Over recent decades, the tourism industry has demonstrated progressive expansion, driven by advancements in aviation technologies and shifting consumer interests. In this context, online flight itinerary ranking has become a pivotal business for Online Travel Platforms (OTPs), which aim to rank flight itineraries by synthesizing real-time flight data provided by airlines with users' individual travel preferences. Currently, most OTPs rely on rule-based methodologies or rudimentary user preference-driven models to address this task. However, these methods are inherently limited by their insufficient consideration of delayed booking behaviors and their neglect of dynamic contextual attributes associated with flight itineraries, thereby undermining their ability to effectively handle the intricacies of flight ranking. To address these shortcomings, this paper introduces the Delayed Conversion Modeling based Personalized Flight Itinerary Ranking Network (DCRNet), designed to improve ranking accuracy by integrating delayed booking patterns and contextual dependencies into the modeling framework. Specifically, DCRNet explores the dynamic associations between users' current contextual information and their historical travel records, and models users' delayed booking behaviors via a masked attention mechanism. Moreover, an enhanced multi-task learning framework is employed to effectively integrate traditional behavioral modeling with delay-aware modeling, thereby improving the overall prediction accuracy and enhancing the system's personalized recommendation capabilities. Extensive offline experiments conducted on real-world datasets from Amadeus and Fliggy demonstrate the superior performance of DCRNet. Furthermore, its successful deployment on Fliggy's online itinerary search system has yielded significant improvements, underscoring its practical effectiveness and scalability.

Price inconsistency between the flight listing page and subsequent booking stages is a critical yet underexplored challenge in large-scale Online Travel Platforms (OTPs). Due to caching latency and real-time inventory dynamics, particularly cabin-class exhaustion, users frequently encounter unexpected fare increases, leading to degraded trust and reduced conversion. While existing ranking and CTCVR models excel at relevance and conversion prediction, they largely ignore the impact of price volatility on user experience. In this work, we formally define the price change aware ranking problem and propose PCANet, Price Change Aware Framework for Mitigating Inconsistencies in Large-Scale Ranking Systems. PCANet integrates three key components: (1) Price Consistency Aligning (PCA), a pre-ranking module that calibrates cached prices using real-time inventory signals; And (2) Price-Sensitive Matching (PSM), a personalized attention mechanism that adapts ranking based on individual user sensitivity to price jumps; Extensive offline experiments on production data and large-scale online A/B tests on Fliggy demonstrate that PCANet significantly improves both ranking accuracy and price consistency, yielding substantial gains in user engagement and booking conversion. To the best of our knowledge, this is the first industrial-scale solution to address price inconsistency in flight ranking systems.

Fuzzy flight search is a vital traffic entry for large-scale platforms like Fliggy, serving over 200k daily active users. Unlike traditional fixed-itinerary searches, fuzzy search involves highly ambiguous intentions and flexible constraints, leading to low conversion rates (1.6% UV-CVR) due to sparse intent signals and complex user trade-offs. Standard ranking models struggle with the domain discrepancy between precise and fuzzy queries, often resulting in a misalignment between offline metrics and online performance. To address these challenges, we propose the Multi-level Cross-scenario Knowledge Distillation (MCKD) framework. MCKD transfers ''dark knowledge'' from a high-capacity Teacher model (trained on precise search data) to a lightweight Student (fuzzy search) model. Our framework introduces three core innovations: 1. Feature-level Hint Learning to align latent semantic representations across heterogeneous feature spaces; 2. Uncertainty-aware Distillation to adaptively weight knowledge transfer based on teacher confidence, mitigating noise propagation; 3. Pairwise Ranking Distillation to explicitly preserve the ranking manifold and relative preference orders. Extensive industrial evaluations and online A/B tests demonstrate that MCKD significantly outperforms state-of-the-art baselines and successfully translates offline gains into substantial online conversion improvements, offering a scalable solution for intent-vague retrieval tasks.

The paradigm of ``pre-training and prompt-tuning", with its effectiveness and lightweight characteristics, has rapidly spread from the language field to the graph field. Several pioneering studies have designed specialized prompt functions for diverse downstream graph tasks based on various graph pre-training strategies. These prompts concentrate on the compatibility between the pre-training pretext and downstream graph tasks, aiming to bridge the gap between them. However, designing prompts blindly to adapt to downstream tasks based on this concept neglects crucial security issues. By conducting covert attacks on downstream graph data, we find that even when the downstream task data closely matches that of the pre-training tasks, it is still feasible to generate highly misleading prompts using simple deceptive techniques. In this paper, we shift the primary focus of graph prompts from compatibility to vulnerability issues in adversarial attack scenarios. We design a highly extensible shield defense system for the prompts, which enhances their robustness from two perspectives:Direct Handling and Indirect Amplification. When downstream graph data contains unreliable biases, the former directly combats invalid information by incorporating hybrid multi-defense prompts to the input graph's feature space, while the latter adopts a training strategy to bypass the invalid components and amplifies valid part. We provide a theoretical derivation that proves their feasibility, indicating that unbiased prompts exist under certain conditions on unreliable data. Extensive experiments across various scenarios of adversarial attacks (including adaptive and non-adaptive attacks) indicate that the prompts within our defense system exhibit enhanced resilience and superiority. This paper explores a new perspective in graph prompt learning, offering a novel option for robust prompt tuning in downstream tasks.

With the flourishing development of aviation and the convenience of booking flights online, nearby flight recommendation has become the core business of Online Travel Platforms (OTPs). Nearby flight addresses the issue of inadequate flight options for travelers by offering more cost-effective alternatives, such as recommending flights from nearby cities or on nearby departure dates. Currently, mainstream OTPs adopt rule-based or simple user preference-based strategies to recommend nearby flights. However, the insufficient emphasis on the user's historical behaviors and the ignorance of nearby flight's context make these existing strategies less effective in solving the nearby flight recommendation. To this end, a Context-based Personalized Deep Net work (CPNet) is proposed in this paper for nearby flight recommendation. In CPNet, a Personalized Preferences Learning (PPL) component is first proposed to encapsulate users' individual preferences, leveraging crucial feature correlations between historical behaviors and target nearby flight. Then, a Historical Cost Learning (HCL) component is designed to learn the price sensitivity of users under the same query and the same nearby flight recommendation. Finally, we present a Context Potential Gain Learning (CPGL) component, where the important cost between target nearby flight and context flights are emphasized and learned. Offline experiments on a production dataset and a world-scale online A/B test at Fliggy. Fliggy: https://www.fliggy.com/ both demonstrate the superiority of the proposed CPNet over baselines. CPNet is now successfully deployed at Fliggy, one of the largest OTPs in China, serving millions of users every day for flight reservations.

In computational social systems, recommendation functionality plays a pivotal role in influencing user behavior, enhancing user experience, and driving engagement. To help recommendation functionality to better suggest relevant content or items, the social platforms usually utilize large-scale knowledge discovery techniques to analyze trends in user interactions and extract patterns from large datasets. Click-through rate (CTR) prediction is crucial in recommendation systems for measuring effectiveness, understanding user behavior, training and optimizing models, impacting business outcomes, enhancing personalization, and identifying issues. It provides actionable insights that assist in continuously refining and improving the recommendation process. Traditional deep learning-based CTR prediction models cannot work well for recommendation in social systems due to the data sparsity and the long-tail data problems since the representation learned from the user behavior is basically dominated by the major part of the data. In this article, we propose a multitask-based self-supervised learning model (MTSSL) that can better deal with sparse and long-tail user interaction data. Specifically, we first transform the CTR prediction task into the multitask joint learning framework with a set of shared subnetworks. Each subnetwork learns a representation of the entire user data, and hence, the sparse and long-tail data would have opportunity to fall into the best matched representation space of historical user behavior. Moreover, two kinds of self-supervision signals are employed to guide the learning of the representations. Extensive experiments over four user interaction datasets demonstrate the superiority of our proposed MTSSL over state-of-art models for recommendations. In terms of online A/B test, our model achieves around 3% better performance than the counterparts.

Recently around half of the world’s current population resides in urban areas and benefit from rich services in the smart city. The majority of smart city services are recommendation-related services, and with the development of Internet, most recommendation services in smart economy are online recommendations. Online travel platforms (OTPs) (like Booking, Airbnb, Ctrip, and Fliggy) provide people sufficient resources and convenient approaches to plan and enjoy their trips in smart city. Hotel recommendation is essential for the success of OTPs. However, it is more challenging compared to item recommendation in typical E-commerce scenarios (e.g., Taobao, Jd, and YouTube). The in-nature characteristics of low-frequency and high unit-price lead to more severe sparse and long-tail data distributions. Moreover, for enhancing user experience and business returns, the recommender system seeks to improve both click-through rate (CTR) and conversion rate (CVR) where the seesaw phenomenon may occur. In order to address the aforementioned shortages in hotel recommendation, a multitask learning (MTL) method with a novel flexible multilevel extraction network [denoted as flexible MTL (FMTL)] is proposed. Particularly, FMTL takes MTL into consideration in a unified representation learning framework and is divided into feature encoding and task prediction. In the feature encoding phase, we introduce a novel multirepresentation extractor with temperature-adjusted gating mechanism (T-MRE) for each task, producing more flexible representations for sparse and long-tail data. Moreover, we fuse different representations for each task with three strategies during the prediction phase and empirically demonstrate that the simple concatenation strategy is superior than other relatively complex gating approaches. Offline and live experiments with regard to both overall metrics and user group analysis based on the scarcity of user behaviors illustrate that without significantly increasing model parameters, our FMTL model outperforms substantially over several state-of-the-art models.