Predicting Novel Location Visits in Human Mobility

30 concepts17 papers1 proposal

Standard sequential transition models—such as Markov chains, recurrent neural networks, and transformers—excel at capturing routine human behaviors. If a person travels from home to work and then to a familiar grocery store every Monday, sequence models easily map these recurring trajectories. However, these models face a fundamental limitation: they treat locations as a fixed vocabulary of discrete tokens. Consequently, they are structurally incapable of predicting a transition to a location that has never appeared in a user's historical data.

Human behavior is inherently driven by neophilia—the desire to seek out novel experiences and explore unfamiliar spaces. When an individual breaks from routine to visit a new restaurant or travel to a different city, standard pattern-matching models fail completely. To capture these spontaneous journeys, a distinct research direction has emerged around Exploration Prediction (EP). This paradigm shifts the focus from purely tracking historical routines to modeling when, why, and where individuals will explore novel locations.

The Fundamental Ceiling of Routine Models

When predictive models rely exclusively on an individual's past location history, they hit a strict theoretical upper bound. As demonstrated by Coco Krumme et al., human mobility includes an element of behavioral innovation that imposes a predictable ceiling on accuracy. No historical sequence can fully account for a sudden deviation into an unknown state.

This limitation is not a minor edge case. Licia Amichi et al. (2023) identified novelty-seeking behavior as an inherently limiting factor that degrades next-place forecasting across standard mobility frameworks. The performance gap is particularly severe for individuals with high location entropy, whose movements are less anchored to rigid schedules (Zongyuan Huang et al.). To quantify this vulnerability, Ye Hong et al. (2025) utilized a causal intervention framework to manipulate exploration tendencies in synthetic data. They discovered that drastically increasing an individual's propensity to explore can collapse the accuracy of next-location neural networks to below 10%.

When an agent exhibits outlier behaviors or travels to unvisited geographic areas, traditional intent-prediction and trajectory-reconstruction models break down because their underlying behavioral assumptions are violated (Philipp Sommer et al., Julien Monteil et al., Alessandro Crivellari et al.).

Framing Exploration as a Two-Stage Problem

To bypass this accuracy ceiling, the Exploration Prediction framework decouples the prediction task into two distinct questions:

  1. When will a user choose to explore rather than return to a familiar place?
  2. Where will they go, given that the destination is not in their history?

Xing Xie et al. formalized the first question as a binary classification task. Instead of guessing the exact next coordinate immediately, the model first determines whether the next visit will be routine or exploratory.

                  [Current Mobility State]
                             │
              ┌──────────────┴──────────────┐
              ▼                             ▼
    [Predict: ROUTINE]             [Predict: EXPLORATION]
              │                             │
              ▼                             ▼
    Markov / Sequence Models       Inductive Spatial-Semantic
    (Predict familiar state)       Models (Predict novel state)

This binary choice relies on the realization that exploratory behavior leaves predictable statistical footprints. The authors showed that a user's likelihood of exploring is strongly influenced by contextual features, including the total number of unique places they have visited, their historical visit entropy, and the novelty of their most recent check-in.

Furthermore, these tendencies are tied to stable, long-term behavioral profiles. By modeling mobility using Finite-State Automata, Licia Amichi et al. (2019) categorized individuals into distinct groups, separating extreme returners (routineers) from extreme explorers (scouters).

By understanding a user's baseline profile and tracking their real-time contextual features, hybrid frameworks can dynamically route predictions (Defu Lian et al.). When the model predicts a routine step, it trusts a standard sequential transition model; when it predicts an exploratory step, it switches to a model optimized for novelty.

Generalizing to Unseen Spaces via Inductive Embeddings

Predicting that a user will explore solves only half the problem. The model must still recommend or predict a specific destination from a set of locations the user has never seen.

The key insight that makes this possible is moving away from discrete location identifiers and toward inductive spatial-semantic representations. If a location is represented purely by an arbitrary ID (e.g., Location_742), a model cannot reason about it without prior training examples. However, if that location is represented by its geographic coordinates and its semantic features (e.g., "a sushi restaurant in a high-foot-traffic commercial district"), the model can evaluate its appeal based on generalized preferences.

Xinglei Wang et al. developed this concept through CaLLiPer, a method that generates inductive spatial-semantic location embeddings. Because these embeddings capture universal geographic and Point-of-Interest (POI) attributes, they possess inductive capabilities. This allows the model to generalize to entirely unseen environments and predict movements into new areas without requiring computationally expensive model re-training.

Similarly, Ye Hong et al. (2023) expanded this capability to predict not just the immediate next destination, but the long-term visit frequencies to new places. By using Graph Convolutional Networks (GCNs) and Multi-Head Self-Attention (MHSA), the authors mapped historical mobility patterns to unvisited nodes in a spatial graph, proving that an individual's past behavior can reliably forecast their future engagement with entirely new locations.

The Scaffolding of Exploration: Context and External Gravity

Exploratory behavior is rarely entirely random; it is frequently guided by external context and digital incentives. Traditional recommendation systems often suffer from a bias toward established habits, making them less effective at suggesting novel locations that align with an exploratory mindset (Hossein A. Rahmani et al.). However, when models successfully leverage external signals, the predictability of exploration increases significantly.

Keiichi Ochiai et al. analyzed the "gravity" exerted by Location-Based Services (LBSs) such as location-based games and social media platforms. They discovered that LBS users exhibit highly distinct mobility patterns, yielding up to a 43.9% increase in novel location prediction accuracy compared to non-users. The digital platform effectively acts as a scaffolding, directing the user's neophilia along predictable paths.

When these external contextual signals are combined with group dynamics—such as coordinating a shared venue for individuals scattered across a city (Jason Shuo Zhang et al.)—models can successfully anticipate serendipitous, out-of-cluster visits that lie entirely outside a user's historical routine.

Key Foundational Work

CEPR (2015) — contributed 6 concept elements to this direction.
The predictability of consumer visitation patterns (2013) — contributed 2 concept elements to this direction.
Mining Location-based Social Networks: A Predictive Perspective (2015) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_exploration_prediction_8541e7b6,
  title = {exploration prediction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/exploration-prediction/rdr_8541e7b66a4a4dc7b846c584bca942d8},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_exploration_prediction_8541e7b6_v1,
  title = {exploration prediction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/exploration-prediction/rdr_8541e7b66a4a4dc7b846c584bca942d8/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake