Posters Galore: 'Improved detection of nuisance calls in-network using machine learning techniques'

Matthew Middlehurst
University of East Anglia

Matthew Middlehurst, University of East Anglia
Email

Nuisance marketing, nuisance sales and fraud calls - generalised as nuisance calls - are a problem which has rapidly expanded in recent years. Yearly, tens of thousands of people lose money to telephone scams and a significantly larger amount deal with the annoyance of spam from call centres.

Previous studies into automated detection of these calls have largely neglected the presence of legitimate call centres. These cases have similar characteristics to nuisance call centres and can cause great harm if incorrectly blocked. 

For a range of calling line identities used by legitimate and nuisance call centres, we track the outgoing call volume for a single day. In recent years, time series classification (TSC) has had a rapid advance in predictive power. Using this time series data we attempt to classify cases as nuisance or legitimate using a variety of TSC techniques. 

We show that the shapelet transform classifier and the time series meta-ensemble HIVE-COTE can differentiate between these centres reasonably well, but may not reach the specificity required for real world application on their own. A further look into cases misclassified as nuisance shows a lot of cases to have patterns resembling robotic callers. This leads us to doubt the certainty of the provided labels, and suspect the classifiers performed better than initially presented. Further analysing the results, we describe the most informative parts of the series for discriminating between legitimate and nuisance using the inherent interpretability provided by shapelets.

View the poster here
View more posters and videos