The paper presents an algorithm for finding a threshold - the value of a discrete attribute - which optimal splits a set of examples into two subsets, where the optimality is defined with regard to a criterion which is not computed over individual examples (like entropy) but on pairs of examples. The algorithm reduces the time complexity from quadratic, which we would get with a brute force algorithm, to linear. We also developed a similar algorithm for subsetting of values of discrete attributes.
COBISS.SI-ID: 7550548
The paper describes the developed formalism and methodology for explaining instance classification for arbitrary model. The formalism is based on the interaction of each subset of feature values. The empirical evaluation has confirmed the intuitiveness of such explanation
COBISS.SI-ID: 7252308
We compare diferent approaches to estimate the reliability of individual predictions in regression. By combining pairs of individual estimates, we compose a combined estimate that performs better than the individual estimates. The results demonstrate the potential of a sensitivity-based estimate, as well as the bagging variance approach, which achieved the best performance with neural networks, bagging and locally weighted regression.
COBISS.SI-ID: 6923604
We have developed an interactive web application called dictyExpress (www.ailab.si/dictyexpress), which offers a complex gene expression data analytics in a easy-to-use graphical interface. The methods behind dictyExpress and the design of the application was done in collaboration with Baylor College of Medicine. dictyExpress is directly linked from D. discoideum's genome home page (www.dictybase.org) and is in frequent use by researchers worldwide since October 2009.
COBISS.SI-ID: 7219028
The search for interactions between single nucleotide polymorphisms (SNPs) generates a multitude of hypotheses. As the follow-up testing is expensive we would like to decrease the false positive rate. We tested whether a previously published methodology (Gayan et al., 2008). On the contrary to Gayan et al. we discovered that the methodology perfomed worse than the naive approach. In our article we discuss why this is so. Additionaly, we set up an environment for the comparison of SNP interaction search methods, which will be useful for further research.
COBISS.SI-ID: 7553364