Its About Time: Analyzing Temporal MicroLevel Behavioral PatternsPresentation Transcript
It s About Time Analyzing Temporal MicroIt’s About Time – Analyzing Temporal Micro‐ Level Behavioral Patterns Chen Yu Indiana University
Temporal Bands of Human Computation (Ballard, Hayhoe, Pook, & Rao, 1997) (B ll d H h P k & R 1997)macro-level ternalbehaviormicro-levelmicro level extbehavior nal intern
Dual Eye Tracking in Child-Parent Interaction ( (with Linda Smith, Damian Fricker, & Linger Xu) , , g ) child’s first person view parent’s first person view from eye camera from head camera from head camera child eye tracker adult eye tracker
Dual Eye Tracking in Child-Parent Interaction
Multimodal data• Vision: 720*480, 30 frames/second, 3 cameras, 18,000 frames per dyads.• Motion tracking: 250 Hz/second, 6 sensors with six dimensions (x,y,z,h,p,r) on each. 900,000 900 000 data points per dyads dyads.• Speech: 44.1Hz• Eye tracking: 30Hz• 12GB per participant
Dealing with data• Synchronization of multiple data streams• Data annotation (automatically or manually)• Data management• Data Mining and Knowledge Discovery
Data Mining• Our goal of analyzing gaze data is to find new patterns and gain new knowledge from such data.• With micro level data, even if we have some predictions micro-level data from our experimental designs, we nonetheless lack precise predictions about the structure and patterns of data at the micro-level.• But how can we discover new and meaningful patterns if we d not know what we are looking for? do k h l k f• Discovering new knowledge requires the ability to detect unknown, surprising, novel, and unexpected patterns.• A particular challenge is not just from the amount of data but p g j from how to extract, select and interpret meaningful patterns from a sea of complex data.
Interactive Data Analysis top‐down knowledge data data pattern pattern visualization extraction• This solution relies on both computational techniques and human domain knowledge. Data visualization and pattern extraction techniques provide candidate patterns through a bottom up way. bottom-up way• Compared with “blind” data mining, what we suggest is that researchers with top down top-down theoretical knowledge need to be in the loop of data mining and data analysis.
face three objects red blue red green red bluechild hildgazeparentgaze face gaze and mutual gaze joint attention
Joint Attention State Joint Attention State60%50%40%30%20%10% 0% mutual gaze t l child‐face/parent‐object parent‐face/child‐object hild f / t bj t t f / hild bj t same object bj t different objects diff t bj t‐10%
child following child following face three objectschildgazeparentgaze parent followingchildggazeparentgaze
What is joint attention made of? when baby looks mom’s face, mom when MOM looks baby’s face, baby looks at baby’s face 54% of the time; when baby looks at an object mom looks at mom’s face 13% of the looks at 32% of timetime; when mom looks at an object baby looks at 14% of time baby lookshild of time at 14% child parent t 54%(face) 32%(object) 13%(face) 14%(object) gaze gaze 55% 56% hand hand 17% 21% probabilities of following an attended object/face
Sequential Patterns From Multimodal Data Stream active face-to- joint attention following facelearner’sgaze 1 1 1teacher’sgaze 2 3 2 3 2 3teacher’shandaction 4 4 4 face three objects
Segmentation and Alignment
Discovering Statistically Reliable Sequential Patterns (Fricker, Zhang, & Yu, 2011)multiple instances naming utterance face look describing utterance sequential prototype durations and timings within and across multiple events
More complex patterns from multiple data streams lti l d t t
From Data To Patterns
From Patterns To Knowledge ……
Interactive Data Mining top‐down knowledge data pattern visualization extraction• Human-in-the-Loop data analysis involves active and iterative examination and exploration of visually displayed patterns to select useful information and guide further knowledge discovery discovery.• “Discovery takes place between the ears” – Ben Shneiderman
Visual Data Mining(Yu, Zhong, Smith, Park, & Huang, 2008, 2009)
Acknowledgement Data collection and pre-processing : Damian Fricker, Amanda p p g , Favata, Char Wozniak, Melissa Elston. Data Processing: Tian Xu, Damian Fricker, Thomas Smith Henry Xu Fricker Smith, Shen. This research was s ppo ted b NSF BCS 0924248 and NIH T32 esea ch as supported by HD 07475 Thanks!