Our Twitter mechanical Turk
The accuracy of a classification algorithm should be measured against a test dataset, meaning a labeled dataset that was not included in the training phase. We do not have access to such a dataset (this is the reason we bootstrapped our model initially), hence we cannot compare the original versus predicted categories. Instead of the true accuracy, we can estimate an overall confidence level by visualizing our results. With all our data on Elasticsearch, we build a Kibana dashboard with an additional plugin for tag cloud visualizations (https://github.com/stormpython/tagcloud).
The following figure shows the number of GDELT articles that were analyzed and predicted on May 1, 2016. Around 18,000 articles have been downloaded in less than 24h (by batch interval of 15 minutes). At each batch, we observe no more than 100 distinct predicted hashtags; this is fortunate as we only kept the top 100 popular hashtags occurring within a 24h time window. Besides, it gives...