Публикации

Nowadays, a lot of scientists’ works aim to improve the quality of people’s life but it could be quite complicated without building a successful collaboration. Productive partnerships can increase research efficiency in many cases and make a huge impact on society. However, today there is no clear way to find such collaborators. In this paper, we propose a recommender system for the scientists from the Higher School of Economics university to help them find co-authors for their prospective studies.

Linear two-timescale stochastic approximation (SA) scheme is an important class of algorithms which has become popular in reinforcement learning (RL), particularly for the policy evaluation problem. Recently, a number of works have been devoted to establishing the finite time analysis of the scheme, especially under the Markovian (non-i.i.d.) noise settings that are ubiquitous in practice. In this paper, we provide a finite-time  analysis for linear two timescale SA. Our bounds show that there is no discrepancy in the convergence rate between Markovian and martingale noise, only the constants are affected by the mixing time of the Markov chain. With an appropriate step size schedule, the transient term in the expected error bound is $o(1/k^c)$ and the steady-state term is ${\cal O}(1/k)$, where $c>1$ and $k$ is the iteration number. Furthermore, we present an asymptotic expansion of the expected error with a matching lower bound of $\Omega(1/k)$. A simple numerical experiment is presented to support our theory.

The chemiluminescence spectra of the samples with nanosized titanium(IV) oxide, both commercial – Hombifine N, Hombikat UV100, Degussa P25, «Anatase I» (Sigma Aldrich), «Anatase II» (Russia), and that prepared in present research by hydrothermal and sol-gel methods (anatase, X-ray amorphous titania, η-phase – TiO2- x×mH2O, and peroxide phase – TiOx(O2)2-x×mH2O) in aqueous media in the presence of luminol or lucigenin under UV-irradiation have been studied. The auxiliary program for data processing and calculating of the Reactive Oxygen Species (ROS: %O2−, %OH radicals, and H2O2) concentration, which correlated with the amount of photoluminescence photons (quantity hν, eV), has been developed. Photocatalytic activity (PCA) of prepared samples has been studied in the reaction of photodegradation of model dyes (methyl orange - MeO and rhodamine 6 G) and harmful organic pollutant (fungicide – difenoconazole, pesticide – thiamethoxam) under UVirradiation. The connection between the rate constant (k, min-1) of photooxidation of MeO and %OH radicals content (in the presence of luminol) and %O2−, H2O2 species content (in the presence of lucigenin) has been described by a transition function of an aperiodic link of the 1st order and a polynomial of 2nd degree, respectively. The maximum k value of the reaction of MeO, difenoconazole, and thiamethoxam photodecomposition under UV irradiation has been achieved in the presence of commercial samples with nanosized anatase, characterized by the optimal crystallite size for photocatalysis (Anatase Sigma-Aldrich) and the maximum specific surface area (Hombifine N «Sachtleben Chemie GmbH»). The k value of the reaction of MeO, difenoconazole, and thiamethoxam photodecomposition under UV irradiation was minimal in the presence of samples containing X-ray amorphous titanium(IV) oxide or η-phase. The role of different factors (sample composition and phase modification, the content of reactive oxygen species, the size of nanoparticles at all levels, specific surface area, etc.) in the exhibition of photocatalytic activity has been discussed.

In this paper, we study the problem of learning graph embeddings for dynamic networks and the ability to generalize to unseen nodes called inductive learning. Firstly, we overview the state-of-the-art methods and techniques for constructing graph embeddings and learning algorithms for both transductive and inductive approaches. Secondly, we propose an improved GSM based on GraphSAGE algorithm and set up the experiments on datasets CORA, Reddit, and HSEcite, which is collected from Scopus citation database across the authors with affiliation to NRU HSE in 2011–2017. The results show that our three-layer model with attention-based aggregation function, added normalization layers, regularization (dropout) outperforms suggested by the respective authors’ GraphSAGE models with mean, LSTM, and pool aggregation functions, thus giving more insight into possible ways to improve inducting learning model based on GraphSAGE model.