Multi-Scale Representation Learning for Spatial Feature Distributions using Grid Cells
Authors: Gengchen Mai, Krzysztof Janowicz, Bo Yan, Rui Zhu, Ling Cai, Ni Lao
ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 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 multiscale 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. |
| Researcher Affiliation | Collaboration | Gengchen Mai1, Krzysztof Janowicz1, Bo Yan2, Rui Zhu1, Ling Cai1 & Ni Lao3 1STKO Lab, University of California, Santa Barbara, CA, USA, 93106 {gengchen_mai,janowicz,ruizhu,lingcai}@ucsb.edu 2Linked In Corporation, Mountain View, CA, USA, 94043 boyan1@linkedin.com 3Say Mosaic Inc., Palo Alto, CA, USA, 94303 ni.lao@mosaix.ai |
| Pseudocode | No | No explicit pseudocode or algorithm blocks are provided in the paper. |
| Open Source Code | Yes | 1Link to project repository: https://github.com/gengchenmai/space2vec |
| Open Datasets | Yes | We utilize the open-source dataset published by Yelp Data Challenge and select all POIs within the Las Vegas downtown area. ... We use two versions of our point space encoder Encpxqpq model (grid, theory) as the location encoder to capture the spatial prior information Ppy | xq. ... Bird Snap:, NABirds:. |
| Dataset Splits | Yes | The POIs are split into training, validation, and test dataset with ratios 80%:10%:10%. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running the experiments. |
| Software Dependencies | No | The paper does not specify software dependencies with version numbers for their own implementation. |
| Experiment Setup | Yes | The hyper-parameters of theory models are based on grid search with dpvq p32, 64, 128, 256q, dpxq p32, 64, 128, 256q, S p4, 8, 16, 32, 64, 128q, and λmin p1, 5, 10, 50, 100, 200, 500, 1kq while λmax 40k is decided based on the total size of the study area. We find out the best performances of different grid cell based models are obtained when dpvq 64, dpxq 64, S 64, and λmin 50. ... The number of layers f and the number of hidden state neurons u of the FFN are selected from f p1, 2, 3q and u p128, 256, 512q. We find out f 1 and u 512 give the best performance for direct, tile, rbf, and theory. |