More Like this

1. Benchmarking Deep Learning Interpretability in Time Series Predictions

   Ismail, Aya; Gunady, Mohamed; Bravo, Héctor; Feizi, Soheil (January 2020, Advances in Neural Information Processing Systems Foundation (NeurIPS))

   null (Ed.)

   Saliency methods are used extensively to highlight the importance of input features in model predictions. These methods are mostly used in vision and language tasks, and their applications to time series data is relatively unexplored. In this paper, we set out to extensively compare the performance of various saliency-based interpretability methods across diverse neural architectures, including Recurrent Neural Network, Temporal Convolutional Networks, and Transformers in a new benchmark of synthetic time series data. We propose and report multiple metrics to empirically evaluate the performance of saliency methods for detecting feature importance over time using both precision (i.e., whether identified features contain meaningful signals) and recall (i.e., the number of features with signal identified as important). Through several experiments, we show that (i) in general, network architectures and saliency methods fail to reliably and accurately identify feature importance over time in time series data, (ii) this failure is mainly due to the conflation of time and feature domains, and (iii) the quality of saliency maps can be improved substantially by using our proposed two-step temporal saliency rescaling (TSR) approach that first calculates the importance of each time step before calculating the importance of each feature at a time step. 

   more » « less
2. A Windowed Temporal Saliency Rescaling Method for Interpreting Time Series Deep Learning Models

   Islam, Md Khairul; Fox, Judy (March 2025, AAAI'25 Workshop: AI for Time Series Analysis: Theory, Algorithms, and Applications (AI4TS))

   Song, Dongjin; Xie, Yao; Purushotham, Sanjay; Chen, Haifeng; Shen, Cong (Ed.)

   Interpreting complex time series forecasting models is challenging due to the temporal dependencies between time steps and the dynamic relevance of input features over time. Existing interpretation methods are limited by focusing mostly on classification tasks, evaluating using custom baseline models instead of the latest time series models, using simple synthetic datasets, and requiring training another model. We introduce a novel interpretation method, Windowed Temporal Saliency Rescaling (WinTSR) addressing these limitations. WinTSR explicitly captures temporal dependencies among the past time steps and efficiently scales the feature importance with this time importance. We benchmark WinTSR against 10 recent interpretation techniques with 5 state-of-the-art deep-learning models of different architectures, including a time series foundation model. We use 3 real-world datasets for both time-series classification and regression. Our comprehensive analysis shows that WinTSR significantly outranks the other local interpretation methods in overall performance. Finally, we provide a novel and open-source framework to interpret the latest time series transformers and foundation models. 

   more » « less
3. Identifying Flare-indicative Photospheric Magnetic Field Parameters from Multivariate Time-series Data of Solar Active Regions

   https://doi.org/10.3847/1538-4365/ad21e4  

   Alshammari, Khaznah; Hamdi, Shah Muhammad; Filali Boubrahimi, Soukaina (March 2024, The Astrophysical Journal Supplement Series)

   Abstract Photospheric magnetic field parameters are frequently used to analyze and predict solar events. Observation of these parameters over time, i.e., representing solar events by multivariate time-series (MVTS) data, can determine relationships between magnetic field states in active regions and extreme solar events, e.g., solar flares. We can improve our understanding of these events by selecting the most relevant parameters that give the highest predictive performance. In this study, we propose a two-step incremental feature selection method for MVTS data using a deep-learning model based on long short-term memory (LSTM) networks. First, each MVTS feature (magnetic field parameter) is evaluated individually by a univariate sequence classifier utilizing an LSTM network. Then, the top performing features are combined to produce input for an LSTM-based multivariate sequence classifier. Finally, we tested the discrimination ability of the selected features by training downstream classifiers, e.g., Minimally Random Convolutional Kernel Transform and support vector machine. We performed our experiments using a benchmark data set for flare prediction known as Space Weather Analytics for Solar Flares. We compared our proposed method with three other baseline feature selection methods and demonstrated that our method selects more discriminatory features compared to other methods. Due to the imbalanced nature of the data, primarily caused by the rarity of minority flare classes (e.g., the X and M classes), we used the true skill statistic as the evaluation metric. Finally, we reported the set of photospheric magnetic field parameters that give the highest discrimination performance in predicting flare classes. 

   more » « less
4. A New Attention Mechanism to Classify Multivariate Time Series

   https://doi.org/10.24963/ijcai.2020/277  

   Hao, Yifan; Cao, Huiping (July 2020, Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence)

   Classifying multivariate time series (MTS), which record the values of multiple variables over a continuous period of time, has gained a lot of attention. However, existing techniques suffer from two major issues. First, the long-range dependencies of the time-series sequences are not well captured. Second, the interactions of multiple variables are generally not represented in features. To address these aforementioned issues, we propose a novel Cross Attention Stabilized Fully Convolutional Neural Network (CA-SFCN) to classify MTS data. First, we introduce a temporal attention mechanism to extract long- and short-term memories across all time steps. Second, variable attention is designed to select relevant variables at each time step. CA-SFCN is compared with 16 approaches using 14 different MTS datasets. The extensive experimental results show that the CA-SFCN outperforms state-of-the-art classification methods, and the cross attention mechanism achieves better performance than other attention mechanisms. 

   more » « less
5. Recovering the Graph Underlying Networked Dynamical Systems under Partial Observability: A Deep Learning Approach

   https://doi.org/10.1609/aaai.v37i7.26085  

   Machado, Sérgio; Sridhar, Anirudh; Gil, Paulo; Henriques, Jorge; Moura, José_M F; Santos, Augusto (June 2023, Proceedings of the AAAI Conference on Artificial Intelligence)

   We study the problem of graph structure identification, i.e., of recovering the graph of dependencies among time series. We model these time series data as components of the state of linear stochastic networked dynamical systems. We assume partial observability, where the state evolution of only a subset of nodes comprising the network is observed. We propose a new feature-based paradigm: to each pair of nodes, we compute a feature vector from the observed time series. We prove that these features are linearly separable, i.e., there exists a hyperplane that separates the cluster of features associated with connected pairs of nodes from those of disconnected pairs. This renders the features amenable to train a variety of classifiers to perform causal inference. In particular, we use these features to train Convolutional Neural Networks (CNNs). The resulting causal inference mechanism outperforms state-of-the-art counterparts w.r.t. sample-complexity. The trained CNNs generalize well over structurally distinct networks (dense or sparse) and noise-level profiles. Remarkably, they also generalize well to real-world networks while trained over a synthetic network -- namely, a particular realization of a random graph. 

   more » « less