Please note: This PhD seminar will take place in DC 1302.
Soheil Johari, PhD candidate
David R. Cheriton School of Computer Science
Supervisor: Professor Raouf Boutaba
Machine Learning (ML) models are increasingly employed for critical network management tasks such as traffic prediction, anomaly detection, root cause analysis, and resource allocation. A major issue in the reliability of these models is data drift, which refers to the discrepancy between training data (source domain) and test/operational data (target domain) caused by changes in network conditions and configurations. Domain adaptation, which aims to develop robust models that can generalize well across different but related domains, is a promising solution to the data drift issue. However, existing domain adaptation methods often fall short in few-shot scenarios, where target domain data is limited due to high data collection costs or restricted operational network access. Furthermore, the existing methods require the network management ML models to be frequently retrained or fine-tuned over time to adapt to the changes in data distributions, leading to high operational costs.
To address these limitations, we propose a novel, model-agnostic domain adaptation approach specifically designed for few-shot scenarios and network data. In our approach, network management ML models are trained exclusively on source domain data with all the input features included, while a two-step method aligns test data samples from the target domain with the source domain during inference. The first step employs our proposed causal-inference-based feature separation (FS) method, which introduces a novel perspective by treating domain shift as soft interventions (interventions that adjust the probability distribution of features rather than making absolute changes) on a set of specific features. FS effectively separates domain-variant and domain-invariant features directly in the input space, even with limited target training data. In the second step of our approach, we propose a Generative Adversarial Network (GAN)-based reconstruction method, trained exclusively on source data, to reconstruct the domain-variant features given the domain-invariant features. During inference, the GAN model maps the domain-variant features of the target domain samples to the source domain distribution, allowing the use of domain-variant features without causing cross-domain performance degradation. Since our approach trains the network management ML models exclusively on source domain data, it eliminates the need for retraining or fine-tuning these models as network conditions or data distributions evolve over time, significantly reducing costs and operational overhead. Comprehensive evaluations on two public 5G network datasets demonstrate an average 52% improvement in mitigating data drift compared to state-of-the-art methods in terms of F1-score.