Source-linked AI summary
Augmenting Organizational Decision-Making with Deep Learning Algorithms: Principles, Promises, and Challenges
Yash Raj Shrestha, Vaibhav Krishna, Georg von Krogh
TL;DR
The paper addresses limited understanding of how deep learning algorithms function and may shape organizational decision-making. It explains deep learning’s algorithmic engine and examines its application through two case studies, finding advantages alongside a need for managerial understanding, reflection, and prudence.
Problem
Scholars need to better understand deep learning algorithms’ functioning and inner workings to examine how they may shape organizational decision-making.
Method
The paper conceptualizes DLADM, explains deep learning principles and algorithmic components, and presents two case studies using image recognition and sentiment analysis on organizationally relevant datasets.
Results
The paper finds advantages of DLADM for firms, with deep learning outperforming traditional machine learning in both image classification and sentiment analysis tasks.
Takeaways & Limitations
Implementing deep learning in organizations demands significant managerial understanding, reflection, and prudence, including attention to ethical issues of opacity and bias.
Takeaways & Limitations
The case studies use only two datasets and are not representative of general DLADM applications.
Abstract
from arXiv · showhide
The current expansion of theory and research on artificial intelligence in management and organization studies has revitalized the theory and research on decision-making in organizations. In particular, recent advances in deep learning (DL) algorithms promise benefits for decision-making within organizations, such as assisting employees with information processing, thereby augment their analytical capabilities and perhaps help their transition to more creative work.
1. Introduction
Organizational decision-making has long been understood as an information-processing problem, and contemporary AI and DL extend this technological support. The paper examines DLADM’s potential for exploiting unstructured data while emphasizing substantial implementation and organizational challenges.
- Organizational decision-making research has long examined how structures, processes, and technologies facilitate or constrain information processing and outcomes.
- Contemporary AI differs from earlier systems through learning, autonomous action, and access to big data and advanced computing capacity.
- The paper opens DL’s algorithmic “black box,” explains its core mechanisms, and examines managerial decision-making through Zalando and Rotten Tomatoes case studies.
- DL algorithms are well suited to extracting patterns and making predictions from unstructured data, which constitutes 80% of existing marketing data.
- DLADM can assist information processing and decision-making, but deployment requires costly data, annotation, computing infrastructure, and safeguards for fairness, accountability, reliability, and trust.
2. DLADM in practice
DLADM is deployed across organizational functions to target customers, monitor environments, detect risks, and support manufacturing scheduling. These applications use digital traces, images, text, financial data, and sensor data to augment managerial decisions.
- DL deployment combines machine-learning procedures with business understanding to address organizational problems.
- 2.1. Targeting: Marketing applications use digital traces, including text, images, login information, and GPS data, to identify customer groups and support product and advertising decisions.
- 2.1. Targeting: Image-based DL applications in fashion include apparel segmentation, recognition, classification, retrieval, and tagging for strategic and marketing decisions.
- 2.2. Monitoring: Real-time monitoring supports proactive forecasting and preventive fraud detection by extracting information from financial, social-media, and environmental data.
- Manufacturing prognostic systems use DL with sensor-generated data for scheduling, resource allocation, and planning when manual analysis becomes infeasible.
- Across functions, DL can augment decision-making through enhanced targeting, monitoring, and scheduling capacity.
3. A brief overview of Machine learning and deep learning subfield
The paper introduces machine learning as an AI subdomain and distinguishes supervised, unsupervised, and reinforcement learning according to data and task goals. It then explains neural-network foundations, deep architectures, and specialized designs for processing complex or unstructured data.
- Overview of ML: Machine learning enables computers to learn from datasets without being explicitly programmed and is categorized as supervised, unsupervised, or reinforcement learning.The categories differ according to the learning goal and whether target variables or environmental interaction are used.
- Overview of ML: Supervised learning maps input features X to targets Y, supporting classification for categorical targets and regression for continuous targets.Fraud detection illustrates classification using labeled historical transactions.
- Overview of ML: Unsupervised learning discovers patterns in input features without a target variable, commonly through clustering or dimensionality reduction.Customer segmentation based on shopping behavior is a representative application.
- Overview of ML: Reinforcement learning determines actions through interaction with an environment to achieve goals, supporting applications such as autonomous systems and game-playing algorithms.The approach is illustrated through robots, cars, and algorithms for Go, poker, and chess.
- Neural-network foundations: Neural networks compute weighted sums of input features and use activation functions to introduce nonlinearity, enabling approximation of complex nonlinear functions.Neurons are connected units organized in layers, while activation functions provide the nonlinear transformation.
- Deep-learning architectures: Deep feedforward networks or multilayered perceptrons use multiple hidden layers to learn complex higher-order features, while specialized architectures address unstructured data.CNNs, RNNs, and GANs were developed for data such as images, videos, and natural language; this paper focuses on supervised learning and unstructured business data.
4. Case studies
The case studies examine DL applications for fashion-image classification and textual sentiment analysis, finding higher accuracy than traditional ML in both tasks and outlining managerial uses for the resulting insights.
- Case studies: The case studies apply DL to image recognition and textual sentiment analysis using fashion-product and movie-review datasets.The fashion case uses Fashion-MNIST from Zalando; the sentiment case uses a Rotten Tomatoes dataset.
- 4.1. Case study 1: Fashion image classification: Fashion-MNIST contains 60,000 training examples and 10,000 test examples, with each example represented as a 28 × 28 grayscale image.Each image contains 784 pixels and belongs to one of ten classes.
- 4.1.2. Results: CNN-based DL reached 92.48% accuracy versus 84% for traditional ML in fashion-image classification, increasing to 93.98% with pretrained ResNet18.The study also reports precision and recall scores and notes that random guessing would yield 10% accuracy across ten classes.
- 4.1.2. Results: Fashion-image detection can inform customer targeting, trend and seasonality analysis, product feedback, and new-product design decisions.The proposed uses include segmenting apparel by patterns, colors, and shapes and linking those segments to buyer demographics and age groups.
- 4.2. Case study 2: Textual sentiment analysis: The textual case evaluates traditional ML and DL with tf-idf embeddings, using an MLP and pretrained BERT for DL.The dataset consists of sampled Rotten Tomatoes sentences labeled with positive or negative polarity.
- 4.2.2. Results: DL achieved 90.4 accuracy versus 79.24 for traditional ML in the sentiment-analysis task.The resulting sentiment patterns could support market-trend analysis, product fixes, and product-development strategy.
5. DLADM challenges and recommendations for managers
DLADM offers decision-support benefits but requires managers to address substantial costs, skills constraints, probabilistic errors, bias, opacity, and privacy risks. The paper recommends tracking techniques that can reduce data, computing, and expertise requirements while developing governance and affordable algorithms.
- Economic and organizational challenges: Cutting-edge DL requires costly annotated data, specialized computing hardware, and substantial data-science capability.Managers may need powerful hardware or cloud resources, large datasets, data-quality safeguards, and specialized skills.
- Recommendations for managers: Transfer learning can accelerate training while reducing data requirements and computational power demands.The authors used transfer learning in both case studies and recommend that managers monitor developments that mitigate implementation costs.
- Economic and organizational challenges: DL's probabilistic nature makes errors possible, requiring contingency planning, failure-risk assessment, governance mechanisms, and human judgment for difficult cases.A human-in-the-loop structure can direct challenging cases to managers to help minimize reputational and other organizational risks.
- Economic and organizational challenges: DL can rapidly reinforce training-data bias, creating detrimental consequences for minorities and introducing new sources of bias in hiring.The passage states that domain expertise can complement DL in mitigating hiring bias.
- Economic and organizational challenges: DL opacity complicates explanations of decisions and creates challenges for organizational trust and accountability.The paper notes that satisfactory solutions to opacity remain unavailable despite ongoing research on interpretable, explainable, and fair AI.
- Economic and organizational challenges: Datafication and outsourcing can create information-privacy, security, control, and surveillance concerns.These concerns are especially salient when firms outsource DLADM elements or run models in the cloud.
- Recommendations for managers: Academic research should develop more financially friendly DL algorithms and multidisciplinary frameworks for organizational challenges.The paper calls for collaboration across diverse academic domains to address these challenges.
6. Discussion
The paper frames DLADM as a way to support organizational information processing, conceptualizes its algorithmic engine through two small-scale case studies, and concludes that implementation requires managerial prudence. It also identifies limited case-study coverage and calls for broader, causal research on where DL improves decision quality.
- Discussion: Organizational scholarship has long examined how technology can support the information processing required for decision-making.The discussion places DLADM within a longstanding organization-theory problem concerning the flow and processing of information.
- Discussion: The paper conceptualizes DLADM, presents core DL principles and algorithmic components, and illustrates them through two small-scale case studies.This describes the paper's central conceptual and illustrative approach.
- Discussion: Despite DLADM's advantages for firms, implementation demands significant managerial understanding, reflection, and prudence.The paper also highlights future work on integrating DL into organizational structures and processes and addressing opacity and bias.
- Discussion: The two case studies use only two datasets and are not representative of general DLADM applications.The authors encourage expanded case studies and randomized controlled trials examining whether DL use improves decision quality.