[Paper Review] Meta-Learning for Relative Density-Ratio Estimation

Atsutoshi Kumagai, Tomoharu Iwata+1 more|arXiv (Cornell University)|Jul 2, 2021

Anomaly Detection Techniques and Applications47 references4 citations

TL;DR

This paper proposes a meta-learning framework for relative density-ratio estimation (DRE) that enables accurate estimation from few-shot instances by leveraging knowledge from related source datasets. It uses neural network-based instance embeddings and a differentiable closed-form linear model to achieve fast, stable adaptation, outperforming existing methods in relative DRE, dataset comparison, and outlier detection tasks.

ABSTRACT

The ratio of two probability densities, called a density-ratio, is a vital quantity in machine learning. In particular, a relative density-ratio, which is a bounded extension of the density-ratio, has received much attention due to its stability and has been used in various applications such as outlier detection and dataset comparison. Existing methods for (relative) density-ratio estimation (DRE) require many instances from both densities. However, sufficient instances are often unavailable in practice. In this paper, we propose a meta-learning method for relative DRE, which estimates the relative density-ratio from a few instances by using knowledge in related datasets. Specifically, given two datasets that consist of a few instances, our model extracts the datasets' information by using neural networks and uses it to obtain instance embeddings appropriate for the relative DRE. We model the relative density-ratio by a linear model on the embedded space, whose global optimum solution can be obtained as a closed-form solution. The closed-form solution enables fast and effective adaptation to a few instances, and its differentiability enables us to train our model such that the expected test error for relative DRE can be explicitly minimized after adapting to a few instances. We empirically demonstrate the effectiveness of the proposed method by using three problems: relative DRE, dataset comparison, and outlier detection.

Motivation & Objective

• To address the challenge of relative density-ratio estimation when only a few instances are available from both distributions.
• To enable effective knowledge transfer from related source datasets to unseen target datasets with minimal support instances.
• To develop a differentiable, closed-form solution for relative DRE that allows end-to-end training via gradient descent.
• To explicitly minimize expected test error after adaptation using meta-learning.
• To demonstrate robust performance across diverse applications including outlier detection and dataset comparison.

Proposed method

• Uses permutation-invariant neural networks to encode each dataset into a latent vector representation from its support instances.
• Maps individual instances into an embedding space using the latent vectors of their respective datasets, enabling instance-specific adaptation.
• Models the relative density-ratio as a linear function in the embedded space, solvable via a closed-form solution for stability and speed.
• Trains the entire model end-to-end by minimizing the expected test error of the relative DRE after adaptation, using gradient-based optimization.
• Employs a meta-learning paradigm where shared parameters across all datasets allow knowledge transfer from source to target tasks.
• Utilizes a differentiable closed-form solution to enable backpropagation through the adaptation process, enabling explicit optimization of generalization performance.

Experimental results

Research questions

• RQ1 Can meta-learning improve relative density-ratio estimation when only a few instances are available from both distributions?
• RQ2 How effectively can knowledge from related source datasets be transferred to enable few-shot relative DRE on unseen target datasets?
• RQ3 Does a differentiable closed-form solution for the linear model in the embedded space lead to better generalization and faster adaptation than numerical optimization?
• RQ4 How does the proposed method compare to existing DRE and outlier detection methods in terms of robustness and performance across diverse data distributions?
• RQ5 What is the impact of architectural components such as the latent embedding function and the adaptation mechanism on overall performance?

Key findings

• The proposed method achieved the highest average AUC of 75.27% across all benchmark datasets in outlier detection, outperforming all 13 comparison methods.
• On the IoT dataset, the method achieved 96.94% AUC with few normal support instances, demonstrating strong few-shot generalization.
• The ablation study showed that removing the latent embedding function (NoLatent) reduced performance to 74.70% average AUC, confirming its importance.
• The method outperformed state-of-the-art PU learning methods and showed robustness across different values of the relative parameter α.
• The closed-form solution enabled faster and more stable adaptation compared to numerical optimization, with improved generalization due to differentiable training.
• The model achieved state-of-the-art performance in both relative DRE and dataset comparison tasks, validating its effectiveness across multiple applications.