Channel 4 Presentation

Channel 4 Presentation

Azeem Iqbal

Weight Lifting Classifier

• Human Activity Recognition
• Six males performing dumbell curls, 4 sensors
• Sensors on arm, forearm, belt and dumbbell

Classes

Dumbbell curl performed in 5 different ways:

• A - Correctly
• B - Throwing elbows forward
• C - Lifting only halfway
• D - Lowering the dumbbell only halfway
• E - Swinging Hips

The Task

Use machine learning to create a classifier:

• Training data set to create model
• Apply model onto test set
• Graded on accuracy of predictions

The Data

12Mb Data set ~20k observations of 160 variables Initially just threw it all into a randomForest model Not Good!

Manipulating Data

Cut down the amount of variables from 160 to 53:

• Got rid of all stats measurements (avg, skew
• Removed other useless data (name, time)
• Left with gyroscope (x,y,z)

Building our Model

Used Random Forests Partitioned training data further:

• 70% for training
• 30% for validation

Building our Model

Model was incredibly accurate (0.5% Error rate) Overfit Tried boosted decision trees, similar accuracy

Model Performance

Random Forest model accurate on validation set 99.35% Accuracy (Close to previous error rate)

Model Performance

Suspected overfit Full marks for test

Conclusion

No idea

Why is this interesting?

Came back to this another time Read documentation of the data Weight lifting coach!

Applications to C4

Cluster your viewers Build a classifier to identify other people into clusters