[ИАД, осень 2025] Foundation models for spatial-time series. Занятие 5

Value of the Information & Strength of the Argument

The video provides valuable insights into two advanced topics: graph neural networks with continuous-time dynamics and Riemannian geometry for time series classification. The coding demonstration effectively illustrates the oversmoothing problem and shows a practical solution. The mathematical presentation is thorough, explaining key concepts such as geodesics, tangent spaces, and the geometric mean, with clear derivations. The argumentation is solid, building from basic intuitions to more complex ideas, and includes practical considerations like the choice of reference point for projection. However, the presentation is informal and lacks rigorous citations, which slightly weakens the scientific rigor.

Scientific Rigor, Source Quality, Title Accuracy

The scientific rigor is moderate. The presenter mentions specific works, such as those by Alexander Rashant and others, but does not provide formal citations or references. The coding demonstration is based on a specific paper, but the details are not fully disclosed. The title is somewhat misleading as it mentions ‘Foundation models’ but the content does not directly address them. The adéquation between title and content is weak, which slightly reduces the overall score. The discussion includes references to historical geodesy and the work of Sergey Dvoenko, but these are anecdotal rather than rigorous sources.

The title mentions 'Foundation models for spatial-time series', but the actual content focuses on graph neural networks and Riemannian geometry for time series classification, with no direct discussion of foundation models. The title is somewhat misleading.

7/10

The video is a technical lecture with a live coding demonstration and a mathematical presentation on Riemannian geometry for time series classification. The content is based on established research and includes references to specific works, but the presentation is informal and lacks rigorous citation of sources. The discussion is expert-level and the mathematical derivations are plausible, but the lack of formal references and the informal setting reduce the overall reliability score.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 00:00Introduction and start of the lecture; Nikita begins his coding demonstration.
• 02:00Nikita explains the oversmoothing problem in graph neural networks and shows code for a standard GNN.
• 05:00Demonstration of the continuous-time diffusion-based GNN and comparison of embeddings.
• 10:00Discussion on the practical use of the framework and potential as a tool for large language models.
• 15:00Daniil starts his presentation on Riemannian geometry for time series classification.
• 20:00Introduction to SPD matrices as a Riemannian manifold, geodesics, and Riemannian distance.
• 25:00Explanation of tangent space, logarithmic and exponential maps, and their use for projection.
• 30:00Discussion on the geometric mean and its iterative computation for SPD matrices.
• 35:00Application of Riemannian geometry to time series classification via covariance matrices.
• 40:00Q&A and discussion on the choice of reference point and regularization of matrices.

• Graph Neural Networks with Continuous-Time Dynamics — Mentioned as the framework for the coding demonstration, but no specific URL provided.
• Riemannian Geometry for SPD Matrices — Referenced as the mathematical foundation for the presentation, but no specific source given.
• Work by Alexander Rashant — Mentioned as an author of approaches for time series classification using Riemannian geometry.

• Riemannian Geometry — Provides background on the mathematical concepts discussed.
• SPD Matrices — Explains the properties of positive definite matrices.