cymcginnis/Student_Persistence_Dashboard

title: Student Success Risk Dashboard emoji: 🎓 colorFrom: blue colorTo: teal sdk: gradio sdkversion: 4.31.0 appfile: app.py pinned: false ---

Student Success Risk Dashboard

A demo Gradio app for university staff, built on the Open University Learning Analytics Dataset (OULAD). It flags students whose early-course engagement and assessment patterns resemble past students who ended up Withdrawn or Failed, and explains why each individual student was flagged using SHAP.

Setup

1. 1.Download the OULAD CSV files and place them in a data/ folder next to app.py:
2. 2.studentInfo.csv
3. 3.studentRegistration.csv
4. 4.studentVle.csv
5. 5.vle.csv
6. 6.assessments.csv
7. 7.studentAssessment.csv
8. 8.Install dependencies: pip install -r requirements.txt
9. 9.Run locally: python app.py, or deploy as-is to a Hugging Face Space.

What it does

• —Cohort Overview — filter by module/presentation, see the top 25 highest-risk students, a historical-outcome bar chart, a risk-level donut chart, and an engagement-vs-risk scatter plot, with a plain-language summary for staff.
• —Student Lookup — look up one student by ID and see a risk gauge plus a SHAP-based breakdown of exactly which factors (low engagement, missed early assessments, late registration, etc.) are pushing their risk score up or down.
• —Ask About the Data — a chat tab (powered by OpenAI) where staff can ask plain-English questions about the dataset, the model, or the currently selected cohort. It's grounded only in cohort-level summary statistics — never individual student rows — and will redirect questions about a named student to the Student Lookup tab instead of guessing.

To enable the chat tab, add an OPENAI_API_KEY secret to your Space (Settings → Repository secrets). Without it, the tab still loads but explains what's missing instead of erroring. Note that each question and a summary of the current cohort are sent to OpenAI's API to generate a response — worth knowing if your data-handling policies require disclosure. The model used defaults to gpt-4o-mini; override it by setting an OPENAI_MODEL secret/variable if you'd rather use a different OpenAI model.

Design

The dashboard opens with a plain-language explainer contrasting a simple statistical rule ("students who miss the first quiz fail 40% of the time") with what this tool actually does (a Random Forest weighing dozens of signals in combination) — aimed at non-technical academic staff, not data scientists. Typography uses Fraunces (headings) and Inter (body) for readability at a larger base font size than Gradio's default.

Model

A RandomForestClassifier trained on registration, VLE (virtual learning environment) clickstream, and early-assessment features from only the first 30 days of each course presentation (configurable via EARLY_WINDOW_DAYS). The model is trained once at startup and cached for the life of the process — it does not retrain on every UI interaction.

Notes / limitations

• —Predictions describe statistical resemblance to historical patterns, not certainty about any individual student. This is intended to support early advising and outreach, not automatic or punitive decisions.
• —Module codes (AAA, BBB, ...) are OULAD placeholders, not real course names.
• —SHAP explanations are computed per-lookup via TreeExplainer, which is fast for a single row but the whole pipeline is retrained only at startup. Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference