1.0 | 1596.790788 |
| 2 | 2.0 | 1548.737929 |
| 3 | 3.0 | 1518.802011 |
| 4 | 4.0 | 1461.298652 |
| 5 | 5.0 | 1489.810459 |
| 6 | 6.0 | 1575.831684 |
basic_pred = pd.merge(lagged_basic_pred,prediction_by_days,on='weekday')
basic_approach = pd.concat([lagged_basic_tr,basic_pred])
plot_basic = np.array(basic_approach[basic_approach['date'] > last_date].sort_values(by='date').Visits)
plt.figure(figsize=(30, 5))
plt.plot(plot_basic)
plt.title('Display the predictions with the Basic model')
plt.show()
df_lagged = basic_approach[['Visits','date']].sort_values(by='date')
df_train = df_lagged[df_lagged['date'] <= last_date]
df_pred = df_lagged[df_lagged['date'] >= last_date]
plt.style.use('ggplot')
plt.figure(figsize=(30, 5))
plt.plot(df_train.date,df_train.Visits)
plt.plot(df_pred.date,df_pred.Visits,color='b')
plt.title('Training time series in red, Prediction on 30 days in blue -- ML Approach')
plt.show()
No optimisation ! No choice between linear, Bagging, boosting or others ! Just with an average by week days and we have a result ! Fast and easily !
V. ARIMA
This part is inspired by: https://www.analyticsvidhya.com/blog/2016/02/time-series-forecasting-codes-python/ Very goodjob with the ARIMA models ! It is more simple when we have directly a stationary Time series. It is not our case…
We will use the Dickey-Fuller Test. More informations here: https://en.wikipedia.org/wiki/Dickey%E2%80%93Fuller_test
# Show Rolling mean, Rolling Std and Test for the stationnarity
df_date_index = times_series_means[['date','Visits']].set_index('date')
def test_stationarity(timeseries):
plt.figure(figsize=(50, 8))
#Determing rolling statistics
rolmean = pd.rolling_mean(timeseries, window=7)
rolstd = pd.rolling_std(timeseries, window=7)
#Plot rolling statistics:
orig = plt.plot(timeseries, color='blue',label='Original')
mean = plt.plot(rolmean, color='red', label='Rolling Mean')
std = plt.plot(rolstd, color='black', label = 'Rolling Std')
plt.legend(loc='best')
plt.title('Rolling Mean & Standard Deviation')
plt.show(block=False)
#Perform Dickey-Fuller test:
print('Results of Dickey-Fuller Test:')
dftest = sm.tsa.adfuller(timeseries['Visits'], autolag='AIC')
dfoutput = pd.Series(dftest[0:4], index=['Test Statistic','p-value','#Lags Used','Number of Observations Used'])
for key,value in dftest[4].items():
dfoutput['Critical Value (%s)'%key] = value
print(dfoutput)
test_stationarity(df_date_index)
Results of Dickey-Fuller Test: Test Statistic -4.292839 p-value 0.000457 #Lags Used 1.000000 Number of Observations Used 48.000000 Critical Value (1%) -3.574589 Critical Value (5%) -2.923954 Critical Value (10%) -2.600039 dtype: float64
Our Time Series is stationary ! it is a good news ! We can to apply the ARIMA Model without transformations.linkcode
Good job ! We have a Time Series Stationary ! We can apply our ARIMA Model !!!
# Naive decomposition of our Time Series as explained above
decomposition = sm.tsa.seasonal_decompose(df_date_index, model='multiplicative',freq = 7)
trend = decomposition.trend
seasonal = decomposition.seasonal
residual = decomposition.resid
rcParams['figure.figsize'] = 30, 20
plt.subplot(411)
plt.title('Obesered = Trend + Seasonality + Residuals')
plt.plot(df_date_index, label='Observed')
plt.legend(loc='best')
plt.subplot(412)
plt.plot(trend, label='Trend')
plt.legend(loc='best')
plt.subplot(413)
plt.plot(seasonal,label='Seasonality')
plt.legend(loc='best')
plt.subplot(414)
plt.plot(residual, label='Residuals')
plt.legend(loc='best')
plt.tight_layout()
plt.show()
We expose the naive decomposition of our time series (More sophisticated methods should be preferred). They are several ways to decompose a time series but in our example we take a simple decomposition on three parts. The additive model is Y[t] = T[t] + S[t] + e[t] The multiplicative model is Y[t] = T[t] x S[t] x e[t] with:
- T[t]: Trend
- S[t]: Seasonality
- e[t]: Residual
An additive model is linear where changes over time are consistently made by the same amount. A linear trend is a straight line. A linear seasonality has the same frequency (width of cycles) and amplitude (height of cycles). A multiplicative model is nonlinear, such as quadratic or exponential. Changes increase or decrease over time. A nonlinear trend is a curved line.A non-linear seasonality has an increasing or decreasing frequency and/or amplitude over time. In ou example we can see it is not a linear model. So it is the reason why we use a multiplicative model.
#from statsmodels.tsa.arima_model import ARIMA
#model = ARIMA(df_date_index, order=(7, 1, 0))
#results_AR = model.fit(disp=-1)
#plt.plot(df_date_index, color = 'blue')
#plt.plot(results_AR.fittedvalues, color='red')
#plt.show()
#forecast = results_AR.forecast(steps = 30)[0]
#plt.figure(figsize=(30, 5))
#plt.plot(pd.DataFrame(np.exp(forecast)))
#plt.title('Display the predictions with the ARIMA model')
#plt.show()
# DataFrame to collect the predictions
#df_prediction_arima = df_date_index.copy()
#list_date = []
#for i in range(31):
# if i >0:
# list_date.append(last_date + pd.to_timedelta(i, unit='D'))
#predictions_arima = pd.DataFrame(list_date,columns = ['Date'])
#predictions_arima['Visits'] = 0
#predictions_arima.set_index('Date',inplace=True)
#predictions_arima['Visits'] = np.exp(forecast)
#df_prediction_arima = df_prediction_arima.append(predictions_arima)
#df_prediction_arima.reset_index(drop=False,inplace=True)
#df_arima = df_prediction_arima[['Visits','index']]
#df_train = df_arima[df_arima['index'] <= last_date]
#df_pred = df_arima[df_prediction_arima['index'] >= last_date]
#plt.style.use('ggplot')
#plt.figure(figsize=(30, 5))
#plt.plot(df_train.index,df_train.Visits)
#plt.plot(df_pred.index,df_pred.Visits,color='b')
#plt.title('Training time series in red, Prediction on 30 days in blue -- ARIMA Model')
#plt.show()
VI. Prophet
Prophet is a forecasting tool availaible in python and R. This tool was created by Facebook. More information on the library here: https://research.fb.com/prophet-forecasting-at-scale/
Compared to the two methods this one will be faster. We can forecast a time series with few lines. In our case we will do a forecast and a display the trend of activity on the period and for a week.
from fbprophet import Prophet
sns.set(font_scale=1)
df_date_index = times_series_means[['date','Visits']]
df_date_index = df_date_index.set_index('date')
df_prophet = df_date_index.copy()
df_prophet.reset_index(drop=False,inplace=True)
df_prophet.columns = ['ds','y']
m = Prophet()
m.fit(df_prophet)
future = m.make_future_dataframe(periods=30,freq='D')
forecast = m.predict(future)
fig = m.plot(forecast)
INFO:fbprophet.forecaster:Disabling yearly seasonality. Run prophet with yearly_seasonality=True to override this. INFO:fbprophet.forecaster:Disabling daily seasonality. Run prophet with daily_seasonality=True to override this.
m.plot_components(forecast);
VI. Keras Starter
In this part we will use Keras without optimization to forecast. It is just a very simple code to begin with Keras and a Time Series. For our example we will try just with one layer and 8 Neurons.
df_dl = times_series_means[['date','Visits']]
train_size = int(len(df_dl) * 0.80)
test_size = len(df_dl) - train_size
train, test = df_dl.iloc[0:train_size,:], df_dl.iloc[train_size:len(df_dl),:]
print(len(train), len(test))
40 10
look_back = 1
def create_dataset(dataset, look_back):
dataX = []
dataY = []
for i in range(len(dataset)-look_back-1):
a = dataset.iloc[i:(i+look_back), 1].values[0]
b = dataset.iloc[i+look_back, 1]
dataX.append(a)
dataY.append(b)
return np.array(dataX), np.array(dataY)
trainX, trainY = create_dataset(train, look_back)
from keras.models import Sequential
from keras.layers import Dense
trainX, trainY = create_dataset(train, look_back)
testX, testY = create_dataset(test, look_back)
model = Sequential()
model.add(Dense(8, input_dim=look_back, activation='relu'))
model.add(Dense(1))
model.compile(loss='mean_absolute_error', optimizer='adam')
model.fit(trainX, trainY, epochs=150, batch_size=2, verbose=0)
Using TensorFlow backend.
Out[44]:
<keras.callbacks.History at 0x7f0da5866470>
trainScore = model.evaluate(trainX, trainY, verbose=0)
print('Train Score: %.2f MSE (%.2f MAE)' % (trainScore, trainScore))
testScore = model.evaluate(testX, testY, verbose=0)
print('Test Score: %.2f MSE (%.2f MAE)' % (testScore, testScore))
Train Score: 65.84 MSE (65.84 MAE) Test Score: 81.51 MSE (81.51 MAE)
trainPredict = model.predict(trainX)
testPredict = model.predict(testX)
# shift train predictions for plotting
trainPredictPlot = np.empty_like(df_dl)
trainPredictPlot[:, :] = np.nan
trainPredictPlot[look_back:len(trainPredict)+look_back, :] = trainPredict
# shift test predictions for plotting
testPredictPlot = np.empty_like(df_dl)
testPredictPlot[:, :] = np.nan
testPredictPlot[len(trainPredict)+(look_back*2)+1:len(df_dl)-1, :] = testPredict
# plot baseline and predictions
plt.plot(np.array(df_dl.Visits))
plt.plot(trainPredictPlot)
plt.plot(testPredictPlot)
plt.title('Predicition with Keras')
plt.show()
VII. Comparison & Conclusion
In Progress…
Important Notice for college students
If you’re a college student and have skills in programming languages, Want to earn through blogging? Mail us at geekycomail@gmail.com
For more Programming related blogs Visit Us Geekycodes . Follow us on Instagram.