Balancing Semantics and Efficiency: A Large‑Scale Personalized Generative Retrieval Method (ICDM 2025)

Background

Generative Retrieval (GR) uses a sequence‑to‑sequence model to decode document IDs (docIDs) directly from a query, simplifying the retrieval pipeline. Existing GR approaches for large‑scale search face two issues: (1) they do not encode efficiency information when producing docIDs, and (2) they ignore the latent relationships among tokens at the same position in a docID.

Proposed Method

The authors introduce a method that balances semantics and efficiency across three stages: representation learning, metric learning, and hierarchical clustering, followed by a position‑aware generative model.

1. Product Vector Representation

A representation‑learning model takes user features, query features, and context features as input and applies an attention mechanism to produce three distinct product vectors: a semantic vector, a generic vector, and an efficiency vector. Similarity scores between the user vector and each product vector are computed (semantic relevance score and efficiency score). The model is trained to minimize the loss between predicted scores and target scores, ensuring that each vector captures its intended meaning.

2. Metric Learning Model

The three vectors are fused by a multilayer perceptron (MLP) to obtain a final product vector used for clustering. A triplet margin loss forces vectors of items with the same label to be closer than those with different labels.

3. Category‑Guided Hierarchical Clustering

The clustering is split into a semantic layer and an efficiency layer. In the semantic layer, items are grouped according to a category hierarchy tree. In the efficiency layer, items are clustered based on their product vectors, and each cluster receives an efficiency score equal to the average CTR of the past 30 days. The encoded results from both layers are concatenated to form the final docID. The algorithmic steps are illustrated in Algorithm 1 (image).

Generative Model with Position‑Aware Loss

Existing GR models (e.g., BART, T5) treat the output as a generic token sequence. The proposed model treats the docID as a structured sequence with three properties: (1) errors in earlier positions have larger impact, (2) tokens at the same position should be semantically close to the target, and (3) tokens at the same position should have similar efficiency. To capture these, a position‑aware loss is defined as the sum of hierarchical loss (weight decay favoring early positions), semantic loss (penalizing semantic distance), and efficiency loss (penalizing efficiency distance). Optimizing this loss aligns the generated docID with both semantic relevance and efficiency.

Key Technical Contributions

Discriminative product vector representation that separates semantic, generic, and efficiency information.

Category‑guided hierarchical clustering that jointly encodes semantic hierarchy and efficiency scores.

Position‑aware loss that integrates hierarchical, semantic, and efficiency penalties.

Incorporation of user personalization both in vector generation and in the generative model input.

Paper Reference

Hi‑Gen: Generative Retrieval For Large‑Scale Personalized E‑commerce Search, arXiv https://arxiv.org/abs/2404.15675.