ASI: Automated Summary and Insight on data cubes

1
(No Transcript)
2
(No Transcript)
3
IBM A large IT services provider organization

• Services, Services, Services
• - to different companies
• Oracle, Intel,
• - across the globe
• US, Canada,India,China,
• - various service lines
• Database, Storage,
• Prism Insight A platform for
• IT Service Metric Monitoring.

4
IT Service Data

• IT Service Metrics
• - MTTR Mean Time to Resolve
• - Opened Tickets
• - Closed Tickets
• WID Geography
• Service line
• workstream
• account

WID
We model the data as matrices
A matrix corresponds to a metric.
TIME
5
Summarizing the IT Service Data

• - Select Representative WIDs

Representative WIDs
6
(No Transcript)
7
(No Transcript)
8
(No Transcript)
9
Connection between Subset Selection and Service
Metric Summary
A
C
TIME
TIME
WID
WID

• Problem min A CCA
  F

C
10
(No Transcript)
11
Subset selection
12
(No Transcript)
13
Summary of deterministic Subset Selection
algorithms
14
Our Clustered Subset Selection (CSS)
WID
Input data matrix
TIME

• clustering
• cluster the columns of the matrix

• Column Selection
• select one column from each cluster

15
Our Clustered Subset Selection (CSS)
Input data matrix
Your favorite matrix

• k-means
• k-median
• spectral clustering
• close to rank 1 clustering

• clustering
• cluster the columns of the matrix

• Select best column by exhaustive enumeration
• use one QR-based deterministic technique

• Column Selection
• select one column from each cluster

16
CSS vs. other deterministic methods

• Comparison of 8 subset selection algorithms over
  4 datasets.
• Our method outperforms the others in all
  datasets.
• Entries show approximation error A CX /
  A - Ak .
• A - Ak is the best rank-k approximation
  with the SVD.

17
(No Transcript)
18
MTTR Mean Time to resolve a problem
MTTR
TIME (490 dates)
WID ( 834 wids)
19
Conclusions Our results

• We proposed the Clustered Subset Selection
  Algorithm
• to solve a classical linear algebraic
  problem
• We demonstrated the effectiveness of our
  algorithm
• in IT service metric summarization.

20
Conclusions One idea to take home

• If you have a summarization problem
• - text summarization
• - video summarization
• - you name it
• There is a lot of work in linear algebra that
  might be useful in your domain.