[Paper Review] Multi-Scale Representation Learning for Spatial Feature Distributions using Grid Cells

Gengchen Mai, Krzysztof Janowicz+4 more|arXiv (Cornell University)|Feb 15, 2020

Advanced Image and Video Retrieval Techniques34 references35 citations

TL;DR

Space2Vec proposes a multi-scale grid-cell–inspired encoder to jointly represent absolute positions and spatial contexts for geographic points, improving POI-type prediction and geo-located image classification over single-scale methods.

ABSTRACT

Unsupervised text encoding models have recently fueled substantial progress in NLP. The key idea is to use neural networks to convert words in texts to vector space representations based on word positions in a sentence and their contexts, which are suitable for end-to-end training of downstream tasks. We see a strikingly similar situation in spatial analysis, which focuses on incorporating both absolute positions and spatial contexts of geographic objects such as POIs into models. A general-purpose representation model for space is valuable for a multitude of tasks. However, no such general model exists to date beyond simply applying discretization or feed-forward nets to coordinates, and little effort has been put into jointly modeling distributions with vastly different characteristics, which commonly emerges from GIS data. Meanwhile, Nobel Prize-winning Neuroscience research shows that grid cells in mammals provide a multi-scale periodic representation that functions as a metric for location encoding and is critical for recognizing places and for path-integration. Therefore, we propose a representation learning model called Space2Vec to encode the absolute positions and spatial relationships of places. We conduct experiments on two real-world geographic data for two different tasks: 1) predicting types of POIs given their positions and context, 2) image classification leveraging their geo-locations. Results show that because of its multi-scale representations, Space2Vec outperforms well-established ML approaches such as RBF kernels, multi-layer feed-forward nets, and tile embedding approaches for location modeling and image classification tasks. Detailed analysis shows that all baselines can at most well handle distribution at one scale but show poor performances in other scales. In contrast, Space2Vec's multi-scale representation can handle distributions at different scales.

Motivation & Objective

• Motivate the need for a general, multi-scale spatial representation that can handle heterogeneous geographic distributions.
• Propose Space2Vec, an encoder–decoder framework using multi-scale sinusoidal position encodings inspired by grid cells.
• Enable unsupervised training to learn distributed representations of point features in space.
• Demonstrate that Space2Vec outperforms RBF, tile, and plain coordinate encoders on POI-type classification and geo-located image tasks.
• Provide qualitative insights into how multi-scale encodings capture spatial structure across scales.

Proposed method

• Encode absolute positions using multi-scale sinusoidal encodings concatenated across 64 scales (Space2Vec theory and grid-inspired encodings).
• Use a two-branch encoder: Enc^(x) for locations and Enc^(v) for point features; concatenate to form e[v] and e[x].
• Two decoders: Dec_s to reconstruct point features from location embedding, and Dec_c to reconstruct a center point's features from the embeddings of nearby context points via a multi-head attention mechanism.
• In Dec_c, compute attention over neighboring points with distance/direction conditioning through displacement encodings and a permutation-invariant aggregation (akin to PointNet).
• Train unsupervised by maximizing the log-likelihood of predicting the true center point’s feature embedding among candidate points, with optional negative sampling.

Experimental results

Research questions

• RQ1 Can a multi-scale, grid-cell–inspired encoding capture both clustered and uniform spatial distributions in GIS data?
• RQ2 Does Space2Vec improve location-aware POI type prediction and geo-located image classification compared to traditional encodings (RBF, tile, wrap) and direct coordinate inputs?
• RQ3 How do multi-scale encodings affect model behavior across POI distribution groups (clustered, middle, even) and spatial scales?
• RQ4 What qualitative patterns emerge in the learned location embeddings and context interactions across scales?

Key findings

• Space2Vec with multi-scale grid-like encodings outperforms baseline encoders (RBF, tile, wrap, and direct coordinate inputs) on POI-type prediction and image classification tasks.
• Single-scale encoders struggle to handle distributions that vary in scale, while Space2Vec effectively integrates information across multiple scales.
• Spatial context modeling with Space2Vec’s grid-based attention improves prediction when context points are used, surpassing specialized non-grid baselines in the test set.
• Qualitative analyses show that Space2Vec learns firing-pattern–like representations that capture spatial structure at different scales and reflect decreasing distance effects via multi-scale encoding.
• The approach leverages an encoder–decoder setup where the location encoder and multi-head attention-based context decoder jointly model absolute positions and spatial relations.