Research map / Large language models / Language, safety and interpretability
Predicting User Behavior in Smart Spaces with LLM-Enhanced Logs and Personalized Prompts
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a personalized soft-prompt approach built from per-user transition/co-occurrence graphs plus LLM-enhanced event descriptions, evaluated for next-event prediction in smart vehicle and smart home logs.
Problem
Predicting a user's next desired function in smart spaces from sparse, cryptic device logs, where rare events are poorly represented and one-size-fits-all models ignore individual preferences.
Approach
Build per-user transition and co-occurrence graphs from interaction histories, convert them via GCN and an MLP into soft continuous prompts prefixed to the event sequence, and enrich abbreviated event names with LLM-generated natural-language descriptions encoded by BERT to share semantics across related rare events.
Key results
- On City A, the proposed method reaches H@1 0.3450 and H@3 0.5529, versus 0.3308 H@1 for the best baseline (DuoRec).
- The ablation shows the LLM-enhanced text embedding raises H@1 on City A from 0.2774 (ID only) to 0.3176, and adding personalized prompts raises it further to 0.3450.
- For events below 1% frequency, LLM-enhanced descriptions improve H@1 by over 40% and H@3 by over 30%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Ours instead of the strongest baseline | Ours · City A | H@1: 0.3308 -> 0.3450 | Helped |
| Use Ours instead of the strongest baseline | Ours · City B | H@1: 0.3517 -> 0.3743 | Helped |
| Use Ours instead of the strongest baseline | Ours · Home A | H@1: 0.4103 -> 0.4293 | Helped |
| Use Ours instead of the strongest baseline | Ours · Home B | H@1: 0.3118 -> 0.3352 | Helped |
| Add enhanced natural-language event descriptions to ID embeddings (ablation) | Ours · City A and City B | H@1: 0.2774 -> 0.3176; 0.3056 -> 0.3427 | Helped |
| Add personalized prompts to ID and Ernie-enhanced log embeddings (ablation) | Ours · City A and City B | H@1: 0.3176 -> 0.3450; 0.3427 -> 0.3743 | Helped |
| Use Qwen-enhanced log embeddings instead of ID embeddings alone (ablation) | Ours · City A and City B | H@1: 0.2774 -> 0.3064; 0.3056 -> 0.3298 | Helped |
| Use Ernie-enhanced log embeddings instead of ID embeddings alone (ablation) | Ours · City A and City B | H@1: 0.2774 -> 0.3097; 0.3056 -> 0.3354 | Helped |
| Add Qwen-enhanced log embeddings to ID embeddings (ablation) | Ours · City A and City B | H@1: 0.2774 -> 0.3140; 0.3056 -> 0.3396 | Helped |
| Add Qwen-enhanced log embeddings and personalized prompts to ID embeddings (ablation) | Ours · City A and City B | H@1: 0.2774 -> 0.3439; 0.3056 -> 0.3722 | Helped |
| Use LLM-enhanced descriptions for events below 1% frequency | Ours · Vehicle datasets, events below 1% frequency | H@1: over 40% | Helped |
| Increase the number of personalized prompts; peak at six prompts | Ours · Smart vehicle datasets | H@3 and N@3 | Helped |
Limitations
- Industrial vehicle datasets cannot be publicly released due to data provider restrictions.
- Evaluated only on four datasets from two domains (two private vehicle datasets, two CASAS smart home datasets).
- Gains over the strongest baseline are moderate on the main datasets (for example, H@1 0.3450 vs 0.3308 for DuoRec on City A).
Experimental setup
Datasets: City A industrial vehicle logs City B industrial vehicle logs Aruba (CASAS) Milan (CASAS)
Benchmarks: City A City B Home A Home B
Metrics: class-wise Hit Rate@K (H@1, H@3) class-wise NDCG@K (N@3)
Baselines: CasLSTM DnnLSTM ConvLMLC ASGen BERT4Rec DuoRec CL4SRec RecFormer
Builds on: prompt tuning BPR loss chain-of-thought BERT graph convolutional network
Closest papers by meaning
- HYDRA: Model Factorization Framework for Black-Box LLM Personalization (NeurIPS 2024 · similarity 0.84)
- FingerTip 20K: A Benchmark for Proactive and Personalized Mobile LLM Agents (ICLR 2026 · similarity 0.83)
- BehaviorBench: Modeling Real-World User Decisions from Behavioral Traces (NeurIPS 2026 · similarity 0.83)
- How Does Personalized Memory Shape LLM Behavior? Benchmarking Rational Preference Utilization in Personalized Assistants (NeurIPS 2026 · similarity 0.83)
- Tell me about yourself: LLMs are aware of their learned behaviors (ICLR 2025 · similarity 0.82)
- PersonalLLM: Tailoring LLMs to Individual Preferences (ICLR 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.