Computer Vision News - April 2020

3 Summary A Step Towards Explainability 15 Weight Feature 0.0754 ± 0.0445 thalassemia_reversable defect 0.0459 ± 0.0321 max_heart_rate_achieved 0.0393 ± 0.0334 num_major_vessels 0.0295 ± 0.0245 st_depression 0.0197 ± 0.0382 thalassemia_fixed defect 0.0197 ± 0.0245 rest_ecg_normal 0.0164 ± 0.0000 exercise_induced_angina_yes 0.0131 ± 0.0131 sex_male 0.0131 ± 0.0131 cholesterol 0.0066 ± 0.0334 age 0 ± 0.0000 chest_pain_type_non-anginal pain 0 ± 0.0000 thalassemia_normal 0 ± 0.0000 fasting_blood_sugar_lower than 120mg/ml 0 ± 0.0000 rest_ecg_left ventricular hypertrophy 0 ± 0.0000 st_slope_upsloping 0 ± 0.0000 resting_blood_pressure -0.0033 ± 0.0131 chest_pain_type_atypical angina -0.0131 ± 0.0131 chest_pain_type_typical angina -0.0328 ± 0.0207 st_slope_flat predictions but actually how to do so: this is the partial dependence plot (PDP ). We’ll make a small diversion now from our heart model, to have a look at the fact that PDP’s are calculated after a model has been fit. The advantage of this is that the model is fit on real data without them having been manipulated in any way. It shows the marginal effect on the predicted outcome of a machine learning model that one or two features have (J. H. Friedman 2001). In other words, the relationship between the target and the feature can be shown as linear, monotonic or more complex. As a visualisation example, to better understand PDP, let’s think of a cervical cancer classification where a random forest is fit to predict whether a woman might get cervical cancer based on risk factors. The permutation importance factor listed on that table suggests that thalassemia may be a result of ‘reversable effect’. On the same time, a ‘max heart rate achieved’ is highly important according to the analysis, which indeed makes sense as this is an immediate, subjective state of the patient at the time of the examination (versus, let’s say, the resting blood pressure). Another tool is helpful to further investigate not just what variable most affect