[Paper Review] PROMISSING: Pruning Missing Values in Neural Networks

Seyed Mostafa Kia, Nastaran Mohammadian Rad+6 more|arXiv (Cornell University)|Jun 3, 2022

Machine Learning in Healthcare4 citations

TL;DR

PROMISSING proposes a novel method to handle missing values in neural networks by treating them as informative signals of uncertainty rather than errors, using a learnable representation to neutralize their effect on activations. The method achieves performance comparable to imputation techniques while enabling models to become increasingly indecisive with more missing data—mimicking human-like uncertainty awareness.

ABSTRACT

While data are the primary fuel for machine learning models, they often suffer from missing values, especially when collected in real-world scenarios. However, many off-the-shelf machine learning models, including artificial neural network models, are unable to handle these missing values directly. Therefore, extra data preprocessing and curation steps, such as data imputation, are inevitable before learning and prediction processes. In this study, we propose a simple and intuitive yet effective method for pruning missing values (PROMISSING) during learning and inference steps in neural networks. In this method, there is no need to remove or impute the missing values; instead, the missing values are treated as a new source of information (representing what we do not know). Our experiments on simulated data, several classification and regression benchmarks, and a multi-modal clinical dataset show that PROMISSING results in similar prediction performance compared to various imputation techniques. In addition, our experiments show models trained using PROMISSING techniques are becoming less decisive in their predictions when facing incomplete samples with many unknowns. This finding hopefully advances machine learning models from being pure predicting machines to more realistic thinkers that can also say "I do not know" when facing incomplete sources of information.

Motivation & Objective

• Address the challenge of missing data in real-world machine learning, especially in clinical and multi-modal settings where data collection is costly and incomplete.
• Overcome the limitations of traditional imputation methods, which assume missing data mechanisms and may introduce bias or require extensive preprocessing.
• Develop a method that allows neural networks to naturally handle missing values without imputation or removal, preserving model robustness and interpretability.
• Enable models to express uncertainty when faced with incomplete inputs, aligning more closely with human-like reasoning in uncertain environments.
• Facilitate counterfactual interpretation of model decisions by using missing values as a tool to probe feature importance at the individual patient level.

Proposed method

• Introduce a learnable representation for missing values that is applied during forward pass to neutralize their effect on neuron activations.
• Modify the forward pass of a neural network to treat missing values (NaN) as a distinct input type, using a learned embedding to represent the absence of information.
• Apply a pruning mechanism that suppresses the contribution of missing features to the activation of neurons, effectively 'pruning' their influence.
• Train the model end-to-end with standard backpropagation, allowing the network to learn when and how to downweight uncertain inputs.
• Extend the method to multi-modal data by applying the same mechanism independently to each modality, enabling flexible handling of mixed data types (continuous, binary, categorical).
• Enable counterfactual interpretation by artificially masking features as missing to assess their impact on predictions, providing local explainability at the individual level.

Experimental results

Research questions

• RQ1 Can a neural network be trained to handle missing values directly without imputation or removal, while maintaining predictive performance?
• RQ2 How does the model's confidence in predictions change as the number of missing values increases when using PROMISSING versus imputation-based methods?
• RQ3 Can PROMISSING enable more interpretable and clinically meaningful model decisions by allowing counterfactual analysis through artificial missingness?
• RQ4 Does PROMISSING improve model calibration by reflecting uncertainty in predictions when information is incomplete?
• RQ5 How does PROMISSING perform across diverse data types and in real-world clinical datasets with high-dimensional, multi-modal, and incomplete features?

Key findings

• PROMISSING achieves prediction performance comparable to multiple imputation techniques, including MICE and deep generative models, across classification and regression benchmarks.
• Models trained with PROMISSING become increasingly indecisive as the number of missing values increases, with predictions monotonically converging to 0.5 (random chance), reflecting calibrated uncertainty.
• The mPROMISSING variant (with learned embedding for missingness) corrects the slight divergence from chance performance observed in standard PROMISSING, improving reliability.
• PROMISSING enables effective counterfactual interpretation: by masking features as missing, the method reveals which modalities (e.g., PANSS, MINI) are most influential in model decisions.
• In clinical applications, PROMISSING supports explainable AI by answering 'Why' questions at the individual patient level, enhancing trust and utility in precision psychiatry.
• The method is flexible across data types (continuous, binary, categorical) and requires no data preprocessing, making it suitable for complex multi-modal datasets without additional engineering.