Welcome to the documentation for the Pattern Extraction module of MORE!¶
More-Pattern-Extraction¶
A python library for timeseries Pattern Extraction and Handling.
About More-Pattern-Extraction¶
More-Pattern-Extraction is a python library for implementing utilities which aim to fulfill a full Pattern Extraction workflow for analysing timeseries data and extracting meaningful patterns in the form of motifs, discords, changepoints, etc. More-Pattern-Extraction implements a series of training features, for the input time series, running variations of motif discovery algorithms in order to identify different types of patterns, including motifs, discords and changepoints, in order to find the most fitting model for the data at hand. Further, it it encapsulates input/output utilities which assist in storing or loading the output of the aforementioned processing steps, and utilities for plotting the output or depicting intermediate computations.
Description¶
For the development process, several python libraries will be used. In particular: STUMPY, MatrixProfile and Pyscamp. The aforementioned libraries are implemented in python3 (or provide python3 bindings to C++ code), thus the Pattern Extraction module will adopt the same programming language. An installation of the CUDA toolkit8 is necessary for deploying the GPU-accelerated versions of the aforementioned libraries. The module consists of the following steps:
Input/Output
This group of utilities includes all functions which are responsible for loading and storing time series to the disk. Moreover, it includes functions for loading/storing necessary metadata which are produced by any of the methods of this module and can be further processed.
Preprocessing
Utilities regarding the handling and processing of the input data. Such functions handle filtering, resampling, adding noise, normalization and filling out missing values.
Statistics
Utilities regarding the computation of useful statistics about the dataset, that implement variations of standard error measures, evaluating the performance of a regression model, ranking time series.
Plotting
Functions that are used for the plotting of the results are presented, also include functions for plotting auxiliary information and helper functions for switching between different data formats.
Basic pattern extraction
In this section, we discuss the implementation of functions related to pattern extraction. This includes functions for computing a matrix profile, but also functions related to the post-processing of a matrix profile. The latter refers to functions for discovering motifs/discords or segmenting a time series.
Similarity
Utilities related to similarity search refer to functions which are responsible for performing proximity queries to a dataset. Queries that can be handled include finding the k-nearest neighbors of a given vector or time series. One example of such a query is computing the top k-nearest neighbors of a given subsequence in a multivariate time series, among all (or a subset of) subsequences in the same time series.
Learning
In this set of utilities, we implement functions that build upon standard machine learning algorithms. In particular, functions contained in this module, handle tasks related to approximating a time series by means of a regression model. Those functions are needed, for example, when we are trying to discover trends or changepoints.
Pattern extraction tools¶
The execution of the project starts with the presentation of some initial tools. Later those tools will become more interactive through visual modules.
Pattern Discovery
The pattern discovery tool aims to detect interesting patterns in time series with labeled time intervals. In particular, given a time series which is divided into a sequence of discrete segments, each one assigned to one of two classes, the goal is to detect patterns which correspond to distinguishable characteristics of one of the classes. The intuition is that patterns that are representative of their respective class, should mostly appear in the time regions corresponding to that class. Link to the notebook: Pattern Discovery
Time Series Segmentation
Given as input a time series, the time series segmentation tool finds points where a behavioral change is observed. Those points, partition the time series into discrete segments. The user can specify the number of changepoints/segments which will be returned by the tool. In a scenario where ground truth, labelled, time series exist, the user has the ability to calculate a cost, defined essentially as a distance function between the changepoints returned by the tool and the original changepoints (labels). This allows for optimizing the parameters inherited by the matrix profile library in use cases where labelled data are available. Link to the notebook: Time Series Segmentation
Changepoint Detection
In this section, we present a tool for detecting and recognizing changepoints in a time series. Given a set of segments of the input time series, corresponding to periods of time where a changepoint may occur, our tool essentially ranks those segments with respect to the possibility of containing a changepoint. This can be seen as a classification task, where segments are classified with respect to whether they contain a changepoint or not. Link to the notebook: ChangePoint Detection
Deviation Detection
In this section, we present a tool for detecting segments of a time series where the behavior of a given target variable deviates much from the usual. We assume that certain segments of the time series are labelled, representing the usual behavior of the variable in question. Our main ingredient is regression. We fit a regression model on the labelled segments, aiming to capture the “usual” behavior of the target variable. Then, we compare the values of the trained model with the real values. Link to the notebook: Deviation Detection
Documentation¶
Source code documentation is available from GitHub pages
