CtrlCE: Bridging Personalization and User Control in Scientific Personalized Search
Paper: https://arxiv.org/abs/2411.02790 Authors: Sheshera Mysore, Garima Dhanania, Kishor Patil, Surya Kallumadi, Andrew McCallum, Hamed Zamani Published: 2024-11-05
Problem and Motivation
Personalized search learns user preferences from per-user historical interaction data, and the inferred preferences help personalized ranking models surface more relevant documents. However, personalization is widely seen as opaque in how it uses historical interactions, it is not amenable to users' control, and it limits the diversity of information users are exposed to. Automatically diversifying results does little to fix the underlying lack of control over personalization.
The Approach: CtrlCE
The authors propose a personalized search model that lets users proactively control personalized rankings. The system has two main components:
1. CtrlCE (controllable cross-encoder): a novel cross-encoder augmented with an editable memory built from the user's historical interactions. The editable memory makes cross-encoder personalization efficient and gives users direct control over how their history influences ranking.
2. Calibrated mixing model: since not all queries need personalization, a calibrated model determines *when* personalization is necessary, so users exercise control via their editable memory only when it matters.
Evaluation
The paper evaluates CtrlCE along three axes:
- Empirical ranking performance across four domains of science;
- The mixing model's ability to selectively request user control, assessed via a calibration evaluation;
- The degree of control afforded by the editable memory, assessed in a user study.
- Architecture: editable user memory is a lightweight mechanism to inject personalization into cross-encoders without retraining per user.
- Selectivity: personalization should be applied conditionally; a calibrated gating model avoids unnecessary personalization and reduces exposure-diversity harms.
- Transparency and control: giving users an inspectable, editable representation of their preferences addresses opacity concerns that standard learned personalization cannot.
- Evaluation practice: beyond relevance metrics, evaluating calibration of when-to-personalize and running user studies strengthens claims about real user benefit.
- A Survey of Personalization: From RAG to Agent (arXiv:2504.10147)
- Can Large Language Models Understand Preferences in Personalized Recommendation? (arXiv:2501.13391)
- IntentRec: Predicting User Session Intent with Hierarchical Multi-Task Learning (arXiv:2408.05353)
- Unified Embedding Based Personalized Retrieval in Etsy Search (arXiv:2306.04833)
- User-LLM: Efficient LLM Contextualization with User Embeddings (DOI: 10.1145/3701716.3715463)
Quantitative results should be checked against the original PDF tables.
Takeaways for Search and Personalization
Limitations
As with much of this literature, evaluation scale, domain coverage, and alignment between offline benchmarks and live user populations may constrain generalization; readers citing numerical results should consult the original paper.