A Generic Multilevel Architecture for Time Series Prediction
A Generic Multilevel Architecture
for Time Series Prediction
Abstract:
Rapidly evolving
businesses generate massive amounts of time-stamped data sequences and cause a
demand for both univariate and multivariate time series forecasting. For such
data, traditional predictive models based on autoregression are often not
sufficient to capture complex non-linear relationships between multidimensional
features and the time series outputs. In order to exploit these relationships
for improved time series forecasting while also better dealing with a wider
variety of prediction scenarios, a forecasting system requires a flexible and
generic architecture to accommodate and tune various individual predictors as
well as combination methods. In reply to this challenge, an architecture for
combined, multilevel time series prediction is proposed, which is suitable for
many different universal regressors and combination methods. The key strength
of this architecture is its ability to build a diversified ensemble of
individual predictors that form the input to a multilevel selection and fusion
process before the final optimised output is obtained. Excellent generalization
ability is achieved due to the highly boosted complementarity of individual
models further enforced through crossvalidation-linked training on exclusive
data subsets and ensemble output post-processing. In a sample configuration
with basic neural network predictors and a mean combiner, the proposed system
has been evaluated in different scenarios and showed a clear prediction
performance gain.
Existing System:
RECENT e-revolution has
led to a situation in which many businesses and organisations continuously
generate massive amounts of data which constitute univariate and multivariate
time series. Predicting future values of such time series is vital for gaining
competitive advantage in the case of businesses. Time series forecasting is a
very challenging signal processing problem, as in real situations, it is typically
a function of a large number of variables most of which are unknown or
inaccessible at the time of prediction. Although these series usually appear as
very noisy, non-stationary and non-linear signals, their histories may carry a
significant evidence that can be used to build a predictive model.
Disadvantages:
1. The main disadvantage of this
system is to increasing
volume of time series data exhibiting complex non-linear relationships between
its multidimensional features and outputs.
2. It
combines a multilevel architecture of highly robust and diversified individual prediction
models with operators for fusion and selection that can be applied at any level
of the structure.
3) A number of highly correlated
forecasts naturally do not produce a good combination result, however,
diversity has to be traded off with individual accuracy.
Proposed
System:
Performance of time series
forecasting not only depends on the method used, data pre- and post processing
also plays a crucial role. This work presents a generic architecture for
combined, multilevel time series prediction that can accommodate a number of
different individual predictors and combination methods. The architecture
covers a complete prediction cycle from feature selection to output
post-processing and enforces building of a highly diversified ensemble of
individual predictors. These models are then further subjected to multilevel
selection and fusion processes eventually leading to the system output. The
learning process is integrated with a cross-validation method that ensures
exclusive training sets for individual classifiers to encourage their complementarily
and thereby boosting ensemble fault tolerance. Among other key features, the
architecture performs post-processing of the predicted output which is
extremely useful for elimination of output noise.
Advantages:
1) Recent e-revolution has led to a
situation in which many businesses and organisations continuously generate
massive amounts of data which constitute univariate and multivariate time
series. Predicting future values of such time series is vital for gaining
competitive advantage in the case of businesses.
2) It combines a multilevel
architecture of highly robust and diversified individual prediction models with
operators for fusion and selection that can be applied at any level of the
structure.
3) multilevel time series prediction
that can accommodate a number of different individual predictors and
combination methods.
Architecture:
Fig.
1 NN ensemble model building process.
Individual neural networks are trained on disjoint subsets of data and
evaluated via a 10-fold cross-validation process. The predictions P from these
networks are grouped into k disjoint subsets and the outputs from the best
performing network from each group are propagated to the next level at which they are combined by an average
operator.
Software
Requirements Specification:
Software
Requirements:
Front End : java Jsp,Servlet
Back End :
Oracle 10g
IDE : my eclipse 8.0
Language : java (jdk1.6.0)
Operating
System : windows XP
Hardware
Requirements:
System : Pentium IV 2.4 GHz.
Hard Disk
: 80 GB.
Floppy Drive : 1.44 Mb.
Monitor
: 14’ Colour Monitor.
Mouse
: Optical Mouse.
Ram : 512 Mb.
Keyboard :
101 Keyboards.
Module
Description:
- Feature Generation and selection
- Predictor identification information
- Model diversification
- Post processing and tuning
Feature Generation and Selection
The temporal dimension of
the data increases the potential scope of the M-feature space to M ·L dimensions
where L denotes the length of the time series. This means that whatever
set of M features describes the actual problem, its temporal variability
also enforces consideration of the whole available history of feature series as
potential inputs to the predictive model. Careful selection of features is
therefore of much greater importance in comparison to the static-data
prediction problem. On the other hand, temporal feature selection depends
strongly on the availability of features in their temporal relation to outputs
as well as the depth of outputs prediction.
Predictor identification information:
Due to the continuous
nature of typical time series output variables, the choice of individual
predictive models should include universal and flexible regression models
capable of handling multiple inputs and multiple outputs. Neural Networks (NN)
are considered to be a universal non-linear regression model with the ability
to control its complexity and high predictive diversity that can be further
encouraged by varying network architectures and initialisation conditions,
cross-training and even simple injection of noise to the data. Given all these
advantages, we decided to choose a simple Feed forward Multilayer Perceptron
(MLP) as a base model that would be used to test the presented architecture
against standard predictors and combiners.
Model Diversification:
Diversity has many
different forms and could be encouraged or enforced in a number of ways. A
non-exhaustive list of typical diversification techniques is presented below:
Training individual models on different data subsets.
1.
Training
individual models on different feature sets. Injecting noise to different versions of
the training data.
2.
Random
sampling for different versions of the training data.
5.
Selecting
different models for the ensemble.
6.
Varied
initialization and parameter settings of different models.
7.
Selecting
complementary predictors on one or many levels for combination
8.
Varied
combination methods for multi-stage prediction architectures.
Post
Processing and tuning
Combined architecture
described above indicated that the predicted series on average fits the true
series quite well, although it exhibits a significant error component with
particularly harmful occasional peaks of a signal. To reduce the impact of such
errors and any other problems that the predicted series may exhibit, we
introduced a post-processing stage into our architecture. In our case, the post
processing covers an original smoothing technique applied to the predicted
series obtained for the validation set. The parameters of this smoothing have
been fitted to minimize the error rate on the validation set and once fixed,
they are kept constant to deliver smoothed predicted series on the testing set.
Algorithm: Time series forecasting and forecast
combination algorithms
Categories:
0 comments:
Post a Comment