Scholarlink

Forecasting small sequences with missing values poses significant challenges in machine learning, particularly when dealing with limited training data and incomplete information. In this paper we present a novel unified methodological standpoint, which comparatively evaluates the effectiveness of various machine learning models in predicting sequential data under constrained conditions. We investigate four model families: ARIMA, Gradient Boosting, Fully Connected Neural Networks, and Transformers, examining their performance through both frequentist and Bayesian implementations. Utilizing five diverse datasets representing different sequencing patterns, we systematically analyze model behavior under varying data availability and missingness scenarios. By randomly reducing dataset sizes and introducing missing values, we explore how model complexity and Bayesian probabilistic approaches impact predictive accuracy. The experimental results demonstrate that Bayesianization consistently improves model performance across different datasets, with an average SMAPE reduction of 1–5%. Notably, neural network models, particularly Fully Connected Neural Networks, showed the most significant improvements through Bayesian techniques. This research provides insights into handling small, sparse sequential data and highlights the potential of Bayesian methods in enhancing predictive modeling under data-constrained conditions.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Bayesianization of ML Models in Forecasting Small Data Sequences with Missing Values

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Aleksandra Vatian,
  • Ivan Tomilov,
  • Keram Goguev,
  • Oksana Romakina,
  • Anna Arsenyeva,
  • Dmitry Dobrenko,
  • Natalia Gusarova

摘要

Forecasting small sequences with missing values poses significant challenges in machine learning, particularly when dealing with limited training data and incomplete information. In this paper we present a novel unified methodological standpoint, which comparatively evaluates the effectiveness of various machine learning models in predicting sequential data under constrained conditions. We investigate four model families: ARIMA, Gradient Boosting, Fully Connected Neural Networks, and Transformers, examining their performance through both frequentist and Bayesian implementations. Utilizing five diverse datasets representing different sequencing patterns, we systematically analyze model behavior under varying data availability and missingness scenarios. By randomly reducing dataset sizes and introducing missing values, we explore how model complexity and Bayesian probabilistic approaches impact predictive accuracy. The experimental results demonstrate that Bayesianization consistently improves model performance across different datasets, with an average SMAPE reduction of 1–5%. Notably, neural network models, particularly Fully Connected Neural Networks, showed the most significant improvements through Bayesian techniques. This research provides insights into handling small, sparse sequential data and highlights the potential of Bayesian methods in enhancing predictive modeling under data-constrained conditions.