Dmitry I. Ignatov

Waleed Khalid, Dmitry Ignatov, Radu Timofte

Large language models (LLMs) excel in program synthesis, yet their capacity for neural architecture design -- balancing syntactic reliability, performance, and structural novelty -- remains underexplored. We present a closed-loop architecture synthesis pipeline within the NNGPT framework, in which a code-oriented LLM evolves over 22 supervised fine-tuning cycles. At each cycle, the LLM synthesizes PyTorch convolutional networks, validated via low-fidelity performance signals and filtered via a MinHash--Jaccard criterion to prevent structural redundancy before being incorporated into the LEMUR dataset. High-performing candidates with novel architectures are converted into prompt--code pairs for parameter-efficient LoRA fine-tuning. This feedback loop drives a measurable distributional shift, progressively internalizing empirical architectural priors such that valid and high-performing outputs evolve from scarce to dominant across cycles. On CIFAR-10, the valid generation rate stabilizes at 50.6% (peaking at 74.5%), mean first-epoch accuracy rises from 28.1% to 51.0%, and candidates exceeding 40% accuracy grow from 2.0% to 96.8%. Cross-dataset transfer to CIFAR-100 and SVHN confirms that improved validity, shifted accuracy distributions, and sustained novelty generalize across benchmarks of varying difficulty and visual domain. Across 22 cycles, 455 unique architectures absent from the original corpus are admitted under the novelty filter. By grounding synthesis in execution feedback and novelty filtering, we demonstrate that iterative self-supervised fine-tuning reshapes an LLM into a task-specialized architectural prior -- improving generation reliability, proxy performance, and structural diversity -- offering a reproducible, annotation-free alternative to hand-crafted search spaces.

Roman Kochnev, Arash Torabi Goodarzi, Zofia Antonina Bentyn et al.

Optimal hyperparameter selection is critical for maximizing the performance of neural networks in computer vision, particularly as architectures become more complex. This work explores the use of large language models (LLMs) for hyperparameter optimization by fine-tuning a parameter-efficient version of Code Llama using LoRA. The resulting model produces accurate and computationally efficient hyperparameter recommendations across a wide range of vision architectures. Unlike traditional methods such as Optuna, which rely on resource-intensive trial-and-error procedures, our approach achieves competitive or superior Root Mean Square Error (RMSE) while substantially reducing computational overhead. Importantly, the models evaluated span image-centric tasks such as classification, detection, and segmentation, fundamental components in many image manipulation pipelines including enhancement, restoration, and style transfer. Our results demonstrate that LLM-based optimization not only rivals established Bayesian methods like Tree-structured Parzen Estimators (TPE), but also accelerates tuning for real-world applications requiring perceptual quality and low-latency processing. All generated configurations are publicly available in the LEMUR Neural Network Dataset (https://github.com/ABrain-One/nn-dataset), which serves as an open source benchmark for hyperparameter optimization research and provides a practical resource to improve training efficiency in image manipulation systems.

Dmitry Lyutkin, Andrey Soloviev, Dmitry Zhukov et al.

The need for skilled medical support is growing in the era of digital healthcare. This research presents an innovative strategy, utilizing the RuBERT model, for categorizing user inquiries in the field of medical consultation with a focus on expert specialization. By harnessing the capabilities of transformers, we fine-tuned the pre-trained RuBERT model on a varied dataset, which facilitates precise correspondence between queries and particular medical specialisms. Using a comprehensive dataset, we have demonstrated our approach's superior performance with an F1-score of over 92%, calculated through both cross-validation and the traditional split of test and train datasets. Our approach has shown excellent generalization across medical domains such as cardiology, neurology and dermatology. This methodology provides practical benefits by directing users to appropriate specialists for prompt and targeted medical advice. It also enhances healthcare system efficiency, reduces practitioner burden, and improves patient care quality. In summary, our suggested strategy facilitates the attainment of specific medical knowledge, offering prompt and precise advice within the digital healthcare field.

Robiul Islam, Dmitry I. Ignatov, Karl Kaberg et al.

This study investigates classifier performance across EEG frequency bands using various optimizers and evaluates efficient class prediction for the left and right hemispheres. Three neural network architectures - a deep dense network, a shallow three-layer network, and a convolutional neural network (CNN) - are implemented and compared using the TensorFlow and PyTorch frameworks. Results indicate that the Adagrad and RMSprop optimizers consistently perform well across different frequency bands, with Adadelta exhibiting robust performance in cross-model evaluations. Specifically, Adagrad excels in the beta band, while RMSprop achieves superior performance in the gamma band. Conversely, SGD and FTRL exhibit inconsistent performance. Among the models, the CNN demonstrates the second highest accuracy, particularly in capturing spatial features of EEG data. The deep dense network shows competitive performance in learning complex patterns, whereas the shallow three-layer network, sometimes being less accurate, provides computational efficiency. SHAP (Shapley Additive Explanations) plots are employed to identify efficient class prediction, revealing nuanced contributions of EEG frequency bands to model accuracy. Overall, the study highlights the importance of optimizer selection, model architecture, and EEG frequency band analysis in enhancing classifier performance and understanding feature importance in neuroimaging-based classification tasks.

Léonard Kwuida, Dmitry I. Ignatov

Machine Learning (ML) provides important techniques for classification and predictions. Most of these are black-box models for users and do not provide decision-makers with an explanation. For the sake of transparency or more validity of decisions, the need to develop explainable/interpretable ML-methods is gaining more and more importance. Certain questions need to be addressed: How does an ML procedure derive the class for a particular entity? Why does a particular clustering emerge from a particular unsupervised ML procedure? What can we do if the number of attributes is very large? What are the possible reasons for the mistakes for concrete cases and models? For binary attributes, Formal Concept Analysis (FCA) offers techniques in terms of intents of formal concepts, and thus provides plausible reasons for model prediction. However, from the interpretable machine learning viewpoint, we still need to provide decision-makers with the importance of individual attributes to the classification of a particular object, which may facilitate explanations by experts in various domains with high-cost errors like medicine or finance. We discuss how notions from cooperative game theory can be used to assess the contribution of individual attributes in classification and clustering processes in concept-based machine learning. To address the 3rd question, we present some ideas on how to reduce the number of attributes using similarities in large contexts.

Dmitry Egurnov, Dmitry I. Ignatov, Dmitry Tochilkin

In this paper, we describe versions of triclustering algorithms adapted for efficient calculations in distributed environments with MapReduce model or parallelisation mechanism provided by modern programming languages. OAC-family of triclustering algorithms shows good parallelisation capabilities due to the independent processing of triples of a triadic formal context. We provide the time and space complexity of the algorithms and justify their relevance. We also compare performance gain from using a distributed system and scalability.

Dmitry I. Ignatov, Gennady V. Khvorykh, Andrey V. Khrunin et al.

Missing genotypes can affect the efficacy of machine learning approaches to identify the risk genetic variants of common diseases and traits. The problem occurs when genotypic data are collected from different experiments with different DNA microarrays, each being characterised by its pattern of uncalled (missing) genotypes. This can prevent the machine learning classifier from assigning the classes correctly. To tackle this issue, we used well-developed notions of object-attribute biclusters and formal concepts that correspond to dense subrelations in the binary relation $\textit{patients} \times \textit{SNPs}$. The paper contains experimental results on applying a biclustering algorithm to a large real-world dataset collected for studying the genetic bases of ischemic stroke. The algorithm could identify large dense biclusters in the genotypic matrix for further processing, which in return significantly improved the quality of machine learning classifiers. The proposed algorithm was also able to generate biclusters for the whole dataset without size constraints in comparison to the In-Close4 algorithm for generation of formal concepts.

Taisia Glushkova, Alexey Machnev, Alena Fenogenova et al.

DaNetQA, a new question-answering corpus, follows (Clark et. al, 2019) design: it comprises natural yes/no questions. Each question is paired with a paragraph from Wikipedia and an answer, derived from the paragraph. The task is to take both the question and a paragraph as input and come up with a yes/no answer, i.e. to produce a binary output. In this paper, we present a reproducible approach to DaNetQA creation and investigate transfer learning methods for task and language transferring. For task transferring we leverage three similar sentence modelling tasks: 1) a corpus of paraphrases, Paraphraser, 2) an NLI task, for which we use the Russian part of XNLI, 3) another question answering task, SberQUAD. For language transferring we use English to Russian translation together with multilingual language fine-tuning.

Dmitry I. Ignatov, Polina Ivanova, Albina Zamaletdinova

This paper is related to the problem of finding the maximal quasi-bicliques in a bipartite graph (bigraph). A quasi-biclique in the bigraph is its "almost" complete subgraph. The relaxation of completeness can be understood variously; here, we assume that the subgraph is a $γ$-quasi-biclique if it lacks a certain number of edges to form a biclique such that its density is at least $γ\in (0,1]$. For a bigraph and fixed $γ$, the problem of searching for the maximal quasi-biclique consists of finding a subset of vertices of the bigraph such that the induced subgraph is a quasi-biclique and its size is maximal for a given graph. Several models based on Mixed Integer Programming (MIP) to search for a quasi-biclique are proposed and tested for working efficiency. An alternative model inspired by biclustering is formulated and tested; this model simultaneously maximizes both the size of the quasi-biclique and its density, using the least-square criterion similar to the one exploited by triclustering \textsc{TriBox}.

Artem M. Grachev, Dmitry I. Ignatov, Andrey V. Savchenko

Recurrent neural networks have proved to be an effective method for statistical language modeling. However, in practice their memory and run-time complexity are usually too large to be implemented in real-time offline mobile applications. In this paper we consider several compression techniques for recurrent neural networks including Long-Short Term Memory models. We make particular attention to the high-dimensional output problem caused by the very large vocabulary size. We focus on effective compression methods in the context of their exploitation on devices: pruning, quantization, and matrix decomposition approaches (low-rank factorization and tensor train decomposition, in particular). For each model we investigate the trade-off between its size, suitability for fast inference and perplexity. We propose a general pipeline for applying the most suitable methods to compress recurrent neural networks for language modeling. It has been shown in the experimental study with the Penn Treebank (PTB) dataset that the most efficient results in terms of speed and compression-perplexity balance are obtained by matrix decomposition techniques.

Artem M. Grachev, Dmitry I. Ignatov, Andrey V. Savchenko

In this paper, we consider several compression techniques for the language modeling problem based on recurrent neural networks (RNNs). It is known that conventional RNNs, e.g, LSTM-based networks in language modeling, are characterized with either high space complexity or substantial inference time. This problem is especially crucial for mobile applications, in which the constant interaction with the remote server is inappropriate. By using the Penn Treebank (PTB) dataset we compare pruning, quantization, low-rank factorization, tensor train decomposition for LSTM networks in terms of model size and suitability for fast inference.

Dmitry Ignatov, Andrey Ignatov

Various modifications of decision trees have been extensively used during the past years due to their high efficiency and interpretability. Tree node splitting based on relevant feature selection is a key step of decision tree learning, at the same time being their major shortcoming: the recursive nodes partitioning leads to geometric reduction of data quantity in the leaf nodes, which causes an excessive model complexity and data overfitting. In this paper, we present a novel architecture - a Decision Stream, - aimed to overcome this problem. Instead of building a tree structure during the learning process, we propose merging nodes from different branches based on their similarity that is estimated with two-sample test statistics, which leads to generation of a deep directed acyclic graph of decision rules that can consist of hundreds of levels. To evaluate the proposed solution, we test it on several common machine learning problems - credit scoring, twitter sentiment analysis, aircraft flight control, MNIST and CIFAR image classification, synthetic data classification and regression. Our experimental results reveal that the proposed approach significantly outperforms the standard decision tree learning methods on both regression and classification tasks, yielding a prediction error decrease up to 35%.

Dmitry I. Ignatov, Alexander Semenov, Daria Komissarova et al.

Multimodal clustering is an unsupervised technique for mining interesting patterns in $n$-adic binary relations or $n$-mode networks. Among different types of such generalized patterns one can find biclusters and formal concepts (maximal bicliques) for 2-mode case, triclusters and triconcepts for 3-mode case, closed $n$-sets for $n$-mode case, etc. Object-attribute biclustering (OA-biclustering) for mining large binary datatables (formal contexts or 2-mode networks) arose by the end of the last decade due to intractability of computation problems related to formal concepts; this type of patterns was proposed as a meaningful and scalable approximation of formal concepts. In this paper, our aim is to present recent advance in OA-biclustering and its extensions to mining multi-mode communities in SNA setting. We also discuss connection between clustering coefficients known in SNA community for 1-mode and 2-mode networks and OA-bicluster density, the main quality measure of an OA-bicluster. Our experiments with 2-, 3-, and 4-mode large real-world networks show that this type of patterns is suitable for community detection in multi-mode cases within reasonable time even though the number of corresponding $n$-cliques is still unknown due to computation difficulties. An interpretation of OA-biclusters for 1-mode networks is provided as well.

Dmitry I. Ignatov, Bruce W. Watson

Being an unsupervised machine learning and data mining technique, biclustering and its multimodal extensions are becoming popular tools for analysing object-attribute data in different domains. Apart from conventional clustering techniques, biclustering is searching for homogeneous groups of objects while keeping their common description, e.g., in binary setting, their shared attributes. In bioinformatics, biclustering is used to find genes, which are active in a subset of situations, thus being candidates for biomarkers. However, the authors of those biclustering techniques that are popular in gene expression analysis, may overlook the existing methods. For instance, BiMax algorithm is aimed at finding biclusters, which are well-known for decades as formal concepts. Moreover, even if bioinformatics classify the biclustering methods according to reasonable domain-driven criteria, their classification taxonomies may be different from survey to survey and not full as well. So, in this paper we propose to use concept lattices as a tool for taxonomy building (in the biclustering domain) and attribute exploration as means for cross-domain taxonomy completion.

Sheikh Muhammad Sarwar, Mahamudul Hasan, Dmitry I. Ignatov

In this paper we describe our machine learning solution for the RecSys Challenge, 2015. We have proposed a time efficient two-stage cascaded classifier for the prediction of buy sessions and purchased items within such sessions. Based on the model, several interesting features found, and formation of our own test bed, we have achieved a reasonable score. Usage of Random Forests helps us to cope with the effect of the multiplicity of good models depending on varying subsets of features in the purchased items prediction and, in its turn, boosting is used as a suitable technique to overcome severe class imbalance of the buy-session prediction.

Dmitry I. Ignatov, Denis Kornilov

We propose a new algorithm for recommender systems with numeric ratings which is based on Pattern Structures (RAPS). As the input the algorithm takes rating matrix, e.g., such that it contains movies rated by users. For a target user, the algorithm returns a rated list of items (movies) based on its previous ratings and ratings of other users. We compare the results of the proposed algorithm in terms of precision and recall measures with Slope One, one of the state-of-the-art item-based algorithms, on Movie Lens dataset and RAPS demonstrates the best or comparable quality.

Yury Kashnitsky, Dmitry I. Ignatov

The paper briefly introduces multiple classifier systems and describes a new algorithm, which improves classification accuracy by means of recommendation of a proper algorithm to an object classification. This recommendation is done assuming that a classifier is likely to predict the label of the object correctly if it has correctly classified its neighbors. The process of assigning a classifier to each object is based on Formal Concept Analysis. We explain the idea of the algorithm with a toy example and describe our first experiments with real-world datasets.

Rustam Tagiew, Dmitry I. Ignatov

This paper presents an analysis of data from a gift-exchange-game experiment. The experiment was described in `The Impact of Social Comparisons on Reciprocity' by Gächter et al. 2012. Since this paper uses state-of-art data science techniques, the results provide a different point of view on the problem. As already shown in relevant literature from experimental economics, human decisions deviate from rational payoff maximization. The average gift rate was $31$%. Gift rate was under no conditions zero. Further, we derive some special findings and calculate their significance.

Anastasia Bezzubtseva, Dmitry I. Ignatov

In this paper we present a review of the existing typologies of Internet service users. We zoom in on social networking services including blogs and crowdsourcing websites. Based on the results of the analysis of the considered typologies obtained by means of FCA we developed a new user typology of a certain class of Internet services, namely a collaboration innovation platform. Cluster analysis of data extracted from the collaboration platform Witology was used to divide more than 500 participants into six groups based on three activity indicators: idea generation, commenting, and evaluation (assigning marks) The obtained groups and their percentages appear to follow the "90 - 9 - 1" rule.

Elena Nenova, Dmitry I. Ignatov, Andrey V. Konstantinov

We propose a new approach for Collaborative Filtering which is based on Boolean Matrix Factorisation (BMF) and Formal Concept Analysis. In a series of experiments on real data (Movielens dataset) we compare the approach with the SVD- and NMF-based algorithms in terms of Mean Average Error (MAE). One of the experimental consequences is that it is enough to have a binary-scaled rating data to obtain almost the same quality in terms of MAE by BMF than for the SVD-based algorithm in case of non-scaled data.

Jonas Poelmans, Paul Elzinga, Alexey Neznanov et al.

Concept Relation Discovery and Innovation Enabling Technology (CORDIET), is a toolbox for gaining new knowledge from unstructured text data. At the core of CORDIET is the C-K theory which captures the essential elements of innovation. The tool uses Formal Concept Analysis (FCA), Emergent Self Organizing Maps (ESOM) and Hidden Markov Models (HMM) as main artifacts in the analysis process. The user can define temporal, text mining and compound attributes. The text mining attributes are used to analyze the unstructured text in documents, the temporal attributes use these document's timestamps for analysis. The compound attributes are XML rules based on text mining and temporal attributes. The user can cluster objects with object-cluster rules and can chop the data in pieces with segmentation rules. The artifacts are optimized for efficient data analysis; object labels in the FCA lattice and ESOM map contain an URL on which the user can click to open the selected document.

Dmitry I. Ignatov, Jonas Poelmans, Vasily Zaharchuk

In this paper we propose two new algorithms based on biclustering analysis, which can be used at the basis of a recommender system for educational orientation of Russian School graduates. The first algorithm was designed to help students make a choice between different university faculties when some of their preferences are known. The second algorithm was developed for the special situation when nothing is known about their preferences. The final version of this recommender system will be used by Higher School of Economics.