Haphazard Inputs as Images in Online Learning

TL;DR

This work tackles online learning with haphazard inputs—data arriving with varying feature spaces and missing or obsolete features—by transforming each instance's observed features into a fixed-dimension, color-coded bar-plot image $I^t\in\mathbb{R}^{3\times224\times224}$. The method HI2 converts $X^t$ to $I^t$ on the fly and applies pre-trained vision models (e.g., ResNet-34, ViT) in an online gradient-descent loop, using streaming $Z$-score or Min-Max normalization to produce stable inputs. Across four public datasets, HI2 consistently outperforms classical and deep baselines on AUPRC, AUROC, and Balanced Accuracy, particularly as the fraction of unavailable features grows, demonstrating robustness and scalability. The approach enables a model-agnostic, image-based pathway for handling varying feature spaces in online learning, with practical implications for real-time systems and IoT-enabled domains, and a provided codebase for reproducibility.

Abstract

The field of varying feature space in online learning settings, also known as haphazard inputs, is very prominent nowadays due to its applicability in various fields. However, the current solutions to haphazard inputs are model-dependent and cannot benefit from the existing advanced deep-learning methods, which necessitate inputs of fixed dimensions. Therefore, we propose to transform the varying feature space in an online learning setting to a fixed-dimension image representation on the fly. This simple yet novel approach is model-agnostic, allowing any vision-based models to be applicable for haphazard inputs, as demonstrated using ResNet and ViT. The image representation handles the inconsistent input data seamlessly, making our proposed approach scalable and robust. We show the efficacy of our method on four publicly available datasets. The code is available at https://github.com/Rohit102497/HaphazardInputsAsImages.

Figures (2)

• Figure 1: The advantages of the proposed solution over existing models in the haphazard inputs field (or varying feature space in online learning).
• Figure 2: HI2 concept: Initially, we receive the streaming data as shown by a snapshot of magic04 data (see Table \ref{['tab:data_description']}) in the first box. The values in this figure are rounded to 2 decimal digits for ease of visualization. Next, for each new feature, a unique color is generated in the color-mapping storage. The colors corresponding to the observed features are selected for image creation. Subsequently, the raw data is normalized in a streaming manner. The normalized features are then converted into a bar graph of 224$\times$224 dimensions with corresponding features color. Finally, an image classifier is used to give a prediction.