Uncertainty Propagation in Long-Term Structured Regression on Evolving Networks
Authors: Djordje Gligorijevic, Jelena Stojanovic, Zoran Obradovic
AAAI 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The obtained empirical results and use case analysis provide evidence that the new approach allows better uncertainty propagation as compared to published alternatives. To evaluate the quality of the proposed method, we compare to the several benchmarks from the group of unstructured iterative models: iterative Linear Regression (ILR) (Smith 2013) and iterative Gaussian Processes (IGP) (Girard et al. 2003; Candela et al. 2003). Results show evidence of benefits of using structure information for prediction and uncertainty propagation estimation in multiple steps ahead setup. Set-up of the experiments conducted on two real-world datasets from the medical and climate domains, and results in terms of predictive error (Mean Squared Error MSE) and plots of predictions with propagating uncertainty are reported in this section. |
| Researcher Affiliation | Academia | Djordje Gligorijevic, Jelena Stojanovic and Zoran Obradovic Center for Data Analytics and Biomedical Informatics, Temple University, Philadelphia, PA 19122 USA {gligorijevic, jelena.stojanovic, zoran.obradovic}@temple.edu |
| Pseudocode | Yes | Algorithm 1 Multiple steps ahead GCRF regression Input: Test data X, model (αk, βl, θk, ψl) 1. Initialize ΣX for each node in a graph with all zeroes 2. Make a one step ahead prediction of ˆy T +1 for k = 2...K do 3. Update inputs according to the previous predictions ˆy T +k 1 4. Update {ΣX } for the previously introduced noisy input using Eq. (13) 5. Predict following time step ˆy T +k using Eq. 3 and Eq. 9 end for |
| Open Source Code | Yes | The supplement materials with additional experiments and theoretical derivations, as well as the Matlab implementation are available for download at the authors websites. |
| Open Datasets | Yes | We used the State Inpatient Databases (SID)1 California database provided by the Agency for Healthcare Research and Quality and is included in the Healthcare Cost and Utilization Project (HCUP) . This dataset contains 35,844,800 inpatient discharge records over 9 years (from January 2003 to December 2011) collected from 474 hospitals. HCUP State Inpatient Databases (SID). Healthcare Cost and Utilization Project (HCUP). 2005-2009. Agency for Healthcare Research and Quality, Rockville, MD. www.hcupus.ahrq.gov/sidoverview.jsp. A dataset of precipitation records from meteorological stations across the USA has been acquired from NOAA s National Climate Data Center (NCDC) (Menne, Williams, and Vose 2009). |
| Dataset Splits | No | The paper specifies training and testing periods (e.g., "trained our models on 36 monthly snapshots and iteratively predicted for the following 48 months" and "trained on 48 months of data and tested on the following 96 months") but does not explicitly mention a validation set or a specific validation split methodology. |
| Hardware Specification | No | The paper does not provide specific hardware details such as CPU/GPU models, memory amounts, or types of computing resources used for running its experiments. |
| Software Dependencies | No | The paper mentions "Matlab implementation" but does not provide specific version numbers for Matlab or any other software libraries or dependencies used to replicate the experiment. |
| Experiment Setup | No | The paper describes the data splits for training and prediction horizons (e.g., "trained our models on 36 monthly snapshots and iteratively predicted for the following 48 months"; "models were learned using 12 previous values of precipitation variables as input attributes"), but it does not specify concrete hyperparameter values (e.g., learning rate, batch size, epochs) or other system-level training settings. |