Abstract
Changes of users' query volume in online flight ticketing systems indicate the changes of requirements in civil aviation markets. By analyzing the big data of users' online query behaviors, we can timely and accurately discover abnormal civil aviation requirements. This ability is very conducive for airlines and agencies for taking immediate and effective marketing actions. In this paper, we propose a novel method to discover abnormal civil aviation requirements based on time-series curves of users' query volumes. In addition, we utilize the domestic airline route network to optimize the anomaly detection results from the perspective of a global network rather than that of a single airline. We conduct experiments on real-world users' query datasets collected from an online ticketing site. The experimental results demonstrate that the proposed method can effectively discover abnormal civil aviation requirements from users' online query logs.
Introduction
In the civil aviation industry, demand is rarely static. While there is a predictable baseline, "abnormal" demand—triggered by holidays, major sporting events, social incidents, or natural disasters—can catch airlines off guard.
The challenge is that traditional anomaly detection usually looks at a single flight route in isolation. However, query data is noisy; a spike on one route might be a bot or a random fluke. To truly understand if a market shift is happening, researchers must determine if similar patterns are appearing across related routes (e.g., all flights heading into a specific city for a festival). This paper addresses the gap between local time-series analysis and global network intelligence.
Methodology
The proposed approach consists of two primary stages designed to filter noise and leverage network-wide correlations.
1. Multi-dimensional Local Detection
For every sliding window of query data on a single route, an initial anomaly score is calculated based on three factors:
- Vertical Comparison: Measuring how much the current query volume deviates from the historical "standard curve" for that specific route.
- Horizontal Comparison: Comparing the route's behavior to all other routes during the same time period to filter out national-level trends.
- Self-Complexity: Using Seasonal-Trend decomposition (STL) to isolate the "random noise" component. If a curve is naturally volatile, a small spike is treated as less suspicious than a spike on a normally smooth curve.
2. Network-based Iterative Optimization
This is the paper's core innovation. The aviation system is modeled as a graph where nodes are airports and edges are routes. Assuming that routes sharing an origin or destination should exhibit correlated demand, an iterative algorithm (similar to PageRank) is used to "smooth" anomaly scores across the network. If multiple routes into a specific city show high anomaly scores, the algorithm boosts the confidence that a real event is occurring.
Data and Experimental Setup
The study utilizes a massive real-world dataset from a major Chinese online travel platform, covering 159 cities and 23,416 unique flight routes across four specific query periods in 2015. To validate the algorithm, a subset of 6 routes (connecting Beijing, Kunming, and Xining) was manually labeled by civil aviation revenue analysts, providing a "gold standard" test set of 1,392 samples (378 true anomalies and 1,014 normal instances).
Key Results and Insights
The proposed method outperformed traditional curve-similarity benchmarks across several metrics:
- Enhanced Precision and Recall: While the multi-dimensional approach alone improved results, the network iteration provided the most significant boost, increasing Precision@K and Recall@K by 10% to 20% in many scenarios.
- Queries as Leading Indicators: Because travelers search for flights days or weeks before booking, query volume acts as a "sensor" for future demand, allowing airlines to react proactively before seats are sold out.
- Network Synergy: The model successfully leverages the fact that demand is a collective phenomenon. An anomaly at a specific airport "radiates" through all connected routes, allowing the system to ignore isolated noise and focus on systemic shifts.
Conclusion and Future Work
While the paper demonstrates a significant leap in detection accuracy, several areas remain for future exploration. Integrating external news or social media data could provide root cause analysis for identified anomalies. Additionally, adapting this iterative network approach for real-time, streaming query data and optimizing the computational load for global-scale deployment are key next steps for industry application.
Summary updated