Text from the interactive explanation
Tabular Data Adapters: Pseudo-Labeling Unlabeled Private Tabular Data for Outlier Detection
Dayananda Herurkar, Jörn Hees, Vesselin Tzvetkov, Andreas Dengel
Original paperIEEE Access · 2026 · Herurkar, Hees, Tzvetkov & Dengel
Fraud and faults hide as rare odd rows in private tables, and nobody has labelled which ones they are. Tabular Data Adapters borrow an answer key: translate each private row into the format of a similar public table that has already been solved, and let that table’s tuned detector draft a verdict.
① Your private table no answer key: 420 unlabelled rows
② The adapter translates each row private encoder → shared middle → public decoder
③ A public table’s tuned detector judges it “Plant measurements” and its tuned boundary
pseudo-labels drafted 0 / 420 · flagged as outliers 0
toy simulationAn illustration of the mechanism, not the paper’s data. Each dot is a row, placed by its table’s two strongest patterns.
Chapter 1 · The cold start
Which alarm do you trust?
Meet your private table, the Ledger: 420 card payments described by eight columns such as amount, hour of day and distance from home. A few rows are fraud. Nothing tells you which.
Chapter 2 · Dataset similarity
A fingerprint that needs no labels
Chapter 3 · Dataset transformation
The adapter: translating rows between tables
Chapter 4 · Pseudo-labels
Borrowed judges vote
Chapter 5 · Measured results
What the 50-dataset experiment found
Chapter 6 · Back to your table
Draft labels are a starting point
Topics in this iPaper
Computer science
we introduce Tabular Data Adapters (TDA), a method for generating pseudo-labels for unlabeled tabular data in outlier detection (OD) tasks
Abstract, page 25691
outlier detection
a method for generating pseudo-labels for unlabeled tabular data in outlier detection (OD) tasks
Abstract, page 25691
pseudo labeling
our approach enables the generation of weak labels
Abstract, page 25691
tabular data
Tabular data plays a vital role in numerous industries
Section I, page 25691
autoencoder
leverages an autoencoder-based neural network to transform samples from private datasets into public datasets
Section III-A, page 25694
principal component analysis
we apply PCA to the private dataset and calculate the mean reconstruction error
Section III-A, page 25694
transfer learning
a principled way to transfer knowledge from public data to private data
Section I, page 25692