Article Preview
Top1. Introduction
In near future, the data figure will double at least other two years. Few examples that show this massive and fast increase are: Google processes data of 100 PB, Facebook generates data of 600 TB, Baidu, a Chinese company, processes data of 10-100 PB, and Taobao, a subsidiary of Alibaba, generates data of around 10 TB per day from online transactions (Jin, Wah, Cheng, & Wang, 2015). As per reports of IDC, marketing of big data will reach to $32.4 billion by 2017 (International Data Corporation [IDC], 2013). Big datasets typically include masses of structured and unstructured data that require analysis to be performed in real-time. Accessing data, domain knowledge of data, privacy of data, computing and mining data are the major challenges of big data as identified by Wu et al. (2014).
Big data analytics will require a redesigning of existing data mining algorithms and execution in parallel frameworks. Among many alternatives, the MapReduce model and its distributed file system (Zhao, & Pjesivac-Grbovic, 2009; Dean & Ghemawat, 2008) offers a scalable and fault tolerant framework to address big datasets analysis. Hadoop has been the most relevant implementation for MapReduce (Lam, 2009) over the last few years. However, despite its various good properties, Hadoop-MapReduce framework has some drawbacks like the insufficient handling of iterative jobs and slow performance when combining data from multiple sources. Several alternatives like Spark (Zaharia, Chowdhury, Franklin, Shenker, & Stoica, 2010) and Flink (Alexandrov et al., 2014) are looked upon as the new standards.
Pre-processing followed by an appropriate classification technique is one of the most significant approaches in data mining. Classification is the process of classifying data into structured classes or groups. A variety of classification models may be constructed such as fuzzy logic, classification rules, support vector machine, neural networks and fuzzy neural etc. For big data classification, these traditional algorithms need to be modified using various ways. A comprehensive review of scalable machine learning algorithms for big data has been done by Gupta et al. (2016). The authors concluded that most of the previous works have focused on making traditional serial techniques scalable. New techniques build specifically for big data are not worked upon much. The present paper discusses various classification approaches for big data sets. The paper is well organized into five different sections. Section 1 outlines with general introduction of big data and big data classification techniques. Section 2 explores various problems of big data classification. Section 3 elaborates the employed research methodology emphasizing as on what basis the research papers have been selected and also the criteria for their inclusion in our work. It also presents the various research questions to be addressed by our research. Sections 4 and 5 provide a detailed survey and comparative analysis of various classification approaches for big data. Sections 6 and 7 contain concluding remarks and future scope.