SoM-TP architecture. An aggregated output ofall pooling P is passed to the attention block to calculate the attention score A. In the attention block, a weighted pooling output M is formed by multiplication ofP and a learnable weight vector Ao. After M passes through the convolutional layer 00, the attention score A is drawn out as an encoded weight vector. By an index ofthe highest attention scor

Detailed experimental settings.

Temporal Pooling Performances. The best performances that SoM-TP beat others bolded and the best performances of other temporal poolings are underlined.

Table presents a comparison of models that leverage temporal information, models that exploit scale-invariant properties, and SoM-TP. Firstly, in comparison to state-of-the-art models utilizing temporal information, SoM-TP exhibited the highest performance. Furthermore, when compared to scale-invariant models, SoM-TP consistently outperformed them in terms of overall performance for univariate dat

Comparison ofInput attribution results between fixed and diverse perspective learning. Thefigure results are from three different datasets in the UCR repository; CricketZ, Fungi, and WordSynonyms by each respective row in order. The accuracy results are highly dependent on which pooling 1S used and each pooling shows different attributions result with different perspectives Note that the circles a

Decision boundary of temporal poolings on ToeSegmentationl dataset in UCR repository. To clarify that SoM-TP is a novel classification pooling method, we examine whether pooled vectors of SoM-TP are better clustered than the other temporal poolings. We use t-SNE [54 to investigate decision boundaries. While other temporal pooling vectors are not clearly clustered by classes, SoM-TP's decision boun

Ensemble performance on FCN-MAX.

The \ ablation study for SoM-TP. I is the decay value of the perspective loss, which is one of the most important hyper-parameter in SoM-TP. The ablation study is done with X along the range of [1, 1e-5] with 11 intervals. The blue and red line represents the used pooling operation type, respectively MAX and AVG. And the optimal x with the highest performance is circled in green.

Convolutional stack architecture. We use FCN and ResNet which are specially designed for TSC [57]. Here, the embedding dimensions of each convolutional layer before the pooling and FC layers for the classification decision are drawn.

Performance graph on UCR and UEA by accuracy; FCN and ResNet architec- tures with MAX type pooling. The area under the diagonal line has more datasets, which means SoM-TP is better than the other poolings. However, high pooling complexity resulting from selection- ensemble leads to performance degradation in a few simple TSC datasets. Black dots are independent datasets in the repository and the X

Pooling classification model architecture. The input attribution of best tempora. pooling by fixed-perspective learning is first calculated with LRP: 2+ rule for[이 and 6rule forf11+1:12 Then multivariate input with raw time series and input attribution goes through the model to classify the best pooling methods for each dataset. In this network, the simple global pooling layer is used.

Attention analysis. DPL attention plays a critical role on SoM-TP by selecting appro- priate pooling on each data type. Thefigure shows how DPL attention is trained and formulated: Adiac and OSULeafin UCR repository. First, to consider dataset characteristics, attention weight Ao cumu- latively trained by every batch. We can see that Ao highlights the important feature in one temporal pooling alon

SoM-TP on different time lengths and data sizes. With different datasets which contain different data sizes and time lengths, the SoM-TP works robustly and dynamically selects appro- priate pooling both duringtraining and inference procedures. With the first column, the x-axis, we can see the different time lengths ofeach dataset. Also, with the same 20 epoch and 8 batch size, the batch number is

SoM-TP on different time lengths, data sizes, and dimensions. With different datasets which contain different dimensions in the UEA repository, the SoM-TP works robustly and dynamically selects appropriate pooling both at training and inference procedures. Here, the multivariate datasets with different dimensions and data sizes are shown and the detailed selection of pooling by each data with SoM-

ECG and EEG samples. First, ECG is univariate time series data with 5 classes as shown in (a) with the x-axis oftime and the y-axis of values. Second, EEG is multivariate time series data with 2 classes, non-alcoholic VS alcoholic, as shown in (b) with the x-axis of64 channels and the y-axis of values.

Detailed data description and experimental settings.

Model performances. The best performance of SoM-TP is bolded and the best perfor mances ofother temporal poolings are underlined.