This project is part of Data Mining & Machine Learning elective course in Computer Engineering & Informatics Department of University of Patras for Spring Semester 2024 (Semester 8).
https://archive.ics.uci.edu/dataset/779/harth
The Human Activity Recognition Trondheim (HARTH) dataset contains recordings of 22 participants wearing two 3-axial Axivity AX3 accelerometers for around 2 hours in a free-living setting. One sensor was attached to the right front thigh and the other to the lower back. The provided sampling rate is 50Hz.
Each subject's recordings are provided in a separate .csv file. One such .csv file contains the following columns:
1. **timestamp**: date and time of recorded sample
2. **back_x**: acceleration of back sensor in x-direction (down) in the unit g
3. **back_y**: acceleration of back sensor in y-direction (left) in the unit g
4. **back_z**: acceleration of back sensor in z-direction (forward) in the unit g
5. **thigh_x**: acceleration of thigh sensor in x-direction (down) in the unit g
6. **thigh_y**: acceleration of thigh sensor in y-direction (right) in the unit g
7. **thigh_z**: acceleration of thigh sensor in z-direction (backward) in the unit g
8. **label**: annotated activity code
The dataset contains the following annotated activities with the corresponding coding:
1: **walking**
2: **running**
3: **shuffling**
4: **stairs (ascending)**
5: **stairs (descending)**
6: **standing**
7: **sitting**
8: **lying**
13: **cycling (sit)**
14: **cycling (stand)**
130: **cycling (sit, inactive)**
140: **cycling (stand, inactive)**
-
Statistic study of the features of the dataset. Our main goal is to reveal probable "hidden" dependancies between the data.
-
Training of 3 types of classifiers that predict the activity of every participant for every timestamp of the testing set based on the N previous values (lags) of the accelometers. It is also noted that the original dataset is split into training set (80% of the original dataset) and testing set (20% of the original dataset). Our main goal is to make a close approximation for every case, not to make perfectly accurate predictions.
- Neural Networks: We chose to work with an Artificial Neural Network that is trained for 50 epochs, using the tensorflow library and keras. For increased accuracy we can use a greater amount of epochs, however this increase in epochs would require a significantly greater amount of time.
- Naive Bayesian Networks
- Random Forests: We chose to work with 30 decision trees (estimators) to create the Random Forest, using the sklearn.ensemble module. We can also use more estimators for increased accuracy like the case of the Neural Network.
-
We also used some performance metrics to evaluate the efficiency of the training of the models, such as: Accuracy, F1 - score, Precision, Recall e.t.c. These metrics are also shown graphically, so that we can have a better visualisation of the performance of the three classifiers.
-
Splitting of the 22 participants into clusters based on their physical activity history using the following three clustering techniques. To determnine the most suitable amount of clusters, we used the Elbow Method w/ Squared Sum Error (SSE) and Bayesian Information Criterion (BIC) and conlcuded that 3-4 clusters should be generated by every algorithm.
- k - means Algorithm (Partitioning Clustering): For the execution of this algorithm we used the sklearn.preprocessing and sklearn.cluster modules.
- Kohonen Network (Net Clustering): For the execution of this algorithm we used the minisom module. The training of the Self Organizing Map (SOM) is random with a neighbor function of variance equal to 0.2. Also the grid of the network is a 2x2 frame and the learning rate is also 0.2.
- Gaussian Mixture (Non-Hierarchical Clustering)
Finally, for the evaluation of the efficiency of clustering we used Silhouette Score metric.
Further details of our implementations can be found in the technical report in the Documentation folder.
Python, Jupyter Notebook