CausalPOI: Causal Framework for POI Forecasting

• The paper introduces a novel spatio-temporal graph-based causal representation framework that predicts check-ins for newly added POIs under cold-start conditions.
• It employs structurally aligned treatment and control graphs using semantic embeddings from BERT and spatial decay to capture functional and urban interactions.
• Empirical results show significant reductions in RMSE and MAE over baselines, offering interpretable, intervention-aware forecasts for urban planning.

CausalPOI is a spatio-temporal graph-based causal representation learning framework for forecasting check-ins at newly introduced Points of Interest under cold-start conditions and quantifying intervention effects. It explicitly models both the factual scenario, in which the POI is introduced, and the counterfactual scenario, in which the same urban context is considered without the POI’s semantic influence, thereby enabling interpretable, intervention-aware forecasts at POI-level granularity. The framework was introduced together with a new problem setting—cold-start POI check-in forecasting—in which a new POI has no pre-introduction behavioral history and only its semantic attributes and spatial context are available (Zhang et al., 3 Jun 2026).

1. Problem definition and causal estimand

CausalPOI formalizes a POI $p$ by its geographic coordinates $p.g=(lat_p, lon_p)$, textual tags $p.t=(t_1,t_2,\ldots,t_i)$, and check-in sequence $p.c=\{c_1,c_2,\ldots,c_W\}$, where each $c=(w,n)$ denotes $n$ visits in week $w$. In the cold-start setting, the target is a newly introduced POI with no historical check-ins. The forecasting objective is to predict its future weekly check-in sequence over the next $W$ weeks, $\hat p.c=\{\hat c_1,\ldots,\hat c_W\}$, while training and evaluation use the ground-truth future sequence $p.c=\{c_1,\ldots,c_W\}$ as supervision (Zhang et al., 3 Jun 2026).

The causal component is defined through a networked potential outcome framework with localized interference. For each target POI, the method derives two predicted weekly sequences: $p.g=(lat_p, lon_p)$0, corresponding to the factual scenario in which the POI is introduced, and $p.g=(lat_p, lon_p)$1, corresponding to the counterfactual scenario without the POI’s semantic influence. Weekly uplift is computed as $p.g=(lat_p, lon_p)$2, and the framework interprets this as an estimate of the individual treatment effect of introducing the POI. The intervention context is explicitly urban: adding a new POI may alter local competitive and complementary interactions within an existing spatial-functional environment.

The motivation for this formulation is that correlation-driven models trained on observational data do not directly answer the counterfactual question “what if the POI had not been introduced?”, and proximity-only graphs miss functional dependencies such as competition and complementarity. CausalPOI therefore treats POI introduction as a localized urban intervention rather than only a forecasting target. At the same time, the paper states that conditioning on rich pre-treatment covariates mitigates endogeneity but cannot fully remove unobserved confounding; consequently, the estimated effects are localized counterfactual estimates, not randomized causal effects.

2. Spatio-temporal functional interaction graph

The core structural object in CausalPOI is the Spatio-Temporal Functional Interaction Graph (ST-FIG). For each newly added POI $p.g=(lat_p, lon_p)$3 and each week $p.g=(lat_p, lon_p)$4 in the prediction horizon, the framework constructs a weekly 1-hop subgraph $p.g=(lat_p, lon_p)$5 that includes only neighbors active in week $p.g=(lat_p, lon_p)$6. The node set and neighborhood representation therefore evolve over time, matching the dynamic urban context (Zhang et al., 3 Jun 2026).

Neighborhood membership is determined by a Haversine distance threshold: $p.g=(lat_p, lon_p)$7 with $p.g=(lat_p, lon_p)$8. Each node, including the target POI and its neighbors, is initialized with a textual embedding derived from POI tags using BERT: $p.g=(lat_p, lon_p)$9

Edges are defined from each neighbor $p.t=(t_1,t_2,\ldots,t_i)$0 to the target $p.t=(t_1,t_2,\ldots,t_i)$1, while message passing during encoding is bidirectional. The edge weight combines functional interaction and spatial decay: $p.t=(t_1,t_2,\ldots,t_i)$2 where $p.t=(t_1,t_2,\ldots,t_i)$3 is a bandwidth parameter. Functional interaction is learned by contrastive pretraining over category descriptions using BERT and an InfoNCE loss. Competitive POIs are treated as positives, and complementary POIs as negatives. The paper defines functional dissimilarity as

$p.t=(t_1,t_2,\ldots,t_i)$4

and gives the functional influence weight, normalized to $p.t=(t_1,t_2,\ldots,t_i)$5, as

$p.t=(t_1,t_2,\ldots,t_i)$6

The paper further notes that the intended normalization is $p.t=(t_1,t_2,\ldots,t_i)$7.

A central design choice is the use of structurally aligned treatment and control graphs. The treatment graph $p.t=(t_1,t_2,\ldots,t_i)$8 is the 1-hop subgraph centered at $p.t=(t_1,t_2,\ldots,t_i)$9, with semantic embeddings and functional-weighted spatial decay. The control graph $p.c=\{c_1,c_2,\ldots,c_W\}$0 keeps the same topology but zero-masks the target POI’s embedding and removes functional interaction from the edge weights, using only spatial decay. The stated rationale is structural alignment: keeping topology fixed avoids conflating changes due to node removal with semantic effects. The adjacency weights are locally normalized so that $p.c=\{c_1,c_2,\ldots,c_W\}$1, and $p.c=\{c_1,c_2,\ldots,c_W\}$2 is used as attention bias in the GATv2 encoder.

3. Representation learning architecture and objective function

CausalPOI uses a shared graph-temporal backbone across treatment and control branches. For each week $p.c=\{c_1,c_2,\ldots,c_W\}$3 and treatment indicator $p.c=\{c_1,c_2,\ldots,c_W\}$4, a shared GATv2 encoder processes the weekly graph: $p.c=\{c_1,c_2,\ldots,c_W\}$5 where $p.c=\{c_1,c_2,\ldots,c_W\}$6 is the base embedding dimension and $p.c=\{c_1,c_2,\ldots,c_W\}$7 is the number of neighbors. Edge weights act as attributes influencing attention (Zhang et al., 3 Jun 2026).

From the encoded graph, the framework extracts the target POI embedding and the mean neighbor embedding for each week. It augments these with a sinusoidal geographic positional encoding over latitude and longitude, using positional dimension $p.c=\{c_1,c_2,\ldots,c_W\}$8 and scaling factor $p.c=\{c_1,c_2,\ldots,c_W\}$9. The target, neighborhood, and positional embeddings are concatenated and passed through LayerNorm to form the weekly representation $c=(w,n)$0. A temporal GRU then aggregates the sequence $c=(w,n)$1 into a treatment-specific temporal embedding $c=(w,n)$2.

Outcome prediction is handled by treatment-aware heads: $c=(w,n)$3 where $c=(w,n)$4 and $c=(w,n)$5 are parameter-shared decoders with treatment-specific output heads. This yields two weekly forecasts, one for the factual branch and one for the counterfactual branch.

The causal regularization mechanism includes a propensity model. A pre-treatment embedding is constructed by concatenating a zero embedding for the target node, the neighborhood embedding, and the positional encoding, followed by LayerNorm. A sigmoid projection then produces the propensity estimate $c=(w,n)$6. Training minimizes a combined objective $c=(w,n)$7 consisting of squared temporal prediction error on the observed branch and a CrossEntropy term for propensity regularization, weighted by $c=(w,n)$8. The paper reports that removing propensity regularization by setting $c=(w,n)$9 yields small but consistent degradation, suggesting a stabilizing role in optimization.

4. Training pipeline, inference procedure, and computational profile

The algorithmic workflow begins with aggregation of SafeGraph POI metadata and weekly check-ins. Newly added POIs are identified as those emerging within the observation window and persisting for at least 40 weeks. For each target POI, the experimental focus sets the prediction horizon to $n$0 weeks after introduction (Zhang et al., 3 Jun 2026).

For each target-week pair, the framework builds the 1-hop ST-FIG by selecting active neighbors within $n$1, computing node text embeddings with BERT, deriving positional encodings from coordinates, and constructing aligned weekly treatment and control graphs. In treatment, edge weights use functional-weighted spatial decay; in control, the target semantics are zero-masked and edge weights remove functional interaction. The locally normalized weights are passed to the shared GATv2 encoder.

The training loop encodes each weekly graph, extracts the target and neighbor-mean embeddings, concatenates positional encoding, applies LayerNorm to obtain $n$2, aggregates over time with a GRU, and decodes the two weekly sequences $n$3 and $n$4. The observed-branch prediction is formed using the observed treatment $n$5, and the model is trained by minimizing $n$6.

At inference time for a new cold-start POI, the method builds ST-FIGs for the $n$7 future weeks using active neighbors and semantic information. It then runs both treatment and control branches, producing the forecast $n$8 together with the counterfactual sequence $n$9. Weekly uplift is the difference $w$0.

The computational profile is explicitly localized. Because each weekly graph is 1-hop, the graph is small; a single GATv2 pass has cost $w$1 with $w$2, while the temporal GRU scales as $w$3. The paper characterizes the approach as efficient for localized neighborhoods.

5. Empirical evaluation and quantitative findings

Experiments use SafeGraph POIs and weekly check-ins from September 2018 to January 2020, covering 74 weeks and four U.S. macro regions: Northeast, Midwest, South, and West. The reported numbers of new POIs are 16,959, 19,116, 31,894, and 21,611, respectively. Features include POI tags and categories encoded with BERT, geographic coordinates encoded positionally, Haversine distances, contrastive-pretrained category embeddings for functional interactions, and dynamically active neighbor sets per week (Zhang et al., 3 Jun 2026).

The baselines span multiple methodological families. Time-series GNN-based baselines are DCRNN, GraphWaveNet, AGCRN, and LightST. TrafficStream represents node-update-based methods. SVGP and LCFM are statistical baselines. TIME-LLM and TimeCMA are LLM-based baselines. GCIM is the causal representation baseline, and KGDiff is the generative synthesis baseline. For baseline methods that require temporal inputs, cold-start handling uses a proxy sequence from neighbor-averaged check-ins scaled by a training-calibrated attenuation factor $w$4 to avoid overestimation of early-stage activity.

Forecasting performance is evaluated by RMSE and MAE over the four-week horizon. The paper also includes causal sanity checks based on uplift behavior in high- versus low-competition groups, together with placebo tests using pseudo interventions and shifted introduction weeks.

Region	RMSE	MAE
Northeast	5.36 ($w$5)	1.84 ($w$6)
Midwest	5.87 ($w$7)	2.00 ($w$8)
South	5.88 ($w$9)	1.98 ($W$0)
West	6.17 ($W$1)	1.95 ($W$2)

Across all regions, CausalPOI attains the lowest errors. Relative to the strongest baseline, KGDiff, the reported reductions reach up to 57.8% in RMSE and 34.3% in MAE in the most challenging region. Ablation studies show that removing ST-FIG or removing the causal module increases RMSE and MAE consistently, which the paper interprets as confirming the importance of both functional interaction modeling and counterfactual estimation. Sensitivity analysis over $W$3, $W$4, and $W$5 shows typical trade-offs, with the chosen configuration balancing causal regularization, spatial decay, and neighborhood extent.

The uplift sanity checks align with the framework’s causal interpretation. High-competition neighborhoods exhibit lower relative uplift than low-competition neighborhoods. In placebo tests, uplift is largest under real interventions—for example, Northeast $W$6—and near zero under pseudo interventions—for example, Northeast $W$7. Shifting introduction timing reduces uplift—for example, Northeast $W$8—which the paper presents as evidence of intervention specificity.

6. Interpretability, use cases, limitations, and reproducibility

CausalPOI is designed to produce interpretable causal insights in addition to forecasts. Its uplift estimates contextualize the incremental effect of opening a POI given nearby competition and complementarity. The paper gives the example that new gyms in dense competitive clusters show smaller uplift, whereas gyms near complementary businesses such as smoothie bars show higher uplift (Zhang et al., 3 Jun 2026).

The practical uses identified in the paper are urban intervention analysis, site selection, co-location strategy, and early-stage operations. For site selection, the recommendation is to prioritize locations with favorable functional neighborhoods. For co-location strategy, the framework supports enhancing complementary mixes, such as cafés near coworking spaces. For early-stage operations, it can inform staffing and promotion adjustments based on an expected four-week uplift trajectory. A plausible implication is that the model is intended not only for demand prediction but also for intervention-aware urban decision support.

The limitations are explicit. Observational data cannot fully eliminate unobserved confounding. The localized interference assumption may underrepresent broader spillovers. Functional interactions are derived from category semantics and may therefore miss nuanced relationships. The paper also states that the method is portable to other cities and POI domains with tag or category metadata and weekly activity logs, although retraining the contrastive encoders and ST-FIG parameters may be needed to reflect local semantics.

Reproducibility details are concrete. The implementation uses bert-base-uncased with hidden size 768 for tags and category descriptions, together with PyTorch for GATv2 and GRU components and HuggingFace Transformers for BERT. Reported hyperparameters are learning rate $W$9, batch size $\hat p.c=\{\hat c_1,\ldots,\hat c_W\}$0, embedding dimension $\hat p.c=\{\hat c_1,\ldots,\hat c_W\}$1, positional scaling $\hat p.c=\{\hat c_1,\ldots,\hat c_W\}$2, spatial decay $\hat p.c=\{\hat c_1,\ldots,\hat c_W\}$3, $\hat p.c=\{\hat c_1,\ldots,\hat c_W\}$4 m, causal weight $\hat p.c=\{\hat c_1,\ldots,\hat c_W\}$5, and $\hat p.c=\{\hat c_1,\ldots,\hat c_W\}$6. Training is conducted on a V100-SXM2 GPU until convergence. Source code is provided at https://github.com/ZZQ-NTU/CausalPOI, and the data source is SafeGraph POI and weekly patterns with region splits following U.S. Census macro regions.