Abstract:
|
In this dissertation, a method of feature selection in machine learning, and
more particularly supervised learning is presented. Supervised learning is a machine
learning task that infers answers from a training data set. In machine learning, training
datasets are employed in order to create a model which enables reasonable
predictions, while in supervised learning, each training example is a training set
consisting of instances and labels, and the learning objective is to be able to predict
the label of a new unseen instance with as few errors as possible. In recent years,
many proposed learning algorithms that perform fairly well have been proposed. The
factors to accomplish successful model building depend on many aspects such as
noise and size of data. Most often for learning algorithms, it is assumed that training
data is represented by a vector of numerical data for which each measurement is a
feature, and an important question related to machine learning is how to represent
instances using vectors of these to yield high learning performance.
Nowadays, data volumes are tremendously large in terms of aspects such as
the number of features and most machine learning and data mining techniques may
not be productive for high dimensional data, query accuracy and efficiency lessen
swiftly as the dimension increases, the so-called curse of dimensionality. One of the
requirements of good representation is conciseness since representation that uses too
many features incurs major computational difficulties and may lead to poor prediction
performance. Attribute selection is one of the significant methods in which the
objective is to choose a small subset to predict the target sufficiently well. Feature
selection selects the most importance features, eliminates irrelevant and redundant
features from the entire set of attributes, reduces the computational complexity of any
learning and prediction algorithm used in the process, and reduces cost by excluding
unselected features.
A floating search is commonly used for the searching process. They are
heuristic search methods which dynamically change the number of attributes included
or eliminated at each step; they have produced very good results. The principal
improvement of this thesis is focused on filter-based feature selection using genetic
algorithm technique. Filters are normally less computationally intensive than wrapper
method because wrappers apply a predictive model to score feature subsets. This
approach is selected to be fast to reckon, whereas rooted to spot apprehending the
goodness of the feature subsets. GA method can help to gain more diversity of
population and provides us a way of reducing search space. Moreover, the
contributions related to improves the contemporary sequential forward floating
selection algorithm. In this thesis, an improving feature step using genetic algorithm is
proposed as an additional step in a floating search. The objective is to eliminate weak
features and replace a predominant one at each sequential step. From the research
observations, the proposed method was discovered to be beneficial in selecting
features that can boost the accuracy of data classification. Moreover, the experimental
outcomes show that the proposed method with the genetic algorithm enhanced
classification correctness and cut down data dimensionality for supervised learning
problems.
|
