TL;DR
This paper introduces SessionPath, a novel neural network model for personalized category suggestions in e-commerce type-ahead. SessionPath dynamically predicts category paths, balancing precision and recall by considering both user queries and in-session browsing behavior, which is crucial for improving search relevance in large catalogs.
What Problem Does This Paper Address?
Modern e-commerce search engines face the challenge of presenting relevant results from massive product catalogs. While type-ahead suggestions help narrow down search intent, they often lack personalization and can lead to a poor user experience if they don't accurately reflect the user's needs. Traditional methods struggle to balance precision (showing only highly relevant results) and recall (ensuring all relevant results are discoverable), especially when users might browse multiple pages or re-order products based on custom criteria. This paper addresses the need for a dynamic and personalized approach to category suggestions that can adapt to real-time user behavior and the complexity of product taxonomies.
Methods
The core contribution is SessionPath, a personalized model for facet prediction in type-ahead suggestions. SessionPath employs an encoder-decoder architecture. The encoder processes both the user's query and their in-session browsing history (represented by product embeddings) to create a contextualized representation. The decoder then generates a probability distribution over possible category nodes at each step of the product taxonomy.
A key innovation is the dynamic path generation and a decision module. Instead of predicting a single category, SessionPath predicts a sequence of categories (a path) and uses a decision module to determine the optimal depth of this path. This decision module, parameterized by a confidence threshold, allows for a tunable trade-off between precision and recall. The authors also explore various methods for encoding linguistic input (query) and session context, including word2vec, BERT, and a novel "Search2Prod2Vec" approach that leverages product click data.
Datasets
The authors evaluated their models on data from two partnering e-commerce shops (Shop 1 and Shop 2). These datasets consist of search logs, including queries with context and the associated products. The data was split into training (June-August 2019) and testing (September 2019) sets. No new benchmarks were introduced; instead, the authors focused on evaluating their model against existing count-based and neural network baselines.
Key Insights and Ideas
- Personalization is Key: Simply relying on query text isn't enough. In-session browsing behavior significantly informs a user's true intent, and SessionPath effectively leverages this.
- Dynamic Path Depth Matters: The "paradox of choice" suggests that too many options can be overwhelming. SessionPath's ability to dynamically decide how deep to go in the category taxonomy (e.g., "shoes" vs. "shoes/basketball") is crucial for providing a focused yet comprehensive search experience.
- Confidence-Driven Decisions: The decision module, controlled by a confidence threshold, allows businesses to tune the model's aggressiveness. This provides a principled way to balance precision and recall based on specific business needs (e.g., favoring cleaner result pages vs. broader discovery).
- Hierarchical Structure is Powerful: SessionPath's ability to model the hierarchical nature of product taxonomies allows for more efficient learning and better generalization, especially when compared to models that treat categories as flat labels.
- Domain-Specific Embeddings Shine: While pre-trained language models like BERT are useful, embeddings derived from shop-specific data (like Search2Prod2Vec) can outperform them, suggesting that understanding the nuances of a particular e-commerce domain is highly beneficial.
Key Results
SessionPath variants consistently outperformed count-based (CM) and multi-layer perceptron (MLP) baselines across various metrics and depths. For instance, SP+SV (SessionPath with Search2Prod2Vec) showed up to a 150% increase over CM and a 67% increase over MLP in accuracy. The paper also demonstrated that SessionPath can achieve good performance even with significantly reduced training data (e.g., 1/10th of the dataset), highlighting its sample efficiency. The decision module's tuning was shown to effectively control the precision-recall trade-off, with higher confidence thresholds leading to higher precision and lower recall.
Limitations and Future Work
The authors acknowledge that their evaluation was based on offline experiments using historical search logs. They suggest that A/B testing on a variety of target clients is the next crucial step to validate the model's real-world impact. They also propose several future directions:
- Improving Deep Architecture: Further optimization of the encoder-decoder architecture and exploring advanced linguistic generalization techniques (e.g., transfer learning across tasks or clients).
- Personalized Query Scoping: Applying SessionPath's principles to personalize query scoping directly, potentially improving click-through rates.
- End-to-End Path Depth Modeling: Integrating path depth prediction directly into the decoder, or using reinforcement learning (contextual bandits) to learn the optimal stopping point.
- User Studies: Conducting user studies to gain a deeper understanding of the ideal decision mechanism and how users perceive the suggested category paths.
Summary updated