Black-Box Time-Series Domain Adaptation via Cross-Prompt Foundation Models

TL;DR

BBTSDA remains challenging due to privacy constraints and noisy pseudo-labels from inaccessible source models. This work introduces Cross-Prompt Foundation Models (CPFM), a prompt-tuning framework built on a time-series foundation model (MOMENT) with dual prompts and reconstruction-based domain alignment to adapt unlabeled target data. The method combines prompt reconstruction and input reconstruction in a dual-branch setup, plus an entropy-based multi-source weighting scheme, achieving state-of-the-art MF1 scores on HAR, SSC, and MFD datasets with notable margins. The results demonstrate the effectiveness and privacy-preserving potential of a foundation-model–driven approach for time-series domain adaptation, with clear benefits from multi-source collaboration and ablation-supported components.

Abstract

The black-box domain adaptation (BBDA) topic is developed to address the privacy and security issues where only an application programming interface (API) of the source model is available for domain adaptations. Although the BBDA topic has attracted growing research attentions, existing works mostly target the vision applications and are not directly applicable to the time-series applications possessing unique spatio-temporal characteristics. In addition, none of existing approaches have explored the strength of foundation model for black box time-series domain adaptation (BBTSDA). This paper proposes a concept of Cross-Prompt Foundation Model (CPFM) for the BBTSDA problems. CPFM is constructed under a dual branch network structure where each branch is equipped with a unique prompt to capture different characteristics of data distributions. In the domain adaptation phase, the reconstruction learning phase in the prompt and input levels is developed. All of which are built upon a time-series foundation model to overcome the spatio-temporal dynamic. Our rigorous experiments substantiate the advantage of CPFM achieving improved results with noticeable margins from its competitors in three time-series datasets of different application domains.

Figures (6)

• Figure 1: Lower bound vs Upper bound performance in the HAR dataset
• Figure 2: (a) CPFM learning policy where only the prompts and classification heads are tuned leaving the backbone network frozen; (b) Prompt Reconstruction is devised to deliver distinct prompts in each network branch; (c) Input Reconstruction learns the structure of the target domain without any label information.
• Figure 3: T-SNE of SSC Dataset by network branch 1 before training
• Figure 4: T-SNE of SSC Dataset by network branch 1 after training
• Figure 5: T-SNE of SSC Dataset by network branch 2 before training