This project applies human activity recognition to the weight lifting domain. Six participants taught by a professional and equiped with measurement captors were asked to perform barbell lifts correctly and incorrectly in 5 different ways.
Machine learning is used to build a model from a dataset of measurements provisioned by the set of captors installed on the participants, and to classify the correctness of the same exercises executed by anyone without the support of a professional of the domain. The measure of correctness or the outcome of the predictions is given by the class category variable which values belong to the enumeration A,B,C,D and E.
The data for this project comes from this source: http://groupware.les.inf.puc-rio.br/har
training <- read.csv("pml-training.csv", na.strings = c("NA",""))
testing <- read.csv("pml-testing.csv", na.strings = c("NA",""))The training set has 160 variables and 19622 observations and the testing set has 160 variables and 20 observations.
Missing data can be detected in several predictors of the training and the testing datasets, and includes all their observations.
range(colSums(is.na(training)))## [1] 0 19216hist(colSums(is.na(training)), plot = F)$counts## [1] 60 0 0 0 0 0 0 0 0 100range(colSums(is.na(testing)))## [1] 0 20hist(colSums(is.na(testing)), plot = F)$counts## [1] 60 0 0 0 0 0 0 0 0 100Consequently a removal of these predictors is preferred to any data imputation. Only the healthy predictors are considered.
training <- training[,colSums(is.na(training)) == 0]
testing <- testing[,colSums(is.na(testing)) == 0]The predicting power of predictors based on identity and timing features is very low. Only the predictors having a significant predicting power on the outcome are considered.
names(training)[1:7]## [1] "X" "user_name" "raw_timestamp_part_1"
## [4] "raw_timestamp_part_2" "cvtd_timestamp" "new_window"
## [7] "num_window"training <- training[, -c(1:7)]
testing <- testing[, -c(1:7)]The training set has now 53 variables and 19622 observations and the testing set has now 53 variables and 20 observations.
Both sets have the same 53 variables but the last one, respectively classe and problem_id. The training set is significantly larger than the testing one.
A split of the cleaned training set into a training dataset and a validation dataset allows to compute out-of-sample errors during model building evaluation. A 70% / 30% partition is considered and a fix seed is required for reproducibility.
set.seed(33833)
inTrain <- createDataPartition(training$classe, p=0.7, list = F)
training_data <- training[inTrain,]
validation_data <- training[-inTrain,]The outcome to be predicted is the classe variable of the training dataset. A prediction model based on the other variables of the training dataset aims in modelling the behaviour of the data and in predicting the outcome when applied first of all to the validation dataset and finally to the testing set.
Classification trees and Random forests, both with K-fold cross validation, are candidate techniques for this classification problem.
The Random forests algorithm applied to the training dataset using a 5-fold cross validation:
control <- trainControl(method = "cv", number = 5)
model_fit_rf <- train(classe ~ ., data = training_data, method = "rf", trControl = control)
model_fit_rf$results## mtry Accuracy Kappa AccuracySD KappaSD
## 1 2 0.9907553 0.9883039 0.002249986 0.002846600
## 2 27 0.9906093 0.9881209 0.001244782 0.001573247
## 3 52 0.9838390 0.9795558 0.001227610 0.001554533The prediction applied to the validation dataset, followed by its evalution using the confusion matrix:
prediction_rf <- predict(model_fit_rf, validation_data)
confusion_rf <- confusionMatrix(validation_data$classe, prediction_rf)
confusion_rf$table## Reference
## Prediction A B C D E
## A 1672 2 0 0 0
## B 1 1137 1 0 0
## C 0 18 1008 0 0
## D 0 0 25 937 2
## E 0 0 4 3 1075confusion_rf$overall[1]## Accuracy
## 0.9904843The accuracy of the prediction and the out-of-sample error (0.0095157) are pretty good, although the computation of the algorithm is not efficient and takes a long time.
Classification tree algorithm applied to the training dataset using a 5-fold cross validation:
model_fit_rpart <- train(classe ~ ., data = training_data, method = "rpart", trControl = control)
model_fit_rpart$results## cp Accuracy Kappa AccuracySD KappaSD
## 1 0.03722917 0.5128475 0.36415374 0.02031529 0.02753661
## 2 0.06133659 0.4435392 0.25428003 0.06697676 0.11252072
## 3 0.11484081 0.3319451 0.07275564 0.04353741 0.06651522Prediction applied to the validation dataset, followed by its evalution using the confusion matrix:
prediction_rpart <- predict(model_fit_rpart, validation_data)
confusion_rpart <- confusionMatrix(validation_data$classe, prediction_rpart)
confusion_rpart$table## Reference
## Prediction A B C D E
## A 1517 26 128 0 3
## B 484 369 286 0 0
## C 494 28 504 0 0
## D 435 172 357 0 0
## E 162 135 294 0 491confusion_rpart$overall[1]## Accuracy
## 0.4895497The accuracy of the prediction based on a classification tree is very low compared to the one obtained using the Random forests method, and the out-of-sample error (0.5104503) is very high.
The resulting prediction model based on the Random forests algorithm applied to the testing dataset:
predict(model_fit_rf, testing)## [1] B A B A A E D B A A B C B A E E A B B B
## Levels: A B C D EThe Random forests method applied to this dataset returns far better prediction results than the classification tree method, but is computationaly intensive. A usual 10-fold cross validation is not necessary and would make the execution time even longer.