Online Transfer Learning for Concept Drifting Data Streams

Helen McKay, Nathan Griffiths, Phillip Taylor, Theo Damoulas, Zhou XuBigMine @ KDD 2019 · OpenReview

TL;DR

This paper introduces BOTL, a novel framework for online transfer learning that accounts for concept drift in both source and target data streams. By enabling bi-directional knowledge transfer and incorporating model culling, BOTL significantly improves prediction accuracy in dynamic environments compared to existing methods.

What Problem Does This Paper Address?

In many real-world applications, data arrives as a continuous stream, and the underlying patterns (concepts) can change over time – a phenomenon known as concept drift. Traditional machine learning models struggle with this, requiring frequent retraining. Transfer Learning (TL) offers a solution by leveraging knowledge from a "source" domain to improve learning in a "target" domain, especially when the target domain has limited data. However, existing online TL methods often assume the source domain is static or "offline," failing to account for concept drift in the source itself. This limitation is critical because if the source data's concept drifts, the transferred knowledge can become outdated or even detrimental. This paper addresses the challenge of performing effective transfer learning in scenarios where both the source and target data streams are dynamic and susceptible to concept drift.

Methods

The core of the proposed solution is the Bi-directional Online Transfer Learning (BOTL) framework. Key aspects include:

  • Bi-directional Transfer: Unlike previous methods, BOTL treats all domains (source and target) as online and capable of learning and transferring knowledge to each other. This allows for mutual improvement.
  • Concept Drift Detection: BOTL utilizes an adaptation of the RePro algorithm for regression to detect concept drift within each domain. RePro is chosen for its ability to maintain a history of concepts and avoid relearning recurring ones, and its adaptation uses R² performance instead of classification accuracy.
  • Model Culling (BOTL-C.I and BOTL-C.II): To manage the potentially large number of transferred models in multi-domain scenarios and prevent overfitting, two model culling variants are introduced:
  • BOTL-C.I: Temporarily removes models whose performance drops below a threshold (Acperf).
  • BOTL-C.II: Employs a more sophisticated approach, culling models based on both performance (Acperf) and diversity (using mutual information, ACMI). This aims to retain useful knowledge while discarding redundant or conflicting models.
  • Meta-Learner: An Ordinary Least Squares (OLS) regressor acts as a meta-learner to combine the current target model and the transferred models into an overarching predictive function.

Datasets

The paper evaluates BOTL on three types of datasets:

  • Drifting Hyperplane: A synthetic dataset generator modified from existing work to include sudden, gradual, and recurring drifts.
  • Heating Simulation: A simulated dataset based on UK weather data and user schedules, designed to mimic complex, noisy real-world scenarios.
  • Following Distance: A real-world dataset derived from vehicle telemetry, used to predict Time To Collision (TTC).

Key Insights and Ideas

  • The Importance of Bi-directional Transfer: The most significant insight is that assuming only the target is online is insufficient. Allowing source domains to also adapt and transfer knowledge back leads to substantial performance gains.
  • Model Culling is Crucial for Scalability: In scenarios with many domains, the sheer volume of transferred models can overwhelm the meta-learner. Culling strategies are essential to prevent overfitting and maintain performance.
  • Diversity Matters in Culling: BOTL-C.II's success, particularly in scenarios with many domains, highlights that simply removing poorly performing models isn't enough. Considering model diversity alongside performance is key to retaining valuable, complementary knowledge.
  • Concept Drift Detection is Independent: The paper emphasizes that concept drift detection should occur locally within each domain, independent of knowledge transfer, to ensure accurate drift identification.

Key Results

  • Superior Performance: BOTL consistently outperformed baseline methods like RePro (no transfer) and GOTL (offline source) across all tested datasets, achieving statistically significant improvements in R² performance.
  • Drifting Hyperplane: BOTL achieved higher R² scores and lower RMSE compared to RePro and GOTL. BOTL-C variants also performed well, though slightly below BOTL, indicating the trade-off between aggressive culling and performance.
  • Heating Simulation: BOTL showed a significant performance boost over RePro, demonstrating its effectiveness in complex, noisy environments.
  • Following Distance: BOTL and its variants outperformed RePro and GOTL. Notably, BOTL-C.II showed the best performance, especially as the number of domains increased, suggesting its scalability.
  • Scalability: Figure 3 illustrates that as the number of domains increases, BOTL's performance degrades due to overfitting, but BOTL-C.II maintains performance by effectively managing the model set.

Limitations and Future Work

  • Parameter Tuning: The underlying RePro algorithm and the culling strategies (Acperf, ACMI) still require domain expertise for parameter tuning.
  • Homogeneous Domains: The current BOTL framework assumes homogeneous domains (same feature space).
  • Future Work: The authors suggest investigating task relatedness to identify similar concepts without parameterised thresholds, reducing the dependency on domain expertise. They also propose incorporating domain adaptation to enable knowledge transfer across heterogeneous domains.

Summary updated

BibTeX
@inproceedings{mckay2019online,
  author = {Helen McKay and Nathan Griffiths and Phillip Taylor and Theo Damoulas and Zhou Xu},
  title = {{Online Transfer Learning for Concept Drifting Data Streams}},
  booktitle = {BigMine @ KDD},
  year = {2019},
  url = {https://openreview.net/forum?id=jKpkYs0tKp},
}
Figure