Current members

Carlos R. Rivero
Associate Professor in CS@RIT
   

Narayanan Asuri Krishnan
PhD student
Email: nk1581

Bhaskar Krishna Gangadhar
MS student
Email: bg4437

Md Towhidul Chowdhury
PhD student
(Main advisor: Naveen Sharma)
Email: mac9908

Former members

Iti Bansal

Angel Cambero

Andrea Cimmino

Maheen Riaz Contractor

Mayur Jawalkar

Victor J. Marin

Wilberto Z. Nunez

Michael Peechatt

Tobin Pereira

Aishwarya Rao

Srinivas Sridharan

Sudhanshu Tiwari

Gabriella Wolf

Mainstream topics

Knowledge graphs

Even though knowledge graphs have proven very useful for several tasks, they are marked by incompleteness. Completion algorithms aim to extend knowledge graphs by predicting missing (subject, predicate, object) triples, usually by training a model to discern between correct (positive) and incorrect (negative) triples. However, under the open-world assumption in which a missing triple is not negative but unknown, negative triple generation is challenging. We aim to thoroughly evaluate knowledge graph completion.

Program comprehension

Many software-related activities require comprehending programs that other programmers wrote. In the majority of such scenarios, program comprehension is a manual process since current approaches cannot adequately handle program variability and, in particular, interleaved tasks, i.e., sets of non-contiguous program statements with specific semantic purposes. We apply recent advances in graph databases to programs modeled as system dependence graphs. We rely on subgraph patterns to model tasks and subgraph matching to compute pattern occurrences.
Additional info: System Dependence Graph builder, ARCC.

Efficient subgraph matching

Subgraph matching is a very hard problem and it has many important applications in graph databases to retrieve data represented as graphs, such as biological networks, chemical compounds, geographic maps, social networks and more. There exist several algorithms that exploit different heuristics and indexing structures to perform subgraph matching efficiently in practice. In general, it is not clear how these algorithms compare in terms of performance and/or accuracy. We aim to build a framework on top of Neo4j to fairly compare side by side current and future subgraph matching algorithms. This framework will allow to build intelligent query engines able to estimate, given a subgraph query and a data graphs, which algorithm will have the best performance.

Other topics

• Link discovery. In the context of the Web of Data, there are datasets that usually store entities that refer to the same real-world concept. Link discovery focuses on the generation of high-level rules that indicate how such entities relate. Existing techniques exploit genetic programming algorithms to compute the rules based on positive and negative examples of related entities. In general, genetic programming algorithms are very sensible to configuration parameters and different parameter combinations may result in the computation of completely different rules. We aim to perform in-depth evaluations of existing algorithms regarding these combinations, and to develop genetic programming operators that take sets of entities into account.
• Data exchange. Due to their distributed nature, there exist a large amount of heterogeneous RDF datasets and there is a strong need for exchanging data among them, i.e., populating a target dataset using data that come from one or more source datasets. Data exchange may be implemented using correspondences that are later transformed into executable mappings; however, this is not a trivial task, so mechanisms to aid exchanging data among RDF datasets are a must. Correspondences are hints that specify which entities in the source and target RDF datasets correspond to each other, i.e., are somewhat related. Our algorithm automatically generates SPARQL executable mappings in the context of RDF datasets that rely on a subset of the OWL 2 Lite profile.
• URL classification. Unsupervised web page classification refers to the problem of clustering the pages in a web site so that each cluster includes a set of web pages that can be classified using a unique class. Our algorithm automatically builds a number of URL patterns that represent the different classes of pages in a web site, so further pages can be classified by matching their URLs to the patterns. Its salient features are that it is based on lightweight crawling to avoid interfering with the normal operation of the web site, it is unsupervised to avoid the need for a training set of pre-classified pages, and it uses features from outside the page to be classified to avoid downloading it.

Contact

We are in the third floor of the Golisano College of Computing and Information Sciences. If you visit us, you need to get a temporary parking permit at the Welcome Center. The closest lots are J and S.