Human-Assisted Graph Search: It’s Okay to Ask Questions Reported by Qi Liu.

Post on 15-Jan-2016

218 views 0 download

Tags:

Transcript of Human-Assisted Graph Search: It’s Okay to Ask Questions Reported by Qi Liu.

Human-Assisted Graph Search: It’s Okay to Ask Questions

Reported by Qi Liu

Scenes

• Human Computation• Crowding-sourcing service(Amazon’s

Mechanical Turk)

An example

• Image classification

Terminology of the Problem

• HumanGS: abbreviation for human-assisted graph search

• Taxonomy: DAG(directed acyclic graph)• Category: Node• Question: Reachability• Tricks: Not leaves, Not root, Just middle!• Challenge: High Latency

More Applications

• Manual Curation(insert a web into web-graph)• Question: Is the item a kind of x?

Apps(cont.)

• Debugging of Workflows• Question: Is the outputfragment at point x wrong?

Apps(cont.)

• Filter Synthesis• Question: Do youwant all data items satisfying conditionx to be part of theresult?

Apps(cont.)

• Interactive Search• Question: Do you want more results like

concept x?

Dimensions of the problem

• Single/Multi (target set)• Bounded/Unlimited (question set)• DAG/Downward-Forest/Upward-Forest

Define the Problem

DAG property

Candidate Set

An Example

Q(nissan,maxima)=yes=>

Cand(nissan,maxima)={nissan,maxima,sentra}Q(mercedes,maxima)=no

=>Cand(mercedes,maxima)=V/{mercedes}

Q(car,maxima)=yes=>

Cand(car,maxima)=V/{vehicle}

Extending to|N|> 1

Goal: Picking N set

Single Target Node

Single-Bounded

Single-Bounded: DAG

Conclusion: A NP-hard max-cover problem

Single-Bounded: Downward-Forest

Equivalence to the Partition Problem

Show the equivalence

An example

Candidate Set : Partition

Induction and Conclusion

• Minimum wcase(N) => the size of the largest partition that can be induced by N.

• Solved in PTIME!

Single-Bounded: Upward-Forest

Single-Unlimited

• For DAG, the question numbers vary from O(log n) to O(n)

Single-Unlimited: Downward-Forest

Single-Unlimited: Upward-Forest

Multiple Target Nodes

• Multi-Bounded: DAG• Lower-bound: NP-hard in n and k• Upper-bound: • Multi-Bounded: Downward/Upward-Forest– DP algorithm: O(k^2*n*6)

Experiments

The End