Title: Which Kinds of Trend Metrics Are More Effective for Emerging Trend Detection
1Which Kinds of Trend Metrics Are More Effective
for Emerging Trend Detection?
- Yuen-Hsien Tseng
- National Taiwan Normal UniversityYu-I Lin
- Taipei Municipal Univ. of Education
Chun-Hsien Kuo Yi-Yang Lee Science Technology
Policy Research and Information Center
Taipei, Taiwan, R.O.C. 106
This presentation is based on the work to appear
in Scientometrics.
2Introduction ETD
- Monitoring research trends has always been a
concern of policy makers - it helps resource allocation and technology
forecast. - Increasingly important research topics are of
particular interest to those policy makers - They have been also termed as hot topics, upward
trends, or emerging trends. - ETD (Emerging Trend Detection)
3But how to detect them effectively?
- Domain experts are often consulted
- good at identifying interesting research trends
- But their observations do not generalize
effectively to the fields beyond their expertise - when a large number of research topics need to be
prioritized, inconsistent decision may result - Automatic mechanism for monitoring research
trends in a large stream of upcoming publications
would be of great help
4Detecting trends in scientometrics
- Noyons and van Raan pointed out that
- Domain experts are often hard to find, due to
busy schedules and lack of affinity with
scientometrics studies - Policy makers are often too much overwhelmed by
the amount of resulting information
5Motivations
- In past trend analysis,
- different year spans may be used to create the
time sequence - different indices were chosen for trend
observation - Simple count of publications is suspicious to get
good trend sequences Chi et al, 2006 - The effectiveness of these choices
- was unknown quantitatively and comparatively
- This work provides clues to better interpret the
results when a certain choice was made
6Questions ?
- For effective trend detection, which options
should be used?
Different year spans!
Data are from Smeaton et al 2003 ACM SIGIR Forum
Simple count to create a sequence suspicious for
ETD
Different trend orderings due to different
criteria!
7Simple Trend vs EigenTrend
Chi, Tseng, Tatemura, 2006, CIKM challenged
the validity of the simple accumulation of
published documents over time
- Simple authority
- Simple trend
- (First) Authority U1
- (First) Eigen-Trend s11V1
- Error
Break down by sources
DUSVT
8Outline of the following talk
- ETD methodology
- Trend metrics to be compared
- Evaluation method
- Data sets for evaluating ETD
- Safety agriculture (SA)
- Information retrieval (IR)
- Evaluation results
- Conclusions and implications
9ETD methodology
- Documents (terms) were clustered to yield topics
- For each topic, a time series of number of
publications over time was created - Topics were then ranked by a trend metric
- an IR-based metaphor
- Input
- a set of publications (each with PY, TI, AU, C1,
SO, ) - Output
- a ranked list of topics in decreasing order of
interest
10Trend metrics to be compared (1/2)
- api (average percentage of increase)
- used in a foresight survey in Japan (STFC, 2004)
- used by Noyons et al when n2
- slp slope of the linear regression line that
best fits the data in the time series - slpz same as slp, but the sequence is first
z-score transformed (zi(di-avg)/stderr )
11Trend metrics to be compared (2/2)
- slppi a combination of api and slp.
- d1, d2, , dngtpi1, pi2, , pin-1,
pii(di1-di)/di - may be ideal for sharp increasing trend detection
- slpc eigen-trend break down by C1
- C1 first authors country
- slpj eigen-trend break down by SO (journal)
12Evaluation method NAP Pre_at_R
- Assume A-E and V-Z are ten items to be ordered
and A-E are relevant while V-Z are not. - Ordering S1 is the best by
- NAP Non-interpolated Average Precision rate
- Pre_at_R Precision rate at Recall position
- Pre_at_R r/R, where r is the number of
relevant items in the top R items - With NAP and Pre_at_R, we can evaluate which trend
orderings are best
Pre_at_R0.603/5 NAP0.68(1/12/33/54/75/9)/5
13Data set SA
- Six research domains regarding safety agriculture
(SA) were enumerated by a group of experts from
the Science Technology Policy Research and
Information Center (STPI) - food security, crop protection, livestock,
fishery, agroforestry, and environment - for each domain, a query was formulated to search
the ISIs Web of Science database - 72500 records between 1996 and 2005 were
downloaded
14Topic detection for Safety agriculture
- Clustering analysis was based on controlled terms
- 179 SC terms each occurs in more than 10 docs.
- 3632 DE terms of this kind
- Terms from each field (SC or DE) co-occurred in
the same records were counted - Similarity based on this count was used in a
complete link clustering algorithm - 80 clusters (topics) were found for SC terms
- 1617 clusters for DE terms
15Trend Type Labelling by Experts (1/2)
- We sampled 50 of clusters from SC and 10 from
DE for experts to judge their trend types - 6 professors, 2 researchers, 1 admin. manager
- Trend types
- sharp increasing
- increasing
- fluctuation
- - decreasing
- -- sharp decreasing
- ? inconclusive
16Trend Type Labelling by Experts (2/2)
- Experts were advised to judge the type of each
cluster based on their knowledge - If this did not help, the time series of the
cluster can be consulted - If this did not help either, the documents in the
cluster can be examined. - If all these efforts failed, the cluster was
labeled inconclusive
17Experts feedback
Data are from 72500 documents in safety
agricultural area.
Sharp increase Increase
Controlled terms clustering Different fields
undecidable
- Decrease -- Sharp decrease
18Date set Information retrieval (IR)
- 853 papers from the first ACM SIGIR conference to
the 25th were clustered by a commercial software
package called Clustan Graphics by Smeaton et al
ACM SIGIR Forum 2003 - 29 non-overlapping clusters were generated
- They then inspected each cluster manually and
assigned a topic description to reflect the theme
of the majority of the papers in each cluster - Topics are sorted approximately in order of a
combination of the year of their first
appearance, and the number of papers published
19Clustering and ordering of SIGIR papers by topics
made by Smeaton
20Hot topics predicted by Seamton et al
- The ideal paper title expected by Smeaton et al
to appear in SIGIR 2003 is - "Evaluation of a Language Model Implementation of
a Topic-Based, Cross-Lingual Question-Answering
and Summarisation System
21Fourteen session titles (topics) in the SIGIR
2003 conference
22Evaluation results SA
Avg is the average of the values in the SC and
DE rows
23Prediction effectiveness when year span varies
from 1, 2, to 5
(x1, x2)gt ((x1-avg)/stderr, (x2-avg)/stderr)(
(x1-x2)/2 / (x1-x2)/2, (-x1x2)/2 / (x1-x2)/2
). Thus only 3 values result (-1, 1), (1, -1),
(0, 0), which in turn yield only 3 possible
slopes 2, -2, and 0.
24Prediction based on less data
Percentage of performance drop for slp using only
the first n years of data, where n10, 8, 6, 4,
and 2.
Pre_at_R
NAP
25Evaluation results IR
26Conclusions
- Which metrics (methods) perform best for ETD?
- api average percentage of increase
- slp slope of the linear regression line
- eigen-trends
- Smeatons chronological ordering
- Our answer is slp, because it performs well
under - different year spans (1, 2, 5)
- different observation durations (10 vs. 25 years)
- different domains (SA vs IR)
- different collection scales (72500 vs 853 papers)
- api only works for n2 (so Noyons work still
valid)
27Conclusions
- Our goal is to explore the best way to predict
upward trends in an environment where a large
number of topics are to be monitored. - If a good trend index is used, the inspection in
the order sorted by the index should be efficient - Our work is important to know which metric is the
best under a certain condition.
28Implications
- The IR based method for evaluating the trend
index performance suggests a relatively objective
and repeatable procedure to indentify better
indices and to gather evidence to support (or
invalidate) our current results.