an automated prompting system for smart environments
TRANSCRIPT
![Page 1: An Automated Prompting System for Smart Environments](https://reader034.vdocuments.mx/reader034/viewer/2022052622/558cc711d8b42a0c7c8b461b/html5/thumbnails/1.jpg)
An Automated Prompting System for Smart Environments
Presented by:
Barnan DasCenter for Advanced Studies in Adaptive Systems (CASAS)
Washington State University
June 21, 2011
Barnan Das, Chao Chen, Adriana M. Seelye, Diane J. Cook
![Page 2: An Automated Prompting System for Smart Environments](https://reader034.vdocuments.mx/reader034/viewer/2022052622/558cc711d8b42a0c7c8b461b/html5/thumbnails/2.jpg)
2
By 2030, 20% of US population will be >65 years of age
![Page 3: An Automated Prompting System for Smart Environments](https://reader034.vdocuments.mx/reader034/viewer/2022052622/558cc711d8b42a0c7c8b461b/html5/thumbnails/3.jpg)
3
Avg. cost of skilled nursing home care: ~$70,000/person/year.
![Page 4: An Automated Prompting System for Smart Environments](https://reader034.vdocuments.mx/reader034/viewer/2022052622/558cc711d8b42a0c7c8b461b/html5/thumbnails/4.jpg)
4
Innovative health care technology can help sustain independent lifestyle.
![Page 5: An Automated Prompting System for Smart Environments](https://reader034.vdocuments.mx/reader034/viewer/2022052622/558cc711d8b42a0c7c8b461b/html5/thumbnails/5.jpg)
5
“Prompting Systems”
![Page 6: An Automated Prompting System for Smart Environments](https://reader034.vdocuments.mx/reader034/viewer/2022052622/558cc711d8b42a0c7c8b461b/html5/thumbnails/6.jpg)
The Problem
6
Please turn off the burner.Please turn off the burner.
You just picked up the wrong vessel.You just picked up the wrong vessel.
It would be a good idea to take a walk.It would be a good idea to take a walk.
You look tired, why don’t you take a nap.You look tired, why don’t you take a nap.
Sugar is in the cupboard.Sugar is in the cupboard.
Sam is trying to get in touch with you.Sam is trying to get in touch with you.
Its John’s birthday, you wanna write a card?Its John’s birthday, you wanna write a card?
Please take a look at the Wattage of the light bulb.Please take a look at the Wattage of the light bulb.
Its time to take medicine.Its time to take medicine.
Sugar is in the cupboard.Sugar is in the cupboard.
Sam is trying to get in touch with you.Sam is trying to get in touch with you.
Automatic delivery of verbal or non-verbal interventions that would help a smart home inhabitant in successful completion of daily tasks.
Its time to take medicine.Its time to take medicine.
![Page 7: An Automated Prompting System for Smart Environments](https://reader034.vdocuments.mx/reader034/viewer/2022052622/558cc711d8b42a0c7c8b461b/html5/thumbnails/7.jpg)
Our Solution
7
PUCK
Prompting Users and Control Kiosk Under development at
CASAS, WSU
Automated Prompting System Based on Supervised
Learning
![Page 8: An Automated Prompting System for Smart Environments](https://reader034.vdocuments.mx/reader034/viewer/2022052622/558cc711d8b42a0c7c8b461b/html5/thumbnails/8.jpg)
System Architecture
8Figure 1: System Architecture of PUCK
![Page 9: An Automated Prompting System for Smart Environments](https://reader034.vdocuments.mx/reader034/viewer/2022052622/558cc711d8b42a0c7c8b461b/html5/thumbnails/9.jpg)
Experimental Setup
9
• Testbed: 2 story apartment in WSU campus
• Sensors: Motion, door, object, temperature, power
• Participants: 128 older adults with mild cognitive disorder
• Activities: Sweeping, Medication, Writing birthday card, Watching DVD, Water plants, Phone call, Cooking, Selecting outfit
• Activities are subdivided into steps.
• Activities monitored via web cam. Experimenter remotely plays (in)direct audio/video cues when an error is detected.
• Human annotators annotate datasets for activities and activity steps.
![Page 10: An Automated Prompting System for Smart Environments](https://reader034.vdocuments.mx/reader034/viewer/2022052622/558cc711d8b42a0c7c8b461b/html5/thumbnails/10.jpg)
Feature Generation
10
• Generate features considering each step of an activity as an instance in the training dataset.
Feature # Feature Name Description
1 stepLength Length of the step in time (seconds)
2 numSensors Number of unique sensors involved with the step
3 numEvents Number of sensor events associated with the step
4 prevStep Previous step
5 nextStep Next step
6 timeActBegin Time (seconds) elapsed since the beginning of the activity
7 timePrevStep Time (seconds) difference between the last event of the previous step and the first event of the current step
8 stepsActBegin Number of steps visited since the beginning of the activity
9 activityID Activity ID
10 stepID Step ID
11 M01 … M51 All of M01 to M51 are individual features denoting the frequency of firing of these sensors associated with the step
12 Class Binary class. 1-”Prompt”, 0-”No-Prompt”
![Page 11: An Automated Prompting System for Smart Environments](https://reader034.vdocuments.mx/reader034/viewer/2022052622/558cc711d8b42a0c7c8b461b/html5/thumbnails/11.jpg)
Experimentation
11
• Training and Testing: 10-fold Cross Validation
• Classifiers used:• Decision Tree (J48)• k – Nearest Neighbor (IBk)• Support Vector Machines (SMO)
• Performance Metrics:• True Positive Rate (TP Rate)• True Negative Rate (TN Rate)• Area Under ROC Curve (AUC)• Accuracy (Acc)
![Page 12: An Automated Prompting System for Smart Environments](https://reader034.vdocuments.mx/reader034/viewer/2022052622/558cc711d8b42a0c7c8b461b/html5/thumbnails/12.jpg)
Performance of Baseline Classifiers
12
Figure 3: Accuracy Performance for Baseline Classifiers Figure 4: TP and TN Rates for Baseline Classifiers
![Page 13: An Automated Prompting System for Smart Environments](https://reader034.vdocuments.mx/reader034/viewer/2022052622/558cc711d8b42a0c7c8b461b/html5/thumbnails/13.jpg)
Failure of Baseline Classifiers
13
Problem: Highly imbalanced class distribution.
Cause: Vast majority of training situations do not require prompts.
Total # unique steps: 53# steps recognizable by annotators: 38# prompt instances: 149 (3.74% of total # of instances)
![Page 14: An Automated Prompting System for Smart Environments](https://reader034.vdocuments.mx/reader034/viewer/2022052622/558cc711d8b42a0c7c8b461b/html5/thumbnails/14.jpg)
Handling Imbalanced Class Distribution
14
Under-sampling Over-sampling
Imbalanced DataSampling: Rebalancing the dataset
![Page 15: An Automated Prompting System for Smart Environments](https://reader034.vdocuments.mx/reader034/viewer/2022052622/558cc711d8b42a0c7c8b461b/html5/thumbnails/15.jpg)
SMOTE
15
Boosting prompt situations in the training set without under/over representation.
Technique: Synthetic Minority Over-sampling Technique or SMOTE.
Over-samplingi. Compute the difference between the feature vector (sample) under consideration(belonging to minority class) and its nearest neighbor (which is also assumed to belong to the minority class).ii. Multiply this difference by a random number between 0 and 1.iii. Add the product to the feature vector under consideration.
Under-samplingRandom under-sampling
![Page 16: An Automated Prompting System for Smart Environments](https://reader034.vdocuments.mx/reader034/viewer/2022052622/558cc711d8b42a0c7c8b461b/html5/thumbnails/16.jpg)
SMOTE-Variant
16
Why can’t we use SMOTE directly?
• Minority class instances are small in absolute number (149 in our case).• No nearest neighbor with same step of an activity in some cases.
SMOTE-Variant:i.Randomly pick a minority class instance.ii.Consider activityID and stepID to find nearest neighbor.iii.Randomly choose any one nearest neighbor.iv.Synthesize new data point in the same way as SMOTE.
![Page 17: An Automated Prompting System for Smart Environments](https://reader034.vdocuments.mx/reader034/viewer/2022052622/558cc711d8b42a0c7c8b461b/html5/thumbnails/17.jpg)
Sampling
17
What is the ideal class distribution?Vary the percentage of minority class from 5-95% and test its performance using J48 Decision Tree.
Figure 5: Effect of Class Distribution
![Page 18: An Automated Prompting System for Smart Environments](https://reader034.vdocuments.mx/reader034/viewer/2022052622/558cc711d8b42a0c7c8b461b/html5/thumbnails/18.jpg)
Results
18
Figure 3: Comparison of TP Rate Figure 4: Comparison of TN Rate
![Page 19: An Automated Prompting System for Smart Environments](https://reader034.vdocuments.mx/reader034/viewer/2022052622/558cc711d8b42a0c7c8b461b/html5/thumbnails/19.jpg)
Results
19
Figure 3: Comparison of AUC
Classifier Baseline Sampled
J-48 96.206 91.55
K-NN 94.2965 92.05
SMO 96.2312 91.35
Table 1: Comparison of Accuracy
![Page 20: An Automated Prompting System for Smart Environments](https://reader034.vdocuments.mx/reader034/viewer/2022052622/558cc711d8b42a0c7c8b461b/html5/thumbnails/20.jpg)
Conclusion
20
• Introduced PUCK
• Discussed framework in which the prompting system is developed.
• Challenges of using supervised machine learning methods without pre-processing.
• Proposed a variant of an existing sampling method.
![Page 22: An Automated Prompting System for Smart Environments](https://reader034.vdocuments.mx/reader034/viewer/2022052622/558cc711d8b42a0c7c8b461b/html5/thumbnails/22.jpg)
22