Published Papers

Emotion attribution in social graphs requires inferring directed emotional attitudes between entities in complex, multi-turn dialogues. While transformer models dominate the field, they often lack the transparency required for social science applications. We present a systematic comparison of three modeling paradigms for this task: a fully interpretable rule-based system, a fine-tuned RoBERTa-large model, and a few-shot Llama-3-8B. Utilizing the DialogRE dataset, we demonstrate that incorporating a 3-turn conversational context significantly improves attribution accuracy across paradigms. Crucially, our results show that the interpretable rule-based system achieves a competitive F1 score, making it statistically indistinguishable from the state-of-the-art RoBERTa model. In contrast, we find that few-shot large language models exhibit poor performance in emotion attribution of semantic relations, performing below the rule-based baseline. We provethat interpretability does not necessitate a performance trade-off in social emotion analysis

Modern automated resume screening systems are typically based on neural text classification models that encode a resume as a feature representation and predict a discrete label corresponding to candidate category, suitability level, or job role. Such models commonly produce class logits parameterized by model weights, which are converted into class probabilities via the softmax function over the target classes. These models are typically trained using cross-entropy loss and deployed as the first stage of automated candidate filtering. Despite their effectiveness, resume classifiers may encode implicit bias through correlations between predictions and non-job-related or proxy textual features. To study this effect, we analyze feature influence using Integrated Gradients, which assign an attribution score to each input feature based on the path integral of partial derivatives between a baseline representation and the actual input. This analysis reveals systematic dependencies on features that should be irrelevant to candidate evaluation. Building on these observations, we evaluate multiple debiasing techniques and propose an interpretability-guided framework for bias mitigation. We compare six methods spanning in-processing approaches (GroupDRO, Focal Loss, Label Smoothing, Adversarial debiasing) and attribution-based approaches (Data Scrubbing, Attention Regularization) that leverage the interpretability findings directly. This allows explainable analysis to guide the development of fairer resume screening models.

Artificial intelligence systems are now integral to virtually every facet of our lives, exhibiting an ability to reason and solve problems within defined formal frameworks. However, challenges remain, particularly the issue of hallucination—where AI systems generate incorrect or misleading information. This paper proposes a task-based approach to building reliable AI systems, focusing on the task itself and the criteria necessary for its resolution. Our objective is to ensure that AI systems not only provide solutions but also possess an understanding of the underlying limitations of the problem. This includes identifying the axioms and theorems involved, allowing the solution process to be informed by a clear comprehension of the problem’s structure and constraints.

We study intraday canonicalization and auditing of implied-volatility (IV) surfaces for S&P 500 index (SPX) options when the available observations are sparse, trade-based open–high–low–close–volume (OHLCV) bars. Such panels exhibit strongly non-uniform cross-sectional support—dense near at-the-money (ATM) and short maturities, and thin in the wings and long tenors—so naive fixed-grid completion can create visually smooth surfaces with unstable calibration and misleading feasibility diagnostics. We propose a reliability-aware canonicalization pipeline that (i) constructs forward and discount proxies from liquid futures strips, (ii) quantifies local information content via an effective sample size (ESS) diagnostic induced by kernel receptive fields, and (iii) calibrates a per-timestamp Surface SVI (SSVI) total-variance surface directly in price space under robust losses. Our main methodological contribution is a reliability-weighted vega-normalized robust objective that downweights weakly supported marks while retaining the interpretability and tractability of price-space calibration. We compare three calibration objectives—robust price residuals, vega-normalized residuals, and reliability-weighted vega residuals—using paired timestamp-level inference with Holm-adjusted randomization tests, and we evaluate economic plausibility with a unified static-arbitrage audit that reports both violation rates and hinge-type severity measures. Empirically, price-residual calibration minimizes price root-mean-squared error (RMSE), whereas reliability-weighted vega residuals yield the most consistent reductions in out-of-sample (OOS) arbitrage severity and in average any-rule violation rates. These results support reliability-weighted robust objectives as default canonicalizers for economically plausible intraday surfaces built from sparse transaction bars and for downstream learning tasks.

This paper presents a methodology for forming an adaptive investment portfolio structure based on a multi-module architecture implemented using an ensemble of neural network models. The developed system comprises five specialized analytical modules, each responsible for processing a specific class of factors: macroeconomic, fundamental, technical, and credit. The central element of the architecture is a neural network model for forecasting the Bank of Russia key interest rate, which defines scenario conditions for the valuation of all asset classes. Fundamental equity analysis, technical analysis of market dynamics, credit risk assessment of debt instruments, and an adapted Markowitz portfolio optimization model are integrated within a digital twin that reconciles the outputs of individual modules. The digital twin performs the functions of coordinating and aggregating the ensemble results, identifying inconsistencies between partial recommendations, and generating a coherent investment decision. The system produces probabilistic investment recommendations, thereby enabling scenario analysis and quantitative risk assessment. The proposed approach can be used as a decision-support tool in both individual and institutional asset management, taking into account changes in the macroeconomic environment.

Decision-making in omnichannel logistics networks requires balancing multiple performance dimensions under dynamic and uncertain operating conditions. In practice, key performance indicators (KPIs) are frequently applied for ex-post evaluation, while their direct integration into decision formulation remains limited, leading to a structural separation between performance management and alternative selection. This study proposes a KPI-driven decision framework that embeds multidimensional performance indicators at the modeling stage. Alternatives are represented within a unified multi-criteria structure, where KPI dimensions are normalized and aggregated into a weighted composite index. The decision problem is formulated as the maximization of expected composite performance over discrete alternatives under scenario uncertainty. A simulation-based operationalization is developed to incorporate stochastic variation in key indicators while preserving a consistent comparison logic across scenarios. The resulting structure enables robust ranking and selection of alternatives based on expected multidimensional performance. The study provides an interpretable and computationally reproducible framework for multi-indicator decision support in omnichannel logistics networks.

The article examines the problem of formalizing investment cash flow in a distributed ledger environment. Within the framework of the digital transformation of financial relations, the cash flow of an investment project can be represented as a digital twin, recorded in the distributed ledger infrastructure and implemented through smart contracts. The aim of the study is to develop a mathematical model of the digital twin of investment cash flow and an algorithm for its forecasting using neural networks. Theoretical approaches to the interpretation of digital twins are systematized, and the limitations of the classical discounted cash flow model in relation to the digital environment are analyzed. A formalized model of digital cash flow is proposed, taking into account transaction fees of the distributed ledger, algorithmically accrued income, and an extended discount rate structure including technological and regulatory risk premiums. An algorithm for neural network forecasting of the digital twin is developed based on a feature vector integrating financial and infrastructure parameters. A comparative analysis of the digital and classical models is performed, which allowed establishing the structural modification of the investment process in the digital environment. The obtained results can be used in the valuation of digital financial assets and the construction of adaptive systems for forecasting their cash flows.