Bonfring International Journal of Networking Technologies and Applications, Vol. 11, Issue 1, February 2024 17 Modified Fuzzy Neural Network Approach for Academic Performance Prediction of Students in Early Childhood Education Marwah Hameed Abstract--- Modern education relies heavily on educational technology, which provides students with unique learning opportunities and enhances their ability to learn. For many years now, computers and other technological tools have been an integral part of education. However, compared to other educational levels, the incorporation of educational technology in early childhood education is a more recent trend. It is because of this that materials and procedures tailored to young children must be created, implemented, and studied. The use of artificial intelligence techniques in educational technology resources has resulted in better engagement for students. Early childhood special education students' academic achievement is predicted using a Modified Fuzzy Neural Network (MFNN). Before constructing the classifier, the dataset had to be preprocessed to remove any extraneous information. As a follow-up, this study will put to the test an organized approach to the implementation of customized fuzzy neural networks for the prediction of academic achievement in early childhood settings. Considerations for the analysis of academic achievement in early childhood education are discussed in this article, including recommendations for the implementation of proposed modified fuzzy neural networks. In terms of evaluation metrics such as Precision, recall, accuracy, and the F1 coefficient, the proposed model outperforms conventional machine-learning (ML) techniques. Keywords--- Early Childhood Special Education, Computer-based Learning System, Artificial Intelligence, Modified Fuzzy Neural Network. I. INTRODUCTION E DUCATIONAL technology includes computer-based learning. Computers have been used in education since the 1950s, and students and teachers can utilize them independently or in teams (Wolery, M., et al., 2002). However, educational technology generally combines resources other than computers to maximize each resource's unique traits and benefits. Particularly in early childhood education (McConnell, S. R. (2000)). Besides computers, interactive whiteboards and programmable toys are commonly employed in early childhood education. Using instructional technology has several benefits. Educational technology may motivate pupils to learn by attracting their attention and encouraging innovative actions (Lifter, K., et al., (2011)). The utilization of technology allows for unique instructional characteristics such as multimedia- based engagement and problem-solving process visualization. Technology also fosters collaborative learning and constructivism (Odom, S. L., & Wolery, M. (2003)). Teaching pupils about educational technology helps them understand the Information Society. Finally, technology may help schools connect with their communities. Artificial Intelligence (AI) is used in many fields. Educational technology is an interesting topic for AI (Warren, S.F. (2000)). Since the 1970s, Artificial Intelligence has been used in instructional technologies. E-learning is a broad phrase [6]. It is the use of instructional technology to meet specific educational needs. The emphasis on new resources in educational technology often ignores older but still important tools (Warren, S.F., & Walker, D. (2005)). The major goal is to assist students and teachers over traditional techniques. Using instructional technology in the classroom might be difficult. The integration process should address concerns specific to a student group (Schwartz, I. S. (2000)). Technology can help solve specific educational issues or offer the infrastructure for activities that would not be possible without it. Creating a good prediction model improves the forecast range. Predicting academic performance in early childhood special education using MFNN. Section 2 describes the recommended technique for the rest of the research. Section 3 presents the findings. Section 4 concludes and plans future work. II. PROPOSED METHODOLOGY For predicting academic success in early childhood special education, an MFNN is presented. Primarily, preprocessing eliminates unnecessary data from the dataset, increasing the classifier's prediction performance. Another goal of this research is to apply modified fuzzy neural networks to predict academic achievement in early childhood education. Considerations for analyzing academic success in early childhood education are discussed in this article. Figure 1 depicts the suggested methodology's overall procedure. Marwah Hameed, Department of Computer Science, College of Computer Science and Information Technology, University of Kirkuk, Kirkuk, Iraq. DOI: 10.9756/BIJNTA/V11I1/BIJ24007 ISSN 2320 - 5377 | © 2024 Bonfring Bonfring International Journal of Networking Technologies and Applications, Vol. 11, Issue 1, February 2024 Input Data Base (ICFES) Data Preprocessing - Z-score normalization Classification by Modified Fuzzy Neural Network (MFNN) Performance Measures Figure 1: The Overall Process of the Proposed Methodology 1. Data Preprocessing Using Z-Score Normalization Each experiment's basic intensity data were normalized by computing the average intensity for each dataset, then the average of the averages [9]. This grand average was used to compute normalization factors for each experiment. The grand average was then equaled by all normalized data. A z-normal score's distribution curve. There is a -3 standard deviation (far left of the normal distribution curve) to a +3 standard deviation range (fall to the far right of the normal distribution curve). In order to use a z-score, you need to know the mean μ and also the population standard deviation σ. Esp, let xi (i = 1, 2, · · ·, D) represents the i-th component of each feature vector x ∈ R D. The mean and the standard deviation of these D components are evaluated as: 𝜇𝑥 = 1 1 ∑𝐷 𝑥 , 𝜎 𝐷 𝑖=1 𝑖 𝑥 = = 𝑍𝑁(𝑥) = 𝑥−𝜇𝑥 1 √ ∑𝐷 (𝑥 𝐷 𝑖=1 𝑖 2 − 𝜇𝑥 ) (1) Z-score normalization is then applied as, 𝑥 (𝑧𝑛) 𝜎𝑥 18 calculates the net value using an LR type fuzzy number and so does not presume criteria independence. The FBP algorithm also avoids oscillations and falls into local minima. The convergence of the FBP algorithm for single-output networks with single and multiple training patterns is proven. 3. Fuzzy Backpropagation Algorithm (FBP Algorithm) Different neuro-fuzzy approaches have recently been presented for calculating the net value of the ith neuron's inputs. The mapping is mathematically represented by Sugeno's fuzzy integral, which is based on a psychological foundation. Step 1: Randomly create the initial weight sets w for the input hidden layer in which each 𝑤𝑗𝑖 = (𝑤𝑚𝑗𝑖 , 𝑤𝛼𝑗𝑖 , 𝑤𝛽𝑗𝑖 ) is an LR-type fuzzy number. And create the weight set w’ for the hidden output layer ′ ′ ′ ′ Here 𝑤𝑘𝑗 = (𝑤𝑚𝑘𝑗 , 𝑤𝛼𝑘𝑗 , 𝑤𝛽𝑘𝑗 ) 𝑤𝑗𝑖 = (𝑤𝑚𝑗𝑖 , 𝑤𝛼𝑗𝑖 , 𝑤𝛽𝑗𝑖 ) ′ ′ ′ ′ 𝑤𝑘𝑗 = (𝑤𝑚𝑘𝑗 , 𝑤𝛼𝑘𝑗 , 𝑤𝛽𝑘𝑗 ) Step 2: Consider (𝐼𝑝 , 𝐷𝑝 ) 𝑝 = 1, 20 … 𝑁 input-output pattern set. In which 𝐼𝑝 = (𝐼𝑝0 , 𝐼𝑝1 , 𝐼𝑝1 ) also every 𝐼𝑝𝑖 is an LR-type fuzzy number. Step 3: Allocate values for α and η; Alpha=0.1 Neta =0.9 Step 4: Acquire next pattern set (𝐼𝑝 , 𝐷𝑝 ) Assign (𝑂𝑝𝑖 = 𝐼𝑝𝑖 , i=1,2,3..1 Step 5: Calculate the input to hidden neurons ′ ′ 𝑂𝑝𝑗 = 𝑓(𝑁𝐸𝑇𝑝𝑗 ), 𝑗 = 1,2 … . , 𝑚; 𝑂𝑝0 =1 Where 𝑁𝐸𝑇𝑝𝑗 = 𝐶𝐸 (∑ 𝑊𝑗𝑖 𝑂𝑝𝑖 ) Step 6: Evaluate the hidden to output neurons 𝑂’’𝑝𝑘 = 𝑓 (𝑁𝐸𝑇’𝑝𝑘 ), 𝑘 = 1,2, . . 𝑛; (2) Based on these calculations, z-score normalization extends the original feature vectors along the 1 vector to a hyperplane including the origin and being perpendicular to √ 1. These vectors are then adjusted to have a similar length as D, resulting in final normalized vectors that lie on a hypersphere of radius √D. Next the preprocessing of the given data, the feature selection procedure is carried out, as explained in the following section. 2. Classification Using MFNN Neuronal networks and fuzzy logic are emerging technologies that could be used in pharmaceutical formulation and processing (Yang, B., et al., (2007)). ANNs and evolutionary algorithms work well together to forecast and optimize formulation conditions. Fuzzy-neural systems seem to have flourished more than other methods of symbolic connectionism. A fuzzy neural network has three layers: an input layer (fuzzification), a hidden layer (fuzzy rules), and an output layer (fuzzification) (defuzzification). Sometimes a five-layer network containing sets in the second and fourth layers can be found. In practice, the criteria are connected. The linear evaluation function cannot capture inter-criteria relationships. To solve the SBP algorithm's disadvantage. This paper proposes a Fuzzy Backpropagation (FBP) technique. It Where 𝑁𝐸𝑇’𝑝𝑘 = 𝐶𝐸 (∑ 𝑊𝑗𝑖 𝑂’𝑝𝑗 ) Step 7: Evaluate modification of weights ∆ w’(t) for the hidden output layer as below Evaluate ∆𝐸𝑝 (𝑡) = (𝜕𝐸𝑝 /𝜕𝑤’𝑚𝑘𝑗 , 𝜕𝐸𝑝 /𝜕𝑤𝛼𝑘𝑗, 𝜕𝐸𝑝/𝜕𝑤’𝛽𝑘𝑗 ) Evaluate ∆𝑤’(𝑡) = −𝜂∆𝐸𝑝 (𝑡) + 𝛼∆𝑤’(𝑡 − 1) The modified weight i of hidden to output neuron is 𝑊’(𝑡) = 𝑊’(𝑡 − 1) + ∆𝑊’(𝑡) Step 8: Calculate modification of the weights ∆ w’(t) for the input hidden layer as follows Let 𝛿𝑝𝑚𝑘 = −(𝐷𝑝𝑘 − 𝑂’’𝑝𝑘 )𝑂’’𝑝𝑘 (1 − 𝑂’’𝑝𝑘 ).1 1 𝛿𝑝𝑚𝑘 = −(𝐷𝑝𝑘 − 𝑂’’𝑝𝑘 )𝑂’’𝑝𝑘 (1 − 𝑂’’𝑝𝑘 ). (− ) 3 1 𝛿𝑝𝑚𝑘 = −(𝐷𝑝𝑘 − 𝑂’’𝑝𝑘 )𝑂’’𝑝𝑘 (1 − 𝑂’’𝑝𝑘 ). ( ) 3 ISSN 2320 - 5377 | © 2024 Bonfring Bonfring International Journal of Networking Technologies and Applications, Vol. 11, Issue 1, February 2024 Evaluate ∆𝐸𝑝 (𝑡) = (𝜕𝐸𝑝 /𝜕𝑤’𝑚𝑗𝑖 , 𝜕𝐸𝑝 /𝜕𝑤𝛼𝑗𝑖 , 𝜕𝐸𝑝/𝜕𝑤’𝛽𝑗𝑖 ) Table 1: Performance results of the proposed and existing prediction methods Calculate ∆𝑤’(𝑡) = −𝜂∆𝐸𝑝 (𝑡) + 𝛼∆𝑤’(𝑡 − 1) Step 9: Modify weight for the input-hidden-output layer as, 𝑊(𝑡) = 𝑊(𝑡 − 1) + ∆𝑊(𝑡) 𝑊’(𝑡) = 𝑊’(𝑡 − 1) + ∆𝑊’(𝑡) 19 Metrics SVM ANN FNN Accuracy 91.24 93.58 99.57 Precision 71.45 84.67 91.58 Recall 78.24 86.57 92.51 F-measure 87.24 92.57 98.24 Table 1. tabulate the performance results of the proposed and existing prediction methods. Step 10: 𝑝 = 𝑝 + 1; if (p<=N) go to step 5 Step11: output w’ and w’’ the final weight sets. cos ((𝜋⁄𝐺𝑚𝑎𝑥 )×𝑇)+2.5 4 cos ((𝜋×𝑇 ⁄𝐺𝑚𝑎𝑥 ))×2.5 4 100 78.24 86.57 92.51 71.45 84.67 91.58 0 ANN MFNN METHODS (3) In which T is the number of iterations. Assume G max= 40, the changing curve of value K arrived. Formula (3) is described below: 𝑣𝑖𝑑 = ( Recall 200 SVM ) × [𝑣𝑖𝑑 + 2 × 𝑟𝑎𝑛𝑑() × (𝑝𝑖𝑑 − 𝑥𝑖𝑑 ) + 2 × 𝑅𝑎𝑛𝑑() × (𝑝𝑔𝑑 − 𝑥𝑖𝑑 )] (4) Here 𝑉𝑖𝑑 is the regularity distribution factor, and the decreasing 𝑉𝑖𝑑 value is dispersed in combination with the rand function. The modified number of leaders per iteration is 𝑉𝑖𝑑 ·N and the number of followers is equal to 1 − 𝑉𝑖𝑑 ·N. III. RESULTS AND DISCUSSION The ICFES collected the information for this research. Approximately 200,000 Colombian university students took the SABER PRO test in 2016. These included data on each student's SABER 11 test results, socioeconomic status, childhood school characteristics, and academic status. The original data set included student gender, age, and academic program. The pupils' identities were kept secret because they were coded in the ICFES data collection. TP, FP, TN, and FN rates are used to determine various performance measures. The first performance metric was precision or the fraction of relevant retrieved occurrences. Remember that recall is defined as the proportion of relevant instances retrieved. The measurements of accuracy and recall are both significant in evaluating a prediction approach's success. So these two metrics can be merged with equal weights to get the F-measure. Accuracy is the proportion of accurately predicted instances to all expected instances. Figure 2: Precision and recall results between the proposed and existing methods Figure 2. shows the proposed MFNN technique gives high value of Precision and recall than the existing classifier. From the results it is identified that the proposed algorithm is highly effective. So the performance of the proposed model will be higher compared to other classifier built on previously generated model. F-measure Prediction Percentage (%) 𝐾= Prediction Percentage(%) 4. Regularity Distribution Factor To reduce the likelihood of failure in iterations, the regularity distribution factor should select a convex function in the early iterations, allowing the population to find an optimal solution over a large range. In the late phase, a concave function should be chosen so that the regularity distribution factor can gradually change to the minimum for local development to occur. It ensures the algorithm's convergence. The functional regularity distribution factor structuring on the basis of the cosine function is demonstrated in formula (3): precision Accuracy 100% 91.24 93.58 99.57 87.24 92.57 98.24 50% 0% SVM ANN MFNN METHODS Figure 3: Accuracy results between the proposed and existing methods Figure.3. show the relationship between the experimental and the MFNN based learning predicted results on the SVM and ANN-based methods. The result indicates that the proposed MFNN based learning can greatly improve the accuracy prediction among the different methods. ISSN 2320 - 5377 | © 2024 Bonfring Bonfring International Journal of Networking Technologies and Applications, Vol. 11, Issue 1, February 2024 IV. CONCLUSION This study examines the use of Artificial Intelligence in early childhood education. For predicting academic success in early childhood special education, an MFNN is suggested. It was designed to classify students' academic achievement using numerous MFNNs. This conclusion may be explained by the fact that different academic programs attract students with diverse abilities and interests. The substance of each academic program may also have influenced student preparation and performance. Thus, the predictive efficacy of selected academic performance predictors may vary by discipline. The suggested model outperforms the existing techniques in terms of prediction accuracy. So more topologies with different learning paradigms should be investigated. Furthermore, determining MFNN confidence and prediction intervals requires more research. REFERENCES M. Wolery and D.B. Bailey Jr, “Early childhood special education research”, Journal of early intervention, Vol. 25, No. 2, Pp. 88-99, 2002. [2] S.R. McConnell, “Assessment in early intervention and early childhood special education: Building on the past to project into our future”, Topics in Early Childhood Special Education, Vol. 20, No. 1, Pp. 43-48, 2000. [3] K. Lifter, S. Foster-Sanda, C. Arzamarski, J. Briesch and E. McClure, “Overview of play: Its uses and importance in early intervention/early childhood special education”, Infants & Young Children, Vol. 24, No. 3, Pp. 225-245, 2011. [4] S.L. Odom and M. Wolery, “A unified theory of practice in early intervention/early childhood special education: Evidence-based practices”, The Journal of Special Education, Vol. 37, No. 3, Pp. 164-173, 2003. [5] S.F. Warren, “The future of early communication and language intervention”, Topics in early childhood special education, Vol. 20, No. 1, Pp. 33-37, 2000. [6] J.J. Carta, “An early childhood special education research agenda in a culture of accountability for results”, Journal of Early Intervention, Vol. 25, No. 2, Pp. 102-104, 2002. [7] S.F. Warren and D. Walker, “Fostering early communication and language development”, Handbook of research methods in developmental science, 249-270, 2005. [8] I.S. Schwartz, “Standing on the shoulders of giants: Looking ahead to facilitating membership and relationships for children with disabilities”, Topics in early childhood special education, Vol. 20, No. 2, Pp. 123-128, 2000. [9] C. Cheadle, M.P. Vawter, W.J. Freed and K.G. Becker, “Analysis of microarray data using Z score transformation”, The Journal of molecular diagnostics, Vol. 5, No. 2, Pp. 73-81, 2003. [10] B. Yang, L. Yao and H.Z. Huang, “Early software quality prediction based on a fuzzy neural network model”, In Third International Conference on Natural Computation (ICNC 2007), Vol. 1, Pp. 760-764, 2007. [1] ISSN 2320 - 5377 | © 2024 Bonfring 20