Technical Summary: CausalPOI

Problem Definition

The paper introduces a novel research problem termed Cold-Start POI Check-in Forecasting. Unlike traditional urban forecasting tasks that rely on historical data or aggregate regional flows, this problem focuses on predicting the future check-in patterns of a newly introduced Point of Interest (POI) that possesses no prior behavioral observations.

The core challenge lies in modeling the temporal evolution and functional interactions of a new POI within a structured urban spatial context. The objective is twofold:

1. Forecasting: Predict the future weekly check-in sequence ($\hat{p}.c$) for the new POI over a horizon of $W$ weeks.
2. Causal Estimation: Estimate the Individual Treatment Effect (ITE) of introducing the new POI. This involves comparing the factual outcome (with the POI present) against a counterfactual outcome (where the POI is absent) to understand the causal impact of the intervention on the local urban ecosystem.

Existing methods often fail here because they rely on proximity-based graphs that ignore semantic functional dependencies (e.g., competition vs. complementarity) and lack mechanisms to simulate counterfactual scenarios in structured spatial domains.

Methodology: CausalPOI

The authors propose CausalPOI, a spatio-temporal graph-based causal representation learning framework. The architecture consists of two primary modules designed to handle the cold-start nature of the problem and the need for causal inference.

1. Spatio-Temporal Functional Interaction Graph (ST-FIG) Module

This module constructs a dynamic graph to capture the evolving relationships between the target POI and its neighbors.

• Graph Construction: For a new POI $p$, a 1-hop neighborhood $N$ is defined based on geographic proximity (Haversine distance $\le$ 1000m).
• Node Representation: Nodes are initialized with textual embeddings derived from POI tags using a pre-trained BERT encoder.
• Edge Weighting: Edges are weighted by a combination of spatial decay and functional interaction strength ($\alpha(p,n)$).
  • Functional strength is learned via contrastive pretraining (InfoNCE loss). Functionally competitive POIs (similar categories) are treated as positive pairs, while complementary POIs are negative pairs. This allows the model to distinguish between competitive suppression and complementary boosting effects.
  • The edge weight formula combines spatial distance decay with the learned functional similarity score.
• Temporal Dynamics: The graph is constructed for each week, updating the node set to include only active POIs, thereby capturing time-varying neighborhood contexts.

2. Causal Inference Module

This module explicitly models the counterfactual scenario to estimate the ITE.

• Treatment and Control Graphs: For each new POI, the framework constructs a pair of structurally aligned graphs:
  • Treatment Graph ($G^{(1)}$): The standard ST-FIG with the target POI's semantic representation active.
  • Control Graph ($G^{(0)}$): Shares the exact same topology but replaces the target POI's embedding with a zero vector. This simulates the absence of the POI's functional influence while preserving the spatial structure, avoiding structural mismatch issues associated with physically removing nodes.
• Feature Extraction: A shared GATv2 encoder processes both graphs to generate node embeddings, incorporating sinusoidal positional encodings for geographic coordinates. A GRU then models the temporal evolution of these embeddings over $W$ weeks.
• Prediction Head: The model employs treatment-aware decoders to generate two parallel predictions: $\hat{y}_1$ (factual) and $\hat{y}_0$ (counterfactual). The difference between these outputs represents the estimated ITE.
• Training Objective: The model is trained using a temporally extended loss function that combines:
  • Prediction Loss: Mean Squared Error (MSE) between predicted and actual check-in volumes (for the factual scenario).
  • Propensity Loss: Cross-entropy loss to estimate the treatment propensity score based on pre-treatment covariates, ensuring unbiased estimation by balancing treatment and control representations.

Key Contributions

The paper outlines three primary contributions:

1. Problem Definition: The formalization of Cold-Start POI Check-in Forecasting, addressing the gap in predicting activity for new POIs without historical data, a critical need for urban planning and commercial decision-making.
2. Framework Proposal: The introduction of CausalPOI, the first framework to systematically incorporate causal estimation at the POI level. It uniquely integrates:
   • A Spatio-Temporal Functional Interaction Graph that uncovers latent competitive and complementary relationships beyond mere distance.
   • A Causal Inference Module that simulates factual and counterfactual outcomes via structurally aligned graphs to estimate ITE.
3. Empirical Validation: Extensive experiments on real-world SafeGraph datasets (covering four US regions) demonstrating that CausalPOI significantly outperforms state-of-the-art baselines, including time-series GNNs, statistical models, LLM-based approaches, and other causal models.

Experimental Results

The authors evaluated CausalPOI on four US regions (Northeast, Midwest, South, West) using data from September 2018 to January 2020.

• Performance: CausalPOI achieved the lowest Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE) across all regions.
  • On the most challenging region, it achieved up to a 57.8% reduction in RMSE and a 34.3% reduction in MAE compared to the best-performing baseline.
• Ablation Studies: Removing either the ST-FIG module or the Causal Module resulted in consistent performance degradation, confirming that both functional semantic modeling and counterfactual reasoning are essential. The propensity loss was also shown to provide a stabilizing effect.
• Validity Checks:
  • Uplift Sanity Check: The model correctly predicted that POIs in highly competitive environments exhibit smaller relative uplifts compared to those in less competitive settings, aligning with urban economic intuition.
  • Placebo Tests: When interventions were randomly assigned to existing POIs or shifted in time, the estimated uplift dropped significantly (near-zero for random assignments), confirming that the model learns specific causal effects tied to actual new POI introductions rather than spurious correlations.

Significance and Claims

The paper claims that CausalPOI provides a more interpretable and actionable foundation for urban intervention analysis. By moving beyond correlation-driven modeling to causal reasoning, the framework allows stakeholders (e.g., business operators, urban planners) to:

• Accurately forecast demand for new facilities before they open.
• Understand the causal impact of a new POI on surrounding businesses (e.g., whether a new gym will cannibalize or boost traffic to nearby smoothie bars).
• Make data-driven decisions regarding site selection and resource allocation in dynamic urban environments.

The authors emphasize that while the framework mitigates endogeneity by conditioning on rich pre-treatment contexts, the estimated causal effects are interpreted as localized counterfactual estimates rather than fully randomized intervention effects, acknowledging the observational nature of the data.