Challenges for the Computational Discovery of Scientific Knowledge Pat Langley School of Computing and Informatics Arizona State University Tempe, Arizona Institute for the Study of Learning and Expertise Palo Alto, California
Thanks to K. Arrigo, D. Billman, M. Bravo, S. Borrett, W. Bridewell, S. Dzeroski, and L. Todorovski for their contributions to this research, which is funded by a grant from the National Science Foundation.
Drawbacks of Scientific Data Mining Because it borrows from work on commercial applications, most work on scientific data mining: • generates models in forms inappropriate to most sciences • makes incorrect assumptions about the available inputs • focuses on convenient algorithmic issues, not scientists’ needs We need to redirect attention toward a broader range of discovery tasks that actually arise in scientific fields. Data-mining researchers would benefit from looking at the older literature on computational scientific discovery.
Claim 1: Scientific Notations Traditional data-mining notations are not easily understood by or communicated to domain scientists. Most sciences state and communicate models in formalisms they have used for decades. We need more work on discovering scientific knowledge cast in communicable forms (Dzeroski & Todorovski, 2007). Ecosystem model
Gene regulation model
NPPc = Σmonth max (E · IPAR, 0) E = 0.56 · T1 · T2 · W T1 = 0.8 + 0.02 · Topt – 0.0005 · Topt2 T2 = 1.18 / [(1 + e 0.2 · (Topt – Tempc – 10) ) · (1 + e 0.3 · (Tempc – Topt – 10) )] W = 0.5 + 0.5 · EET / PET PET = 1.6 · (10 · Tempc / AHI)A · PET-TW-M if Tempc > 0 PET = 0 if Tempc < 0 A = 0.00000068 · AHI3 – 0.000077 · AHI2 + 0.018 · AHI + 0.49 IPAR = 0.5 · FPAR-FAS · Monthly-Solar · Sol-Conver FPAR-FAS = min [(SR-FAS – 1.08) / SR (UMD-VEG) , 0.95] SR-FAS = (Mon-FAS-NDVI + 1000) / (Mon-FAS-NDVI – 1000)
+
NBLR
+
DFR
NBLA psbA1
+ + Light
RR
-
psbA2 cpcB
-
PBS Health
+
+
Photo
Claim 2: Background Knowledge Scientists often have initial knowledge that should influence the discovery process. Ignoring this knowledge can produce models that scientists reject as nonsensical (Pazzani et al., 2001). Observations
NBLR
+
+
Model Revision NBLR
+
+ DFR
+ Light
NBLA psbA1
+
RR
-
psbA2 cpcB
-
PBS
psbA1
+ Light
-
RR
× -
psbA2
-
PBS
+ Health
+
×
Photo
cpcB
Health
+ +
DFR
NBLA
Revised model
Photo
Initial model
Claim 3: Small Data Sets Most data-mining work assumes that large data sets are available. But in many scientific domains, data are rare and hard to obtain. Discovering scientific knowledge from small data sets raises an entirely different set of challenges (Lee et al., 1998). We need more research on this important aspect of discovery. Ecosystem model Number of variables Number of equations Number of parameters Number of samples
Gene regulation model 8 11 20 303
Number of variables Number of initial links Number of possible links Number of samples
9 11 70 20
Claim 4: Scientific Explanation Most work on data mining finds models that, although accurate, merely describe the observations. However, scientists often want models that explain their data using familiar concepts. Explanatory models can include theoretical entities and processes that link back to domain knowledge (Langley et al., 2002). Ecosystem model
Gene regulation model
NPPc NBLR E
+
+
IPAR DFR
e_max
W
T2
T1
SOLAR
psbA1
FPAR
+
A
PET
EET
Topt
+
SR
Light AHI
PETTWM
Tempc
NDVI
VEG
NBLA
RR
-
-
PBS Health
+ -
psbA2
cpcB
+
Photo
Claim 5: Interactive Discovery Most data-mining work focused on entirely automated algorithms. But most scientists want computational aids rather than systems that would replace them. We need more work on interactive discovery (Bridewell et al., 2007). Domain user NBLR
+
+
Model Revision
Observations NBLR
+
+ DFR
+ Light
NBLA psbA1
+
RR
-
psbA2 cpcB
-
PBS
psbA1
+ Light
-
RR
× -
psbA2
-
PBS
+ Health
+
×
Photo
cpcB
Health
+ +
DFR
NBLA
Revised model
Photo
Initial model
The PROMETHEUS System (Bridewell et al., 2007)