Technical Summary: Think Before You Act (IntentPOI)

Problem Statement

The paper addresses the task of Next Point-of-Interest (POI) Prediction, which involves recommending the location a user is likely to visit next based on historical check-in trajectories. While recent methods leveraging Large Language Models (LLMs) have demonstrated strong reasoning capabilities, the authors identify a critical limitation: existing approaches typically formulate the task as a one-step trajectory-to-location mapping problem.

These methods (both prompt-based and token-based) directly map historical check-ins to a target POI or re-rank candidates. Consequently, they are prone to:

1. Shallow trajectory correlations: Relying heavily on implicit pattern matching rather than deep behavioral understanding.
2. Historical frequency bias: Defaulting to frequently visited locations even when they are semantically inconsistent with the current context (e.g., predicting a gym visit during a time slot typically used for dining).

The authors argue that human mobility decisions are rarely made directly at the POI level. Instead, users first form a traveling intention (e.g., dining, socializing, shopping) and then select specific locations that satisfy that intention under spatial and temporal constraints. Current LLM-based methods fail to explicitly model this intermediate reasoning step.

Methodology: IntentPOI

To address these limitations, the authors propose IntentPOI, a two-stage intention-guided reasoning framework operating on the principle of "Think Before You Act." The framework decouples intention inference from location prediction.

1. Thinking Stage: Multi-Rationale Intention Inference

In this stage, the system infers the user's latent traveling intention before selecting a location. The inference is grounded in multi-rationale evidence:

• User Profile ($P_n^F$): Generated by an LLM ($M_F$) analyzing the user's historical trajectory ($X_n$) and statistical summaries (frequency distributions of hours, POIs, and categories). This provides a foundation of long-term mobility patterns.
• Peer Behaviors ($P_n^S$): To mitigate data sparsity, the system identifies similar users based on:
  • Semantic Similarity: Cosine similarity of text embeddings derived from user profiles.
  • Geographic Similarity: Jaccard index of visited geohash cells.
    The top-$k$ similar users' check-ins within a temporal window ($\tau$) are aggregated to provide cross-user mobility evidence.
• Temporal Context: The query trajectory ($X_n^q$) provides immediate context.

An LLM ($M_I$) synthesizes these inputs to generate a structured Intention ($P_n^I$). This output includes the likely POI categories (e.g., "Wine Bar," "Mall") and a natural-language reasoning trace explaining why this intention is inferred based on the evidence.

2. Acting Stage: Intention-Guided Reasoning for POI Prediction

Once the intention is established, the system performs candidate recommendation:

• Candidate Pool Construction: A compact pool ($H_n$) is built by combining:
  • Historical Candidates: All POIs previously visited by the user.
  • Spatial Candidates: The $Z$ nearest POIs to the user's last known location.
• Hybrid Filtering: To ensure efficiency, the pool is filtered to size $B$ using a hybrid strategy:
  • Selecting top candidates based on a visiting score (frequency, day-of-week, hour-bucket matches).
  • Selecting top candidates based on spatial proximity.
• Intention-Grounded Reasoning: A reasoning LLM ($M_R$) takes the query trajectory, user profile, the inferred intention ($P_n^I$), and the filtered candidate pool ($\bar{H}_n$) as input. The intention acts as a reasoning scaffold, guiding the LLM to rank candidates that best align with the inferred goal, rather than just the most frequent or closest ones. The output is a ranked list of POIs with supporting reasons.

Key Contributions

1. Reformulation of the Task: The paper identifies the lack of explicit intention modeling as a fundamental gap in existing LLM-based POI prediction. It reformulates the problem from a single-step mapping to a two-stage reasoning process (Intention Inference $\rightarrow$ Location Determination).
2. IntentPOI Framework: The proposal of a "Thinking-Then-Acting" framework that explicitly infers user intentions using multi-rationale evidence (user profile, peer behaviors, temporal context) and uses this intention as a scaffold for downstream reasoning.
3. Empirical Validation: Extensive experiments on three real-world datasets (NYC, TKY, CA) demonstrate that IntentPOI consistently outperforms eleven state-of-the-art baselines, including deep learning models (e.g., DeepMove, GETNext) and other LLM-based methods (e.g., CoMaPOI, QT-Mob).

Experimental Results

• Performance: IntentPOI achieved the best results across all metrics (HR@5, HR@10, N@5, N@10, MRR) on all three datasets. Notably, on the TKY and CA datasets, it improved HR@5 by 16.66% and 11.71% respectively over the best baseline.
• Ablation Studies:
  • Removing the user profile from the thinking stage caused a massive performance drop (HR@1 decreased by ~49.8%), confirming its role as foundational evidence.
  • Removing the intention signal from the acting stage resulted in an 11.2% drop in HR@1, validating the intention as a critical reasoning scaffold.
  • Removing peer behaviors caused a 9.5% drop, highlighting the value of cross-user evidence, especially for sparse data.
  • The hybrid filtering strategy was shown to be significantly more efficient than random filtering or using only spatial/historical candidates.
• Sparse Data Robustness: The framework showed particular resilience on sparse user trajectories, where peer behaviors played an even more critical role in compensating for limited individual history.
• Case Studies: Analysis of successful cases showed that the multi-rationale approach correctly identified intentions (e.g., "evening leisure outing") that pure frequency-based models missed. Failure cases were attributed to scenarios where the ground truth deviated so significantly from all available signals (profile, peers, history) that the inferred intention could not serve as an effective scaffold.

Significance and Claims

The paper claims that by explicitly decoupling intention inference from location prediction, IntentPOI transforms next POI prediction from frequency-driven trajectory matching into intention-grounded reasoning.

• Interpretability: Unlike "black-box" trajectory matching, IntentPOI provides auditable reasoning traces, explaining why a location was recommended based on specific evidence (profile, peers, time).
• Generalization: The explicit modeling of intention allows the model to adapt to complex urban environments and diverse user behaviors where simple pattern matching fails.
• Paradigm Shift: The work suggests a shift in LLM-based mobility modeling from treating LLMs as direct predictors to using them as reasoning agents that first understand the "why" (intention) before determining the "where" (location).

The authors acknowledge limitations, noting that the framework relies on the quality of LLM-generated intentions and currently focuses on short-term intentions without modeling their long-term evolution. Future work is proposed to explore more efficient models and dynamic intention evolution.