Global Journal of Computer Science and Technology, C: Software & Data Engineering, Volume 22 Issue 2

Fake News Detection: Covid-19 Perspective Global Journal of Computer Science and Technology Volume XXII Issue II Version I 7 Year 2022 ( ) C © 2022 Global Journals correctly predicted to all positive results predicted. • Recall: Recall is the ratio of rightly predicted positive observation of all outcomes in actual class-yes. Fig. 7: Data Preprocessing • F1-Score: F1-Score is the weighted average of Precision and Recall. P erformance E volution We tested the detection system using two different types of feature extraction one is a count vectorizer and another is TF-IDF. As we have only one hundred seventy fake news so we merged them with four different scales (Table 1) of true data to have a depth look at the detection system. Mixing Fake Data with Four Different Scales of True Data. In set-1 we have merged 170 fake news with 180 true news. In set-2 we have 80 more true news than fake news, 58% of total data is true news. In set-3, 66% of data is truly labeled, and set-4 having 76% of true news. Then we have divided our data set into the training set and test set. [’keep’,’smiling’,’beacuse’,’life’, ’is’,’beautiful’,’life’,’is’,’short’, ’and’,’is’,’here’,’to’,’lived’.] keep smiling because life is beautiful life is short and is here to lived Keep smiling because Life is Beautiful Life is short and is here to Lived Keep smiling because Life is Beautiful. Life@ is short and is here to Lived. [’keep’,’smile’,’beacus’,’life’, ’beauti’,’life’,’short’,’live’.] keep smile because life beauti life short life Initial Data Special Character removed Lower character Converted the data into a set of array Stemming and stop words removed Final Data Fake News True News Total Set-1 170 180 350 Set-2 170 230 400 Set-3 170 330 500 Set-4 170 530 800 Precison = TP TP + FP Recall = TP TP + FN F1-Score = 2 * Pre * Re Pre + Re • Precision: Precision is the ratio of positive results Table 1: V. TFIDF. We can see our final result in Table 2. accuracy of 89% on set-1 than other three algorithms using TF-IDF. Also shows stable values on precision, recall, and f1-score in fake news detection. In other sets, the test accuracy as the amount of true data is much greater than the fake data. In set 4 it seems that the accuracy of 81% is much better than the previous two sets. But we can see that the recall (8%) and f1-score (14%), which is too much low in detecting fake news. In Table 3, we have our final result vectorizer. We have a different scenario than TF-IDF while using a count vectorizer. In Table 3, we can see First, we have Multinomial bayes using Multinomial bayes shows the highest test declined of Multinomial bayes using count of Multinomial bayes Naïve Naïve Naïve Naïve that the recall and f1-score drastically decreased while using TF-IDF. But in the count vectorizer, the accuracy increased as expected but the other three values did not drastically decrease shown stable value 3. In set 2 and set 3 Multinomial bayes shown the same accuracy of 88% (Table 3). Naïve On set 4 Passive aggressive classifier shows highest test accuracy of 92% using TF-IDF than other