Nudging Past the Peak: Exposure-Aware Scheduling of Early Academic Interventions

Early warning systems in universities predict which students are likely to fail and then send them help, usually emails or prompts pointing back to course material. The standard way to spend a limited intervention budget is to rank students by predicted risk each week and contact the top of the list. This proposal argues that the policy has a hidden flaw. The same students stay at the top of the risk ranking week after week, so they receive nudge after nudge. Work on online social networks suggests that the response to a repeated prompt is often not a steadily declining benefit but a curve that peaks after one or two exposures and then falls. If student responses follow that shape, risk-ranked scheduling keeps pushing its most-contacted students past the peak. It may even make later, more important outreach less effective. Meanwhile, students just below the cutoff are never contacted at all.

We propose to measure the shape of the nudge-response curve in an educational setting, which has not been done. We then propose an allocator that decides who to nudge each week from three quantities: each student's current risk, the number of nudges that student has already received, and an explicit cost for spending future responsiveness. The central, falsifiable claim is that this allocator beats both risk-ranked and frequency-capped schedules, and beats a myopic uplift-greedy scheduler, when the response curve is non-monotone, and only then. When the curve is monotone, myopic greedy allocation is already near-optimal, and we predict no gain.

Core Claim

Three bodies of work meet here, and each plays a different role. Non-monotonic social influence supplies the mechanism. Forwarding and acceptance probabilities peak at the second exposure and decline afterward in microblogging (Shen et al.) and in e-commerce invitations (Xu et al.). Book purchases fall after the second recommendation from the same contact (Leskovec et al.). Excessive notifications suppress response altogether (Gunaratne et al.).

The most consequential finding for scheduling comes from Feng et al.. Their model lets activation probability decay with the number of prior exposures, $f(n)=n^{-\gamma}$. Under that decay, adding exposures can lower total reach: an early, weak exposure "immunizes" a target against later, stronger ones. The value of a nudge now therefore depends on what you plan to do later, and the problem stops being one where greedy choice is safe.

Early academic intervention supplies the application and its one piece of direct evidence. Zhang et al. sent relational feedback emails to at-risk programming students at two fixed points in the semester. Within two weeks, 31.8% of flagged students re-engaged with material they had never touched. Sixty percent of survey respondents asked for more frequent contact, and the authors flagged notification fatigue as the risk of doing so. That study fixed the cadence and never varied it, so it cannot say where the peak lies.

Popularity-bias mitigation supplies a lens rather than a mechanism. A risk-ranked early warning system concentrates attention on a fixed "head" of students in the same way a recommender concentrates exposure on head items. The corrective tools transfer:

  • Redistribution away from saturated users. At LinkedIn, Tu et al. redirected feedback from saturated power users to under-served creators, using a concave utility, and raised creator retention by 5.26%.
  • Measuring concentration directly. Following the session-calibration argument of CSBR, we report how evenly nudges are spread across students, not only accuracy-style outcomes.
  • Separating baseline propensity from the treatment effect. Following MACR, which subtracts the direct effect of item popularity from a recommender's score, we treat a student's baseline tendency to engage as a direct effect to be separated from the causal effect of a nudge. In observational intervention logs, the students who were nudged are exactly the high-risk ones. A naive response model would therefore learn that nudges are associated with failure.

What is new relative to Tu et al. is that we model reversal, not just diminishing returns, and we plan over time so that immunization is priced in. What is new relative to Zhang et al. is that cadence becomes a per-student decision variable, measured under randomization.

Research Program

Let $r_s(t)$ be student $s$'s predicted risk in week $t$. Let $n_s$ be the number of nudges $s$ has received so far. Let $h(n_s, x_s)$ be the probability that the next nudge produces re-engagement, given those prior nudges and student features $x_s$. A myopic scheduler contacts the $B$ students with the largest $r_s(t)\,h(n_s,x_s)$, where $B$ is the weekly budget. Our scheduler instead ranks students by

$$\Delta_s(t) = r_s(t)\,h(n_s,x_s) - \lambda\,\big[V_s(n_s) - V_s(n_s+1)\big],$$

where $V_s(n)$ is the expected value of the best remaining policy for student $s$ after $n$ nudges, and $\lambda \ge 0$ weights future against present gains. The bracketed term is the responsiveness a nudge "spends." It is near zero when $h$ is flat or rising and large when the next exposure moves the student onto the declining side of the curve. We compute $V_s$ per student by backward induction over the remaining weeks, using forecast risk trajectories. Because each student's state is just a nudge count and a week index, this computation is cheap. It avoids the global non-submodular search that made Feng et al.'s pruned greedy algorithms necessary at graph scale.

The program has three parts.

  1. Measurement. A randomized cadence trial in a large course, in a setting like Zhang et al.'s. Flagged students are randomly assigned to receive zero to four nudges over the semester. This directly estimates $h(n)$ and tests whether it is unimodal. It also tests whether a dip appears at $n=2$, as in social settings, and whether a failed early nudge lowers the response to a later one, which would be educational immunization. Randomization handles the exposure confounding that MACR-style subtraction can only approximate.
  2. Estimation. We fit $h$ as a shape-constrained function of $n$. Following the monotone decay constraints of DC-UCB, we impose unimodality instead of monotonicity, with heterogeneity across students entering through $x_s$. This also addresses DC-UCB's stated limitation that decay depends only on attempt count. The model is fit both on trial data and on observational logs with direct-effect subtraction, so we can quantify how wrong the observational estimate is.
  3. Allocation. The scheduler above, evaluated offline and then prospectively.

Baselines

The naive reference is no intervention, which gives the natural re-engagement rate of flagged students. The second reference is Zhang et al.'s fixed two-pulse schedule. It is the only deployed cadence in the file and reports 31.8% two-week re-engagement among flagged students. All other policies are run with the same total number of nudges that schedule uses, so each comparison is budget-matched.

The practitioner baseline is weekly top-$B$ by risk. In the offline study, risk comes from a gradient boosting model on Open University Learning Analytics Dataset clickstreams. Aljohani et al. report 0.894 dropout-prediction accuracy for gradient boosting on that dataset (OULAD) and HarvardX/MITx. We use the same feature family so that the risk model is held fixed and only the scheduling policy varies. A variant adds a frequency cap of one nudge per two weeks.

The strongest competitor is a myopic uplift-greedy bandit. It is DC-UCB adapted from influence maximization to students: each student is an arm, response probability is constrained to not increase with attempt count, and nudges go to those with the highest optimistic expected uplift. Kong et al. report 8% to 25% higher average reward than the next-best online baseline on NetHEPT, a citation graph. There is no reported result on Lacuna for any adaptive nudge-allocation method on an educational dataset.

We also include a feedback-shaping re-ranker in the style of Tu et al.. It scores students by risk plus $\alpha$ times a concave diminishing-returns utility, which tests whether modeling saturation alone is enough without modeling reversal.

The key ablation sets $\lambda=0$, which isolates the lookahead term. A second ablation replaces the unimodal $h$ with a monotone fit, which isolates the curve shape.

Every policy gets the same risk model, the same budget, and the same tuning budget of 50 configurations, selected on a held-out cohort. To beat the uplift-greedy bandit, the scheduler must achieve higher budget-matched recovery, meaning the fraction of flagged students who re-engage and later complete. The gap must have a confidence interval excluding zero, in the regimes where the measured or simulated curve is non-monotone.

Why not just…?

Why not just cap frequency at one email every two weeks? A cap ignores whether the last nudge worked and treats all students alike. The 60% of Zhang et al.'s respondents who wanted more contact suggests that responsiveness varies widely across students. Still, if the measured $h$ turns out to be flat for $n \le 3$ and the trial shows no immunization, a well-chosen cap will be hard to beat. The cap is therefore a first-class baseline, and the trial decides whether it is enough.

Why not just run an uplift model or contextual bandit with prior-nudge count as a feature? This is the real threat. A myopic uplift learner already stops nudging students whose uplift has fallen. Where it falls short, in theory, is immunization. If a low-value nudge now lowers the response to a later high-value one, greedy allocation spends that future responsiveness without accounting for it, which is exactly Feng et al.'s non-monotone failure. The $\lambda=0$ ablation settles this directly. If lookahead adds nothing in the trial-calibrated simulator, we will report that uplift-greedy suffices, and the contribution reduces to the measurement study.

Why not just vary the message content so fatigue never sets in? It may be that redundancy, not count, drives the decline. That is the explanation Shuai et al. offer for their drop at two spreaders (Shuai et al.). The trial therefore crosses nudge count with varied versus repeated content. If varied content flattens $h$, the count-based curve is the wrong state variable, and the scheduler should track content history instead.

Proposed Experiments

Offline study. We build a semester simulator on OULAD weekly clickstreams, with risk forecast by the gradient boosting model. Response curves come from four families drawn from the file:

  • concave saturation, as in Tu et al.;
  • peak at the second exposure, as in Shen et al. and Xu et al.;
  • power decay with immunization, as in Feng et al.;
  • a dip at two exposures, as in Shuai et al.

The predicted pattern is that the scheduler and uplift-greedy tie under concave saturation, while the scheduler wins under the other three families. That pattern is the decisive internal check. Outcomes are simulated recovery, nudges per recovered student, and the Gini coefficient of nudges across students, the concentration measure used in popularity-bias work.

Fairness check. Following Cheng et al., we report demographic parity of who gets contacted. Spreading nudges toward borderline students could shift attention across groups in either direction, so this is an outcome to watch, not a guaranteed gain.

Prospective study. The randomized cadence trial both estimates $h$ and evaluates policies by off-policy evaluation on the logged randomized assignments. A following-year cohort tests whether the curve transfers. This matters because Zhang et al. saw risk-model AUC drop from about 0.98 within the training cohort to 0.72–0.82 on the next year's students.

The proposal fails cleanly in either of two cases: if the trial finds a monotone curve with no immunization, or if the $\lambda=0$ ablation matches the full scheduler.