Evidence based training has been around for for while, where data is collected and pre-defined measures are used to evaluate the training session. Skisens AB are presenting new methods to evaluate a session which can be compared between sessions in a fair way, without having outside factors influencing the results by, measuring the force generated by the skier. Measuring the force generated involves customized handles that changes the dimensions of the handles and in the extension the ergonomics of the handles. This thesis aims to try to accurately predict the force generated in each stroke from the skier, in order for Skisens AB not having to measure the force using the customized handles. We propose an unsupervised detection algorithm, for detecting when a stroking motion is performed as well as a few model design for achieving the best predictive results. Contents: 1 Introduction 1 1.1 Skisens . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.2 Objective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 2 Data Overview 3 2.1 Description of the data . . . . . . . . . . . . . . . . . . . . . . . . . . 3 2.2 Unsupervised event detection . . . . . . . . . . . . . . . . . . . . . . 10 2.2.1 Identifying an event . . . . . . . . . . . . . . . . . . . . . . . . 13 2.3 Feature engineering . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 2.3.1 Response variable . . . . . . . . . . . . . . . . . . . . . . . . . 19 2.3.2 features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20 2.4 Correlations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 2.5 Assumptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 3 Theory review 27 3.1 Decision trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 3.2 Random forest . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 3.2.1 Variable Importance . . . . . . . . . . . . . . . . . . . . . . . 30 4 Study design 33 4.1 Design 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33 4.2 Design 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 4.3 Design 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 5 Results 37 5.1 Parameter tuning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 5.1.1 Result of tuning for the models in design 1 . . . . . . . . . . . 38 5.1.2 Result of tuning for the models in design 2 . . . . . . . . . . . 39 5.2 Performance of models . . . . . . . . . . . . . . . . . . . . . . . . . . 41 5.2.1 Performance of the models in design 1 . . . . . . . . . . . . . 41 5.2.2 Performance of the models in design 2 . . . . . . . . . . . . . 41 5.2.3 Performance of the models in design 3 . . . . . . . . . . . . . 42 5.3 Feature importance . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 6 Conclusions 45 6.1 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46 Bibliography 49 A Appendix I A.1 Right side variables when positive force . . . . . . . . . . . . . . . . . I A.2 Distribution of a few features conditioned of style . . . . . . . . . . . III A.3 Result of tuning for the models in design 3 . . . . . . . . . . . . . . . IV A.4 Diagnostic plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . VI A.4.1 Models design 1 . . . . . . . . . . . . . . . . . . . . . . . . . . VI A.4.2 Models design 2 . . . . . . . . . . . . . . . . . . . . . . . . . . VII A.4.3 Models design 3 . . . . . . . . . . . . . . . . . . . . . . . . . . VIII
© Copyright 2019 Published by University of Gothenburg, Department of Mathematical Sciences. All rights reserved.