The analysis found that 2.85% of all tweets sent to women MPs between 1 January and 8 June were identified as abusive amounting to a total of 25 688 abusive tweets out of 900 223 total tweets.
The true positive rate, i.e. the tweets labelled as abusive by the model that are truly abusive, is around 64%. The remaining tweets are false positives, i.e. not truly abusive. This means that we expect the number of truly abusive tweets within the 25,688 tweets to be 16,440 (= 0.64 X 25,688 ).
However, a large number of abusive tweets are not captured by the model. Only about 44% of truly abusive tweets are detected and the rest are false negatives, i.e. tweets not labelled as abusive but are truly abusive. This means that our estimate of the actual number of abusive tweets out of the 900 223 total number of tweets is around 37 364 truly abusive tweets (= 16,440 / 0.44)). For the purpose of this publication, we have used the 25,688 figure.