Automatic Optimization of Deep Learning Training through Feature-Aware-Based Dataset Splitting
<p>Raw data of grapevine leaves composed of 6 classes (phenotypes), acquired over a white background—also documented in previous work [<a href="#B2-algorithms-17-00106" class="html-bibr">2</a>]. From left to right: <span class="html-italic">Touriga Nacional</span>, <span class="html-italic">Tinto Cão</span>, <span class="html-italic">Códega</span>, <span class="html-italic">Moscatel</span>, <span class="html-italic">Tinta Roriz</span>, and <span class="html-italic">Rabigato</span>.</p> "> Figure 2
<p>Different types of defects commonly found in bridges, which are also documented in previous work [<a href="#B1-algorithms-17-00106" class="html-bibr">1</a>].</p> "> Figure 3
<p>External image sets DL models extended assessments: (<b>a</b>) features a set of images for grapevine variety identification through leaf analysis; (<b>b</b>) showcases examples related to the detection of bridge defects. A careful alignment of the classes/labels that compose contextually corresponding sets (grapevine and bridge defects group) was ensured.</p> "> Figure 4
<p>Data splitting general workflow.</p> "> Figure 5
<p>KFCV process.</p> "> Figure 6
<p>General workflow of the split process.</p> "> Figure 7
<p>Key steps of PCA.</p> "> Figure 8
<p>Visualization of centroid point of features for the class of “<span class="html-italic">absence of joint cover plate</span>” for <math display="inline"><semantics> <mrow> <msubsup> <mrow> <mi>D</mi> </mrow> <mrow> <mi>T</mi> <mi>r</mi> <mo>−</mo> <mi>V</mi> <mi>a</mi> <mi>l</mi> </mrow> <mrow> <mi>B</mi> <mi>D</mi> <mi>I</mi> </mrow> </msubsup> </mrow> </semantics></math>, for exemplification purposes. The blue dots represent the distribution of features, while the red cross is the global centroid.</p> "> Figure 9
<p>FCDIS process for setting up training and validation subsets: (<b>a</b>) presents the main pipeline of the process; (<b>b</b>) depicts the pipeline’s general result-oriented concept using 3 abstract clusters as examples, where in red, yellow, green represents the farthest, intermediate, and closer distance to centroid, respectively. Considering the distance of each cluster centroid to a global centroid, the <math display="inline"><semantics> <mrow> <msub> <mrow> <mi>D</mi> </mrow> <mrow> <mi>T</mi> <mi>r</mi> </mrow> </msub> </mrow> </semantics></math> and <math display="inline"><semantics> <mrow> <msub> <mrow> <mi>D</mi> </mrow> <mrow> <mi>V</mi> <mi>a</mi> <mi>l</mi> </mrow> </msub> </mrow> </semantics></math> are assembled, grounded in the following diversity rule: the farthest the cluster centroid is, the more images are used to compose the <math display="inline"><semantics> <mrow> <msub> <mrow> <mi>D</mi> </mrow> <mrow> <mi>T</mi> <mi>r</mi> </mrow> </msub> </mrow> </semantics></math>.</p> "> Figure 9 Cont.
<p>FCDIS process for setting up training and validation subsets: (<b>a</b>) presents the main pipeline of the process; (<b>b</b>) depicts the pipeline’s general result-oriented concept using 3 abstract clusters as examples, where in red, yellow, green represents the farthest, intermediate, and closer distance to centroid, respectively. Considering the distance of each cluster centroid to a global centroid, the <math display="inline"><semantics> <mrow> <msub> <mrow> <mi>D</mi> </mrow> <mrow> <mi>T</mi> <mi>r</mi> </mrow> </msub> </mrow> </semantics></math> and <math display="inline"><semantics> <mrow> <msub> <mrow> <mi>D</mi> </mrow> <mrow> <mi>V</mi> <mi>a</mi> <mi>l</mi> </mrow> </msub> </mrow> </semantics></math> are assembled, grounded in the following diversity rule: the farthest the cluster centroid is, the more images are used to compose the <math display="inline"><semantics> <mrow> <msub> <mrow> <mi>D</mi> </mrow> <mrow> <mi>T</mi> <mi>r</mi> </mrow> </msub> </mrow> </semantics></math>.</p> "> Figure 10
<p>Example of K-means-based clustering process plot perspective, using a computational method for the determination of K value. In (<b>a</b>), there is a chart depicting the value determined from the elbow technique regarding the “<span class="html-italic">absence of joint cover plate</span>” class belonging to <math display="inline"><semantics> <mrow> <msubsup> <mrow> <mi>D</mi> </mrow> <mrow> <mi>T</mi> <mi>r</mi> <mo>−</mo> <mi>V</mi> <mi>a</mi> <mi>l</mi> </mrow> <mrow> <mi>B</mi> <mi>D</mi> <mi>I</mi> </mrow> </msubsup> </mrow> </semantics></math>, for exemplification purposes. The resulting clustering operation, based on a predetermined K, is illustrated in (<b>b</b>).</p> "> Figure 11
<p>Visualization of a few samples of each cluster for the “absence of joint cover plate” class concerning <math display="inline"><semantics> <mrow> <msubsup> <mrow> <mi>D</mi> </mrow> <mrow> <mi>T</mi> <mi>r</mi> <mo>−</mo> <mi>V</mi> <mi>a</mi> <mi>l</mi> </mrow> <mrow> <mi>B</mi> <mi>D</mi> <mi>I</mi> </mrow> </msubsup> </mrow> </semantics></math>: (<b>a</b>) depicts the images grouped by K-means clustering; (<b>b</b>) presents a summary of the images collected for <math display="inline"><semantics> <mrow> <msub> <mrow> <mi>D</mi> </mrow> <mrow> <mi>T</mi> <mi>r</mi> </mrow> </msub> </mrow> </semantics></math>, considering each cluster’s distance and size.</p> "> Figure 12
<p>FSCIS process for setting up training and validation subsets: (<b>a</b>) presents the main pipeline of the process; (<b>b</b>) depicts the result of the pipeline behavior, in which (i) images are firstly sorted by their distance to the feature space centroid, (ii) subgroups are then created, (iii) examples for training are iteratively selected, based on the median element of each created subgroup, until a percentage of 80% of the to <math display="inline"><semantics> <mrow> <msub> <mrow> <mi>D</mi> </mrow> <mrow> <mi>T</mi> <mi>r</mi> <mo>−</mo> <mi>V</mi> <mi>a</mi> <mi>l</mi> </mrow> </msub> </mrow> </semantics></math> is reached, and, finally, (iv) the remaining images are assigned to <math display="inline"><semantics> <mrow> <msub> <mrow> <mi>D</mi> </mrow> <mrow> <mi>V</mi> <mi>a</mi> <mi>l</mi> </mrow> </msub> </mrow> </semantics></math>. Black vertical lines illustrate the referred distance to the feature space centroid, white squares represent the images that have not yet been assigned; orange squares depict the images for <math display="inline"><semantics> <mrow> <msub> <mrow> <mi>D</mi> </mrow> <mrow> <mi>T</mi> <mi>r</mi> </mrow> </msub> </mrow> </semantics></math>; and, finally, purple squares represent the images that have been assigned for <math display="inline"><semantics> <mrow> <msub> <mrow> <mi>D</mi> </mrow> <mrow> <mi>V</mi> <mi>a</mi> <mi>l</mi> </mrow> </msub> </mrow> </semantics></math>.</p> "> Figure 13
<p>Learning curves, generated to compare the learning behavior of models over datasets produced based on both traditional techniques and FCDIS/FSCIS methods: (<b>a</b>) Xception/Nadam for <math display="inline"><semantics> <mrow> <msubsup> <mrow> <mi>D</mi> </mrow> <mrow> <mi>T</mi> <mi>r</mi> <mi>V</mi> <mi>a</mi> <mi>l</mi> </mrow> <mrow> <mi>B</mi> <mi>D</mi> <mi>I</mi> </mrow> </msubsup> </mrow> </semantics></math> bridge defects datasets; (<b>b</b>) Xception/Nadam for <math display="inline"><semantics> <mrow> <msubsup> <mrow> <mi>D</mi> </mrow> <mrow> <mi>T</mi> <mi>r</mi> <mo>,</mo> <mi>V</mi> <mi>a</mi> <mi>l</mi> </mrow> <mrow> <mi>G</mi> <mi>L</mi> <mi>I</mi> </mrow> </msubsup> </mrow> </semantics></math>.</p> "> Figure 14
<p>Visualization of the Grad-CAM evaluation approach: (<b>a</b>,<b>b</b>) show the saliency maps over <math display="inline"><semantics> <mrow> <msubsup> <mrow> <mi>D</mi> </mrow> <mrow> <mi>E</mi> <mi>x</mi> <mi>t</mi> <mo>_</mo> <mi>T</mi> <mi>s</mi> <mi>t</mi> </mrow> <mrow> <mi>G</mi> <mi>L</mi> <mi>I</mi> </mrow> </msubsup> </mrow> </semantics></math>. and <math display="inline"><semantics> <mrow> <msubsup> <mrow> <mi>D</mi> </mrow> <mrow> <mi>E</mi> <mi>x</mi> <mi>t</mi> <mo>_</mo> <mi>T</mi> <mi>s</mi> <mi>t</mi> </mrow> <mrow> <mi>B</mi> <mi>D</mi> <mi>I</mi> </mrow> </msubsup> </mrow> </semantics></math>, respectively. In each one, the first row corresponds to the Grad-CAM visualization, while the second row presents the ground-truth.</p> "> Figure 15
<p>Attention visualization for <math display="inline"><semantics> <mrow> <msubsup> <mrow> <mi>D</mi> </mrow> <mrow> <mi>E</mi> <mi>x</mi> <mi>t</mi> <mo>_</mo> <mi>T</mi> <mi>s</mi> <mi>t</mi> </mrow> <mrow> <mi>G</mi> <mi>L</mi> <mi>I</mi> </mrow> </msubsup> </mrow> </semantics></math> is performed using the top Xception/Nadam model trained with the documented splitting techniques. Model1, Model2, Model3, and Model4 correspond, respectively, to the models trained with the <math display="inline"><semantics> <mrow> <msubsup> <mrow> <mi>D</mi> </mrow> <mrow> <mi>E</mi> <mi>x</mi> <mi>t</mi> <mo>_</mo> <mi>T</mi> <mi>s</mi> <mi>t</mi> </mrow> <mrow> <mi>G</mi> <mi>L</mi> <mi>I</mi> </mrow> </msubsup> </mrow> </semantics></math> variants split using SCDIS, FSCIS, HO, and KFCV10. In the first column, the red rectangle represents the most prominent attention area determined through Grad-CAM, while the green rectangle corresponds to the hand-crafted ground-truth.</p> "> Figure 15 Cont.
<p>Attention visualization for <math display="inline"><semantics> <mrow> <msubsup> <mrow> <mi>D</mi> </mrow> <mrow> <mi>E</mi> <mi>x</mi> <mi>t</mi> <mo>_</mo> <mi>T</mi> <mi>s</mi> <mi>t</mi> </mrow> <mrow> <mi>G</mi> <mi>L</mi> <mi>I</mi> </mrow> </msubsup> </mrow> </semantics></math> is performed using the top Xception/Nadam model trained with the documented splitting techniques. Model1, Model2, Model3, and Model4 correspond, respectively, to the models trained with the <math display="inline"><semantics> <mrow> <msubsup> <mrow> <mi>D</mi> </mrow> <mrow> <mi>E</mi> <mi>x</mi> <mi>t</mi> <mo>_</mo> <mi>T</mi> <mi>s</mi> <mi>t</mi> </mrow> <mrow> <mi>G</mi> <mi>L</mi> <mi>I</mi> </mrow> </msubsup> </mrow> </semantics></math> variants split using SCDIS, FSCIS, HO, and KFCV10. In the first column, the red rectangle represents the most prominent attention area determined through Grad-CAM, while the green rectangle corresponds to the hand-crafted ground-truth.</p> "> Figure 16
<p>Attention visualization for <math display="inline"><semantics> <mrow> <msubsup> <mrow> <mi>D</mi> </mrow> <mrow> <mi>E</mi> <mi>x</mi> <mi>t</mi> <mo>_</mo> <mi>T</mi> <mi>s</mi> <mi>t</mi> </mrow> <mrow> <mi>B</mi> <mi>D</mi> <mi>I</mi> </mrow> </msubsup> </mrow> </semantics></math> is performed using the top Xception/Nadam model trained with the documented splitting techniques. Model1, Model2, Model3, and Model4 correspond, respectively, to the models trained with the <math display="inline"><semantics> <mrow> <msubsup> <mrow> <mi>D</mi> </mrow> <mrow> <mi>E</mi> <mi>x</mi> <mi>t</mi> <mo>_</mo> <mi>T</mi> <mi>s</mi> <mi>t</mi> </mrow> <mrow> <mi>B</mi> <mi>D</mi> <mi>I</mi> </mrow> </msubsup> </mrow> </semantics></math> variants split using SCDIS, FSCIS, HO, and KFCV10. In the first column, the red rectangle represents the most prominent attention area determined through Grad-CAM, while the green rectangle corresponds to the hand-crafted ground-truth.</p> ">
Abstract
:1. Introduction
2. Related Work
3. Materials and Methods
3.1. Imagery Used for the Empirical Assessment of the Proposed Dataset Splitting Methodology
3.2. Complementary Imagery for Extended Assessments—Models Inference Consistency and Attention Map Analysis
- D represents the dataset;
- i represents a specific dataset chosen from a set of datasets I = {bridge defects imagery (BDI), grapevine leaves imagery (GLI)};
- n represents the specific sets N = {training-validation (Tr-Val), train (Tr), validation (Val), test (Tst), and external test (Ext_Tst)}.
3.3. General Workflow
- Accuracy is a fundamental metric in the field of machine learning, measures the proportion of correctly predicted instances out of the total number of instances. It is usually used to evaluate how well the model classified and predicted classes over a testing subset. More specifically, it provides an overall assessment of a model’s correctness (Equation (1)).
- IOU is a spatial overlap metric commonly used in tasks involving object detection and segmentation. It quantifies the degree of overlap between the predicted regions and the ground truth regions. Specifically, IOU calculates the ratio of the intersection area between the predicted and actual regions to the union area of those regions. IOU allows the capture of the spatial alignment and precision of the model’s output in relation to the true object regions. To compute it, the most prominent attention area of CNN’s gradient-weighted class activation mapping (Grad-CAM) is used as a bounding box predictor (Equation (2)).
3.4. Hardware and Software Tools
- Processor: an Intel(R) Core (TM) i7-8700 CPU 3.20 GHz 3.19 GHz (Intel Co., Santa Clara, CA, USA);
- Random access memory (RAM): 16.0 GB (Corsair, Fremont, CA, USA);
- Graphic card: NVIDIA GeForce GTX 1080, 16.0 GB (NVIDIA Co., Santa Clara, CA, USA);
- Storage: 500 SSD (Samsung Electronics, Suwon, Republic of Korea);
- Operative system: Windows 10 Pro (Microsoft Co., Redmond, WA, USA).
3.5. Setting Up the Standards: Traditional Splitting Methodology
3.6. Proposed Diversity-Oriented Data Split Approaches
3.6.1. Preliminary Feature Engineering
3.6.2. Feature Clustering Distance-Based Image Selection (FCDIS)
3.6.3. Feature Space Center-Based Image Selection (FSCIS)
4. Experimental Results
4.1. Consistency of the Training Metrics: Analyzing the Learning Curves
4.2. Assessment with the Data Reserved for Testing
4.3. Assessment with External Imagery Sources
4.4. Attention Mechanisms Assessment with External Imagery Sources
5. Discussion, Summary, and Conclusions
Author Contributions
Funding
Data Availability Statement
Conflicts of Interest
References
- Shahrabadi, S.; Gonzalez, D.; Sousa, N.; Adao, T.; Peres, E.; Magalhaes, L. Benchmarking Deep Learning Models and Hyperparameters for Bridge Defects Classification. Procedia Comput. Sci. 2023, 219, 345–353. [Google Scholar] [CrossRef]
- Adão, T.; Pinho, T.M.; Ferreira, A.; Sousa, A.; Pádua, L.; Sousa, J.; Sousa, J.J.; Peres, E.; Morais, R. Digital Ampelographer: A CNN Based Preliminary Approach. In Proceedings of the EPIA Conference on Artificial Intelligence, Vila Real, Portugal, 3–6 September 2019; Lecture Notes in Computer Science (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). Springer: Berlin/Heidelberg, Germany, 2019; Volume 11804 LNAI, pp. 258–271. [Google Scholar]
- Shahrabadi, S.; Carias, J.; Peres, E.; Magalhães, L.G.; Lopez, M.A.G.; Silva, L.B.; Adão, T. Image-Based Lung Analysis in the Context of Digital Pathology: A Brief Review. In Proceedings of the Hcist—International Conference on Health and Social Care Information Systems and Technologies (HCist), Porto, Portugal, 8–10 November 2023. [Google Scholar]
- Tran, T.-O.; Vo, T.H.; Le, N.Q.K. Omics-Based Deep Learning Approaches for Lung Cancer Decision-Making and Therapeutics Development. Brief. Funct. Genomics 2023, elad031. [Google Scholar] [CrossRef]
- Yuan, Q.; Chen, K.; Yu, Y.; Le, N.Q.K.; Chua, M.C.H. Prediction of Anticancer Peptides Based on an Ensemble Model of Deep Learning and Machine Learning Using Ordinal Positional Encoding. Brief. Bioinform. 2023, 24, bbac630. [Google Scholar] [CrossRef] [PubMed]
- Shahrabadi, S.; Rodrigues, J.; Margolis, I.; Evangelista, L.; Sousa, N.; Sousa, E.; Guevara Lopéz, M.A.; Magalhães, L.G.; Peres, E.; Adão, T. Digital Tools for Aircraft Maintenance: Prototyping Location-Aware AOI for Engine Assessment and Cable Routing Solutions. In Proceedings of the International Conference on Graphics and Interaction (ICGI), Tomar, Portugal, 2–3 November 2023. [Google Scholar]
- Oliveira, J.; Gomes, R.; Gonzalez, D.; Sousa, N.; Shahrabadi, S.; Guevara, M.; Ferreira, M.J.; Alves, P.; Peres, E.; Magalhães, L.; et al. Footwear Segmentation and Recommendation Supported by Deep Learning: An Exploratory Proposal. Procedia Comput. Sci. 2023, 219, 724–735. [Google Scholar] [CrossRef]
- Simonyan, K.; Zisserman, A. Very Deep Convolutional Networks for Large-Scale Image Recognition. arXiv 2014, arXiv:1409.1556. [Google Scholar]
- He, K.; Zhang, X.; Ren, S.; Sun, J. Deep Residual Learning for Image Recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA, 27–30 June 2016; pp. 770–778. [Google Scholar] [CrossRef]
- Chollet, F. Xception: Deep Learning with Depthwise Separable Convolutions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, HI, USA, 21–26 July 2017; pp. 1251–1258. [Google Scholar]
- Tredennick, A.T.; Hooker, G.; Ellner, S.P.; Adler, P.B. A Practical Guide to Selecting Models for Exploration, Inference, and Prediction in Ecology. Ecology 2021, 102, e03336. [Google Scholar] [CrossRef]
- Xu, C.; Coen-Pirani, P.; Jiang, X. Empirical Study of Overfitting in Deep Learning for Predicting Breast Cancer Metastasis. Cancers 2023, 15, 1969. [Google Scholar] [CrossRef]
- Mathur, J.; Baruah, B.; Tiwari, P. Prediction of Bio-Oil Yield during Pyrolysis of Lignocellulosic Biomass Using Machine Learning Algorithms. Can. J. Chem. Eng. 2023, 101, 2457–2471. [Google Scholar] [CrossRef]
- Montesinos López, O.A.; Montesinos López, A.; Crossa, J. Multivariate Statistical Machine Learning Methods for Genomic Prediction; Springer International Publishing: Berlin/Heidelberg, Germany, 2022. [Google Scholar]
- Nematzadeh, Z.; Ibrahim, R.; Selamat, A. Comparative Studies on Breast Cancer Classifications with K-Fold Cross Validations Using Machine Learning Techniques. In Proceedings of the 2015 10th Asian Control Conference (ASCC), Kota Kinabalu, Malaysia, 31 May–3 June 2015; pp. 1–6. [Google Scholar]
- Nakatsu, R.T. Validation of Machine Learning Ridge Regression Models Using Monte Carlo, Bootstrap, and Variations in Cross-Validation. J. Intell. Syst. 2023, 32, 20220224. [Google Scholar] [CrossRef]
- Pal, K.; Patel, B.V. Data Classification with K-Fold Cross Validation and Holdout Accuracy Estimation Methods with 5 Different Machine Learning Techniques. In Proceedings of the Proceedings of the 4th International Conference on Computing Methodologies and Communication, ICCMC 2020, Erode, India, 11–13 March 2020; Institute of Electrical and Electronics Engineers Inc.: Piscataway, NJ, USA, 2020; pp. 83–87. [Google Scholar]
- Haq, A.U.; Li, J.P.; Khan, J.; Memon, M.H.; Nazir, S.; Khan, G.A.; Ali, A. Intelligent Machine Learning Approach for Effective Recognition of Diabetes in E-Healthcare Using Clinical Data. Sensors 2020, 20, 2649. [Google Scholar] [CrossRef] [PubMed]
- Lakshmanan, V.; Robinson, S.; Munn, M. Machine Learning Design Patterns; O’Reilly Media, Inc.: Newton. MA, USA, 2020. [Google Scholar]
- Reitermanová, Z. Data Splitting. WDS’10 Proceedings of Contributed Papers(Part I); MatfyzPress: Prague, Czech Republic, 2010; pp. 31–36. [Google Scholar]
- Fox, M.; Schoeffmann, K. The Impact of Dataset Splits on Classification Performance in Medical Videos. In Proceedings of the ICMR 2022—2022 International Conference on Multimedia Retrieval, Newark, NJ, USA, 27–30 June 2022; Association for Computing Machinery, Inc.: Times Square, NYC, USA, 2022; pp. 6–10. [Google Scholar]
- Leibetseder, A.; Petscharnig, S.; Primus, M.; Kietz, S.; Münzer, B.; Schoeffmann, K.; Keckstein, J. Lapgyn4: A Dataset for 4 Automatic Content Analysis Problems in the Domain of Laparoscopic Gynecology. In Proceedings of the 9th ACM Multimedia Systems Conference, Amsterdam, The Netherlands, 12–15 June 2018; pp. 357–362. [Google Scholar]
- Shin, H.; Oh, S. Feature-Weighted Sampling for Proper Evaluation of Classification Models. Appl. Sci. 2021, 11, 2039. [Google Scholar] [CrossRef]
- Kang, D.; Oh, S. Balanced Training/Test Set Sampling for Proper Evaluation of Classification Models. Intell. Data Anal. 2020, 24, 5–18. [Google Scholar] [CrossRef]
- Eliane Birba, D. A Comparative Study of Data Splitting Algorithms for Machine Learning Model Selection. Master’s Thesis, KTH Royal Institute of Technology, Stocksund, Sweden, 2020. [Google Scholar]
- Farias, F.; Ludermir, T.; Bastos-Filho, C. Similarity Based Stratified Splitting: An Approach to Train Better Classifiers. arXiv 2020, arXiv:2010.06099. [Google Scholar]
- Nurhopipah, A.; Hasanah, U. Dataset Splitting Techniques Comparison For Face Classification on CCTV Images. Indonesian J. Comput. Cybern. Syst. 2020, 14, 341. [Google Scholar] [CrossRef]
- Lakshmi, M.J.; Nagaraja Rao, S. Effect Of K Fold Cross Validation on Mri Brain Images Using Support Vector Machine Algorithm. Int. J. Recent Technol. Eng. 2019, 7, 2277–3878. [Google Scholar]
- Sharma, R.C.; Hara, K.; Hirayama, H. A Machine Learning and Cross-Validation Approach for the Discrimination of Vegetation Physiognomic Types Using Satellite Based Multispectral and Multitemporal Data. Scientifica 2017, 2017, 9806479. [Google Scholar] [CrossRef] [PubMed]
- Varma, S.; Simon, R. Bias in Error Estimation When Using Cross-Validation for Model Selection. BMC Bioinform. 2006, 7, 91. [Google Scholar] [CrossRef] [PubMed]
- Vabalas, A.; Gowen, E.; Poliakoff, E.; Casson, A.J. Machine Learning Algorithm Validation with a Limited Sample Size. PLoS ONE 2019, 14, e0224365. [Google Scholar] [CrossRef]
- Kahloot, K.M.; Ekler, P. Algorithmic Splitting: A Method for Dataset Preparation. IEEE Access 2021, 9, 125229–125237. [Google Scholar] [CrossRef]
- McInnes, L.; Healy, J.; Astels, S. Hdbscan: Hierarchical Density Based Clustering. J. Open Source Softw. 2017, 2, 205. [Google Scholar] [CrossRef]
- He, X.; Cai, D.; Shao, Y.; Bao, H.; Han, J. Laplacian Regularized Gaussian Mixture Model for Data Clustering. IEEE Trans. Knowl. Data Eng. 2011, 23, 1406–1418. [Google Scholar] [CrossRef]
- Amidan, B.G.; Ferryman, T.A.; Cooley, S.K. Data Outlier Detection Using the Chebyshev Theorem. In Proceedings of the 2005 IEEE Aerospace Conference, Big Sky, MT, USA, 5–12 March 2005; pp. 3814–3819. [Google Scholar]
- Szegedy, C.; Vanhoucke, V.; Ioffe, S.; Shlens, J.; Wojna, Z. Rethinking the Inception Architecture for Computer Vision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Boston, MA, USA, 7–12 June 2015. [Google Scholar]
- Doan, Q.H.; Mai, S.H.; Do, Q.T.; Thai, D.K. A Cluster-Based Data Splitting Method for Small Sample and Class Imbalance Problems in Impact Damage Classification [Formula Presented]. Appl. Soft Comput. 2022, 120, 108628. [Google Scholar] [CrossRef]
- Christias, P.; Mocanu, M. A Machine Learning Framework for Olive Farms Profit Prediction. Water 2021, 13, 3461. [Google Scholar] [CrossRef]
- Huang, S.; Liu, W.; Qi, F.; Yang, K. Development and Validation of a Deep Learning Algorithm for the Recognition of Plant Disease. In Proceedings of the 21st IEEE International Conference on High Performance Computing and Communications, 17th IEEE International Conference on Smart City and 5th IEEE International Conference on Data Science and Systems, HPCC/SmartCity/DSS 2019, Zhangjiajie, China, 10–12 August 2019; Institute of Electrical and Electronics Engineers Inc.: Piscataway, NJ, USA, 2019; pp. 1951–1957. [Google Scholar]
- Durai, S.K.S.; Shamili, M.D. Smart Farming Using Machine Learning and Deep Learning Techniques. Decis. Anal. J. 2022, 3, 100041. [Google Scholar] [CrossRef]
- Alibabaei, K.; Gaspar, P.D.; Lima, T.M.; Campos, R.M.; Girão, I.; Monteiro, J.; Lopes, C.M. A Review of the Challenges of Using Deep Learning Algorithms to Support Decision-Making in Agricultural Activities. Remote Sens. 2022, 14, 638. [Google Scholar] [CrossRef]
- Rao, M.S.; Singh, A.; Reddy, N.V.S.; Acharya, D.U. Crop Prediction Using Machine Learning. In Proceedings of the 2020 Third International Conference on Smart Systems and Inventive Technology (ICSSIT), Tirunelveli, India, 20–22 August 2020; IOP Publishing Ltd.: Bristol, UK, 2022; Volume 2161. [Google Scholar]
- Suescún, M.F.R. Machine Learning Approaches for Tomato Crop Yield Prediction in Precision Agriculture. Master’s Thesis, Universidade Nova de Lisboa, Lisbon, Portugal, August 2020. [Google Scholar]
- Akiba, T.; Sano, S.; Yanase, T.; Ohta, T.; Koyama, M. Optuna: A Next-Generation Hyperparameter Optimization Framework. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, Anchorage, AK, USA, 4–8 August 2019. [Google Scholar]
- Nguyen, Q.H.; Ly, H.B.; Ho, L.S.; Al-Ansari, N.; Van Le, H.; Tran, V.Q.; Prakash, I.; Pham, B.T. Influence of Data Splitting on Performance of Machine Learning Models in Prediction of Shear Strength of Soil. Math. Probl. Eng. 2021, 2021, 4832864. [Google Scholar] [CrossRef]
- Zhang, Y.; Bakasa, W.; Viriri, S. VGG16 Feature Extractor with Extreme Gradient Boost Classifier for Pancreas Cancer Prediction. J. Imaging 2023, 9, 138. [Google Scholar] [CrossRef]
- Lachmann, M.; Rippen, E.; Rueckert, D.; Schuster, T.; Xhepa, E.; von Scheidt, M.; Pellegrini, C.; Trenkwalder, T.; Rheude, T.; Stundl, A.; et al. Harnessing Feature Extraction Capacities from a Pre-Trained Convolutional Neural Network (VGG-16) for the Unsupervised Distinction of Aortic Outflow Velocity Profiles in Patients with Severe Aortic Stenosis. Eur. Heart J. Dig. Health 2022, 3, 153–168. [Google Scholar] [CrossRef] [PubMed]
- Sharma, S.; Guleria, K.; Tiwari, S.; Kumar, S. A Deep Learning Based Convolutional Neural Network Model with VGG16 Feature Extractor for the Detection of Alzheimer Disease Using MRI Scans. Meas. Sensors 2022, 24, 100506. [Google Scholar] [CrossRef]
- Deng, J.; Dong, W.; Socher, R.; Li, L.J.; Li, K.; Fei-Fei, L. ImageNet: A Large-Scale Hierarchical Image Database. In Proceedings of the 2009 IEEE Conference on Computer Vision and Pattern Recognition, Miami, FL, USA, 20–25 June 2009; Institute of Electrical and Electronics Engineers (IEEE): Piscataway, NJ, USA, 2010; pp. 248–255. [Google Scholar]
- Lan, T.; Erdogmus, D.; Black, L.; Van Santen, J. A Comparison of Different Dimensionality Reduction and Feature Selection Methods for Single Trial ERP Detection. In Proceedings of the 2010 Annual International Conference of the IEEE Engineering in Medicine and Biology, Buenos Aires, Argentina, 31 August–4 September 2010. [Google Scholar] [CrossRef]
- Velliangiri, S.; Alagumuthukrishnan, S.; Thankumar Joseph, S.I. A Review of Dimensionality Reduction Techniques for Efficient Computation. In Proceedings of the Procedia Computer Science, Surabaya, Indonesia, 23–24 July 2019; Elsevier B.V.: Amsterdam, The Netherlands, 2019; Volume 165, pp. 104–111. [Google Scholar]
- Salem, N.; Hussein, S. Data Dimensional Reduction and Principal Components Analysis. In Proceedings of the Procedia Computer Science, Surabaya, Indonesia, 23–24 July 2019; Elsevier B.V.: Amsterdam, The Netherlands, 2019; Volume 163, pp. 292–299. [Google Scholar]
- Tzutalin LabelImg: Image Annotation Tool. Available online: https://github.com/tzutalin/labelImg (accessed on 20 November 2023).
- Syakur, M.A.; Khotimah, B.K.; Rochman, E.M.S.; Satoto, B.D. Integration K-Means Clustering Method and Elbow Method for Identification of the Best Customer Profile Cluster. In Proceedings of the IOP Conference Series: Materials Science and Engineering, Surabaya, Indonesia, 9 November 2017; Institute of Physics Publishing, Ltd.: Bristol, UK, 2018; Volume 336. [Google Scholar]
- Marutho, D.; Handaka, S.H.; Wijaya, E. The Determination of Cluster Number at K-Mean Using Elbow Method and Purity Evaluation on Headline News. In Proceedings of the 2018 International Seminar on Application for Technology of Information and Communication, Semarang, Indonesia, 21–22 September 2018; pp. 533–538. [Google Scholar]
Class Name | Códega | Moscatel | Rabigato | Tinta Roriz | Tinto Cão | Touriga Nacional |
Size | 80 |
Class name | Absence of Joint Cover Plate | Cracks | Paint Deterioration | Pavement Crack | Peeling on Concrete | Plant |
Size | 149 | 461 | 406 | 523 | 101 | 232 |
Dataset Split Approaches | Xception | ||||
---|---|---|---|---|---|
Nadam | SGD | ||||
Acc | ESE | Acc | ESE | ||
SCDIS | 0.90 | 100 | 0.86 | 100 | |
FSCIS | 0.88 | 100 | 0.81 | 100 | |
Hold out | 0.89 | 100 | 0.82 | 100 | |
KFCV10 | Fold 0 | 0.89 | 71 | 0.86 | 100 |
Fold 1 | 0.88 | 100 | 0.84 | 100 | |
Fold 2 | 0.87 | 100 | 0.82 | 100 | |
Fold 3 | 0.90 | 100 | 0.81 | 100 | |
Fold 4 | 0.88 | 100 | 0.82 | 100 | |
Fold 5 | 0.88 | 100 | 0.85 | 100 | |
Fold 6 | 0.89 | 100 | 0.83 | 100 | |
Fold 7 | 0.89 | 100 | 0.82 | 100 | |
Fold 8 | 0.88 | 89 | 0.84 | 100 | |
Fold 9 | 0.88 | 100 | 0.83 | 100 | |
Mean KFCV10 | 0.88 | 0.73 |
Dataset Split Approaches | Xception | ||||
---|---|---|---|---|---|
Nadam | SGD | ||||
Acc | ESE | Acc | ESE | ||
SCDIS | 0.70 | 100 | 0.67 | 100 | |
FSCIS | 0.76 | 100 | 0.74 | 100 | |
Hold out | 0.75 | 100 | 0.69 | 100 | |
KFCV10 | Fold 0 | 0.70 | 100 | 0.67 | 100 |
Fold 1 | 0.65 | 100 | 0.70 | 100 | |
Fold 2 | 0.70 | 98 | 0.68 | 100 | |
Fold 3 | 0.74 | 100 | 0.71 | 100 | |
Fold 4 | 0.68 | 94 | 0.68 | 100 | |
Fold 5 | 0.68 | 100 | 0.65 | 100 | |
Fold 6 | 0.65 | 92 | 0.67 | 100 | |
Fold 7 | 0.61 | 48 | 0.60 | 100 | |
Fold 8 | 0.65 | 84 | 0.67 | 100 | |
Fold 9 | 0.67 | 100 | 0.71 | 100 | |
Mean KFCV10 | 0.67 | 0.67 |
Dataset Split Approaches | Xception | ||
---|---|---|---|
Nadam | SGD | ||
Acc | Acc | ||
SCDIS | 0.75 | 0.68 | |
FSCIS | 0.75 | 0.72 | |
Hold out | 0.72 | 0.70 | |
KFCV10 | Fold 0 | 0.72 | 0.73 |
Fold 1 | 0.72 | 0.77 | |
Fold 2 | 0.77 | 0.70 | |
Fold 3 | 0.73 | 0.73 | |
Fold 4 | 0.68 | 0.75 | |
Fold 5 | 0.77 | 0.67 | |
Fold 6 | 0.75 | 0.75 | |
Fold 7 | 0.75 | 0.70 | |
Fold 8 | 0.73 | 0.70 | |
Fold 9 | 0.72 | 0.68 | |
Mean KFCV10 | 0.73 | 0.72 |
Dataset Split Approaches | Xception | ||
---|---|---|---|
Nadam | SGD | ||
Acc | Acc | ||
SCDIS | 0.32 | 0.27 | |
FSCIS | 0.25 | 0.22 | |
Hold out | 0.22 | 0.20 | |
KFCV10 | Fold 0 | 0.25 | 0.25 |
Fold 1 | 0.25 | 0.22 | |
Fold 2 | 0.25 | 0.28 | |
Fold 3 | 0.23 | 0.28 | |
Fold 4 | 0.32 | 0.28 | |
Fold 5 | 0.23 | 0.27 | |
Fold 6 | 0.23 | 0.23 | |
Fold 7 | 0.25 | 0.25 | |
Fold 8 | 0.27 | 0.23 | |
Fold 9 | 0.25 | 0.30 | |
Mean KFCV10 | 0.25 | 0.26 |
Class | SCDIS | FSCIS | HO | KFCV10—Fold3 | ||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
µAcc | µIoU | µGCS | µDis | µAcc | µIoU | µGCS | µDis | µAcc | µIoU | µGCS | µDis | µAcc | µIoU | µGCS | µDis | |
Codega | 0.00 | 0.12 | 4393 | 85 | 0.00 | 0.13 | 4880 | 82 | 0.00 | 0.10 | 3541 | 87 | 0.00 | 0.11 | 4139 | 86 |
Moscatel | 0.90 | 0.173 | 7349 | 69 | 0.50 | 0.16 | 6311 | 72 | 0.70 | 0.14 | 5396 | 75 | 0.00 | 0.15 | 4996 | 70 |
Rabigato | 0.60 | 0.19 | 7678 | 67 | 0.80 | 0.19 | 7478 | 70 | 0.50 | 0.15 | 6099 | 76 | 0.70 | 0.15 | 6000 | 72 |
Tinta Roriz | 0.20 | 0.24 | 8231 | 63 | 0.10 | 0.23 | 7731 | 62 | 0.10 | 0.24 | 7956 | 62 | 0.00 | 0.24 | 8526 | 62 |
Tinto Cao | 0.00 | 0.21 | 7961 | 58 | 0.00 | 0.23 | 8911 | 58 | 0.00 | 0.16 | 6467 | 62 | 0.10 | 0.15 | 6415 | 65 |
Touriga Nacional | 0.20 | 0.23 | 7707 | 67 | 0.10 | 0.21 | 7652 | 70 | 0.00 | 0.28 | 8784 | 63 | 0.00 | 0.22 | 7915 | 70 |
Total mean | 0.38 | 0.19 | 7220 | 68 | 0.25 | 0.19 | 7160 | 69 | 0.22 | 0.18 | 6374 | 71 | 0.13 | 0.17 | 6332 | 71 |
Class | SCDIS | FSCIS | HO | KFCV10—Fold3 | ||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
µAcc | µIoU | µGCS | µDis | µAcc | µIoU | µGCS | µDis | µAcc | µIoU | µGCS | µDis | µAcc | µIoU | µGCS | µDis | |
AJCP | 0.90 | 0.18 | 42,969 | 139 | 0.90 | 0.20 | 48,068 | 137 | 0.80 | 0.19 | 46,502 | 140 | 0.80 | 0.18 | 45,516 | 141 |
Cracks | 0.70 | 0.21 | 39,203 | 126 | 0.70 | 0.21 | 38,771 | 129 | 0.40 | 0.18 | 38,429 | 132 | 0.60 | 0.20 | 37,225 | 127 |
Painting | 0.50 | 0.20 | 36,071 | 124 | 0.50 | 0.18 | 32,077 | 117 | 0.90 | 0.20 | 33,226 | 107 | 0.70 | 0.21 | 36,871 | 118 |
PC | 0.90 | 0.16 | 27,457 | 149 | 0.90 | 0.16 | 26,962 | 148 | 0.90 | 0.15 | 28,107 | 147 | 1.00 | 0.15 | 28,070 | 147 |
PoC | 0.50 | 0.17 | 30,665 | 125 | 0.50 | 0.17 | 30,473 | 126 | 0.50 | 0.17 | 31,801 | 125 | 0.50 | 0.18 | 31,065 | 124 |
Plant | 1.00 | 0.17 | 49,259 | 144 | 1.00 | 0.16 | 48,878 | 146 | 0.50 | 0.16 | 48,549 | 145 | 0.80 | 0.17 | 49,576 | 145 |
Total mean | 0.75 | 0.18 | 37,604 | 135 | 0.75 | 0.18 | 37,538 | 134 | 0.67 | 0.18 | 37,769 | 133 | 0.73 | 0.18 | 38,054 | 134 |
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. |
© 2024 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/).
Share and Cite
Shahrabadi, S.; Adão, T.; Peres, E.; Morais, R.; Magalhães, L.G.; Alves, V. Automatic Optimization of Deep Learning Training through Feature-Aware-Based Dataset Splitting. Algorithms 2024, 17, 106. https://doi.org/10.3390/a17030106
Shahrabadi S, Adão T, Peres E, Morais R, Magalhães LG, Alves V. Automatic Optimization of Deep Learning Training through Feature-Aware-Based Dataset Splitting. Algorithms. 2024; 17(3):106. https://doi.org/10.3390/a17030106
Chicago/Turabian StyleShahrabadi, Somayeh, Telmo Adão, Emanuel Peres, Raul Morais, Luís G. Magalhães, and Victor Alves. 2024. "Automatic Optimization of Deep Learning Training through Feature-Aware-Based Dataset Splitting" Algorithms 17, no. 3: 106. https://doi.org/10.3390/a17030106
APA StyleShahrabadi, S., Adão, T., Peres, E., Morais, R., Magalhães, L. G., & Alves, V. (2024). Automatic Optimization of Deep Learning Training through Feature-Aware-Based Dataset Splitting. Algorithms, 17(3), 106. https://doi.org/10.3390/a17030106