Semi-Supervised Time Series Classification & DTW-D REPORTED BY WANG YAWEN

Outline Semi-Supervised Time Series Classification(KDD_06) ◦Background ◦Time Series Classification ◦Semi-Supervised Time Series Classification ◦Empirical Evaluation DTW-D: Time Series Semi-Supervised Learning from a single sample(KDD_13) ◦Introduction ◦DTW-D ◦Two Key Assumptions ◦Algorithm ◦Experiment

Semi-Supervised Time Series Classification LI WEI, EAMONN KEOGH DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING UNIVERSITY OF CALIFORNIA, RIVERSIDE {WLI,

Semi-Supervised Time Series Classification Background ◦Labeled training data  difficult or expensive to obtain ◦Use the value of unlabeled data

Semi-Supervised Time Series Classification Background ◦Learning from both labeled and unlabeled data  SSL ◦Semi-supervised technique for building time series classifiers that take advantage of the large collections of unlabeled data ◦Only a handful of labeled examples to construct accurate classifiers ◦Note: the usefulness of unlabeled data depends on the critical assumption that the underlying models / features / kernels / similarity functions match well with the problem at hand

Semi-Supervised Time Series Classification Time Series Classification

Semi-Supervised Time Series Classification Time Series Classification

Semi-Supervised Time Series Classification Time Series Classification under realistic situation ◦It is typically not the case that we have two or more well defined classes ◦Positive class with some structure ◦Negative examples have little or no common structure ◦Can not in general assume that subsequences not belonging to the positive class look similar to each other ◦It is typically the cade that positive labeled examples are rare, but unlabeled data is abundant ◦  building binary time series classifiers for extremely imbalanced class distributions, with only a small number of labeled examples from the positive class

Semi-Supervised Time Series Classification ◦One-nearest-neighbor with Euclidean distance classifier

Semi-Supervised Time Series Classification ◦Training the classifier(train itself) ◦Step 1: train on the initial training set, where all labeled instances are positive and all unlabeled instances are regarded as negative ◦Note: the size of the training set never changes during the training process, but the labeled set is augmented ◦Step 2: classify the unlabeled data in the training set ◦Step 3: among all the unlabeled instances, the one we can most confidently classify as positive is the instance which is closet to the labeled positive examples, which will be added into the positive set. Go back to Step 1. ◦Look BackLook Back

Semi-Supervised Time Series Classification

◦Stooping criterion ◦Precision-recall breakeven point ◦Distance between the closest pair in the labeled positive set ◦Decrease gradually: space gets denser ◦Stabilizing phase: incorporate closest pair ◦Drop: negative example add

Semi-Supervised Time Series Classification Empirical Evaluation ◦Precision-recall breakeven point ◦The value at which precision and recall are equal

Semi-Supervised Time Series Classification Empirical Evaluation ◦ECG dataset

Semi-Supervised Time Series Classification Empirical Evaluation ◦Word Spotting dataset

Semi-Supervised Time Series Classification Empirical Evaluation ◦Gun dataset

Semi-Supervised Time Series Classification Empirical Evaluation ◦Wafer dataset

Semi-Supervised Time Series Classification Empirical Evaluation ◦Yoga dataset

DTW-D: Time Series Semi-Supervised Learning from a Single Example YANPING CHEN, BING HU, EAMONN KEOGH, GUSTAVO E.A.P.A BATISTA1 UNIVERSITY OF CALIFORNIA, RIVERSIDE 1UNIVERSIDADE DE SÃO PAULO-USP {YCHEN053, BHU002,

DTW-D Introduction ◦Unlabeled members of a circumscribed positive class may be closer to some unlabeled members of a diverse negative class than to the labeled positive class

DTW-D ED is an upper bound to DTW

DTW-D

Why did DTW not solve the problem? ◦There are other differences between P1 and U2, including the fact that the first and last peaks have different heights, DTW cannot mitigate this ◦Simple shape tend to be close to everything ◦Smooth, flat or least very slowly changing time series tend to be surprisingly close to other objects ◦Observation: if a class is characterized by the existence of intra-class warping(possibly among other distortions), then we should expect that moving from ED to DTW reduces distances more for intra-class comparisons than interclass caparisons.

DTW-D

Two Key Assumptions ◦Assumption 1: the positive class(the target concept) contains time warped versions of some platonic ideal(some prototypical shape), possibly with other types of noise/distortions. ◦Assumption 2: the negative class may be very diverse, and occasionally by chance produces objects close to a member of the positive class, even under DTW.

DTW-D Assumption 1 is mitigated by large amounts of labeled data ◦Our noted weakness of semi-supervised learning happens when the nearest instance to a labeled positive exemplar is a negative instance. With more labeled positive instances this becomes less and less likely to happen.

DTW-D Assumption 2 is compounded by a large negative dataset ◦If the negative class is random and/or diverse, then the larger the negative class is, the more likely it is that it will produce an instance that just happens to be close to a labeled positive item

DTW-D Assumption 2 is compounded by low complexity negative data ◦A complex time series is one that is not well approximated by few DFT coefficients or by a low degree polynomial. ◦Low complexity data  close to everything

DTW-D Algorithm Details ◦One-class classifier with no training examples from the negative class

DTW-D Why is DTW-D better? ◦Training process(choosing) ◦DTW-D selects better labeled objects than DTW ◦Evaluation process(classification) ◦DTW-D is better at selecting the top K nearest neighbors during evaluation process

DTW-D Comparison to rival methods ◦We favor our rival approaches by offering them with fifty more initial labeled examples ◦We start with a single labeled example

DTW-D Experiment ◦Learning dataset ◦Labeled dataset P: a single positive example ◦Unlabeled dataset U: the rest of objects in the learning dataset ◦Holdout dataset ◦Test the accuracy of the learned classifier

DTW-D Experiment ◦Insect Wingbeat Sound Detection

DTW-D Experiment ◦Historical Manuscript Mining

DTW-D Experiment ◦Activity Recognition

Conclusion Semi-Supervised learning framework when only a small set of labeled examples is available Simple idea that dramatically improves the quality of SSL in time series domain. Future work ◦Revisit the stopping criteria issue in light of DTW-D ◦Consider other avenues where DTW-D may be useful