Please finish the chapters 4 to 7 in total 12000 of the report,
Chapter 1-3 are already written
No SPSS needed
result in the need to use Python with tree base model analysis- random forest, XGboost, SVM to analysis
-result and discussion with the following angle to write:
Angles
Machine Learning positioning: 2. Model prediction ability as a product? -> is machine learning capable to predict injury in the future? -> from binary classification(this project) to multi-class classification (e.g. predict the injury site instead of injury or not, predict the injury risk level e.g. green, yellow, red)
Model findings on the data thru training and prediction process? -> feature importance -> whats insight from the importance, why the model treat some of the features more importanct than others after learning from the data
Feature importance usage: 1.1. compare 3 models and using the result and feature importance ONLY from the best model? ->by what score we decide the model is the best? precision? recall? roc auc score? accuracy? different intepretation ->recall: we dont mind about false positive case? can we treat those who falsely predicted injury people as higher risk patients? in this case, can the model be used as 1st line of screening? but if ONLY focus recall, model may lose the basic ability of classify injury or not, e.g. recall with 1 can mean that model classify all patient as injury.
1.2. Look into the result of 3 models and find out the common ground ->Compare the top 10 important features of 3 models and find out the common features appeared in all 3 ->Most common features means they all important despite the difference of 3 model algo, as different algo may have different apporach/bias to look into the data ->Are those common features align with your domain knowledge? e.g. Right hip – right knee alignment bla bla ->Find the connection between those commonly important features and your domain knowledge to create your story/justify your domain knowledge with these feature findings
Data Bias: 2.1. Notice that right Hip and right knee usually higher importance than left ->data bias? the majority of data are right hander that their main leg is left? ->as main leg is left, right hip strength is weaker and right knee align is worse? ->As the majority of data input are “right” so the model is biased to right ->biased model breed biased result and biased feature important ->improvment: shd we not differentiate 2 sides? or shd we change the target of the model from (Yes vs No) to (left side injury vs right side injury vs no injury, which is multi-class classification mentioned above)
