Multi-agent Auto-Bidding with Latent Graph Diffusion Models

TL;DR

This work tackles auto-bidding in large-scale, multi-agent auctions with KPI constraints by introducing LGD-AB, a framework that combines learnable graph embeddings of impression opportunities and agents with a planning-based latent diffusion model for trajectory forecasting and bid planning. The two-tier approach first encodes auction dynamics via a graph representation and multi-agent context, then uses a latent diffusion model to plan bids while aligning outcomes to KPI targets through reward alignment and offline value-function guidance. Key contributions include a graph-based representation of IO interdependencies, a diffusion-based planning module, and a KPI-aligned training loop demonstrated on real-world AuctionNet data and synthetic auctions, yielding improvements in CPA, ROI, win rate, and social welfare, along with better forecast accuracy under partial observability. The results suggest LGD-AB enables more effective, KPI-compliant auto-bidding in large, open multi-agent auction environments and provides a scalable path toward richer graph-driven planning in practice.

Abstract

This paper proposes a diffusion-based auto-bidding framework that leverages graph representations to model large-scale auction environments. In such settings, agents must dynamically optimize bidding strategies under constraints defined by key performance indicator (KPI) metrics, all while operating in competitive environments characterized by uncertain, sparse, and stochastic variables. To address these challenges, we introduce a novel approach combining learnable graph-based embeddings with a planning-based latent diffusion model (LDM). By capturing patterns and nuances underlying the interdependence of impression opportunities and the multi-agent dynamics of the auction environment, the graph representation enable expressive computations regarding auto-bidding outcomes. With reward alignment techniques, the LDM's posterior is fine-tuned to generate auto-bidding trajectories that maximize KPI metrics while satisfying constraint thresholds. Empirical evaluations on both real-world and synthetic auction environments demonstrate significant improvements in auto-bidding performance across multiple common KPI metrics, as well as accuracy in forecasting auction outcomes.

Figures (5)

• Figure 1: An illustration of the LGD-AB framework. At every time-step, a graph-based embedding is computed for agent $i$ using a bipartite graph of agent and IO nodes. The nodes of a sub-graph $\mathcal{G}_t^i$ consist of two virtual nodes, connecting to IOs exposed and not exposed to agent $i$ respectively. Sub-graphs for other agents are similarly formed and connected accordingly. The graph is processed using a GNN to generate an embedding vector $x^i_t$. The diffusion model then learns a posterior of the temporal sequence of embedding vectors to forecast auction dynamics.
• Figure 2: Inverse Dynamics Model for Bid Computation.
• Figure 3: Agent $i$'s Approximation of Other Agents' Graph $\omega^i(\mathcal{G}_t^{-i})$ by creating $h$ subgraphs used in student model $\mathcal{S}$.
• Figure 4: Computing Joint Embedding Vector using Self-Attention Network.
• Figure 5: KPI Alignment Learning Curve on Synthetic Auction with multi-KPI alignment (shown in red) compared to optimizing each KPI criterion independently (shown in blue). For social welfare, we compare the training between EC and non-EC variants, as shown in the dotted lines.