Dynamic Time Warping Dissimilarity Research Articles

Fuzzy clustering with spatial–temporal information

Clustering geographical units based on a set of quantitative features observed at several time occasions requires to deal with the complexity of both space and time information. In particular, one should consider (1) the spatial nature of the units to be clustered, (2) the characteristics of the space of multivariate time trajectories, and (3) the uncertainty related to the assignment of a geographical unit to a given cluster on the basis of the above complex features. This paper discusses a novel spatially constrained multivariate time series clustering for units characterized by different levels of spatial proximity. In particular, the Fuzzy Partitioning Around Medoids algorithm with Dynamic Time Warping dissimilarity measure and spatial penalization terms is applied to classify multivariate Spatial–Temporal series. The clustering method has been theoretically presented and discussed using both simulated and real data, highlighting its main features. In particular, the capability of embedding different levels of proximity among units, and the ability of considering time series with different length.

Vegetable classification in Indonesia using Dynamic Time Warping of Sentinel-1A dual polarization SAR time series

This study investigates the potential of Sentinel-1A (S1A) dual polarization SAR time series data for vegetable classification in Indonesia. Vegetables are characterized by the temporal changes of observables extracted from time series of S1A data. We extracted observables regarding both backscatter (VH and VV) coefficients and decomposition features (i.e., entropy, angle, and anisotropy). The vegetable classification is based on a time-weighted Dynamic Time Warping dissimilarity measure that is calculated with SPRING strategy for subsequence searching, referred to as twDTWS. This study focuses on three main vegetable types widely planted in Indonesia, namely chili, tomato, and cucumber. We conducted vegetable classification in two areas, Malang and Lampung, using time series of S1A data covering the dry season in 2017. Our results show that the twDTWS method provides a promising means to classify vegetables using time series of S1A data for the dry season, while the features decomposed from dual polarization S1A data have little influence on the classification accuracy. Moreover, the twDTWS method with query sequences (namely reference temporal profiles) defined on the Malang dataset produced an overall accuracy of 0.80 for the classification of chili and cucumber from the Lampung dataset when the query sequences correspond to the growth cycles of vegetables. The variation in the length (i.e., the number of observations) of query sequences can affect the classification accuracy. We conclude that the twDTWS method has a high potential for classifying vegetables in different areas when constructing the query sequences of vegetables based on their growth cycles.

Beyond Features for Recognition: Human-Readable Measures to Understand Users’ Whole-Body Gesture Performance

ABSTRACTUnderstanding users’ whole-body gesture performance quantitatively requires numerical gesture descriptors or features. However, the vast majority of gesture features that have been proposed in the literature were specifically designed for machines to recognize gestures accurately, which makes those features exclusively machine-readable. The complexity of such features makes it difficult for user interface designers, non-experts in machine learning, to understand and use them effectively (see, for instance, the Hu moment statistics or the Histogram of Gradients features), which reduces considerably designers’ available options to describe users’ whole-body gesture performance with legible and easily interpretable numerical measures. To address this problem, we introduce in this work a set of 17 measures that user interface practitioners can readily employ to characterize users’ whole-body gesture performance with human-readable concepts, such as area, volume, or quantity. Our measures describe (1) spatial characteristics of body movement, (2) kinematic performance, and (3) body posture appearance for whole-body gestures. We evaluate our measures on a public dataset composed of 5,654 gestures collected from 30 participants, for which we report several gesture findings, e.g., participants performed body gestures in an average volume of space of 1.0 m3, with an average amount of hands movement of 14.6 m, and a maximum body posture diffusion of 5.8 m. We show the relationship between our gesture measures and recognition rates delivered by a template-based Nearest-Neighbor whole-body gesture classifier implementing the Dynamic Time Warping dissimilarity function. We also release BOGArT, the Body Gesture Analysis Toolkit, that automatically computes our measures. This work will empower researchers and practitioners with new numerical tools to reach a better understanding of how users perform whole-body gestures and thus, to use this knowledge to inform improved designs of whole-body gesture user interfaces.

Modified multidimensional scaling approach to analyze financial markets.

Detrended cross-correlation coefficient (σDCCA) and dynamic time warping (DTW) are introduced as the dissimilarity measures, respectively, while multidimensional scaling (MDS) is employed to translate the dissimilarities between daily price returns of 24 stock markets. We first propose MDS based on σDCCA dissimilarity and MDS based on DTW dissimilarity creatively, while MDS based on Euclidean dissimilarity is also employed to provide a reference for comparisons. We apply these methods in order to further visualize the clustering between stock markets. Moreover, we decide to confront MDS with an alternative visualization method, "Unweighed Average" clustering method, for comparison. The MDS analysis and "Unweighed Average" clustering method are employed based on the same dissimilarity. Through the results, we find that MDS gives us a more intuitive mapping for observing stable or emerging clusters of stock markets with similar behavior, while the MDS analysis based on σDCCA dissimilarity can provide more clear, detailed, and accurate information on the classification of the stock markets than the MDS analysis based on Euclidean dissimilarity. The MDS analysis based on DTW dissimilarity indicates more knowledge about the correlations between stock markets particularly and interestingly. Meanwhile, it reflects more abundant results on the clustering of stock markets and is much more intensive than the MDS analysis based on Euclidean dissimilarity. In addition, the graphs, originated from applying MDS methods based on σDCCA dissimilarity and DTW dissimilarity, may also guide the construction of multivariate econometric models.

On the verification of triangle inequality by dynamic time-warping dissimilarity measures

The verification of the Triangle Inequality (TI) by Dynamic Time Warping (DTW) Dissimilarity Measures (DM) seems to be a fairly important prerequisite for the application of some new and promising methods recently proposed for reducing the number of DTW computations in Isolated Word Recognition. The degree of satisfaction of this property for real speech samples has been empirically studied by various authors, who have reported rather controversial results for the different DTW and DMs speech data utilized by each of them. Therefore, a systematic study seemed to be necessary of the impact of the different factors upon which the DTW-based DMs depend. This paper is concerned with such work. Throughout the extensive experiments carried out, it was found that the TI never failed for any of the many combinations of speech parameters. DTW productions, local metrics and speech data taken into account; moreover, this property was always observed as being “loosely” satisfied. In view of these results, a prospective study of some possible causes of TI violation was carried out. The conclusions suggest the use of time-compressing preprocessing techniques and the application of suboptimal DTW procedures (and especially if gross end-point detection errors are involved) as the most likely causes of the TI dissatisfaction rates reported elsewere.