System and method for predicting response time of an enterprise system

Patent metadata

Field	Value
Publication number	US-2017185902-A1
Application number	US-201615272025-A
Country	US
Kind code	A1
Filing date	Sep 21, 2016
Priority date	Dec 29, 2015
Publication date	Jun 29, 2017
Grant date	—

Abstract

System and method for predicting enterprise system response time is disclosed. System pre-processes causal variables of historical output time series data to select subset of causal variables by applying regression techniques to obtain significant causal variables. Historical output time series data shows response time of enterprise system. System derives dummy variables from historical output time series data using threshold based method. Dummy variables are specific to peak detection and trough detection in historic output time series data. System trains predictive model using historical output time series data, significant causal variables, and dummy variables to generate trained predictive model and predictive model designed using machine learning technique selected based on forecast methodology used for forecasting input time series data. System predicts enterprise system response time by using trained predictive model, input time series data or lag between input time series data and historical output time series data.

First claim

What is claimed is:

1 . A method for predicting response time of an enterprise system, the method comprising:
pre-processing, by a processor, a plurality of causal variables of an historical output time series data, affecting a response time of an enterprise system, to select a subset of the plurality of causal variables by applying one or more regression techniques to obtain significant causal variables, wherein the historical output time series data shows the response time of the enterprise system, and deriving, by the processor, a plurality of dummy variables from the historical output time series data, by using a threshold value based method, and wherein the plurality of dummy variables are specific to at least one of a peak detection and a trough detection in the historical output time series data; and training, by the processor, at least one predictive model using 1) the historical output time series data, 2) the plurality of significant causal variables, and 3) the plurality of dummy variables to generate at least one trained predictive model, wherein the at least one predictive model is designed using at least one machine learning technique; and predicting, by the processor, the response time of the enterprise system by using the at least one trained predictive model and an input time series data or a lag between the input time series data and the historical output time series data

2 . The method of claim 1 , wherein the input time series data is forecasted based on a type of the causal variables of the input time series data, wherein the type of the causal variables is either numerical or categorical.

3 . The method of claim 1 , wherein the historical output time series data is pre-processed to add one or more missing values in the historical output time series data.

4 . The method of claim 1 , wherein the historical output time series data is a multivariate time series data of the enterprise system indexed by ‘t’ with independent variables (input) labeled as X i (t) and the dependent variable (output) labeled as Y(t), where ‘i’ takes values from 1 to n with ‘n’ being the total number of independent variables that affect the response time Y(t).

5 . The method of claim 1 , wherein the historical output time series data is measured at predefined sampling intervals or computed using an aggregation scheme for long intervals, and the plurality of causal variables are measured simultaneously with respect to a time measurement frame of the historical output time series.

6 . The method of claim 3 , wherein the one or more missing values are added to the historical output time series data at one or more historical instants depending on an enterprise system state at that instant, or at one or more previous instants, or at subsequent instants in the historical output time series data.

7 . The method of claim 1 , wherein the threshold values are either preset, based on historical knowledge, or design specifications, or dynamically adjusted as data evolves with time.

8 . The method of claim 1 , wherein the plurality of dummy variables are selected based on capturing temporal and seasonal patterns in the historical output time series data, and the plurality of dummy variables further assist in capturing the temporal and seasonal patterns in the historical output time series data while predicting the response time of the enterprise system.

9 . The method of claim 1 , wherein applying the one or more regression techniques comprises selecting the subset of the plurality of causal variables by evaluating the temporal relationship among the plurality of causal variables.

10 . The method of claim 1 , wherein the plurality of dummy variables, the significant causal variables and threshold values, and the at least one machine learning technique are used to train the at least one predictive model to forecast the output time series data.

11 . The method of claim 1 , wherein using the threshold based method comprises automatic computation of one or more thresholds, wherein a first threshold and a second threshold from the one or more thresholds correspond to a first set of values of Y(t), and a third threshold and a fourth threshold from the one or more thresholds correspond to a second set of values of Y(t), and wherein the first threshold, second threshold, third threshold and the fourth threshold are calculated after removing outliers from Y (t), and wherein computation of the outliers is based on a multiple of the standard deviation on either side of the mean value of Y(t) until t.

12 . The method of claim 1 , wherein the predictive models are designed based on a first technique comprising using a direct relationship between input variables X(t) and Output variables Y(t) of the output time series data, and wherein the input variables X(t) include the original input variables Xi(t) that are the significant causal variables of the output time series data and the derived inputs Xdi(t) that are the plurality of dummy variables of the output time series data.

13 . The method of claim 1 , wherein forecasting the input time series of numerical data is based on a seasonal mean of the input time series data, wherein the seasonal mean is the average of the input variable values at the corresponding time instants in history.

14 . The method of claim 1 , wherein forecasting the input time series of categorical data is based on a mode of the input time series data at the corresponding time instants in history.

15 . The method of claim 1 wherein derivation of the plurality of dummy variables is based on a mode of the plurality of dummy variables at the corresponding time instants in the history.

16 . The method of claim 1 , wherein the at least one predictive model is designed based on a second technique comprising use of a modified naive method for predicting the output Y (t) using a time period lagged version of the inputs variables along with a lagged output variable.

17 . A system 102 for predicting a response time of an enterprise system, the system 102 comprising:
a processor 202 ; and
a memory 204 coupled to the processor 202 , wherein the processor 202 executes a plurality of modules 208 stored in the memory 204 , and wherein the plurality of modules 208 comprises:
a pre-processing module 210 to,
pre-process, a plurality of causal variables of a historical output time series data affecting a response time of an enterprise system, to select a subset of the plurality of causal variables affecting the response time of the enterprise system by applying one or more regression techniques to obtain significant causal variables, wherein the historical output time series data shows the response time of the enterprise system; and

a variable generating module 212 to,
derive, a plurality of dummy variables from the historical output time series data, by using a threshold based method, and wherein the plurality of dummy variables are specific to at least one of a peak detection and a trough detection in the historical output time series data; and

a prediction module 214 to,
train at least one predictive model using 1) the historical output time series data, 2) the plurality of significant causal variables, and 3) the plurality of dummy variables to generate at least one trained predictive model, wherein the at least one predictive model is designed using at least one machine learning technique; and
predict, the response time of the enterprise system by using the at least one trained predictive model, and an input time series data or a lag between the inpu

Assignees

• Tata Consultancy Services Ltd

Inventors

• Kumar Kriti
• Thokala Naveen Kumar
• Karumanchi Ravikumar
• Chandra Mariswamy Girish
• Kamakolanu Guru Kalyan Prathap
• Upparapalli Suresh
• Chowdary Madhusudhan Kamma Chavala
• Kulkarni Prasanna Madhavrao
• Sharda Pareshkumar Bhawanishankar

Classifications

• G06N20/00 Primary

  Machine learning · CPC title

• G06N5/022 Primary

  Knowledge engineering; Knowledge acquisition · CPC title

• G06F18/2135

  based on approximation criteria, e.g. principal component analysis · CPC title

• G06Q10/04

  Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem" (market predictions or forecasting for commercial activities G06Q30/0202) · CPC title

• G05B23/024

  Quantitative history assessment, e.g. mathematical relationships between available data; Functions therefor; Principal component analysis [PCA]; Partial least square [PLS]; Statistical classifiers, e.g. Bayesian networks, linear regression or correlation analysis; Neural networks · CPC title