Збережено в:
| Автор: | |
|---|---|
| Формат: | Recurso digital |
| Мова: | |
| Опубліковано: |
Zenodo
2026
|
| Предмети: | |
| Онлайн доступ: | https://doi.org/10.5281/zenodo.20078428 |
| Теги: |
Додати тег
Немає тегів, Будьте першим, хто поставить тег для цього запису!
|
Зміст:
- <p class="MsoBodyText">Personalized<span> </span>recommender<span> </span>systems<span> </span>play<span> </span>a<span> </span>central<span> </span>role<span> </span>in<span> </span>modern<span> </span>digital<span> </span>platforms,<span> </span>yet<span> </span>many<span> </span>high-performing models<span> </span>operate<span> </span>as<span> </span>opaque<span> </span>systems,<span> </span>limiting<span> </span>user<span> </span>trust<span> </span>and<span> </span>practical<span> </span>adoption.<span> </span>Explainable<span> </span>artificial<span> </span>intelligence has<span> </span>emerged<span> </span>as<span> </span>a<span> </span>promising<span> </span>approach<span> </span>to<span> </span>address<span> </span>this<span> </span>challenge<span> </span>by<span> </span>enhancing<span> </span>transparency while<span> </span>maintaining predictive capability. This study investigates the impact of integrating explainable features into a deep learning-based personalized recommender system to improve both performance and interpretability. Using the<span> </span>REASONER<span> </span>dataset,<span> </span>which<span> </span>includes<span> </span>user-item<span> </span>interactions<span> </span>enriched<span> </span>with<span> </span>user<span> </span>attributes,<span> </span>personality<span> </span>traits, and multi-aspect tags, a logistic regression model was implemented as a baseline and compared with a feedforward<span> </span>neural<span> </span>network.<span> </span>Model<span> </span>performance<span> </span>was<span> </span>evaluated<span> </span>using<span> </span>accuracy,<span> </span>precision,<span> </span>recall,<span> </span>F1<span> </span>score,<span> </span>and ROC-AUC,<span> </span>with<span> </span>threshold<span> </span>tuning<span> </span>and<span> </span>feature<span> </span>importance<span> </span>analysis<span> </span>applied<span> </span>to<span> </span>optimize<span> </span>and<span> </span>interpret<span> </span>the<span> </span>neural network.<span> </span>The<span> </span>results<span> </span>show<span> </span>that<span> </span>the<span> </span>neural<span> </span>network<span> </span>achieved<span> </span>superior<span> </span>discriminative<span> </span>performance<span> </span>with<span> </span>a<span> </span>ROC-AUC<span> </span>of<span> </span>0.729<span> </span>and<span> </span>improved<span> </span>minority-class<span> </span>recall<span> </span>after<span> </span>optimization.<span> </span>Feature<span> </span>importance<span> </span>analysis<span> </span>revealed<span> </span>that interest tags, video tags, and reason-based attributes were the most influential predictors, whereas demographic variables contributed less significantly. These findings indicate that incorporating explainable, behavior-driven features enhances both the effectiveness and transparency of recommendation models. Overall, the study highlights the importance of combining deep learning with explainable inputs to develop more reliable, user-centered recommender systems.</p>