SSET: Swapping-Sliding Explanation for Time Series Classifiers in Affect Detection

TL;DR

SSET is a swapping--sliding decision explanation for multivariate time series classifiers, called SSET, which measures the importance of different variables over time in a novel way characterized by multiple factors.

Abstract

Local explanation of machine learning (ML) models has recently received significant attention due to its ability to reduce ambiguities about why the models make specific decisions. Extensive efforts have been invested to address explainability for different data types, particularly images. However, the work on multivariate time series data is limited. A possible reason is that the conflation of time and other variables in time series data can cause the generated explanations to be incomprehensible to humans. In addition, some efforts on time series fall short of providing accurate explanations as they either ignore a context in the time domain or impose differentiability requirements on the ML models. Such restrictions impede their ability to provide valid explanations in real-world applications and non-differentiable ML settings. In this paper, we propose a swapping--sliding decision explanation for multivariate time series classifiers, called SSET. The proposal consists of swapping and sliding stages, by which salient sub-sequences causing significant drops in the prediction score are presented as explanations. In the former stage, the important variables are detected by swapping the series of interest with close train data from target classes. In the latter stage, the salient observations of these variables are explored by sliding a window over each time step. Additionally, the model measures the importance of different variables over time in a novel way characterized by multiple factors. We leverage SSET on affect detection domain where evaluations are performed on two real-world physiological time series datasets, WESAD and MAHNOB-HCI, and a deep convolutional classifier, CN-Waterfall. This classifier has shown superior performance to prior models to detect human affective states. Comparing SSET with several benchmarks, including LIME, integrated gradients, and Dynamask, we found..

Figures (10)

• Figure 1: SSET schema showing the three main components.
• Figure 2: Swapping stage of SSET, providing salient signals. In this stage, the black-box predicts class $c$ (shown by the red color) with the score $y_i^c$ for the series of interest $x_i$ (red star). We extract the train data with opposite classes to $c$ (colored circles), denoted as $X_{tc}$, from which a few neighbors $X_{neighbors}$ are randomly sampled (see Fig. \ref{['fig.neighborhood']}). To detect the important variables, the values of each signal $s$ in $x_i$ is swapped with its counterparts in $X_{neighbors}$, resulting in $X_{swp_s}$. The corresponding signals of those $X_{swp_s}$ that cause a performance drop maximally and below a certain threshold (orange series), are considered as salient. Otherwise, either further sampling and updating of the neighborhood scope are performed, or the DualSignals component is activated.
• Figure 3: Determining the neighboring scope. Starting from the instance to be explained (the red star), a region of radius $l$ (white space) is specified around the instance. If no neighboring samples (colored circles) fall in this space or satisfy the dropping constraint, the scope is updated by shifting the start point by $\delta$.
• Figure 4: Sliding stage of SSET, providing salient sub-sequences of the important signals. To simplify the visualization, we only show one salient signal (EDA) in $x_i$ using the orange color. Starting with a context (neighboring step) size of $1$, a window is slid over EDA, including the current, backward, and forward time steps. The corresponding values are replaced by their counterparts in $X_{swp_s}$, while the values of other time steps in $x_i$ are kept unchanged. Provided that the performance of the black-box drops in the manipulated instances, the corresponding sub-sequences are designated as salient, and their importance degree is measured for each time step. Otherwise, the context size is increased and the process is repeated iteratively.
• Figure 5: Dynamask (a), IG (b), LIME (c), and SSET (d) explanations of instance $41$ on WESAD. The explanations are in the form of importance scores, ranging between 0 and 1 and mapped to blue colors in the heatmap visualization. A darker color indicates higher importance. As can be seen, none of the models except SSET provide a meaningful explanation. The Dynamask scores are locked at $0.5$ for all features, IG assigns low scores to most features, and LIME concludes that nearly all features are unimportant. However, SSET presents the sub-sequence of ACC2 between time steps 7 and 29 as highly important in the detection of the neutral state, aligning with the domain knowledge.