Machine Learning. Part II.

Non-residential building occupancy modeling. Part II. Occupancy classification

So dataset was taken from this place. The dataset comprised of different sources: surveys, data logging from sensors, weather, environment variables. Total feature list consist of 118 features and can be grouped as general (occupancy, time), environment (indoor, outdoor), personal characteristics (age, office type, accepted sensation range etc), comfort/productivity/satisfaction, behavior (clothing, window, interaction with thermostat etc ), personal values (choices on different set points). It contains data on 24 occupants whether it private office or joint one, the first task is to implement binary classification of each occupant using some input data from sensors and time.
For rapid protoyping I will use python Tensor Flow wrapper Keras along Anaconda framework.

• First, loading all required libraries

from keras.models import Sequential  
from keras.layers import Dense  
import numpy  
import matplotlib.pyplot as plt  
# fix random seed for reproducibility  
numpy.random.seed(15) 

• Next loading required data and splitting for X and Y (input matrix and output matrix)

In my case I took Time, Temperature, CO2 input values

# load pima indians dataset  
  dataset = numpy.loadtxt("sample.csv", delimiter=",")  
  # split into input (X) and output (Y) variables  
  X = dataset[:,0:3] # Take 3 columns from 0 to 2 as inputs 
  Y = dataset[:,3]  #Third index or fourth column from input file is occupancy value 0 or 1
  # either present or absent

• After, we create three layer nerual network by creating a model and adding layers one by one:

# create model  
 model = Sequential()# In this example, we will use a fully-connected network structure with three layers  
 model.add(Dense(7, input_dim=7, activation='tanh'))# First layer 7 neurons, 8 inputs, rectifier activation func  
 model.add(Dense(3, activation='tanh'))# hidden layer   
 model.add(Dense(1, activation='sigmoid'))  

• Finally, we compiling model and training it:

# Compile model  
 model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])  
 # We can train or fit our model on our loaded data by calling the fit() function on the model.  
 # Fit the model or Model Training  
 model.fit(X, Y, epochs=150, batch_size=4)  
'''
  In this case we are fitting the whole our data (X,Y) and running it for 150 iterations.
  However if you want to split your data into training and validation parts you 
  can write: 
  model.fit(X, Y,validation_split=0.33, epochs=150, batch_size=10)
  thus, model will split data into 77% training, 33% validation
'''
 # calculate predictions  
 predictions = model.predict(X)  
 rounded = [round(x[0]) for x in predictions]  

Some thoughts about model's accuracy and loss

While analyzing mine dataset I tried to figure out what input data I may choose so that model will give the highest accuracy result. Of course, there is not solid and general approach towards defining correct inputs for your prediction model, but we always can experiment with it. So in my case experimented with different combinations such as [temperature, relative humidity] - accuracy 74%, [temperature, CO2] - accuracy 79.16%, [time, temperature, CO2] - accuracy 91.54%. Finally, it is possible to notice how particular data input affect accuracy. Comparing 1st and 2nd set you see that CO2 increeses accuracy on ~ 5%, therefore we may suggest that CO2 reading are preferable over relative humidity data. Big thanks to Dr.Jason Brownlee and his site about ML.