A Joint Auction Framework with Externalities and Adaptation

Recently, joint advertising has gained significant attention as an effective approach to enhancing the efficiency and revenue of advertising slot allocation. Unlike traditional advertising, which allocates advertising slots exclusively to a single advertiser, joint advertising displays advertisements from brands and stores that have established a joint selling relationship within the same advertising slot. However, existing approaches often struggle to accommodate both joint and traditional advertising frameworks, thereby limiting the revenue potential and generalizability of joint advertising. Furthermore, these methods are constrained by two critical limitations: they generally neglect the influence of global externalities, and they fail to address the bidding variability stemming from multi-party advertiser participation. Collectively, these limitations present substantial challenges to the design of joint auction mechanisms. To address these challenges, we propose a Joint Auction Framework incorporating Externalities and Adaptation, and leverage the automated mechanism design (AMD) method through our proposed JEANet to compute joint auction mechanisms that satisfy the conditions of individual rationality (IR) and approximate dominant strategy incentive compatibility (DSIC). As the first AMD method to integrate global externalities into joint auctions, JEANet dynamically adapts to the bidding characteristics of multi-party advertiser and enables unified auctions that integrate both joint and traditional advertising. Extensive experimental results demonstrate that JEANet outperforms state-of-the-art baselines in multi-slot joint auctions.

💡 Research Summary

**
The paper tackles the emerging problem of joint advertising, where a store and a brand cooperate to bid for the same ad slot, and proposes a novel automated mechanism design (AMD) framework called JEANet (Joint Auction with Externalities and Adaptation). Existing solutions either focus on traditional single‑advertiser auctions or on joint auctions without accounting for two critical issues: (1) global externalities arising from the interaction between ads and organic items in a hybrid list, and (2) the heterogeneous bid distributions of the three participant types (store‑only, brand‑only, and joint advertisers). These gaps limit revenue potential and the applicability of joint advertising in real‑world platforms.

JEANet’s architecture consists of four main components. First, a bid‑extraction module learns type‑specific embeddings from advertisers’ contextual features and historical bids, thereby capturing the distinct statistical patterns (e.g., log‑normal versus normal) observed in real data. Second, a context encoder transforms ad and organic‑item features into high‑dimensional representations, which are combined with slot‑specific click‑through rates (α_k). Third, an externality‑modeling layer defines a probabilistic allocation function a_i(b, ue, c_ad, c_na, α) that explicitly quantifies how placing an ad (or organic item) in a given slot influences the expected user experience (ue) of all other items in the list. Finally, an allocation‑payment network jointly optimizes slot assignments and payments to maximize a weighted sum of platform revenue (expected clicks multiplied by bids) and user‑experience scores, while satisfying two economic constraints: individual rationality (IR) and approximate dominant‑strategy incentive compatibility (approx‑DSIC). The latter is enforced by adding a regret loss term that penalizes any advertiser’s incentive to deviate from truthful bidding beyond a small tolerance ε.

Formally, the mechanism M = (a, p) maps bids, user‑experience vectors, ad/organic contexts, and slot CTRs to allocation probabilities and payments. The learning objective combines revenue, user‑experience, IR penalties, and regret penalties, all of which are differentiable, enabling end‑to‑end stochastic gradient training.

Experiments are conducted on both synthetic datasets (with controlled bid distributions and varying numbers of slots) and a large‑scale industrial dataset from Meituan. Baselines include classic auction rules (VCG, Myerson, GSP), recent AMD models (RegretNet, TICNet), and joint‑advertising specific methods (JAMA, JRegNet). Across multiple metrics—platform revenue, user‑experience score, regret magnitude, and robustness to reduced candidate‑ad sets—JEANet consistently outperforms the baselines. In a 5‑slot setting, JEANet achieves 12–18 % higher revenue and improves user‑experience by over 5 % relative to the best baseline. Notably, when the pool of eligible ads shrinks (a common situation in joint advertising because only stores with brand partnerships qualify), JEANet’s adaptive bid module maintains stable performance, whereas other methods suffer significant drops.

The paper’s contributions are threefold: (1) it introduces the first joint‑auction mechanism that explicitly models global externalities between ads and organic items, (2) it designs an adaptive bid‑extraction component that learns and exploits heterogeneous bid distributions, and (3) it provides an AMD framework that is compatible with both joint and traditional advertising, thereby offering a unified solution for hybrid lists.

Limitations are acknowledged: the current architecture is evaluated on a modest number of slots (K ≤ 5), and the inference cost is higher than that of lightweight rule‑based auctions. Moreover, the quality of the user‑experience distribution estimates directly affects performance, suggesting a need for more robust online learning of ue.

Future work includes scaling JEANet to larger slot counts, incorporating multi‑objective optimization to balance advertiser‑specific KPIs (e.g., conversion rates) with platform revenue, and developing online updating mechanisms for the externality model based on real‑time user behavior.

Overall, JEANet represents a significant step toward practical, revenue‑optimal, and user‑friendly joint advertising systems, demonstrating how modern deep‑learning‑based AMD can reconcile the complex economic and externality considerations inherent in today’s hybrid ad platforms.