Download Sequence Mining Application in Smart Homes

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts
no text concepts found
Transcript
CS 548 – Spring 2016
Sequence Mining
Showcase by:
Daniel Duhaney, Scott Judson, Michaela Kachadoorian
Showcasing work of:
Daniel Schweizer, Michael Zehnder, Holger Wache, Hans-Friedrich Witschel, Danilo
Zanatta, Miguel Rodriguez
on
“Using consumer behavior data to reduce energy
consumption in smart homes”
References
1.
Daniel Schweizer, Michael Zehnder, Holger Wache, Hans-Friedrich Witschel, et al.
“Using consumer behavior data to reduce energy consumption in smart homes.” In
2015 IEEE 14th International Conference on Machine Learning and Applications
(ICMLA), pp. 1123-1129. IEEE, 2015.
2.
King, N. “Smart home. A definition” Intertek Research and Testing Center, pp. 1-6,
2003.
3.
C. Baumann. "Smart energy case study." In Proceedings of the Fourth ACM
Workshop on Embedded Sensing Systems for Energy-Efficiency in Buildings, pp. 3638. ACM, 2012.
4.
D. Schweizer. "Learning frequent and periodic usage patterns in smart homes.“
Master’s thesis, School of Business, University of Applied Sciences and Arts
Nortwestern Switzerland (FHNW) Feb, 2014 https://www.digitalstrom.org/wpcontent/uploads/2014/01/Daniel_Schweizer-2014Learning_frequent_and_periodic_usage_patterns_in_smart_homes_Final.pdf
5.
M. Zehnder. "Energy saving in smart homes based on consumer behaviour data."
Master’s thesis, School of Business, University of Applied Sciences and Arts
Nortwestern Switzerland (FHNW) Jan, 2015.https://www.digitalstrom.org/wpcontent/uploads/2014/01/Michael-Zehnder-2015-Energy-saving-in-smart-homesbased-on-consumer-behavior-data.pdf.
6.
M Zehnder. "Energy saving in smart homes based on consumer behavior: A case
study." In Smart Cities Conference (ISC2), 2015 IEEE First International, pp. 1-6.
IEEE, 2015
http://ieeexplore.ieee.org/xpl/login.jsp?tp=&arnumber=7366231&url=http%3A%2F
%2Fieeexplore.ieee.org%2Fxpls%2Fabs_all.jsp%3Farnumber%3D7366231
7.
Nizar R. Mabroukeh and C. I. Ezeife. “A taxonomy of sequential pattern mining
algorithms.” ACM Comput. Surv. Vol. 43, No. 1, Article 3. December 2010.
http://doi.acm.org/10.1145/1824795.1824798
Worcester Polytechnic Institute
Smart Home
• A “dwelling incorporating a communications
network that connects the key electrical
appliances and services, and allows them to be
remotely controlled, monitored or accessed” [1].
• One of the major benefits of a smart home is the
ability to maximize the efficiency of energy
consumption.
Worcester Polytechnic Institute
http://domoticawinkel.nl/media/catalog/product/cache/1/image/512x512/9df78eab33525d08d6e5fb8d27136e95/d/i/digitalst
rom.apps_and_the_connected_home_2__2_3_1.jpg
Worcester Polytechnic Institute
Sequence mining consumer behavior to
improve energy conservation
M. Zehnder, et al. "Energy saving in smart homes based on consumer behaviour data." PhD diss., Master’s
thesis, School of Business, University of Applied Sciences and Arts Nortwestern Switzerland (FHNW) Jan, 2015.
Worcester Polytechnic Institute
Training data set
• Previous data collected between 12/8/02-6/25/14
─33 homes
─ 3521 devices in the 33 homes
─ 4,331,443 total events in all 33 homes
─ 6829 unique events or scenes
5
1
3
Events 5 + 1 + 3 = One Pattern
Worcester Polytechnic Institute
Algorithm Criteria
• Find both frequent sequential patterns and
periodic sequential patterns
• Find wildcarded patterns and output where the
wildcard is positioned in the pattern
• Process the continuous stream of data coming
from a smart home (process events in real time)
Worcester Polytechnic Institute
Window Sliding with De-Duplication
(WSDD)
• Window size = 5 events
2002
2014
Adapted from Figure 8: The difference between overlapping and non-overlapping patterns by D.
Schweizer, et al. "Learning frequent and periodic usage patterns in smart homes."
Worcester Polytechnic Institute
Window Sliding with De-Duplication
(WSDD)
• Brute Force Method
─ Loop 1: Build possible patterns
─ Loop 2: Count the support
• Improved brute force method using hash tree & post-pruning
HashMap-Keys
HashMap-Values
513
0.14
128
0.21
278
0.19
─ Post-processing to eliminate infrequent patterns that do not constitute
normal behavior
Worcester Polytechnic Institute
WSDD basic pattern mining algorithm
smartHomeList = [sh0,sh1,…shi]
eventi = {ev0,ev1,…evj} where evij = {startTime, endTime, sourceID, sceneID}
minPatternLength, maxPatternLength, minSupportCount defined by user
For each smartHome in smartHomeList
download all of the events for sh
sort events by startTime and eventID
lastPosition = number of events in database for smartHomeID
for position for range(startPosition, lastPosition)
for patternLength in range(minPatternLength, maxPatternLength
patternk = ev0 + ev1 + … +evcurrentLength
hash(patternk):
if hash(patternk) exists: increment supportCountk in hash tree
else: supportCountk = 1 in hash tree
patternList for smartHome = patterns where supportCount > minSupportCount
Worcester Polytechnic Institute
Algorithm Results
Learning frequent and periodic usage patterns in smart homes
Conclusion
• WSDD is a competitive algorithm when compared
to other sequential pattern mining algorithms
• WSDD has good run times due to relatively small
number of different patterns in a smart home
• Wildcarding is not necessary for smart home event
datasets
Figure 2. Benchmark of run times for data mining algorithms
FIGURE 61: ADJUSTED RUN TIME BENCHMARK FOR MINSUP=0.005, OVERLAP=TRUE, LENGTH=2-5
As can be seen, WSDD is the best choice regarding run times with those settings. This corresponds with
D. Schweizer, et al. "Using consumer behavior data to reduce energy consumption in
all the other benchmarks conducted for this research project and from a run
time point
of view
it can
smart homes." arXiv preprint arXiv:1510.00165 (2015).[To be presented at IEEE
Worcester
Polytechnic
Institute
International
of Machine
and Applications
(ICMLA,like
Dec.PrefixSpan
2015)
therefore Conference
be concluded
that Learning
using elaborate
algorithms
or GapBIDE is not only an
The architecture of the recommender system developed in this project (illustrated in Figure 26) can be
divided in tree main parts:
The storage of the association rules.
Recommender
System
The event stream of the current behaviour data inside the smart home.
The matching algorithm for both previous points.
Figure 26 Architecture of the recommender system
The rule database stores the association rules, which were deduced from the relevant pattern.
Polytechnic Institute
occurrence.
time of their
The event stream contains the current events from the smart home, ordered byWorcester
M. Zehnder, et al. "Energy saving in smart homes based on consumer behaviour data." PhD diss., Master’s thesis, School of
Business, University of Applied Sciences and Arts Nortwestern Switzerland (FHNW) Jan, 2015.
Design of the recommender system
Recommender-Finite State Machine
Figure 27 Example for the state machine created from an association rule
The example for the state machine in Figure 27 is for an association rule deduced from a pattern, where
M. Zehnder, et al. "Energy saving in smart homes based on consumer behaviour data." PhD diss., Master’s thesis, School of Worcester Polytechnic Institute
the action
was
at Sciences
the end
of aNortwestern
pattern.
Therefore,
the2015.
last condition before sending the recommendation
Business,
University of
Applied
and Arts
Switzerland
(FHNW) Jan,
m (FS
SM)
houses agreed to participate in
i the evaluation of this work
including both single- and
a
multi-inhabitant houses.
Recommendations were sent peer SMS to the mobile devices of
the inhabitants. An example off such recommendation is shown
in Fig. 5.
Response to Recommendations
ociaation rules, which
t paatterns. The event
rom
m the smart home,
ore component of the
rulees and the event
matching algorithm is
rmiinistic finite state
Fig
g. 4, which reflects
E. A new instance of
thee stream. If there is
ancee is removed from
thee next event is not
omm
mendation.
tem allows more than
me. In order to avoid
we propose
p
to weight
zatio
on criterion.
We ran the evaluation in tw
wo phases, which are described in
the following sections.
TABLE II.
KEY RESULTS
R
OF EVALUATION
Parameter
# days evaluated
Phase
1
2
14
34
160
120
76
55
7
5
69
50
Ratio useful/answered
9.21%
9.10%
Number of active rules
54
46
Number of rules that resulted in recommendations
23
17
5
3
Recommendations sent
Answered recommendations
Voted useful
Voted not useful
Number of rules with 10 negative feeedbacks
A. Phase 1
The aim of the first phase was to provide a large basis of
data for evaluation and furtheer improvement of the system.
Polytechnic Institute
Thehomes
analysis
the data
collec
cted
phase
should
help
Zehnder, et al. "Energy saving in smart
based onof
consumer
behaviour
data."
PhD during
diss., Master’s
thesis,1
School
of Worcester
d M.
fro
om
the log files of
Business, University of Applied Sciences and Arts Nortwestern Switzerland (FHNW) Jan, 2015.
to improve the recommender sy
ystem in terms of decreasing the
lem
mented on a cloud
achine). The VM was
Questions?
Related documents