Your Smartphone Knows What You're Doing

A data science project that took a Kaggle dataset and created a logistic regression model that predicted with 98% accuracy one of six activities performed by a group of test participants.

I'm no data scientist nor do I play one on YouTube, but this was a great project to extend my knowledge in how data from gyroscopes and accelerometers can be used in the real world.

Transcript

1. YOUR SMARTPHONE KNOWS
   WHAT YOU’RE DOING
   C Todd Lombardo
   

   View Slide

2. OUTLINE
   ● Problem statement
   ● Performance summary
   ● Process
   ○ Exploratory data analysis
   ○ Model selection
   ○ Feature selection
   ○ Model tuning
   ○ Error analysis
   ○ Tried and failed
   ● Next steps
   

   View Slide

3. Problem Statement
   

   View Slide

4. Problem statement
   Hypothesis
   By examining accelerometer and gyroscope sensor data from a smartphone, a
   model can classify which activity was performed by a person
   Goals
   Accurately predict a human movement from accelerometer and gyroscope
   smartphone data, both provided in the data repo and captured by a mobile app.
   Risks and limitations
   Feature engineering may need to go beyond the scope of the course content. One
   is the possible exploration of a principal component analysis and other
   feature reduction techniques.
   

   View Slide

5. Performance Summary
   

   View Slide

6. Results: Classify by logistic regression
   Logistic Regression
   Accuracy = 0.98640 on the
   test data set
   With PCA, the 561 features
   could be reduced to 120
   principal components
   Sitting and Standing were
   the most difficult to
   discern: 20 false positives
   among them
   

   View Slide

7. Process
   

   View Slide

8. Target variable: Activity
   The model will need to accurately predict one
   of these movements:
   Walking
   Walking_upstairs
   Walking_downstairs
   Sitting
   Standing
   Laying
   dynamic
   static
   

   View Slide

9. Exploratory Data Analysis
   

   View Slide

10. About the dataset
    Human Activity Recognition w/Smartphone (Sources Kaggle, UC Irvine)
    1. Inertial sensor data
    a. Raw triaxial signals from the accelerometer & gyroscope of all
    the trials with participants
    b. The labels of all the performed activities
    2. Records of activity windows. Each one composed of:
    a. A 561-feature vector with time and frequency domain variables.
    b. Its associated activity label
    c. An identifier of the subject who carried out the experiment.
    The experiments were carried out with a group of 30 volunteers within an age bracket of 19-48 years. Each person performed six activities (WALKING,
    WALKING-UPSTAIRS, WALKING-DOWNSTAIRS, SITTING, STANDING, LAYING) wearing a smartphone (Samsung Galaxy S II) on the waist. Using its embedded accelerometer
    and gyroscope, we captured 3-axial linear acceleration and 3-axial angular velocity at a constant rate of 50Hz. The obtained dataset has been randomly
    partitioned into two sets, where 70% of the volunteers was selected for generating the training data and 30% the test data.
    

    View Slide

11. About the dataset
    Features will be selected and/or engineered from:
    Time-series signals
    ▸ tBodyAcc-XYZ
    ▸ tGravityAcc-XYZ
    ▸ tBodyAccJerk-XYZ
    ▸ tBodyGyro-XYZ
    ▸ tBodyGyroJerk-XYZ
    ▸ tBodyAccMag
    ▸ tGravityAccMag
    ▸ tBodyAccJerkMag
    ▸ tBodyGyroMag
    ▸ tBodyGyroJerkMag
    Fourier Transformed Signals
    ▸ fBodyAcc-XYZ
    ▸ fBodyAccJerk-XYZ
    ▸ fBodyGyro-XYZ
    ▸ fBodyAccMag
    ▸ fBodyAccJerkMag
    ▸ fBodyGyroMag
    ▸ fBodyGyroJerkMag
    Features are normalized and bounded within [-1,1].
    Each feature vector is a row on the text file. Some feature derivations are included in the dataset.
    The units used for the accelerations (total and body) are 'g's (gravity of earth -> 9.80665 m/seg2).
    The gyroscope units are rad/seg.
    

    View Slide

12. About the dataset
    Count of the Features
    

    View Slide

13. Data: Basic statistics
    hua.describe()
    

    View Slide

14. So about those correlations...
    

    View Slide

15. Data: By activity
    hua.groupby(Activity).count()
    

    View Slide

16. Which correlations matter?
    Feature_1 Feature_2 Correlation Abs_Corr
    tBodyAccJerk-energy()-X fBodyAccJerk-energy()-X 0.999999 0.999999
    fBodyAccJerk-energy()-X tBodyAccJerk-energy()-X 0.999999 0.999999
    fBodyAcc-bandsEnergy()-1,24 fBodyAcc-energy()-X 0.999878 0.999878
    fBodyAcc-energy()-X fBodyAcc-bandsEnergy()-1,24 0.999878 0.999878
    fBodyGyro-energy()-X fBodyGyro-bandsEnergy()-1,24 0.999767 0.999767
    fBodyGyro-bandsEnergy()-1,24 fBodyGyro-energy()-X 0.999767 0.999767
    fBodyAcc-bandsEnergy()-1,24.1 fBodyAcc-energy()-Y 0.999661 0.999661
    fBodyAcc-energy()-Y fBodyAcc-bandsEnergy()-1,24.1 0.999661 0.999661
    tBodyAccJerkMag-mean() tBodyAccJerk-sma() 0.999656 0.999656
    tBodyAccJerk-sma() tBodyAccJerkMag-mean() 0.999656 0.999656
    corr_values = hua[feature_cols].corr()
    corr_values[(corr_values.Correlation < 1.0) & (corr_values.Correlation > 0.9)]
    

    View Slide

17. What do the activities look like for one subject?
    

    View Slide

18. What do the activities look like for all subjects?
    

    View Slide

19. What does one activity look like for different subjects?
    

    View Slide

20. Can we separate static and dynamic activities?
    Yes ——> If tBodyAccMag < -0.5 then it’s a static activity
    If tBodyAccMag > -0.5 then it’s a dynamic activity
    dynamic
    static
    

    View Slide

21. Is it possible to separate specific activities?
    

    View Slide

22. Is it possible to separate all the activities? (t-SNE)
    

    View Slide

23. Model Selection
    

    View Slide

24. Model Selection: “Naive” (No feature selection, no tuning)
    Algorithm Train Cross-Validation Score Test Accuracy Score
    DecisionTreeClassifier 0.840870 0.852392
    RandomForestClassifier 0.917985 0.922973
    KNNeighborsClassifier 0.897175 0.900238
    LogisticRegression 0.933495 0.957923
    Naive running is fitting models with no feature selection, tuning, or optimization. Yes, I ran all 561 features. And yes, it took a while.
    Must be classification algorithm, ran four
    

    View Slide

25. Feature Selection
    

    View Slide

26. Principal Component Analysis: Visualization of PC1 & PC2
    

    View Slide

27. PCA
    

    View Slide

28. PCA – Project Back to the Features
    

    View Slide

29. Map back to actual features?
    Signal Value
    angle(tBodyGyroJerkMean,gravityMean) 0.479798
    tBodyAccJerk-mean()-X 0.201355
    tGravityAcc-energy()-Y 0.115365
    fBodyAcc-kurtosis()-Z 0.098975
    fBodyAcc-skewness()-Z 0.098076
    tGravityAcc-correlation()-X,Y 0.085849
    angle(tBodyGyroMean,gravityMean) 0.075399
    angle(Z,gravityMean) 0.074471
    tGravityAcc-min()-Y 0.059661
    tGravityAcc-mean()-Y 0.055262
    pd.Series(pc1, index=hua_pca.columns).sort_values(ascending=False)
    Which are the most
    predictive signals?
    

    View Slide

30. This matches up to the corr heatmap...
    

    View Slide

31. Model Tuning
    

    View Slide

32. How many PCA features needed to increase accuracy?
    Plotting all the accuracy
    scores by increasing the
    number of PCA variables in
    the model
    

    View Slide

33. Which are the most important PCA Features?
    coef variable abscoef
    6.015424 PC3 6.015424
    3.503071 PC22 3.503071
    2.330216 PC48 2.330216
    1.976726 PC32 1.976726
    -1.844982 PC26 1.844982
    1.678715 PC23 1.678715
    1.291059 PC45 1.291059
    1.269009 PC13 1.269009
    -1.153241 PC6 1.153241
    1.136382 PC24 1.136382
    coefs_vars.sort_values('abscoef', ascending=False, inplace=True)
    Which are the most
    important PCA
    Features?
    

    View Slide

34. Error Analysis
    

    View Slide

35. Confusion Matrix
    120 PCA features
    With 20 false positives
    

    View Slide

36. Tried and Failed
    

    View Slide

37. These didn’t work so well
    1. Scaling the data with StandardScaler: Accuracy dropped to 0.911
    (“features are normalized”)
    2. Further feature reduction of PCA or Signal features
    a. a challenge to find specific signals, it seems the
    combination of signals is far more useful
    b. even when using top 20 principal components
    3. Ridge — struggled to get this to work properly
    

    View Slide

38. Next Steps
    

    View Slide

39. With more time
    1. Horn's parallel analysis
    2. Optimize Lasso and Ridge, would a penalty improve the model?
    3. Dig further into t-SNE for feature selection & reduction
    4. Run experiment with my own phone
    a. Perform the six activities with a sensor recorder app
    b. Process the raw data, run the model and score it
    5. Try this with data from a smartwatch!
    

    View Slide

40. View Slide

41. References & credits
    https://www.kaggle.com/uciml/human-activity-recognition-with-smartphones
    http://www.cis.fordham.edu/wisdm/public_files/sensorKDD-2010.pdf
    https://upcommons.upc.edu/bitstream/handle/2117/101769/IWAAL2012.pdf
    https://www.elen.ucl.ac.be/Proceedings/esann/esannpdf/es2013-11.pdf
    https://github.com/anas337/Human-Activity-Recognition-Using-Smartphones
    https://github.com/markdregan/K-Nearest-Neighbors-with-Dynamic-Time-Warping
    http://www.cis.fordham.edu/wisdm/dataset.php
    With code lovingly borrowed from many GA Jupyter and Kaggle notebooks
    

    View Slide