Extended Deep Adaptive Input Normalization for Preprocessing Time Series Data for Neural Networks

Data preprocessing is a crucial part of any machine learning pipeline, and it can have a significant impact on both performance and training efficiency. This is especially evident when using deep neural networks for time series prediction and classification: real-world time series data often exhibit irregularities such as multi-modality, skewness and outliers, and the model performance can degrade rapidly if these characteristics are not adequately addressed. In our work, we propose the EDAIN (Extended Deep Adaptive Input Normalization) layer, a novel adaptive neural layer that learns how to appropriately normalize irregular time series data for a given task in an end-to-end fashion, instead of using a fixed normalization scheme. This is achieved by optimizing its unknown parameters simultaneously with the deep neural network using back-propagation. Our experiments, conducted using synthetic data, a credit default prediction dataset, and a large-scale limit order book benchmark dataset, demonstrate the superior performance of the EDAIN layer when compared to conventional normalization methods and existing adaptive time series preprocessing layers.

In this repository we provide an implementation of the Extended Deep Adaptive Input Normalization (EDAIN) layer using PyTorch. We also provide all the necessary code and instructions for reproducing the results in our paper. The associated thesis can be found here and our paper is available at arXiv.

Reproducing the results
Examples
- Using the EDAIN preprocessing layer
- Generating synthetic multivariate time-series data
Known issues
License information

Reproducing the results

First, make sure all the python libraries listed in requirements.txt are installed. Note that the cudf and cupy are only needed to setup the American Express default prediction dataset.

Setting up the Amex dataset

Download raddar's deanonymized dataset as well as the raw dataset:

kaggle datasets download -d raddar/amex-data-integer-dtypes-parquet-format
kaggle competitions download -c amex-default-prediction

The dataset directory should look like:

data_dir/
  * raw/
    * train_labels.csv
    * train_data.csv
    * test_data.csv
    * sample_submission.csv
  * derived/
    * train.parquet
    * test.parquet
    * processed-splits/

Run the following to split the datasets into suitable partitions

from src.lib import initial_preprocessing

initial_preprocessing.compress_csvs_to_feather(DATA_DIR)

test_path = os.path.join(DATA_DIR, 'derived', 'test.parquet')
train_path = os.path.join(DATA_DIR, 'derived', 'train.parquet')
save_path = os.path.join(DATA_DIR, 'derived', 'processed-splits')

initial_preprocessing.split_raddars_parquet(test_path, save_path, num_splits=20)
initial_preprocessing.split_raddars_parquet(train_path, save_path, num_splits=10)

Change the train_split_data_dir key in the experiment config (src/experiments/configs/) to point to the directory with the processed splits generated in step 3

Setting up the FI-2010 LOB dataset

Download the preprocessed data from here (courtesy of passalis's DAIN repo)
Change the preprocessed_lob_path in the experiment config (src/experiments/configs/) to point to the lob.h5 file downloaded

Running the experiments

Make sure config.yaml has been updated to match your filesystem, and add the project root to your python path with export PYTHONPATH=$(pwd).

Default prediction dataset experiments

To generate the history files for the different preprocessing methods on the American Express default prediction, run the following commands from the project root:

No preprocessing: python3 src/experiments/run_experiments.py --experiment_config=src/experiments/configs/experiment-config-alpha.yaml --device=0 --dataset=amex --model=gru-rnn --preprocessing_method=identity --num_cross_validation_folds=5 --random_state=42 --experiment_name=no-preprocess-amex-RECENT
$z$-score scaling: python3 src/experiments/run_experiments.py --experiment_config=src/experiments/configs/experiment-config-alpha.yaml --device=0 --dataset=amex --model=gru-rnn --preprocessing_method=standard-scaler --ignore_time=True --num_cross_validation_folds=5 --random_state=42 --experiment_name=standard-scaling-no-time-1
CDF inversion: python3 src/experiments/run_experiments.py --experiment_config=src/experiments/configs/experiment-config-alpha.yaml --device=0 --dataset=amex --model=gru-rnn --preprocessing_method=cdf-invert --num_cross_validation_folds=5 --random_state=42 --experiment_name=cdf-inversion-amex
EDAIN-KL: python3 src/experiments/run_experiments.py --experiment_config=src/experiments/configs/experiment-config-alpha.yaml --device=2 --dataset=amex --model=gru-rnn --preprocessing_method=min-max --edain_kl=True --num_cross_validation_folds=5 --random_state=42 --experiment_name=amex-edain-kl-preprocessing-1 --override='edain_bijector_fit:scale_lr:10.0 edain_bijector_fit:shift_lr:10.0 edain_bijector_fit:outlier_lr:100.0 edain_bijector_fit:power_lr:0.0000001'
EDAIN (global-aware): python3 src/experiments/run_experiments.py --experiment_config=src/experiments/configs/experiment-config-alpha.yaml --device=0 --dataset=amex --model=gru-rnn --adaptive_layer=edain --preprocessing_method=standard-scaler --num_cross_validation_folds=5 --random_state=42 --experiment_name=edain-preprocessing-1
EDAIN (local-aware): python3 src/experiments/run_experiments.py --experiment_config=src/experiments/configs/experiment-config-alpha.yaml --device=0 --dataset=amex --model=gru-rnn --adaptive_layer=edain --preprocessing_method=standard-scaler --num_cross_validation_folds=5 --random_state=42 --experiment_name=edain-local-aware-amex-RECENT --override='edain_bijector_fit:scale_lr:1.0 edain_bijector_fit:shift_lr:1.0 edain_bijector_fit:outlier_lr:10.0 edain_bijector_fit:power_lr:10.0'
DAIN: python3 src/experiments/run_experiments.py --experiment_config=src/experiments/configs/experiment-config-alpha.yaml --device=1 --dataset=amex --model=gru-rnn --adaptive_layer=dain --preprocessing_method=standard-scaler --num_cross_validation_folds=5 --random_state=42 --experiment_name=amex-dain-preprocessing-1
BIN: python3 src/experiments/run_experiments.py --experiment_config=src/experiments/configs/experiment-config-alpha.yaml --device=0 --dataset=amex --model=gru-rnn --adaptive_layer=bin --preprocessing_method=standard-scaler --num_cross_validation_folds=5 --random_state=42 --experiment_name=amex-bin-preprocessing-1

Financial forecasting dataset (FI-2010) experiments

To generate the history files for the different preprocessing methods on the American Express default prediction, run the following commands from the project root:

No preprocessing: python3 src/experiments/run_experiments.py --experiment_config=src/experiments/configs/experiment-config-beta.yaml --device=5 --dataset=lob --model=gru-rnn --preprocessing_method=identity --num_cross_validation_folds=99 --random_state=42 --experiment_name=no-preprocess-lob-RECENT
$z$-score scaling: python3 src/experiments/run_experiments.py --experiment_config=src/experiments/configs/experiment-config-beta.yaml --device=0 --dataset=lob --model=gru-rnn --preprocessing_method=standard-scaler --num_cross_validation_folds=100 --random_state=42 --experiment_name=LOB-standard-scaling-experiment-final
CDF inversion: python3 src/experiments/run_experiments.py --experiment_config=src/experiments/configs/experiment-config-beta.yaml --device=6 --dataset=lob --model=gru-rnn --preprocessing_method=cdf-invert --num_cross_validation_folds=99 --random_state=42 --experiment_name=cdf-inversion-lob-v2
BIN: python3 src/experiments/run_experiments.py --experiment_config=src/experiments/configs/experiment-config-beta.yaml --device=0 --dataset=lob --model=gru-rnn --adaptive_layer=bin --preprocessing_method=identity --num_cross_validation_folds=100 --random_state=42 --experiment_name=LOB-BIN-experiment-final
DAIN: python3 src/experiments/run_experiments.py --experiment_config=src/experiments/configs/experiment-config-beta.yaml --device=0 --dataset=lob --model=gru-rnn --adaptive_layer=dain --preprocessing_method=identity --num_cross_validation_folds=100 --random_state=42 --experiment_name=LOB-DAIN-experiment-final
EDAIN (local-aware): python3 src/experiments/run_experiments.py --experiment_config=src/experiments/configs/experiment-config-beta.yaml --device=0 --dataset=lob --model=gru-rnn --adaptive_layer=edain --preprocessing_method=identity --num_cross_validation_folds=100 --random_state=42 --experiment_name=LOB-EDAIN-experiment-final-v1 (after changing the $\beta$ config to batch_aware=True and using the learning rate modifiers for local-aware)
EDAIN (global-aware): python3 src/experiments/run_experiments.py --experiment_config=src/experiments/configs/experiment-config-beta.yaml --device=0 --dataset=lob --model=gru-rnn --adaptive_layer=edain --preprocessing_method=standard-scaler --num_cross_validation_folds=100 --random_state=42 --experiment_name=LOB-EDAIN-global-experiment-final-v1
EDAIN-KL: python3 src/experiments/run_experiments.py --experiment_config=src/experiments/configs/experiment-config-beta.yaml --device=0 --dataset=lob --model=gru-rnn --preprocessing_method=standard-scaler --edain_kl=True --num_cross_validation_folds=100 --random_state=42 --experiment_name=LOB-EDAIN-KL-experiment-final-v1

Synthetic data experiments

Run the script python3 src/experiments/misc/synthetic_data_performance_compare.py

Reproducing the tables and plots

To reproduce the plot in Figure 4, run python3 scripts/plots/amex_performance_convergence.py.
To reproduce the latex code for Table 2, run python3 scripts/plots/amex_performance_convergence.py.
To reproduce the latex code for Table 1, run python3 scripts/plots/synthetic_data_table.py
To reproduce the latex code for Table 3, run python3 scripts/plots/lob_performance_table.py.

Examples

Using the EDAIN preprocessing layer

Below is a minimal example on how to incorporate the global-aware EDAIN layer into a basic RNN sequence model for time-series binary classification. The example also includes optimising the model using sublayer-specific learning rate modifiers.

import torch
import torch.nn as nn
from src.preprocessing.normalizing_flows import EDAIN_Layer

class ExampleModel(nn.Module):
    def __init__(self, input_dim):
        super(ExampleModel, self).__init__()

        # initialise the global-aware EDAIN layer
        self.edain = EDAIN_Layer(
            input_dim=input_dim,
            # This is used by the EDAIN-KL version
            invert_bijector=False,
            # Add the skip connection to the outlier mitigation sublayer
            outlier_removal_residual_connection=True,
            # change to True to use the local-aware version of EDAIN
            batch_aware=False,
            outlier_removal_mode='exp',
        )

        self.gru_layer = nn.GRU(
            input_size=input_dim,
            hidden_size=128,
            num_layers=2,
            batch_first=True,
            dropout=0.2,
        )

        self.classifier_head = nn.Sequential(
            nn.Linear(128, 64),
            nn.ReLU(),
            nn.Linear(64, 32),
            nn.ReLU(),
            nn.Linear(32, 1),
            nn.Sigmoid(),
        )

    def forward(self, X):
        """
        Input tensor shape (N, T, input_dim)
        """
        # preprocess
        X = self.edain(X)

        # RNN layers
        h0 = torch.zeros(2, X.size(0), 128, device=X.device)#.required_grad_()
        X, _ = self.gru_layer(X, h0.detach())
        X = X[:, -1, :]

        # classifier head
        return self.classifier_head(X)

### Example of inference ###

model = ExampleModel(144)
example_input = torch.normal(0, 1, size=(1024, 13, 144)) # sequence length = 13
example_output = model(example_input)

### Training with sublayer learning rate modifiers ###

optimizer = torch.optim.Adam(
    [
        {'params' : model.gru_layer.parameters(), 'lr' : 1e-3},
        {'params' : model.classifier_head.parameters(), 'lr' : 1e-3},
    ] +
    model.edain.get_optimizer_param_list(
        base_lr=1e-3,
        # These modifiers should be further tuned for your specific dataset
        scale_lr=0.01,
        shift_lr=0.01,
        outlier_lr=100.0,
        power_lr=10.0,
    ), lr=1e-3)

# ...

Generating synthetic multivariate time-series data

Below is an example of generating a time series dataset with three predictor variables, each distributed according to some irregular PDFs:

from scipy import stats
from src.lib.synthetic_data import SyntheticData

import numpy as np

D = 3
T = 10
# lower bound, upper bound, and unormalized PDF
bounds = [(-8, 10), (-30, 30), (-1, 7)]
# The PDFs from which to generate samples
f1 = lambda x: 10 * stats.norm.cdf(10 * (x+4)) * stats.norm.pdf(x+4) + 0.1 * np.where(x > 8, np.exp(x - 8), 0) * np.where(x < 9.5, np.exp(9.5 - x), 0)
f2 = lambda x: np.where(x > np.pi, 20 * stats.norm.pdf(x-20), np.exp(x / 6) * (10 * np.sin(x) + 10))
f3 = lambda x: 2 * stats.norm.cdf(-4 * (x-4)) * stats.norm.pdf(x - 4)
# The MA theta parameters for setting the covariance structure within each time-series
thetas = np.array([
    [-1., 0.5, -0.2, 0.8],
    [-1., 0.3, 0.9, 0.0],
    [-1., 0.8, 0.3, -0.9],
])
CROSS_VAR_SIGMA = 1.4
RESPONSE_NOISE_SIGMA = 0.5
RESPONSE_BETA_SIGMA = 2.0
RANDOM_STATE = 42
NUM_DATASETS = 100
NUM_EPOCHS = 30
NUM_SAMPLES = 50000

synth_data = SyntheticData(
    dim_size=D,
    time_series_length=T,
    pdfs = [f1, f2, f3],
    ar_q = thetas.shape[1] - 1,
    ar_thetas=thetas,
    pdf_bounds=bounds,
    cross_variables_cor_init_sigma=CROSS_VAR_SIGMA,
    response_noise_sigma=RESPONSE_NOISE_SIGMA,
    response_beta_sigma=RESPONSE_BETA_SIGMA,
    random_state=RANDOM_STATE,
)

# generate a dataset
X_raw, y_raw = synth_data.generate_data(n=NUM_SAMPLES, return_uniform=False, random_state=42)
print(X_raw.shape, y_raw.shape) # (50000, 10, 3), (50000,)

Known issues

The gradients in power transformation EDAIN sublayer might produce NaNs during optimisation. This is known to occur when either the power transform sublayer-specific learning rate modifier is too high or the input values are extreme. If this is encountered, it is recommended to apply $z$-score scaling to the data before passing it to the global-aware EDAIN layer for further preprocessing.

License information

In our work, we used the following assets:

The dain preprocessing layer developed by Passalis et al. (2019). No license information is specified in their repository
The bin preprocessing layer developed by Tran et al. (2020). Their implementation is publicly released under the Apache License 2.0.
The benchmark dataset for mid-price forecasting of limit order books (FI-2010), available here. This dataset is released under the creative commons attribution 4.0.
The American Express default prediction dataset was published online for use in this Kaggle competition. We have received permission to use this dataset in our research by American Express.

marcusGH / edain_paper