[Paper Review] A Survey of Classification Techniques in the Area of Big Data

Praful Koturwar, Sheetal Girase+1 more|arXiv (Cornell University)|Mar 25, 2015

Data Stream Mining Techniques3 references22 citations

TL;DR

This paper surveys supervised classification techniques for Big Data, analyzing their advantages and limitations in handling large, complex, unstructured datasets. It evaluates algorithms like decision trees, Naïve Bayes, SVM, and K-NN, emphasizing their scalability and performance in big data environments, offering a comparative overview for researchers and practitioners.

ABSTRACT

Big Data concern large-volume, growing data sets that are complex and have multiple autonomous sources. Earlier technologies were not able to handle storage and processing of huge data thus Big Data concept comes into existence. This is a tedious job for users unstructured data. So, there should be some mechanism which classify unstructured data into organized form which helps user to easily access required data. Classification techniques over big transactional database provide required data to the users from large datasets more simple way. There are two main classification techniques, supervised and unsupervised. In this paper we focused on to study of different supervised classification techniques. Further this paper shows a advantages and limitations.

Motivation & Objective

• To examine the state of the art in supervised classification techniques tailored for Big Data environments.
• To identify and analyze the strengths and weaknesses of key supervised learning algorithms in processing large, complex datasets.
• To provide a comparative overview of classification methods suitable for big data workloads, focusing on scalability and efficiency.
• To guide researchers and practitioners in selecting appropriate classification techniques based on data characteristics and system constraints.

Proposed method

• Systematic review of supervised classification techniques such as decision trees, Naïve Bayes, support vector machines (SVM), and k-nearest neighbors (K-NN).
• Analysis of algorithmic behavior in high-volume, high-velocity, and high-variety data environments.
• Evaluation of computational complexity and scalability of each technique in big data frameworks.
• Comparison of accuracy, training time, and resource utilization across different classification algorithms.
• Incorporation of empirical insights from prior studies cited in the survey to assess real-world applicability.
• Use of tabular and graphical summaries (2 tables, 3 figures) to compare performance metrics and structural characteristics.

Experimental results

Research questions

• RQ1 Which supervised classification techniques are most effective for processing Big Data in terms of accuracy and scalability?
• RQ2 How do traditional classification algorithms perform when applied to large-scale, unstructured datasets?
• RQ3 What are the key limitations of existing supervised classification methods in big data environments?
• RQ4 How do different classification algorithms compare in terms of training time, memory usage, and prediction accuracy on big data?
• RQ5 What factors influence the selection of a classification technique for big data applications?

Key findings

• Decision trees show strong interpretability and moderate scalability, making them suitable for semi-structured data.
• Naïve Bayes achieves high speed and efficiency in high-dimensional data but assumes feature independence, which may limit accuracy.
• Support Vector Machines (SVM) offer high accuracy in small to medium datasets but face scalability challenges with very large data volumes.
• K-Nearest Neighbors (K-NN) performs well in low-dimensional spaces but suffers from high computational cost and memory usage in big data settings.
• The survey identifies a trade-off between accuracy and scalability, with simpler models like Naïve Bayes and decision trees often preferred in big data pipelines.
• No single algorithm dominates across all big data scenarios; selection depends on data size, dimensionality, and computational constraints.