ὅδε οἶκος, ὦ ἑταῖρε, μνημεῖον ἐστιν ζῴων τῶν σοφῶν ἀνδρῶν, καὶ τῶν ἔργων αὐτῶν

**Seminar on Computer Science and Applied Mathematics**

**PROGRAM**

Knez Mihajlova 36

Fakultet organizacionih nauka, Univerzitet u Beogradu,

Jove Ilica 154

IEEE Chapter Computer Science (CO-16) Belgrade, Republic of Serbia

MI SANU, Knez Mihailova 36, sala 301f

Upravni odbor Matematickog instituta SANU je na nedavnoj sednici doneo odluku da se dosadasnji Seminar za primenjenu matematiku, sada nazove Seminar za racunarstvo i primenjenu matematiku, a u cilju potenciranja znacaja racunarstva kao jedne od oblasti delatnosti Instituta. Istovremeno, Upravni odbor doneo je odluku o osnivanju Odeljenja za racunarstvo i primenjenu matematiku i vezao rad novog odeljenja za rad Seminara za racunarstvo i primenjenu matematiku.

**PETAK, 01.09.2017. u 14:15, Sala 301f, MI SANU, Kneza Mihaila 36**

*Zoran Obradović, L.H. Carnell Professor of Data Analytics, Temple University, Philadelphia, USA*
**STRUCTURED REGRESSION IN LARGE TEMPORAL NETWORKS**

In the first part of this talk we will present a novel sampling-based structured regression method for prediction on top of temporal networks.
The algorithm allows efficient learning of an ensemble model by automatically skiping the entire re-training or some phases of the training process in an evolving environment.
In conducted experiments the new method was about 140 time faster than alternative structured regression approaches while it was also more accurate as evident on modeling the H3N2 Virus Influenza network.
The second part of the talk will describe an efficient algorithm to uncover the underlying dependency structure in high dimensional data.
This is achieved by relaying on Cholesky decomposition to learn a sparse Gaussian Markov Random Field.
The new method is applied to discover the connectivity structure among gene expressions in septic patients.

Results reported in this talk are published at:

- Pavlovski, M., Zhou, F., Stojkovic, I., Kocarev, Lj., Obradovic, Z., "Adaptive Skip-Train Structured Regression for Temporal Networks", Proc. European Conf. Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML-PKDD), Sept. 2017.
- Stojkovic, I., Jelisavcic, V., Milutinovic, V., Obradovic, Z., "Fast Sparse Gaussian Markov Random Fields Learning Based on Cholesky Factorization", Proc. 26th Int’l Joint Conf. Artificial Intelligence (IJCAI), Aug. 2017.

(

Over the last years, Data Envelopment Analysis (DEA) has turned into probably the most widely used econometric tool for efficiency evaluation, while its "cousin" Free Disposal Hull (FDH) has received far less attention for practical applicability reasons. In this work some recent advances addressing practical applicability of these methods is presented. It is first shown how "granular" versions of Multiple Data Envelopment Analysis (MDEA) and Multiple Free Disposal Hull (MFDH) may enhance the performance of these methods. Next, the Jackstrap technique (combination of Jackknife and Bootstrap) for identification of outliers is addressed, and finally several parallelization strategies are compared.

Large-scale graphs are stored and managed in the graph database systems where each arc is represented by a triple: (subject, predicate, object). The scalability of the storage system and the query processing engine for managing from Tera towards Peta triples is currently possible solely by using the distribution of data into the large shared-nothing clusters. The query execution system in such environment must be able to employ various types of parallelism to allow simultaneous execution of queries.

The seminar will present the problems and some solutions in the design and the implementation of large-scale distributed graph database system big3store. The system is implemented in Erlang programming environment. It is based on the dataflow architecture of query processing---each query is a tree of algebra operations that is dynamically mapped to the tree composed of processes interconnected by streams of graphs. The scheduler that maps query trees to the set of processes balances the computation load among the servers of the cluster. One of the leading ideas that has shaped the architecture of big3store is the use of the higher-level semantic data stored in the graph database, i.e., the so-called knowledge graph, to define the distribution of the graph database, as well as, to lead the query processing. The development of the big3store system is a joint project between Yahoo Japan Research and University of Primorska.

In this lecture we present recent advances in optimization software packages. In particular, we will cover the JuMP, which consists an open-source modeling language that allows users to express a wide range of optimization problems (linear, mixed-integer, quadratic, conic-quadratic, semidefinite, and nonlinear) in a high-level, algebraic syntax. JuMP takes advantage of advanced features of the Julia high-level, high-performance, dynamic programming language for numerical computing. We also show here, how to solve mixed-integer optimization problems using the state-of-the-art Gurobi Optimizer, which exploits modern architectures and multi-core processors.

Related references:

- Dunning I., Huchette J., and Lubin M., “JuMP: A modeling language for mathematical optimization”, SIAM Review, Vol. 59, No. 2, pp. 295-320, 2017.
- Changhyun Kwon, Julia Programming for Operations Research: A Primer on Computing, CreateSpace Independent Publishing Platform, 2016.

** RUKOVODIOCI SEMINARA **

** MI SANU Vera Kovačević-Vujčić Milan Dražić FON Zorica Bogdanovic Marijana Despotovic-Zrakic IEEE Bozidar Radenkovic **