Data Recognition and Classification (KPI lecture course, 2011–2013)

Yaroslav Yu. Dorohyi and N. M. Pishchaieva prepared this lecture course for the Department of Automation and Control in Technical Systems at the National Technical University of Ukraine “Kyiv Polytechnic Institute.” The cover identifies a 2010 course packet; the surviving PDFs were exported or revised from September 2011 through November 2013. The archive combines lectures 1–3, 5–6, and 8–16 in numerical order. Lectures 4 and 7 are absent from the supplied folder, so the combined file contains 14 lecture documents and 309 pages, not 15 documents.

The course is admissible as instructor-authored lecture notes. It is broad enough to connect classical statistics and pattern recognition, but its age and sparse citations make it a secondary source: GTK uses it for mechanism, derivations, terminology, and historical structure, while current primary papers, standards, and official documentation determine present practice.

Complete folder audit #

The supplied RKU/ folder contains three source classes:

The one admissible authored source contains 309 pages and was extracted and visually inspected in full. The excluded add-on and Orange manuals are intentional exclusions, not coverage gaps.

Lecture-by-lecture concept map #

Principal-components and factor analysis are only incidental in the admissible lecture set. The dedicated treatment comes from current mathematical sources in principal components and factor analysis. The ML dimensionality-reduction chapter takes PCA as a prerequisite and does not repeat its derivation.

Mechanisms worth preserving #

Representation fixes the problem the algorithm sees #

Lectures 1, 2, and 6 make one continuous argument even though they appear in separate files. A physical or social object becomes a record only after an observation process chooses attributes and measurement scales. Coding and transforming those attributes then determines what differences are representable. Finally, a distance or similarity function converts those differences into geometry. A nearest-neighbour or clustering method therefore does not discover an intrinsic geometry: it operates on the one produced by measurement, coding, scaling, and missing-data decisions.

That causal chain is more useful than a catalogue of distance formulas. For example, treating an unordered three-level category as the numbers 0, 1, and 2 silently asserts that level 1 lies midway between the others. Euclidean distance then enforces that false relation. The GTK chapters preserve the course’s catalogue but make the modelling assertion explicit and add current handling for learned embeddings, leakage, grouped data, and deployment drift.

Experimental design creates identifiable contrasts #

The design and ANOVA lectures connect operational choices to algebra. Randomization breaks systematic alignment between treatment and nuisance variables; replication estimates the variation remaining under nominally identical conditions; blocking removes known nuisance structure before it inflates the error term. Factorial designs vary several factors together so an interaction can be estimated rather than absorbed into residual noise. ANOVA then partitions total variation into components induced by that design.

The important lesson is causal, not ceremonial: an $F$ statistic cannot rescue a design that never produced the comparison of interest. Conversely, a good design makes the estimand visible in a contrast and gives the residual term a defensible interpretation. GTK adds modern estimand language, randomization-based reasoning, multiplicity, effect sizes, diagnostics, and hierarchical alternatives.

Hierarchical clustering is a sequence of irreversible merges #

Lecture 5’s worked dendrograms show why linkage is part of the model. After the closest pair is merged, the algorithm must define the distance from that new cluster to every remaining cluster. Single linkage measures the nearest pair and can chain through bridges; complete linkage measures the farthest pair and favours compact groups; average linkage compromises between them; Ward’s criterion chooses the merge with the smallest increase in within-cluster dispersion. Because each merge changes the candidate distances and is never undone, an early bridge or outlier can restructure the whole dendrogram.

Discrimination separates estimation from decision #

Lecture 8 constructs discriminant directions from within- and between-class scatter, then turns scores into a class decision. Lecture 15 adds priors and losses. Together they expose a distinction that is frequently blurred: estimating a posterior or discriminant score is not the same as choosing an action. The decision threshold depends on class priors, error costs, and any abstention option. Modern calibration and classification-metric chapters keep those layers separate.

Measurement uncertainty propagates through sensitivity #

Lecture 2 treats an indirect measurement as a function $y=f(x_1,\ldots,x_p)$. Small input perturbations change the output approximately through the gradient, so covariance is transported by the local linear map. This is the durable mechanism behind first-order propagation. GTK replaces the course’s older “error” vocabulary with the current distinction between correction, standard uncertainty, covariance, coverage, and traceability from the BIPM vocabulary and GUM framework, and it states when linear propagation fails.

Current-use boundaries #

• The course predates today’s distinction between prediction performance, calibrated probabilities, causal effects, and operational utility. GTK defines the estimand and decision before selecting an algorithm.
• Several hypothesis-test chapters present “accept the null” language. A non-rejection is not evidence that a model is true; the modern entries discuss power, effect size, model criticism, parameter estimation, and multiplicity.
• Classical stepwise regression and in-sample adequacy checks are retained as historical mechanisms, not recommended as reliable feature-selection or generalization procedures.
• Neural-network coverage stops at perceptron-style updates. Backpropagation, initialization, normalization, adaptive optimization, convolutional and recurrent networks, transformers, and present evaluation practice require current primary sources.
• Biometric material illustrates a recognition pipeline but predates current threat models, demographic evaluation, privacy practice, and standards. Its numerical claims and deployment recommendations are not reused.
• The PDFs contain author-applied colour highlights and edits. Highlighting is presentation metadata, not a confidence indicator or independent evidence.

Visual audit #

All 309 pages of the combined PDF were rendered after merging and inspected at contact-sheet scale; load-bearing pages were also inspected at full resolution during the source audit.

1. Data and measurement visuals: attribute-type trees, scatterplots, histograms, transformation curves, and direct/indirect measurement diagrams explain how observations become variables. GTK preserves their mechanisms through new examples, equations, and tables rather than copying low-resolution figures.
2. Clustering and discrimination geometry: linkage diagrams, distance matrices, dendrograms, projected group clouds, decision boundaries, and scatter matrices are information-bearing. Clean theme-aware diagrams in the clustering and discriminant chapters preserve the topology and remain readable at article width.
3. Experimental-design and ANOVA tables: factor matrices, response tables, sums-of-squares decompositions, and interaction layouts support the worked derivations. GTK rebuilds the necessary examples in accessible tables and equations so every label and unit is explicit.
4. Recognition mechanisms: perceptron blocks, decision regions, voting structures, contour codes, morphological operations, and the biometric pipeline illustrate the flow from feature extraction to decision. Only the general mechanisms survive; the original coloured diagrams are not required for the modern account.
5. Missing source figures: merged pages 173–179 preserve captions and empty frames for several regression figures that are already absent in RKU_L009_UKR.pdf. They are unusable as evidence and were not reconstructed from inference or from excluded laboratory material.

No original visual was copied into a topic entry. Redraws were chosen only where they improve causal explanation and legibility; current primary sources support the redrawn mechanism.

Excluded Orange manuals #

The following files are instructional laboratory manuals. They were inventoried by filename only and deliberately excluded:

Their omission from the library, examples, and visuals is intentional. No concept is considered uncovered merely because it appears only in these manuals.