Slide 1people.csail.mit.edu/teevan/work/publications/talks/sigird… · PPT file · Web viewThe...

37
Jaime Teevan MIT, CSAIL The Re:Search Engine

Transcript of Slide 1people.csail.mit.edu/teevan/work/publications/talks/sigird… · PPT file · Web viewThe...

Jaime TeevanMIT, CSAIL

The

Re:Search Engine

“Pick a card, any card.”

Case 1 Case 2 Case 3 Case 4 Case 5 Case 6

Your Card is GONE!

People Forget a Lot

Change Blindness

http://www.usd.edu/psyc301/ChangeBlindness.htm

Change Blindness

http://www.usd.edu/psyc301/ChangeBlindness.htm

Re:Search Engine

?

Merge Old and New ResultsOld

New

Merged

We still need magic!

Overview

♠ Memorability study♠ Recognition study♠ Assumptions♠ Implementation issues♠ Evaluation issues♠ Choose your own adventure

Memorability Study♠ Participants issued

self-selected query♠ After an hour, asked

to fill out a survey♠ 129 people

remembered something

Data Analysis

Probability of being remembered♥ Anything? # of words? # of fields?♥ Features

♣Result features: clicked, not clicked, last clicked, rank, dwell time, frequency of visit, etc.

♣Query features: query type, query length, # of search in session, elapsed time, etc.

♠ Remembered rank v. real rank♥ Map remembered rank to real rank

“Memorability”

00.10.20.30.40.50.60.70.8

1 2 3 4 5 6 7 8 9 10

Rank - R

P(R

emem

|R,C

)

Clicked - C Not clicked

Remembered Results Ranked High

-2

0

2

4

6

8

10

12

-2 0 2 4 6 8 10 12

Remembered Rank

Act

ual R

ank

Recognition Study

♠ Same set-up as Memorability Study♠ Follow-up survey: Results the same?

♥ Case 1: Old results♥ Case 2: New results♥ Case 3: Clicked to top♥ Case 4: Intelligent merging

♠ 92 people have completed both steps

16%

74%

65%

17%

Assumptions♠ Re-search v. search♠ Memorable v. relevant♠ Results change v. stay the same♠ Hide change v. show change♠ Forget v. remember as forgettable♠ Merge v. identify old or new

Why? How to test? What if I’m wrong?

Implementation Issues

♠ Page of cached result may disappear♠ Multiple result pages♠ Identifying repeat queries

♥ User identified♥ Search sessions are not repeat queries♥ Exact query may be forgotten

Evaluation Issues

♠ Various goals to test♥ Does a merged list look like the original?♥ Does merging make re-finding easier?♥ Is search improved overall?

♠ Lab study♥ How to set up re-finding task?♥ Timing differences significant enough?

♠ Longitudinal study – What to measure?♠ What are good baselines?

Choose Your Own Adventure♠ Re-search v. search♠ Memorable v. relevant♠ Results change v. stay the same♠ Hide change v. show change♠ Forget v. remember as forgettable♠ Merge v. identify old or new

♠ Implementation issues♠ Evaluation issues

Choose Your Own Adventure♠ Re-search v. search♠ Memorable v. relevant♠ Results change v. stay the same ♠ Hide change v. show change♠ Forget v. remember as forgettable♠ Merge v. identify old or new

♠ Implementation issues♠ Evaluation issues (Done)

Hide Change v. Show Change

♠ Why I think change should be hidden♥ Example: dynamic menus

♠ How to prove♥ New results better, called the same or worse♥ Baseline for testing – 2 lists, change explicit

♠ What if we should show change?♥ Memorability suggests changes to highlight♥ Other applications where want to hide change

(Done)

Memorable v. Relevant

♠ Why I think memorability is important♥ Relevance at a future date is what matters♥ Necessary to hide change

♠ How to prove♥ Baseline for lab study with target first

♠ What if relevance is what’s important?♥ Mapping between memorable and relevant♥ Useful related work on implicit feedback

(Done)

Re-search v. Search

♠ Why I think people repeat searches♥ Information seeking literature♥ Re-finding consistently reported as a problem

♠ How to prove♥ Study shows prefer to follow known paths♥ Search log analysis

♠ What if people just want to search?♥ Memorable results ranked first♥ Other domains where list consistency matters

(Done)

Merge v. Identify Old and New

♠ Why I think results should be merged♥ Information need not necessarily one or other♥ People don’t like to do extra work

♠ How to prove♥ Search log analysis♥ Look at what people do in longitudinal study♥ Lab study – timing becomes an issue

♠ What if people want to identify query type?♥ Other applications where merging is useful

(Done)

Results Change v. Stay the Same

♠ Why I think results change♥ How search engines work♥ Personalization and dynamic content

♠ How to prove♥ Track query results

♠ What if results don’t change?♥ Probably will in future applications♥ Existing applications where lists change

(Done)

Forget v. Remember as Forgettable

♠ Why I think people forget♥ Visual analogy

♠ How to prove♥ Lab study – Do people find new information?♥ Longitudinal study – Ever click on new result?

♠ What if remember as forgettable?♥ Build better model of memorability♥ Highlight important changes

(Done)

Implementation Issues

♠ Page of cached result may disappear♠ Multiple result pages♠ Identifying repeat queries

♥ User identified♥ Search sessions are not repeat queries♥ Exact query may be forgotten

(Done)

Evaluation Issues

♠ Various goals to test♥ Does a merged list look like the original?♥ Does merging make re-finding easier?♥ Is search improved overall?

♠ Lab study♥ How to set up re-finding task?♥ Timing differences significant enough?

♠ Longitudinal study – What to measure?♠ What are good baselines?

(Done)

Jaime [email protected]

Teleporting

Orienteering

Strategies for Finding

Why Do People Orienteer?

♠ Easier than saying what you want♠ You know where you are♠ You know what you find

♠ The tools don’t work

All must be the same to re-find the information!

Structural Consistency Important

New name

Absolute Consistency Unnecessary

New name

Focus on search result lists

Query Changes

♠ Most changes are simple♥ Capitalization♥ Phrasing♥ Word ordering♥ Word form♥ New queries shorter

♠ What about longer time horizons?♠ Recognition v. recall

Result List Changes

♠ Tracked 10 queries on Google for a year+♠ 1.18 of top 10 disappear each week♠ Rate of change likely to increase, e.g.:

♥ Raw personalization♥ Relevance feedback

♠ People forget their queries♥ 28% of queries forgotten within an hour

Example: “neon signs”

0

2

4

6

8

10

12

14