[Paper Review] Accurate Local Estimation of Geo-Coordinates for Social Media Posts

Derek Doran, Swapna S. Gokhale+1 more|arXiv (Cornell University)|Oct 17, 2014

Human Mobility and Location-Based Analysis20 references3 citations

TL;DR

This paper proposes a content-based methodology for accurately estimating the geo-coordinates of social media posts within a known broad region, using ensemble language models trained on non-overlapping sub-regions and geo-smoothing to refine location estimates. Evaluated on 82,863 NYC tweets, the approach achieves an average estimation error of just 2.15 km, significantly improving upon prior methods that only locate posts within broad areas.

ABSTRACT

Associating geo-coordinates with the content of social media posts can enhance many existing applications and services and enable a host of new ones. Unfortunately, a majority of social media posts are not tagged with geo-coordinates. Even when location data is available, it may be inaccurate, very broad or sometimes fictitious. Contemporary location estimation approaches based on analyzing the content of these posts can identify only broad areas such as a city, which limits their usefulness. To address these shortcomings, this paper proposes a methodology to narrowly estimate the geo-coordinates of social media posts with high accuracy. The methodology relies solely on the content of these posts and prior knowledge of the wide geographical region from where the posts originate. An ensemble of language models, which are smoothed over non-overlapping sub-regions of a wider region, lie at the heart of the methodology. Experimental evaluation using a corpus of over half a million tweets from New York City shows that the approach, on an average, estimates locations of tweets to within just 2.15km of their actual positions.

Motivation & Objective

• To address the lack of accurate geo-tagging in social media posts, which limits applications in disaster response, advertising, and sociological research.
• To improve upon existing methods that only estimate locations at the city level or within broad regions (e.g., 80–100 km radius).
• To enable fine-grained, local geo-tagging of social media content using only textual content and prior knowledge of the broader origin region.
• To evaluate the effectiveness of geo-smoothing techniques in enhancing location estimation accuracy through sub-region language modeling.
• To demonstrate that high-precision geo-location is feasible when the broader region is already known, even without GPS or explicit location tags.

Proposed method

• The method partitions the known broad region into a g×g grid of non-overlapping sub-regions to capture local linguistic variations in social media content.
• It builds an ensemble of interpolated bigram language models, one per sub-region, trained on tweets originating from that sub-region.
• Geo-smoothing is applied by combining language model probabilities from neighboring sub-regions, weighted by their spatial distance, using a polynomial decay function with parameter d.
• The final location estimate for a tweet is the sub-region with the highest smoothed probability, computed as Pℓi(T) = α × Pℓi(T) + (1−α) × Σ Pℓj(T) × w(d), where w(d) decays with distance d and α=0.9.
• The approach leverages the fact that prior knowledge of the broader region (e.g., Manhattan) limits the search space, enabling high-precision estimation from content alone.
• The method is evaluated using a corpus of over 500,000 tweets from NYC, with test sets of 82,863 tweets used to measure estimation error.

Experimental results

Research questions

• RQ1 Can language models trained on localized sub-regions of a city improve the accuracy of geo-coordinate estimation for social media posts?
• RQ2 To what extent does geo-smoothing—incorporating neighboring sub-regions’ language models—reduce estimation error?
• RQ3 How does the smoothing diameter d affect the distribution and magnitude of estimation errors?
• RQ4 Is the proposed method capable of achieving sub-5 km accuracy in real-world urban settings?
• RQ5 Does increasing the smoothing radius beyond a certain threshold yield diminishing returns in accuracy?

Key findings

• The proposed method achieves an average estimation error of 2.15 km across 82,863 test tweets from New York City, representing just 4% of the total region size.
• Mean estimation error decreases with increasing smoothing diameter d, dropping from 3.38 km (d=1) to 2.15 km (d≥5), with convergence observed at d≥5.
• For d≥3, the error density sharply peaks around 1.75 km, indicating high concentration of accurate estimates near the true location.
• The model estimates geo-coordinates within 4 km of the true location with 80% probability, within 2 km with over 50% probability, and within 1 km with 20% probability at d=3.
• Error distributions become increasingly exponential and monotonic with increasing d, confirming that larger d consistently improves accuracy without special-case degradation.
• The method shows diminishing returns beyond d=2, suggesting that smoothing over neighbors more than two sub-regions away offers minimal additional benefit.