Download
1 / 30
300 likes | 670 Views
sDTW : Computing DTW Distances using Locally Relevant Constraints based on Salient Feature Alignments. K. Selçuk Candan , Maria Luisa Sapino Xiaolan Wang , Rosaria Rossini.
E N D
sDTW: Computing DTW Distances using Locally Relevant Constraints based on Salient Feature Alignments K. SelçukCandan, Maria Luisa Sapino XiaolanWang, Rosaria Rossini This work is supported by the NSF Grant 1043583 MiNC: NSDL Middleware for Network- and Context- aware Recommendations.
DynamicTimeWarping Dynamic time warping (DTW) tries to find the best mapping with the minimum distance between two sequences. The method is called ”time warping” since we need to compress or expand in time in order to find the best mapping.
DTW Grid and WarpPath Given two time series, X and Y, of length N and M, alternative warping strategies can be compactly represented as an NxM grid. A warp path between X and Y then can be visualized as a path from the lower-left corner of the NxM grid to its upper-right corner.
Reducing the costof DTW The major cost of the DTW algorithm is filling the NxM grid. To reduce cost, Sakoe and Chiba band constraints and Itakura Parallelogram define a narrow band around the diagonal from which the warp paths can pass.
sDTW Idea Our key observation is that time series often carry structural features that can be used for identifying locally relevant constraints to eliminate redundant work.
ProposedAdaptive fixed core&adaptive width (fc,aw): the width is adapted on local features; adaptive core&fixed width (ac,fw): the core is not necessarily on the diagonal; adaptive core&adaptive width (ac,aw):both the core and the width are adapted;
sDTWProcess • Search for salient features on the input time series. • Find salient alignments of a given pair of time series by matching the descriptors of the salient features. • Use these salient alignments to compute locally relevant constraints to prune the warp path search. sDTW leverages salient alignment evidences to improve the effectiveness of the pruning constraints.
First Question To search for robust local features, we adapt the scale-invariant feature transform (SIFT). How can we locate robust local features of time series?
Scale-InvariantFeatureTransform The scale-invariant feature transform (SIFT)4 is a computer vision algorithm used to detect and describe local features in 2D images. The algorithm identifies salient point of a given image (and their descriptors) that are invariant to: • Imagescaling; • Translation; • Rotation; • Different illuminations and noise. We adapted the 2D SIFT algorithm in a way that captures characteristics of 1D time series.
1D Scale-Invariant Feature Transform Steps The salient feature extraction algorithm relies on a three step process to identify such salient features and their robust local feature descriptors in time series: • Scale-spaceextrema detection; • Featuredescriptorcreation. • Featurefiltering and localization Each step is designed to capture characteristics of time series.
Featuredescriptorcreation 8 gradients for a time series describe the magnitude of the changes at different distances from the salient point. The same descriptor size would cover different temporal ranges at different time scale.
2nd Step: Features Matching & Inconsistency Pruning • the distance between the feature descriptors of each pair of salient features using Euclidean Distance; • the size of scopes between the features of the matching pairs. The Feature Matching Step is performed by compare: A good match haslessdistancedescriptors and similarsize-scopesbetweentwofeatures.
InconsistencyPruning For each pair of matching features we look on: • Alignmentfeature pairs with large temporal length and close to each other in time; • Similaritypairs of features that have both similar descriptors and similar average amplitudes; • Combined score bothallignment and similarity.
InconsistencyPruning Given two time series, consider the scope boundaries of the pairs of matching salient features in descending order of their μcomb scores and prune those that imply inconsistent ordering of start and end points.
InconsistencyPruning Pairs of matching salient points (pre-inconsistency removal) Pairs of matches remaining after inconsistency pruning. Scope boundaries of the matching salient features after inconsistency elimination
3trd StepofsDTWProcess LocallyRelevant DTW Constraints: WIDTH. A sets of matching consistent of salient features after the inconsistent pruning. Thus, our intuition is that each corresponding interval pair in the two time series corresponds to similar regions and thus must have similar characteristics.
AdaptiveWidthConstrains Adaptive width constraints use the widths of the resulting intervals to choose a different locally relevant width for each point (details in the paper).
3trd StepofsDTWProcess LocallyRelevant DTW Constraints: CORE. Sample point alignments within interval E. We can use the starting and ending points of the corresponding intervals in the two time series to associate to each point in one series a candidate point in the other time series.
AdaptiveCoreConstrains The core of the band is aligned with the diagonal; instead, the core follows a path that reflects the alignments implied by the scopes of the salient features.
Adaptive Core & Adaptive Width Adaptive Width Contraints can easely be combined with Adaptive Core Constraints.
EvaluationCriteria For estimate the efficiency and effectiveness of sDTW, it is tested on the time series data sets:
Experiment Set Up • Intel Core 2 • Quad CPU 3GHz machine, 8Gb RAM • Ubuntu 9.10(64bit) • Matlab 7.8.0 For the baseline fc&fw schemes we used the DTW code of Sakoe-Chiba.
EvaluationCriteria In order to assess the effectiveness of various DTW algorithms, we use the following measures: • RetrievalAccuracy • Distanceaccuracy • Classificationaccuracy
ExecutionTimeAnalysis The overheadof the matchingstepfor the adaptive algorithmsisnegligible.
Top-kRetrievalAccuracy fixed core&fixed width (fc,fw) ⇒ the larger is the value of w, the more accurate the results are; adaptive core&fixed width (ac,fw) ⇒ has significant gains in accuracy is obtained when the core is adapted; adaptive core&adaptive width (ac,aw) ⇒ the accuracy is further boosted when both the width and the core are adapted.
ClassificationAccuracy The figure re-confirms the results by focusing on the classification accuracy. Similarly to the top-k retrieval and DTW distance accuracy results, adaptive core, and adaptive width algorithms improve the classification accuracy.
DistanceAccuracy The figure confirms the results by focusing on the errors in the DTW estimates provided by the various algorithms for the data set.
Conclusions We recognize that the time series that are being compared often carry structural evidences that can help identify locally relevant constraints that can prune unnecessary work during dynamic time warp (DTW) distance computation. Experiment results have shown that the proposed locally-relevant constraints based on salient features help improve accuracy in DTW distance estimations. We have proposed three different constraint types based on different assumptions on the structural variations in the data set.
