Context-adaptive deep learning for sustainable product recommendation
Application to concrete
Bibliographic Data
| ID | 22434006 |
|---|---|
| Authors | Roger Vergés (0009-0001-5887-4785, The University of Melbourne, corresponding author), Núria Forcada (0000-0003-2109-4205, Universitat Politècnica de Catalunya), Kàtia Gaspar (0000-0003-3842-1401, Universitat Politècnica de Catalunya), M Reza Hosseini (0000-0001-8675-736X, The University of Melbourne) |
| Year | 2026 |
| Volume | 67 |
| Pages | 119-144 |
| Publication date | 2026-09-01 |
| Open Access | Yes |
| Type | ARTICLE |
| Venue | Sustainable Production and Consumption (JOURNAL) |
| Journal identifiers | ISSN: 2352-5509 • E-ISSN: 2352-5509 |
| Publisher | Elsevier BV (PUBLISHER) |
| DOI | 10.1016/j.spc.2026.06.011 |
| OpenAlex | W7167065834 |
| Language | EN |
| References cited | 74 |
The increasing availability of sustainability and circularity data for construction products contrasts with the absence of automated multi-criteria decision-making tools, forcing practitioners to rely on cost and availability heuristics that systematically underutilise sustainable alternatives. This study presents a context-adaptive deep learning model that scores and ranks construction products, with concrete serving as the application domain. A Set Transformer architecture employing multi-head self-attention captures relational dependencies among competing alternatives, supporting variable-size candidate sets of 2–5 alternatives. Each product is encoded through 66 input features covering environmental, circularity, economic, and performance indicators alongside stakeholder preferences and application-specific contextual inputs. Indicator relevance is designed to shift dynamically according to the supplied context. The model was trained on a hybrid data set of 42,874 labelled scenarios combining deterministic control cases, large language model-generated synthetic labels, and expert annotations weighted by source reliability. Performance on held-out test data yields mean reciprocal rank values between 0.98 and 0.89 and normalised discounted cumulative gain values between 0.97 and 0.89 across two- to five-alternative scenarios, with a stable mean absolute error of 0.05. External evaluation against six domain experts across 32 real-product scenarios provides preliminary evidence of alignment between model outputs and expert reasoning, and model analyses reveal that indicator relevance responds consistently and interpretably to variation in stakeholder preferences and application context. The framework's compatibility with environmental product declaration and performance schemas and its modular architecture suggest transferability to other construction product categories, positioning it as a scalable decision-support tool that enables clients to make objective early-stage procurement decisions.
Artificial neural network · Deep learning · Sustainable development · Sustainable production · Machine Learning in Materials Science · Recommender Systems and Techniques · Sentiment Analysis and Opinion Mining
Advantages of the mean absolute error (MAE) over the root mean square error (RMSE) in assessing average model performance
ReCiPe2016
Impact 2002
Evaluating collaborative filtering recommender systems
A survey on Image Data Augmentation for Deep Learning
A product-level framework for prioritizing circular and sustainable construction in building design
Large-scale analysis of environmental and circularity metrics in construction products through automated data extraction using large language models
Beyond carbon
Assessment of environmental product declaration and databases
| Highly cited | No |
|---|