CoS: Towards Optimal Event Scheduling via Chain-of-Scheduling

TL;DR

The paper tackles the NP-hard problem of recommending temporally and geographically feasible event schedules in EBSNs. It introduces Chain-of-Scheduling (CoS), a structured three-stage framework (Exploration, Verification, Integration) that guides LLMs, augmented by Knowledge Distillation from traditional search-based teachers into lightweight, fast models. CoS demonstrates near-theoretical-optimal effectiveness with high efficiency and strong generalization, including zero-shot transfer to unseen cities, by distilling rich spatiotemporal reasoning into LLMs and applying a post-processing validity check. The approach offers interpretable scheduling traces, scalable performance, and practical impact for real-world event recommendation systems. The work also provides NP-hardness proof for Event Scheduling and extensive appendices detailing prompts, analyses, and supplemental experiments.

Abstract

Recommending event schedules is a key issue in Event-based Social Networks (EBSNs) in order to maintain user activity. An effective recommendation is required to maximize the user's preference, subjecting to both time and geographical constraints. Existing methods face an inherent trade-off among efficiency, effectiveness, and generalization, due to the NP-hard nature of the problem. This paper proposes the Chain-of-Scheduling (CoS) framework, which activates the event scheduling capability of Large Language Models (LLMs) through a guided, efficient scheduling process. CoS enhances LLM by formulating the schedule task into three atomic stages, i.e., exploration, verification and integration. Then we enable the LLMs to generate CoS autonomously via Knowledge Distillation (KD). Experimental results show that CoS achieves near-theoretical optimal effectiveness with high efficiency on three real-world datasets in a interpretable manner. Moreover, it demonstrates strong zero-shot learning ability on out-of-domain data.

Figures (12)

• Figure 1: A illustration of event scheduling. (a) is a specific example of event scheduling. (b) denotes the challenges of event scheduling, i.e., existing methods face a trade-off among effectiveness, efficiency, and interpretability while CoS can achieve all-round excellent via exploration, verification, and integration.
• Figure 2: The illustration of the Chain-of-Scheduling (CoS). (a) denotes the basic process of teaching the LLM to generate CoS. Distillation CoS from Combinatorial Optimization, then supervised fine-tune (SFT) the (prompt, CoS) pairs, enabling the LLM to autonomously generate CoS. (b) is a specific running example. Small LLM (e.g., Qwen2.5-7B-Instruct) can quickly arrive at the correct answer via CoS, while directly using a oversized LLM (e.g., DeepSeek-R1) will fall into an overly lengthy and time-consuming thinking process, ultimately leading to an incorrect answer.
• Figure 3: A post-processing example.
• Figure 4: Hyperparameter Sensitivity Analysis.
• Figure 5: Events Utility Score Distribution. We investigate the distribution of utility scores for events users actually attended and for randomly sampled unattended events. The utility scores for attended events are concentrated in a higher range (0.60-0.75), indicating that BERT effectively captures user preferences.