Three Amap Papers at CIKM 2026: Route Planning, Generative Search & Recommendation, and Causal Inference

ACM International Conference on Information and Knowledge Management (CIKM) is a top‑tier venue for research on information retrieval, knowledge management, data mining, and database systems. The conference now emphasizes large language models, generative AI, and trustworthy AI, reflecting their growing influence on information access and integration.

Paper 01 – GenMRP: A General Multi‑Route Planning Framework for Efficient and Personalized Real‑Time Industrial Navigation

Industrial navigation must deliver millisecond‑level responses on massive road networks. Existing approaches fall into two groups: (1) pre‑computed road weights with heuristic generation of alternatives, which struggle to provide personalized and diverse routes; and (2) generative methods that build routes segment by segment but cannot meet the computational efficiency required for large‑scale real‑time scenarios.

GenMRP addresses these gaps with an iterative framework. The first iteration computes the optimal route; subsequent iterations apply a novel Correctional Boosting mechanism that balances route quality and diversity. Each iteration feeds network features, user click history, and previously generated routes into a Link Cost Model to update edge weights, then uses a bidirectional Dijkstra algorithm for path finding (training replaces Dijkstra with sampled optimal routes). Experiments on both offline benchmarks and online Amap traffic show high computational efficiency, strong personalization, and improved diversity. Deployed for over a year in the Amap app, the system increased user trajectory coverage by more than 0.5%.

Paper 02 – IntSR: A Deployed Generative System for Integrated Search and Recommendation at Amap

Separating search and recommendation on internet platforms raises engineering costs and fragments user experience. IntSR unifies these core tasks with a single generative model, handling explicit queries and implicit interest signals within one architecture.

The system introduces two key innovations: (1) a high‑efficiency computing architecture called QDB , designed for massive deployment and capable of meeting sub‑hundred‑millisecond latency requirements; and (2) a time‑aware negative‑sampling strategy that adapts to rapidly changing online products and services, ensuring the model always trains on the most recent valid information.

IntSR has been running stably in the Amap app for two years. Large‑scale online A/B tests demonstrate substantial business impact: +9.34% growth in digital‑asset GMV, +2.76% increase in POI recommendation click‑through rate, and +7.04% improvement in travel‑mode suggestion accuracy. The paper also provides a detailed blueprint for designing, optimizing, and deploying large‑scale unified generative models in complex production environments.

Paper 03 – FunnelCausalNet: Funnel‑Aware Joint Conversion‑Revenue Uplift for Multi‑Tier Coupon Allocation

Coupon allocation must estimate not only conversion probability but also the resulting transaction amount under a subsidy budget. Gross Merchandise Value (GMV) exhibits strong zero‑inflation and a long‑tail distribution, and follows a natural “no conversion, no amount” funnel, making direct GMV modeling ineffective.

FunnelCausalNet tackles this by sharing representations to jointly learn conversion probabilities and conditional amounts across multiple coupon tiers. It applies a hard‑funnel composition to generate expected GMV for each tier, then computes uplift by contrasting with a control group. During training, the conversion head uses the full dataset while the amount head focuses on converted samples; optional consistency and monotonicity constraints can be added per scenario. This structure mitigates sparsity and long‑tail challenges.

The model completes an “estimate‑audit‑decision” pipeline: it uses RCT‑based tier‑wise mean calibration, conformal intervals, and conflict screening for risk audit, followed by a scalable budget‑constrained allocator for joint decision making. Experiments confirm that FunnelCausalNet improves GMV uplift estimation in zero‑inflation settings and provides a practical, end‑to‑end modeling path for multi‑objective intelligent marketing.

These three papers illustrate Amap’s research contributions spanning route planning, generative search/recommendation, and causal inference, each validated through extensive experiments and real‑world deployments.