1. Figures in High Resolution

2. Hamming distance for all sliding words using Average Link Three clusters of equal diameter when K = 20 { whoid, davud, njoin, dovid, david } { ified, frien, oined, oiden, viden, vidan } { todow, todov, sonof } Three clusters of equal diameter when K = 40 { avuds, ovida, avide } { davud, dovid, david } { nofpe, nedpe, andpe } Figure 1 Idea Behind the Figure Representative partitional clusters from dataset D for two settings of K.

3. 248 50 words Adiac Beef CBF Coffee ECG200 Face All Face Four FISH Gun Point Lighting2 Lighting7 OliveOil OSULeaf Swedish Leaf Synthetic Control Trace Two Patterns wafer yoga 163264 128256512 1024204840968192 163843276865536 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 248 2 24 2 32 Classification Accuracy Classification accuracy on 18 time series datasets as a function of the data cardinality. Even if we reduce the cardinality of the data from the original 4,294,967,296 to a mere 64 (vertical bar), the accuracy does not decrease. Figure 2

4. Figure 3 050100150200250 A 0 B C D Four time series of length 250 and with a cardinality of 256. Naively all require 250 bytes to represent, but they have different description lengths.

5. Figure 4 050100150200 250 0 B H B’ which is B-H, denoted as B’ is B given H B’ = (B|H) Time series B can be represented exactly as the sum of the straight line H and the difference vector B'. We can store B in many ways: 1) keep B without any encoding, it requires 250*log 2 (256) = 2000 bits. 2) keep B using entropy coding (Huffman), it requires 250*7.29 = 1822 bits. 3) keep B by encoding with H, it requires DL(H) + DL(B│H) = DL(H) + DL(B’) = (2 *8) + (250* 2.51) = 644 bits

6. Figure 5 0.511.522.53 x 10 5 0 50100150200250300 0 50100150200250300 0 Two interwoven bird calls featuring the Elf Owl, and Pied-billed Grebe are shown in the original audio space (top), and as a time series extracted by using MFCC technique (middle), and then clustered by our algorithm (bottom). The original calls can be download from AZFO Bird Sounds Library.AZFO Bird Sounds Library

7. Figure 6 Step 1: Create a cluster from top-1 motif Step 2: Create another cluster from next motif Step 3: Add subsequence to an existing cluster Step 4: Merge 2 clusters(rejected) SubsequencesCenter/Hypothesis 1234 -4 -2 0 2 Step of the clustering process bitsave per unit Clustering stops here Create Add Merge A trace of our algorithm on the bird call data shown in Figure 5.bottom.

8. Figure 7 20040060080010001200 0 20040060080010001200 0 12345678910 Step of the clustering process bitsave per unit 0 1 2 Clustering stops here Create Add Merge top) 29.8 seconds of an audio snippet of poem “The Bells” by Edgar Allan Poe, represented by the first coefficient in MCFF space, and then annotated with colors to reflect the clusters (middle). A trace of the steps use to produce the clustering (bottom).

9. Figure 8 Dimension U1 of the Winding dataset (top), and time series encoding with color shown the clustering created by our algorithm (bottom) A trace of the clustering steps produced by our algorithm. 5001000150020002500 0 dropouts spikes 1234567891011121314 -2 0 2 Step in clustering process bitsave per unit Clustering stops here Create Add Merge Representative clusters obtained.

10. Figure 10 200400600800100012001400160018002000 1234 Step of the clustering process Bitsave per unit 0 1 2 3 Clustering stops here, because there is essentially no data left to cluster top) The same 2,000 datapoints from Koski-ECG as used in Figure 9. middle.right) A trace of the clustering steps produced by our algorithm. middle.left) the single cluster discovered has five members (stacked) bottom) all five subsequences in the cluster. Note that all subsequences are quantized to 64 cardinality.

11. 01000200030004000500060007000800090001000011000120001300014000150001600017000 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 x 10 4 Size of input time series Running time (sec) Figure 11 The running time of our algorithm on Koshi-ECG (Figure 10) data when s = 350.