The document presents a prototype of a context and user aware smart notification system. The goal is to develop a system that can filter incoming notifications using machine learning based on notification information, environment status, user context, and user habits. The proposed architecture is modular and aware of these factors. Preliminary results show support vector machines and decision trees had the best performance at predicting the correct notification delivery method, with support vector machines being the most accurate but slowest. Future work includes collecting real notification data and evaluating additional machine learning algorithms.

1. A Context and User Aware
Smart Notification
System
Fulvio Corno
Luigi De Russis
Teodoro Montanaro*
http://jol.telecomitalia.com/j
olswarm/
http://elite.polito.it/

2. 2
Outline
1. Context and Motivation
2. Goal
3. Architecture
4. Prototype
5. Preliminary results
6. Conclusion
7. Future work

3. 3
Context
Context
Infographic from "The Connectivist": growing of IoT connected devices
(http://www.theconnectivist.com/2014/05/infographic-the-growth-of-the-internet-of-things/)

4. 4
Motivation
Motivation

5. 5
Motivation
Motivation
IoT devices can generate, receive and show different kinds of notifications

6. 6
Motivation
Motivation
IoT devices can generate, receive and show different kinds of notifications

7. 7
Motivation
Motivation
IoT devices can generate, receive and show different kinds of notifications

8. 8
Motivation
Motivation
IoT devices can generate, receive and show different kinds of notifications
The number of notifications is growing

9. 9
Motivation
Motivation
IoT devices can generate, receive and show different kinds of notifications
The number of notifications is growing

10. 10
Motivation
Motivation
IoT devices can generate, receive and show different kinds of notifications
Nowadays the same notification is replicated on all available devices
The number of notifications is growing

11. 11
Motivation
Motivation
IoT devices can generate, receive and show different kinds of notifications
Nowadays the same notification is replicated on all available devices
The number of notifications is growing

12. 12
Motivation
Motivation
IoT devices can generate, receive and show different kinds of notifications
Nowadays the same notification is replicated on all available devices
The number of notifications is growing
The benefit of displaying the same
notification on all available devices
could put user patience to a hard test

13. 13
Analyze how machine learning approach can improve
IoT notification user experience
Goal
Goal

14. 14
Analyze how machine learning approach can improve
IoT notification user experience
Goal
Goal

15. 15
Analyze how machine learning approach can improve
IoT notification user experience
Goal
Goal
Develop a system able to filter incoming notifications
depending on:
• Notification information
• Environment status
• User context
• User habits

16. 16
Analyze how machine learning approach can improve
IoT notification user experience
Goal
Goal
Develop a system able to filter incoming notifications
depending on:
• Notification information
• Environment status
• User context
• User habits

17. 17
Analyze how machine learning approach can improve
IoT notification user experience
Goal
Goal
Develop a system able to filter incoming notifications
depending on:
• Notification information
• Environment status
• User context
• User habits
Evaluate machine learning approach

18. 18
We propose:
Architecture
Architecture

19. 19
We propose: A modular architecture
Architecture
Architecture

20. 20
We propose: A modular architecture
Architecture
Architecture

21. 21
We propose: A modular architecture
Architecture
Architecture

22. 22
We propose: A modular architecture
aware of
Architecture
Architecture

23. 23
We propose:
Environment status
(e.g., weather information,
current date and time)
A modular architecture
aware of
Architecture
Architecture

24. 24
We propose:
Environment status
(e.g., weather information,
current date and time)
User context (e.g.,
location, status, current
activity),
A modular architecture
aware of
Architecture
Architecture

25. 25
We propose:
Environment status
(e.g., weather information,
current date and time)
User context (e.g.,
location, status, current
activity),
User habits
A modular architecture
aware of
Architecture
Architecture

26. 26
We propose: A modular architecture
Architecture
Architecture

27. 27
We propose:
Decision maker: Machine Learning
algorithm makes decisions (best devices
+ best modes + best moment).
Architecture
Architecture

28. 28
Architecture: example
Architecture

29. 29
Architecture: example
Architecture
Mario is in a
meeting

30. 30
Architecture: example
Architecture
Mario is in a
meeting

31. 31
Architecture: example
Architecture
Every meeting lasts at least 2
hours
Mario is in a
meeting

32. 32
Architecture: example
Architecture
Every meeting lasts at least 2
hours
Mario is in a
meeting

33. 33
Architecture: example
Architecture
Time: 17:00
Meeting started at
16:00
Wife is at home
Every meeting lasts at least 2
hours
Mario is in a
meeting

34. 34
Architecture: example
Architecture
Time: 17:00
Meeting started at
16:00
Wife is at home
Every meeting lasts at least 2
hours
Mario is in a
meeting

35. 35
Architecture: example
Architecture
Notification:the toilet
paper has just finished
Time: 17:00
Meeting started at
16:00
Wife is at home
Every meeting lasts at least 2
hours
Mario is in a
meeting

36. 36
Architecture: example
Architecture
Notification:the toilet
paper has just finished
Time: 17:00
Meeting started at
16:00
Wife is at home
Every meeting lasts at least 2
hours
Mario is in a
meeting

37. 37
Architecture: example
Architecture
Notification:the toilet
paper has just finished
Time: 17:00
Meeting started at
16:00
Wife is at home
Every meeting lasts at least 2
hours
Mario is in a
meeting

38. 38
Architecture: example
Architecture
Notification:the toilet
paper has just finished
Notify:
• At 18:10
• on his personal
smartphone
• Sound
Time: 17:00
Meeting started at
16:00
Wife is at home
Every meeting lasts at least 2
hours
Mario is in a
meeting

39. 39
Architecture: example
Architecture
Notification:the toilet
paper has just finished
Notify:
• At 18:10
• on his personal
smartphone
• Sound
Time: 17:00
Meeting started at
16:00
Wife is at home
Every meeting lasts at least 2
hours
Mario is in a
meeting

40. 40
Architecture: example
Architecture
Notification:the toilet
paper has just finished
Notify:
• At 18:10
• on his personal
smartphone
• Sound
Time: 17:00
Meeting started at
16:00
Wife is at home
Every meeting lasts at least 2
hours
Mario is in a
meeting

41. 41Prototype
Prototype implementation

42. 42Prototype
Prototype implementation
Aim: evaluate machine learning approach to decide
• who should receive an incoming notification;
• the best moment to show the notification;
• the best device(s)
• the best mode to notify the incoming notification

43. 43Prototype
Prototype implementation
Aim: evaluate machine learning approach to decide
• who should receive an incoming notification;
• the best moment to show the notification;
• the best device(s)
• the best mode to notify the incoming notification

44. 44Prototype
Prototype implementation
Preliminary version of
Aim: evaluate machine learning approach to decide
• who should receive an incoming notification;
• the best moment to show the notification;
• the best device(s)
• the best mode to notify the incoming notification

45. 45Prototype
Prototype implementation
Preliminary version of
Aim: evaluate machine learning approach to decide
• who should receive an incoming notification;
• the best moment to show the notification;
• the best device(s)
• the best mode to notify the incoming notification

46. 46Prototype
Prototype implementation

47. 47Prototype
Prototype implementation
Preliminary version of

48. 48Prototype
Prototype implementation
Preliminary version of
Dataset

49. 49Prototype
Prototype implementation
Preliminary version of
Dataset Algorithms

50. 50
Prototype implementation
Prototype
Dataset

51. 51
Prototype implementation
Prototype
94 people over 9 months
monitored through
smartphones in 2004:
• Sender
• Receiver
• Type of notification
• Timestamp of receipt
• User current location
Dataset

52. 52
Prototype implementation
Prototype
94 people over 9 months
monitored through
smartphones in 2004:
• Sender
• Receiver
• Type of notification
• Timestamp of receipt
• User current location
Dataset
Synthetic data:
• User current
activity
• Available devices
for the user
• Target device.

53. 53
Prototype implementation
Prototype
94 people over 9 months
monitored through
smartphones in 2004:
• Sender
• Receiver
• Type of notification
• Timestamp of receipt
• User current location
Dataset
Synthetic data:
• User current
activity
• Available devices
for the user
• Target device.
Real + synthetic dataset:
165,289 samples, almost one per
each hour of the day
(the missing samples are related
to hours in which users turned
off their smartphones)

54. 54
Prototype implementation
Prototype
94 people over 9 months
monitored through
smartphones in 2004:
• Sender
• Receiver
• Type of notification
• Timestamp of receipt
• User current location
Dataset
Synthetic data:
• User current
activity
• Available devices
for the user
• Target device.
Real + synthetic dataset:
165,289 samples, almost one per
each hour of the day
(the missing samples are related
to hours in which users turned
off their smartphones)
Information collected by Decision Maker in previous example
{
“notification“:{
“senderName“:“mySmartHome“,
“type“:“smart Home Notification“,
“receiptTimestamp“:“1447347600“
},
“userStatus“: {
“senderId“:“359“,
“currentActivity“:“STILL“,
“currentActivityConfidence“:“50%“,
“availableDevices”:[“deviceId”:”23”]
},
“deviceStatus“:{
“deviceId“:”23”,
“category“:”Smartphone”,
“currentStatus“:”On”,
“currentMode“:”Ring”,
“wifiStatus“:” Connected through MOBILE”,
“batteryLevel“:” 57%”,
“batteryStatus“:”BATTERY_STATUS_NOT_CHARGING”
}
}

55. 55
Prototype implementation
Prototype
Machine learning algorithms
Dataset
Real + synthetic data (165,289
samples)

56. 56
Prototype implementation
Prototype
Machine learning algorithms
Dataset
Real + synthetic data (165,289
samples)

57. 57
Prototype implementation
Prototype
Machine learning algorithms
Dataset
Real + synthetic data (165,289
samples)

58. 58
Prototype implementation
Prototype
Machine learning algorithms
Dataset
Real + synthetic data (165,289
samples)
Training dataset: 80% of data
Tests dataset: 20% of data

59. 59
Prototype implementation
Prototype
Simplified version of the Decision
maker:
• only one device as receiver;
• only one available mode for each
device;
• no decision about the best time
to deliver the notification;
• not aware of environment
context
Machine learning algorithms
Dataset
Real + synthetic data (165,289
samples)
Training dataset: 80% of data
Tests dataset: 20% of data

60. 60
Prototype implementation
Prototype
Machine learning algorithms
Dataset
Real + synthetic data (165,289
samples)
Training dataset: 80% of data
Tests dataset: 20% of data

61. 61
Prototype implementation
Prototype
Machine learning algorithms
Dataset
Real + synthetic data (165,289
samples)
Training dataset: 80% of data
Tests dataset: 20% of data

62. 62
Prototype implementation
Prototype
Three machine learning
algorithms:
1. SupportVector Machine
2. Gaussian Naïve Bayes
3. Decision Trees.
Machine learning algorithms
Dataset
Real + synthetic data (165,289
samples)
Training dataset: 80% of data
Tests dataset: 20% of data

63. 63
Preliminary results
Preliminary results

64. 64
Preliminary results
Preliminary results
96,10%
83,40%
93,90%
0,00%
20,00%
40,00%
60,00%
80,00%
100,00%
Support
Vector
Machine
Gaussian
Naive Bayes
Decision Trees
Percentage of corrected predictions obtained
with used algorithms

65. 65
Preliminary results
Preliminary results
CPU time (in seconds) for a training phase with
33058 samples
96,10%
83,40%
93,90%
0,00%
20,00%
40,00%
60,00%
80,00%
100,00%
Support
Vector
Machine
Gaussian
Naive Bayes
Decision Trees
5801,1
12,9 13,9
1
10
100
1000
10000
Support
Vector
Machine
Gaussian
Naive Bayes
Decision Trees
Percentage of corrected predictions obtained
with used algorithms

66. 66
Preliminary results
Preliminary results
Average CPU time (in milliseconds) for each
notification classification
SupportVector Machine 40,22 ms
Gaussian Naive Bayes 0,31 ms
Decision Trees 0,001 ms
96,10%
83,40%
93,90%
0,00%
20,00%
40,00%
60,00%
80,00%
100,00%
Support
Vector
Machine
Gaussian
Naive Bayes
Decision Trees
Percentage of corrected predictions obtained
with used algorithms
CPU time (in seconds) for a training phase with
33058 samples
5801,1
12,9 13,9
1
10
100
1000
10000
Support
Vector
Machine
Gaussian
Naive Bayes
Decision Trees

67. 67
Conclusion
Obtained results demonstrated that our system uses a promising technique to
manage the problem of overwhelming notifications.
Specifically, the machine learning approach was tested through 3 different
algorithms and SVM and DT seem to be the most promising one.
Conclusion
Future work:
Define a new dataset to include all the needed real information
Development of a system to collect real data and real notifications
Careful evaluation of the machine learning algorithms
Enhancement of prototype to include unconsidered blocks

68. 68
Thank you
Future work
Notification Collector (beta):
Android app to collect real data
https://goo.gl/pLMWSG
To contribute: download it!
Requirement: Android 5 (Lollipop)
We collect (anonymously):
• Incoming notification info (no
content)
• User current activity
• User current location
• Device status
• User feedback