Scribd
Upload
27 views12 pages

University of Computer Studies, Mandalay (UCSM)

This document discusses techniques for using data mining to design neural network time series predictors. It describes two categories of prior knowledge extraction: direct and indirect embedding. Direct embedding uses information theory, dynamical systems analysis, and stochastic modeling to provide insight into neural network architecture, parameters, and data sampling. Indirect embedding exploits known target function properties and non-stationarity detection to enlarge data sets and avoid unnecessary model redesign. The document provides examples of applying these techniques to improve time series prediction accuracy and efficiency.

Uploaded by

Hsu Let Yee Hnin
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
27 views12 pages

University of Computer Studies, Mandalay (UCSM)

This document discusses techniques for using data mining to design neural network time series predictors. It describes two categories of prior knowledge extraction: direct and indirect embedding. Direct embedding uses information theory, dynamical systems analysis, and stochastic modeling to provide insight into neural network architecture, parameters, and data sampling. Indirect embedding exploits known target function properties and non-stationarity detection to enlarge data sets and avoid unnecessary model redesign. The document provides examples of applying these techniques to improve time series prediction accuracy and efficiency.

Uploaded by

Hsu Let Yee Hnin
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
You are on page 1/ 12

University of Computer Studies, Mandalay

(UCSM)

Faculty of Computer Science


Chapter 10 Data Mining Techniques
for Designing Neural Network Time
Series Predictors
 A real-life time series prediction system is usually subject
to two constraints- accuracy and time, meaning that a
sufficiently accurate prediction has to be provided in an
imposed time frame.
Cont’d
 Direct knowledge embedding methods, based on
information theoretical modeling, dynamical system
analysis, and stochastic modeling.
 It is illustrated that direct methods can produce a wealth
of prior information regarding the choice of an
appropriate neural network architecture, data sampling
rates, as well as starting values for the model
parameters, which otherwise have to be found as a
result of costly trial-and-error procedures.
Cont’d
 Indirect embedding methods which exploit known
properties of the target function and non-stationarity
detection techniques.
 The use of known properties of the target function can
enlarge scarce data sets or enforce more accurate
learning through constrained optimization.
 Non-stationarity analysis can considerably improve the
computational efficiency of time series forecasting by
avoiding the neural network model redesign more often
than needed.
Introduction
 The outcomes of a phenomenon over time form a time
series. Time series are encountered in sciences as well
as in real life.
 Example of time series:
 the voltage measured every second across a resistor in an electrical
circuit,
 the number of cars passing a marker on a highway every minute,
 the yearly power consumption of the United States,
 the hourly exchange rate of German mark versus U.S. dollar,
 the daily car production of Chrysler corporation
 The observation of past outcomes of a phenomenon in
order to anticipate its future behavior represents the
essence of forecasting (prediction).
Cont’d
 The design of an appropriate neural network for time series
prediction problems with high data arrival rates (e.g. Internet traffic
predictions or financial intra-day predictions) can be a challenging
task, due to time consuming trial-and error architecture selection,
non-linear parameter optimization and the need to devise new
prediction models whenever the underlying data distribution
changes.
 For these reasons, any prior knowledge that could be extracted from
a time series under study can dramatically decrease the design time
of a predictor and also improve its prediction accuracy significantly.
Cont’d
 Two categories of prior knowledge extraction techniques,
which are embeddable into neural network prediction
models.
 The first category, denoted as direct knowledge embedding
encompasses information theory, non-linear dynamics, and
stochastic analysis.
 These techniques of exploratory data analysis can provide prior
knowledge regarding appropriate neural network architecture, initial
network parameters and adequate data sampling rate.
 A real-life time series (compressed video traffic data), as well as an
artificial, non-linear, chaotic time series (Mackey-Glass data) are
used to illustrate the embedding of prior knowledge extracted from
stochastic analysis into the neural network design process.
Cont’d
 The second category, indirect knowledge embedding,
addressed, includes the use of known properties of
target functions and non-stationarity detection.
 The use of known properties of the target function can
enlarge scarce data sets by creating additional artificial
training examples, or enforce more accurate learning
through constrained optimization.
 Non-stationarity analysis can considerably improve the
computational efficiency of time series forecasting by
avoiding the neural network model re-design more often
than needed.
Direct Information Extraction
Procedures
 Information Theory
 Dynamical System Analysis
 Stochastic Analysis
Indirect Information Extraction
Procedures
 Knowledge of Properties of the Target Function
 Non- Stationarity Detection
Knowledge of Properties of the
Target Function
 Some properties of the function to be approximated are
known ahead of time.
 The use of these properties, called hints, is of major
importance especially in problems with scarce, costly to
obtain, or noisy data, like financial forecasting problems
in which hints can improve the model’s accuracy
dramatically.
 A non-valid hint can deteriorate the performance of the
model considerably, so care must be taken in order to
analyze the validity of hints.
Cont’d
 Two modalities for incorporating hints in the neural
network learning process are proposed:
 creating additional “virtual” training examples; .
 imposing constraints on the learning process by
modifying the cost function.

You might also like