💻 computer science

A hybrid PLS-SEM and convolutional neural network analysis of artificial intelligence adoption intention in SMEs

This study introduces a unified two-stage framework that integrates Partial Least Squares Structural Equation Modeling (PLS-SEM) with a Convolutional Neural Network (CNN) to analyze AI adoption in Nigerian SMEs, demonstrating that while PLS-SEM confirms theoretical relationships, the CNN significantly enhances out-of-sample predictive accuracy when both models are rigorously benchmarked on identical data folds.

Original authors: Ayodeji Idowu, Yemisi Tomilola Babalola, Yacob Haliso, Oyebayo Ridwan Olaniran

Published 2026-10-22

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ayodeji Idowu, Yemisi Tomilola Babalola, Yacob Haliso, Oyebayo Ridwan Olaniran

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Small businesses are the backbone of the global economy, employing half the world's workforce and forming the vast majority of companies everywhere. Yet, when it comes to adopting new technologies like artificial intelligence, these smaller firms often lag behind their larger counterparts. The gap is not usually because the technology is unavailable or too expensive; rather, it is a matter of willingness. Decision-makers in small companies must first understand what the technology can do and then decide if they want to use it. Researchers have long tried to map out the factors that influence this decision, such as whether a business owner trusts the technology, if they have the necessary digital skills, or if they feel pressure from competitors. Traditionally, scientists have used statistical tools to draw straight lines between these factors and the final decision to adopt, assuming that the relationship is simple and direct. However, human behavior and business environments are rarely so straightforward, and the tools used to study them often struggle to capture the messy, non-linear reality of how people actually make choices.

A team of researchers from Nigeria and the United Kingdom set out to test whether a more sophisticated approach could reveal a clearer picture of how small businesses decide to use artificial intelligence. They focused on a specific group of 1,210 small and medium-sized enterprises across Nigeria, surveying owners and managers about their awareness of various AI tools, their trust in the technology, and their intention to adopt it. The researchers combined two different methods of analysis. The first was a standard statistical approach that tests a pre-defined theory about how different factors connect. The second was a type of artificial intelligence known as a convolutional neural network, a system designed to find complex patterns in data that standard statistics might miss. Crucially, the team did not just run these two methods one after the other on different sets of data. Instead, they built a unified system where the results of the first method shaped the input for the second, and they tested both models on the exact same chunks of data to ensure a fair comparison.

The study began by confirming a set of fifteen different factors that might influence a business's decision. These included technological aspects like how useful or easy the tools seemed, organizational factors like the support of top management and the digital literacy of the staff, and environmental factors like government support and competitive pressure. The researchers also measured how much the business owners trusted AI and how aware they were of specific tools available in the market. When they analyzed the data, they found that almost every single one of these fifteen factors played a role. The strongest drivers were trust in the technology and the level of digital literacy within the company. Interestingly, the researchers discovered that these factors rarely influenced the decision to adopt AI directly. Instead, they worked by first increasing the business owner's awareness of what AI could do. Once a decision-maker became aware of the technology, that awareness became the primary gateway to the intention to adopt. Trust was the only factor that worked in two ways: it increased awareness, but it also directly boosted the willingness to adopt, independent of awareness.

The researchers then tested whether their complex artificial intelligence model could predict the adoption intention better than the standard statistical model. They found that the neural network did indeed perform slightly better, reducing the error in its predictions by about two percent compared to the standard model. This improvement was statistically significant, meaning it was unlikely to be a fluke. However, the study also revealed a surprising limitation. When the researchers compared the neural network to other simpler machine learning tools, such as a random forest algorithm or a basic feed-forward network, the neural network did not perform any better. This suggests that the improvement came not from the specific, complex architecture of the neural network, but simply from the fact that it allowed for non-linear relationships, which the standard statistical model could not capture. The study also found that using a validated measurement model to group the survey questions was essential. When they tried to predict adoption using all the raw survey questions without grouping them first, the results were significantly worse, proving that organizing the data according to theory is just as important as the prediction method itself.

Ultimately, the research highlights that for small businesses, the path to adopting artificial intelligence is paved with awareness. Contextual factors like government support, financial readiness, or competitive pressure do not push a business to adopt AI directly; they first push the business owner to become aware of the technology. Once that awareness is established, the intention to adopt follows. The study also serves as a methodological lesson for the scientific community. It demonstrates that while advanced machine learning tools can offer a slight edge in prediction, they must be tested fairly against simpler alternatives to understand what is actually driving the improvement. In this case, the gain came from allowing the data to speak in its own complex, non-linear voice rather than forcing it into a straight line, but the specific complexity of the neural network itself was not the magic ingredient. The findings confirm that theory still matters: organizing data around established concepts helps even the most advanced algorithms make better sense of the world.

Technical Summary: A Hybrid PLS-SEM and Convolutional Neural Network Analysis of AI Adoption Intention in SMEs

Problem Statement
Empirical research on technology adoption, particularly Artificial Intelligence (AI) in Small and Medium-sized Enterprises (SMEs), has historically relied on Partial Least Squares Structural Equation Modelling (PLS-SEM). While effective for hypothesis testing and explaining in-sample variance, this approach faces three critical limitations:

Predictive Validity: Models are often evaluated based on in-sample explanatory power ( $R^2$ ) rather than out-of-sample predictive performance, leading to potential overestimation of a model's transportability.
Functional Form Assumptions: Conventional PLS-SEM assumes linear relationships, failing to capture potential non-linear dynamics (e.g., capability thresholds or diminishing returns) inherent in adoption behaviors.
Flawed Hybrid Methodologies: Existing hybrid approaches that combine SEM with Artificial Neural Networks (ANN) typically train the neural network on construct scores derived from the structural model and evaluate them on separate data splits. This methodology prevents a formal, statistically rigorous comparison between the structural and predictive models, rendering claims of "mutual validation" untestable.

Methodology
The study proposes a unified, two-stage evaluation framework applied to data from 1,210 Nigerian SMEs. The design integrates confirmatory and predictive stages through three explicit architectural couplings:

Stage 1: Confirmatory PLS-SEM
- Objective: Validate the measurement model, estimate latent constructs, and test hypotheses based on an integrated Technology-Organization-Environment (TOE), Diffusion of Innovations (DOI), and behavioral framework.
- Model: A structural model with 15 constructs (13 antecedents, 1 mediator, 1 outcome) and 15 hypotheses.
- Output: Validated construct scores and standardized path coefficients serve as the basis for the next stage.
Stage 2: Predictive Convolutional Neural Network (CNN)
- Input Representation: Instead of using raw indicators or simple construct scores, the study encodes each respondent as a 2D tensor ( $14 \times 6$ ). Rows represent the 14 validated constructs (13 antecedents + mediator), and columns represent their retained indicators.
- Architecture:
  - Within-Construct Convolution: 1D kernels (specifically 1×3) slide along the indicator axis to learn an aggregation analogue of the PLS-SEM "outer model."
  - Across-Construct Convolution: 1D kernels (specifically 3×1) slide along the construct axis to learn inter-construct relationships, serving as a learned analogue of the "inner model."
- Evaluation Protocol: Crucially, the CNN, the PLS-SEM structural model, and four alternative benchmarks (Feed-Forward ANN, Random Forest, Construct-Score CNN, and Ordinary Least Squares) are evaluated using identical 10-fold cross-validation splits. This ensures commensurate out-of-sample comparisons.
- Statistical Rigor: Performance differences are assessed using paired t-tests, Wilcoxon signed-rank tests, and effect sizes (Cohen's $d$ ).

Key Results

Structural Findings: All 15 hypotheses were supported. The model explained 61.1% of the variance in AI Awareness and 54.7% in AI Adoption Intention.
- Mediation: AI Awareness fully mediated the effects of 12 out of 13 antecedents on adoption intention.
- Trust: Trust in AI was the only construct exhibiting complementary partial mediation, influencing intention both directly and indirectly through awareness.
- Predictive Relevance: While the PLS-SEM model showed positive predictive relevance ( $Q^2_{predict}$ ), it failed to outperform a linear regression benchmark on the final outcome variable (Adoption Intention), highlighting a limitation in relying solely on $Q^2$ without explicit benchmarking.
Predictive Performance:
- The item-tensor CNN achieved a 2.06% reduction in out-of-sample RMSE compared to the PLS-SEM structural benchmark ( $p = 0.032$ ).
- However, the CNN did not significantly outperform a Feed-Forward Neural Network or a Random Forest. This suggests the performance gain stems from relaxing linear functional form assumptions rather than the specific convolutional architecture or network depth.
- Dimensionality Reduction Value: An OLS model trained on all 68 raw indicators performed significantly worse (2.97% higher RMSE) than the two-predictor structural model. This confirms that theory-guided dimensionality reduction via a validated measurement model is essential for predictive performance at this sample size.
Importance Analysis: Permutation importance from the CNN and standardized path coefficients from PLS-SEM showed broad concordance (Spearman $\rho = 0.503$ ), with AI Awareness and Trust identified as the dominant predictors in both frameworks.

Significance and Claims
The paper claims to advance hybrid SEM-ANN research by addressing methodological transparency and rigor:

Architectural Integration: It formalizes a hybrid architecture where the measurement model defines the deep network's input representation, making the linkage between confirmatory and predictive stages architectural rather than merely sequential.
Falsifiable Testing: By evaluating all models on identical cross-validation folds with paired statistical testing, the study provides a level of inferential transparency rarely found in hybrid literature. It demonstrates that claims of "neural confirmation" of structural models are only valid if the neural model is tested against the structural model on the same data partitions.
Role of Measurement Theory: The study provides empirical evidence that in survey-based research with moderate sample sizes ( $N \approx 1,200$ ), theory-guided dimensionality reduction (via PLS-SEM) preserves predictive information better than flexible estimators applied to raw, high-dimensional data.
Substantive Insight: The findings reinforce the "staged decision logic" in diffusion theory, identifying AI Awareness as a near-universal gateway through which contextual factors influence adoption intention, with Trust acting as a unique dual-path driver.

The authors maintain a modest stance, acknowledging that the observed predictive gains over linear models are modest (a few percentage points) and that the specific CNN architecture offered no advantage over simpler non-linear baselines like Random Forests in this context. The primary contribution is the methodological framework for rigorous comparison rather than the superiority of deep learning itself.

Technical Summary: A Hybrid PLS-SEM and Convolutional Neural Network Analysis of AI Adoption Intention in SMEs

More like this