Abstract
Route recommendation (RR) is a core task of route planning in the Amap app, with the goal of recommending the optimal route among candidate routes to users. Unlike traditional recommendation methods, insights into the local quality of routes and comparisons between candidate routes are crucial for enhancing recommendation performance but often overlooked in previous studies. To achieve these, we propose a novel model called Comprehensive Comparison Network (CCN). CCN not only uses query-level features (e.g. user features) and item-level features (e.g. route features, item embedding) that are common in traditional recommendations, but also introduces comparison-level features which describe the non-overlapping segments between different routes to capture the local quality of routes. The key component Comprehensive Comparison Block (CCB) in CCN is designed to enable comparisons between routes. CCB includes a Comprehensive Comparison Operator (CCO) and a multi-scenario MLP, which can update the representations of candidate routes based on a comprehensive comparison. By stacking multiple CCBs, CCN can determine the final scores of candidate routes and recommend the optimal one to the user. Additionally, since routes directly affect the costs and risks experienced by users, the RR model must be interpretable for online deployment. Therefore, we designed an interpretable pair scoring network to achieve interpretability. Both offline and online experiments demonstrate that CCN significantly improves RR performance and exhibits strong interpretability. CCN has been fully deployed in the Amap app for over a year, providing stable and optimal benefits for route recommendations.
Introduction
In modern navigation applications, selecting the "best" route is a two-stage process: recall (generating a set of candidate paths) and ranking (selecting the one most likely to be chosen by the user). Traditional recommendation systems, such as those used for media or e-commerce, rely heavily on unique item IDs. However, routes are dynamic paths through billions of possible road segments, making ID-based embeddings impractical.
Furthermore, standard "attention" mechanisms are designed to find co-occurrence correlations, but route selection is fundamentally a competitive process. Users evaluate routes by contrasting specific attributes—for instance, choosing a path that is slightly longer but avoids heavy traffic or tolls. Existing models often fail to capture these subtle, competitive differences between candidate routes, particularly the local quality of specific road segments.
Methodology: The Comprehensive Comparison Network (CCN)
The Comprehensive Comparison Network (CCN) shifts the focus from individual route evaluation to modeling the competitive relationships between all candidates. The architecture integrates three levels of features to provide a holistic view of the recommendation task:
- Query-level Features ($X^s, X^u, X^h$): These capture the broader context, including scenario features (e.g., time of day), user profiles, and historical trip sequences.
- Item-level Features ($X^r$): These represent individual route attributes, such as total distance and estimated time of arrival (ETA).
- Comparison-level Features ($X^c$): This novel feature set focuses on the road segments where candidate routes diverge. By extracting features from these non-overlapping sections, the model learns to distinguish between routes based on local nuances like spatial diversity or local detours.
Key Components
- Comprehensive Comparison Operator (CCO): This operator systematically organizes features so that every route is evaluated against every other candidate, facilitating "full candidate interaction."
- Pair Scoring Network (PSN): To ensure the model is suitable for online deployment, the PSN organizes features into physical fields (e.g., Time, Toll, Distance, Road Capacity). It calculates preference scores for each field, allowing the system to explain why a specific route was recommended (e.g., "This route is recommended because it has fewer traffic lights").
- Multi-input CCB: These blocks are stacked to iteratively refine the representations of route candidates, allowing the model to converge on the optimal recommendation through repeated comparison.
Data and Experimental Setup
The study utilizes two major datasets to validate the CCN framework:
- MSDR: A public dataset focused on user responses to route changes during navigation.
- The New Amap Dataset: A massive pre-navigation dataset introduced by the authors, featuring data from 175 million users, 512 million samples, and 6 billion routes across 370 cities. The primary metric used is the "Coverage Rate" (CR), which measures the overlap between the recommended route and the user's actual path.
Key Insights and Results
- Contrast is King: Route recommendation is a zero-sum game between candidates. By explicitly modeling the differences in non-overlapping segments, CCN focuses on the features that actually drive user choice, avoiding the scalability issues of traditional ID-based embeddings.
- Interpretability and Performance: The Pair Scoring Network demonstrates that explainability does not have to come at the cost of accuracy. By forcing the model to respect physical attributes, it learns more robust representations that align better with real-world logic.
- Benchmarks: CCN outperformed all baselines on the MSDR dataset, raising the AUC from 75.04% to 76.72%. On the Amap dataset, it achieved an offline CR of 85.70%.
- Online Impact: In live A/B testing at Amap, the model showed a +1.2% improvement in online CR and a -0.8% reduction in the "Deviation Rate" (how often users ignore the suggested route).
- Efficiency: Despite its complexity, the model maintains an average response time of 15 milliseconds, meeting the strict latency requirements for real-time industrial applications.
Conclusion
CCN demonstrates that modeling local quality and candidate interactions is vital for high-performance route recommendation. The framework's ability to provide interpretable results while maintaining high accuracy has led to its successful long-term deployment at Amap. Future work will explore incorporating more granular real-time environmental data directly into the comparison blocks to further enhance recommendation precision.
Summary updated