Empowering Non-Coders: The Future of Data Science Education with AI

I'd like to share insights from a recent study on the intersection of AI and data science education, titled "Generative AI for Data Science 101: Coding Without Learning To Code" by Jacob Bien and Gourab Mukherjee. This paper showcases an innovative approach to teaching data science to non-technical students, leveraging the power of generative AI tools like GitHub Copilot to bridge the gap between complex data analysis tasks and those without coding expertise.

The implications of this study are profound, hinting at a future where access to data science and technical education is democratised, making it more inclusive and accessible to a broader audience. It challenges traditional educational paradigms, encouraging a shift towards a focus on conceptual understanding and problem-solving skills over rote learning of syntax.

As we navigate this new era of education, it's crucial for educators, students, and industry leaders to explore and embrace these innovative tools. This approach not only opens up new pathways for learning and teaching but also prepares a more diverse and creative cohort of future data scientists.

I invite you to delve into the details of this fascinating study and join me in pondering the transformative potential of generative AI in education. Think how we can leverage these advancements to foster an inclusive, engaging, and practical learning environment for all.

Read the full paper here and let's explore the future of data science education together.

Introduction

Data science educators face a significant dilemma: should coding be a mandatory part of the curriculum for non-technical students? This question is especially pertinent in introductory statistics and data science classes, where the primary goal is to impart a foundational understanding of statistical principles. Traditionally, the inclusion of coding in these courses has been contentious. On one hand, the ability to code is seen as essential for engaging directly with data, allowing students to apply theoretical concepts to real-world datasets. On the other hand, the steep learning curve associated with programming languages can be daunting for beginners, potentially detracting from the core statistical lessons.

Enter the innovative study "Generative AI for Data Science 101: Coding Without Learning To Code" by Jacob Bien and Gourab Mukherjee. This paper presents an approach that seeks to resolve this educational conundrum by leveraging generative Artificial Intelligence (AI) tools. Specifically, it explores the use of GitHub Copilot, an AI-powered code completion tool, in a required introductory data science course for full-time MBA students at the Marshall School of Business, University of Southern California. The course, designed for students without technical backgrounds, aimed to introduce them to data science and statistics within a business context.

The central thesis of Bien and Mukherjee's paper is both simple and revolutionary: it is possible to empower students to perform complex data science tasks through English-language prompts that are converted into executable R code by AI, thus bypassing the need for traditional coding instruction. This method not only demystifies data science for non-coders but also opens up new pedagogical possibilities for integrating AI tools into education. The study represents a case in point for a broader discussion about the future of data science education, highlighting the potential of generative AI to bridge the gap between technical and non-technical learners.

As we delve deeper into the implications of this study, we must consider the broader context in which this experiment was conducted. The rise of large language models and their application in generating code presents an unprecedented opportunity for educational innovation. By examining Bien and Mukherjee's approach, we can gain insights into how generative AI might be utilised to make data science more accessible, engaging, and applicable to a wider audience, ultimately reshaping the landscape of data science education for the better.

The Experiment: A New Approach to Teaching Data Science

In late 2023, a pioneering experiment was conducted at the Marshall School of Business, University of Southern California, aimed at redefining the approach to teaching data science to non-technical students. The course, "Data Science for Business," was part of the full-time MBA program and designed to introduce students to the fundamentals of data science and statistics, particularly within the context of business applications. The traditional challenge of such courses has been finding the right balance between teaching the statistical concepts necessary for data science and the technical coding skills required to apply these concepts to real-world data.

The innovative solution proposed and tested by Jacob Bien and Gourab Mukherjee involved the use of GitHub Copilot, a generative AI tool developed by OpenAI and GitHub. GitHub Copilot functions as a sophisticated code completion tool, capable of translating English-language prompts into executable R code. This approach allowed students who had little to no programming experience to engage directly with data science projects without the prerequisite of learning a programming language's syntax.

The primary goal of this experiment was to democratise access to data science by removing one of the most significant barriers to entry: the need to code. By leveraging GitHub Copilot, students were able to formulate their analytical questions in plain English, which the AI then translated into R code. This not only facilitated a direct interaction with data but also enabled students to focus on the conceptual understanding of data science methods and their applications in business, without getting bogged down by the complexities of coding syntax.

This novel teaching approach represents a significant departure from conventional data science education, which often requires students to spend considerable time and effort learning a programming language before they can start analysing data. Instead, Bien and Mukherjee's method places students in the driver's seat from the outset, allowing them to experiment with data, formulate hypotheses, and see the results of their inquiries with minimal delay. It embodies a shift towards a more inclusive and accessible data science education, potentially setting a new standard for how such courses are taught in the future.

Key Findings from the Paper

The experiment conducted by Jacob Bien and Gourab Mukherjee on integrating GitHub Copilot into data science education for non-technical students yielded several compelling findings, fundamentally challenging traditional pedagogical approaches in this domain. The core results of their study underscore the transformative potential of using generative AI tools in educational settings, particularly for subjects that traditionally require a strong technical foundation.

Main Findings

1. Successful Engagement with Data Science Tasks: One of the most significant outcomes highlighted in the paper was the ability of students to effectively engage with complex data science tasks through the use of English-language prompts. By interacting with GitHub Copilot, students could translate their analytical questions into executable R code without needing to write the code themselves. This method allowed for direct, hands-on involvement with data analysis, fostering a deeper conceptual understanding and appreciation for the field of data science among students who might otherwise have been sidelined due to a lack of coding skills.
2. Reduced Learning Curve and Intimidation Factor: Traditionally, the steep learning curve associated with mastering a programming language has been a significant barrier to entry for many students interested in data science. The use of GitHub Copilot in the classroom effectively mitigated this challenge. By abstracting away the syntax and allowing students to focus on the logic and objectives of their analyses, the approach significantly reduced the intimidation factor and made the learning process more accessible and engaging. Students could concentrate on the "what" and "why" of data science, rather than getting bogged down by the "how" of programming syntax.

Examples from the Paper

Several examples from the paper vividly illustrate the practical applications and benefits of this approach:

• Loading Data: Students were tasked with loading a dataset into their working environment—a foundational step in any data analysis project. Instead of writing the code, students simply provided an English prompt such as "Load the data in housing-prices.csv," and GitHub Copilot generated the appropriate R code to execute this task.
• Performing Logistic Regression: In a more complex example, students were able to perform logistic regression analysis to predict outcomes based on certain variables in the dataset. A student might use a prompt like "Fit a logistic regression model to predict ‘one_month_on_market‘ based on ‘year_built‘ and ‘neighborhood‘." GitHub Copilot then provided the R code necessary to create the model, demonstrating how students could engage in sophisticated analytical techniques without prior coding knowledge.

These examples highlight the paper's key finding that generative AI tools like GitHub Copilot can effectively democratise data science education, making it accessible and engaging for a broader audience. This approach not only facilitates a more inclusive learning environment but also encourages a more profound engagement with the data science process, allowing students to focus on the analytical thinking and decision-making skills that are crucial in the field.

Examples in Python

I don’t have R installed so I thought I’d try it out using Python, using simple examples.

Load Data

Summarise Data

Clean the data

List the kinds of analysis i could perform on this data

The?describe()?function in pandas provides a summary of the central tendency, dispersion, and shape of a dataset's distribution. Based on this, here are some analyses you could perform on your?housing_data:

1. Descriptive Statistics: Analyze the count, mean, standard deviation, minimum, and maximum of each column.
2. Correlation Analysis: Determine the relationship between different variables in your dataset. You can use the?corr()?function in pandas for this.
3. Histogram Analysis: Plot histograms of various columns to understand the distribution of values in those columns.
4. Box Plot Analysis: Create box plots to understand the statistical summary of various columns, which includes minimum, first quartile, median, third quartile, and maximum.
5. Scatter Plot Analysis: Create scatter plots between different pairs of columns to understand the relationships between different variables.
6. Regression Analysis: If your dataset includes a target variable (like house price), perform regression analysis to understand how other variables affect the target variable.
7. Outlier Detection: Detect outliers in your dataset that could potentially skew your analysis.

Remember, the type of analysis you perform depends on the questions you're trying to answer with your data.

Let's do a regression analysis to understand how other variables affect the price

import statsmodels.api as sm

# Assuming 'price' is your dependent variable and all other columns are independent variables
X = housing_data.drop('price', axis=1)
y = housing_data['price']

# Add a constant to the independent variables
X = sm.add_constant(X)
X = X.select_dtypes(include='number')

# Create an OLS model
model = sm.OLS(y, X)

# Fit the model
results = model.fit()

# Print the summary
print(results.summary())
        

                            OLS Regression Results                            
==============================================================================
Dep. Variable:                  price   R-squared:                       0.562
Model:                            OLS   Adj. R-squared:                  0.558
Method:                 Least Squares   F-statistic:                     138.1
Date:                Thu, 07 Mar 2024   Prob (F-statistic):           4.37e-94
Time:                        12:27:54   Log-Likelihood:                -6103.4
No. Observations:                 545   AIC:                         1.222e+04
Df Residuals:                     539   BIC:                         1.224e+04
Df Model:                           5                                         
Covariance Type:            nonrobust                                         
==============================================================================
                 coef    std err          t      P>|t|      [0.025      0.975]
------------------------------------------------------------------------------
const      -2081.9213   3523.350     -0.591      0.555   -9003.101    4839.259
lotsize        4.7302      0.380     12.448      0.000       3.984       5.477
bedrooms    2397.2827   1184.753      2.023      0.044      69.984    4724.581
bathrms      1.62e+04   1697.548      9.541      0.000    1.29e+04    1.95e+04
stories     7827.7116    984.207      7.953      0.000    5894.361    9761.063
garagepl    5394.2327    954.345      5.652      0.000    3519.542    7268.924
==============================================================================
Omnibus:                       69.095   Durbin-Watson:                   1.533
Prob(Omnibus):                  0.000   Jarque-Bera (JB):              149.339
Skew:                           0.704   Prob(JB):                     3.73e-33
Kurtosis:                       5.143   Cond. No.                     2.66e+04
...
Notes:
[1] Standard Errors assume that the covariance matrix of the errors is correctly specified.
[2] The condition number is large, 2.66e+04. This might indicate that there are
strong multicollinearity or other numerical problems.
        

Benefits and Challenges Identified

The innovative teaching approach explored by Jacob Bien and Gourab Mukherjee in their experiment with GitHub Copilot in a data science course for non-technical students revealed both significant benefits and notable challenges. These insights contribute to the ongoing discourse on the integration of AI tools in education, especially in fields that traditionally require a high level of technical proficiency.

Benefits Identified

1. Democratising Access to Data Science: A primary benefit of using GitHub Copilot in the classroom is the democratisation of data science education. By lowering the barrier to entry associated with learning coding syntax, this approach makes data science accessible to a wider range of students, irrespective of their technical background. This inclusivity could lead to a more diverse pool of individuals entering the field, bringing varied perspectives and insights.
2. Enhancing Problem-Solving Skills: The experiment demonstrated that students could focus more on solving data science problems rather than grappling with the syntax of coding. This focus on problem-solving enhances critical thinking and analytical skills, as students learn to articulate their questions clearly and think logically about the steps needed to arrive at a solution.
3. Fostering a Deeper Interest in the Subject: By removing the intimidation factor associated with coding, students were more likely to engage deeply with the subject matter. The ability to directly interact with data and see the immediate application of data science concepts in solving real-world problems can spark curiosity and foster a lasting interest in the field.

Challenges and Limitations Noted

Despite these significant benefits, the experiment with GitHub Copilot also revealed some challenges and limitations:

1. Randomness of Copilot's Outputs: One of the challenges encountered was the variability and randomness in the responses generated by GitHub Copilot. Because the AI's suggestions are based on a vast corpus of code and not a fixed set of rules, different prompts can lead to different outputs, and the same prompt can produce different results at different times. This unpredictability required instructors and students to adapt and sometimes refine their prompts to achieve the desired outcome.
2. Importance of Writing Specific Prompts: The effectiveness of GitHub Copilot as a teaching tool heavily relies on the ability of users to write clear, specific prompts. This necessitates a certain level of skill in articulating the data science task at hand. Developing this skill is essential, as vague or overly broad prompts can lead to incorrect or unhelpful code suggestions. This challenge underscores the need for teaching students how to effectively communicate with AI tools, a skill that is becoming increasingly important as these technologies become more prevalent in various industries.

While the integration of generative AI tools like GitHub Copilot into data science education offers promising benefits such as increased accessibility and a focus on problem-solving, it also presents new challenges. Educators considering this approach must be prepared to guide students in navigating the unpredictability of AI-generated code and in developing the skills necessary to use these tools effectively. Despite these challenges, the potential of such AI tools to transform educational practices and make technical subjects more accessible to a broader audience is undeniable.

Implications for Data Science Education and Industry

The findings of the paper have profound implications for the future of data science education and the industry at large. The use of AI tools like GitHub Copilot to bridge the gap between non-technical students and data science tasks represents a paradigm shift in how we approach teaching technical subjects. This evolution carries significant potential to alter educational practices, skill requirements in the industry, and the inclusivity of technical fields.

Changing Educational Practices

The successful integration of AI tools in teaching data science underscores the potential for a broader application of similar technologies across various technical subjects. By enabling students to focus on conceptual understanding and problem-solving without the initial barrier of learning complex coding syntax, educators can make these subjects more accessible and engaging. This approach could lead to more adaptive and personalised learning experiences, where students use AI as a tool to complement their learning process, allowing for a more hands-on and exploratory approach to education.

Impact on Skills Required in the Industry

As AI tools become more integrated into the educational process, the skills required in the data science industry may also evolve. The ability to effectively communicate with AI to translate business problems into data science tasks could become as crucial as traditional coding skills. This shift does not diminish the value of understanding programming languages but rather adds a new layer of competency in leveraging AI tools for data science tasks. Such a development could make the field more inclusive, opening up opportunities for individuals with diverse backgrounds and strengths, particularly those who excel in analytical thinking and problem-solving but may not have formal training in programming.

Balancing Code Understanding and AI Tool Leveraging

The paper also touches on an essential debate about the balance between understanding code and leveraging AI tools. While AI tools like GitHub Copilot can significantly reduce the entry barriers to data science, there remains a fundamental value in understanding the underlying principles of coding and data science methodologies. This comprehension ensures that practitioners can critically evaluate the AI-generated code, understand the limitations of AI tools, and make informed decisions based on the outputs. Therefore, the future of data science education may lie in a hybrid model that combines the foundational knowledge of coding with the strategic use of AI tools, preparing students to navigate a landscape where both skills are indispensable.

The implications of incorporating AI tools into data science education extend beyond the classroom, potentially influencing the entire data science industry. By making data science more accessible and fostering a diverse pool of talent, we can drive innovation and creativity in the field. However, the key to unlocking this potential lies in finding the right balance between traditional coding skills and the use of generative AI tools, ensuring that the next generation of data scientists is equipped to tackle the challenges of the future with a comprehensive toolkit.

Looking Ahead

The innovative approach presented in "Generative AI for Data Science 101: Coding Without Learning To Code" by Jacob Bien and Gourab Mukherjee marks a significant milestone in the journey toward making data science education more accessible and inclusive. By harnessing the capabilities of generative AI tools like GitHub Copilot, the authors have demonstrated a powerful method to empower non-coders, allowing them to engage meaningfully with data science tasks. This approach not only democratises access to data science but also highlights the potential for AI to transform educational methodologies across various technical disciplines.

As we stand on the brink of a new era in education, the role of generative AI in shaping learning paradigms cannot be overstated. The success of this experiment prompts us to reimagine the boundaries of traditional education, where the emphasis shifts from rote learning of syntax to fostering a deeper understanding of concepts and enhancing problem-solving skills. This shift has the potential to cultivate a more diverse cohort of data scientists, equipped not just with technical know-how but with the creativity and critical thinking skills necessary to drive innovation.

This moment serves as a call to action for educators, students, and industry leaders alike. The future of education is being rewritten, and it is incumbent upon us to explore and embrace these innovative tools. By integrating AI into our learning and teaching methodologies, we can unlock new possibilities for students of all backgrounds, making the field of data science richer and more varied. Let us seize this opportunity to make education more engaging, practical, and inclusive, ensuring that everyone has the chance to contribute to and benefit from the data-driven decisions shaping our world.

References

Jacob Bien and Gourab Mukherjee.

• Generative AI for Data Science 101: Coding Without Learning To Code - [2024-02-01]