Mid Sweden University

Temporal data analysis and mining has attracted substantial interest due to theproliferation and ubiquity of time series in many fields. Time series clustering isone of the most popular mining methods, and many time series clustering algorithmsprimarily focus on detecting the clusters in a batch fashion that will use alot of memory space and thus limit the scalability and capability for large timeseries.The BIRCH algorithm has been proven to scale well to large datasets,which is characterized by an incrementally clustering data objects using a singlescan. However the Euclidean distance metric employed in BIRCH has beenproven to not be accurate for time series and will degrade the accuracy performance.To overcome this drawback, this work proposes an extended BIRCH algorithmfor large time series. The BIRCH clustering algorithm is extended bychanging the cluster feature vector to the proposed modified cluster feature, replacingthe original Euclidean distance measure with dynamic time warping andemploying DTW barycenter averaging method as the centroid computation approach,which is more suitable for time-series clustering than any other averagingmethods. To demonstrate the effectiveness of the proposed algorithm, weconducted an extensive evaluation of our algorithm against BIRCH, k-meansand their variants with combinations of competitive distance measures. Experimentalresults show that the extended BIRCH algorithm improves the accuracysignificantly compared to the BIRCH algorithm and its variants, and achievescompetitive and similar accuracy as k-means and its variant, k-DBA. However,unlike k-means and k-DBA, the extended BIRCH algorithm maintains the abilityof incrementally handling continuous incoming data objects, which is thekey to cluster large time-series datasets. Finally the extended BIRCH-based algorithmis applied to solve a subsequence time-series clustering task of a simulationmulti-variate time-series dataset with the help of a sliding window.