Zhara Ferdowski
Data mining to find socioeconomic character of modern cities
Chicago modern data. (survey from 2000)
K-means clustering
data mining vs. statistical techniques.
Chicago - 860 census tracks (what is a census track?)
Zhara sorted them down to five clusters.
1. mainly hispanic. renters, low income, low edu.
2. mainly white. middle class.
3. high income.
color map- slides. orange (3) = north side.
4.
5. blue (I guess- poor?)
Questions:
How did you determine the # of clusters?
A: tried # of clusters from 2-20, based on 2 criteria. got opinions of experts on which is more "real"
Q: the program wouldn't work in an environment where you're spring-injecting variables, would it?
A: static analysis tool. Can be used to retrofit exisiting programs. (Was this an answer?)
Query-based data mining for the web
Barbara Poblette (researcher at Yahoo. in Chile, santiago, barcelona, etc.)
info retrieval, data mining.
High data dimensionality. Focus: user behavior. usage logs. (how did she get this? eep.)
automatic organization is important for IR
Data mining to find socioeconomic character of modern cities
Location: Dunwoody 9:30
Chicago modern data. (survey from 2000)
K-means clustering
data mining vs. statistical techniques.
building permits- studying renovation leading to gentrification (is that the right cause-effect?)
also using the 1980, 90 census data.
Data preprocessing- all sets in different formats.
also using the 1980, 90 census data.
Data preprocessing- all sets in different formats.
Chicago - 860 census tracks (what is a census track?)
Zhara sorted them down to five clusters.
1. mainly hispanic. renters, low income, low edu.
2. mainly white. middle class.
3. high income.
color map- slides. orange (3) = north side.
4.
5. blue (I guess- poor?)
Questions:
How did you determine the # of clusters?
A: tried # of clusters from 2-20, based on 2 criteria. got opinions of experts on which is more "real"
Q: What was the hardest part of the proj?
A: collecting data. there was no consistent data... different formats... not complete data set.
Divya- PhD @ Penn.
9:48
advisor: Trent Jaeger(sp?)
Retrofitting Authorization in legacy systems
full room!
X server system
mediation statements-
- declassifyers(don't put in plaintext- encrypt before writing to file (entering password))
- type 2??
- access hooks: user must have permission to read/write to file inorder to use...
example: logrotate (code! also small red lines.)
goal: automate the process to re-secure old programs or at least detect fails.
security lattice- secrecy, integrety, authoization.
information flow graph. (model program using one.)
place mediators at min-cut vertices.
"completeness guarantee"
can mark lib files as "must not" locations
if there's already a bit that encrypts, all flows through that can be dismissed.
heuristically better mediators: avoid putting them in functions that are accessed often.
applied this to ~8 programs
profilecode at runtime to gather this info.
hoisting: reducing field sensitivity
trying to write it in C. Not done yet.
example: tinySQL
A: collecting data. there was no consistent data... different formats... not complete data set.
Divya- PhD @ Penn.
9:48
advisor: Trent Jaeger(sp?)
Retrofitting Authorization in legacy systems
full room!
X server system
mediation statements-
- declassifyers(don't put in plaintext- encrypt before writing to file (entering password))
- type 2??
- access hooks: user must have permission to read/write to file inorder to use...
example: logrotate (code! also small red lines.)
goal: automate the process to re-secure old programs or at least detect fails.
security lattice- secrecy, integrety, authoization.
information flow graph. (model program using one.)
place mediators at min-cut vertices.
"completeness guarantee"
can mark lib files as "must not" locations
if there's already a bit that encrypts, all flows through that can be dismissed.
heuristically better mediators: avoid putting them in functions that are accessed often.
applied this to ~8 programs
profilecode at runtime to gather this info.
hoisting: reducing field sensitivity
trying to write it in C. Not done yet.
example: tinySQL
Q: the program wouldn't work in an environment where you're spring-injecting variables, would it?
A: static analysis tool. Can be used to retrofit exisiting programs. (Was this an answer?)
Query-based data mining for the web
Barbara Poblette (researcher at Yahoo. in Chile, santiago, barcelona, etc.)
info retrieval, data mining.
High data dimensionality. Focus: user behavior. usage logs. (how did she get this? eep.)
automatic organization is important for IR
"Right now they have human editors, and that's not a scalable approach."
2" query dociument model, ___
traditional: vector space model i.e. bag of words
query model: bag of uery terms
query-set model: bag of query sets (how of you determine a query set?)
query - transaction which happens n times.
also, # of clicks from query to doc.
2 types of web graphs: click and hyperlink.
link graph: susceptible to spam(?)
combine them, find best-of.
"Queries are an excellent way to reduce the feature space of the web."
Q: implications = ? how did you gather this data?
A: Interning at yahoo, access to yahoo access log. "they have all the queries..." big log... 144 mil documents...
Talk 3:
thesis
relational data mining.
classification of
"A new link-based naive bayes classifier"
Bahareh Bina
Oliver Schulte
scenario 1: social network. goal: predict whether a person smokes. check smoking habits of friends.
scenario 2: goal: will customer buy product x or no? tradiationally, use only attributes of customer. however, now can use attributes of their friends sigh.
link-based classification:
predict the class label give attributes of obj, related obj, and of the links
(marketing :p)
prev work: graph-NB
all assumptions together = standad relational NBC.
our algorithm: LNBC
consider weight and number of rows (to normalize the weight)
improve classification accuracy on benchmark datasets
2" query dociument model, ___
traditional: vector space model i.e. bag of words
query model: bag of uery terms
query-set model: bag of query sets (how of you determine a query set?)
query - transaction which happens n times.
also, # of clicks from query to doc.
2 types of web graphs: click and hyperlink.
link graph: susceptible to spam(?)
combine them, find best-of.
"Queries are an excellent way to reduce the feature space of the web."
Q: implications = ? how did you gather this data?
A: Interning at yahoo, access to yahoo access log. "they have all the queries..." big log... 144 mil documents...
Talk 3:
thesis
relational data mining.
classification of
"A new link-based naive bayes classifier"
Bahareh Bina
Oliver Schulte
scenario 1: social network. goal: predict whether a person smokes. check smoking habits of friends.
scenario 2: goal: will customer buy product x or no? tradiationally, use only attributes of customer. however, now can use attributes of their friends sigh.
link-based classification:
predict the class label give attributes of obj, related obj, and of the links
(marketing :p)
prev work: graph-NB
all assumptions together = standad relational NBC.
our algorithm: LNBC
consider weight and number of rows (to normalize the weight)
improve classification accuracy on benchmark datasets