Analysis of requirements for AI education dataset

To analyze requirements for AI education datasets, we first investigated current dataset usage trends. The UCI ML Repository, which provides various types of datasets for AI modeling research, offers 664 different datasets. As shown in Fig. 3, users can check information such as appropriate modeling algorithms, number of variables, and access frequency. Based on access frequency, the most frequently used datasets were identified as ‘Iris’, ‘Dry Bean Dataset’, ‘Heart Disease’, ‘Rice’, and ‘Adult’³⁴.

To analyze the usage of specific datasets for educational purposes, we examined the dataset usage status in “Entry,” South Korea’s representative educational programming language that provides practical functions for AI education. As shown in Fig. 4, Entry is a visual programming language. This visual programming language helps reduce the difficulties associated with learning syntax and maintains students’ interest while they understand and learn the basic concepts of AI^35,36. Additionally, Entry provides basic resources for AI education and offers 19 datasets through the “Data Analysis” feature. In the Entry platform (https://playentry.org), you can view datasets by clicking the following menu options sequentially: [Create] – [Analyze data] – [Load tables] – [Add tables] – [Select tables]. It also enables AI modeling practices such as data visualization, linear regression, binary classification, multi-class classification, and clustering, and provides datasets with specified purposes for AI modeling, such as ‘Iris’, ‘Boston Housing’, ‘Palmer Penguins’, and ‘Titanic’³⁷.

Educational programming language entry and datasets in the visual environment.

We analyzed program outputs utilizing AI modeling features and datasets in Entry between December 31, 2020, and December 31, 2021, deriving the total dataset usage count from these artifacts. To analyze overall usage patterns, we visualized the utilization status of the top 10 most frequently used datasets as shown in Fig. 5.

Usage of AI education datasets in entry.

The top 10 most frequently used datasets in Entry were identified as ‘Iris’, ‘Population by City’, ‘Boston Housing’, followed by ‘Consumer Price Index’ in descending order. Detailed analysis of the visualized chart for AI modeling datasets reveals consistent usage of Iris and Boston Housing datasets, with Iris being used 7499 times and Boston Housing 6619 times during the study period. This figure demonstrates significantly higher usage compared to other AI modeling datasets that failed to rank within the top 10.

Both the UCI ML repository (a platform providing datasets for AI modeling) and Entry (an educational programming language platform) showed Iris as the most utilized dataset. The Iris dataset, composed of continuous independent variables and categorical dependent variables, is particularly suitable for multiclass classification tasks. Its high usage frequency suggests it serves as a representative dataset for AI modeling practice.

While widely-used datasets for AI modeling practice offer the advantage of easily accessible examples for various AI modeling and computing activities, they exhibit limitations including lack of relevance to students’ daily lives, difficulty in connecting to real-world contexts, and inability to provide authentic practical experiences. Notably, the Boston dataset has been discontinued in major machine learning libraries like scikit-learn due to ethical concerns, necessitating the development of alternative datasets³⁷.

Design and implement for AI education datasets

This study develops datasets for AI education by benchmarking Entry, a widely adopted educational programming platform with significant classroom impact. Through analysis of Entry’s technical specifications, we established requirements for datasets suitable for supervised and unsupervised learning implementations, as detailed in Table 10. The platform supports essential modeling algorithms including Linear Regression, Logistic Regression, k-Nearest Neighbors (kNN), Support Vector Machines (SVM), Classification and Regression Trees (CART), and Clustering. For model configuration, users can define up to 6 continuous variables as independent features, while dependent variables may incorporate either continuous or categorical data types depending on the learning task.

To provide AI educational datasets contextualized to students’ daily lives, we explored and structured preliminary dataset drafts as shown in Table 11, incorporating contextual frameworks from PISA 2022 Mathematics. We investigated public data platforms and diverse dataset sources while verifying appropriate variable inclusion for respective modeling methods. All datasets explicitly specify applicable licenses for educational purposes, with particular emphasis on exploring publicly available data centered around daily life topics likely to engage student interest. For certain datasets requiring specialized context, researchers directly collected and structured original data to complete the draft dataset compositions.

The draft datasets were systematically restructured according to AI modeling methodologies and educational objectives to facilitate effective utilization in AI education, as detailed in Table 12.

The datasets were primarily restructured according to AI modeling methods and educational objectives. For Linear Regression datasets, it was necessary to ensure completeness by specifying independent and dependent variables. The Seoul Mosquito Activity Index and Synoptic Meteorological Observation datasets were joined by date after establishing variable relationships through synthesis of prior research on meteorological environments and mosquito populations^42,43. Key variables were extracted and reorganized to enhance student comprehension and provide successful modeling experiences¹⁶.

The baseball game results dataset was collected from sports information websites, then completely synthesized using statistical simulation methods based on original data to minimize team-specific bias and enhance objectivity⁴⁴. The dataset was further restructured by extracting key variables influencing the dependent variable for student accessibility.

The body measurements and t-shirt size dataset, initially collected from students, was replaced with a fully synthetic version to address privacy concerns and improve size appropriateness through synthetic data generation techniques⁴⁴. This approach enhanced objectivity while resolving issues with limited original data scale.

The earthquake location dataset was restructured by removing entries below magnitude 2, which are classified as non-impactful seismic events based on domain expertise, to improve size appropriateness⁴¹.

Testing AI education dataset

Experts review

The evaluation results of the draft datasets through data quality assessment and authentic activity characteristics analysis, along with group interview findings, are summarized in Table 13.

Regarding overall feedback on the datasets, experts frequently noted that the developed AI education datasets showed high applicability due to their relevance to students’ daily lives, while emphasizing the need to provide concrete usage examples. Several reviewers suggested intentionally incorporating elements like data preprocessing activities to encourage diverse approaches and outcome variations among students.

The detailed specifications of the finalized AI education datasets, reflecting expert interview outcomes, are presented in Table 14.

In the ‘Mosquito activity index’ dataset, some data fields were found to contain uniformly input values during the data collection process. While some experts recommended preprocessing these values before providing the dataset to students, we ultimately preserved the uniformly input values to facilitate practical data preprocessing exercises in educational settings⁴⁴.

For the ‘Baseball game results’ dataset, we removed the ‘Team name’ column containing categorical information and revised variable names based on expert recommendations to enhance student comprehension.

The ‘T-shirt sizes’ dataset was recognized as particularly suitable for introductory AI education, especially for transparent understanding of decision tree models. Experts noted that variables like BMI showed high correlation with other factors, potentially causing multicollinearity issues if used as independent variables. Since addressing this through preprocessing might exceed students’ current capabilities, we simplified the dataset to essential ‘Height’ and ‘Weight’ variables. Additionally, we modified some data points to create overlapping size categories, addressing concerns about excessive model accuracy from overly distinct clusters.

For the ‘Earthquake information’ dataset, we addressed structural simplicity concerns by reintroducing preprocessed data points below magnitude 2 (previously excluded) and structuring the dataset to demonstrate clustering differences through preprocessing activities.

Review AI modeling accuracy

The AI education datasets developed through a constructivist lens, which are closely connected to students’ daily lives, must be effectively utilized for their intended educational purposes and should ultimately lead to the development of integrated intelligent systems as tangible outcomes of learners’ computational activities²². Prior to implementation, it is essential to evaluate the accuracy and usability of outputs – key factors that often hinder effective education using real-world data¹⁹. To address this, we conducted comprehensive testing of the developed datasets through modeling and evaluation using appropriate performance metrics including accuracy measures.

The ‘Mosquito activity index’ dataset is designed for linear regression analysis using continuous dependent and independent variables. In the Entry programming environment, setting one dependent and one independent variable allows visual confirmation of results, significantly enhancing students’ understanding of AI modeling principles. We selected ‘average mosquito activity index’ as the dependent variable and ‘average ground temperature’ as the independent variable based on their statistically significant correlation, implementing the model using Scikit-Learn’s LinearRegression. To validate the modeling results, we visualized the data and regression line as shown in Fig. 6, reserving 20% of the data for testing. We employed Mean Squared Error (MSE) and R-squared (R²) values, standard metrics for linear regression accuracy assessment, with results detailed in Table 15. Notably, we compared model accuracy between the original dataset containing uniformly input values and its preprocessed version to validate our initial dataset construction rationale.

Visualization results of linear regression of mosquito activity index dataset.

The comparative analysis revealed enhanced performance on test data after preprocessing, demonstrating that the refined linear model exhibits greater generalizability and explanatory power (R²-Test = 0.81). This dataset’s structure allows for modeling with various combinations of two or more independent variables, enabling comparative analysis of results and encouraging diverse student outcomes through multiple analytical approaches. These characteristics confirm the dataset’s effectiveness for both linear regression applications and comprehensive education about regression techniques, including preprocessing considerations.

The ‘Baseball game results’ dataset proves suitable for binary classification tasks using various dependent variables to predict game outcomes. Within the Entry programming environment, we implemented binary classification through TensorFlow, offering optional use of Adam Optimizer or SGD (Stochastic Gradient Descent) Optimizer. Through correlation analysis and variance inflation factor examination, we selected six independent variables (‘runs scored’, ‘triples’, ‘home runs’, ‘stolen bases’, ‘strikeouts’, and ‘double plays’) while excluding those showing multicollinearity. Using Keras framework, we constructed a neural network comprising a single fully connected layer with 32 neurons and a Sigmoid activation function. We evaluated both optimization approaches by reserving 20% of data for testing, visualizing accuracy/loss trajectories in Fig. 7. To ensure rigorous validation, we employed comprehensive metrics including accuracy, precision, recall, and F1-Score, supplemented by averaged results from 1,000 iterative modeling trials as detailed in Table 16.

An accuracy and loss graph according to the optimizer of the binary classification model.

The analysis revealed consistently high accuracy across all available optimization functions in the Entry programming environment. The 1,000 iterative measurements demonstrated robust mean accuracy with low standard deviation, confirming the dataset’s effectiveness for teaching binary classification concepts while allowing students to freely configure independent variables and explore diverse modeling approaches.

The ‘T-shirt sizes’ dataset is optimized for multiclass classification using categorical dependent variables. The Entry environment implements CART (Classification and Regression Tree) methodology for this purpose, where we designate the categorical ‘t-shirt size’ variable as the dependent feature and reserve 20% of data for testing. Using Scikit-Learn’s DecisionTreeClassifier, we established modeling parameters by setting the minimum leaf node count to 5 (matching the unique category count in the dependent variable) and systematically increasing maximum tree depth from 1 to 10. For each depth configuration, we performed 1,000 modeling iterations to calculate mean, maximum, and minimum accuracy values, as detailed in Table 17.

Analysis of the decision tree models revealed that maximum tree depth plateaued at 7, with no further depth increases observed beyond this threshold. Accuracy evaluation demonstrated two distinct patterns: shallow trees (depth = 1) showed limited classification capability across all dependent variables (accuracy = 0.42), while deeper configurations (depth ≥ 4) achieved peak performance (accuracy = 0.87). This progression confirms the ‘T-shirt sizes’ dataset’s effectiveness for teaching decision tree principles and implementing multiclass classification models.

The ‘Earthquake information’ dataset serves as an unsupervised learning resource featuring magnitude estimates for seismic intensity and geospatial coordinates (latitude/longitude) for cluster analysis. Using the Entry platform’s k-Means implementation with Scikit-Learn, we conducted cluster modeling experiments with varying group quantities (2–9 clusters). To objectively determine optimal clustering, we calculated inertia values—the sum of squared distances between cluster centers and their member points. We performed comparative analysis using both raw data and a preprocessed subset containing only seismically significant events (magnitude ≥ 2.0), with visualization results shown in Fig. 8.

Visualization of cluster and Inertia values through k-Means before and after preprocessing.

Visual analysis of inertia values revealed distinct clustering patterns and centroid positions between preprocessed and raw data when using 5–7 clusters. This demonstrates the dataset’s educational value for implementing AI-driven decision-making processes in classroom settings, as students can critically compare different clustering outcomes. The dataset’s effectiveness for cluster modeling education was thereby confirmed.

Maintanance AI education dataset

To enhance accessibility and educational utility of the developed datasets, we implemented distribution through the Entry programming platform following standardized procedures. As shown in Fig. 9, educators and students can access datasets through Entry’s practice interface using the workflow: [Table] → [Load Data Table] → [Add Table], ensuring consistency with other educational datasets available on the platform.

Publishing datasets via entry.

The dataset interface incorporates expert recommendations from the testing phase, particularly addressing dataset quality assessment and practical application requirements. As demonstrated in Fig. 10, each dataset includes: (1) Basic description, (2) Key variable explanations, (3) Column/row metadata, and (4) Usage examples—implementing expert guidance that “datasets should be easily understandable from a quality assessment perspective” and “must enable creation of complete outputs reflecting real-world activities”³⁶.

Dataset Presented via [Table] Menu of the Entry.

We established a maintenance framework featuring multiple feedback channels: an integrated bulletin board within Entry and a dedicated web portal with usage guides. This infrastructure allows users to submit improvement suggestions, which researchers can implement through collaborative review with the Connect Foundation (Entry’s governing organization). Approved modifications undergo immediate integration into the programming environment through automated deployment pipelines.