Slide 1BackgroundProblemProposed SolutionRelated WorkIntegration PointsOpportunistic Feature Vector MergingOpportunistic Feature Vector MergingOpportunistic Feature Vector MergingSocial Network Driven SharingSocial Network Driven SharingProof of Concept ExperimentResultsResultsResultsConclusionsQuestions?Nicholas D. Lane, Hong Lu, Shane B. Eisenman, and Andrew T. CampbellPresenter: Pete ClementsCooperative Techniques Supporting Sensor-based People-centric InferencingBackgroundMetroSenseAndrew T. CampbellCollaboration between labs at Dartmouth & Columbia UniversityProjects IncludeSoundSenseCenceMeSensor SharingBikeNetAnonySenseSecond Life SensorProblemPeople-centric sensor-based applications need models to provide custom experienceLearning inference models is hampered by Lack of labeled training dataInsufficient training dataDisincentive due to time and effortAppropriate feature inputsHeterogeneous devicesInsufficient data inputsProposed SolutionOpportunistic feature vector mergingSocial-network-driven sharing of Model training data Models themselvesRelated WorkSharing training sets in machine learning nomenclature known as co-trainingSeveral successful systems using collaborative filtering (similar users can predict for each other)However, none keyed specifically on sharing data of users in same social networkIntegration PointsOpportunistic Feature Vector MergingMotivation - the accuracy of models increase as the sensor inputs from more capable cell phones are used to generate better modelsShareable CapabilitiesSensor configurationAvailable memoryCPU/DSP characteristicsAnything not highly person, device or location specificEssentially necessary sensor data not available through low end phone is opportunistically borrowed from more capable phoneOpportunistic Feature Vector MergingDirect SharingBorrowed from user in proximityLender broadcasts data sources, not featuresBorrowers request features of specific data sourceIndirect SharingBy matching common features to similar users with more capable featuresCentral server collects data, looks for merging opportunitiesOpportunistic Feature Vector MergingChallengesSharing not available when you need itMaintain multiple models based on feature availabilityUse algorithms more resilient to missing dataPrivacyUser configures shareable featuresTruly anonymous data exchange ongoing researchSocial Network Driven SharingMotivationAccurate models require lots of training data, and sharing data reduces this loadChallengesSharing data reduces accuracyUncontrolled collection methodHeterogeneous devicesSimple global model not the answerSocial Network Driven SharingTraining Data SharingAssume known social graphsModels trained from individual data and high ranking people in individual social graphLabel consistency issues addressed with clusteringModel sharingTest models in social network to discover best performingMix and match model componentsProof of Concept ExperimentSignificant places classifier that infers and tags locations of importance to a user based on sensor data gathered from cell phonesPhone capabilities ignored as needed to produce four capability classesBluetooth OnlyBluetooth + WiFiBluetooth + GPSBluetooth + WiFi + GPSResultsResults•Global Model•Pools training data from all participants equally•User Model•Training data sourced from user only•Instance Sharing•Training data source from user and users from social graph•Model Sharing•Selects best performing per-user model from self, global and users from social graphResultsPhone survey results indicate higher label recognition among members of same social groupConclusionsThere is opportunity to leverage both device heterogeneity, and social relationships when sharing data and models in the support of more accurate and timely model buildingQuestions?Thank
View Full Document