Engagement counts help sentiment models only after they are filtered and weighted.
Numeric engagement data is tempting precisely because it is easy to collect. A like is a number, a retweet is a number, and both arrive without any parsing. The trouble is that a number is not an opinion, and a model that leans on it too hard will learn the wrong thing.
The approach in this study does the two steps in that order. Fisher Score first ranks which engagement features are worth keeping, cutting the ones that say little about sentiment. Only then does the model decide how much each family of features should count.
That second decision takes the form of two weights, alpha for the text side and beta for the engagement side. Rather than guessing, the team ran the experiment across combinations and read the F1 score at each setting.
The winning setting kept the larger share on the text, at 0.6 against 0.4, and produced an F1 score of 0.838. The gap between the two weights is not dramatic, and that is the interesting part. Engagement data adds something real, but it does not take over the decision.
Anyone working with social data can carry one habit out of this. A reaction metric measures the audience as much as the message, so it belongs in the model beside the words rather than in place of them.
Rizal Setya Perdana and Aryo Pinandito (2018). Combining Likes-Retweet Analysis and Naive Bayes Classifier within Twitter for Sentiment Analysis. Journal of Telecommunication, Electronic and Computer Engineering.
