10.30.26 - Corey Harper

10.30.26 - Corey Harper

Title history
  • 10.30.26 - Corey Harper — observed on 20261101-0213 UTC (YouTube)

by Corey Harper · View the channel on PSI

Applied Sciences & Engineering Arts & Architecture AVMusicAVPMusicians, singers, bands and groups
🎙 Corey Harper 👥 116 📅 October 30, 2026 ⏱ 53 min 👁 2 📄 original study 🧭 2026-11-01
Available in: English (current) Français

Keywords

AVEVlatent class analysisSHAPassociation rule mining

Summary

Corey Harper, associate professor at Carnegie Mellon, presents his research on advancing toward a smarter and safer transportation system. He first discusses the potential of connected and automated vehicles (AVs) to reduce crashes caused by human error. Using national road fatality statistics, he shows that fatalities have risen since 2020 and that AVs could help. He then details a published study in the Journal of Safety Research that analyzes level 4 AV crashes in California using NHTSA Standing General Order data. The study employs latent class analysis to identify three distinct crash scenarios: low-speed daytime, moderate-speed, and low-speed nighttime crashes. Machine learning models (XGBoost, CatBoost) with SHAP interpretations reveal factors such as non-motorists, adverse weather, and interactions with heavy vehicles that increase injury risk. He provides policy recommendations like speed limits in high-pedestrian areas and friction-aware controls. Next, he introduces ongoing work on electric vehicle (EV) crashes using PennDOT data, clustering five crash scenarios and applying association rule mining to find factor combinations linked to injuries, such as motorcycle involvement and aggressive driving. The talk emphasizes scenario-dependent effects rather than average relationships, and concludes that further research is needed on time-of-day effects and EV-specific crash dynamics.

196 words

Critical Evaluation

Value of the Information & Strength of the Argument

The presentation provides substantial value by moving beyond aggregate analyses to scenario-specific crash factor identification. The use of latent class analysis to segment crashes into distinct profiles is methodologically sound and yields actionable insights. The argumentation is clear and well-supported with data visualizations and statistical results. The speaker effectively contrasts his approach with prior research, highlighting how average effects can mask important differences. He also discusses practical implications and acknowledges limitations, such as the need to explore time-of-day effects further. The logic flows logically from problem statement to methodology to results to policy recommendations.

Scientific Rigor, Source Quality, Title Accuracy

The talk is technically rigorous, relying on official crash databases (NHTSA Standing General Order, PennDOT). The published study in the Journal of Safety Research indicates peer-review quality. The citation of SAE levels and prior research shows awareness of the field. The title of the video is inadequate as it merely gives a scheduling code, not reflecting the substantive content. The talk would benefit from a descriptive title, but this does not affect the scientific content presented.

185 words

Title / Content Match

The title is a scheduling ID ('10.30.26 - Corey Harper') and does not reflect the content. The talk is a technical presentation on AV and EV crash analysis.

Quality & Reliability

9/10

The talk presents rigorous research based on official NHTSA Standing General Order crash data and PennDOT data, with transparent methodology (latent class analysis, SHAP, association rule mining) and published in a peer-reviewed journal. The speaker is an associate professor at Carnegie Mellon, adding credibility.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

  • National Highway Traffic Safety Administration Standing General Order — Used as primary data source for AV crash analysis in California
  • Journal of Safety Research publication on level 4 AV crashes — Study discussed in detail during the talk, presenting the latent class and SHAP analysis
  • Society of Automotive Engineers (SAE) Levels of Automation — Referenced to explain automation levels, especially level 4 vs 5
  • PennDOT statewide crash data — Data source for the ongoing electric vehicle crash analysis

Concurring Sources

  • NHTSA Standing General Order reports — Official crash reports provided by manufacturers, consistent with the speaker's use.
  • Existing literature on AV crash factors — Earlier studies have examined AV crashes but focused on average effects, aligning with the speaker's point.

Contribution & Novelties

The talk’s original contribution lies in using scenario segmentation (latent class analysis) and explainable AI (SHAP) to uncover heterogeneous crash factors for level 4 autonomous vehicles, rather than averaging across all crashes. This approach reveals that injury risks differ substantially by crash scenario, enabling targeted policy interventions. The ongoing EV study similarly applies latent class analysis and association rule mining to identify specific factor combinations, moving beyond average effects.

Pour aller plus loin :

116 words

Radar Profile

The radar profile shows consistently high scores across all dimensions, indicating a well-balanced and technically strong presentation. The highest score is in level of technical detail and information quantity, reflecting the methodological depth. The lowest (but still high) score is in global reliability, likely due to reliance on self-reports from manufacturers and the ongoing nature of EV research.

Reliability 9/10