On the trasductive arguments in statistics

Ya’acov RitovarXiv 2010 · DOI · arXiv

Abstract: The paper argues that a part of the current statistical discussion is not based on the standard firm foundations of the field. Among the examples we consider are prediction into the future, semi-supervised classification, and causality inference based on observational data.

Introduction: Induction, Deduction, and the Rise of Transduction

In classical statistical theory, we typically operate within the realms of induction—using a sample to describe a population—or deduction—inferring parameters of a population based on a known model. However, modern machine learning and statistics have increasingly embraced transduction. Transduction involves making inferences about specific instances or structures that are not strictly supported by the observed data distribution or the underlying model.

Ya’acov Ritov’s 2010 paper offers a foundational critique of this trend. He identifies a growing gap between "statistical conclusions" (precise statements about error and confidence) and "experimenter’s conclusions" (practical guesses based on intuition). Ritov argues that many modern practices require a "leap of faith" regarding unverifiable assumptions that standard statistical theory cannot justify.

The Limits of Prediction: PP1 vs. PP2

Ritov distinguishes between two types of prediction problems in time series analysis. PP1 involves predicting a single future event based on the past. He argues this is often a "non-statistical task" because a single event cannot be meaningfully repeated in a "Gedankenexperiment" (thought experiment) to define valid confidence intervals.

In contrast, PP2 involves predicting a stream of values. This allows for self-correction and calibration. Ritov introduces "slowly walking" predictors on a grid that can be proven to be unbiased over time, suggesting that statistics is on firmer ground when dealing with sequences where errors can be monitored and corrected.

The Semi-Supervised Trap

Semi-supervised learning (SSL) uses unlabeled data to improve a classifier. The intuitive appeal is clear: unlabeled data reveals the structure of the feature space ($X$). However, Ritov’s Meta-Theorem 5.1 invokes semi-parametric bounds to show that knowing the distribution of $X$ does not inherently help estimate the relationship between features and labels ($Y|X$) without assuming a specific, unverifiable link.

Figure 1: Classification with labeled data (a), unlabeled data (b), and the Bayes line (c).

As shown in Figure 1, unlabeled data can make a classification boundary look obvious. We often assume that labels follow the clusters of the unlabeled data—the "manifold assumption." But if this assumption is wrong, the unlabeled data will mislead the model. Ritov warns against mistaking this mathematical "beauty" for truth. He proposes a "safe" transduction: use unlabeled data only to suggest potential predictors, but rely exclusively on labeled data to select the final model.

Causality as Metaphysics

Ritov views counterfactual causality—the "what if" scenarios—as a form of metaphysics when applied to purely observational data. Without manipulated experimental data, any causal conclusion is essentially "wishful thinking" disguised as an assumption. He argues that statisticians should be honest about when they are extrapolating outside the observed model.

Conclusion: A Call for Statistical Honesty

The paper concludes that while transduction is often unavoidable in practice, the field must distinguish between "statistical safe ground" and ungrounded belief. Future work should focus on robust, self-correcting procedures that do not fail catastrophically when their underlying transductive assumptions are invalid.

Summary updated

Abstract

The paper argues that a part of the current statistical discussion is not based on the standard firm foundations of the field. Among the examples we consider are prediction into the future, semi-supervised classification, and causality inference based on observational data.
BibTeX
@misc{ritov2010trasductive,
  author = {Ya’acov Ritov},
  title = {{On the trasductive arguments in statistics}},
  year = {2010},
  eprint = {1003.1513},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1003.1513},
}
Figure