TL;DR
IntentGR is a novel end-to-end generative retrieval framework that prevents "intent drift" in personalized search by reinforcing query semantics at every stage of the model pipeline. By explicitly aligning query intent with item identifiers and enriching long-tail query data, it achieves up to an 11.96% improvement in retrieval accuracy over state-of-the-art baselines.
What Problem Does This Paper Address?
Generative Retrieval (GR) is a rising paradigm where models directly "generate" the identifier (SID) of a relevant item, rather than ranking a pre-defined list. While end-to-end GR—which jointly learns how to tokenize items and how to retrieve them—has shown promise in recommendation systems, it often falters in personalized search.
The core issue is intent drift. In search, the model must balance a user's long-term preferences (e.g., "I like electronics") with their immediate intent (e.g., "I need a sleeping bag"). Current models often treat the search query as a passive input feature. Consequently, the generated identifiers tend to drift toward high-frequency token patterns (popular items) rather than the specific intent of the query. This is especially problematic for "long-tail" queries that appear infrequently in training data.
Methods
The authors propose IntentGR, which injects query intent into four complementary stages of the end-to-end pipeline:
- Query-Aware Semantic Behavior Representation (QBR): Beyond the user's personal history, QBR adds a "query-level behavior sequence"—a list of items globally popular for that specific query. This acts as a form of query-level collaborative filtering.
- Query-Semantic Alignment (QSA): This introduces a shared contrastive objective. It forces the model to align the representation of the query with the reconstructed representation of the target item, ensuring the learned "language" of identifiers is grounded in search intent.
- Long-Tail Query Enrichment (LQE): To combat the dominance of frequent queries, LQE uses repeated sampling and query rewriting (generating semantic variants of rare queries) to give the model more exposure to diverse, low-frequency intents.
- Query-conditioned Attention Modulation (QAM): During the actual generation of the item ID, QAM explicitly injects global query information into the decoder's attention mechanism. This ensures the model doesn't "forget" the query as it generates long sequences of ID tokens.
The model follows an alternating training strategy, where it switches between optimizing the item tokenizer and the generative retriever before a final fine-tuning phase.
Datasets
The framework was evaluated on two distinct datasets:
- AOL4PS: A public benchmark for personalized search derived from AOL web search logs.
- AE: A massive, real-world industrial dataset from AliExpress, containing over 11 million users and 9 million items. This dataset is particularly challenging due to the "lexical gap" between colloquial user queries and keyword-heavy product titles.
Key Insights and Ideas
- Query-Level Evidence is a Game Changer: The authors found that simply adding query-level behavioral evidence (QBR) provides a much stronger signal for target items than user history alone, especially for inactive users.
- Generalization vs. Memorization: By using QSA and QAM, the model moves from simply "memorizing" which items go with which queries to "abstracting" a semantic relationship. This allows the model to perform better even when it hasn't seen a specific user-query pair before.
- Balancing the Tail: Long-tail queries represent the vast majority of unique intents (87% in AOL4PS). LQE proves that you don't need more data to improve these; you just need to utilize the existing "tail" data more effectively through augmentation and balanced supervision.
Key Results
IntentGR consistently outperformed both discriminative (like BM25) and generative (like TIGER and ETEGRec) baselines:
- Recall@10 Improvements: It achieved an 8.11% gain on AOL4PS and a 11.96% gain on the industrial AE dataset compared to the strongest baseline.
- Long-Tail Success: On the AE dataset, IntentGR improved Recall@10 for the "Low" frequency query bucket by a staggering amount, nearly doubling the performance of the baseline in some metrics.
- Robustness: Hyper-parameter analysis showed that the model is relatively stable across different weights for its alignment and attention modulation components.
Limitations and Future Work
While IntentGR significantly improves intent-awareness, the authors note a few areas for growth:
- Evidence Sources: The current model relies on text and click history. Future work could incorporate multimodal signals (like product images) or session-level context (the sequence of queries in a single sitting) to further refine intent.
- Scalability: While tested on 9 million items, the computational cost of end-to-end training on even larger, rapidly changing corpora remains a challenge for all generative retrieval methods.
Summary updated