Candidate: Parisa Rashidi Advisor: Diane J. Cook 1.

Candidate: Parisa Rashidi Advisor: Diane J. Cook 1

Agenda Introduction Challenges Solutions Sequence mining Stream mining Transfer Learning Active learning Results Conclusions & future directions 2

Smart Homes Sensors & actuators integrated into everyday objects Knowledge acquisition about inhabitant 3 Environment Agent Percepts (sensors) Actions (controllers)

Applications Energy efficiency Security Achieving more comfort Monitoring well-being of residents In home monitoring Monitor daily activities Check for anomalies Help by giving prompts and cues 4

Activity Recognition A vital component of smart homes Recognizing activities from stream of sensor events 5 An Activity (Sequence of sensor events) A Sensor Event

Agenda Introduction Challenges Solutions Sequence mining Stream mining Transfer learning Active learning Results Conclusions & future directions 6

Why it is difficult? Human activity is erratic and complex Discontinuous (interrupting events) Step order might vary each time Inter-subject and intra-subject variability The algorithm should be scalable Data annotation Costly and laborious Training for each new space? 7

Unsolved Challenges Many methods proposed Hidden Markov models, conditional random fields, nave Bayes, Current methods Consider many simplifying assumptions Mostly are supervised Data annotation problem Even if unsupervised Trained for each new setting from scratch Ignore activity variations or interruptions 8

Our Solutions Discovering complex activities Sequence mining Discovery activities from stream Stream sequence mining Transferring activity models to new spaces Transfer learning Guiding activity annotation Active learning 10

Sequence Mining Sequence Ordered set of items Examples Speech: sequence of phonemes DNA sequence: AAGCTACGTAA Network: sequence of packets Our data: sequence of sensor events Goal Finding repetitive sequential patterns in data Many methods proposed GSP, PrefixSpan, SPADE, 12

Activity Sequence Mining Problem Data: a single sequence with no boundaries Unlike transaction data We are looking for activity sequence patterns With discontinuous steps Variations of the same activity 13 Transaction IDItems 1{Milk, Egg, Bread} 2{Bread, Beer} 3{Soap, Milk, Egg} MDMDACDF Item-set boundary No boundaries !

From Sequence Mining to Activity Recognition Find activity patterns Discontinuous Varied Sequence Mining (DVSM) Continuous, varied Order, Multi Threshold (COM) Cluster similar patterns Cluster centroid is a representative activity. Recognize activities Hidden Markov Model 14

DVSM Finds general patterns/variations in several iteration During each iteration Finds increasing length patterns Extend by prefix and suffix at each iteration Checks if it is a variation of a general pattern At the end of each iteration Retain only interesting patterns according to MDL principle 15 Pattern Instances {b,x,a} {a,b,q} {a,u,b} General Pattern Continuity Compression

DVSM Continuity Pattern Variations Instances Events Prunes patterns/variations with low compression values Highly discontinuous Infrequent Prunes non-maximal patterns Prune irrelevant variations using mutual information and sensor 16

Improve DVSM: COM Different sensor frequencies for Different regions of home Different types of sensor Rare item problem A global min-support doesnt work! Use multiple support thresholds 17

Clustering Grouping similar objects together There are many different clustering methods Partition based (k-Means) Hierarchal (CURE) Density based (DBSCAN) Model based (EM) 18

Similarity Measure How similarity is determined? Our activity similarity measure 19 Total Similarity Start Time Similarity Duration Similarity Structure Similarity Location Similarity =+ + +

Activity Recognition Basically a sequence classification problem Different than ordinary classification problems Variable length records Order Probabilistic methods are the most widely used Markov chains Hidden Markov models Dynamic Bayesian Networks Conditional random fields 20

Hidden Markov Model A statistical model Markovian property A number of observed & hidden variables Their transition probabilities We automatically build HMM from cluster centroids 21

Stream Mining Many emerging applications IP network traffic Scientific data Process data as it arrives We cannot store all data One pass Approximate and randomization answers E.g. relaxed support threshold Some proposed methods Frequent itemset mining Lossy counting [Manku 2002], SpaceSaving algorithm [Metwally 2005], Frequent sequence mining SPEED algorithm [Raissi 2005],.. 23

Tilted Time Model Uses a set of time-tilted windows to keep frequency of items Finer details for more recent time frame Coarser details for older time frames Shifting history into older time frames as data arrives 24 Month day hour *C. Giannella, J. Han, J. Pei, X. Yan, and P. S. Yu, Mining Frequent Patterns in Data Streams at Multiple Time Granularities. MIT Press, 2003, ch. 3.

Tilted Time Model Minimum support: Maximum support error: An itemset can be Frequent Sub-frequent Infrequent Pruning itemsets (tail pruning) 25

StreamCOM Extending COM into a stream mining method Using tilted time model 26 COM Titled Time Model StreamCOM

Finds general patterns/variations in several iteration During each iteration Finds increasing length patterns Extend by prefix and suffix at each iteration Checks if it is a variation of a general pattern At the end of each iteration Retain only interesting patterns according to MDL principle 27 Discovering Patterns {b,x,c,a}{a,b,q}{a,u,b} General Pattern Variation

General Pattern T (a) g Interesting ( g s ) T (a) < g Sub- interesting Otherwise uninteresting Variation i T (a i ) Interesting ( v ) T (a i ) < Sub- interesting Otherwise uninteresting 28 Interesting Patterns Average compression of all variations

Tail pruning 29 Pruning Patterns General Pattern Variation

Tail Pruning To reduce the number of frequency records in the tilted-time windows Prune old frequency records of an itemset 30

Transfer Learning Apply skills learned in previous tasks to novel tasks Chess Checkers Math CS 32 Traditional ML Transfer Learning training items test items training items test items

Transfer Learning Methods Transfer Learning Labeled Target Data? Non-Inductive Transfer Learning Labeled Source Data? Unsupervised Transfer Learning Transductive Transfer Learning Same domains? Sample Selection/Covariance Shift Domain Adaptation Inductive Transfer Learning Labeled Source Data? Self Taught LearningMulti-Task Learning 33 Yes No Yes No Yes No * S. Pan; Q.Yang;, "A Survey on Transfer Learning, IEEE TKDE, vol.22, no.10, pp.1345-1359, Oct. 2010

Why in Smart Homes? Why transfer learning? Supervised methods Requires annotation Unsupervised methods Requires lots of data 34

Our Transfer Learning Solutions Activity Transfer Transfer from one resident to another Different residents, space layouts, sensors Transfer from a single physical source to a target Transfer from multiple physical source to a target Domain selection 35

Multi Resident Transfer Learning 1. Find interesting target patterns using DVSM 2. Cluster discovered patterns 3. Map cluster centroids to source activities 36

Multi Home Transfer Learning (MHTL) 1. Find activity models in both spaces Source: extract activity model Target: location based mining, incremental clustering Activity consolidation, sensor selection 2. Map activity models from source to target Map Sensors Map activities 3. Map Labels 4. Use labels for recognition! 37

MHTL Architecture 38

EM Framework Updating sensor mappings probabilities Updating activity mapping probabilities 39

Label Assignment 40

Domain Selection Our previous works Assumed all sources are equal Not all sources are equal Some sources are more equal! Select top N sources Efficiency: do not use all sources Accuracy: negative transfer effect 41 Some animals are more equal... George Orwell Animal Farm

Domain Similarity How to measure difference between two distributions? 42

Domain Similarity Conventional similarity measures Kullbeck Leibler divergence (KL), Jensen Shannon divergence (JSD), L 1 or L p norms Kifer et al [2004] proposed H distance Later Ben David et al [2007] proved that It is exactly the problem of minimizing the empirical risk of a classifier that discriminates between instances drawn from the two domain! 43

Demonstration of H Distance 44 H-distance: 0.1, small! *Shai Ben-David, John Blitzer, Koby Crammer, and Fernando Pereira. Analysis of representations for domain adaptation. In NIPS, 2007.

Domain Similarity Kifer et al [2004] proposed H distance Later Ben David et al [2007] proved that It is exactly the problem of minimizing the empirical risk of a classifier that discriminates between instances drawn from the two domain! 45

Our Domain Selection Method Find similarity of domains activity-wise Overall similarity: average activity-wise similarity Select n top sources 46

Active Learning The learning algorithm can query for the label of a point Ask the oracle! Proposed methods Uncertainty sampling, committee based, 48

A Problem! Traditional active learning methods Ask overly specific queries 49 What is the class label if (sex= female) and (age =39) and (chest pain type =3) and (serum cholesterol = 150.2 mg/dL) and (fasting blood sugar = 150 mg/dL)... and (electrocardiographic result = 1) and (maximum heart rate achieved = 126) and (exercise induced angina = 90) and (heart old peak = 2.3) and (number of major vessels colored by fluoroscopy = 3)? vs. What is the class label if (age > 65) and (chest pain type = 3) and (serum cholesterol > 240 mg/dL) ?

Template Based Queries Select the most informative instances Select friends (+) and enemies (-) = Select relevant and weakly relevant features in Build a template query using relevant and weakly relevant features 50

RIQY 51 RIQY: Rule Induced active learning QuerY method Select the most informative instances Select friends (+) and enemies (-) = Use rule induction to build generic queries

Details The most informative instance 52

Can we discover activities? DVSM vs. COM 54

Activity Discovery Confusion matrix for various activities in apartment 1 55

Some Discovered Patterns 56

StreamCOM Taking medication activity 57

Transferring Activities 58

Transferring Activities 59

What about active learning? 60 Wisconsin breast cancer dataset -UCI repository Kyoto smart apartment dataset -CASAS

Conclusions Two novel sequence mining methods DVSM COM A novel stream data mining method StreamCOM A couple of transfer learning methods Between residents Between one/multiple smart homes Source selection Two novel active learning methods Template based active learning RIQY 61

Future Work Anomaly detection in sequences Exploiting more temporal information Order of activities Change detection in patterns 62

Publications Published/Accepted Parisa Rashidi and Diane J. Cook. Mining and Monitoring Patterns of Daily Routines for Assisted Living in Real World Settings. Proceedings of International Health Informatics Conference (IHI). 2010. Parisa Rashidi and Diane J. Cook. Transferring learned activities in smart environments between different residents. Proceedings of International Conference on Intelligent Environments (IE), volume 2, pages 185-192. Springer-Verlag, 2009. Parisa Rashidi and Diane J. Cook. Multi Home Transfer Learning for Resident Activity Discovery and Recognition. Proceedings of International Workshop on Knowledge Discovery from Sensor Data (KDD), pages 53-63, 2010. Parisa Rashidi, Diane J. Cook, "Home to home transfer learning", Proceedings of AAAI Plan, Activity, Intention Recognition Workshop (AAAI), 2010. 63

Publications Published/Accepted Parisa Rashidi, Diane J. Cook, "Transferring Learned Activities and Cues between Different Residential Spaces", Journal of Pervasive and Mobile Computing (PMC). March 2010. Maureen Schmitter-Edgecombe, Parisa Rashidi, Diane J. Cook, Larry Holder. Discovering and Tracking Activities for Assisted Living, The American Journal of Geriatric Psychiatry. In Press, 2010. Parisa Rashidi, Diane J. Cook,, Larry Holder, Maureen Schmitter- Edgecombe. Discovering Activities to Recognize and Track in a Smart Environment, IEEE Transaction of Data and Knowledge Engineering (TKDE). In Press, 2010. Parisa Rashidi, Diane J. Cook, Mining Sensor Streams for Discovering Human Activity Patterns Over Time. Proceedings of International Conference on Data Mining (ICDM), 2010. 64

Publications Submitted Parisa Rashidi, Diane J. Cook. Domain Selection and Adaptation in Smart Homes. ICOST 2011, January 2011, submitted. Parisa Rashidi, Diane J. Cook. Template Based Active Learning. AAAI 2011, February 2011. Submitted. Parisa Rashidi, Diane J. Cook. Ask Me Better Questions. Rule Induction Based Active Learning. KDD 2011, February 2011. Submitted. 65

Publications Invited/To be submitted Parisa Rashidi, Diane J. Cook. Mining and Monitoring Patterns of Daily Routines for Assisted Living in Real World Settings. ACM Transactions special issue on Intelligent Systems for Health Informatics. Invited. April 2011 Parisa Rashidi, Diane J. Cook. Generic Active Learning Queries. TKDE or JMLR. May 2011. To be submitted. 66

Questions? 67

Candidate: Parisa Rashidi Advisor: Diane J. Cook 1.

Documents

sequence mining sequence

activity sequence patterns

stream stream sequence

activity sequence of

single sequence

sequence of packets

activity models

activity recognition