Sustainability Assessment of Coffee By-Product Valorization Using Explainable Neural Networks in Smallholder Food Systems

Coffee by-product valorization offers resource recovery opportunities, although implementation depends on smallholder capacities. This study developed a multidimensional assessment using a multitask neural network. Literature informed the conceptual framework, while the implementation generated 260 producer profiles and reference targets for 12 technologies. Sampling with replacement produced 3600 records containing 2120 distinct producer–technology pairs, rather than independent empirical observations. Environmental, economic, social, and technological scores were constructed through predefined rules, alongside suitability and generated adoption labels. Evaluation separated producer identifiers across partitions. Suitability predictions yielded MAE = 0.0492 and RMSE = 0.0628, equivalent to 4.92 and 6.28 points on a 0–100 scale. R2 = 0.5538 indicated that predictions accounted for 55.38% of reference-score variation, rather than a 55.38% probability of correct prediction. Performance was moderate, leaving substantial variation unaccounted for. Adoption discrimination reached AUC = 0.7087. Technological feasibility was most strongly associated with reference suitability, partly reflecting target construction. For phenolic extraction, budget and technical skills jointly shaped predicted compatibility. These results support exploratory comparison within the computational setting; usefulness remains unverified. Future research should validate outcomes against independent producer records, compare simpler models, and incorporate producer priorities into weighting before using predictions to screen valorization pathways or inform investment decisions.