Visually Mining and Monitoring Massive Time Series

1
Visually Mining and Monitoring Massive Time Series
Lin, J, Keogh, E., Lonardi, S., Lankford, J.P.
and Nystrom, D.M.In Proceedings of the 10th ACM
SIGKDD International Converence on Knowledge
Discovery and Data Mining, 2004.

• Amy Karlson
• V. Shiv Naga Prasad
• 15 February 2004
• CMSC 838S

Images courtesy of Jessica Lin and Eamonn Keogh
2
What are Time Series?

• Simply
• Observations of a variable made over time
• Typical across a wide variety of domains
• Medicine
• Physiology
• Finance
• Microbiology
• Meteorology
• Surveillance

3
Motivation Critical Decision Making

• Domains
• Spacecraft Launch
• Medicine
• Research Directions
• Mining Archives
• Extract rules, patterns, regularities
• Visualizing Streams
• Novel visualization and interaction for
• Query by content
• Motif discovery
• Anomaly detection

4
Some Visual Time Series Systems

• Time Searcher
• Direct Manipulation Pattern Query
• Theme Rivers
• Theme strengthover time
• Spirals
• Periodic Data withknown period

Hochheiser and Shniederman
dot.com stocks 1999-2002
Havre, Hetzler, Whitney Nowell InfoVis 2000
Weber et. al
5
VizTree

• Construct a subsequence tree to span the space of
  subsequences of a given time series.
• Use this to collect statistics about the series.
• Size of the structure is independent of the
  length of the series.

6
VizTree Approach - Overview

• Place windows along the time series to obtain
  subsequences.
• Quantize along time and value dimension to obtain
  sequences of discrete symbols.
• Construct a subsequence tree to represent all
  possible such sequences.
• Collect frequencies of traversal of the branches
  of the subsequence tree.
• Use these for motif and anomaly detection, and
  for comparing time series.

7
Subsequences
Place windows along the time series to obtain
subsequences.
8
Discretization

• Subsequences are patterns.
• Take windows along time series
• length of window length of subsequence.
• Discretize the range of data - one symbol for
  each quantum.
• Divide window into segments represent one
  segment with one symbol.

9
Symbolic Aggregate approXimation(SAX)
Representative symbols
Quantization levels
Segments
One subsequence
Discrete version acdcbdba
10
Subsequence Tree - example
a
b

• symbolsa,b,c
• segments per window2
• Tree spans the space of subsequences.
• Branch factor symbols (size of alphabet)
• Depth segments per window
• Branch thickness freq. of occurrence of
  subsequence.

a
c
a
b
b
c
a
c
b
c
11
VisTree Tool
Demo
12
Query by Content Subsequence Matching

• Finding known patterns
• Chunking
• Breaking a time series into individual series
• Methods
• Time (e.g. power usage)
• Shape (e.g. heart beats)
• Search Approaches
• Exact - Slow
• Approximate - Fast
• Exploration
• Hypothesis Testing

---------
VizTree
---------------------
VizTree
13
Motif Discovery

• Finding unknown patterns
• Not exact matches
• VisTree allows exploration at varying levels of
  precision
• E.g., cc vs. ccac

14
Anomaly Detection

• Finding abnormal patterns.
• Use data already seen to identify anomalies
• Identified by thin branches

15
Comparing Series Diff Tree

• Same parameters ? same tree structure
• Compare the test branch frequencies with respect
  to reference branch frequencies
• Blue underrepresented
• Green overrepresented
• Red equivalent
• Thickness magnitude

16
Thoughts on VizTree (Vis.)

• Most of discovery is implicit
• Manual search
• Parameter setting might be an issue
• Automation might help
• Tree Visualization
• Use of real estate?
• Effective?
• Intuitive?
• Alternatives?

17
Thoughts on VizTree (HCI)

• Primarily a tool to for researchers now
• (Also, we might have an outdated version)
• Even so, some HCI suggestions
• Indication of how tree detail relates to tree
  overview
• Zoom into a specific area of the time series
  (rather than zoomscroll)
• Selection in subsequence detail relates to
  subsequence overview
• Unfortunately, least interesting patterns are
  most easily accessed (branches at root)
• snap to branch or snap to intersection ?
• Ability to turn off highlighting (undo)

18
Summary Unique Contributions

• Fundamental support for aperiodic series
• Scalable
• Resource requirements do not grow linearly with
  length series
• Rich visual feature set
• Global summaries
• Diff-trees between multiple series
• Local patterns and anomalies