Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..

Generating Origin-Destination Matrices in Neural Spatial Interaction Models

Authors: Ioannis Zachos, Mark Girolami, Theodoros Damoulas

NeurIPS 2024 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically test our framework on both synthetic and real-world data from Cambridge, UK and Washington, DC, USA. We compare GENSIT against SIM-MCMC [13], SIM-NN [17], SIT-MCMC [47] and the Geo-contextual Multitask Embedding Learner (GMEL) [27].
Researcher Affiliation Academia 1Department of Engineering, Cambridge University, Cambridge, CB2 1PZ. 2The Alan Turing Institute, London, NW1 2DB. 3Departments of Statistics & Computer Science, University of Warwick, Coventry, CV4 7AL.
Pseudocode Yes Algorithm 1 : Generating Neural Spatial Interaction Tables. O(NE(τJ + IJ))
Open Source Code Yes Codebase found at https://github.com/Yannis Za/Ge NSIT
Open Datasets Yes We empirically test our framework on both synthetic and real-world data from Cambridge, UK and Washington, DC, USA. In the Cambridge dataset, the ground truth ODM is a 69 13 contingency table with 33, 704 agents. We apply our method to the Washington dataset, where the ground truth ODM is a 179 179 contingency table with 200, 029 agents. Both y, D + have been sourced from the UK s population census dataset provided by the Office of National Statistics. Our codebase and the real-world data we used are accessible from the Supplementary Material.
Dataset Splits Yes In the case of the Washington DC data, we employ the same train/test/validation test split as in [27].
Hardware Specification Yes All experiments were run using a 32-core CPU machine with 128GB memory.
Software Dependencies No The paper mentions software like "Py Torch [26]" and "Adam optimizer [23]" but does not specify their version numbers, which is required for reproducibility.
Experiment Setup Yes The input layer is set to the observed log-destination attractions y RJ... The output layer is two-dimensional due to the parameter vector θ R2. For both datasets we set the number of hidden layers to one and number of nodes to 20. The hidden, output layers have a linear and absolute activation functions, respectively. The NN parameters W are initialised by sampling uniformly over the region [0, 4](J+1) 20+21 2. We use the Adam optimizer [23] with 0.002 learning rate. Bias is initialised uniformly at [0, 4]. an Euler-Maruyama numerical solver ϕHW is employed throughout the paper with a time discretisation step of t = 0.01 and number of steps τ = 1. The low and high SDE noise levels correspond to σ = 0.014 and σ = 0.141, respectively... we set the responsiveness parameter ϵ = 1, and the parameter δ relating to the job availability of a destination where no agents travel to 0... We follow [13, 47] in fixing σd = 0.03 and σT , σΛ to 0.07...