AQOURSNet: Time2Graph Rework

Ziyuan Chen, Zhirong Chen | July 2022
Summer Research @ Yang Yang Lab, Zhejiang University

This work is protected under the MIT License.
Copyright (c) 2022 Ziyuan Chen & Zhirong Chen unless otherwise noted.

Running the Program

$ pip install -r Requirements.txt
$ python main.py ucr_dataset/dataset [--argument ARGUMENT]

For instance,

$ python main.py ucr_dataset/Strawberry --smpratio 1 --tail mlp --lr 5e-4 --dtw --kmedians --amp

Possible arguments are presented below.

Category	Argument	Description	Default
Dataset	dataset	Name of dataset	Required
	--seed	Random seed	42
	--device	Device to use	cuda if available else cpu
Shapelet & Graph	--nshapelet	Number of shapelets to extract	30
	--nsegment	Number of segments for mapping	20
	--smpratio	Pos/Neg ratio for up/downsampling (set to 0 = disable biased sampling, forced to 0 for multi-class datasets)	0
	--maxiter	Max number of KMeans iterations	300
	--tol	Tolerance of KMeans	0.0001
	--percent	Percentile for pruning weak edges	30
GAT***	--dhidden	Hidden dimension	256
	--dembed	Embedding dimension of graph (output dimension of GAT)	64
	--nlayer	Number of layers	4
	--nhead	Number of attention heads	8
	--negslope	Negative slope of LeakyReLU	0.2
	--dropout	Dropout rate	0.5
	--tail	Type of prediction tail (One of none, linear, mlp, resnet)	linear
Training	--nepoch	Number of epochs	100
	--nbatch	Number of mini-batches	16
	--optim	Optimization algorithm for learning (See torch.optim algorithms for a list)	Adam
	--lr	Learning rate	0.001
	--wd	Weight decay	0.001
	--amp	Switch* for using Automatic Mixed Precision (Forced to False unless device is cuda)	False
	--f1	Switch* for reporting F1 score in place of loss (Forced to 0 for multi-class datasets)	False
Enhancements
	--ts2vec	Switch* for using TS2Vec**	False
	--ts2vec-dhidden	Hidden dimension of TS2Vec encoder	64
	--ts2vec-dembed	Embedding dimension of TS2Vec encoder	320
	--ts2vec-nlayer	Number of layers in TS2Vec encoder	10
	--dtw	Switch* for using Dynamic Time Warping	False
	--dtw-dist	Pointwise distance function of DTW (See scipy.spatial.distance.cdist for a list)	euclidean
	--dtw-step	Local warping step pattern of DTW (See dtw/stepPattern.py for a list)	symmetric2
	--dtw-window	Windowing function of DTW (One of none, sakoechiba, itakura, slantedband)	none
	--kmedians	Switch* for using KMedians in place of KMeans in clustering	False

* Switches have action='store_true': their presence means True, and absence means False.
  Usage like ... --switch True or ... --switch False would result in a parsing error.

** The condensed ts2vec.py (--ts2vec options) has not been thoroughly tested. Use with caution.
  In case it fails, delete ts2vec.py, and clone yuezhihan/ts2vec under the same folder.

*** For fine-tuned GAT and out-of-the-box TS2Vec, refer to rong-hash/Time2GraphRework.

Model Pipeline

0. Data preparation
   • read_dataset
1. Time series ---extract---> Shapelets
   • extract_shapelets (Wrapper)
   • kmeans
   • TS2VEC (Enhancement)
2. Time series & shapelets ---embed---> Series embedding
   • embed_series
   • dtw (Enhancement)
3. Series embedding ---construct---> Graph
   • adjacency_matrix
4. Graph ---embed---> Graph embedding
   • GraphDataset (Wrapper)
   • graph_dataloader (Wrapper)
   • NeuralNetwork
   • GAT
5. Graph embedding ---predict---> Predicted classes
   • MultilayerPerceptron
   • FCResidualNetwork
   • train, test (Wrapper)

Call Hierarchy

• main.py
  • utils.py (#0, #4 Wrapper)
  • construct_graph.py (#1 Wrapper, #1, #2, #3)
    • ts2vec.py (#1 Enhancement)
    • dtw (#2 Enhancement)
  • network.py (#4, #5)
    • xgboost - TO BE IMPLEMENTED
  • process.py (#4 Wrapper, #5 Wrapper)

Folder Structure

• Stanford_CS224w - a prerequisite course
  • *.py - condensed code for GNNs like GCN, GraphSAGE, and GAT, adapted from the course materials
  • dataset - datasets required by the demos (raw only)
  • deepsnap - auxiliary code from snap-stanford@GitHub. Copyright (c) 2019 DeepSNAP Team
• ref_papers - papers associated with shapelets providing essential background knowledge
  • traditional - evolution since 2009: Logical, ST, Unsupervised, FS, LS, Random, Forest, Random forest, etc.
  • neural_network - SOTA DNNs like Dynamic, Adversarial, Adv. Dynamic, ShapeNet, BSPCover, etc. (Review)
• cached_programs - historical versions and experiments of KMeans, SVM, MLP, ResNet, Time2Vec, hierarchies, etc.
  • WARNING: Codes in the cache are not optimized for environment compatibility and may not run properly.
• affiliated_licenses - LICENSEs for code segments from yuezhihan, subhadarship, DTW, and pyg-team.
• ucr_dataset - a neatly formatted version of the UCR Dataset in compressed .npz
  • The numpy arrays contained in each file have keys train_data, train_label, test_data, test_label
  • *_data has shape (num_samples, num_features), *_label has shape (num_samples,)
• presentation - presentation materials including slides and diagrams

Credits

• Original Time2Graph model by Cheng et al., 2020 & Cheng et al., 2021, code inspired by petecheng@GitHub
• Time2Vec algorithm (essence of AQOURS) by Kazemi et al., 2019, code adapted from yuezhihan@GitHub
  • Copyright (c) 2022 Zhihan Yue
• KMeans acceleration on PyTorch by subhadarship@GitHub, with adaptations
  • Copyright (c) 2020 subhadarshi
• DTW algorithm implemented by DynamicTimeWarping@GitHub
  • Copyright (c) 2019 Toni Giorgino
• GAT structure by Veličković et al., 2017, code inspired by Stanford CS224w, pyg-team, DGraphXinye@GitHub
  • Copyright (c) 2021 Matthias Fey, Jiaxuan You (pyg-team)
• MLP & ResNet (prediction tail) structures and hyperparameters inspired by Wang et al., 2017
• PyTorch implementation of F1 Score originally written by Michal Haltuf on Kaggle, adapted by SuperShinyEyes@GitHub Gist
• Data from the UCR Time Series Classification Archive, as documented in Dau et al., 2018
• Python modules: torch, torch_geometric, kmeans, dtw, xgboost

Easter Egg

Some of you may wonder where the name "AQOURSNet" actually comes from……