This page restates the published work in a repository-native format: the research problem, architecture, experimental setup, quantitative results, module-level attention analysis, and repository evidence that supports each claim.
Traditional ERP systems are operationally central but often rigid, difficult to scale cleanly, and weak at producing real-time insight when module dependencies become complex.
The paper argues that AI and cloud computing together provide a better foundation: AI for predictive automation and dependency modeling, cloud for scalability and infrastructure elasticity.
The proposed framework uses a BiLSTM-Attention architecture to evaluate ERP module relationships and support modular analysis in cloud-hosted enterprise systems.
The reported intent is better system efficiency, resource optimization, and intelligent decision support without changing the core meaning of the published work.
This section focuses on what the local paper bundle directly supports: dataset description, preprocessing rules, model architecture, and hyperparameters.
| Parameter | Value |
|---|---|
| Embedding Dimension | 128 |
| LSTM Hidden Size | 128 |
| Attention Vector Size | 64 |
| Dropout Rate | 0.3 |
| Learning Rate | 0.001 |
| Optimizer | Adam |
| Epochs | 50 |
| Batch Size | 32 |
| Loss Function | Categorical Crossentropy |
The BiLSTM-Attention model is the top performer across accuracy, precision, recall, and F1-score. Random Forest is the strongest traditional baseline, and GRU is the strongest non-proposed sequence baseline.
The paper emphasizes that the proposed architecture improves not only peak accuracy but also consistency across evaluation metrics.
The published figures show a steady improvement in training and validation accuracy and a simultaneous reduction in both training and validation loss.
Because the raw logs are absent, this repository keeps the original images and separately provides a clearly labeled reconstructed curve series for interactive use.
| Model | Precision | Recall | F1-score | Accuracy |
|---|---|---|---|---|
| Logistic Regression | 84.0 | 79.5 | 81.7 | 82.7 |
| Random Forest | 87.5 | 83.0 | 85.2 | 85.3 |
| Support Vector Machine | 80.2 | 77.0 | 78.6 | 79.4 |
| Gated Recurrent Unit | 85.0 | 86.5 | 85.7 | 83.2 |
| BiLSTM-Attention | 89.5 | 90.8 | 90.1 | 91.2 |
| Module | Score | Interpretation |
|---|---|---|
| Module_1 | 0.07 | Low importance; likely non-critical table |
| Module_2 | 0.09 | Moderate relevance |
| Module_3 | 0.05 | Minimal impact |
| Module_4 | 0.13 | High dependency weight |
| Module_5 | 0.17 | Most influential; likely a central or shared table |
| Module_6 | 0.11 | Strong contextual role |
| Module_7 | 0.04 | Negligible attention; possibly isolated module |
| Module_8 | 0.12 | Important secondary dependency |
| Module_9 | 0.10 | Supportive of prediction context |
| Module_10 | 0.12 | High final-context relevance |
The paper explicitly discusses a cloud-native ERP integration pathway, but not a completed real-world production deployment. This repository preserves that boundary instead of overstating operational evidence.
Publication metadata on this site follows the IEEE indexing record. The original LaTeX bundle still contains an outdated first-page header from an earlier template.