Explanations with SHAP SHAP is an external model

  • Slides: 5
Download presentation
Explanations with SHAP • SHAP is an external model for explaining the predictions of

Explanations with SHAP • SHAP is an external model for explaining the predictions of any classifier ML – it uses trained classifier – for each sample for each feature it assigns SHAP value – SHAP values describe the influence of the particular feature to the final prediction 1

Application to the lumis with know kind of problem • Let’s take for example

Application to the lumis with know kind of problem • Let’s take for example run 275757 [[104, 122]] with problem: “low DCS: HBHE. nothing on PF plot” • Results for all others (14 runs): https: //github. com/helgako/cmsdqm/blob/master/notebooks/shap-new. ipynb • Presented here results illustrate feature importance for all lumisections from [[104, 122]] • Note: not all runs were found in the dataset; lumis for analysis appeared to be from the TRAIN set (timewise splitting was used) 2

1. Results of application. Several ways to look at the prediction 1. To sum

1. Results of application. Several ways to look at the prediction 1. To sum SHAP values for a feature for all lumis, look at 20 features with the biggest sums. ['q. PFJet 4 CHSEta_3', 'q. EEchi 2_1', 'q. PFJet. EIEta_3’, 'q. SCEta. Width_0', 'q. SCEta. Width 5 x 5_1', 'q. EEenergy_4', 'q. ESenergy_3’, ' q. EEtime_3', 'q. CCEta 5 x 5_0, ' ' q. PFMet. Pt_5', 'q. EBchi 2_0', 'q. CCEta 5 x 5_4’, 'q. CCEn 5 x 5_5', 'q. SCEta. Width 5 x 5_3', 'q. PFMet. Phi_0', 'q. ESenergy_0’, ' q. CCEta 5 x 5_5', 'q. HBHEtime_3', 'q. PFMet. Phi_4’] 2. Take 50 the most important features from each lumisection, look at features which are important for all lumis from the given range. ['q. PFJet 4 CHSEta_3', 'q. EEchi 2_1', 'q. PFJet. EIEta_3’, 'q. SCEta. Width_0', 'q. SCEta. Width 5 x 5_1', 'q. EEenergy_4', 'q. ESenergy_3’, 'q. EEtime_3', 'q. CCEta 5 x 5_0', 'q. EBchi 2_0', 'q. SCEta. Width 5 x 5_3’, 'q. PFMet. Phi_0', 'q. HBHEtime_3', 'q. CCEta 5 x 5_4', 'q. CCEta 5 x 5_5’, 'q. CCEn 5 x 5_5', 'q. ESenergy_0', 'q. EBchi 2_1', 'q. EBtime_3', 'q. HFiphi_1’, 'q. PFMet. Pt_0', 'q. Pre. Sh. YEta_5’] 3

SHAP values distribution 4

SHAP values distribution 4

2. Results of application. From the channels’ point of view Importance for channel =

2. Results of application. From the channels’ point of view Importance for channel = sum of SHAP values for features of particular type PF calo muons photons 71. 0820. 40 -8 0. 172 -0. 827 The problem mostly affected PF channel. Note: there are features not belonging to any of these 4 channels. 5