時間超參數對交通事故頻率預測及交通事故特徵關聯性之多重時間效應

Multiple Temporal Effects of Temporal Hyperparameters on Traffic Crash Frequency Prediction and Associations between Traffic Crash Characteristics

吳元維(Yuan-Wei Wu)

指導教授 ： 許添本

國立臺灣大學/工學院/土木工程學研究所/博士(2022年)

https://doi.org/10.6342/NTU202201295

摘要

2017至2020年，臺灣交通事故的死亡及受傷人數持續攀升，顯示出居高不下的肇事違規率與成效有限的交通安全政策無法遏止交通安全的惡化，用路人不遵守交通規則，以及與時俱變的交通事故特性間的關聯性，均值得深入探討。透過瞭解交通執法的效果以及交通事故特性間的關聯性，有助於研擬具體有效的政策以促進交通安全。交通安全分析師可探討交通事故特性間的關聯性，並對未來的交通事故頻率進行預測，以發展切合實際的交通安全對策並妥善分配有限資源。
近年來諸多文獻指出，由於分析時間點的不同，會導致交通安全研究結果呈現顯著的時間不穩定性。在此脈絡下，本研究認為特徵期間（Feature period）、反應期間（Response period）、訓練期間（Training period）、資料蒐集尺度（Data collection level），以及分析時間點（Timepoint of analysis）等時間超參數（Temporal hyperparameters）所產生的多重時間效應（Multiple temporal effects），會使得交通安全研究結果不穩定，進而使其無法成為跨越時間的通用知識與觀點。因此，本研究從交通事故頻率預測及交通事故特性間的關聯性等不同角度，引入各種時間超參數，藉此探討多重時間效應的影響。在導入各種時間超參數產生的多重時間效應的實驗架構下，本研究利用隨機森林（Random forest）及梯度推進回歸樹（Gradient boosting regression tree）等以決策樹為基礎的機器學習方法，獲得理想的交通事故頻率預測表現，同時避免了預測變數間的多重共線性問題，並利用部分相依（Partial dependence）及累積局部效應（Accumulated local effects）等事後解釋技術（Post-hoc）分析所建立的預測模型。此外，交通事故特性間的關聯性則透過多重對應分析（Multiple correspondence analysis）及卡方距離分析（Chi-squared distance analysis）等探索式資料分析方法（Exploratory data analysis）進行展示與探討。
實驗結果顯示，多重時間效應明顯影響交通事故頻率預測模型的表現，以及交通事故特性間的關聯性。在交通事故頻率預測的任務中，相較於特徵期間，模型的表現對反應期間更為敏感，較長的訓練期間原則上易產生較為穩定的預測表現，但較短的訓練期間則有機會獲得更優良的預測準確率。部分相依和累積局部效應分析的結果，分別提供了可提高預測模型準確度的特徵期間與反應期間長度，以及評估交通執法活動改善交通安全的效果。在交通故特性關聯性分析的部分，研究結果顯示，多重對應分析和卡方距離分析所發現的交通事故特性關聯性，同樣明顯地受到多重時間效應的影響，並可從研究結果中同時觀察到交通事故特性關聯性短期的週期性變化與長期的趨勢。
本研究主要的貢獻在於探討既有文獻尚未深究的時間超參數及其所產生的多重時間效應，研究成果可協助交通安全分析師，瞭解交通事故頻率預測準確度的限制與改善方案，並可辨識出多重時間效應下關鍵的交通安全議題，據此研擬合適的交通事故頻率預測方法、交通執法專案，以及相應的交通安全改善對策；此外，本研究也幫助交通安全分析師對交通安全研究結果保持著審慎的態度進行解讀，並針對不同時間尺度下發掘的交通安全議題研提有效的策略，以達成各種時間尺度下的交通安全目標、妥善分配有限資源以改善交通安全。

關鍵字

時間超參數 ； 多重時間效應 ； 交通執法 ； 交通事故頻率預測 ； 機器學習 ； 探索式資料分析 ； 交通事故特性

並列摘要

Traffic safety performance in Taiwan deteriorated over the 2017–2020 due to the high at-fault crash rate and weakened safety policy. Thereby, the disregard of traffic regulations by road users and temporally unstable associations between crash characteristics deserve substantial attention. Understanding the effects of traffic enforcement and associations between crash characteristics facilitates the development of policies for improving traffic safety. For the development of appropriate traffic safety countermeasures and effective allocation of limited resources, traffic analysts could investigate associations between traffic crash characteristics as well as predict traffic crash frequencies in the future.
Significant temporal instability in the results of traffic safety research due to different timepoints of analysis has been studied recently. Extending from this context, this dissertation argues that multiple temporal effects derived from the interactions of temporal hyperparameters, including the feature period, response period, training period, data collection level, and timepoints of analysis, entail instability in the results of traffic safety research, such that generalized insights beyond time might not be obtained from traffic safety analyses affected by multiple temporal effects. This dissertation investigated the influences of multiple temporal effects from the perspectives of traffic crash prediction and associations between traffic crash characteristics with the implementation of various temporal hyperparameters. Under an experimental structure with multiple temporal effects derived from various temporal hyperparameters, tree-based machine learning approaches, including random forest and gradient boosting regression tree algorithms, were used to obtain desirable predictive performance and address the issue of multicollinearity among predictors, while post-hoc techniques, including analyses of partial dependence and accumulated local effects, were employed to analyze the tree-based machine learning prediction models. In addition, associations between crash characteristics were demonstrated by exploratory data analysis methods, including multiple correspondence analysis and chi-squared distance analysis.
The results of experiments in this dissertation show that multiple temporal effects substantially affect the predictive performances of the proposed crash frequency prediction models as well as the level of associations between crash characteristics. In crash frequency prediction tasks, predictive performances are more sensitive to the response period than to the feature period. Longer training period leads to more stable performances in general, while shorter training periods provide possibilities to achieve higher prediction accuracies. Results of partial dependence and accumulated local effects analyses suggest appropriate lengths of feature periods and response periods for the improvement of prediction accuracy and also provide evaluations for effects of traffic enforcement activities to improve traffic safety. Regarding the analysis of associations between crash characteristics, results show that multiple temporal effects substantially affect the associations between crash characteristics identified by multiple correspondence analysis and chi-squared distance analysis. Furthermore, both short-term cyclical variations and long-term trends of associations between crash characteristics were observed under multiple temporal effects.
This dissertation contributes to the literature by investigating multiple temporal effects of temporal hyperparameters to which insufficient attention has been paid. The results facilitate the understanding of limitations and solutions to improve crash frequency prediction accuracy and to identified critical traffic safety issues under the consideration of multiple temporal effects. With these insights, traffic analysts and agencies could develop appropriate crash frequency prediction schemes, traffic enforcement programs, and corresponding measures to improve traffic safety. Moreover, attentive interpretations of the results of traffic safety research and the identification of traffic safety issues at various temporal levels could be obtained for the development of effective strategies targeting objectives at various temporal scales and the appropriate allocation of limited resources to improve traffic safety.

並列關鍵字

temporal hyperparameter ； multiple temporal effects ； traffic enforcement ； traffic crash frequency prediction ； machine learning ； exploratory data analysis ； traffic crash characteristic

參考文獻

Abdel-Aty, M., Lee, J., Siddiqui, C., Choi. K., 2013. Geographical Unit Based Analysis in the Context of Transportation Safety Planning. Transportation Research Part A: Policy and Practice 49, 62–75.

Al-Bdairi, N.S.S., Behnood, A., 2021. Assessment of temporal stability in risk factors of crashes at horizontal curves on rural two-lane undivided highways. Journal of Safety Research 76, 205–217.

Al-Bdairi, N.S.S., Behnood, A., Hernandez, S., 2020. Temporal stability of driver injury severities in animal-vehicle collisions: A random parameters with heterogeneity in means (and variances) approach. Analytic Methods in Accident Research 26, 100120.

Ali, Q., Yaseen, M. R., Khan, M. T. I., 2019. The causality of road traffic fatalities with its determinants in upper middle income countries: A continent-wide comparison. Transportation Research Part A: Policy and Practice. 119, 301–312.

Alnawmasi, N., Mannering, F., 2019. A statistical assessment of temporal instability in the factors determining motorcyclist injury severities. Analytic Methods in Accident Research 22, 100090.