Zharko Aleksovski

Using Google distance to weight approximate ontology matches

Discovering mappings between concept hierarchies is widely regarded as one of the hardest and most urgent problems facing the Semantic Web. The problem is even harder in domains where concepts are inherently vague and ill-defined, and cannot be given a crisp definition. A notion of approximate concept mapping is required in such domains, but until now, no such notion is vailable. The first contribution of this paper is a definition for approximate mappings between concepts. Roughly, a mapping between two concepts is decomposed into a number of submappings, and a sloppiness value determines the fraction of these submappings that can be ignored when establishing the mapping. A potential problem of such a definition is that with an increasing sloppiness value, it will gradually allow mappings between any two arbitrary concepts. To improve on this trivial behaviour, we need to design a heuristic weighting which minimises the sloppiness required to conclude desirable matches, but at the same time maximises the sloppiness required to conclude undesirable matches. The second contribution of this paper is to show that a Google based similarity measure has exactly these desirable properties. We establish these results by experimental validation in the domain of musical genres. We show that this domain does suffer from ill-defined concepts. We take two real-life genre hierarchies from the Web, we compute approximate mappings between them at varying levels of sloppiness, and we validate our results against a handcrafted Gold Standard. Our method makes use of the huge amount of knowledge that is implicit in the current Web, and exploits this knowledge as a heuristic for establishing approximate mappings between ill-defined concepts.

Matching unstructured vocabularies using a background ontology

Existing ontology matching algorithms use a combination of lexical and structural correspondence between source and target ontologies. We present a realistic case-study where both types of overlap are low: matching two unstructured lists of vocabulary used to describe patients at Intensive Care Units in two different hospitals. We show that indeed existing matchers fail on our data. We then discuss the use of background knowledge in ontology matching problems. In particular, we discuss the case where the source and the target ontology are of poor semantics, such as flat lists, and where the background knowledge is of rich semantics, providing extensive descriptions of the properties of the concepts involved. We evaluate our results against a Gold Standard set of matches that we obtained from human experts.

Identifying disease-centric subdomains in very large medical ontologies: a case-study on breast cancer concepts in SNOMED CT. or: Finding 2500 Out of 300.000

Modern medical vocabularies can contain up to hundreds of thousands of concepts. In any particular use-case only a small fraction of these will be needed. In this paper we first define two notions of a disease-centric subdomain of a large ontology. We then explore two methods for identifying disease-centric subdomains of such large medical vocabularies. The first method is based on lexically querying the ontology with an iteratively extended set of seed queries. The second method is based on manual mapping between concepts from a medical guideline document and ontology concepts. Both methods include concept-expansion over subsumption and equality relations. We use both methods to determine a breast-cancer-centric subdomain of the SNOMED CT ontology. Our experiments show that the two methods produce a considerable overlap, but they also yield a large degree of complementarity, with interesting differences between the sets of concepts that they return. Analysis of the results reveals strengths and weaknesses of the different methods.

Ontology mapping using background knowledge

In this paper, we report on a method for aligning two lists of terms using structure-rich ontologies as background knowledge. The results of the method can be seen as suggested mapping candidates to users that perform an ontology alignment task. We applied the method on lists of medical terms, and we discuss the outcome.

Iterative node localization for intelligent street lighting

In intelligent street lighting systems where light posts communicate wirelessly, location awareness is necessary for the system to provide context-aware services. To solve the specific localization problem, we developed an iterative algorithm based on the relation between the received signal strength and the distance to assign the known locations to the nodes. Our experimental results show small error rates in the number of wrongly localized nodes, which indicates the algorithm is applicable in practice.

Identifying Breast Cancer Concepts in SNOMED-CT Using Large Text Corpus

Large medical ontologies can be of great help in building a specialized clinical information system. First step in their use is to identify the subset of concepts which are relevant to the specialty. In this paper we present a method to automatically identify the breast cancer concepts from the SNOMED-CT ontology using large text corpus as source of knowledge. In addition to finding them, the concepts are also assigned relevance values.

Semantic Mapping by Approximation

We address the problem of semantic coordination, namely finding an agreement between the meanings of heterogeneous semantic models. We propose a new approximation method to discover and assess the “strength” (preciseness) of semantic mappings between different concept hierarchies. We apply this method in the music domain. We present the results of tests on mapping two music concept hierarchies from actual sites on the Internet.

Using lexical and logical methods for the alignment of medical terminologies

Standardized medical terminologies are often used for the registration of patient data. In several situations there is a need to align these terminologies to other terminologies. Even when the terminologies cover the same domain, this is often a non-trivial task. The task is even more complicated when the terminology does not contain much structure. In this paper we describe the initial results of a procedure for mapping a terminology with little or no structure to a structure-rich terminology. This procedure uses the knowledge of the structure-rich terminology and a method for semantic explicitation of concept descriptions. The first results shows that, when compared to approaches based on syntactic analysis only, the recall can be greatly improved without sacrificing much of the precision.

Abstract 3651: Normal versus tumor and subtype prediction in renal cell carcinoma TCGA data sets

Background: In clinical practice it is a common challenge to correctly classify disease against normal cases and to identify disease subtypes. In complex diseases such as cancer where patterns are heterogeneous, highly complex interaction of pathways are involved, and continuous multi-level genomic changes occur, this is a challenging task. Using genomics data, analyzed by un-supervised and supervised machine learning tools, we demonstrate the ability to quantitatively and accurately describe a patient9s tumor. Design: We used Level III RNASeq gene expression data from 20531 genes in 889 tumor samples of RCC across three subtypes - clear cell renal cell carcinoma (CCRCC), papillary renal cell carcinoma (PRCC), chromophobe carcinoma (ChRCC) and 129 normal samples from the Cancer Genome Atlas (TCGA). We developed a computational framework for feature (gene/transcript) selection and subtype predictive model construction. This framework relies on a well-known “random forest” (RF) method with iterative feature selection and 10-fold cross-validation. We performed a series of (1) tumor vs normal tissue experiments for each subtype; (2) pairwise subtype comparison, and finally (3) all three subtypes comparison and predictive genes identification. Results: In each computational run our method detected 2054 (10%) top varying genes, and estimated the predictive power of each of the selected genes using RF. On average this method demonstrated 93-97% accuracy. We identified genes known to play a role in renal cell carcinoma for example, CA9, LOX, SFRP1, SLC4A1, CDKN2A, KISS1R, EGF and others. In addition, our analysis uncovered genes that may represent characteristic patterns for subtyping and differentiation from normal renal tissue cells, for example TCF21, IRX1, STC2, UMOD, AQP2, ANGPTL4, BSND and FABP7, genes not previously associated with renal cell carcinoma. In three different experiments we differentiated each of the three subtypes from normal tissue, and performed enrichment analysis for the top most significant genes in each case. We observed that both CCRCC and PRCC have genes involved in the “glycosaminoglycan biosynthesis - heparan sulfate” (HS6ST2, HS6ST3) and “riboflavin metabolism” (ACPP) pathways. Whereas ChRCC is more strongly associated with the “glycosphingolipid biosynthesis - lacto and neolacto series” pathway (B3GNT3, FUT6) and have five genes (B3GNT3, CYP2B6, CYP2J2, FUT6, UGT2A3) involved in other metabolic pathways. Changes in glycosaminoglycan and glycolipid were also previously reported for associations with RCC. Conclusions: We demonstrate the effectiveness of a computational framework and predictive power of gene expression data for tumor subtyping in RCC. Our framework is generic and can be applied in combination with other types of data such as different modalities of genomic data (copy number variations, methylation) as well as clinical data. Citation Format: Konstantin Volyanskyy, Yong Mao, Yee Him Cheung, Balaji Santhanam, Vlado Menkovski, Zharko Aleksovski, Minghao Zhong, John T. Fallon, Nevenka Dimitrova. Normal versus tumor and subtype prediction in renal cell carcinoma TCGA data sets. [abstract]. In: Proceedings of the 107th Annual Meeting of the American Association for Cancer Research; 2016 Apr 16-20; New Orleans, LA. Philadelphia (PA): AACR; Cancer Res 2016;76(14 Suppl):Abstract nr 3651.

Learning to sort by using evolution

This paper present a work where Genetic Programming (GP) was used to the task of evolving imperative sort programs. A variety of interesting lessons were learned. With proper selection of the primitives, sorting programs were evolved that are both general and non-trivial. Unique aspect of our approach is that we represent the individual programs with simple assembler code, rather than usual tree like structure. We also report the effect of different parameters on quality of the programs and time needed for finding the solution.