Multi-Source Transfer Learning for Non-Stationary Environments

Honghui Du, Leandro L. Minku, Huiyu Zhou

Abstract

In data stream mining, predictive models typically suffer drops in predictive performance due to concept drift. As enough data representing the new concept must be collected for the new concept to be well learnt, the predictive performance of existing models usually takes some time to recover from concept drift. To speed up recovery from concept drift and improve predictive performance in data stream mining, this work proposes a novel approach called Multi-sourcE onLine TrAnsfer learning for Non-statIonary Environments (Melanie). Melanie is the first approach able to transfer knowledge between multiple data streaming sources in non-stationary environments. It creates several sub-classifiers to learn different aspects from different source and target concepts over time. The sub-classifiers that match the current target concept well are identified, and used to compose an ensemble for predicting examples from the target concept. We evaluate Melanie on several synthetic data streams containing different types of concept drift and on real world data streams. The results indicate that Melanie can deal with a variety drifts and improve predictive performance over existing data stream learning algorithms by making use of multiple sources.

Introduction

Predictive modeling in data streams, such as fraud detection or sensor monitoring, is complicated by concept drift—changes in the underlying data distribution over time. When a drift occurs, traditional models often experience a significant drop in accuracy and require a substantial amount of new data to adapt. While transfer learning is a staple in offline machine learning for leveraging knowledge from related tasks, its application to non-stationary data streams has been limited. Most existing methods rely solely on the target stream's history.

This paper introduces Melanie (Multi-sourcE onLine TrAnsfer learning for Non-statIonary Environments), the first framework designed to transfer knowledge from multiple streaming sources to a target stream in the presence of drift.

The Melanie Framework

Melanie is an ensemble-based approach that dynamically selects and weights knowledge from various sources. Its architecture consists of several key components:

  • Multi-Source Pool: Melanie maintains a collection of online ensembles (e.g., Online Bagging or Boosting) for every available source and the target stream.
  • Drift Detection: Utilizing the Drift Detection Method (DDM), Melanie monitors each stream. Upon detecting a drift, it initializes a new ensemble while preserving historical models in the pool, allowing for the reuse of "old" knowledge if concepts recur.
  • Dynamic Sub-Classifier Selection: Rather than transferring monolithic models, Melanie evaluates individual base learners (sub-classifiers) from all pools. Each sub-classifier is assigned a weight based on its recent performance on the target stream.
  • Time-Forgetting Factor ($\theta$): To ensure agility in non-stationary environments, a forgetting factor prioritizes recent performance, enabling the ensemble to adapt rapidly as the target concept evolves.
  • Negative Transfer Mitigation: To prevent performance degradation from irrelevant sources, Melanie sets the weight of any sub-classifier to zero if its accuracy falls to or below random guessing levels.
Initial accuracy comparison on a target stream, showing Melanie's advantage with multiple sources when data is scarce.

Experimental Evaluation

The effectiveness of Melanie was tested against state-of-the-art baselines, including Adaptive Random Forest (ARF) and Dynamic Weighted Majority (DWM), using both synthetic and real-world datasets:

  • Synthetic Streams: Gaussian distributions simulating no drift, abrupt drift, and incremental drift.
  • ELEC2 (Electricity): A real-world dataset of electricity pricing in New South Wales, characterized by frequent concept drifts.
  • Weather: A numeric dataset used for rain prediction.

Key Results and Insights

The experimental results highlight several critical advantages of the Melanie framework:

  • Immediate Performance Gains: Transfer learning is most impactful when target data is scarce—specifically at the start of a stream or immediately following a drift. As shown in Figure 2, Melanie with multiple sources achieves higher initial accuracy than non-transfer methods.
  • Rapid Recovery from Drift: In scenarios with abrupt concept drift, Melanie significantly outperforms standard drift-handling algorithms. By leveraging sub-classifiers that may already match the new concept, it bypasses the lengthy "relearning" phase.
  • Internal History as a Source: Even without external sources, Melanie's ability to treat the target's own pre-drift models as sources allows it to outperform methods like ARF by better managing historical knowledge.
Performance recovery after an abrupt concept drift at 10,000 examples.

In a comprehensive Friedman Rank test, Melanie (Online Boosting) with sources achieved the top average rank (3.42), demonstrating its robustness across diverse drift types and data volumes.

Conclusion and Future Work

Melanie demonstrates that multi-source transfer learning is a powerful tool for enhancing the resilience of online models against concept drift. While the current framework focuses on binary classification, future research will explore extensions to multi-class problems, regression, and explicit handling of class imbalance. Additionally, further optimization of hyperparameter sensitivity and computational overhead will broaden Melanie's applicability to resource-constrained streaming environments.

Summary updated

Abstract

In data stream mining, predictive models typically suffer drops in predictive performance due to concept drift. As enough data representing the new concept must be collected for the new concept to be well learnt, the predictive performance of existing models usually takes some time to recover from concept drift. To speed up recovery from concept drift and improve predictive performance in data stream mining, this work proposes a novel approach called Multi-sourcE onLine TrAnsfer learning for Non-statIonary Environments (Melanie). Melanie is the first approach able to transfer knowledge between multiple data streaming sources in non-stationary environments. It creates several sub-classifiers to learn different aspects from different source and target concepts over time. The sub-classifiers that match the current target concept well are identified, and used to compose an ensemble for predicting examples from the target concept. We evaluate Melanie on several synthetic data streams containing different types of concept drift and on real world data streams. The results indicate that Melanie can deal with a variety drifts and improve predictive performance over existing data stream learning algorithms by making use of multiple sources.
BibTeX
@inproceedings{du2019multi,
  author = {Honghui Du and Leandro L. Minku and Huiyu Zhou},
  title = {{Multi-Source Transfer Learning for Non-Stationary Environments}},
  booktitle = {IJCNN},
  year = {2019},
  url = {https://openreview.net/forum?id=WBF9r6lP1A},
}
Figure