Share the Intention, Personalize the Follow-Through: Federated Models of the Intention–Action Gap

People who say they intend to exercise, take medication, or use a mental-health app often don't follow through. Health psychology has a well-tested account of why. Behavior unfolds in two phases. In the first, a person forms an intention by weighing risks, expected outcomes, and social pressure. In the second, the person turns that intention into action through planning, coping with obstacles, and fitting the action into a daily routine. Survey studies find that the first phase is highly predictable and the second is not. An integrated model explained 83% of the variance in exercise intention among people with substance use disorders but only 12% of actual exercise (Meng et al.). On a real mental-health website, planning explained 0.1% of whether people actually engaged (Rouvere et al.).

We propose to treat this asymmetry as a design rule for federated learning, which trains one model across many phones without collecting anyone's raw data. The claim is simple. The mapping from a person's context to their intention is largely shared across people, so it should be learned jointly. The mapping from intention to action is idiosyncratic, so it should stay on each person's phone. We build a model with two factors to match. A shared intention factor is supervised by sparse daily self-reports of intention. A private follow-through factor predicts sensed action given intention and the phone's passive context. The novel part is that we test where cross-user differences actually live. If they sit in the follow-through factor, as the theory predicts, the model should need fewer personal parameters and less per-user data than generic personalization. If they do not, the theory's prediction fails, and we learn that.

Core Claim

Mobile-health federated learning already personalizes. Its standard move is to share a feature extractor and fine-tune a per-user output layer: FedHealth does this for activity recognition (FedHealth), and FFGG does it with convergence guarantees under heavily varied client data (Mishchenko et al.). These methods decide what to personalize by network depth. They never use the intermediate labels that behavioral science says matter. A scoping review of 26 mobile-health federated studies found none that used a stage or health-behavior model (JMIR scoping review).

Simply renaming the shared layers "motivational" and the personal head "volitional" would add nothing. What we add is a factorization of the prediction target, with separate supervision for each factor. Let $x_t$ be the passive sensor context on day $t$ for client $m$ (one user's phone), $i_t$ the self-reported intention, and $a_t$ the sensed action. The model is

$$ p(a_t = 1 \mid x_t) = \sigma(g_\theta(x_t))\, h_{w_m}(x_t) + \big(1 - \sigma(g_\theta(x_t))\big)\, b_{w_m}(x_t). $$

In this equation, $g_\theta$ is the shared intention factor, with parameters $\theta$ averaged across clients. The term $h_{w_m}$ is the private probability of acting given intention. The term $b_{w_m}$ is a private "bypass" probability of acting without intention, which covers habit. Meng et al. found that perceived behavioral control reached behavior directly, without passing through intention. The vector $w_m$ never leaves the phone.

Each client minimizes the action loss on every sensed day, plus $\lambda$ times the intention loss on the days it has an intention self-report. Intention labels and personal routine parameters both stay local. Only gradients for $\theta$ are shared, which also shrinks the number of shared parameters that must be protected with differential privacy (adding calibrated noise to shared updates).

Research Program

The first deliverable is the model above, trained with an FFGG-style split: local fine-tuning of $w_m$, then a global gradient step on $\theta$. This reuses the optimization that Mishchenko et al. showed avoids client drift, the tendency of each user's local updates to pull the shared model toward that user's data.

The second deliverable is a diagnostic for where heterogeneity lives. Start from a converged global model and measure, for each held-out user, the accuracy gain from locally fine-tuning only the intention factor and, separately, only the follow-through factor. Under matched parameter counts, the theory predicts that almost all of the personalization gain comes from the follow-through factor.

The third deliverable is a cold-start analysis. Local-only depression models need weeks of a person's data before they become useful (MAIKI forecasting study). If the intention factor really transfers across people, a new user should get a usable model after a few days, because only the small follow-through factor must be learned.

We use the existing literature in different roles. The survey studies by Meng et al. and Rouvere et al. motivate the asymmetry, though they are cross-sectional self-report and cannot show it in sensed behavior. FFGG and FedHealth supply the optimization and the shared/personal split. Sadilek et al. constrain the deployment setting. They showed that differentially private federated averaging reproduces centralized health estimates, reaches 99% of its best performance with 2% of devices participating per round, and stays viable at N=159 (Sadilek et al.). Those results set realistic targets for client participation and sample size. Lin et al. motivate the separation on the intervention side: effective planning interventions must target both capability beliefs and concrete planning (Lin et al.).

Data

No dataset in the file pairs repeated intention self-reports with sensed actions, so we say this plainly and run the work in two stages.

Stage one uses StudentLife: 48 students, 10 weeks of phone sensing, and daily surveys on stress, mood, and sleep. The sparse daily survey plays the role of the upstream self-reported state, and sensed physical activity and mobility serve as the downstream behavior. This stage tests the mechanism (sparse self-report factor shared, dense sensed factor private), not the psychological theory itself. We also train standard stress prediction under federation, to place our pipeline against published centralized numbers. If access is granted, MAIKI (65 depression outpatients) gives a clinical replication, with mood surveys upstream and time at home and mobility downstream.

Stage two is a purpose-built 8-week on-device study of roughly 150–200 participants, sized against Sadilek's N=159 viability result. Each participant answers a one-item daily intention question (for example, "do you intend to walk 30 minutes today?") alongside passively counted steps and phone context. This is the real test of the theory.

Baselines

The naive reference is a per-user base rate: predict each person's historical action frequency. It costs nothing, and on sparse, highly variable behavioral data it is often hard to beat. Any model that fails to beat it by a clear margin in area under the ROC curve (AUROC, a ranking accuracy score where 0.5 is chance) is not worth deploying. Purely local models, one per phone with no sharing, are the second reference. They define what federation must improve on, especially in the first two weeks.

Plain federated averaging (FedAvg) trains one global model with no personalization. FFGG reports 51.43% for FedAvg versus 64.93% for FFGG on the Google Landmarks v2 image benchmark (GLDv2) (Mishchenko et al.). We expect a similar gap on behavioral data, but no reported result was found on Lacuna for either method on behavioral sensing.

The strongest competitor is FFGG itself, run with the same backbone and the same number of personal parameters as our follow-through factor. We also give it the intention labels as an auxiliary output on the shared body, so that it has every advantage except the factorization. A FedHealth-style variant serves as a second generic-personalization baseline: a shared encoder with a locally fine-tuned head aligned to the user's feature statistics. All federated methods get the same rounds, the same client participation rate, the same differential-privacy budget, and the same hyperparameter search size.

For grounding we report centralized upper bounds where they exist. On StudentLife stress, CALM-Net reports F1 0.594 (CALM-Net) and STGNN reports F1 65.4 ± 3.1 (STGNN). On MAIKI, the best reported daily PHQ-2 error (MAE) is 0.801 and the best one-day-ahead error is 1.349 (MAIKI study). No reported federated result on these datasets was found on Lacuna.

Beating FFGG means two things. Our model must match or exceed it on held-out-day action AUROC with no more personal parameters. It must also clearly beat it on the cold-start curve over the first seven days of a new user.

The key ablations isolate the factorization. The first keeps the two-factor model but removes intention supervision ($\lambda = 0$). The second swaps the roles, making the intention factor personal and the follow-through factor shared. The third makes both factors shared, and the fourth makes both personal. The swap is the sharpest test: the theory says it should do clearly worse than the original assignment.

Why not just…?

Why not just use FFGG with intention as an auxiliary task? This is the real threat. If the network's depth already separates shared from personal structure, the explicit factorization buys nothing. Our proposal is built to find out. The matched-capacity FFGG-plus-auxiliary baseline and the $\lambda = 0$ ablation directly measure what the factorization adds beyond supervision and capacity. If they tie on accuracy and cold start, we report that generic personalization suffices. The heterogeneity diagnostic would still tell practitioners whether the theory's split matches where the data actually differ.

Why not just feed self-reported intention in as an input feature? Intention is reported on only some days and is unavailable at the moment a reminder must be timed. The factored model predicts intention from passive context on unreported days. An input-feature model must impute it. We include that variant, with intention imputed by carrying the last report forward, and compare on days without a report.

Why not just centralize de-identified data, or keep everything local? Continuous location and communication logs re-identify people even after names are removed, and Sadilek et al. show that federation preserves scientific estimates, so centralizing is not needed for validity. Centralized CALM-Net-style training is still reported as an upper bound, to measure the cost of privacy. Local-only models suffer the multi-week warm-up seen in MAIKI. The cold-start curve shows whether sharing the intention factor removes that warm-up.

Decisive Checks

Three results would settle the proposal. The first is the heterogeneity diagnostic on the purpose-built study: does fine-tuning the follow-through factor alone capture most of the per-user gain while fine-tuning the intention factor alone captures little? The second is the swap ablation: does assigning sharing and personalization against the theory clearly hurt accuracy? The third is the cold-start curve: does the factored model beat matched FFGG on new users within their first week, at equal privacy budget? A yes on all three would turn a decades-old psychological finding into a working rule for where federated health models should personalize. A no on the first would be the first evidence from sensed behavior that the motivational phase is as person-specific as the volitional one.