Machine Learning

Machine Learning

Machine Learning — Predictive Modelling and Data-Driven Insight

Introduction and Value of Machine Learning

Machine Learning (ML) is the technology that allows data to generate useful and actionable knowledge. Unlike systems based on fixed rules, ML learns patterns directly from real data and builds models capable of predicting, classifying or recommending automatically.

It is an iterative process that evolves over time and adapts to changing environments.

At amsafis, Machine Learning is one of the three core pillars — together with Expert Systems and RAG-LLM — and acts as a natural complement to both:

  • it generates reliable predictions,
  • identifies non-trivial relationships,
  • and provides a solid mathematical foundation that can feed expert rules, diagnostic systems or inference engines.

A key element is the treatment of nonlinear models, which allow capturing complex interactions that cannot be expressed with simple equations. Much of the relevant information in real-world data — healthcare, industrial processes, finance, quality control or IoT — is intrinsically nonlinear, and ML is the appropriate tool to extract its full value.

Machine Learning enables:

  • prediction of future behaviour,
  • robust classification,
  • anomaly detection,
  • segmentation and discovery of hidden patterns,
  • process optimisation,
  • intelligent automation,
  • decision support.

ML is especially valuable for small and medium-sized organisations, where data are often heterogeneous, limited or imperfect. Through flexible and nonlinear techniques — such as GAM, multilevel models, additive regressions, functional data analysis, dependence methods, Möbius integrals or neural networks — it is possible to obtain reliable insight and solid prediction even in complex environments.

This pillar complements RAG-LLM (for understanding or generating content) and Expert Systems (for formalising rules and knowledge), closing a complete cycle from data to knowledge, and from knowledge to action.

Approaches, Methods and Advanced Models

Machine Learning integrates a wide spectrum of methodologies, ranging from classical statistical models to advanced optimisation and deep learning approaches. This diversity allows addressing very different problems, from numerical prediction to pattern mining or structural modelling.

Nonlinear Models and Flexible Learning

Most real phenomena are nonlinear. ML incorporates methods capable of capturing curvature, interactions and complex effects.

Generalized Additive Models (GAM)

Models curved relationships in an interpretable way, ideal for capturing nonlinear dependencies without sacrificing transparency.

Multilevel (Hierarchical) Models

Integrate variability between groups, clients, sensors or centres, enabling prediction while respecting the real structure of the data.

Functional Data Analysis

Designed for data represented as functions, curves or continuous signals — such as movement trajectories, temporal profiles or biomedical signals.

Differential Equations and Dynamical Systems

Particularly useful when temporal evolution is governed by internal dynamics, such as biological, physical or industrial processes.

Association and Pattern Mining

Another fundamental block is the discovery of unsupervised patterns and relationships.

Association Rules

Identify combinations of indicators, products or behaviours that appear together.
They are widely used in markets, diagnostics, industrial processes and customer-pattern detection.

These rules connect naturally with Expert Systems, where they can be transformed into explicit decision logic.

In healthcare, this type of analysis also enables the identification of associations between genetic markers and adverse reactions. An example is the study of SNPs related to Penicillin Allergy:

➡️ Example: Association rules identifying genetic patterns linked to penicillin allergy.

Nonlinear Integrals, Capacity and Möbius Theory

Advanced methods based on capacity theory and non-additive aggregation:

Choquet, Sugeno and Möbius Integrals

Powerful tools for nonlinear aggregation in situations where variables interact strongly and cannot be combined using simple sums or weights.

These techniques are especially useful in quality analysis, risk modelling, complex data mining and decision-making.

Neural Networks and Advanced Optimisation

Deep learning enables modelling highly complex structures:

  • neural networks for nonlinear prediction,
  • optimisation,
  • anomaly detection,
  • time series modelling.

They can be integrated with Structural Equation Models (SEM) when it is necessary to identify relationships between variables while fitting nonlinear models through numerical optimisation.

This combination is particularly effective when:

  • causal dependencies exist,
  • relationships are nonlinear,
  • latent structures are present,
  • and a robust yet flexible model is required.

Relation with Expert Systems and RAG-LLM

Machine Learning can feed:

  • automatically derived expert rules,
  • inference engines,
  • hybrid models combining prediction (ML), rules (Expert Systems) and structured knowledge (RAG-LLM).

This integration enables the construction of complete decision platforms for business, healthcare, research or IoT, including incident prediction, anomaly detection and automated decision-making.