Personalization can improve activity-recognition performance, but participant-specific gains are heterogeneous, and every additional calibration label has an acquisition cost. This study presents HB-PVI, a hierarchical Bayesian personalization and value-of-information framework jointly modeling participant heterogeneity, the benefit and harm of four personalization mechanisms, and the economic value of an additional label, for the 47-participant MUSIC-CAR complex-activity cohort. A leakage-safe, leave-one-participant-out evaluation combines a sequential-Monte-Carlo participant-effect updater with a Student-t hierarchical gain model and a one-step expected-value-of-sample-information (EVSI) stopping rule. Adapter personalization produced small positive mean F1 gains, growing from 0.00099 at one label to 0.00198 at ten, while adapter-plus-head and prototype-residual personalization were negative on average. Under the primary practical-benefit threshold (Δmin=0.01) and cost setting, one-step EVSI was zero at every decision state, so the policy purchased no labels and retained population inference for all 47 participants, matching always-stop exactly (region-of-practical-equivalence probability =1). Relative to fixed ten-shot adapter personalization, this reduced labeling by 100% while keeping the posterior mean F1 loss at 0.00217 (95% credible interval, 0.00048 to 0.00389), with posterior probability 0.9992 of remaining below the 0.005 tolerance. HB-PVI was utility-optimal in 199 of 216 cost-threshold settings and in every setting at or above the primary label cost. These results argue for a population-first deployment policy whenever personalization gains are small relative to labeling, computation, and harm costs, and show that value-of-information reasoning, not raw predictive accuracy, should drive personalization decisions in health-sensing applications.
Figures & tables
Action
Shot
Mean gain
Pos. (%)
Ben. (%)
Harm (%)
Adapter
1
+0.00099
61.7
8.5
4.3
10
+0.00198
72.3
10.6
4.3
Head
1
+0.00109
51.1
14.9
8.5
10
+0.00037
55.3
19.1
10.6
Adapter+head
1
−0.00236
44.7
17.0
21.3
10
−0.00449
48.9
21.3
25.5
TABLE I: Action-specific personalization gain at 1 and 10 purchased labels (percentages of 47 participants)
Fig. 1: Action-specific personalization outcomes across purchased labels. (A) Mean sample-level F1 gain relative to population inference, with 95% intervals. (B) Percentage of participants with meaningful benefit ( g>0.01 ). (C) Percentage with meaningful harm ( g<−0.01 ).
Study
Backbone
Personalization
Mean F1
MUSIC-CAR baseline [ 4 ]
Feature+XGBoost
None
0.588
P-HART [ 1 ]
P-HART transf.
None
0.880
P-HART [ 1 ]
P-HART transf.
6-shot adapter
0.926
HB-PVI (this study)
P-HART-style a
None (population)
0.904
HB-PVI (this study)
P-HART-style a
Adapter, 6 shots
0.905
HB-PVI (this study)
P-HART-style a
Adapter, 10 shots
0.906
TABLE II: Population and personalized mean F1 across related studies on the same cohort or sensor modality
Action
MAE
RMSE
Cov. (%)
Brier g>0
Brier g>.01
Adapter
0.00565
0.00824
91.9
0.2695
0.0915
Head
0.00958
0.01321
91.1
0.2721
0.1257
Adapter+head
0.01396
0.01857
88.9
0.2940
0.1318
Prototype-resid.
0.00325
0.00546
95.3
0.1860
0.0002
TABLE III: Held-out gain-model prediction and calibration by action
Fig. 2: Population-only F1 versus adapter F1 at (A) 1, (B) 6, and (C) 10 purchased labels, all 47 participants (Table ). Points scatter closely around the y=x line at every shot count, and meaningful-benefit (green) and meaningful-harm (red) cases both occur at low and high baseline F1 alike, illustrating that realized gain is not predictable from population-only performance ( r=−0.06 at 10 shots.
Fig. 3: Fixed-order one-step EVSI and stopping decisions. (A) Mean EVSI under the primary setting; no state-level decision continued. (B) Labels purchased across the three practical-gain thresholds. (C) Action selected at stopping under the primary Δmin=0.01 setting.
Comparator
Util.
Δ mean
95% CrI
P(Δ>0)
Fixed adapter@1
−0.00434
+0.00387
(0.00054, 0.00723)
0.9892
Fixed adapter@6
−0.00894
+0.00820
(0.00492, 0.01157)
1.0000
Fixed adapter@10
−0.01129
+0.01017
(0.00703, 0.01343)
1.0000
TABLE IV: Realized comparator utility and posterior paired-utility difference (HB-PVI minus comparator)
cL
0
0.0005
0.001
0.0025
0.005
0.01
HB-PVI optimal (%)
61.1
91.7
100.0
100.0
100.0
100.0
TABLE V: Percentage of the 36 computation-and-harm-penalty settings, per label cost cL , in which HB-PVI was utility-optimal (across all Δmin )
Fig. 4: Cost sensitivity and F1 preservation with zero calibration labels. (A) Percentage of computation-cost and harm-penalty settings in which HB-PVI was utility-optimal, by label cost and practical-gain threshold; the dotted vertical line marks the primary label cost. (B) Bayesian and bootstrap estimates of mean F1 loss from purchasing zero labels rather than using fixed adapter personalization at 1, 6, or 10 shots. The dashed vertical line marks the F1-loss tolerance of 0.005, and the annotations report the posterior probability that the mean loss remained below this tolerance.
Table S2: Validation gates for the sequential updater, evaluated on all 47 outer folds
Predictor
Domain
Shot count
Primary predictor
Ensemble mutual information (mean)
Ensemble uncertainty
Ensemble predictive entropy (mean)
Ensemble uncertainty
Ensemble predictive variance (mean)
Ensemble uncertainty
Purchased-label cardinality (mean)
Purchased-label summary
Purchased-label pattern diversity
Purchased-label summary
Appendix
Table S3: Sequential-state predictor dictionary used in Model II
Quantity
Value(s)
Practical-gain thresholds Δmin
{0,0.005,0.01}
Eligibility constraint
P(g>Δmin)≥0.90 , P(g<0)≤0.10
Label-cost grid cL
{0,0.0005,0.001,0.0025,0.005,0.01}
Adapter computation-cost grid cC
{0,0.0005,0.001}
Population computation cost
0 (fixed)
Harm-penalty grid cR
{0,0.005,0.01,0.02}
Appendix
Table S4: EVSI cost specification
Action
Shot
Mean F1 (95% CI)
Gain
Ben. > .01 (%)
Harm <− .01 (%)
Adapter
1
0.9052 (0.8810, 0.9272)
+0.00099
8.5
4.3
2
0.9053 (0.8792, 0.9271)
+0.00109
8.5
4.3
3
0.9053 (0.8797, 0.9270)
+0.00105
10.6
6.4
4
0.9054 (0.8810, 0.9274)
+0.00121
8.5
4.3
5
0.9055 (0.8807, 0.9270)
+0.00124
8.5
4.3
6
0.9054 (0.8808, 0.9262)
+0.00118
8.5
4.3
Appendix
Table S5: Action-specific mean F1 and gain by shot (all 10 shots; percentages of 47 participants)
Shot
Min
Q1
Median
Q3
Max
1
−0.02188
−0.00168
+0.00125
+0.00369
+0.02083
6
−0.02486
−0.00214
+0.00164
+0.00469
+0.01932
10
−0.02650
−0.00058
+0.00203
+0.00533
+0.01678
Appendix
Table S6: Distribution of realized adapter gain across 47 participants, by shot count
Fold
Population F1
Adapter F1@1
Adapter F1@6
Adapter F1@10
Gain@10
000
0.5664
0.5606
0.5621
0.5777
+0.0113
001
0.7462
0.7505
0.7471
0.7482
+0.0020
002
0.6804
0.6811
0.6772
0.6699
-0.0105
003
0.7883
0.7887
0.7902
0.7885
+0.0002
004
0.7906
0.7687
0.7669
0.7641
-0.0265
005
0.9203
0.9158
0.9183
0.9203
+0.0000
Appendix
Table S7: Exact participant-specific population and adapter F1 scores, all 47 participants
Figure S1: Held-out gain-model prediction and calibration. (A) Prediction error (MAE, RMSE) by action. (B–C, E) Reliability curves comparing predicted and observed frequencies for positive gain, gain >0.01 , and negative gain (five probability bins per action-event combination; bins with <10 held-out observations omitted). (D) Empirical coverage of the nominal 95% predictive interval by action, with the 95% reference line.
Comparator
Div.
Max treedepth
Max R
Min bulk ESS
Min tail ESS
Always-stop
—
—
—
—
— a
Fixed adapter@1
0
0
1.0012
5871
5802
Fixed adapter@6
0
0
1.0007
5224
5909
Fixed adapter@10
0
0
1.0008
4099
3495
Appendix
Table S8: Paired-comparison posterior diagnostics (HB-PVI minus comparator)
Δmin
cL
HB-PVI best (%)
Runner-up policy (count)
0.000
0.0000
66.7
fixed_adapter_10 (4)
0.000
0.0005
91.7
fixed_adapter_1 (1)
0.000
0.0010
100.0
—
0.000
0.0025
100.0
—
0.000
0.0050
100.0
—
0.000
0.0100
100.0
—
Appendix
Table S9: HB-PVI utility-optimality by practical-gain threshold and label cost
Figure S2: HB-PVI utility-optimality (%) across the full label-cost ( cL ) and harm-penalty ( cR ) grid, averaged over the three computation-cost ( cC ) settings, for each practical-gain threshold. HB-PVI is never preferred only in the single cell where both costs are zero; it is preferred in a minority of settings only when exactly one of the two costs is zero and the other is at its lowest tested nonzero level; everywhere else in the grid, HB-PVI is optimal in 100% of cases.
Reference
Post. τ
Post. ν
Max R
Min bulk ESS
Min tail ESS
Fixed adapter@1
0.00490
4.52
1.0011
3001
3300
Fixed adapter@6
0.00468
3.70
1.0009
4014
4088
Fixed adapter@10
0.00510
5.75
1.0007
4159
4349
Appendix
Table S10: Label-efficiency posterior diagnostics (Bayesian Student- t model)
cL
Mean EVSI at stop
Max EVSI observed
Mean labels purchased
0.0000
0.00405
0.00411
10.0
0.0005
0.00405
0.00411
10.0
0.0010
0.00405
0.00411
10.0
0.0025
0.00956
0.01012
1.0
0.0050
0.00956
0.01012
1.0
0.0100
0.00956
0.01012
1.0
Appendix
Table S11: Positive one-step EVSI decisions, grouped by label cost (all at Δmin=0 ; n=12 decisions per row)
#
Objective
Status
1
P (mean utility > best existing policy) >0.95 without worse calibration
Sensor-based Human Activity Recognition (HAR) models often degrade on unseen users due to domain shifts caused by individual movement patterns and sensor placement. Practical wearable HAR systems therefore require personalization methods that are lightweight, applicable across diverse calibration scenarios, and robust under limited calibration data. We present BayaHAR, a gradient-free framework that repurposes pretrained HAR classifiers as Prototypical Networks using prior prototypes that preserve zero-shot performance while regularizing adaptation. For labeled calibration data, we introduce closed-form Bayesian prototype estimation and extend the same principle to weakly labeled data, requiring only knowledge of which activities were performed. With only 3 seconds of calibration (one shot) per activity, supervised adaptation improves test macro-F1 on unseen users by +2.76 to +33.44 percentage points across four datasets, while weakly supervised adaptation improves by +0.56 to +32.13 points. Since adaptation requires only closed-form prototype updates, the framework enables efficient and robust on-device personalization of preexisting HAR classifiers.
Maximilian Burzer, Till Riedel, Michael Beigl +1
Karlsruhe Institute of Technology Karlsruhe, Germany · IPAI Foundation gGmbH Heilbronn, Germany
Over the past couple of decades, many active learning acquisition functions have been proposed, leaving practitioners with an unclear choice of which to use. Bayesian active learning offers principled objectives with explainable intuition, including Expected Error Reduction (EER), Expected Predictive Information Gain (EPIG), and Bayesian Active Learning by Disagreements (BALD). A key challenge of such methods is the difficulty of extending to the batch setting, leading to either computational challenges (BatchBALD) or dramatic performance drops (top-B selection). In this work, we extend active learning's Bayesian Decision Theory framework to derive Partial Batch Label Sampling (ParBaLS) for the EPIG algorithm. We show experimentally that ParBaLS EPIG gives superior performance for a variety of datasets with different budgets using Bayesian logistic regression on embeddings from large pre-trained models. Our code is available at https://github.com/ADDAPT-ML/ParBaLS.
Kangping Hu, Stephen Mussmann
Department of Computer Science, Georgia Institute of Technology, Atlanta, GA 30332
General-purpose models promise sensor-based decisions without training a task-specific classifier, which could reduce the dependence of Human Activity Recognition (HAR) on labeled data. Yet it remains unclear whether such models can directly interpret deterministic descriptions of physical sensor signals well enough to replace or complement trained HAR models. We study this question using Jev, a fixed general-purpose probabilistic decision model, on 1,800 class-balanced accelerometer windows from WISDM, UCI341, and PAMAP2. Jev receives no labeled examples, retrieval context, or HAR-specific parameter updates. We evaluate three deterministic sensor representations and compare 5,400 Jev decisions with a generative baseline and three supervised HAR models. Jev remains far below supervised recognition, with its strongest representation reaching macro-F1 of 0.038, 0.118, and 0.089 across the three datasets, compared with 0.686 to 0.907 for the supervised models. More numerical features do not improve Jev. Instead, they reduce recognition on all three datasets, while augmenting the same numerical evidence with a deterministic semantic rendering partially recovers performance, although the experiment does not isolate semantics from the accompanying serialization and redundancy changes. Jev is fast and inexpensive to query, but its probabilities are not reliably calibrated for recognition. A post-hoc fusion analysis finds a small improvement on WISDM that does not replicate on UCI341 or PAMAP2. These results show that training-free sensor decisions depend not only on the information available in the signal, but also on whether the model can use the representation through which that information is exposed. The sensor-to-model interface should therefore be treated as part of the model evaluation rather than as a neutral preprocessing step.
Orhan Konak
Hasso Plattner Institute, University of Potsdam Potsdam, Germany