Personalized federated learning (PFL) may improve client-level performance, but deeper personalization increases deployment cost. We formulate personalization-depth selection as a utility-based empirical decision problem in which client utility equals Macro-F1 minus a cost-sensitivity weight times deployment cost, over six personalization depths labeled P0 through P5. We compare fixed policies with an offline per-client Oracle reference and report allocation maps over Dirichlet data heterogeneity and cost sensitivity. Although evaluated on this hierarchy, the formulation can accommodate any finite candidate set. On CIFAR-10, CIFAR-100, and a PTB-XL subset, personalization shows empirical, setting-dependent decision regimes rather than a universal benefit: higher cost sensitivity shifts Oracle mass toward the global-only depth P0; low cost sensitivity can favor deeper actions. Patterns are grid observations (three seeds per cell), not theoretical boundaries. Best-uniform is competitive with Oracle assignment (6.5 percent plus or minus 7.8 percent relative regret; 95 percent confidence interval 5.9 to 7.1), whereas random and personalize-all incur much larger regret (189.3 percent plus or minus 367.1 percent and 409.0 percent plus or minus 803.2 percent). Under the same utility, Ditto full-model personalization is substantially suboptimal once cost is counted. The contribution is an offline characterization of when personalization depth is worth its cost, not a new PFL optimizer or deployment-time scheduler.
Figures & tables
Action
Locally trainable components
P0
None; global-only model
P1
Classifier head
P2
Adapter and head
P3
Last encoder block, adapter, head
P4
Last two encoder blocks, adapter, head
P5
Full model personalization
TABLE I: PERSONALIZATION HIERARCHY ADOPTED IN THIS STUDY (EXPERIMENTAL A ).
Comp.
Meaning
Pre-weight scale
Weight
Params
Trainable parameters at a
÷105
0.25
FLOPs
Local train+eval estimate
÷108
0.20
Time
Local training wall time (s)
—
0.25
Comm
Model-delta bytes
÷106
0.15
Mem
Param+activation proxy (MB)
÷100
0.15
TABLE II: COST COMPONENTS, PRE-WEIGHT SCALING (APPLIED BEFORE WEIGHTING), AND DEFAULT WEIGHTS.
Fig. 1: Empirical offline-Oracle allocation maps over Dirichlet heterogeneity α and cost sensitivity λ . Each row shows CIFAR-10, CIFAR-100, and PTB-XL for one metric: (a) P0 mass (%), (b) deep-action mass P3 – P5 (%), (c) Oracle gain over the best-uniform policy (%), (d) P5 -for-all penalty (color scale; cell labels omitted due to extreme magnitude range). The maps show regime-dependent allocation patterns on the finite experimental grid under A={P0,…,P5} .
Policy
Regret (%)
Oracle win (%)
95% CI
Random
189.3±367.1
99.7±5.8
[160.6,219.8]
Best-uniform
6.5±7.8
82.5±38.0
[5.9,7.1]
Global-only
28.2±29.6
83.3±37.3
[25.9,30.6]
Personalize-all
409.0±803.2
99.2±9.1
[346.2,472.9]
TABLE IV: COST-AWARE REGRET VS. OFFLINE PER-CLIENT ORACLE ( n=594 ). LOWER IS BETTER; 95% BOOTSTRAP CIS.
Policy
Regret (%)
95% CI
Oracle win (%)
Ditto
454.9±875.8
[380.7,534.7]
94.0
Best-uniform
6.4±8.0
[5.7,7.1]
80.7
Global-only
27.9±29.7
[25.3,30.6]
81.7
TABLE V: DITTO UNDER THE SAME UTILITY (P5-EQUIVALENT COST).
Dataset
Regret (%)
Win (%)
95% CI
CIFAR-10
6.4±10.4
81.7±38.8
[4.7,8.3]
CIFAR-100
5.5±8.5
63.9±48.2
[4.4,6.8]
CIFAR-10 aux.
7.1±6.7
90.7±29.1
[5.9,8.4]
PTB-XL subset
7.2±5.0
96.7±18.0
[6.5,7.9]
TABLE VI: BEST-UNIFORM REGRET VS. OFFLINE ORACLE BY DATASET.
Dataset
P0 ratio
Deep ratio
Unif. gap
P5 gap
CIFAR-100
0.620±0.383
0.006±0.016
9.3±12.5
136±32
CIFAR-10 aux.
0.326±0.297
0.074±0.092
7.1±4.4
1850±3376
PTB-XL
see Table VIII
Overall
0.499±0.317
0.069±0.096
9.1±8.3
695±2082
TABLE VII: ORACLE SPARSITY AT λ=0.3 (SPARSITY-VALIDATION CELLS).
Department of Computer Engineering, Istanbul Technical University, Istanbul, Turkey · Department of Computer Engineering, Marmara University, Istanbul, Turkey