Twitter timelines fill up with more than conversation. Promotions, duplicate posts, and links to nothing useful arrive alongside the posts a user actually wants. A 2017 project built a way for Android developers to filter that noise out of their own apps.
The idea is a reusable framework rather than one finished app. A developer drops the classifier in as a module, and the app can label each incoming tweet as spam or legitimate before it reaches the timeline. The heavy work sits behind that interface.
The team collected tweets through the Twitter search API around trending topics in Indonesia, all written in Indonesian. One hundred tweets were labelled by hand for training, and another hundred were held back for testing.
Naive Bayes reached 82 percent accuracy on the test set. A nearest neighbour classifier, weighted with a term frequency measure, came in at 71 percent. Both beat guessing, and the simpler of the two was clearly ahead.
The value of the work is in that packaging. A small classifier that runs on the phone, trained on a hundred labelled examples, is a far easier thing for an app team to adopt than a research pipeline.
The honest limits are the dataset size and the language. Two hundred tweets from one country is a narrow base, and spam patterns shift quickly. The authors present the framework as a starting point developers can retrain on their own data.
Pinandito et al. (2017). Spam Detection Framework for Android Twitter Application Using Naive Bayes and K-Nearest Neighbor Classifiers. Proceedings of the 2017 International Conference on Software and Computer Applications.
