TL;DR
This paper introduces DMSN, a novel deep learning framework for predicting users' intended travel destinations on online platforms. By fusing diverse user behavior sequences and incorporating temporal dynamics, DMSN significantly improves destination prediction accuracy, which is crucial for personalized travel recommendations.
What Problem Does This Paper Address?
Online travel platforms aim to provide personalized travel plans, but a key challenge is accurately predicting a user's true destination intention. This is difficult because user behavior is often sparse, infrequent, and influenced by various factors like travel status and the type of product they interact with (e.g., flights, hotels, visa applications). Existing methods often struggle to capture these nuances, leading to less relevant recommendations. The authors highlight that a user might search for one destination but their actual intention lies elsewhere, making accurate destination prediction a critical first step for effective travel planning.
Methods
The core contribution is the Deep Multi-Sequences fused neural Networks (DMSN) architecture. DMSN addresses the sparsity and complexity of user behavior by:
- Multi-Sequence Fusion: It fuses historical behavior sequences from five different categories (train, flight, hotel, vacation item, search) into a single, dense "global cityID sequence." This creates a richer representation than individual sequences.
- Attention Mechanisms: To account for the temporal nature of travel and user interests, DMSN employs attention mechanisms. This allows the model to dynamically weigh the importance of different historical behaviors based on their relevance to the target destination and the time they occurred.
- Dual Architectures: Two variants are proposed:
- DMSN-ATRNN: Uses a Recurrent Neural Network (RNN) within the attention layer to capture sequential correlations and long-term preferences.
- DMSN-ATMC: Employs a multi-grained Convolutional Neural Network (CNN) layer instead of an RNN. This allows for capturing both short-term and long-term preferences simultaneously by using convolution kernels of different sizes.
The overall architecture follows a typical deep learning paradigm for recommendation systems, including an Embedding Layer, an Attention-based Neural Network Layer, an MLP Layer, and a Rank Loss Layer.
Datasets
The authors used real-world log data from Fliggy, a prominent online travel platform. This dataset includes user behaviors across various business lines (train tickets, flight tickets, hotel bookings, and vacation item purchases). The data was split into training, validation, and testing sets. The paper provides statistics on the number of samples, positive/negative samples, users, vacation items, and destinations.
Key Insights and Ideas
- Behavior Fusion is Key: Simply using individual behavior sequences isn't enough. Fusing them into a global sequence provides a more comprehensive view of user intent, especially for sparse travel data.
- Time Matters (A Lot): User preferences for destinations are not static; they change based on travel periods. Incorporating temporal information via attention is crucial for accurate prediction. Recent behaviors are often more indicative of immediate intent.
- Short-term vs. Long-term Preferences: Different users and situations require capturing both immediate interests (e.g., a last-minute booking) and long-standing preferences. The DMSN-ATMC's multi-grained CNN approach is particularly effective at this.
- CityID Representation: Using cityIDs as a representation for behaviors helps alleviate sparsity issues compared to using product IDs, as there are fewer cities than products.
Key Results
- Superior Destination Prediction: Both DMSN-ATRNN and DMSN-ATMC significantly outperformed state-of-the-art models like AutoInt and DIN in intention destination prediction, achieving higher AUC scores.
- DMSN-ATMC Leads: The DMSN-ATMC model, utilizing multi-grained CNNs, achieved the highest AUC among all competitors. This highlights the benefit of its ability to capture both short-term and long-term preferences effectively.
- Vacation Item Prediction Improvement: When integrated into a two-stage vacation item recommendation system, DMSN models (particularly DMSN-ATMC) also led to improved prediction accuracy, demonstrating the downstream benefits of accurate intention prediction.
- Real-world Impact: An online A/B test on Fliggy showed that the proposed ATMC&ATMC model achieved significant CTR improvements across various business lines, with the highest improvement on train tickets (+3.4%).
Limitations and Future Work
- Data Sparsity: While DMSN addresses sparsity, it remains a challenge in travel data, especially for users who travel infrequently.
- Hyperparameter Sensitivity: The performance of DMSN-ATMC can be influenced by hyperparameter choices like the number of kernels and hidden layers, requiring careful tuning.
- Future Work: The authors suggest exploring more sophisticated methods for fusing multi-category behavior sequences and further investigating the impact of temporal dynamics. They also plan to release their code and a large-scale dataset to facilitate future research.
Summary updated