A MapReduce Opinion Mining for COVID-19-Related Tweets Classification Using Enhanced ID3 Decision Tree Classifier

Abstract

Opinion Mining (OM) is a field of Natural Language Processing (NLP) that aims to capture human sentiment in the given text. With the ever-spreading of online purchasing websites, micro-blogging sites, and social media platforms, OM in online social media platforms has picked the interest of thousands of scientific researchers. Because the reviews, tweets and blogs acquired from these social media networks, act as a significant source for enhancing the decision making process. The obtained textual data (reviews, tweets, or blogs) are classified into three different class labels which are negative, neutral and positive for analyzing and extracting relevant information from the given dataset. In this contribution, we introduce an innovative MapReduce improved weighted ID3 decision tree classification approach for OM, which consists mainly of three aspects: Firstly We have used several feature extractors to efficiently detect and capture the relevant data from the given tweets, including N-grams or character-level, Bag-Of-Words, word embedding (GloVe, Word2Vec), FastText, and TF-IDF. Secondly, we have applied a multiple feature selector to reduce the high feature's dimensionality, including Chi-square, Gain Ratio, Information Gain, and Gini Index. Finally, we have employed the obtained features to carry out the classification task using an improved ID3 decision tree classifier, which aims to calculate the weighted information gain instead of information gain used in traditional ID3. In other words, to measure the weighted information gain for the current conditioned feature, we follow two steps: First, we compute the weighted correlation function of the current conditioned feature. Second, we multiply the obtained weighted correlation function by the information gain of this current conditioned feature. This work is implemented in a distributed environment using the Hadoop framework, with its programming framework MapReduce and its distributed file system HDFS. Its primary goal is to enhance the performance of a well-known ID3 classifier in terms of accuracy, execution time, and ability to handle the massive datasets. We have carried out several experiences that aims to assess the effectiveness of our suggested classifier compared to some other contributions chosen from the literature. The experimental results demonstrated that our ID3 classifier works better on COVID-19_Sentiments dataset than other classifiers in terms of Recall (85.72 %), specificity (86.51 %), error rate (11.18 %), false-positive rate (13.49 %), execution time (15.95s), kappa statistic (87.69 %), F1-score (85.54 %), classification rate (88.82 %), false-negative rate (14.28 %), precision rate (86.67 %), convergence (it convergent towards the iteration 90), stability (it is more stable with mean deviation standard equal to 0.12 %), and complexity (it requires much lower time and space computational complexity).

A Parallel Approach for Sentiment Analysis on Social Networks Using Spark

“…When the DOM findings were compared to human annotations, the average accuracy for long and short text was determined to be 76.32 percent. Supervised sentiment analysis using MapReduce was also proposed in [14]. For identifying the sentiment of large corpora, it suggested employing WordMap, a lexicon dictionary, and natural language processing rules in MapReduce operations.…”

Evaluation of different extractors of features at the level of sentiment analysis

“…Most conventional research papers on sentiment analysis has employed supervised machine learning approaches as the primary module for classification or clustering [11]. These approaches typically exploit the Bag-Of-Words, Word2vec, GloVe, FastText, N-Gram and TF-IDF models to extract the essential features of the text containing user-generated sentiments [12].…”

“…P, R, F1 and A of our approach with other approaches selected from the existing literature From the results shown in the table12, we remark that our approach (CNN+FastText) obtained the strongest performances in terms of accuracy (91.32%), precision (93.43%), recall (90.89%), and F1 measures (92.14%) compared to other chosen classifiers from the literature which are Naresh et al[14], Carvalho et al[15], Avinash et al[16], Kumar et al[17] and Zainuddin et al[18].VII. CONCLUSIONFeature extraction is needed to get good performance in sentiment classification.…”

Application of machine learning models based on decision trees in classifying the factors affecting mortality of COVID-19 patients in Hamadan, Iran

“…Also, the structure of decision trees requires less execution time in data classification compared to other machine learning classification techniques [ 19 ]. There are several different approaches to decision trees, including the LMT, C4.5, C5.0, and CART trees, in a variety of research areas such as basic science studies [ 20 ], medicine [ 21 ], and classification images [ 22 ] have been utilized. The random forest is a conventional machine learning algorithm for solving complex problems which is one of the supervised learning methods and its structural model is based on the tree and is used in issues such as classification and regression.…”