LACRM: LLM-Guided Mining of Directed Advertiser Competition Relationships in E-Commerce Advertising

The gist In e-commerce advertising, ad platforms typically fail to model the externalities among advertisers when making optimization decisions, and this paper proposes LACRM, a novel framework that constructs directed competition graphs with continuous edge intensities via a three-stage pipeline.

Problem Formulation

The objective is to learn the scoring function s(·) to construct the dense competitive strength matrix W ∈ R N ×N, which serves as a foundational marketstructure representation for downstream tasks The framework is grounded in two core principles: Continuous Intensity and Latent Discovery Competition is characterized by the intensity of traffic diversion from vi attributable to vj, operationally approximated using auction co-occurrence signals The competition score is inherently asymmetric (s(vi → vj) ≠ s(vj → vi)) because 77.41% of all edges are strictly unidirectional

Methodology

LACRM converts structured merchant profiles into continuous asymmetric competition intensities across three sequential stages: Dual-Profile Generation, Multi-Route Retrieval, and Dynamic Behavioral Alignment

Dual-Profile Generation:

The framework structurally decouples each advertiser’s semantic footprint into an asymmetric dual-view representation by defining the LLM-guided Advertiser Representation (AR) and Query Representation (QR) AR is a factual profile that captures who the merchant is to serve as the invariant retrieval anchor, whereas QR characterizes a strategic, multi-view trajectory identifying which cohort of rivals the merchant targets to serve as the variable retrieval query

Multi-Route Retrieval:

The system decomposes the query space into a taxonomy of five fine-grained competitive views (V), derived from brand relation and category relation axes, augmented by a temporal overlay For each view ω ∈ V with active fields F(ω) ⊆ F, we define QR(i) = τ(ω)∥ Ê6 k=1 ψκ fk (xi) The resulting dense query embedding q(ωi) and target anchor embedding a j are extracted as follows: q(ωi) = ΦQR(ωi), IQ, a j = ΦAR j, IA

Dynamic Behavioral Alignment:

The final refinement stage is formulated as a fine-grained listwise Learning-to-Rank (LTR) paradigm implemented with LambdaMART The final competition score is produced by a gradient-boosted decision tree ensemble: sfinal(vi → vj) = HLTR(xi j) This ranker learns how the different competition views should be weighted and combined for each candidate pair, together with their interactions with observed market behavior

Experimental Validation

The evaluation is conducted along three dimensions: (1) offline candidate retrieval against realized rivalries, (2) verification of latent rivalries, and (3) online A/B testing in a production advertising system

Offline Retrieval:

The historical Realized Ground Truth graph Greal = (V, Ereal) is established if and only if the share of vi’s traffic diverted by vj within identical auction funnels exceeds a threshold τe = 0.05 over an observation window The novelty fraction indicates that a substantial fraction of evaluation-week competition consists of edges absent from recent history LACRM achieves emerging-edge Recall of 8.6%, 77.9%, and 86.0% at K ∈ 50, 100, respectively

Latent Rivalry Verification:

We assess their legitimacy through two independent and complementary lenses: an ensemble of LLM-as-a-Judge scorers [18] to rate directed competition intensity at scale, and a blind human expert study to guard against shared LLM priors The LLM judges clearly separate the realized and negative controls, with strong cross-judge agreement (ICC = 0.918, mean pairwise Spearman ρ = 0.946; p < 10−16) The latent cohort lies significantly above the negative controls (Mann–Whitney p < 10−16 for each judge) while remaining distinct from realized rivals (KS p ≈ 10−9), with a 61.3% ensemble positive rate

Online Experiments

In the Competition-Aware Bidding Experiment, the LACRM-guided condition yields a 1.73% increase in ad revenue and a 0.91 percentage-point improvement in the daily budget-target fulfillment rate In the Competition-Aware Incentive Experiment, the LACRM-derived competition signals can improve incentive precision and encourage advertisers to express their valuations more actively through bids and budget commitments, yielding a 1.14% increase in ad revenue relative to the existing incentive policy

The results demonstrate that explicit competition modeling provides a reusable market representation for advertiser-facing optimization tasks This framework demonstrates the practical value of explicit competition modeling in large-scale advertising systems. The results suggest that LACRM-derived competition relationships provide a more effective competition signal for auction-level bidding coordination and advertiser budget utilization than the rule-based baseline.

The paper concludes that LACRM enables the capture of heterogeneous competition patterns, the distinction of directional competitive pressure, and the discovery of latent rivals beyond historical auction interactions. The framework successfully combines the broad coverage of LLM-derived semantic representations with the precision of observed behavioral evidence. The results demonstrate that explicit competition modeling in large-scale advertising systems yields measurable gains in advertising revenue and budget fulfillment.

The paper is supported by numerous references including those on auction design [2], market structure surveillance [30], and LLM-based recommendation [42]. The authors contributed equally to this work. The results show that LACRM outperforms its ablated variants and two directed GNN baselines in realized-rival retrieval The deployment of LACRM-guided bidding and incentive allocation increases advertising revenue by 1.73% and 1.14%, respectively.

The paper is supported by numerous references including those on auction design [2], market structure surveillance [30], and LLM-based recommendation [42]. The authors contributed equally to this work. The results show that LACRM outperforms its ablated variants and two directed GNN baselines in realized-rival retrieval. The deployment of LACRM-guided bidding and incentive allocation increases advertising revenue by 1.73% and 1.14%, respectively.

The paper is supported by numerous references including those on auction design [2], market structure surveillance [30], and LLM-based recommendation [42]. The authors contributed equally to this work. The results show that LACRM outperforms its ablated variants and two directed GNN baselines in realized-rival retrieval<ref:

1. Based on LACRM (LLM-driven Advertiser Competition Relationship Mining), an AI system can construct a directed, weighted graph that quantifies asymmetric competitive intensities among advertisers, providing a foundational market structure representation for downstream tasks. This allows the system to move beyond symmetric similarity by using LLM-driven role-separated profiling to uncover both realized and latent rivals.

2. The improved system can perform asymmetric multi-route retrieval to surface both realized and latent competitors, capturing competition across multiple facets. This is achieved by decomposing the query space into five fine-grained competitive views—such as Intra-Brand Cross-Channel Friction or Cross-Category Functional Substitution—to ensure that one vector does not conflate all competition types.

3. The AI can accurately model directed competition intensity by employing a listwise learning-to-rank aligned with temporal auction behavior. This module, using LambdaMART, calibrates semantic candidates against observed market dynamics to produce the final score, as described in the formula: sfinal(vi → vⱼ) = HLTR(xiⱼ).

4. The system can enhance its ability to discover novel competition by implementing a verification protocol. This involves using an ensemble of LLM-as-a-Judge scorers and a blind human expert study to validate latent relations, ensuring that recovered competitors are not merely model-generated artifacts.

5. In production, the system can improve advertising efficiency by integrating these signals into decision-making. Specifically, LACRM can be used for LACRM-guided bidding and incentive allocation, which has been shown to increase advertising revenue by 1.73% and 1.14%, respectively, compared to rule-based baselines.