When processing text or metadata, standard Information Retrieval (IR) metrics like Term Frequency-Inverse Document Frequency (TF-IDF) rely on document boundaries to determine relevance. A word is deemed important if it appears frequently within a specific document but rarely across the broader corpus.
However, when the goal shifts to predicting geographic locations—such as geolocating a tweet, an image's tags, or a regional dialect—document boundaries lose their meaning. As Van Laere et al. point out, traditional statistical feature selection methods fail to capture the spatial relevance required for accurate geographic language modeling.
Consider the words "morning" and "Eiffel." Both might appear frequently in a dataset of social media posts. A standard TF-IDF model treats them similarly if their document distributions match. Yet, "morning" is globally ubiquitous, offering zero spatial signal, whereas "Eiffel" is highly concentrated in a single geographic point. Without a mechanism to prioritize geographic concentration over raw frequency, models suffer from high localization errors because common, non-specific tags carry the same weight as unique regional landmarks (Cao et al.).
To solve this, researchers use geospatial term weighting (or spatial entropy weighting): a family of techniques designed to mathematically measure and prioritize features based on their geographic concentration rather than their textual frequency.
Measuring Spatial Entropy and Variance
Instead of measuring how evenly a word is distributed across text files, spatial entropy measures how evenly a word is distributed across physical space.
If a term or topic is distributed uniformly across geographic coordinates, its spatial entropy is high, indicating it is spatial noise. Conversely, a low spatial entropy signifies a tight geographic cluster. As demonstrated by Kordopatis-Zilos et al., calculating the spatial entropy of topics generated through global topic modeling allows systems to identify and filter out non-informative metadata.
To refine this further, Kordopatis-Zilos et al. introduced a mechanism that applies a Gaussian function to filter metadata based on geographic distribution. This technique suppresses both globally ubiquitous terms and extremely rare noise, focusing the model on features with a moderate, highly informative level of geographic specificity. On a continuous scale, Scherrer et al. similarly prioritize linguistic features with high spatial variance to map text along a geographic continuum, ensuring that regional dialect variations drive the localization.
High Spatial Entropy (Noise): [ . . . . . . . . ] -> Down-weighted (e.g., "morning")
Low Spatial Entropy (Signal): [ . . . ] -> Up-weighted (e.g., "Eiffel")
Distance and Ratio-Based Weighting Schemes
Beyond entropy, geographic specificity can be captured using physical distance metrics or explicit frequency ratios across regions.
One intuitive approach incorporates the physical distance between occurrences of a feature. Vidyasagaran et al. implemented a "tag spread measure" that calculates the Euclidean distance between spatial clusters of keywords. By applying a mathematical penalty to terms that appear in geographically distant clusters, the system filters out global noise in favor of locally specific indicators.
Alternatively, models can compute explicit ratios between local and global frequencies. As surveyed by Zheng et al., geospatial term weighting often calculates the ratio of a term's frequency within a specific region to its frequency across all regions. This statistical approach isolates regional dialects and local entities from general language.
This ratio-based logic mirrors Inverse Document Frequency but operates on a spatial scale:
- Pavlopoulos et al. utilized a modified term-weighting scheme to discover regional dialect markers in Greek proverbs by penalizing terms that appear frequently across multiple locations.
- Jin et al. designed an IDF-inspired weighting scheme to penalize information that triggers widespread, non-specific engagement across multiple regions, thereby amplifying the signal of hyper-local occurrences.
Isolating Location-Indicative Words
Why does explicitly weighting or filtering these terms work so well? The core intuition is that the vast majority of human language is spatially neutral. Including every word in a spatial classifier introduces immense noise that degrades precision.
Han et al. demonstrated that focusing on a small subset of "Location Indicative Words" (LIWs) improves geolocation accuracy by over 10% while reducing memory and processing requirements. They utilized the Information Gain Ratio to rank spatial features, which normalizes for word frequency and prevents the model from biasing toward ultra-rare terms that lack statistical power.
While some systems explicitly filter words using these metrics, discriminative models can also learn these weights implicitly. Wing et al. showed that hierarchical discriminative classifiers automatically identify and weight geographically significant terms like regional slang and local landmarks without requiring external geographic databases or gazetteers. Furthermore, Salehi et al. showed that geolocation models drastically improve by selectively using only high-precision entity categories (such as local facilities or sports teams) while discarding noisy categories like celebrities or products that lack specific spatial associations.
Extension to Visual Features
The intuition behind geospatial term weighting extends naturally beyond text into computer vision. In large-scale image retrieval and localization, models face the challenge of "geometric burstiness"—where unrelated locations share highly similar geometric arrangements of visual features (such as repetitive building facades or windows).
To solve this, Sattler et al. introduced a drop-in weighting scheme that weights visual features based on their distribution across distinct geographic locations rather than their raw image frequency.
By using geographic metadata to cluster database images into distinct locations beforehand, a visual feature is only penalized if it matches across multiple different geographic sites. This preserves the importance of unique local landmarks that appear in many photos of the same site, while successfully suppressing generic architectural patterns that appear across unrelated parts of a city.
Key Foundational Work
3D face recognition using sparse spherical representations (2008) — contributed 1 concept elements to this direction.
Inferring Long-term User Properties Based on Users' Location History (2007) — contributed 1 concept elements to this direction.
The UPV at GeoCLEF 2007 (2007) — contributed 1 concept elements to this direction.