all about tasks building an automated task list

all about tasks building an automated task list + interrupt manager electronic Max ph. d. proposal overview aire, mit csail

all about tasks outline • no time to think (motivation: 4) • tasks & interruptions (background: 4) • the task box: a thought experiment (applications: 8) • inside the task box (methods: 3) – mapping actions to tasks (8) – unsupervised methods to sequences (2) • discussion

0: no time to think how technology brings knowledge workers less rest

the decline and fall of the american knowledge worker US Knowledge workers are feeling increasingly overworked, stressed and dissatisfied with their jobs. Since 1980’s, longer hours, working weekends, mandatory overtime Breakdown in barriers between work and leisure Increase in reports of worker dissatisfaction Rise in occupational stress-related illnesses Affecting life outside of work -- families, dating

the decline and fall of the american knowledge worker Is technology at fault? “work anywhere, anytime”! more social and professional connections : more roles, commitments individual empowerment : less human support distributed work requires more effort for coordination rise in collaboration: # of primary and secondary roles in projects

what’s the solution? People are aware of the problem. . . EU Working Time Directive (WTD) 48 hrs/wk, 4 wk break But in the US: Solutions are stymied in conflict ! Knowledge work : 60% of the US economy Bush wants to eliminate overtime protection, 40 -hr week for knowledge workers (see FLSA exemption for “profesisonal”, “admin” & computer workers”) Management reluctant: policies too general, expensive, Little is being done about it -- “Just a factor of modern life” Take Back Your Time Day (Oct 24) ?

individual empowerment: self-help david allen’s algorithm for “stress-free productivity” : how to (1) organize and prioritize tasks and (2) your materials around your tasks how to handle interrupts

1: tasks & interruptions

call mom to wish her a happy thanksgiving make slides for aire group meeting seek out an xbox 360 game for Willy for Christmas review unlockedgroove. 001 release for pitchfork try albert’s pymozilla suggestion make playlist for Dark. BOT 11. 25 proposal: write section on learning-related challenges • technology => more tasks test jrex stability when using privoxy obtain tickets to morr music event • more tasks => need tools coordinate COUHES approval form submission with scotty • today’s tools are primitive talk to larry re: KIMONO “glorified scraps of paper” get back to Bo @ PARC regarding soundtube solutions learn about latent semantic analysis pay speakeasy bill buy tickets home for the holidays ask jacob re: parameter smoothing for estimation “hypertasking” burn CD of darkbot cc for elizabeth set up a meeting with albert and scotty re: Ok-net for next week send out a HCI-seminar reminder for paul next week go pick up zipcard from 1 st st office pay electricity bill set up training session with thrashcore zach finish albatross tutorial set up tech night for WMBR playlist team engineering training for ryan proposal: write bibliography proposal: contact rcm proposal: find other readers for my committee why tasks?

V. Bellotti et al @ PARC “What-A-To-Do: Studies of Task Management, Towards the Design of a Personal Task List Manager” (CHI’ 04) - mean active to-dos: 75 - kept in explicit list : 14% - mean ‘to-do reminder tools’ : 12 - most prominent todo reminder: email (online calendar) Conclusions: - task vistas - flexible task spec: low effort of input - task histories - time monitoring - support priority assessment - support multi-bird killing “value extensions”

why interrupts? “interrupt-driven lifestyle” Gloria Mark: “No Task Left Behind? Examining the Nature of Fragmented Work”, CHI 2005 knowledge work is highly fragmented internal vs external interrupts, “working spheres” time spent on work task before interrupt: mean: 11 min (std: 18 mins) task resumption: 77% same-day, mean 2. 3 tasks later co-located vs distributed: co-located people spend longer in a working sphere (11 min vs 9 min) but interrupts take longer the sphere, the longer the interrupts

2: restoring calm using The Task Box: a thought experiment

user task identifier user { t 0, t 1, t 2, . . . tm } taski

task-based user interfaces, proactive help user task identifier calm: reduce clutter in HFAs improving search, information retrieval

off the desktop. . . user task identifier calm: context: < time, location, task > ubicomp apps that provide context-based proactive assistance

atlim ui resource watcher proposal p 735 -bellotti. pdf atlim p 347 -bellotti. pdf thesis proposal chi 2005. pdf http: //portal. acm. org/library/511234 http: //portal. acm. org/library/232323 http: //citeseer. isu. pst. edu/library/22 user task identifier tthesis proposal hacking watson calm: organize materials around projects simplify access to all relevant resources for a task (Task. Master, UMEA)

user task identifier user calm: improve user awareness of their time thru feedback: workflow patterns, accurate time spent, task fragmentation

{ t 0, t 1, t 2, . . . tm } task clusterer/ analyzer user task identifier user calm: t 0: 25 min t 1: 15 min t 2: 13 min. . assist users with task prioritization: predict how long a task might take, given similar tasks in the past

interrupts user task identifier t 4 interrupt barrier interrupt queue relevant to t 4 user calm: filtering interrupts: reduce task-switching by allowing only interrupts that are relevant to the current task, or high priority

task stack user task identifier t 3 t 21 t 16 calm: help users recover from interrupts by keeping a task stack, and reminding users of previous work state(s) higher resumption rate, reduced time and effort required to resume (Kimura)

3: inside the box

learning correspondences user task identifier user observational input { t 0, t 1, t 2, . . . tm } task specification taski

input 0 : : task specification : : initial representations of tasks 1. 2. 3. 4. 5. 6. Task description Task type / tags Priority Deadline Related people Related documents

input 1 : : observations : : event streams from applications

3. 2 a: inside the box: matching actions to tasks

matching actions to tasks: formally speaking. . { t-k, t-k+1, t-k+2, t-k+3, . . . t }, f { t 0, t 1, t 2, t 3, t 4, . . . t|T| } { tunknown } T f ( [ t-k, t-k+1, t-k+2, t-k+3, . . . t ] ) => [ task( t-k), task( t-k). . . task( t) ] primary task simplification: each action maps to at most 1 task in T

directly mapping actions to tasks: a supervised learning formulation Labeled examples { 0, 1, 2, 3, . . . m }, { t 0, t 1, t 2, t 3, t 4, . . . tm } New examples f { 20, 21, 22, 23, . . . 30 }, { t 18, t 26, t 32, t 33, t 34, . . . } two generalizations: actions and tasks

one more time, if only our tasks were a fixed set. . . we’d have a standard, |T|-classification problem: given a training set, { 0, 1, 2, 3, . . . m }, { t 0, t 1, t 2, t 3, t 4, . . . tm } we can use a bayesian classifier: f( ) = argmaxti T p(ti| ) p(ti | ) p( | ti ) p( ti ) estimate class-conditional density p( | ti ) any way we please (gaussians, parzen windows, etc). . . but they aren’t.

undaunted, we try a representing tasks by their features instead: for each action, { 20, 21, 22, 23, . . . 30 }, suppose we write each i feature vector in N : C R E R R ‘max’ ‘http: /’ ‘thesi’ ‘http: /’ ‘bell’ 0. 01 0. 02 0. 01 0. 42 0. 01 0. 02 0. 01 0. 23 0. 2 0. 43 0. 04 0. 02 0. 06 0. 04 0. 02 0. 24 0. 11 0. 24 0. 61 0. 23 0. 89 0. 00 0. 89 0. 23 0. 89 0. 04. . . 20 21 22 23 24 25 26 C ‘http: /’ 0. 42 0. 06 0. 24 0. 89. . . 30 likewise, for each task, { t 0, t 1, t 21, . . . t 192 }, suppose we write each ti feature vector in M : A R D G Z Y X ‘food’ ‘sally’ ‘thesi’ ‘http: /’ ‘http: ’ ‘couh’ ‘bell’ 0. 01 0. 2 0. 01 0. 42 0. 01 0. 02 0. 01 0. 3 0. 2 0. 93 0. 23 0. 17 0. 43 0. 24 0. 84 0. 02 0. 96 0. 64 0. 04 0. 92 0. 02 0. 14 0. 11 0. 24 0. 72 0. 74 0. 1 0. 3 0. 13 0. 20 0. 59 0. 23 0. 89 0. 5. . . . . t 0 t 1 t 21 t 24 t 25 t 26 . . . M ‘http: /’ 0. 42 0. 06 0. 24 0. 89. . . t 192

f( ) -> t f( ) = argmaxt T F( , t) Where F( , t) specifies the joint distribution over actions and tasks. However, F( , t) is N M N+M. . . what to do ? max’s mini-insight: compare with other problems with structured output ; rather than a class label. in particular. . .

Perceptron learning for structured output spaces : M. Collins “Discriminative Training Methods for Hidden Markov Models” and Experiments with Perceptron Methods, EMNLP, 2002. . Theory later generalized to: Tsochantaridis, I. , Joachims, T. : “Support Vector Machine Learning for M. Collins, “Parameter Statistical Parsing Models: Interdependent and. Estimation Structuredfor Output Spaces”, ICML 2004. Theory and Practice of Distribution-free Methods” 2004. two tricks: 1. dimensionality reduction using a feature function of input space (x, y) to some d-dimensional feature space : (x, y) : N M D 2. definition of learned function: f(x) = argmax y GEN(x) < (x, y), W> where W is learned via a modified Rosenblatt’s perceptron learning algorithm ** task identification: tasks are structured representations too! (x, y) useful for pulling out features in common between i tj how to choose (x, y) ?

(x A, y T) D. . . is where all the domain insight comes in 0(x, y): [0. . 1] normalized TF-IDF dot product between Descr(y) and About(x) 1(x, y): [0. . 1] normalized TF-IDF dot product across documents RD(y) and About(x) 2(x, y): {0, 1} indicating whether the Related. People(y) match in Who: (x) 3(x, y): [0, 1] indicating normalized Priority(y) 4(x, y): [0, 1] indicating normalized Proximityof. Deadline(y). . . D(x, y): [0. . 1] indicating how close User’s location is to Who(y) Perceptron learning algorithm selects weights W for < W, (x, y) > according to predictive power of features of (x, y)

3. 2 b: inside the box: unsupervised methods for sequences of actions

f( ) = argmaxti T p( ti | activity_subgraph(C( )) )) we can directly estimate task change ? C 1 C 3 C 5 C 4 C 10 C 15 C 11 C 12 (sorted by “monitoring” behavior? “thrashing” behavior? cluster )

4. Discussion

questions. . • will any of this really work at all ? – Benefit worth the trouble of labeling data? – Performance that is useful? • how can we boost the performance? – input processing: Harmful ? Task representation: ? – choice of features • will perceptual features boost performance? • feasibility of training across users? • method: can we incorporate better priors • how can we build a task / interrupt manager around our box? • how can I get famous and graduate ?