Yulong Zhao

Revealed actions whose effect on task success reverses over task time, a pattern masked by aggregate action counts, in 19,089 events from 2,668 users of an educational problem-solving assessment (OECD PISA 2012). Developed the interpretable model behind it, which estimates how each action's effect changes with when it occurs and classifies every action as having no effect, a constant effect, or a time-varying effect. The model also improved prediction over a count-feature logistic baseline: 50% lower log loss (0.237 to 0.118) and accuracy from 93.5% to 96.1% in 10-fold cross-validation, winning every fold. See the interactive curve at the top.

More than doubled the share of failing users caught at least 2 actions before submission, from 14% to 34% at a 10% false-alarm budget, compared with a published early-window XGBoost baseline, on OECD PIAAC assessment data, by adding users' behavior on earlier tasks. Built the online early-warning system behind it: a neural sequence model (GRU) updates each user's risk of failing after every action and raises an alert before the user submits, while there is still time to intervene. Designed a model-agnostic alert rule calibrated on each user's peak risk over the whole sequence, with a provable guarantee that the share of successful users ever falsely alerted stays within a chosen budget.

Recovered interpretable structure without labels on 1,129 users of an email-and-spreadsheet task: 3 skills and 2 behavioral states learned from 48 action types, with distinct skills for search, sort, and email actions, skills with opposing effects on related actions, and state transitions consistent with the task logic. All 14 learned skill-to-action links reappeared in at least 94% of bootstrap resamples. Developed the structured latent-variable model behind it, which jointly estimates persistent user skills and evolving behavioral states from interaction sequences, learning sparse skill-to-action mappings without prespecified assignments or labeled states.

Reduced forecast MSE by 38% on average versus a historical-mean baseline, beating it in all 18 settings, with a nonlinear regression pipeline (PCA + kernel SVR) forecasting biological change from paleo-environmental records, evaluated strictly out-of-sample (walk-forward). Evaluated environmental regime shifts using recurrence plots, multiple-testing correction, and preprocessing sensitivity analysis, distinguishing robust findings from preprocessing-dependent results.