Raneen Younis, Zahra Ahmadi, Abdul Hakmeh, Marco Fisichella
{younis,ahmadi,mfisichella@l3s.de}
Increasing privacy concerns have led to decentralized and federated machine learning techniques that allow individual clients to consult and train models collaboratively without sharing private information. Some of these applications, such as medical and healthcare, require the final decisions to be interpretable. One common form of data in these applications is multivariate time series, where deep neural networks, especially convolutional neural networks based approaches, have established excellent performance in their classification tasks. However, promising results and performance of deep learning models are a black box, and their decisions cannot always be guaranteed and trusted. While several approaches address the interpretability of deep learning models for multivariate time series data in a centralized environment, less effort has been made in a federated setting. In this work, we introduce FLAMES2Graph, a new horizontal federated learning framework designed to interpret the deep learning decisions of each client. FLAMES2Graph extracts and visualizes those input subsequences that are highly activated by a convolutional neural network. Besides, an evolution graph is created to capture the temporal dependencies between the extracted distinct subsequences. The federated learning clients only share this temporal evolution graph with the centralized server instead of trained model weights to create a global evolution graph. Our extensive experiments on various datasets from well-known multivariate benchmarks indicate that the FLAMES2Graph framework significantly outperforms other state-of-the-art federated methods while keeping privacy and augmenting network decision interpretation.
- The Baydogan’s archive, contains 13 multivariate time series classification datasets. In this work, we choose five datasets from this archive (UWave, ECG, AUSLAN, NetFlow, and Wafer).
- Human Activity Recognition (HAR) dataset recorded daily activities from 30 volunteers and produced six different labels of these activities (walking, walking upstairs, walking downstairs, standing, sitting, and lying).
- PAMAP2 Physical Activity Monitoring (PAM) dataset recorded 18 different daily activities using 17 sensors.
In order to simulate federated learning as it occurs in a real-world scenario, we employ NVIDIA FLARE. We set up a server with four clients for each experiment.
The python packages needed are:
- numpy
- pandas
- sklearn
- scipy
- matplotlib
- tensorflow
- keras
In the following examples, we demonstrate additional cases by taking a closer look at the interpretability and effectiveness of our proposed FLAMES2Graph framework for the ECG (2-dimensional) and UWave (3-dimensional) datasets.
This example demonstrates the MHAP evolution graph for ECG data and the temporal transitions for a Class 1 data set sample. The framework identifies the important subsequences of the multivariate time series data and represents the temporal transition through the MHAP evolution graph. Figure (a) shows the three MHAPs extracted from the class 1 sample that serve as nodes in the graph shown in (b). The transitions between the nodes show the order of occurrence of the MHAPs in the input data. To understand the decision of the network in classifying time series, our framework relies on the important subsequences of the input data (MHAPs) and their temporal order, which is crucial. For instance, when interpreting ECG data, a physician needs to examine the data in a specific temporal sequence to determine whether the patient's ECG recording is normal or has pathological problems.
Our framework is tested on two samples from the UWave dataset, as shown in this example. The first sample (a) includes three representative patterns (nodes 1, 2, and 3) that were learned from the network, while the second sample (b) consists of representative MHAP nodes 4, 5, and 2. The graph represents the relationship between these nodes based on their temporal occurrence order in (c). Notably, node 2 is present in both samples, but its temporal order is different in the two classes. Our graph representation plays a crucial role in identifying the representative patterns and their chronological order, aiding the classifier in making a specific decision.
As we analyze sample (a), it becomes apparent that the first output class label is activated in both the x and y-axis signals of the UWave dataset. This label corresponds to a gesture pattern that starts at a specific point, moves upwards, towards the left, and finally downwards, as shown by the 'output label shape' in (a). Although this information provides insight into the pattern recognition process, a domain knowledge expert can provide a more nuanced interpretation of the outputs. Our framework is designed to help the user visualize the significant input subsequences that contribute to a specific decision, thereby facilitating precise interpretation by domain experts.
If you re-use this work, please cite:
@inproceedings{younis2023flames2graph,
title={FLAMES2Graph: An Interpretable Federated Multivariate Time Series Classification Framework},
author={Younis, Raneen and Ahmadi, Zahra and Hakmeh, Abdul and Fisichella, Marco},
booktitle={Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
pages={3140--3150},
year={2023}
}