Evaluation of the methodology


Three regions of the United Kingdom are considered in our scheme: Central England & Wales, South England and North England. The experiments are carried out on data from 22/06/2009 to 28/03/2010 (overall 40 weeks or 280 days). The datasets are comprised by vector space representations of the Twitter corpus using a vocabulary of 2675 candidate features and flu rates published by the Health Protection Agency (HPA) denoting the GP consultations per 105 citizens where the diagnosis result was Influenza-like Illness (ILI).

MAE: Mean Absolute Error Magenta Line: Inferred flu rates from the Twitter corpus
LCC: Linear Correlation Coefficient Red Line: Flu rates from the HPA



  • We train on days 1-259 (weeks 1-37) on all 3 regions and test our inferences on days 260-280 (weeks 38-40) per region. Here is the list of selected markers derived by applying Bolasso.

  • Central England & Wales South England North England
    MAE: 5.27 MAE: 4.26 MAE: 2.18



  • We train on days 22-280 (weeks 4-40) on all 3 regions and test our inferences on days 1-21 (weeks 1-3) per region. Here is the list of selected markers derived by applying Bolasso.

  • Central England & Wales South England North England
    MAE: 18.34 MAE: 9.38 MAE: 27.29
    LCC: 0.94 -- P-value: 3.48e-10 LCC: 0.84 -- P-value: 2.18e-06 LCC: 0.87 -- P-value: 4.01e-07



  • As an overall performance quantification, 10-fold cross validation is performed where each fold is formed by 4 contiguous weeks; the MAE is on average equal to 11.1 with a standard deviation of 10.04. The MAE for folds 1-10 is respectively equal to 30.4592, 27.4488, 4.5617, 4.4625, 8.8856, 7.7987, 14.8908, 3.1253, 3.5159 and 5.8573.