Skip to content

Latest commit

 

History

History
64 lines (49 loc) · 2.5 KB

readme.md

File metadata and controls

64 lines (49 loc) · 2.5 KB
Raw

Automatically segment a Timeseries

Installation

> pip install git+https://github.com/oboulant/timesegment.git

Example

import pandas as pd
import numpy as np

from timesegment import Partition_tree

import matplotlib.pyplot as plt

# Read data
data = pd.read_csv('data_sample.csv')
# Invert time axis (specific for this data)
data = data.iloc[::-1]

# Segment the 256 most recent points
my_obj = Partition_tree(np.array(data['value'])[data.shape[0] - 256:], # data as numpy array
                             -1,  # Max depth of the partitionning tree
                             1,   # Early Stop
                             30,  # The number of segment desired after pruning
                             0.0, # A Complexity parameter
                             1)   # Tau : Minimum number of observations within a segment

# Build the partition tree
res = my_obj.split()
# Tree pruning
my_obj.weakest_link_pruning()
# Get predictions
preds = my_obj.get_predictions()
# Get segments durations
durations = my_obj.get_durations()
print(durations)

# Plot raw data alongside with prediction
plt.plot(np.arange(np.array(data['date'])[data.shape[0] - 256:].shape[0]),
             np.array(data['value'])[data.shape[0] - 256:], 'k',
             np.arange(np.array(data['date'])[data.shape[0] - 256:].shape[0]),
             preds, 'ro')
plt.show()

Object Partition_tree parameters

  • signal : The timeseries to be segmented. It should be a numpy array of shape = [n_samples]
  • max_depth : The maximum depth of the partitioning tree. If -1, then no depth constraint exists on the tree
  • early_stop : Early Stop. If 1, then stop splitting a node if no MSE improvment is found. Otherwise, the best split is performed, even if it induces a MSE increase.
  • nb_segments : The number of segments desired when performing the Weakest Link Pruning
  • delta_complexity : A complexity parameter. Only perform the best split if np.abs(MSE_CurrentNode - min(MSE_LeftChild + MSE_RightChild)) / MSE_CurrentNode <= delta_complexity Which in human language reads as : "only perform the best split if it decreases the MSE by more than delta_complexity percentage"
  • tau : The minimum number of observations within a segment. If the current segment has less than 2*tau observation, we do not split. Otherwise, we split in two segments, both of which of duration greater than tau

Still to be done

  • Change the call to np.append() in
    • Partition_node.get_predictions()
    • Partition_node.get_durations()