By Ran Liu
Chief Data Scientist at MARi
One of the ultimate goals of education is to effectively prepare students for long-term career and workforce success. Most existing intelligent systems in education focus on tutoring for specific academic subjects and deliver personalized learning on a relatively short timescale. As we face an ever-changing 21st century job market due to increased automation and AI technologies, it is more important and challenging than ever to deliver learning and mentorship that leads to long-term success in the workforce. In my webinar, I discuss some of the unique challenges presented by this goal as well as promising approaches that can address the challenges and move us closer to building effective, workforce-driven intelligent mentoring systems.
One prominent challenge lies in collecting and storing fine-grained, longitudinal data that tracks an individual’s skills and abilities from K12-through-career. These data are needed to develop intelligent models that can provide mentorship with respect to long-term goals. It is difficult to find or collect such datasets, as this typically involves integrating data from many disparate sources, over long periods of time, and that come at multiple grain sizes. There are some existing sources of longitudinal data, primarily from state education departments (e.g., Virginia and Nevada). These datasets, however, track high-level metrics of academic achievement and post-K12 success such as college enrollment and completion, and wages from first jobs. In order to amass the kinds of fine-grained data that will enable us to connect K12 education to career and workforce competencies in a deep and meaningful way, we need a flexible, interoperable data architecture as well as intelligent ways to connect in and map data from a wide array of sources. The Common Education Data Standards is a Department of Education initiative that has strong potential as a standardized data framework supporting K-through-Career data integration and storage. Supplemented by the application of Natural Language Processing (NLP) approaches to mapping skills across different educational and competency standards frameworks, I believe that this interoperable data framework will make it possible to collect organized, integrated, and granular longitudinal data.
Another challenge is the need to measure and quantify the variety of non-academic skills that are likely to be critical for job/workforce success. These include skills such as critical thinking, self-efficacy, grit, creativity, problem-solving, and reliability. The quantification of such skills will be necessary in order for intelligent mentoring systems to identify skill gaps and deliver personalized recommendation/remediation. There is a history of research-validated survey- and self-report-based methods to measure these skills and recently increased efforts to combine them into various “21st Century Skills” frameworks. Although survey-based methods remain the current standard for assessing such traits, new technologies resulting from the “Internet of Things” movement, combined with active research on Multi-Modal Learning Analytics (MMLA), show promise for more naturalistic and context-sensitive ways to assess these personal characteristics. The field of MMLA has produced promising methods for capturing, processing, and fusing natural rich modalities of communication, such as speech, writing, and nonverbal interaction (e.g., movements, gestures, facial expressions, gaze, biometrics, etc.) during genuine learning activities as they occur in or outside the classroom. However, the development of the models that could use such information to reliably measure 21st century skills is still in its infancy.
Lastly, I discuss two primary aspects of adapting personalized learning models to operate longitudinally. The first aspect involves identifying the constituent skills needed for workforce-relevant goals, identifying each individual’s specific skill gaps based on measurement data, and recommending resources to close the skill gaps. This primarily involves mapping the content of job postings and career competency profiles to the fine-grained skills for which there’s data to bring to bear. Topic modeling and other NLP approaches can intelligently eliminate the burden on human effort to accomplish such a mapping. The second aspect involves developing models that can find common “clusters” of longitudinal pathways and learn the common event contingencies that lead to desirable outcomes observed far down the event chain. In addition to sequential clustering techniques, one class of models that may prove promising are Markov Decision Processes (MDPs). I discuss conceptual example of how we might utilize other student’s historical K-through-career pathways to train MDPs that intelligently help future students navigate their own.