Machine Learning Report on Human Activity Classification

Introduction

Predictive modeling is a concept of coming up with a model that has the ability to make predictions. The model usually includes a machine learning algorithm that is involved in learning specific properties from training data sets in order to make these predictions (Witten et al., 2016). It is further classified into regressive and pattern classification. Regression models are basically based on relationship analysis between variables as well as trends in order to make predictions on variables which are continuous. Regression approach assigns discrete class labels to particular observation outcome of the prediction.

An accelerometer is a device that is electromechanical that is used to determine the acceleration force. The force can be static, or continuous like that of gravity. The device senses the vibrations and movement. This device works on the principle that displacement of small proof mass attached into the silicon of an IC and suspended by small beams. On the other hand, a gyroscope is a device that has a wheel mounted in it so that it spins rapidly about an axis and itself is free to alter in any direction. It uses earth gravity to determine orientation. This orientation is not impacted by tilting the mounting meaning can also be used to create stability and maintain direction reference during navigation (Zhang et al., 2011).

This report presents algorithm comparisons, while using human activity data, in deciding, which is the most accurate algorithm to use in classifying g human activity, based on the ones to be presented in the results section. The data consists of employee activities including walking, laying, and standing, which have been detected and captured using accelerometer and gyroscope.

Methods

In machine learning, classification is the problem of identifying to which of the categories set a new observation belongs in the context of training the data having the observations or an instance whose type membership is known. Additionally, it is a learning approach which is supervised where computer program learns from the data input given to it and uses the learning to categorize new observations.

Classification tree predicts the response to data. To do so, one follows the decisions in the tree from the root nodes down to a leaf node. The leaf usually contains the response. The classification tree to be specific gives the response in a nominal way such as ‘false’ or ‘true’ (Kotsiantis, 2013).

On the other hand, the support vector machine (SVM) is a type of machine learning classification algorithm (Yang et al., 2012). Simple (SVM) linearly separates data into two dimensions where the algorithm tries to find a boundary that divides the data in a way that it reduces the miscalculation error. At advanced level is the kernel SVM which separates non- linearly data where a straight line cannot be used as a boundary decision. The model projects the separable and non-linearly data lower dimension to linearly separable data in dimensions which are higher in a way that data points belong to different categories are given different categories.

Lastly, on differentiation, KNN is a supervised learning algorithm family. It deals with given labelled datasets that consist of training observations such as (a, b) that one would like to capture the relationship between the two groups. It is classified as non-parametric and instance based which does not explicit learning model and instead memorizes the training instance which is subsequently utilized as knowledge for the phase of prediction. This model works in a way that K nearest neighbour algorithm boils down to form majority vote between the given unseen observational and K-most similar instances.

Evaluation

The human activity detection has been represented in the figures below:

Human Detection

The first two figures indicate the differences in the classification of activities and imply that walking has more classifications as compared to standing and lying. Consequently, laying has more classifications than standing. The third figure above shows how, altogether, the various predicted human activities are difficult to compare based on classification.

Fine Tree

Scatter Plot

Confusion Matrix

SVM Fine Gaussian

Scatter Plot

Confusion Matrix

Fine KNN

Scatter Plot

Confusion Matrix

As regards the plots, it is displayed that all the human activities detected behave similarly, and the predictions seem correctly as well as incorrectly. Therefore, there is the need to use the confusion matrices to ascertain accuracy.

Comparing model 2 (SVM Gaussian) and model 3 (Fine KNN), it is clear from the confusion matrices of each model that the third model predicts four activities with the highest level of accuracy as compared to model 2. Also, it does so when compared to the first model. Therefore, recommendations here would be that the third model is the best. This is because its performance is more accurate as compared to the other models.

References

• Kotsiantis, S.B., 2013. Decision trees: a recent overview. Artificial Intelligence Review, 39(4), pp.261-283.
• Witten, I.H., Frank, E., Hall, M.A. and Pal, C.J., 2016. Data Mining: Practical machine learning tools and techniques. Morgan Kaufmann.
• Yang, W., Wang, Y., Vahdat, A. and Mori, G., 2012. Kernel latent SVM for visual recognition. In Advances in neural information processing systems (pp. 809-817).
• Zhang, Y., Markovic, S., Sapir, I., Wagenaar, R.C., and Little, T.D., 2011, May. Continuous functional activity monitoring based on wearable tri-axial accelerometer and gyroscope. In 2011 5th International Conference on Pervasive Computing Technologies for Healthcare (PervasiveHealth) and Workshops (pp. 370-373). IEEE.