Towards Full Candidate Interaction: A Comprehensive Comparison Network for Better Route Recommendation

TL;DR

This work addresses the challenge of ranking candidate routes in large-scale navigation by moving beyond ID-based item representations to comparison-level features that capture route differences. It introduces the Comprehensive Comparison Network (CCN), leveraging a Comprehensive Comparison Operator (CCO) and Multi-input Comprehensive Comparison Blocks (MCCB) to enable cross-route interaction, with an optional interpretable Pair Scoring Network (PSN) for explanation. The method jointly optimizes a ranking loss and a pairwise loss to align recommendations with user behavior on both offline and online metrics, demonstrated on MSDR and a large pre-navigation dataset, achieving state-of-the-art performance and practical interpretability. CCN outperforms baselines, provides field-aware explanations, and has been deployed in AMAP for sustained benefits, marking a meaningful step toward scalable, explainable route recommendation in industrial settings.

Abstract

Route Recommendation (RR) is a core task in route planning within online navigation applications, aiming to recommend the optimal route among candidate routes to users. Industrially, RR adopts the two-stage recall-and-rank framework instead of traditional route planning algorithms primarily for computational efficiency. However, RR fundamentally differs from traditional recommendation systems that follow this paradigm. First, a primary challenge is that route items cannot be assigned unique identifiers. Additionally, RR fundamentally differs from traditional recommendation systems in its approach to feature interaction. These differences render conventional recommendation approaches inadequate for route recommendation scenarios, necessitating specialized methods that can effectively handle route-specific challenges. To address these challenges, we propose a novel method called Comprehensive Comparison Network (CCN) for route recommendation. CCN constructs comparative features by comparing non-overlapping segments between route pairs, enabling difference learning without the infinite scalability issues of ID embeddings. Furthermore, CCN employs a specially designed Comprehensive Comparison Block (CCB) that differs from previous item attention methods to achieve effective cross-interaction between routes using comparison-level features. Moreover, we develop an interpretable Pair Scoring Network (PSN) for route recommendation and introduce a more comprehensive route recommendation dataset to advance research in this field. Experimental results demonstrate the effectiveness of our method, and CCN has been successfully deployed in AMAP for over a year, demonstrating its value in route recommendation.

Figures (5)

• Figure 1: The importance of contrastive cues in route recommendation. (a) Three types of features in recommendation systems. (b) Aggregated item-level route features constrain the model to basic difference identification, while segment comparison reveals that route 2 offers spatial diversity whereas route 3 involves detours compared to route 1. However, existing attention mechanisms are not designed for such comparative features.
• Figure 2: CCN mainly consists of stacked CCBs, which can learn and update route representations cyclically and finally get the route score.
• Figure 3: Illustration of CCO in route interaction. (a) Without route interaction, features are simply tiled and concatenated, resulting in limited context. (b) With CCO, route features are transformed into a comprehensive comparison representation by combining each route feature with others, enabling rich pairwise interactions.
• Figure 4: Feature field organization and interpretable PSN with personalized recommendation examples.
• Figure 5: An example of Pair Scoring Network interpretability.