This project implements the Time2Graph model[1], which focuses on time series modeling with dynamic shapelets.
This project is implemented primarily in Python 3.6, with several dependencies listed below. We have tested the whole framework on Ubuntu 16.04.5 LTS with kernel 4.4.0, and it is expected to easily build and run under a regular Unix-like system.
-
Python 3.6. Version 3.6.5 has been tested. Higher versions are expected be compatible with current implementation, while there may be syntax errors or conflicts under python 2.x.
-
DeepWalk We use a modified version of the original implementation of deepwalk to satisfy the support for directed and weighted graphs. The source codes with minor modifications can be found on weighted_deepwalk.
-
Version 0.4.1 has been tested. You can find installation instructions here. Note that the GPU support is ENCOURAGED as it greatly boosts training efficiency.
-
Version 0.80 has been tested. You can find installation instructions here.
-
Other Python modules. Some other Python module dependencies are listed in
requirements.txt
, which can be easily installed with pip:pip install -r requirements.txt
Although not all dependencies are mentioned in the installation instruction links above, you can find most of the libraries in the package repository of a regular Linux distribution.
Before building the project, we recommend switching the working directory to the project root directory. Assume the project root is at <time2graph_root>
, then run command
cd <time2graph_root>
Note that we assume <time2graph_root>
as your working directory in all the commands presented in the rest of this documentation. Then make sure that the environment variable PYTHONPATH
is properly set, by running the following command (on a Linux distribution):
export PYTHONPATH=`readlink -f ./`
A test script scripts/std_test.py
is available for reproducibility on the benchmark datasets:
python . -h
usage: . [-h] [--dataset] [--n_splits] [--model_cache] [--shapelet_cache] [--gpu_enable]
optional arguments:
-h, --help show this help message and exit
--dataset str, one of `ucr-Earthquakes`, `ucr-WormsTwoClass` and `ucr-Strawberry`,
which we have set the optimal parameters after fine-tuning.
(default: `ucr-Earthquakes`)
--n_splits int, number of splits in cross-validation. (default: 5)
--model_cache bool, whether to use a pretrained model.(default: False)
--shapelet_cache bool, whether to use a pretrained shapelets set.(default: False)
--gpu_enable bool, whether to enable GPU usage. (default: False)
To quickly and exactly reproduce the results that reported in the paper, we highly RECOMMEND that set model_cache
as True, since there are unavoidable randomness in the process of shapelets learning and graph embedding. And if only shapelet_cache
is True, it will learn a new set of shapelet embeddings, which may bring some small fluctuations on the performance. So the easiest way for reproducibility and project testing is to run the following command:
python scripts/std_test.py --model_cache --dataset *OPTION* --gpu_enable
Given a set of time series data and the corresponding labels, the Time2Graph framework aims to learn the representations of original time series, and conduct time series classifications under the setting of supervised learning.
The input time series data and labels are expected to be numpy.ndarray
:
Time_Series X:
numpy.ndarray with shape (N x L x data_size),
where N is the number of time series, L is the time series length,
and data_size is the data dimension.
Labels Y:
numpy.ndarray with shape (N x 1), with 0 as negative, and 1 as positive samples.
We organize the preprocessing codes that load the UCR dataset in the archive/
repo, and if you want to utilize the framework on other datasets, just preprocess the original data as the abovementioned format. Note that the time series data is not needed to be normalized or scaled, since you can set the parameter scaled
as True when initializing Time2Graph model.
Now that the input data is ready, the main script scripts/run.py
is a pipeline example to train and test the whole framework. Firstly you need to modify the codes in the following block (line 46-51) to load your datasets, by reassigning x_train, y_train, x_test, y_test
respectively.
if args.dataset.startswith('ucr'):
dataset = args.dataset.rstrip('\n\r').split('-')[-1]
x_train, y_train, x_test, y_test = load_usr_dataset_by_name(
fname=dataset, length=args.seg_length * args.num_segment)
else:
raise NotImplementedError()
The help information of the main script scripts/run.py
is listed as follows:
python . -h
usage: .[-h] [-- dataset] [--K] [--C] [--num_segment] [--seg_length] [--data_size]
[--n_splits] [--njobs] [--optimizer] [--alpha] [--beta] [--init]
[--gpu_enable] [--opt_metric] [--cache] [--embed] [--embed_size] [--warp]
[--cmethod] [--kernel] [--percentile] [--measurement] [--batch_size]
[--tflag] [--scaled] [--norm] [--no_global]
optional arguments:
-h, --help show this help message and exit
--dataset str, indicate which dataset to load;
need to modify the codes in line 46-51.
--K int, number of shapelets that try to learn
--C int, number of shapelet candidates used for learning shapelets
--num_segment int, number of segment that a time series have
--seg_length int, the segment length,
so the length of a time series is num_segment * seg_length
--data_size int, the dimension of time series data
--n_splits int, number of cross-validation, default 5.
--njobs int, number of threads if using multiprocessing.
--optimizer str, optimizer used for learning shapelets, default `Adam`.
--alpha float, penalty for local timing factor, default 0.1.
--beta float, penalty for global timing factor, default 0.05.
--init int, init offset for time series, default 0.
--gpu_enable bool, whether to use GPU, default False.
--opt_metric str, metric for optimizing out-classifier, default `accuracy`.
--cache bool, whether to save model cache, defualt False.
--embed str, embedding mode, one of `aggregate` and `concate`.
--embed_size int, embedding size in deepwalk, default 256.
--wrap int, warp size in greedy-dtw, default 2.
--cmethod str, candidate generation method, one of `cluster` and `greedy`
--kernel str, choice of outer-classifer, default `xgb`.
--percentile int, distance threshold (percentile) in graph construction, default 10
--measurement str, distance measurement,default `gdtw`.
--batch_size int, batch size, default 50
--tflag bool, whether to use timing factors, default True.
--scaled bool, whether to scale time seriee by z-normalize, default False.
--norm bool, whether to normalize handcraft-features, default False.
--no_global bool, whether to use global timing factor
when constructing shapelet evolution graph, default False.
Some of the arguments may require further explanation:
--K/--C
: the number of shapelets should be carefully selected, and it is highly related with intrinsic properties of the dataset. And in our extensive experiments,C
is often set 10 or 20 times ofK
to ensure that we can learn from a large pool of candidates.--percentile
,--alpha
and--beta
: we have conduct fine-tuning on several datasets, and in most cases we recommend the default settings, although modifying them may bring performance increment, as well as drop.
We include all three benchmark UCR datasets in the dataset
directory, which is a subset of UCR-Archive time series dataset. See Data Sets for more details. Then a demo script is available by calling scripts/run.py
, as the following:
python scripts/run.py --dataset ucr-Earthquakes --K 50 --C 500
--num_segment 21 --seg_length 24 --data_size 1 --embed concate --percentile 5 --gpu_enable
The three benchmark datasets reported in [1] was made public by UCR, which consists of many time series datasets. we select several UCR datasets from many candidates by the following reasons that: 1) to maintain the consistency of evaluation metrics between the real-world and public datasets, we only consider binary-label ones in UCR; 2) we have to make sure that there are enough training cases because we need sufficient samples to capture the normal transitions between shapelets (many binary-label datasets in UCR only have less than 100 training samples), and 3) we omit all datasets categorized as “image”, because the proposed intuition (timing factor, shapelet evolutions) may not be appropriate for time series transformed from images. After filtering based on the abovementioned criterion, and due to space limitation, we only present those three in [1]. We have tested some others such as Ham and Computers, etc., and also achieved competitive results compared with baseline methods.
Furthermore, we apply the proposed Time2Graph model on two real-world scenarios: Electricity Consumption Records (ECR) provided by State Grid of China, and Network Traffic Flow (NTF) from China Telecom. Detailed dataset descriptions can be found in our paper. The performance increment compared with existing models clearly demonstrate the effectiveness of the framework, and below we list the final results along with several popular baselines.
Accuracy on UCR(%) | Earthquakes | WormsTwoClass | Strawberry |
---|---|---|---|
NN-DTW | 70.31 | 68.16 | 95.53 |
TSF | 74.67 | 68.51 | 96.27 |
FS | 74.66 | 70.58 | 91.66 |
Time2Graph | 79.14 | 72.73 | 96.76 |
Performance on ECR(%) | Precision | Recall | F1 |
---|---|---|---|
NN-DTW | 15.52 | 18.15 | 16.73 |
TSF | 26.32 | 2.02 | 3.75 |
FS | 10.45 | 79.84* | 18.48 |
Time2Graph | 30.10 | 40.26 | 34.44 |
Performance on NTF(%) | Precision | Recall | F1 |
---|---|---|---|
NN-DTW | 33.20 | 43.75 | 37.75 |
TSF | 57.52 | 33.85 | 42.62 |
FS | 63.55 | 35.42 | 45.49 |
Time2Graph | 71.52 | 56.25 | 62.97 |
Please refer to our paper [1] for detailed information about the experimental settings, the description of unpublished data sets, the full results of our experiments, along with ablation and observational studies.
[1] Cheng, Z; Yang, Y; Wang, W; Hu, W; Zhuang, Y and Song, G, 2020, Time2Graph: Revisiting Time Series Modeling with Dynamic Shapelets, In AAAI, 2020
@inproceedings{cheng2020time2graph,
title = "{Time2Graph: Revisiting Time Series Modeling with Dynamic Shapelets}",
author = {{Cheng}, Z. and {Yang}, Y. and {Wang}, W. and {Hu}, W. and {Zhuang}, Y. and {Song}, G.},
booktitle={Proceedings of Association for the Advancement of Artificial Intelligence (AAAI)},
year = 2020,
}