Last month’s task was to evaluate possible feature sets to automatically classify tweets as credible / not credible based on the user data I aggregated with the user study.

I identified 25 features and already tested various feature combinations and trained different SVM models. The F1-score which is a very common evaluation metric in Information Retrieval was used to compare the results of the various feature combinations (using a 10-fold cross validation).

To evaluate the feature selection against a different dataset i downloaded the test data from Mediaeval 2016 Conference and employed my model on the data.

The results are promising and the next task is to enhance the results, build a robust model and implement it into the prototype.