Dismiss
InnovationQ will be updated on Sunday, Oct. 22, from 10am ET - noon. You may experience brief service interruptions during that time.
Browse Prior Art Database

AUTOMATED PATTERN RECOGNITION OF FAILURES USING UNSUPERVISED LEARNING

IP.com Disclosure Number: IPCOM000249723D
Publication Date: 2017-Mar-28
Document File: 5 page(s) / 633K

Publishing Venue

The IP.com Prior Art Database

Related People

Eric Chen: AUTHOR

Abstract

Association-rule learning is used in a reverse manner for improved troubleshooting. The most conventionally computation-intensive step is thus transformed into a parallel process. As a result, a high-dimension search space is available to quickly extract failure patterns with combinations of multiple attributes.

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 53% of the total text.

Copyright 2017 Cisco Systems, Inc. 1

AUTOMATED PATTERN RECOGNITION OF FAILURES USING UNSUPERVISED LEARNING

AUTHORS: Eric Chen

CISCO SYSTEMS, INC.

ABSTRACT

Association-rule learning is used in a reverse manner for improved troubleshooting.

The most conventionally computation-intensive step is thus transformed into a parallel

process. As a result, a high-dimension search space is available to quickly extract failure

patterns with combinations of multiple attributes.

DETAILED DESCRIPTION

When a system or software fails, the first step is always to find failure patterns in

order to remedy the failure. However, if failures occur only occasionally with no apparent

way to be reproduced, extracting the pattern can be a very laborious and tedious process

that requires extensive domain knowledge. As such, a method is provided that enables

automated recognition of failure patterns through an unsupervised machine learning

technique.

As Figure 1 below illustrates, the most common approach to troubleshooting is so-

called dimensioning, which divides event logs into cohorts, using a number of key

dimensions (attributes) selected by domain experts. A graph is generated periodically for

each dimension to help identify dominant failure patterns. However, to be successful with

this approach, one needs to choose the right dimensions to monitor. Pattern detection is

done through either “eyeballing” or some fixed threshold, and only one dimension can be

monitored at a time. Multi-dimensional analysis, while also possible, can be

computationally expensive as the cost increases exponentially beyond the first dimension.

Copyright 2017 Cisco Systems, Inc. 2

Figure 1

An existing method called association rule learning is conventionally used to

discover regularities between products in large-scale transaction data. For example, the rule

{A, B} -> {C} found in sales data of a store may indicate that if customers buy item A and

B, they are likely to add item C as well. This algorithm is often used by online shopping

sites to increase sales.

As described herein, this method is modified to induce patterns in a reverse manner.

The basic intuition is to mine for all possible rules that have an outcome (e.g., {C}), and

then extract all associated patterns having the highest statistical significance (e.g., {A, B}).

Figure 2 below illustrates a sample result of applying this invention to collaboration

software. {C} in this case may include definitions of call failures, and a call duration of -1

or 0 may indicate failures. {A, B} on the left hand side may be attributes (e.g., platform,

Copyright 2017 Cisco Systems, Inc. 3

organization, network, Wireless Multimedia Enhancements version, call protocol, AEC

code, Interactive Connectivity Establishment configurations, etc.).

Figure 2

Each dot in the plot in Figure 2 represents a pattern or c...