Small data appears in finance, medicine, climatology, and biology: few observations, thousands of features. In such settings, neural networks overfit, and standard methods fail to separate signal from noise. Related papers on the arXiv already offer maps with honest uncertainty, self-taught algorithms, and neural networks for gravitational maps, but the problem of jointly reducing dimensionality and classifying remains open. This work fits into the line of methods that seek robust structures in small data and offers a closed-form solution with guaranteed convergence.
The authors construct a functional that simultaneously minimizes the reconstruction error after rotating into a lower-dimensional space and the Kullback-Leibler divergence between predicted and true labels. The rotation is given by a matrix with orthonormal columns, and the partition into cells is described by a membership matrix. For each iteration step, an analytical solution is derived: the rotation is found via singular value decomposition, point-to-cell assignment is based on the nearest center with a label correction, and conditional label probabilities are recomputed from frequencies. With discrete partitioning (each point belongs to one cell), all subproblems are solved in closed form, and the total cost per iteration is linear in the number of features and observations.
On synthetic data about worm lifespan, GOAL achieved an AUC of about 0.95 in the large-data regime and about 0.7 in the small-data regime, second only to the eSPA+ algorithm, which uses additional entropy regularization. On forecasting the warm phase of El Niño from 295,200 features and 720 observations, GOAL achieved an AUC of about 0.9 for a 24-month horizon, beating all competitors with significantly lower computation time. On the task of reconstructing gene activity networks from 182 genes, GOAL achieved a mean AUC of 0.910 and an accuracy of 0.828, being the best in 82.7% of 910 scenarios, while random forest and LSTM lagged by 0.1 and 0.17 in AUC, respectively. GOAL's runtime was 0.019 seconds versus 9.84 for random forest and 39.88 for LSTM.
This work differs from its arXiv neighbors by providing a unified solution for rotation, partitioning, and classification with proven convergence and linear cost. Other approaches either handle only dimensionality reduction without considering labels or require expensive hyperparameter tuning. GOAL fills the niche of a fast and transparent tool for small data, where interpretability and reproducibility matter. For science, this means that problems with hundreds of thousands of features and hundreds of observations become solvable on an ordinary laptop.
The topic could evolve toward fuzzy partitioning, where a point belongs to several cells with different probabilities. This might improve accuracy but would require solving convex programming problems instead of using analytical formulas. Another direction is accelerating the rotation step, which is currently the most expensive due to singular value decomposition. Approximate methods for very large matrices or distributed versions of the algorithm may appear.
The algorithm is applicable in climatology for long-term forecasts, in bioinformatics for reconstructing gene networks from limited experimental data, and also in medical imaging and financial analysis, where data are scarce and features are numerous.
The next step is to test GOAL on problems with severe class imbalance and concept drift, where standard methods often fail. It is also worth comparing fuzzy and discrete partitioning on the same datasets to understand when paying computational complexity for accuracy is worthwhile.
In machine learning, the problem of choosing metrics for evaluating models on small data remains unsolved, since standard measures can give biased results due to overfitting. This work does not solve that problem but takes a step toward more robust methods that depend less on data volume. The question of automatic hyperparameter selection also remains open, especially the search range, which is hard to determine in advance for small data.
For the El Niño forecast, 295,200 features were used with only 720 observations, which is like trying to guess a melody from 720 notes when the score has 295,200 lines.