Travel Intent Prediction in E-commerce: Algorithm Strategies, Multi-Source Behavior Modeling, and Experimental Results

The presentation introduces Fliggy, an e‑commerce platform for travel products, and explains why travel recommendation differs from other domains due to low‑frequency user demand, sparse behavior data, long decision cycles, and multi‑source LBS characteristics.

It outlines the specific challenges of travel intent prediction: the need for long historical windows, incorporation of recent real‑time actions, and aggregation of heterogeneous behaviors across tickets, hotels, flights, and other travel services.

The authors describe how multi‑source behavior is collected, normalized, and aligned by time and destination dimensions, forming a unified sequence that captures both local and global patterns.

Model design proceeds in four steps: (1) enriching the fused behavior sequence with side‑information; (2) extracting multi‑scale local features using Multi‑CNN; (3) applying time‑attention pooling with user state as query; and (4) training with a pairwise hinge loss to improve discriminative power.

Extensive experiments compare baseline methods (independent domain, sequential strategies, Autoint, DIN, MLP, ATRNN, ATMC) and demonstrate that the proposed multi‑granularity convolution combined with time‑attention yields stable gains in AUC and CTR/CVR metrics.

Further sections discuss integrating the model with real‑time global click streams via Transformer‑based pooling, and how product and algorithm teams collaborate to refine large‑scale recommendation strategies while respecting business rules and user experience.

The Q&A addresses output formats of travel intent predictions, practical applications in entry‑point and feed‑type recommendations, and considerations for interpreting AUC differences across datasets and scenarios.