Machine learning algorithm for classification of activity of daily life’s
2 likes847 views
The document describes a machine learning approach to classify activities of daily living (ADL) using data from a wrist-worn accelerometer. The approach uses support vector machines (SVM) with feature engineering that includes vector magnitude and singular value decomposition. The model is trained on a dataset containing 11 ADLs performed by 16 volunteers. Hyperparameter tuning is performed to optimize the SVM, achieving up to 86% accuracy on test data when using both vector magnitude and SVD features compared to 71% accuracy using only vector magnitude. The results demonstrate an improved method for detecting ADLs but would benefit from testing on additional datasets.
![Overview of dataset
Accelerometer-2011-04-11-13-28-18-brush_teeth-f1.txt
- Refers to an accelerometer recording that was taken on
March 11, 2011, starting from 13:28.18 p.m.
- The recording refers to the HMP ”brush_teeth" executed
by the volunteer with ID "f1".
Provides equations to convert raw acceleration into real
acceleration values and reduce noise
Raw data from accelerometer for the task “brush_teeth”.
The columns represents the acceleration measured in
X,Y and Z direction.
Each folder contains raw data for the activity.Each file in the dataset has the
following naming convention:
Accelerometer-[START_TIME]-
[ADL]-[VOLUNTEER]](https://image.slidesharecdn.com/machinelearningalgorithmforclassificationofactivityofdailylifes-160923181324/85/Machine-learning-algorithm-for-classification-of-activity-of-daily-life-s-9-320.jpg)
Machine learning algorithm for classification of activity of daily life’s
- 1. Machine learning algorithm for classification of Activity of daily life’s Siddharth Chakravarty
- 2. Overview ADL- What is it? How is it monitored? Technology landscape Problem statement Approach Results
- 3. ADL- Activity of Daily life’s A technique for classification of human activities can be a useful tool to not only classify and monitor our activities, but also improve overall quality of life Image Source: connectingcleveland.net/wp-content/uploads/2014/07/Daily_Life_Wallpaper.jpg
- 4. State of the art technology ADL monitoring 4 • An accelerometer essentially records the acceleration it experiences is the X,Y and Z direction. • Accelerometers are the most commonly used type of sensor for activity recognition with wearable sensors and other consumer electronic devices ranging from Iphones to Wii Walking Running Jumping Sweeping Sleeping Time scale Image courtesy: Jawbone
- 5. Technology landscape Major players • Fitbit - recently announced next generation device • Jawbone- acquired bodymedia to complement it’s technology • Microsoft • Samsung • Pebble • Misfit– Acquired by Fossil • Facebook- acquired Finland-based fitness app maker Protogeo Many more…. Clearly there is a need for accurately monitor and classify ADL’s not just for recreation, but also for other applications. Heart rate Walking Running Sleeping ADL Heart rate monitors X Pedometers X X X Phone X X Wearable X X X X
- 6. Problem statement 6 The objective of the project was to develop machine learning algorithm using SVM to predict ADL’. Emerging Trends in Data Analytics Image courtesy SARK7 The automatic recognition of a set of Activities of Daily Living, is among the most challenging research fields in Ambient Intelligence. Main challenge with wearable technology is classification of use case ADL’
- 7. Data science toolkit What’ the problem statement? Prediction ? Classification? Linear regression K-means Hierarchical clustering Logistic regression Decision Trees Support vector machines Logistic regression Random forestDecision Trees Clustering Clustering ? Support vector machines SVM is a most popular and efficient classification and regression method. Currently four R packages contain SVM related software. For this project the e1071 R package was chosen that supports multi-level classification problems
- 8. The Dataset $Dataset for ADL Recognition with Wrist-worn Accelerometer • The Dataset for ADL Recognition with Wrist-worn Accelerometer is a public collection of labeled accelerometer data recordings to be used for the creation and validation of acceleration models of simple ADL . • It was provided by UCI machine learning repository, Center for Machine learning and Intelligent systems. • The data was collected by using single tri-axial accelerometer attached to the right-wrist of the volunteer. It was carried out by Barbara Bruno, Fulvio Mastrogiovanni, Antonio Sgorbissa from the Laboratory for Ambient Intelligence and Mobile Robotics ,DIBRIS, University of Genova • The Dataset composed of the recordings of 11 simple ADL performed by a total of 16 volunteers. Gender Age Weight M F Min Avg. Max Min Avg. Max 11 5 19 81 57.4 56 85 72.7 Problem statement $Data Data Wrangling Data Exploration Build Models Testing and classification brush_teeth getup_bed walk climb_stairs liedown_bed comb_hair pour_water descend_stairs sitdown_chair drink_glass standup_chair Wrist-worn Accelerometer Image courtesy: Chalkbeat Colorado
- 9. Overview of dataset Accelerometer-2011-04-11-13-28-18-brush_teeth-f1.txt - Refers to an accelerometer recording that was taken on March 11, 2011, starting from 13:28.18 p.m. - The recording refers to the HMP ”brush_teeth" executed by the volunteer with ID "f1". Provides equations to convert raw acceleration into real acceleration values and reduce noise Raw data from accelerometer for the task “brush_teeth”. The columns represents the acceleration measured in X,Y and Z direction. Each folder contains raw data for the activity.Each file in the dataset has the following naming convention: Accelerometer-[START_TIME]- [ADL]-[VOLUNTEER]
- 10. Feature extraction • A key challenge for any classification ML algorithm is feature extraction -.i.e. unique parameters that distinguishes each class (Activity) Ax Ay Az Ax Ay Az Drink glass Ax Ay Az Pour water Pour waterClimb stairs
- 11. Feature extraction approaches Group Methods Time domain Mean 1, Std. Deviation 2, Variance 3, MAD 4, Entropy 5 Frequency domain Fast Fourier transform 6, Discrete cosine transform 7 Other Principal component analysis 8, Linear discriminant analysis 9, Singular value decomposition • The ML algorithm implemented for this project uses Discrete cosine transformation and Singular value decomposition approaches to classify ADL. Data wrangling • Consolidate into a single CSV file. • Convert raw data into acceleration and apply filter • Calculate components of SVD: uvd • Sum vector magnitude (VM) =√(ax)2+(ay)2+(az)2 Data exploration • VM and SVD histograms • Scale data • Under-sample and Oversample data • Split data Build and select model • Build model in SVM using training data • Calculate values for Cost and Gamma • Cross validate model on training data. Validate model • Use test data to validate model. • F measure • Accuracy • Sensitivity • Selectivity
- 12. CONSOLIDATING INFORMATION • The original data comprises 479,288 observations of ax, ay, az, distributed among the 11 Activities • The data is consolidated to 827 observations that contribute to maximum variance in the data. § VM refers to sum of vector magnitude § SVD 1 and SVD2 refer to d1 and d2 components of Singular value decomposition § Act refers to the activities
- 13. What does the data look like? Problem stateme nt $Data Data Wrangling Data Exploration Build Models Testing and classificatio n Histogram of VM Histogram of SVD1 Histogram of SVD2 Two models are evaluated 1. Model 1 uses VD as the dependent variable to predict ADL 2. Model 2 uses VD and SVD to predict ADL
- 14. DATA SAMPLING • The original data set is unbalanced with fewer data points for activities such as brush_teeth, comb_hair, and liedown_bed • The SMOTE function is used to handle unbalanced classification problem. It generates synthetic data sets that addresses the class unbalance problem. • Following which the data set is split into training (70%) and test data (30%)
- 15. Data after balancing and scaling Histogram of VM Histogram of SVD1 Histogram of SVD2 Scaling avoids attributes in greater numeric ranges dominating those in smaller numeric ranges. Another advantage is it reduces numerical difficulties during the calculation
- 16. Tuning of SVM parameters- for Model 1(VD only) • RBF kernel is chosen as it can handle non- linear relations between class and attributes • Cost (C) and Gamma (γ) parameters are the two key parameters for the SVM model • Initial values of cost and gamma for the SVM model are estimated using the “tune function”, and running a coarse grid search. • Following the initial coarse grid search, the value of cost and gamma were fine tuned • The tune function implemented uses a 10 fold cross validation • Parameter tuning of ‘svm’: – Sampling method: 10-fold cross validation – Best parameters: Ø Gamma:20 Ø cost:900 – Best performance: 0.715 (1 – error) • The model performance is first evaluated on training data set, followed by test data. Error Gamma Cost Contour map showing the error rate for different cost and gamma values
- 17. Model performance ROC curves and Confusion matrix (CM) fpr tpr fpr tpr ROC and CM for train data set ROC and CM for test data set TRAINING CONFUSION MATRIX Reference Prediction brush_ teeth climb_ stairs comb_ hair Descend _stairs drink_ glass getup_ bed liedown_ bed pour_ water sitdown_ chair Standup _chair walk 85 brush_teeth 85 0 0 0 0 0 0 0 0 0 0 138 climb_stairs 0 118 0 13 5 9 2 7 0 1 0 47 comb_hair 0 0 42 0 0 0 0 0 0 0 0 60 descend_stairs 0 2 0 35 0 5 0 3 0 0 0 135 drink_glass 0 2 0 1 123 1 0 2 1 1 0 153 getup_bed 0 3 0 3 5 104 2 9 1 2 0 40 liedown_bed 0 2 0 0 1 2 35 0 0 0 0 147 pour_water 0 14 0 7 5 7 2 120 2 1 2 136 sitdown_chair 0 0 0 1 0 1 0 0 129 20 0 144 standup_chair 0 0 0 1 0 0 1 0 12 120 0 161 walk 0 0 4 0 0 2 0 0 0 0 164 TEST DATA SET CONFUSION MATRIX Reference Prediction brush_ teeth climb_ stairs comb_ hair descend_ stairs drink_ glass getup_ bed liedown_ bed pour_ water sitdown_ chair Standup _chair walk 36 brush_teeth 34 0 0 0 0 0 0 0 0 0 0 60 climb_stairs 0 42 0 4 2 5 0 6 0 0 1 19 comb_hair 0 0 14 0 2 0 0 0 0 0 1 26 descend_stairs 0 1 0 10 2 2 0 2 0 0 0 59 drink_glass 0 2 1 1 41 3 0 3 2 2 0 56 getup_bed 0 1 0 2 9 33 0 7 1 0 0 18 liedown_bed 0 4 0 0 0 3 15 1 2 0 0 60 pour_water 0 6 0 7 1 7 0 41 1 0 3 62 sitdown_chair 0 1 0 1 0 1 0 0 47 11 0 62 standup_chair 0 2 0 0 1 0 3 0 9 49 0 71 walk 2 1 4 1 1 2 0 0 0 0 66
- 18. VD model test summary • The initial results using VD’ to predict ADL were satisfactory, but showed deviation from training data set when implemented on test model. • The training model performance could be further improved by tuning the value of cost and gamma ,.i.e. high cost and gamma number. • Prior to implementing model 2, we need to check for collinearity (Strong correlation between two or more predictor variables). • The vif function is used to check for collinearity among the three variables
- 19. Tuning of SVM parameters: Model 2 (VD and SVM) • The initial value of cost and gamma for the SVM model are predicted using the “tune function”, and running a coarse grid search. • Following the initial coarse grid search, the value of cost and gamma were fine tuned • Each set of cost and gamma values are cross validated (10 fold) • Parameter tuning of ‘svm’: - Sampling method: 10-fold cross validation - Best parameters: - gamma:18 - cost:800 - best performance: 0.863 (1 – error) Error Gamma Cost
- 20. ROC curves and Confusion matrix (CM) TEST DATA SET Reference Prediction brush_ teeth climb_ stairs comb_ hair Descend _stairs drink_ glass getup_ bed liedown_ bed pour_ water sitdown_ chair Standup _chair walk 36 brush_teeth 35 0 0 0 0 0 0 0 0 0 0 58 climb_stairs 0 49 0 0 0 1 0 2 1 1 3 15 comb_hair 0 0 15 0 0 1 0 0 0 0 0 25 descend_stairs 0 1 0 23 0 0 0 0 0 1 0 57 drink_glass 0 0 0 0 56 3 0 7 0 0 0 65 getup_bed 0 0 0 0 1 53 0 3 2 0 1 16 liedown_bed 0 0 0 0 0 0 16 0 1 0 0 62 pour_water 0 0 0 0 0 0 0 46 0 0 0 57 sitdown_chair 0 0 0 0 0 0 0 0 66 0 0 61 standup_chair 0 0 0 1 0 0 0 0 3 47 0 69 walk 1 1 0 0 1 6 1 0 3 2 73 TRAINING DATA SET Reference Prediction brush_ teeth climb_ stairs comb_ hair descend_ stairs drink_ glass getup_ bed liedown_ bed pour_ water sitdown_ chair standup_c hair walk 85 brush_teeth 85 0 0 0 0 0 0 0 0 0 0 138 climb_stairs 0 152 0 0 0 0 0 0 0 0 0 47 comb_hair 0 0 50 0 0 0 0 0 0 0 0 60 descend_stairs 0 0 0 54 0 0 0 0 0 0 0 135 drink_glass 0 0 0 0 154 0 0 0 0 0 0 153 getup_bed 0 0 0 0 0 134 0 0 0 0 0 40 liedown_bed 0 0 0 0 0 0 36 0 0 0 0 147 pour_water 0 0 0 0 0 0 0 127 0 0 0 136 sitdown_chair 0 0 0 0 0 0 0 0 140 0 0 144 standup_chair 0 0 0 13 0 0 0 0 0 146 0 161 walk 0 0 0 0 0 0 0 0 0 0 153 fpr tpr fpr tpr ROC and CM for train data set ROC and CM for test data set
- 21. Summary • The results from the study demonstrate a improvised method for detecting ADL’ • The VD and SVM factors when used in conjunction help to classify activities that would otherwise be misclassified. • The study needs to extended to other ADL data sets that are publically available. • While publically available data sets will help tune and validate the model, the efficacy of the model can only be validated using data sets from the industry. • The application of a machine learning algorithm in this setting is quite vast – VR band – Health monitoring – Activity monitoring – Quality of activity, not just quantity – Remote point of care devices – Fall sensors
- 22. References 1. J. Parkka, M. Ermes, P. Korpipaa, J. Mantyjarvi, J. Peltola, and I. Korho- nen, “Activity classification using realistic data from wearable sensors,” IEEE Transactions on Information Technology in Biomedicine, vol. 10, no. 1, pp. 119–128, 2006. 2. E. M. Tapia, S. S. Intille, W. Haskell, K. Larson, J. Wright, A. King, and R. Friedman, “Real-time recognition of physical activities and their intensities using wireless accelerometers and a heart monitor,” in International Symposium on Wearable Computers, 2007. 3. O. D. Lara, A. J. Perez, M. A. Labrador, and J. D. Posada, “Centinela: A human activity recognition system based on acceleration and vital sign data,” Journal on Pervasive and Mobile Computing, 2011. 4. Y.-P. Chen, J.-Y. Yang, S.-N. Liou, Gwo-Yun=Lee, and J.-S. Wang, “Online classifier construction algorithm for human activity detection using a tri-axial accelerometer,” Applied Mathematics and Computation, vol. 205, no. 2, pp. 849–860, 2008. 5. M. Ermes, J. Parkka, and L. Cluitmans, “Advancing from offline to online activity recognition with wearable sensors,” in Engineering in Medicine and Biology Society. 30th Annual International Conference of the IEEE, pp. 4451–4454, 2008. 6. L. Bao and S. S. Intille, “Activity recognition from user-annotated acceleration data,” in Pervasive, pp. 1–17, 2004. 7. K. Altun and B. Barshan, “Human activity recognition using iner- tial/ magnetic sensor units,” in Human Behavior Understanding, Lecture Notes in Computer Science, pp. 38–51, Springer Berlin / Heidelberg, 2010. 8. Z. He and L. Jin, “Activity recognition from acceleration data based on discrete consine transform and svm,” in IEEE International Conference on Systems, Man and Cybernetics, pp. 5041–5044, 2009. 9. Z. He, Z. Liu, L. Jin, L.-X. Zhen, and J.-C. Huang, “Weightlessness feature; a novel feature for single tri-axial accelerometer based activity recognition,” in 19th International Conference on Pattern Recognition, pp. 1–4, 2008. 10. Y.-P. Chen, J.-Y. Yang, S.-N. Liou, Gwo-Yun=Lee, and J.-S. Wang, “Online classifier construction algorithm for human activity detection using a tri-axial accelerometer,” Applied Mathematics and Computation, vol. 205, no. 2, pp. 849–860, 2008. 11. Chih-Wei Hsu,Chih-Chung Chang, and Chih-Jen Lin, “A Practical Guide to Support Vector Classification”, Department of Computer Science,National Taiwan University, Taipei 106, Taiwan 12. Bradley, A. P. (1997). The use of the area under the ROC curve in the evaluation of machine learning algorithms, Pattern Recognition, 30 (7), 1145-1159. 13. Tobias Sing, Oliver Sander, Niko Beerenwinkel, Thomas Lengauer, ROCR: visualizing classifier performance in R.Bioinformatics 21(20): 3940-3941 (2005). 14. A Survey on Human Activity Recognition using Wearable Sensors, O Óscar D.Lara and Miguel A. Labrador, Department of Computer Science and Engineering University of South Florida, Tampa, FL 33620 15. David J. Hand and Robert J. Till (2001). A Simple Generalisation of the Area Under the ROC Curve for Multiple Class Classification Problems. Machine Learning 45(2), p. 171—186.