Source-linked AI summary
Machine learning and deep learning
Christian Janiesch, Patrick Zschech, Kai Heinrich
TL;DR
The paper addresses the need for a fundamental understanding of machine learning and deep learning, including their implementation challenges in electronic markets. It synthesizes their concepts and automated model-building process, concluding that future intelligent systems and AI-as-a-service ecosystems require structured guidance and attention to responsible deployment.
Problem
Electronic-market practitioners and scholars need stronger understanding of machine-learning and deep-learning concepts, processes, and implementation challenges beyond technological hype.
Method
The article provides a broad conceptual introduction to machine learning and deep learning and organizes automated analytical model building around data input, feature extraction, model building, and model assessment.
Results
The paper identifies four fundamental challenges for implementing intelligent systems based on machine learning and deep learning in real-world electronic-market ecosystems.
Takeaways & Limitations
Future electronic-market research and AI-as-a-service ecosystems should address these challenges through structured guidance for data, model performance, and prediction assessment.
Takeaways & Limitations
Responsible transfer-learning guidelines have yet to be established as AI-as-a-service markets are still emerging.
Abstract
from arXiv · showhide
Today, intelligent systems that offer artificial intelligence capabilities often rely on machine learning. Machine learning describes the capacity of systems to learn from problem-specific training data to automate the process of analytical model building and solve associated tasks. Deep learning is a machine learning concept based on artificial neural networks. For many applications, deep learning models outperform shallow machine learning models and traditional data analysis approaches. In this article, we summarize the fundamentals of machine learning and deep learning to generate a broader understanding of the methodical underpinning of current intelligent systems. In particular, we provide a conceptual distinction between relevant terms and concepts, explain the process of automated analytical model building through machine learning and deep learning, and discuss the challenges that arise when implementing such intelligent systems in the field of electronic markets and networked business. These naturally go beyond technological aspects and highlight issues in human-machine interaction and artificial intelligence servitization.
Introduction
Machine learning and deep learning underpin contemporary intelligent systems by automating analytical model building from data. The article introduces their fundamentals and implementation challenges in electronic markets.
- Introduction: Machine learning automatically learns patterns and relationships from examples and observations instead of relying solely on explicitly programmed rules.This approach reduces the need to formalize human knowledge into machine-accessible form.
- Introduction: Intelligent systems use analytical models to generate predictions, rules, answers, recommendations, and similar outcomes.These models support advanced problem solving associated with artificial intelligence.
- Introduction: Deep learning emerged from artificial neural networks as increasingly deep architectures with improved learning capabilities.For specific applications in closed environments, deep learning can achieve superhuman performance.
- Introduction: The article aims to build a fundamental understanding of machine learning and deep learning for analytical model building in electronic markets.It focuses on concepts, processes, and implementation challenges while excluding adoption, policy, and organizational-culture issues.
Conceptual distinction
The paper distinguishes artificial intelligence, machine learning, artificial neural networks, shallow machine learning, and deep learning as related but non-identical concepts. Their relationships involve differences in learning architecture, interpretability, and data suitability.
- Conceptual distinction: The Venn diagram summarizes the hierarchical relationship among machine learning concepts, artificial neural networks, and deep neural networks.The paper notes that advanced machine-learning algorithms are often black boxes unless explained otherwise.
- Conceptual distinction: Artificial intelligence encompasses techniques that mimic or exceed human decision-making for complex tasks with minimal human intervention.Its methods include knowledge-based, rule-based, evolutionary, fuzzy, and multi-agent approaches.
- Conceptual distinction: Machine learning automates analytical model building by iteratively learning from problem-specific training data to perform cognitive tasks.Supervised, unsupervised, and reinforcement learning are distinguished according to the problem and available data.
- Conceptual distinction: Deep learning uses deeply nested neural-network architectures, typically with more than one hidden layer and advanced operations such as convolutions.These characteristics allow raw input data to support automatic representation discovery.
- Conceptual distinction: Shallow machine learning includes simpler neural networks and other algorithms such as decision trees that lack deep-learning functionalities.The literature does not provide an exact demarcation between shallow machine learning and deep learning.
- Conceptual distinction: Deep neural networks generally outperform shallow machine-learning algorithms for large, high-dimensional text, image, video, speech, and audio data.For low-dimensional data with limited training data, shallow machine learning can produce superior and more interpretable results.
Process of analytical model building
The paper frames analytical model building as a process for explicit programming, shallow machine learning, and deep learning. For machine-learning approaches, it emphasizes data input, feature extraction, model building, and model assessment.
- Process of analytical model building: Analytical model building is organized around four aspects: data input, feature extraction, model building, and model assessment.The framework compares explicit programming, shallow machine learning, and deep learning as distinct model-building concepts.
Data input
Electronic markets generate diverse, large-scale data through websites, apps, and social media, creating opportunities for decision support and business automation. Its heterogeneity and scale make algorithmic support necessary.
- Data input: Electronic-market touchpoints generate numerical, unstructured, and non-cross-sectional data, including time series, images, and text.These data assets can support analytical model building for decision support and business automation.
- Data input: Algorithmic support is indispensable when data volume and dimensionality exceed the cognitive capacity of human operators.The challenge is extracting patterns and relationships from large data assets.
- Data input: Time series data contains sequential dependencies and temporal patterns used for forecasting, regression, and trend-classification tasks.Examples include financial-market forecasting and process-behavior prediction.
- Data input: Cross-modal learning combines different data types, such as product images, descriptions, and text metadata, for retrieval and recommendation tasks.It is useful when content has multiple complementary forms of representation.
Feature extraction
Feature extraction converts raw data into representations suitable for analytical model building. Shallow ML depends heavily on manually designed features, whereas deep learning can learn discriminative representations with less human effort.
- Feature extraction: Features are properties derived from raw input to provide a suitable representation for the learning task.Feature extraction aims to preserve discriminatory information and separate relevant factors of variation.
- Feature extraction: Shallow ML performance depends on successfully extracting well-defined features from the data.Applicable techniques vary by data type, including time-domain and frequency-domain features for time series and HOG for images.
- Feature extraction: Manual feature design is time-consuming, labor-intensive, inflexible, and dependent on substantial domain expertise.These constraints arise because application-specific engineering is usually required.
- Feature extraction: Deep neural networks automate feature learning, extracting hierarchical discriminative representations from large-scale, noisy, and unstructured data.Higher-level abstract features are assembled from simpler ones with minimal human effort.
Model building
Model building uses input data and learning algorithms to identify patterns relevant to a task. Deep learning can combine feature learning and model building end to end, or provide representations for other learning systems.
- Model building: Learning algorithms use input data to identify patterns and relationships relevant to the specified learning task.Shallow ML performs this process using well-designed features.
- Model building: Deep learning directly operates on high-dimensional raw input through automated feature learning and model building.Its architectures are often organized as end-to-end pipelines combining both aspects.
- Model building: Deep learning feature representations can also be fed into decision trees, SVMs, or other learning subsystems.This allows other algorithms to exploit the learned representation.
- Model building: Deep learning architectures include CNNs, RNNs, distributed representations, autoencoders, and GANs.Architectures differ in their layers, neural units, and connections, with some better suited to particular data types.
Model assessment
Model assessment considers performance alongside computational resources and interpretability. Evaluation and model selection require task-appropriate metrics, out-of-sample assessment, and comparison of alternative complexities while accounting for overfitting risk.
- Model assessment: Model quality should be assessed using performance, computational resources, and interpretability.For supervised learning, k-fold cross-validation is commonly used to reduce overfitting and estimate out-of-sample performance.
- Model assessment: Classification models are evaluated with accuracy, recall, precision, F1 score, and cost-sensitive measures when error costs are asymmetric.Average cost per predicted observation is useful when different prediction errors have different consequences.
- Model assessment: Suitable prediction models should be selected by comparing alternative model classes and variants with different complexities.Complexity can reflect learning mechanisms, feature construction, and the number of trainable parameters.
- Model assessment: Simpler models may miss relevant nonlinear patterns, while overly complex models increase overfitting risk and reduce reasoning difficulty.Model complexity therefore involves a trade-off between flexibility and generalization.
Challenges for intelligent systems based on machine learning and deep learning
Implementing ML and DL systems in electronic markets involves many design choices, limited guidance, resource trade-offs, and uncertainty about real-world applicability. Business success depends on more than laboratory accuracy, including operational constraints and consistent comparison.
- Challenges for intelligent systems based on machine learning and deep learning: Model development offers nearly endless choices of algorithms or architectures, hyperparameters, and training data.Established guidance for ensuring performance, cost-efficiency, robustness, and privacy remains limited.
- Challenges for intelligent systems based on machine learning and deep learning: Business environments with limited resources require trade-offs such as prediction quality versus computational costs.The relevant balance depends on the application context.
- Challenges for intelligent systems based on machine learning and deep learning: A model with 99.9% accuracy can be unusable when its classification decision arrives too slowly for time-critical applications.The passage gives proactive monitoring and smart-factory quality assurance as examples.
- Challenges for intelligent systems based on machine learning and deep learning: Excellent laboratory or cross-dataset accuracy may not translate into business success in real-world electronic-market environments.Other application-specific factors may outweigh theoretical ML achievements.
Awareness of bias and drift in data
ML and DL models can inherit and amplify biases in human-generated data, while changing relationships between inputs and targets can undermine recommendations over time.
- Human-generated data can introduce cognitive biases that shallow ML and DL models heavily adopt or amplify.Risk increases when training data is poorly selected or pre-processed, class-imbalanced, or reviewed irresponsibly.
- Concept drift occurs when the relationship between input data and the target variable changes over time.Historical data may no longer represent current conditions, including new competitors, production capabilities, or governmental restrictions.
- Concept drift may be incremental, gradual, or recurring, making it difficult to detect.Automated learning techniques using trusted data windows and concept descriptions exist, but discovering and solving business-related problems remains challenging.
Unpredictability of predictions and the need for explainability
Black-box ML and DL models can make predictions difficult to anticipate or understand, motivating explainability methods and context-sensitive model design in electronic markets.
- The complexity of DL and some shallow ML models makes their performance difficult to predict in specific contexts.Users may also struggle to review or understand recommendations produced by these models.
- Explainability may be required by law because users prefer simple explanations when deciding whether to trust and adopt a model.The paper links explainability requirements to both human preferences and legal enforcement.
- Explainable AI augments existing DL models by producing explanations for output predictions.Methods highlight influential image regions, time steps, or words depending on the data type.
- Electronic-market applications should be designed differently according to their criticality and human-interaction requirements.Designers may select an explainable white-box model over a more accurate black-box model or add XAI augmentations.
- Transfer learning specializes models trained on general datasets for specific tasks using considerably smaller problem-specific datasets.AIaaS marketplaces provide cloud applications, platforms, infrastructure, and access to pre-trained models for organizations with limited development resources.
Conclusion
The article introduces ML and DL fundamentals, explains automated analytical model building, and discusses challenges for intelligent systems in electronic markets. It also identifies AIaaS as an emerging market context for future research and methodological guidance.
- The article provides a broad introduction to ML and DL as technologies underlying contemporary and future intelligent systems.
- It conceptualizes ML, shallow ML, and DL, including their algorithms and architectures.
- Automated analytical model building is described through data input, feature extraction, model building, and model assessment.
- The article discusses four fundamental challenges for ML- and DL-based intelligent systems in real-world electronic-market ecosystems.
- AIaaS is presented as a new and unexplored electronic market expected to influence established service platforms.Future research is proposed around applications, roles, business models, and methodological guidance for safe and accessible predictions.