Browsing by Subject "dimensionality reduction"

Now showing 1 - 4 of 4

Access status: Open Access ,
Efficient multi-classifier wrapper feature-selection model. Application for dimension reduction in credit scoring
(Wydawnictwa AGH, 2022) Bouaguel, Waad
The task of identifying the most relevant features for a credit-scoring application is a challenging task. Reducing the number of redundant and unwanted features is an inevitable task for improving the performance of a credit-scoring model. The wrapper approach is usually used in credit-scoring applications to identify the most relevant features. However, this approach suffers from the issue of subset generation and the use of a single classifier as an evaluation function. The problem here is that each classifier may give different results that can be interpreted differently. Hence, we propose an ensemble wrapper featureselection model in this study that is based on a multi-classifier combination. In the first stage, we address the problem of subset generation by minimizing the search space through a customized heuristic. Then, a multi-classifier wrapper evaluation is applied using two-classifier-arrangement approaches in order to select a set of mutually approved sets of relevant features. The proposed method was evaluated on four credit datasets and has shown good performance as compared to individual classifier results.
Access status: Open Access ,
Intrinsic dimensionality detection criterion based on Locally Linear Embedding
(Wydawnictwa AGH, 2018) Meng, Lian; Breitkopf, Piotr
In this work, we revisit the Locally Linear Embedding (LLE) algorithm that is widely employed in dimensionality reduction. With a particular interest to the correspondences of the nearest neighbors in the original and embedded spaces, we observe that, when prescribing low-dimensional embedding spaces, LLE remains merely a weight-preserving rather than a neighborhood-preserving algorithm. Thus, we propose a »neighborhood-preserving ratio« criterion to estimate the minimal intrinsic dimensionality required for neighborhood preservation. We validate its efficiency on sets of synthetic data, including S-curve, Swiss roll, and a dataset of grayscale images.
Access status: Open Access ,
Novel approach for big data classification based on hybrid parallel dimensionality reduction using spark cluster
(Wydawnictwa AGH, 2019) Ali, Ahmed Hussein; Abdullah, Mahmood Zaki
The big data concept has elicited studies on how to accurately and efficiently extract valuable information from such huge dataset. The major problem during big data mining is data dimensionality due to a large number of dimensions in such datasets. This major consequence of high data dimensionality is that it affects the accuracy of machine learning (ML) classifiers, it also results in time wastage due to the presence of several redundant features in the dataset. This problem can be possibly solved using a fast feature reduction method. Hence, this study presents a fast HP-PL which is a new hybrid parallel feature reduction framework that utilizes spark to facilitate feature reduction on shared/distributed-memory clusters. The evaluation of the proposed HP-PL on KDD99 dataset showed the algorithm to be significantly faster than the conventional feature reduction techniques. The proposed technique required >1 minute to select 4 dataset features from over 79 features and 3,000,000 samples on a 3-node cluster (total of 21 cores). For the comparative algorithm, more than 2 hours was required to achieve the same feat. In the proposed system, Hadoop’s distributed file system (HDFS) was used to achieve distributed storage while Apache Spark was used as the computing engine. The model development was based on a parallel model with full consideration of the high performance and throughput of distributed computing. Conclusively, the proposed HP-PL method can achieve good accuracy with less memory and time compared to the conventional methods of feature reduction. This tool can be publicly accessed at https://github.com/ahmed/Fast-HP-PL.
Access status: Open Access ,
Optimizing built-up area extraction in semi-arid regions using Sentinel-2A imagery: a comparative analysis of spectral indices and PCA-based classification in Batna, Algeria
(Wydawnictwa AGH, 2026) Wahiba, Touati; Kalla, Mahdi; Kacha, Lemya
Accurate detection of built-up areas in semi-arid regions is vital for urban planning and environmental monitoring. However, built-up surfaces and bare soils often produce very similar spectral responses. As a result, this similarity causes confusion in satellite image classification. Additionally, spectral overlap among urban materials, bare soil, and sparse vegetation further complicates detection. This study evaluates several spectral indices, including DBSI, NDTI, NDVI, BRBA, and BSI, combined with Principal Component Analysis (PCA) to enhance built-up area extraction from Sentinel-2A imagery. Images captured during the driest season were selected to maximize spectral contrast. Three classification schemes based on Support Vector Machine (SVM) were tested. The first scheme used DBSI, NDTI, and NDVI. The second used BRBA, NDTI, and NDVI. The third relied on PCA-derived components. The results indicate that the PCA-based approach achieved the highest classification accuracy at 95%. In comparison, the DBSI/NDTI/NDVI combination reached 93%, while the BRBA/NDTI/NDVI scheme achieved 92%. Therefore, PCA helps reduce spectral confusion and enhances the identification of built-up areas in semi-arid environments. Overall, combining multiple spectral indices with dimensionality reduction offers a reliable method for urban analysis using Sentinel-2 imagery.