Source-linked AI summary
Algorithmic Monoculture in Machine Learning-based Site-Specific Fertilizer Recommendation
Abdelghani Belgaid
TL;DR
Model-by-model evaluation can miss how shared errors in fertilizer prescriptions aggregate across farms. The paper models adoption shares, shared and farm-specific errors, and cross-model dependence, then calibrates allocation policies to Moroccan cereal benchmarks. Under illustrative parameters and a 10% regret cap, constrained diversification lowers modeled loss in all five systems, but these counterfactuals are not estimates of agronomic or monetary welfare.
Problem
Model-centered evaluation gives limited attention to correlated consequences across farms, although fertilizer recommendations influence nutrient purchases, application rates, and environmental externalities.
Source: §1 · p. 1Method
The paper models shared and farm-specific recommendation errors, links expected social loss to adoption shares, generalizes to cross-model covariance, and evaluates regret-constrained portfolios.
Source: §1 · pp. 1–2, §3 · p. 3Results
10.89–41.53%: constrained portfolios reduce modeled loss across all five systems, averaging 19.41%, under the central calibration.
Source: §5.1 · p. 4Takeaways & Limitations
Ecosystem-level evaluation should consider both predictive-risk safeguards and error covariance alongside externality assumptions.
Source: §8 · p. 6
Abstract
from arXiv · showhide
Machine learning recommendations are often evaluated one model at a time, despite deployment across interconnected decision makers. For continuous NPK fertilizer prescriptions, correlated errors can aggregate across farms and generate systemic costs. A model of shared and farm-specific recommendation errors links expected social loss to model adoption shares; the Herfindahl-Hirschman Index characterizes concentration effects when common model errors are independent and predictive risks are equal. A covariance-generalized formulation captures the dependence between model families. A two-model threshold characterizes when privately optimal monoculture is socially dominated, and a convex portfolio rule limits average private predictive regret. Nine model families across five Moroccan cereal systems provide a benchmark-based calibration. Under illustrative parameters and a 10% regret cap, accuracy-constrained diversification lowers modeled social loss in all five systems by 10.89-41.53% (mean 19.41%). Uniform allocation reduces modeled loss in this setting, but breaches the regret cap in every crop. Correlation sensitivity shows that highly dependent model errors attenuate the benefits of diversification. These counterfactual outcomes are not estimates of agronomic or monetary welfare; the framework identifies adoption architecture and error dependence as ecosystem-level evaluation variables.
1 Introduction
The introduction frames fertilizer recommendations as interconnected decisions whose shared errors can create costs across farms, a dimension missed by model-by-model evaluation. It presents an economic framework, a benchmark calibration, and a boundary on interpreting its counterfactual outcomes.
1 Introduction
- Model-by-model predictive evaluation gives limited attention to correlated consequences when recommendations shape nutrient purchases, application rates, and environmental externalities.For continuous prescriptions, model-common errors can synchronize over- or under-application across farms, with quadratic costs attached to aggregate imbalances.Source: §1 · p. 1
- The framework links adoption shares, shared recommendation errors, and aggregate externalities, with a covariance generalization and an accuracy-constrained portfolio rule.It also gives a closed-form monoculture threshold; the analysis assigns distinct models across farms rather than averaging predictions within a farm.Source: §1 · pp. 1–2
- Nine model families across five Moroccan cereal systems calibrate the framework and illustrate sensitivity to externality intensity, error concentration, and cross-model dependence.Quantitative outcomes are conditional on benchmark predictive risks and stated sensitivity parameters, and modeled loss reductions are not estimated yield, profit, or environmental-damage changes.Source: §1 · p. 2
2 Related literature
The related literature motivates ecosystem-level analysis of shared decision technology and distinguishes deployment-level externalities from prediction quality or single-decision optimization. This paper focuses on correlated multidimensional prescription errors across farms.
2.1 Algorithmic monoculture and predictive multiplicity
- Shared model errors in continuous nutrient prescriptions can synchronize over- or under-application across adopters, unlike congestion or competition mechanisms in ranking and matching.Each farm uses one model, so diversification does not average away its private error but can reduce covariance-driven loss across adopters.Source: §2.1 · p. 2
2.2 Decision-focused learning and fertilizer economics
- The formulation adds cross-farm coupling through aggregate externalities to decision-focused learning and fertilizer economics, targeting deployment-level loss from correlated multidimensional prescription errors.It distinguishes prescription fidelity from economically optimal nitrogen rates and treats fertilizer externalities as potentially different from private costs.Source: §2.2 · p. 2
3 Economic model
The economic model separates shared and farm-specific recommendation errors and connects adoption shares to private and systemic loss. It characterizes concentration and diversification incentives, then constrains portfolio choice by average private predictive regret.
3 Economic model
- Expected loss combines average private predictive risk with a quadratic penalty on normalized aggregate prescription error; farm-specific contributions vanish at rate 1/n while common shocks persist.The model uses a benchmark reference prescription, which is not an identified agronomic optimum.Source: §3 · pp. 2–3
3.1 Concentration and monoculture threshold
- With equal predictive risks and independent model-common shocks, the Herfindahl–Hirschman Index captures the concentration penalty; for two unequal-risk models, the most accurate model need not minimize population loss.Positive cross-model covariance can eliminate the incentive to diversify in the two-model formulation.Source: §3.1 · p. 3
3.2 Accuracy-constrained diversification
- The convex portfolio rule minimizes social loss subject to an average private-regret bound, which constrains population-average predictive risk rather than every farm’s risk.The objective is convex because the cross-model covariance matrix is positive semidefinite; the allocation differs from single-farm ensemble prediction.Source: §3.2 · p. 3
4 Empirical calibration
The empirical calibration uses benchmark recommendation outputs from nine model families across five Moroccan cereal systems. It defines benchmark risk using target-normalized unexplained variance and compares allocation policies under illustrative assumptions.
4.1 Benchmarks and normalized risk
- Nine regression model families are evaluated across rainfed barley, grain maize, silage maize, irrigated wheat, and rainfed wheat using six spatial and soil covariates and crop-specific 80/20 partitions.The benchmark results are associated with turba-models recommendation scenarios; reference NPK targets come from Fertimap workflows and are not observed crop responses.Source: §4.1 · pp. 3–4
- Benchmark risk is the mean target-normalized unexplained variance, with equal variance-normalized nutrient weights used as a calibration convention rather than estimated nutrient-specific externality valuations.On a common evaluation sample, this corresponds to the model’s normalized-error formulation.Source: §4.1 · p. 4
4.2 Counterfactual design
- The counterfactual design compares lowest-risk monoculture, uniform allocation, constrained loss minimization, and unconstrained social-loss minimization across assumed parameter grids.The central illustrative settings are ρ = 0.25, λ = 20, n = 10,000, and an average private-regret cap δ = 0.10; these parameters and cross-model correlations are not identified by aggregate benchmark statistics.Source: §4.2 · p. 4
5 Experimental results
Under the central calibration, accuracy-constrained portfolios reduce modeled social loss in all five crop systems. Sensitivity analyses show that benefits depend on externality intensity and the assumed dependence between model errors.
5.1 Predictive multiplicity and constrained outcomes
- For ρ = .25 and n = 10,000, analytical two-model thresholds range from 0.35 for silage maize to 2.34 for irrigated wheat.Figure 1 identifies the three lowest-risk model families in each crop system; distinct alternative risks determine the private cost of diversification.Source: Figure 1 · p. 4, §5.1 · p. 4
- 10.89–41.53% (mean 19.41%) modeled-loss reductions are achieved by constrained portfolios across the five crop systems.The central calibration uses ρ = .25, λ = 20, n = 10,000, and δ = .10; percentages refer to a dimensionless model objective, not realized welfare changes.Source: §5.1 · p. 4, Table 1 · p. 5
- Constrained allocations assign the lowest-risk model 91.0% of barley farms, 82.1% of grain maize, 57.0% of silage maize, 91.7% of irrigated wheat, and 90.4% of rainfed wheat.Diversified allocations differ substantially in concentration across crops.Source: §5.1 · p. 5
- Uniform allocation lowers modeled loss in the central calibration but exceeds the 10% average private-regret cap in every crop, by 22.99–99.16%.At λ = 0, uniform diversification increases loss because no systemic externality offsets the sacrifice in private predictive risk.Source: §5.1 · p. 5
5.2 Parameter and covariance sensitivity
- 69.0% of 600 crop-by-parameter scenarios yield strictly positive constrained loss reductions; gains vanish at λ = 0 and generally rise with common-error exposure and externality intensity.The 69.0% is a fraction of the prespecified grid, not an estimated probability of benefits; Figure 2 summarizes sensitivity across ρ and λ.Source: Figure 2 · p. 5, §5.2 · p. 5
- With assumed κ = .5, constrained gains range from 1.19–15.94% across crops; with κ ≥ .95, no material gains remain in this calibration.Correlations are not empirically identified, so these results make error dependence a consequential condition for diversification benefits.Source: Figure 3 · p. 5, §5.2 · p. 6
6 Economic interpretation and institutional value
The section interprets diversification as allocation of distinct models across farms, not averaging predictions within a farm, and identifies institutional factors relevant to adoption.
6 Economic interpretation and institutional value
- A centralized advisor could assign distinct models across farms, while decentralized providers could generate adoption shares through individual choices.The regret constraint limits average predictive sacrifice when governance objectives favor diversification; provider and model-version concentration and error dependence are relevant deployment statistics.Source: §6 · p. 6
7 Limitations
The benchmark measures fidelity to upstream recommendations rather than causal agronomic outcomes, and the modeled gains are conditional rather than welfare estimates.
7 Limitations
- The benchmark uses deterministic crop-specific random splits that primarily measure interpolation; nearby locations may be dependent, and maize has fewer locations than winter cereals.Aggregate R2 cannot identify common-error share, externality weight, or model-error covariance; additional grouped or repeated observations would be needed for variance-component identification.Source: §7 · p. 6
- Adoption shares are treated as exogenous, the quadratic externality is stylized, and modeled gains are not estimated profits, emissions reductions, or monetary social welfare.Fertilizer-market effects and environmental damage require separate causal or structural evidence.Source: §7 · p. 6
8 Conclusion
Shared fertilizer-model errors can aggregate across farms into systemic costs. The framework connects adoption patterns and error dependence to modeled social loss, while treating welfare effects as counterfactual rather than observed.
8 Conclusion
- Shared fertilizer-model errors can aggregate across farms and generate systemic costs.The conclusion frames continuous fertilizer prescriptions as an economic mechanism for algorithmic monoculture.Source: §8 · p. 6
- Under stated assumptions, HHI captures concentration risk, while correlated common errors make the covariance-weighted adoption portfolio relevant.Benchmark-calibrated counterfactuals illustrate when regret-constrained allocation can improve modeled social loss and when dependence removes that advantage.Source: §8 · p. 6
- Ecosystem-level evaluation requires predictive-risk safeguards and careful treatment of error covariance and externality assumptions.Source: §8 · p. 6
A Proofs and theoretical details
With equal predictive risks and orthogonal errors, expected loss varies with adoption concentration. For two models, marginal diversification depends on the risk gap, externality parameters, and common-error covariance.
A.1 Concentration proposition
- With equal model risks under Assumption 1, expected loss contains a common-error penalty proportional to HHI, minimized by uniform allocation and maximized by monoculture when λρr > 0.The private term is r/2; the common-error contribution is λρrHHI(q)/2.Source: §A.1 · p. 8
A.2 Accuracy-monoculture threshold and edge cases
- For two models, model 2 has higher risk by Δ = r2 − r1, and its adoption share x defines the marginal shift from monoculture.Source: §A.2 · p. 8
- A beneficial marginal shift requires a positive denominator in Equation (5) and λ above the threshold; correlated errors adjust that denominator through σ12.If the denominator is nonpositive, monoculture remains optimal; the loss objective is convex when Σz ⪰ 0.Source: §A.2 · p. 8
B Benchmark details and sensitivity results
The benchmark section reports selected top-three risks, counterfactual reductions, and constrained adoption shares. The accompanying analysis is reproducible through a notebook and specified benchmark inputs.
B Benchmark details and sensitivity results
- Table 2 lists the top-three benchmark risks, measured as mean 1 − R², selected for each crop; values are dimensionless.Source: Table 2 · p. 9
- Table 3 reports counterfactual social-loss reductions relative to lowest-risk monoculture at ρ = .25, λ = 20, and n = 10,000; uniform and unconstrained allocations do not meet the 10% regret cap.Source: Table 3 · p. 9
- Table 4 gives constrained adoption shares for the first, second, and third models listed in Table 2, with the private-regret constraint binding at 10% in each crop.Source: Table 4 · p. 9
C Reproducibility and compute
- The executable notebook loads the registered dataset and leaderboard and runs CPU-based tabular analysis and constrained quadratic optimization.Inference about agronomic welfare and observed cross-model common shocks is outside the analysis scope.Source: C Reproducibility and compute · p. 9