Travel Intent Prediction in E-commerce: Algorithm Strategies, Multi‑source Behavior Modeling, and Model Design

In this presentation, the speaker from Alibaba explains the problem of recommending travel products, emphasizing that travel demand is low‑frequency, highly sparse, and involves multiple behavior domains such as tickets, hotels, and POIs.

The travel scenario has distinct characteristics: long intervals between user visits, a clear before‑travel → during‑travel → after‑travel state transition, long planning horizons, rich LBS features, and a package‑style demand that spans several behavior domains.

The core prediction target is the user’s destination intent. By aggregating multi‑source travel behaviors over a sufficiently long time window, the model aims to accurately infer the preferred travel destination.

Three main challenges are addressed: (1) the ultra‑low frequency of travel demand requiring long historical windows, (2) the need to capture both long‑term and real‑time user signals, and (3) the multi‑domain nature of travel behavior that must be fused.

To handle multi‑source behavior, the pipeline first normalizes event IDs, aligns timestamps, and merges different behavior streams into a unified sequence. Statistics of sequence lengths for each domain and the fused global sequence are shown.

The model design consists of four steps: (1) enrich the fused sequence with side‑information such as behavior type, (2) extract local features using Multi‑CNN with various window sizes to capture both local and global patterns, (3) apply time‑attention pooling using the user state as a query, and (4) train with a pairwise hinge loss to improve discriminability of intent representations.

Extensive experiments compare several baselines (independent domain, time‑ordered multi‑domain, Autoint, DIN, MLP, ATRNN, ATMC) and demonstrate that the proposed multi‑granular CNN with time‑attention consistently yields stable improvements.

Further, user interest modeling incorporates both real‑time click sequences and exposure‑without‑click sequences across the whole platform using a Transformer plus time‑attention pooling, followed by concatenation of all features and a multi‑layer BN + MLP scorer for CTR/CVR prediction.

The final results show that adding the multi‑granular convolution and time‑attention modules brings a noticeable lift over baselines, confirming the effectiveness of the approach.

In the Q&A, the speaker discusses practical deployment aspects such as output formats (candidate destination lists), collaboration between product and algorithm teams, and considerations for interpreting AUC metrics across different data distributions.

Overall, the talk illustrates how careful feature engineering, multi‑source behavior aggregation, and advanced deep learning architectures can significantly improve travel intent prediction in large‑scale e‑commerce platforms.