A Self-Adaptive ML Pipeline for Sustainable Manufacturing

Chapter 27 A Self-Adaptive ML Pipeline for Sustainable Manufacturing Ramon Angosto Artigues , Andrea Fernández Martínez , Andrea Gregores Coto , and Jonathan Josue Torrez Herrera Abstract The European industry twin green and digital transition is critical to foster sustainable manufacturing practices that minimize and prevent the environmental impact and resource exhaustion, while ensuring European global competitiveness. AI can play a pivotal role in this transitioning helping to optimize resources, enhancing efficiency, and reducing waste. Nonetheless, AI adoption in the industry is still limited due to the skills gap that hinders the management of these advanced systems. This paper proposes a self-adaptive pipeline for ML with the incorporation of a new Overfitting Index (OI) for self-parameter tuning emphasizing overfitting prevention. The pipeline incorporates self-parameter exploration capabilities exploiting surrogate models to improve computational efficiency. The proposed OI for ML function is evaluated in a well-known regression problem, prior to its real evaluation in an aluminum recycling application to support operators selecting the best combination of scraps. Results indicate that configurations with lower OIs demonstrate superior generalization and robustness, with the surrogate model effectively identifying and refining high OI configurations. The proposed methodology provides a baseline for future development and integration of self-adaptive and self-improving ML solutions in the industry. Keywords Machine learning ·Self-adaptive ML ·Sustainable manufacturing ·AI in manufacturing ·Overfitting prevention ·Autonomic computing R. Angosto Artigues · A. Fernández Martínez (B) · A. Gregores Coto AIMEN Technology Centre, O Porriño-Pontevedra, Spain e-mail: [email protected] J. J. Torrez Herrera Ibérica de Aleaciones Ligeras (IDALSA), Remolinos (Zaragoza), Spain © The Author(s) 2025 M. Farmanbar et al. (eds.), Horizons of AI: Ethical Considerations and Interdisciplinary Engagements, Frontiers of Artificial Intelligence, Ethics and Multidisciplinary Applications, https://doi.org/10.1007/978-981-96-7945-4_27 437 438 R. Angosto Artigues et al. 27.1 Introduction The manufacturing industry is a key pillar of the European economy, accounting for up to one quarter of European Union’s (EU) business economy net turnover (“Businesses in the manufacturing sector—statistics explained [online]”). Nevertheless, European industry has been heavily impacted in the last years due to several factors, including rising global competition, the COVID-19 pandemic, and geopolitical crises like the Russian aggression on Ukraine. Initiatives such as the European Industrial Strategy (“European industrial strategy—European Commission [online]”) and the Green Deal Industrial Plan (“EUR-Lex - 52023DC0062 - EN - EUR-Lex [online]”) aim at enhancing the competitiveness of t he European industry by supporting the twin green and digital transitioning of the sector. As a result of this trend, the uptake of advanced technologies has been accelerated in manufacturing environments in the last decade (Calza et al. 2024). In parallel with these policy drivers, there is a growing emphasis on the circular economy and closed-loop manufacturing, where r esources are systematically reused and repurposed to reduce both waste and costs. This focus on sustainability amplifies the relevance of innovative data-driven tools that can monitor, predict, and optimize various industrial processes. Consequently, manufacturing companies are increasingly exploring AI to address resource scarcity, energy consumption, and operational risks. Despite the abundant advantages, the adoption of AI can be slowed by uncertainty regarding model robustness and reliability, which is especially critical in sensitive manufacturing processes where errors can lead to operational downtime or suboptimal resource usage. Artificial Intelligence (AI), considered the fourth largest advanced manufacturing technology (Calza et al. 2024), has stand out as key enablers for sustainable manufacturing. The expected benefits of implementing AI in manufacturing rely on its capacity to enhance resource optimization and operational efficiency, while also reducing energy consumption and waste generation (Fernández Martínez et al. 2024; Waltersmann et al. 2021; Willenbacher et al. 2018) which is particularly important in industrial recycling processes, characterized by these intrinsic features. In the metal industry, a crucial sector in the EU economy that represents around 4% of the EU’s GDP (“Economic and steel market outlook 2023–2024”, 2023), the EU leads metal recycling with a recycling rate of 85% for steel and 75% for aluminum, the second most consumed metal product. In terms of energy savings, the latter sector can save up to 95% of the energy required for primary aluminum production (“New life for recycled aluminium—recycling magazine [online]”). However, industries face particular challenges due the variability and uncertainty of the composition and quality of the scraps that are used as primary feedstock, what affects the efficiency of the overall recycling production (Capuzzi and Timelli 2018). AI can help mitigating these issues by improving the characterization and sorting accuracy of scraps (Capuzzi and Timelli 2018), optimizing processing techniques (Auer et al. 2019), and supporting decision-making activities, among others (“Development of decision 27 A Self-Adaptive ML Pipeline for Sustainable Manufacturing 439 support system based on machine learning and digital twin for aluminium melting furnaces”). Nonetheless, the uptake of AI in industrial environments contends its own barriers, including data availability and quality, lack of standardization, and Human-In-TheLoop (HITL) adoption challenges (“AI watch: AI uptake in Manufacturing—European Commission [online]”). Considering the latter point, initiatives for upskilling and training the workforce have proven to be essential to increase the users buyin of the technology (Auer et al. 2019). However, the increasing complexity of managing advanced analytics software commonly hinders its adoption i n manufacturing environments. To help users overcome these issues, the integration of Autonomic Computing (AC) methods to enhance AI self-management and selfadaptation (Fernández Martínez et al. 2024) has emerged as promising solutions, fostering broader acceptance of AI in the sector. Ensuring efficient synergy between operators and AI to exploit the knowledge coming from human expertise. This paper presents a novel pipeline for self-adaptive ML with the incorporation of an Overfitting Index for ML to enable autonomous, surrogate-based parameter tuning, as illustrated in Fig. 27.1. The model is designed to assist operators in estimating the chemical composition of scrap recipes in aluminum recycling processes. The proposed pipeline is part of a broader Decision Support System that combines an AI Data Pipeline and an Autonomic Manager, enabling adaptive model adjustments via the triggering of self-abilities based on human feedback and process metadata. This work specifically emphasizes the importance of robust overfitting prevention in real-world manufacturing scenarios, where challenges such as out-of-distribution data and limited data availability frequently arise. By leveraging self-adaptive abilities, the pipeline enhances model generalization, ensuring reliable performance under dynamic industrial conditions. The methodology is detailed in Section II, key results are discussed in Section III, and the final conclusions are presented in Section IV. Fig. 27.1 General framework with the Surrogate Adaptive Pipeline (SAP) lying at the core of the AI Data Pipeline (AIDP) providing self-adaptive capabilities with the surrogate-based optimization of the Overfitting Index (OI). The Autonomic Manager triggers the self-abilities in the AIDP 440 R. Angosto Artigues et al. 27.2 Methodology 27.2.1 The Challenge of Industrial Recycling Processes in the Secondary Sector Secondary production is the process of recycling or re-processing scrap materials into valuable products, promoting a more sustainable, energy-efficient manufacturing compared to primary production (“Metal Recycling Factsheet”). However, scrap materials are usually hard to characterize due to its heterogeneity in composition and quality, magnified by the lack of information about its sources, what affects the overall efficiency of t he process (Capuzzi and Timelli 2018). In the context of the aluminum sector, the transition from critical raw materials like bauxite (raw ore) to aluminum scrap also reduces the demand for environmental-harmful activities such as mining, thereby contributing to the preservation of natural resources and reducing the environmental impact (“The raw-materials challenge: How t he metals and mining sector will be at the core of enabling the energy transition | McKinsey [online]”). The secondary aluminum-making process generally starts with the reception of scraps and other materials from different sources and vendors. If needed, these incoming materials might undergo different processing techniques to modify their volume and shape in a separate grinding facility to optimize their future usage. After reception, a sampling of incoming materials can help cross-checking the composition of scraps to properly sort them before their storage in the different silos. According to the planned orders, materials are then selected based on their availability, cost, and the chemical composition obtained during the reception sampling. The set of materials selected for the production of aluminum is usually referred to as recipe of materials. To this day, the selection of materials is commonly made by operators based on human expertise. When the initial sampling of materials is available, mass-balance estimations can also be made about the composition of the proposed mixture to support human decision-making, although these samplings bring their own uncertainties due to the lack of representation of the complete lot and/or incomplete information for the proper characterization of scraps due to limited sensing methods. Data-driven approaches step up as alternative solutions to improve current estimations using historical data to learn about patterns and the behavior of scraps from the process itself. Once materials are selected, they are transferred from the silos to the main industrial facility. Materials are measured by specialized equipment and charged in the main furnaces following a specific order attending to their characteristics. In the primary melting stage, scraps are melted to obtain an aluminum mixture with a desired composition according to the orders of clients. Each order is defined by a product code and a quantity, and each product code has an associated norm that constrains the chemical composition allowed according to national or international normative guidelines, e.g., ISO, UNE. Following primary melting, refining operations are frequently made in a secondary melting stage with the incorporation of pure materials (alloys). Chemical analyses are conducted as needed to guarantee that the 27 A Self-Adaptive ML Pipeline for Sustainable Manufacturing 441 final mixture meets the desired norm. The process finalizes with the molding and cooling of the mixture to obtain t he final aluminum products. These products are ready to be sent to the final clients. 27.2.2 A Self-Adaptive ML Pipeline Integrating Overfitting Prevention ML models have proven to be successful enhancing sustainable manufacturing by optimizing the use of resources, energy-efficiency, and reducing waste (Fernández Martínez et al. 2024; Waltersmann et al. 2021; Willenbacher et al. 2018). However, the complexity of managing ML solutions poses a significant challenge for their deployment in real industrial scenarios. Implementing, debugging, and maintaining complex ML architectures call for specialized knowledge, and operators frequently do not have enough time or training to handle advanced ML, thus raising the barrier for successful AI integration. The principles of Autonomic Computing have gained attention in recent years as a means to provide smart solutions with self-adaptive or self-managing capabilities to overcome these challenges. In any self-managing system, four distinct properties can be distinguished, namely self-configuration, self-healing, self-optimization, and selfprotection (Parashar and Hariri 2005). Considering ML, self-configuration is crucial to facilitate the adjustment of models, including the fine-tuning of their (hyper-) parameters. Nonetheless, enabling this property is not trivial, as the system must be able not only to fine-tune the model, but also evaluate the adjustments made to validate t he updated solution before deployment. Overfitting is a common issue in ML where a model learns not only the underlying patterns in the training data, but also the noise, limiting its ability to generalize and most commonly leading to poor performance in new, unseen data (Montesinos López et al. 2022). However, overfitting might not always be apparent when evaluating model performance for multiple reasons. One of these reasons is an insufficient test dataset size, which often arises from limited data availability. An inadequate test set may fail to represent the full complexity of real-world scenarios, leading to misleading performance indicators. Additionally, a test set too similar in distribution with the training set can lead to overfitted models that can misleadingly indicate good performance, even though they might lack generalization capabilities. Another factor contributing to undetected overfitting is relying on a single evaluation metric rather than employing a multi-metric approach. Using multiple metrics broadens the evaluation scope, offering insights into the model’s performance from different perspectives and reducing the risk of drawing incomplete conclusions. I n order to address the overfitting problem, cross-validation (e.g., K-Fold and Hold-Out validation) and regularization techniques (e.g., L2 and L1 regularization) are traditionally incorporated in the training process. However, these methods do not always guarantee 442 R. Angosto Artigues et al. robust performance if the dataset is highly skewed or if certain scrap compositions are underrepresented. The Overfitting Index (OI) for ML. Recent works have introduced the concept of the so-called Overfitting Index, particularly for advanced models such as Neural Networks (NN) (Aburass 2023). The OI proposed in Aburass (2023) focused on the evaluation of t he accuracy loss in both the training and validation datasets, incorporating the epoch number as a weighted measure to emphasize the overfitting occurring in late training epochs. In this work, we propose an Overfitting Index for ML training procedures based on a holistic evaluation of the traditional Mean Absolute Error (MAE). Unlike prior approaches that assess only the final MAE in the validation set, the proposed OI for ML captures the evolving performance across the training and validation datasets throughout the entire training process by analyzing the differences in their MAE values, that is, the gap between the training and validation MAE curves. Additionally, we introduce a novel term to evaluate the impact of slope trends, reflecting the rate of change in error over time. To ensure a comprehensive and interpretable measure of overfitting, the index is scaled by the total number of data samples used during the model learning process. This scaling ensures the proper penalization of overfitting, particularly in later training stages. Hence, the novelty of our OI for ML lies in the incorporation of two penalization factors. The first factor addresses the gap between the training and validation error curves, while the second factor accounts for slope trends, forming a multi-metric evaluation approach for overfitting detection and prevention. The OI for ML is described in Eq. 27.1. OI = k−1 i=1 ((w1 · Etrain,i − Eval,i + w2 · P(i)) · Ni+1) + Eval,k · Nk(27.1) where •Etrain,i and Eval,i represent the negative mean absolute errors of the training and validation datasets at the ith point of the training process, respectively. •P(i) is a penalization factor for the trends in error slopes described in Eq. 27.2. •Ni+1 is the number of samples used in training at the (i + 1)th step. •Eval,k represents the final negative mean absolute error in the validation dataset at the end of the training process. •Nk represents the total number of samples in the dataset used by the learning curve. •w1 and w2 are weights assigned to the penalization terms. The first term of the equation, w1 · Etrain,i − Eval,i , addresses the discrepancy between the training and validation errors at each step, evaluating in this way the gap between both learning curves. The second term is w2 · P(i) where P(i) is described in more detail in Eq. 27.2, undesirable trends in the error slopes are penalized. Both terms are scaled by the number of samples at the next step Ni+1, reflecting the impact of data volume in the model’s learning progress. The third term refers to the 27 A Self-Adaptive ML Pipeline for Sustainable Manufacturing 443 commonly used validation error at the end of the training process, scaled by the total number of samples to ensure a similar scale with respect to the other two terms. The weights w1 and w2 allow for flexibility in emphasizing t he relevance of performance discrepancy versus slope behaviors versus final validation error. The penalization factor P(i) that adjusts the trends in error slopes is described in Eq. 27.2. Pi = Si · log(1 + Ni+1) · Fi(27.2) where •Si measures the absolute difference between the slopes of the validation and training error curves at the ith point of the training process, that is, slopetrain,i − slopeval,i . •The logarithmic term log(1 + Ni+1) scales the penalty based on the size of the training dataset. •Fi is a conditional penalty factor. The difference between slopes, Si, emphasizes how quickly the errors are changing as new data is introduced in the model. On the other hand, the logarithmic term, log(1 + Ni+1), reinforces the impact of larger datasets in the penalization. The conditional penalty factor Fi adjusts the penalty based on specific conditions observed in the slope trends: •Fi = 20 if both training and validation slopes are positive and the training slope exceeds the validation slope, indicating severe overfitting. •Fi = 10 if the training slope is positive but the validation slope is negative, suggesting the model is learning noise or irrelevant patterns. •Fi = 2 if only the validation slope is positive, indicating rising errors in validation, a milder form of overfitting. Surrogate Models for Optimizing Parameter Exploration. Surrogate models, also known as metamodels, are simplified models that approximate more complex underlying functions of ML models. They are particularly useful in scenarios where evaluation of the original model is computationally expensive or time-consuming. By fitting a surrogate model to the data derived from the original model’s performance across various configurations, it is possible to efficiently explore parameter spaces and predict the performance of new configurations with significantly reduced computational overhead. Incorporating surrogate models into the fine-tuning process enables a more efficient and systematic approach to identifying optimal model adjustments. By leveraging the OI as the objective function, the fine-tuning approach accelerates the identification of optimal model adjustments while minimizing computational and resource costs (Barwey et al. 2023). The proposed methodology supports a selfadaptive pipeline capable of autonomously adjusting model parameters in response to changes in data or performance trends observed in real-world scenarios. This adaptability ensures continuous model optimization with minimal to no user intervention, enhancing both efficiency and robustness. 444 R. Angosto Artigues et al. Fig. 27.2 Model pipeline component of the self-adaptive pipeline (SAP) for machine learning The Self-Adaptive Pipeline (SAP) for ML. The SAP proposed in this work is composed of three main sub-pipelines components. At the core of SAP, a Model Pipeline was defined to pre-process raw data via a Column Transformer component followed by the definition of a Multi-Output Regressor (MOR) model as shown in Fig. 27.2. Column Transformer accounts to the type of data by encoding categorical variables with methods like one-hot encoder, while scaling numerical inputs to ensure uniformity between parameters. The MOR is wrapped around a Random Forest Regressor model capable of predicting multiple target variables from a set of inputs. This setup i s essential for handling the complex relationships often found in material composition data. The second block of the SAP corresponds to the Surrogate Model Optimization Pipeline, shown in the bottom section of Fig. 27.3, which is essential t o reduce computational overhead during the hyperparameter tuning phase. In this stage, the objective function previously defined and the parameter space for fine-tuning are fed to a loop in which a Bayesian optimization is conducted using the OI for ML as key metric. Following this approach, the search of hyperparameters is guided toward configurations that balance model accuracy and generalization. Each iteration of the optimization involves evaluating potential hyperparameters by running the surrogate model, which predicts the OI based on historical data from previous evaluations. Based on this procedure, it is ensured that only the most promising configurations are fully evaluated using the more computationally intensive primary model, thereby streamlining the entire model training process. In the last stage of the SAP, the best set of hyperparameters identified is selected to train the primary model on the full dataset to verify the performance metrics suggested by the surrogate model, as shown in the main pipeline at the upper section of Fig. 27.3. 27 A Self-Adaptive ML Pipeline for Sustainable Manufacturing 445 Fig. 27.3 Workflow of the surrogate model optimization pipeline in the self-adaptive pipeline (SAP) for machine learning 27.2.3 A Metadata-Driven Autonomic Manager to Trigger Self-abilities on an AI Data Pipeline The proposed methodology encapsulates a self-adaptive mechanism whereby the model can dynamically adjust its parameters in response to shifts in data characteristics or performance objectives. This SAP is part of a larger AI Data Pipeline (AIDP) that comprises five components to enable the development and deployment of ML models in real industrial applications (Artigues et al. 2024). These components include Data Ingestion, Data Transformation, Data Exploration, Model Training, and Real-World Usage. Throughout these stages of the AIDP, a dedicated Metadata Logging module records important statistics. Metadata, as its name suggests, corresponds to data from data as a higher-level of abstraction of the components per se. Some of this metadata include human feedback about the performance of the ML models and important statistical information of new data (e.g., data distribution shifts, user overrides, or anomalies in daily scrap deliveries). An external Autonomic Manager component monitors the metadata from the AIDP, analyzes it, plans about the corrective measurements needed, and executes them back in the AIDP through the activation of alarms and triggering of the self-abilities following the MAPE-K architecture. Because this process is event-driven, it allows the entire pipeline to remain dormant when everything is stable, thus avoiding unnecessary computations and system overhead. 452 R. Angosto Artigues et al. of configurations with higher OIs through this surrogate model approach add significant value to t he tuning process, optimizing model performance more efficiently and ensuring its robustness under varying operational conditions. As part of a bigger framework that triggers the activation of autonomous abilities based on real-time human feedback and monitored data, this methodology sets a new standard in machine learning optimization. Its ability to efficiently recommend model adjustments ensures optimal performance while reducing reliance on extensive manual intervention. The approach not only advances ML optimization but also offers a scalable solution for industries seeking to streamline their ML model development lifecycle. Further refinements and enhancements to the Overfitting Index methodology will be explored in future work to optimize its applicability and efficiency in diverse modeling scenarios. Refinements to the OI metric—such as adaptive penalization factors that become stricter over time—could further shield production lines from performance degradation in scenarios of rapid data drift. Besides, integrating these solutions into broader enterprise systems remains a fruitful direction for future research. Additional investigation into multi-task learning, domain adaptation, or advanced sampling for underrepresented scrap types can further enrich the pipeline’s applicability in the future. Acknowledgements This work has been supported by the project ‘self-X Artificial Intelligence for European Process Industry digital transformation’ (s-X-AIPI), which has received funding from the European Union’s Horizon Europe research and innovation program under Grant Agreement No. 101058715. References 7.2. Real world datasets—scikit-learn 1.5.1 documentation [online]. Available: https://scikit-learn. org/stable/datasets/real_world.html#california-housing-dataset. Accessed 6 Aug 2024 Aburass S (2023, Aug) Quantifying overfitting: introducing the Overfitting Index [online]. Available: https://arxiv.org/abs/2308.08682v1.Accessed 5 Aug 2024 AI watch: AI uptake in Manufacturing—European Commission [online]. Available: https://aiwatch.ec.europa.eu/publications/ai-watch-ai-uptake-manufacturing_en. Accessed: 7 Aug 2024 Artigues R, Coto A, Herrera J, Tomás F, Verardi S, Marzano M, Martinez A (2024) An AI-driven user-centric framework reinforced by autonomic computing: A case study in the Aluminium sector. In: Tareq Ahram, Luca Casarotto and Pietro Costa (eds) Human interaction and emerging technologies (IHIET 2024). AHFE (2024) International Conference. AHFE Open Access, vol 1. AHFE International, USA. https://doi.org/10.54941/ahfe1005478 Auer M, Osswald K, Volz R, Woidasky J (2019, Mar) Artificial intelligence-based process for metal scrap sorting [online]. Available: https://arxiv.org/abs/1903.09415v1. Accessed 5 Aug 2024 Barwey S, Kim H, Maulik R (2023, Nov). Interpretable fine-tuning and error indication for graph neural network surrogate models [online]. Available: https://arxiv.org/abs/2311.07548v3. Accessed 5 Aug 2024 Businesses in the manufacturing sector—statistics explained [online]. Available: https://ec.eur opa.eu/eurostat/statistics-explained/index.php?title=Businesses_in_the_manufacturing_sector. Accessed 15 Jul 2024 27 A Self-Adaptive ML Pipeline for Sustainable Manufacturing 453 California housing price predictions [online]. Available: https://www.kaggle.com/code/eswarc handt/california-housing-price-predictions. Accessed 6 Aug 2024 Calza E, Soguero EJ, Fabiani J, De PG (2024, Jun) Advanced manufacturing study. Preliminary findings on EU’s advanced manufacturing industry in the global landscape. Jun, https://doi.org/ 10.2760/798090 Capuzzi S, Timelli G (2018) Apr) Preparation and melting of scrap in aluminum recycling: a review. Metals 8(4):249. https://doi.org/10.3390/MET8040249 Development of decision support system based on machine learning and digital twin for aluminium melting furnaces. https://doi.org/10.5281/ZENODO.5897240 Economic and steel market outlook 2023–2024 First quarter report Data up to, and including, third quarter 2022 Introduction, 2023 EUR-Lex - 52023DC0062 - EN - EUR-Lex [online]. Available: https://eur-lex.europa.eu/legal-con tent/EN/TXT/?uri=CELEX%3A52023DC0062. Accessed 5 Aug 2024 European industrial strategy—European Commission [online]. Available: https://commission.eur opa.eu/strategy-and-policy/priorities-2019-2024/europe-fit-digital-age/european-industrial-str ategy_en. Accessed 7 Aug 2024 Fernández Martínez A et al (2024) A machine learning framework for improving resources, process, and energy efficiency towards a sustainable steel industry. LNNS 1028:3–10. https://doi.org/10. 1007/978-3-031-61905-2_1 Montesinos López OA, Montesinos López A, Crossa J (2022) Overfitting, model tuning, and evaluation of prediction performance. Multivariate Statistical Machine Learning Methods for Genomic Prediction, pp 109–139. https://doi.org/10.1007/978-3-030-89010-0_4 New life for recycled aluminium—recycling magazine [online]. Available: https://www.recyclingmagazine.com/2021/09/02/new-life-for-recycled-aluminium/. Accessed 5 Aug 2024 “Metal Recycling Factsheet”. Parashar M, Hariri S (2005) Autonomic computing: an overview. LNCS 3566:257–269. https://doi. org/10.1007/11527800_20 The raw-materials challenge: How the metals and mining sector will be at the core of enabling the energy transition | McKinsey [online]. Available: https://www.mckinsey.com/industries/met als-and-mining/our-insights/the-raw-materials-challenge-how-the-metals-and-mining-sectorwill-be-at-the-core-of-enabling-the-energy-transition. Accessed 5 Aug 2024 Waltersmann L, Kiemel S, Stuhlsatz J, Sauer A, Miehe R (2021) Jun) Artificial intelligence applications for increasing resource efficiency in manufacturing companies—A comprehensive review. Sustainability 13(12):6689. https://doi.org/10.3390/SU13126689 Willenbacher M, Kunisch C, Wohlgemuth V (2018) Application of methods of artificial intelligence for sustainable production of manufacturing companies, pp 225–236. https://doi.org/10.1007/ 978-3-319-65687-8_20 XGBoost for the California housing dataset | XGBoosting [online]. Available: https://xgboosting. com/xgboost-for-the-california-housing-dataset/. Accessed 7 Aug 2024 454 R. Angosto Artigues et al. Open Access This chapter is licensed under the terms of the Creative Commons AttributionNonCommercial-NoDerivatives 4.0 International License (http://creativecommons.org/licenses/bync-nd/4.0/), which permits any noncommercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license and indicate if you modified the licensed material. You do not have permission under this license to share adapted material derived from this chapter or parts of it. The images or other third party material in this chapter are included in the chapter’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the chapter’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.