Keynote Speakers

Vladimir Stanovov, Generating heuristics for optimization methods via automated algorithm design

Lev Kazakovtsev, Multimodal Fusion in Knowledge Graphs and Fast Data Retrieval

Generating heuristics for optimization methods via automated algorithm design

Creating new algorithms for solving various types of problems is a challenging task for both humans and artificial intelligence systems. In this talk the methods and algorithms for automatic generation of heuristics will be discussed. Two main groups of methods are considered: genetic programming (GP) and large language models (LLMs). Several case studies are considered: designing local search heuristics for dynamic optimization problems, creating new selection mechanisms for genetic algorithm, and creating full search strategies from scratch. The differences between the mentioned methods are discussed, as well as their applicability, and some general recommendations are given. The possible ways of applying these methods for the generation of new heuristics for ILP/MILP problems are discussed as well.

Multimodal Fusion in Knowledge Graphs and Fast Data Retrieval

Knowledge graphs and modern information systems accumulate heterogeneous data: text documents, structured relations, and high-dimensional vector embeddings of objects. Such multimodal repositories are a valuable source of knowledge for intelligent data analysis, and fast retrieval of relevant information from them is one of the central problems in machine learning. Hybrid search combines complementary retrieval signals — sparse lexical matching (BM25, SPLADE) and dense semantic vector search — which differ in scale and semantics and therefore cannot be fused directly. To accelerate retrieval, supplementary index structures, such as the inverted file index (IVF), are usually built, and the results of heterogeneous retrievers are merged by rank-based fusion. In this talk, the author summarizes recent results of the Krasnoyarsk research team in this field: query-dependent fusion of sparse and dense scores based on intrinsic query characteristics (kinematic search), adaptive cascaded search that dynamically allocates candidate count and re-ranking effort per query with early stopping, adaptive IVF search that estimates query complexity on the first search steps, and sparse-vector pruning. The proposed methods are index-agnostic and DBMS-compatible, and demonstrate 20%+ overall speedup at a fixed recall level on standard benchmarks.