Tabular Data Adapters: Improving Outlier Detection for Unlabeled Private Data

TL;DR

The paper tackles the challenge of applying high-performing public outlier detection models to private, unlabeled tabular data by introducing Tabular Data Adapters (TDA). It combines a Dataset Similarity Measure, based on PCA reconstruction errors, with a Dataset Transformation network that maps private data into the input space of public datasets, enabling generation of soft labels through applying public models and fusing their predictions. Two strategies, Top1-DS and TopN-DS, use one or multiple similar public datasets to produce robust soft labels, and are validated across 50 diverse tabular datasets, showing improved labeling accuracy and reduced compute time compared to baselines. The approach offers a practical, scalable bridge between public research and industrial applications, addressing cold-start labeling and domain shift in unlabeled private data, with potential for parallelization and improved fusion methods in future work.

Abstract

The remarkable success of Deep Learning approaches is often based and demonstrated on large public datasets. However, when applying such approaches to internal, private datasets, one frequently faces challenges arising from structural differences in the datasets, domain shift, and the lack of labels. In this work, we introduce Tabular Data Adapters (TDA), a novel method for generating soft labels for unlabeled tabular data in outlier detection tasks. By identifying statistically similar public datasets and transforming private data (based on a shared autoencoder) into a format compatible with state-of-the-art public models, our approach enables the generation of weak labels. It thereby can help to mitigate the cold start problem of labeling by basing on existing outlier detection models for public datasets. In experiments on 50 tabular datasets across different domains, we demonstrate that our method is able to provide more accurate annotations than baseline approaches while reducing computational time. Our approach offers a scalable, efficient, and cost-effective solution, to bridge the gap between public research models and real-world industrial applications.

Figures (6)

• Figure 1: Overview of our proposed soft labeling approach. The private dataset is compared by a "dataset similarity measure" with an index of existing public domain datasets (for each of them the best working ML/OD models are known), resulting in the top $n$ most similar datasets. Method1: Top1-DS performs a "Dataset Transformation" of the private dataset to the top $1$ most similar public dataset to align their formats. Predictions are generated using the known best $m$ optimized ML models for the top $1$ selected public dataset, and soft labels are created through majority voting. Method 2: TopN-DS transforms the private dataset into the formats of the top $n$ similar public datasets. The known best-performing model from each transformed dataset is used to generate predictions, which are fused via majority voting. Both methods aim to efficiently and accurately generate soft labels for unlabeled private datasets.
• Figure 4: Distribution of performance metrics (Balanced Accuracy, F1-score, PR-AUC, and ROC-AUC) across 50 datasets for both methods compared to baseline models. The white dot in each plot represents the mean value. The results demonstrate consistent superiority of the proposed methods in most metrics, with Method 1 and Method 2 showing comparable performance.
• Figure 5: Impact of best $m$ model and top $n$ transformation selection on computation time for both methods. While (single threaded) computation time for Method1 remains largely unaffected by increasing $m$, Method2 shows significant runtime increases, especially for datasets with over 10,000 samples, reflecting the trade-off between accuracy and efficiency.
• Figure 6: Soft label accuracy results for individual datasets in the healthcare and finance domains. The x-axis shows the held out unlabeled (private) datasets, and the y-axis shows performance scores. Both Method1 and Method2 consistently outperform baseline methods across individual datasets, demonstrating their robustness in varied domains.
• Figure 7: Breakdown of (single threaded) computational time for different components of the proposed methods. The dataset transformation step accounts for the majority of the total computation time, particularly in Method2, where multiple transformations are required. The figure highlights the potential for optimization through parallelization to reduce compute time.