TOKYO - A rapidly growing data science program at the University of Tokyo is attracting an unusually wide range of participants, with junior high and high school students studying alongside university students and working adults.
The course, known as GCI, is offered nationwide online and free of charge for students, eliminating barriers even for complete beginners and fueling a surge of interest from across Japan and abroad.
At a recent completion ceremony, organizers reported 10,579 total enrollees and 1,490 graduates, highlighting the program’s rigorous nature with a 14% completion rate. “I thought I might fail the final assignment, but I managed to finish,” said a second-year junior high school student with little programming experience. GCI is held twice a year, with the next session starting in mid-October, and its popularity has gone global, drawing 7,700 applicants from 32 countries and 430 universities for the English-language version.
To explore why the course is so compelling, GCI instructor and AI startup researcher Masayuki Sera walked through its approach, from fundamentals to practical applications. Sera works at Twins, a company spun out of the university’s AI lab, and applies data science to real business problems. “The work is wide-ranging,” he said. “For a telecom company, for example, we might predict whether customers are likely to cancel their contracts and then suggest changes to their plans. We also assess whether current strategies are effective and adjust them if necessary.”
The program’s curriculum follows a structured process: explore and clean data, build models, evaluate results, and iterate. A signature assignment involves the “Home Credit Default Risk” challenge, where students predict whether customers will default on loans based on tabular data such as income, family size, and loan type. The training dataset includes about 170,000 rows and 51 columns, while the test set has around 60,000 rows and 50 columns, with the default labels hidden.
Exploratory data analysis (EDA) is emphasized early on, teaching students to identify missing values, outliers, and skewed distributions. In one example, missing entries in household size and product price had to be filled before modeling. Students also learn how class imbalance—92% repay their loans while 8% default—can distort results and why metrics like AUC are better than raw accuracy. Visualization reveals useful patterns: income distributions become more interpretable after log transformations, and certain features, like education level and loan type, strongly correlate with default rates.
Before modeling, text categories must be encoded as numbers and missing values filled. Although one-hot encoding is generally safer, GCI demonstrates label encoding for simplicity with tree-based models. A basic random forest model trained on a 70/30 split achieves an AUC of around 0.65—“not exceptional but proof the features contain predictive power,” Sera noted.
Students then learn how to improve performance through feature engineering, such as creating new variables like the ratio of loan amount to income (repayment burden) or product price to loan amount (self-financing ratio). These changes can nudge AUC scores upward—sometimes by just 0.5 percentage points, a difference that can significantly impact leaderboard rankings. Other techniques include comparing individual loan amounts to group averages, trying different encoding or imputation strategies, tuning hyperparameters, or even switching algorithms. This iterative cycle—hypothesizing, testing, and refining—is where many learners find themselves “hooked.”
What keeps even teenagers engaged, instructors say, is the immediate feedback and sense of discovery. With only a few lines of Python, beginners can build a competitive model, and a single visualization can reshape their understanding of the data. “You don’t need to master every algorithm to start,” said Sera. “What matters is rigorous analysis, thoughtful feature design, and relentless iteration.”
GCI’s success reflects a broader trend: data science has become the gateway to artificial intelligence. By grounding learners in core skills—predictive modeling, fair evaluation, and careful data preparation—the course demystifies AI and builds practical foundations. For companies, the message is similar: rather than chasing buzzwords, start by examining existing data, asking the right questions, and letting evidence guide strategy.
Source: テレ東BIZ
