EduFusionFormer: A Multimodal Transformer Framework for Analyzing Educational Administrative Behavior Using Simulated Institutional Data

As there is currently no publicly available dataset that connects free-text administrative communication, activity logs, and teacher survey ratings of the same administrators in one dataset, the benchmark that was used for training and testing the EduFusionFormer algorithm was created using a simulation framework. Choosing a simulation framework is a methodological decision: because the data generation process is known, the benchmark provides a guarantee that the classification target is not a trivial function of any single observed feature, allows injecting realistic mechanism-driven missingness, and makes it possible to test the temporal, fusion, and consistency modules against the structure that was specifically created. All figures that are provided in this section - population size, attribute number, missing value number, and class proportions - are design parameters and not statistics calculated on a discovered dataset. The simulation framework generates four data sources correlated on admin_id, month, and month_step variables: the monthly administrative dataset, textual communication, survey aggregations, and administrative logs. 18,000 observations that correspond to 1,500 administrators in 12 months were generated. Monthly administrative dataset contains 91 attributes that describe operational and behavioral activities (number of meetings that were held/attended, number of reports submitted, number of approvals and pending/overdue tasks, projects being worked on, average response and completion times, attendance, punctuality, working hours, decision making, delegated tasks, interactions, budget requests, and complaints solved); its missingness budget was set to 62,862 without duplicates. The textual communication data source contains 35,201 documents that have 15 attributes correlated with an administrator-month and belong to six types of communications (Circular, Official Email, Meeting Minutes, Faculty Communication, Notice, Academic Report) and fifteen topics (e.g. accreditation preparation, examinations scheduling, student welfare, hiring, research projects, faculty review, disciplinary policy, budget allocation, curriculum development, professional development, community outreach, digital learning, quality assurance, admissions, infrastructure upgrade), along with linguistic attributes (urgency level, readability score, negative-word ratio, topic, writing style); 3,699 missing values were injected and no duplicates generated. The survey data source contains 18,000 records with 46 attributes each that aggregate the responses of 20 simulated respondents for one administrator-month. Administrative log data source contains 18,000 records with 25 attributes and missingness budget of 10,273 cells.