TY - GEN
T1 - Feature Discretization and Selection Methods for Supporting Bridge Deterioration Prediction
AU - Liu, Kaijian
AU - El-Gohary, Nora
N1 - Publisher Copyright:
© 2018 American Society of Civil Engineers (ASCE). All roghts reserved.
PY - 2018
Y1 - 2018
N2 - Bridge maintenance decision making largely depends on the ability to predict future bridge deterioration. The increasing availability of large amounts of multi-source bridge data offers opportunities to bridge data analytics for improved bridge deterioration prediction for supporting enhanced maintenance decision making. Such data include National Bridge Inventory (NBI) and National Bridge Elements (NBE) data, bridge inspection data from bridge inspection reports, bridge health condition data from various sensors, and traffic and weather data. However, bridge data have characteristics that add to the challenges of bridge data analytics: (1) the types of data features are mixed (continuous and discrete); and (2) some of the data features are redundant and/or irrelevant. These characteristics could negatively affect the performance of bridge data analytics. Various feature discretization (FD) and feature selection (FS) methods have been proposed in the literature to deal with mixed-type and redundant/irrelevant features, respectively. However, their performance levels vary across domains and applications. There is, thus, a need to evaluate the impact of using FD and FS methods on the performance of data-driven bridge deterioration prediction. To address this need, this paper focuses on evaluating the impact of FD and FS methods on predicting deterioration using NBI data. The preliminary experimental results show that when predicting the condition ratings using NBI data, compared to the baseline: (1) the ModifiedChi2, compared to the other FD methods, achieved the highest average accuracy improvement of 37.6%; and (2) the combination of the ChiMerge FD method and the Lasso FS method, compared to the other combinations, achieved the highest average accuracy improvement of 37.7%. This shows the promise of using the ModifiedChi2 FD method alone for supporting bridge data analytics.
AB - Bridge maintenance decision making largely depends on the ability to predict future bridge deterioration. The increasing availability of large amounts of multi-source bridge data offers opportunities to bridge data analytics for improved bridge deterioration prediction for supporting enhanced maintenance decision making. Such data include National Bridge Inventory (NBI) and National Bridge Elements (NBE) data, bridge inspection data from bridge inspection reports, bridge health condition data from various sensors, and traffic and weather data. However, bridge data have characteristics that add to the challenges of bridge data analytics: (1) the types of data features are mixed (continuous and discrete); and (2) some of the data features are redundant and/or irrelevant. These characteristics could negatively affect the performance of bridge data analytics. Various feature discretization (FD) and feature selection (FS) methods have been proposed in the literature to deal with mixed-type and redundant/irrelevant features, respectively. However, their performance levels vary across domains and applications. There is, thus, a need to evaluate the impact of using FD and FS methods on the performance of data-driven bridge deterioration prediction. To address this need, this paper focuses on evaluating the impact of FD and FS methods on predicting deterioration using NBI data. The preliminary experimental results show that when predicting the condition ratings using NBI data, compared to the baseline: (1) the ModifiedChi2, compared to the other FD methods, achieved the highest average accuracy improvement of 37.6%; and (2) the combination of the ChiMerge FD method and the Lasso FS method, compared to the other combinations, achieved the highest average accuracy improvement of 37.7%. This shows the promise of using the ModifiedChi2 FD method alone for supporting bridge data analytics.
UR - http://www.scopus.com/inward/record.url?scp=85048946446&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85048946446&partnerID=8YFLogxK
U2 - 10.1061/9780784481295.042
DO - 10.1061/9780784481295.042
M3 - Conference contribution
AN - SCOPUS:85048946446
T3 - Construction Research Congress 2018: Infrastructure and Facility Management - Selected Papers from the Construction Research Congress 2018
SP - 413
EP - 423
BT - Construction Research Congress 2018
A2 - Wang, Chao
A2 - Berryman, Charles
A2 - Harper, Christofer
A2 - Lee, Yongcheol
A2 - Harris, Rebecca
T2 - Construction Research Congress 2018: Infrastructure and Facility Management, CRC 2018
Y2 - 2 April 2018 through 4 April 2018
ER -