TLDR: TBGRecall is a new generative AI model for e-commerce recommendations that addresses limitations of traditional models by introducing Next Session Prediction (NSP). Instead of predicting individual items sequentially, NSP focuses on predicting entire user sessions, allowing for more accurate and efficient retrieval of unordered product sets. The model incorporates optimizations like Token-Specific Networks and Multi-Session Prediction, and uses an efficient Partial Incremental Training method. Experiments on large datasets, including Taobao’s, show TBGRecall outperforms existing methods and significantly boosts online transaction metrics.
In the bustling world of online shopping, recommendation systems play a crucial role in helping users discover products they might love. These systems aim to personalize the shopping experience by suggesting relevant items, making e-commerce platforms more engaging and efficient. While generative AI models have shown great promise in various fields, applying them effectively to the “retrieval” stage of recommendation systems – where a large pool of potential items is narrowed down – has presented unique challenges.
Traditional generative models often rely on a method called “autoregressive generation,” which means they predict items one after another in a sequence. This approach works well for tasks like generating text, where order matters. However, in e-commerce, when you open a shopping app, you’re typically presented with a set of products all at once, not one by one in a strict order. This mismatch creates inefficiencies and can limit the accuracy of recommendations.
To tackle this, researchers have introduced a new framework called TBGRecall. This innovative model rethinks how generative AI can be used for e-commerce recommendations, focusing on a core idea: Next Session Prediction (NSP). Instead of predicting the next individual item a user might interact with, TBGRecall predicts the entire “next session” of user interest. This means it understands that items within a single browsing session are often unordered and should not necessarily influence each other sequentially, but rather inform what the user might want in a future session.
The TBGRecall framework achieves this by structuring user interactions into multi-session sequences. Each sequence begins with a “session token” followed by a collection of “item tokens.” During the process of finding recommendations, this session token acts as a guide, helping the system quickly search through a vast catalog of items to find the most relevant ones. To ensure items within a session don’t create artificial dependencies, the model uses a special “Session Mask.” It also incorporates “session-wise rope” for better understanding of positional relationships across sessions.
Beyond its core NSP innovation, TBGRecall includes several smart optimizations. A Token-Specific Network (TSN) helps the model better understand the different types of information associated with session contexts versus individual items. Multi-Session Prediction (MSP) is a particularly impactful enhancement, allowing the model to learn from longer-term user behavior patterns across multiple sessions. The integration of Mixture-of-Experts (MoE) further boosts the model’s ability to learn complex, task-specific knowledge.
Training such a powerful model on the massive datasets of e-commerce platforms like Taobao is a significant undertaking. TBGRecall employs a sophisticated training strategy that combines different types of “loss” functions to ensure it learns to prioritize high-value interactions like clicks and purchases. Crucially, it introduces “Partial Incremental Training (PIT).” This method addresses the challenge of constantly updating models with fresh data. Instead of retraining the entire model on all new data, which can take days, PIT intelligently processes recent data from smaller, rotating segments of users. This allows for daily model updates with significantly fewer computing resources, ensuring that the recommendation system always uses the most up-to-date user preferences without sacrificing performance.
The effectiveness of TBGRecall was rigorously tested on both public benchmarks and a massive industrial dataset from TaoBao, which includes trillions of interactions from half a billion users and items. The results were impressive: TBGRecall consistently outperformed state-of-the-art recommendation methods. On Taobao’s live platform, an online A/B test in the “Guess You Like” section showed a notable increase of 0.60% in transaction count and a significant 2.16% increase in transaction amount. These real-world gains underscore the practical value and scalability of TBGRecall in a large-scale e-commerce environment.
Also Read:
- MOON: A New Generative AI Model for Deeper E-commerce Product Understanding
- New Dataset Enhances Privacy-Preserving Web Search Research
This research represents a significant step forward in generative recommendation systems, bridging the gap between advanced AI modeling and the practical demands of e-commerce. By focusing on session-level understanding and implementing efficient training strategies, TBGRecall offers a robust solution for delivering highly personalized and effective product recommendations. You can read the full research paper for more technical details and insights at this link.
