About

James E. Corter is a Professor of Statistics and Education in the Department of Human Development at Teachers College, Columbia University. He is affiliated with the Center for Decision Sciences at Columbia Business School, the Applied Statistics Center at Columbia University, and the Department of Statistics. Professor Corter earned his Ph.D. in Experimental Psychology from Stanford University in 1983, following graduate study at the L. L. Thurstone Psychometric Laboratory at the University of North Carolina and a B.A. in Psychology with highest honors from the University of North Carolina in 1977. His scholarly interests focus on computational models of human learning and categorization, judgment and decision-making, cognitively diagnostic testing models and methods, statistics expertise and probability problem-solving, as well as clustering and scaling methods for multivariate data. He has made significant contributions to the development of algorithms and software for fitting additive and extended trees, which are used to model similarity and association in data. His work includes the creation of programs such as ADDTREE/P, GTREE, and EXTREE, which implement efficient metric combinatorial algorithms for fitting these tree models, enabling the analysis of complex data structures. Professor Corter's research integrates psychological theory with statistical methodology, advancing understanding in areas such as category learning, probability problem solving, and the use of diagrams in cognitive processes.

Research topics

• Artificial Intelligence
• Computer Science
• Machine Learning
• Data science
• Data Mining
• Political Science
• Epistemology
• Programming language
• Philosophy
• Knowledge management
• Mathematics education
• Psychology
• Chemistry
• Database

Selected publications

• Graph RAG for Automated Short Answer Grading with Feedback: Bridging Pedagogical Needs and Technical Capabilities

  Underline Science Inc. · 2026-01-07

  otherOpen access

  Automated short answer grading with feedback (ASAG-F) systems currently face challenges in transparency, pedagogical alignment, and cost-effectiveness that limit their real-world deployment. We introduce GraphRAG, a knowledge graph-based retrieval-augmented generation framework that addresses these limitations by grounding all large language model (LLM)-generated feedback and scores in instructor-curated atomic facts, ensuring traceability and verifiability. Using the Short Answer Feedback (SAF) dataset with 31 topics, we evaluate GraphRAG on unseen-question and unseen-answer splits. Our systematic evaluation demonstrates that GraphRAG achieves grading accuracy comparable to vector-based RAG and generally superior to a fine-tuned LLM baseline model while providing more transparent source attribution. Additional findings include: (1) Instructing the LLM to discretize continuous scores to match pedagogical rubrics, such as the 0.25 increments common in SAF, improves grading accuracy; (2) LLM-generated feedback exhibits length-dependent quality variations when unconstrained; prompt-based length control substantially enhances feedback quality and its stability, achieving optimal balance of instructional richness and conciseness; (3) Performance scaling analysis reveals that basic models like GPT-4o-mini offer cost-effective performance, while premium models like Claude-Opus-4 show diminishing returns. These results demonstrate that GraphRAG offers a robust, explainable, pedagogy-aligned, and cost-effective solution for large-scale educational applications, enabling transparent automated grading with effective pedagogical feedback and practical deployment costs.

  PublisherDOI

• Graph RAG for Automated Short Answer Grading with Feedback: Bridging Pedagogical Needs and Technical Capabilities

  Proceedings of the AAAI Conference on Artificial Intelligence · 2026-03-14

  articleOpen accessSenior author

  Automated short answer grading with feedback (ASAG-F) systems currently face challenges in transparency, pedagogical alignment, and cost-effectiveness that limit their real-world deployment. We introduce GraphRAG, a knowledge graph-based retrieval-augmented generation framework that addresses these limitations by grounding all large language model (LLM)-generated feedback and scores in instructor-curated atomic facts, ensuring traceability and verifiability. Using the Short Answer Feedback (SAF) dataset with 31 topics, we evaluate GraphRAG on unseen-question and unseen-answer splits. Our systematic evaluation demonstrates that GraphRAG achieves grading accuracy comparable to vector-based RAG and generally superior to a fine-tuned LLM baseline model while providing more transparent source attribution. Additional findings include: (1) Instructing the LLM to discretize continuous scores to match pedagogical rubrics, such as the 0.25 increments common in SAF, improves grading accuracy; (2) LLM-generated feedback exhibits length-dependent quality variations when unconstrained; prompt-based length control substantially enhances feedback quality and its stability, achieving optimal balance of instructional richness and conciseness; (3) Performance scaling analysis reveals that basic models like GPT-4o-mini offer cost-effective performance, while premium models like Claude-Opus-4 show diminishing returns. These results demonstrate that GraphRAG offers a robust, explainable, pedagogy-aligned, and cost-effective solution for large-scale educational applications, enabling transparent automated grading with effective pedagogical feedback and practical deployment costs.

  PublisherOA PDFDOI

• Extracting Latent Dimensions from Multidimensional Response Timing Data

  Underline Science Inc. · 2025-06-18

  otherOpen access

  Computer-based assessments enable the collection of fine-grained response process data through log files. We propose a novel method for extracting latent dimensions from such multidimensional response timing data, based on applying the Weighted MDS (WMDS) model. In our method, dissimilarities among examinees in their response timing vectors are computed, one such matrix for each test item, then WMDS is applied to this collection of matrices. The resulting latent dimensions represent variation among examinees in their patterns of response timing variables, with the dimension weights of the WMDS model reflecting differences across items in the importance of the latent dimensions. Latent dimensions are interpreted via permutation-based importance, correlation analysis and network analysis. Our method is demonstrated using response data from the PISA 2018 Reading and Mathematics assessments. Results show that the extracted latent dimensions are statistically reliable, educationally interpretable, and boost predictive accuracy when used in conjunction with item scores.

  PublisherDOI

• Diagnosing Skills and Misconceptions with Bayesian Networks Applied to Diagnostic Multiple-Choice Tests

  Springer proceedings in mathematics & statistics · 2024-01-01

  book-chapter1st authorCorresponding
  PublisherDOI

• The Effectiveness of Cooperative Activities in Promoting Individual Learning

  2022-10-17

  book-chapterSenior author
  PublisherDOI

• Effects of Induced Mood on Attention and Decision Strategies in Risky Choice

  Psychological Reports · 2022-11-03

  articleSenior author

  The effects of induced incidental moods on patterns of information search and decision outcomes were investigated in a risky choice task with mixed-domain problems. Viewing of short videos was used to induce either happy or sad mood in participants, who then made choices between pairs of options consisting of a probabilistic gain coupled with a probabilistic loss. Eyetracking measures of information search, specifically frequencies of transitions between key aspects of the decision alternatives, were analyzed and related to use of heuristic or analytic compensatory strategies. Data were also gathered in a control condition, where participants were instructed to use an EV-calculation strategy, a prototypical integrative compensatory strategy. Results showed significant differences in choices and attention transitions between the EV-instruction and the induced mood conditions, but minimal differences between the happy and sad induced mood conditions. Participants in the induced mood conditions showed relatively more evidence of heuristic strategy use, but analytic strategies remained the modal strategy in all conditions. Importantly, key types of attention transitions were shown to reliably predict the frequency of observed choices consistent with optimal (EV- maximizing) and certain heuristic strategies.

  PublisherDOI

• Clustering approaches

  Elsevier eBooks · 2022 · 2 citations

  1st authorCorresponding
  • Computer Science
  • Computer Science
  • Data Mining
  PublisherDOI

• EFFECTS OF EXPERIENCE ON DECISIONS FROM DESCRIPTION: EMPIRICAL RESULTS

  Journal of International Business and Economics · 2021-03-01

  articleOpen accessSenior author

  We investigate how repeated trials with experienced outcome feedback affect risk preferences in description-based decisions under risk, and if the observed effects generalize across gain and loss domains.

  PublisherOA PDFDOI

• Modeling group process in collaborative learning and problem-solving using agent-based simulation

  2021

  Senior authorCorresponding
  • Computer Science
  • Artificial Intelligence
  • Computer Science

  Understanding how collaborative groups accomplish their work is a complex undertaking, because group problem-solving is a dynamic process, affected by interactions among many factors including the team’s tasks and goals, the prior skills and capabilities of its individual members, the communication and social dynamics between members, and other aspects of the task and the working environment. In educational contexts an additional crucial question arises: What is the relationship of group performance to individual learning outcomes, and how do group process mechanisms determine that relationship? Also, a critical task variable, discussed in the literature but little studied, is the demonstrability of the correct solution (Shaw, 1963); we investigate its role in moderating learning and performance outcomes. To investigate these questions, we propose a model of how group process facilitates individual learning and task performance in groups, and adopt a simulation approach employing agent-based modeling to assess the importance of and interactions among factors hypothesized to affect group problem-solving process, performance outcomes, and individual learning. Our simulation incorporates a simple cognitively diagnostic measurement model relating prior knowledge and learned skills to performance, a framework that offers conceptual advantages in modeling prior task-relevant knowledge, the demonstrability of correct solutions, individual learning, and process gains from the group work.

  DOI

• Additive Trees

  Wiley StatsRef: Statistics Reference Online · 2020-01-07

  other1st authorCorresponding

  Abstract Additive trees are models for the analysis of proximity data. In such analyses, the data dissimilarities are represented by path‐length distances in the additive tree graph. The additive tree may be characterized as a weighted acyclic graph, in which the weights correspond to the arc lengths. An additive tree may be displayed in rooted or unrooted form. The choice of a root for an additive tree can affect its interpretation in terms of clusters. A number of effective heuristic algorithms have been developed for least‐squares fitting of additive trees to proximities.

  PublisherDOI

Frequent coauthors

• Jeffrey V. Nickerson

  23 shared

• Barbara Tversky

  Stanford University

  16 shared

• Doris Zahner

  12 shared

• Toshihiko Matsuka

  Chiba University

  9 shared

• Lixiu Yu

  Union Hospital

  6 shared

• Yun Jin Rho

  Pearson (United States)

  6 shared

• David L. Mason

  5 shared

• Mark A. Gluck

  5 shared

Education

• B.A.

  University of North Carolina - Chapel Hill

  1977

• Other

  University of North Carolina

• Ph.D., Experimental Psychology

  Stanford University

  1983