Machine Learning Project: Barbell lift

The goal of this project is to use data from various accelerometers to tell if the user performs barbell lifts correctly or incorrectly (5 different ways).

The results are:

A - correct
B - throwing elbows to the front
C - lifting the dumbbell only halfway
D - lowering the dumbbell only halfway
E - throwing hips to the front

Software and Libraries

• Python 3.5.2
• pandas
• scikit-learn

Introduction

The non-preprocessed dataset for this project is provided in the same repository with the name pml-training.csv The dataset consists of 19622 entries of the accelerometers in the x, y and z coordinates making a total of 153 such features. There are six other features making a total of 159 features and hence a table with 19622 rows and 160 columns.

Preprocessing steps

• There are 100 columns that have 19216 missing values each and hence need to be removed from the dataset.
• The columns Unnamed: 0, raw_timestamp_part_1, cvtd_timestamp, user_name, raw_timestamp_part_2, new_window, num_window do not convey any information that is helpful in predictions and hence are removed.
• classe is the target/label/class column and the string classes have been mapped to integers as follows:
  • A -> 0
  • B -> 1
  • C -> 2
  • D -> 3
  • E -> 4

Performance metric

The performance metric used for this project is the standard k-fold cross-validation method with k=10 & 5.

Learning method

The classification method used is a Random Forest using gini method with min_impurty_split or minimum gini level to initiate a split set to 0.1 and a max_depth or maximum permissible depth of the decision tree set to 15. The classifier in this project uses all the available CPUs.

Accuracy score

The k-fold cross-validation used in this project gives the mean accuracy score around 78.79% for k = 10 and 70.05% for k = 5