LSTD: Disentangling Long-Short Term State Under Unknown Interventions for Online Time Series Forecasting (AAAI-25)

🚩News(Dec 26, 2024): After the meeting, we will upload this paper to arXiv.

Motivation

In the industry, since time series data often arrives sequentially and is accompanied by temporal distribution shifts. We observe that nonstationarity is brought by the unknown interventions on short-term states. Moreover, to address the online forecasting task, it is intuitive to find that we should disentangle the long/short-term states from the time series with unknown interventions as shown in Figure 1.

Figure 1. Illustration of sequentially arriving exchange rate data, which is influenced by short-term customs duties and long- term financial revenue. The short-term customs duties are intervened by sudden customs tariff policies.

Model

• To preserve the long-term dependencies in the long-term latent variables, we propose the smooth constraint. $A_{z_h^s}$ and $A_{z_e^s}$ denote the association matrices of the start half and the end half segments, hence we can restrict the long-term dependencies by restricting the similarity of these two matrices.
• We propose the interrupted dependency constraint for the short-term variables. Since the nonstationarity is assumed to be led by the interventions to the short-term latent variables, given $z_{1:H}^d$, if intervention occurs at $\tau$-th time step, and $2<\tau<H-1$, then $\frac{\partial \varepsilon_{H, i}^d}{\partial z_{\tau-1,j}^d}=0$, where $i,j \in { 1,\cdots,n_d}$, hence we can restrict the short-term dependencies by restricting the sparsity of the matrice.
• Our model overview is as shown in Figure 2.

Figure 2. The framework of the proposed LSTD model. The long/short-term latent variables $z_{1:L}^d$ and $z_{1:L}^s$ are extracted from the encoder. And the latent transition module is used to estimated the $z_{L+1:H}^d$ and the $z_{L+1:H}^s$ from $z_{1:L}^d$ and $z_{1:L}^s$, respectively. The long-term and short-term prior networks are used to estimate the prior distributions.

Requirements

• Python 3.8
• torch == 2.3.1
• numpy == 1.23.5
• pandas == 1.5.3
• einops == 0.4.0
• tqdm == 4.64.1

Dependencies can be installed using the following command:

pip install -r requirements.txt

Data

We have already put the datasets in the .\LSTD-main\data\ file.You just should unzip datasets1 and unzip datasets2 and it can be used directly.

Reproducibility

To easily reproduce the results you can run the following commands:

python run_LSTD.py -seed $seed -dataset $dataset -len $1/24/60

Multiple seeds and datasets can be run at one time.The important parameters are in file LSTD_config.py and you can go inside to change the parameters you want.

And we provide explanations for the important parameters:

Parameter name	Description of parameter
data	The dataset name
root_path	The root path of the data file (defaults to ./data/ETT/)
data_path	The data file name (defaults to ETTh2.csv)
features	The forecasting task (defaults to M). This can be set to M,S,MS (M : multivariate predict multivariate, S : univariate predict univariate, MS : multivariate predict univariate)
seq_len	Input sequence length of LSTD encoder (defaults to 60)
label_len	Start token length of LSTD decoder (defaults to 0)
pred_len	Prediction sequence length (defaults to 1)
des	exp description
itr	experiments times
test_bsz	Batch size in test
train_epochs	Epochs in train
online_learning	It is online learning or not
L1_weight	The weights of L1 regularization.
L2_weight	The weights of L2 regularization.
dropout	The magnitude of dropout.
zd_kl_weight	The weight of the prior for the short-term effect.
zs_kl_weight	The weight of the prior for the long-term effect.

More parameter information please refer to main.py.

Results

The main results are shown in table 1.

Citation

If you find this repository useful in your research, please consider citing the following papers:

To be continued...