Abstract:We consider online optimization procedures in the context of logistic regression, focusing on the Extended Kalman Filter (EKF). We introduce a second-order algorithm close to the EKF, named Semi-Online Step (SOS), for which we prove a O(log(n)) regret in the adversarial setting, paving the way to similar results for the EKF. This regret bound on SOS is the first for such parameter-free algorithm in the adversarial logistic regression. We prove for the EKF in constant dynamics a O(log(n)) regret in expectation and in the well-specified logistic regression model.
What problem does this paper attempt to address?
### What problem does this paper attempt to solve?
This paper aims to solve the parameter adjustment problem in Online Logistic Regression, especially the parameter - free algorithm design in the adversarial setting. Specifically:
1. **Background and Motivation**:
- In convex online optimization, parameter adjustment is a crucial issue. Many existing algorithms need to know some constants in advance (such as the exponentiated concavity constant of the loss function and the upper bound of the gradient), but these constants are usually unknown in practical applications.
- The author's goal is to develop algorithms that do not require pre - setting these parameters to improve the robustness and applicability of the algorithms.
2. **Main Contributions**:
- A new second - order algorithm, Semi - Online Step (SOS), is introduced, and its logarithmic regret bound \(O(\log(n))\) in the adversarial setting is proven.
- The Extended Kalman Filter (EKF) is analyzed, and its logarithmic regret bound \(O(\log(n))\) is also proven under specific conditions.
- These results provide a theoretical basis for understanding and improving parameter - free algorithms in online logistic regression.
3. **Specific Problem Description**:
- **Online Logistic Regression Problem**: Given a series of observations \(y_t\in\{- 1,1\}\) and explanatory variables \(X_t\in\mathbb{R}^d\), the goal is to predict the next value \(y_t\) at each time step \(t\).
- **Loss Function**: The logistic loss function \(\ell_t(y_t,\hat{\theta}_t)=\log(1 + \exp(-y_t\hat{\theta}_t^T X_t))\) is used to measure the prediction error.
- **Objective**: Design a recursive update algorithm so that the regret bound of the cumulative loss is as small as possible.
4. **Innovations**:
- The SOS algorithm is proposed, which makes a quadratic approximation in each iteration and can guarantee a logarithmic - level regret bound.
- The EKF is improved and analyzed, and its logarithmic regret bound under certain conditions is proven, which provides theoretical support for the application of EKF in online logistic regression.
5. **Challenges and Solutions**:
- **Challenges**: In the adversarial setting, traditional localization methods are difficult to be directly applied to EKF because the matrix update method of EKF is different from that of SOS.
- **Solutions**: By introducing the SOS algorithm as an intermediate step, the author can gradually approximate the performance of EKF and finally prove the logarithmic regret bound of EKF under specific conditions.
In conclusion, this paper solves the problem of parameter adjustment in online logistic regression, proposes a new algorithm SOS, and conducts in - depth theoretical analysis on it and EKF, proving their effectiveness in the adversarial setting.