This is the implementation of an algorithm to do classification when the distribution of the data generating process changes with time [to be published, https://experiencor.github.io/ftrl_adp.html].

In the following gif, the blue line signifies the decision line to separate two classes. The decision line is adaptive to the centers of the classes, which shift in time.

Code structure

1. ftrl_adp.py => the implementation of FTRL-ADP
2. Algorithm Visualization.ipynb => visualization that shows the online learning process of FTRL-ADP
3. dataset.txt => a simulated classification dataset that contains dynamic concept drifting behavior

Usage

cd ftrl_adp

from ftrl_adp import FTRL_ADP

X_input, Y_label = load_svmlight_file('dataset.txt')

ftrl_adp = FTRL_ADP(L1=1., L2=1., LP = 1., adaptive=True, n_inputs=X_input.shape[1])

for i in xrange(X_input.shape[0]):
    indices = X_input[row].indices
    x = X_input[row].data
    y = y_label[i]
    
    p, decay = classifier.fit(indices, x, y)
    error = [int(np.abs(y-p)>0.5)]