PaperlayerView exact revision on arXiv

Source-linked AI summary

Unsupervised Unlearning of Concept Drift with Autoencoders

André Artelt, Kleanthis Malialis, Christos Panayiotou, Marios Polycarpou, Barbara Hammer

arXiv:2211.12989v2cs.LGcs.AI

TL;DR

Concept drift can make downstream models obsolete, while updating them individually can be costly and difficult to scale. This paper uses unsupervised autoencoders to globally undo drift without modifying downstream models; experiments report significant downstream-model improvement in both datasets.

  • Problem

    Concept drift changes data distributions over time and can make downstream models obsolete; updating models individually can be costly and fail to scale when many models are involved.

    Source: §I · p. 1

  • Method

    The paper uses unsupervised autoencoders for model-agnostic global drift adaptation, transforming data without retraining or adapting downstream models.

    Source: §I · p. 1

  • Results

    Both datasets showed significant downstream-model improvement after applying the proposed function to drifted data, and Hanoi results also showed significantly improved variance.

    Source: C. Results · p. 5

  • Takeaways & Limitations

    The method showed strong performance across evaluated scenarios from different domains without requiring downstream-model retraining or adaptation.

    Source: §VII · p. 6

    The method cannot learn a drift-adaptation function if drift does not produce a sufficiently large increase in autoencoder reconstruction loss.

    Source: §VI · p. 5

Abstract

from arXiv · show

Concept drift refers to a change in the data distribution affecting the data stream of future samples. Consequently, learning models operating on the data stream might become obsolete, and need costly and difficult adjustments such as retraining or adaptation. Existing methods usually implement a local concept drift adaptation scheme, where either incremental learning of the models is used, or the models are completely retrained when a drift detection mechanism triggers an alarm. This paper proposes an alternative approach in which an unsupervised and model-agnostic concept drift adaptation method at the global level is introduced, based on autoencoders. Specifically, the proposed method aims to ``unlearn'' the concept drift without having to retrain or adapt any of the learning models operating on the data. An extensive experimental evaluation is conducted in two application domains. We consider a realistic water distribution network with more than 30 models in-place, from which we create 200 simulated data sets / scenarios. We further consider an image-related task to demonstrate the effectiveness of our method.

I. INTRODUCTION

Concept drift can undermine downstream models, while updating each model may be costly and difficult to scale. The paper proposes unsupervised, model-agnostic global adaptation that transforms data instead of changing the models.

I. INTRODUCTION

  • The paper frames drift as a change in data distribution that can make learning models obsolete, with severe practical consequences if unaddressed.The data distribution evolves over time as samples arrive in a stream.Source: §I · p. 1
  • Sensor faults can make per-model adaptation costly and hard to scale, whereas some drift types, such as water-demand changes, may still call for that traditional approach.The introduction distinguishes sensor faults from demand changes as examples of drift types.Source: §I · p. 1
  • The proposed global method uses autoencoders to undo drift without modifying downstream models or requiring ground-truth labels.It is model-agnostic and operates on the data rather than adapting each model directly.Source: §I · p. 1, §II · p. 2
  • The evaluations cover 200 simulated water-network scenarios with more than 30 downstream models, plus an image-related task.The scenarios represent different drift characteristics.Source: §I · pp. 1–2

II. CONCEPT DRIFT ADAPTATION PROBLEM

  • Local adaptation changes individual models, while global adaptation transforms observed data; the paper proposes the latter as complementary to local adaptation for different drift types.The formalized global approach uses a mapping f to transform samples so model performance increases.Source: §II · p. 2

III. RELATED WORK

Existing concept-drift methods are grouped into three broad categories.

  • Existing concept-drift methods are loosely categorized as concept drift adaptation, transfer learning, and representation learning,.Source: §III · p. 2

1) Concept Drift Adaptation:

Concept-drift adaptation methods are described as passive, active, or hybrid, distinguished by how they update models and respond to drift.

Passive methods

  • Passive methods incrementally update models, often using memory windows or ensembles; methods address memory sizing and class imbalance with adaptive windows, per-class memory, rebalancing,, and bagging,.Learn++.NSE is an example of an ensemble method.Source: 1) Concept Drift Adaptation · p. 2

Active methods

  • Active methods detect drift explicitly and trigger complete retraining, using approaches such as statistical tests,, JIT classifiers, or threshold-based DDM.Active methods discard the existing model and create a new one when a drift alarm is triggered; ensembling has also been used.Source: 1) Concept Drift Adaptation · pp. 2–3

Hybrid methods

  • Hybrid methods combine drift detection with other techniques, including adaptive rebalancing, incremental active learning, and autoencoder-based incremental learning.Source: 1) Concept Drift Adaptation · p. 3

2) Transfer Learning:

The paper reviews transfer and representation learning before presenting an autoencoder-based global adaptation method. The autoencoder models the data distribution, and a transfer function is built after drift detection to make drifted samples resemble pre-drift data.

2) Transfer Learning:

  • Some transfer-learning methods adapt to changed distributions, but many require labeled data; a few unsupervised methods exist –.Source: 2) Transfer Learning · p. 3

3) Representation Learning:

  • Learned representations can themselves be affected by drift and may need adaptation or robustness; little work on adapting representations is noted, including an invariant-representation method for continual learning [30].Source: 3) Representation Learning · p. 3

IV. UNSUPERVISED UNLEARNING OF CONCEPT DRIFT WITH AUTOENCODERS

  • The proposed method has two stages: an autoencoder characterizes the data distribution, then a function is built after drift detection to transform drifted samples toward their pre-drift appearance.The approach assumes an effective drift detector is in place and that the autoencoder can characterize drift; it is agnostic to the detection mechanism.Source: §IV · p. 3

A. Autoencoder for Distribution Learning

B. Unsupervised Transfer Learning using Autoencoders

The method learns an unsupervised mapping from unlabeled post-drift samples to reduce autoencoder reconstruction error, aiming to remove drift’s influence from the data space while leaving downstream models unchanged. It regularizes the mapping to limit feature changes and applies it to samples before the downstream tasks.

B. Unsupervised Transfer Learning using Autoencoders

  • The approach assumes distribution change increases autoencoder reconstruction error and does not require a particular drift-detection mechanism.Using the autoencoder to detect drift is possible, but the adaptation method is stated to be detector-agnostic.Source: B. Unsupervised Transfer Learning… · p. 3
  • The mapping f is learned from unlabeled post-drift samples to reduce autoencoder reconstruction error, used as a proxy for fit to the original distribution.The mapping takes samples from the new distribution as its training data.Source: B. Unsupervised Transfer Learning… · p. 3
  • Chaining f with the autoencoder would increase computational complexity and may be less suited to real-time or resource-limited settings.The paper identifies on-device learning as one such setting.Source: B. Unsupervised Transfer Learning… · p. 3
  • After drift detection, the procedure collects unlabeled post-drift data, learns f, and applies f before the downstream models.The algorithm builds the autoencoder before drift; without drift, it sends samples directly to the downstream tasks.Source: B. Unsupervised Transfer Learning… · p. 4
  • An l1 penalty weighted by C balances the objective while encouraging f to change as few features as possible.The penalty is applied to the difference between f(x_j) and x_j.Source: B. Unsupervised Transfer Learning… · p. 4
  • Global unlearning maps the data back toward a state in which existing models remain applicable, without modifying or considering those downstream models.The section frames this as removing the drift’s influence from the data space so all models working there can perform reasonably well.Source: B. Unsupervised Transfer Learning… · p. 4

V. EXPERIMENTS

The experiments evaluate the proposed global concept-drift adaptation methodology on diverse data sets. The paper also points readers to a publicly available Python implementation.

V. EXPERIMENTS

  • The paper evaluates its global concept-drift adaptation methodology on diverse data sets.The stated evaluation concerns the performance of the proposed methodology.Source: §V · p. 4
  • A Python implementation of the experiments is publicly available on GitHub.Source: §V · p. 4

A. Data

The experiments use two data sets; the supplied details describe a digits task in which drift is introduced by zeroing the upper half of test images. The task uses a logistic-regression classifier and 10-fold cross-validation.

A. Data

1) Digits:

  • The digits task uses the first five digits, 0–4, and a logistic-regression classifier as its single downstream model.Source: 1) Digits · p. 4
  • Digits-test drift is simulated by setting the upper-half pixels to zero, with results assessed using 10-fold cross-validation.The paper also compares with a classifier retrained using labeled training data as an estimate of supervised drift adaptation.Source: 1) Digits · p. 4, Figure 1 · p. 4

2) Hanoi:

The Hanoi experiments model pressure-prediction tasks across a sensor-equipped water network and introduce randomized sensor faults across 200 scenarios. The general procedure trains on pre-drift data, learns an adaptation mapping from post-drift samples, and evaluates downstream models after applying it.

2) Hanoi:

  • The Hanoi network has 32 pressure-sensor nodes and 31 downstream regression tasks that predict pressure at different locations.Source: 2) Hanoi · p. 4
  • The study generates 200 scenarios, each with a randomly selected faulty sensor and random fault parameters.Fault types include constant offset, Gaussian noise, power failure with zero readings, and an offset proportional to the true quantity.Source: 2) Hanoi · pp. 4–5

B. General Setup

  • The setup trains downstream models and an autoencoder on pre-drift data, then learns the adaptation function from a small post-drift window.It evaluates downstream models before and after drift and after applying the adaptation function; the autoencoder reconstruction is also used as a baseline.Source: B. General Setup · p. 5
  • For evaluation, the study uses ground-truth drift timing rather than building a drift detector, and implements the adaptation function as an affine mapping.The autoencoder is implemented as a multilayer perceptron.Source: B. General Setup · p. 5

C. Results

Applying the proposed adaptation significantly improves downstream-model performance in both data sets, with greater stability also reported for Hanoi. The discussion notes that adaptation can be fast with a simple architecture, but depends on drift increasing autoencoder reconstruction loss.

C. Results

  • For Hanoi, results are shown for each downstream task and across all scenarios; Fig. 3 illustrates task 4, while box plots exclude scenarios where autoencoder reconstruction could not be improved, leaving 138 scenarios.Source: C. Results · p. 5, Figure 3 · p. 5
  • Applying f(·) significantly improves downstream-model performance in both data sets; Hanoi results also show more stable solutions.The paper also reports significant improvement for the baseline that uses autoencoder reconstruction to undo drift.Source: C. Results · p. 5

VI. DISCUSSION

  • A relatively simple architecture for f(·) often suffices, requiring few samples and enabling fast training.The time to build f(·) also determines how soon adaptation can continue using the downstream models.Source: §VI · p. 5
  • The approach may not learn a drift-adaptation function unless the drift produces a significantly larger autoencoder reconstruction loss.The authors suggest applying the method only when drift manifests in such a loss increase.Source: §VI · p. 5

VII. CONCLUSION AND FUTURE WORK

The conclusion presents an unsupervised autoencoder-based approach for undoing concept drift without retraining or adapting downstream models, and reports strong performance across scenarios from different domains. Future work includes studying limited-label use, transparency, and adaptation to frequent drift.

Conclusion

  • The proposed unsupervised method aims to undo concept drift without retraining or adapting downstream models, and showed strong performance across scenarios from different domains.Source: §VII · pp. 5–6

Future work

  • The authors propose studying whether a small set of labeled samples could improve downstream-model performance when building f(·).Source: §VII · p. 6
  • Making f(·) transparent, for example by explaining its outputs, could provide insight into observed concept drift.Source: §VII · p. 6
  • For frequent drift, future work could study incremental learning or adaptations of f(·) to speed adaptation; this work considered a single drift.Source: §VII · pp. 6–7

APPENDIX

Table II reports detailed Hanoi data set results across scenarios.

APPENDIX

  • Table II gives detailed results on the Hanoi data set.The table reports median R2 scores over all scenarios, rounded to two decimal places.Source: APPENDIX · p. 8, Table II · p. 8
  • The reported statistic is the median R2 score over all scenarios.Source: APPENDIX · p. 8
  • Scores are rounded to two decimal places.Source: APPENDIX · p. 8
Loading 2211.12989v2…