Abstract:In this paper, we study a class of non-smooth non-convex problems in the form of $\min_{x}[\max_{y\in Y}\phi(x, y) - \max_{z\in Z}\psi(x, z)]$, where both $\Phi(x) = \max_{y\in Y}\phi(x, y)$ and $\Psi(x)=\max_{z\in Z}\psi(x, z)$ are weakly convex functions, and $\phi(x, y), \psi(x, z)$ are strongly concave functions in terms of $y$ and $z$, respectively. It covers two families of problems that have been studied but are missing single-loop stochastic algorithms, i.e., difference of weakly convex functions and weakly convex strongly-concave min-max problems. We propose a stochastic Moreau envelope approximate gradient method dubbed SMAG, the first single-loop algorithm for solving these problems, and provide a state-of-the-art non-asymptotic convergence rate. The key idea of the design is to compute an approximate gradient of the Moreau envelopes of $\Phi, \Psi$ using only one step of stochastic gradient update of the primal and dual variables. Empirically, we conduct experiments on positive-unlabeled (PU) learning and partial area under ROC curve (pAUC) optimization with an adversarial fairness regularizer to validate the effectiveness of our proposed algorithms.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is how to efficiently solve a class of objective functions in a specific form in non - smooth non - convex optimization, namely:
\[
\min_{x \in \mathbb{R}^d} F(x) := \max_{y \in Y} \phi(x, y)-\max_{z \in Z} \psi(x, z)
\]
where:
- \( \Phi(x)=\max_{y \in Y} \phi(x, y) \) and \( \Psi(x)=\max_{z \in Z} \psi(x, z) \) are weakly convex functions.
- \( \phi(x, y) \) and \( \psi(x, z) \) are strongly concave functions in \( y \) and \( z \) respectively.
- \( Y \) and \( Z \) are convex and compact sets.
This type of problem encompasses two sub - classes that have been studied but lack single - loop stochastic algorithms:
1. Difference of Weakly Convex Functions (DWC).
2. Weakly Convex Strongly Concave Min - Max Problems (WCSC).
### Main Contributions
1. **Construct a Unified Framework**: A new framework (DMax optimization) is proposed, which unifies DWC optimization and WCSC min - max optimization.
2. **Single - Loop Stochastic Algorithm**: Based on the Moreau envelope smoothing technique, a single - loop stochastic algorithm (SMAG) is proposed, which is applicable to non - smooth DMax optimization and achieves a convergence rate of \( O(\epsilon^{- 4}) \).
3. **Experimental Verification**: The effectiveness of the proposed algorithm is verified through the applications of positive - unlabeled learning (PU learning) and partial AUC optimization (with adversarial fairness regularization).
### Background and Challenges
- **Non - Smooth Non - Convex Optimization**: The non - smoothness and non - convexity of the objective function make the optimization problem more complex.
- **Maximization Sub - problems**: Computing \( \max_y\phi(x, y) \) and \( \max_z\psi(x, z) \) requires solving maximization sub - problems, which makes it impossible to directly obtain unbiased gradient estimates.
- **Limitations of Existing Methods**: Existing methods usually require a double - loop structure to solve sub - problems, resulting in complex implementation and a large amount of hyper - parameter tuning.
### Solutions
- **Moreau Envelope Smoothing**: Apply the Moreau envelope smoothing technique to each component function respectively, and take their difference as the smoothed approximation of the original objective function.
- **Single - Step Update**: Avoid the use of inner loops by performing only one - step update for each nested optimization problem.
- **Fast Convergence**: Utilize the fast convergence properties of strongly convex/strongly concave problems to prove that single - step update is sufficient to achieve the state - of - the - art convergence rate.
### Experimental Results
- **PU Learning**: Experiments were carried out on multiple datasets, and the results show that the performance of the SMAG algorithm is better than that of the baseline methods.
- **Partial AUC Optimization**: By introducing adversarial fairness regularization, the unfairness problem of the model in some protected groups is solved, and the experimental results verify the effectiveness of the algorithm.
In conclusion, this paper effectively solves the difference of weakly convex functions and weakly convex - strongly concave min - max optimization problems by proposing a new single - loop stochastic algorithm, and it has been verified in practical applications.