Why My AI Project Combines Rules and Machine Learning

In my career pathway adviser, machine learning contributes 10% of the final score. The remaining 90% comes from subject-and-grade scoring and an interest questionnaire.

That balance reflects a central decision in the project: how much influence should each source of information have, given the evidence available?

The system was developed for Mauritian HSC school-leavers. It brings together information about a student's academic background, interests and self-perception, then suggests three broad career fields to explore.

I wanted the scoring to be understandable while leaving room to investigate what machine learning might contribute.

The rule-based part provides a structure I can inspect. Subjects and grades feed into predefined scoring rules. Responses to a RIASEC-style interest questionnaire contribute another score.

This makes it possible to examine how those inputs influence the recommendations. If a field scores unexpectedly highly, I can review the relevant rules and mappings.

However, making a rule visible does not establish that it is the best rule. Decisions about how subjects relate to career fields still involve assumptions. Those assumptions need to be examined and evaluated.

The experimental machine learning component explores a different source of information: self-reported personality and aptitude responses. It learns relationships from examples instead of requiring every relationship to be written as a rule.

For this component, I compared Logistic Regression, Support Vector Machine, Gaussian Naive Bayes and Random Forest.

The main constraint was the data. The experimental classifier used a public demonstration dataset containing 105 records. That provided an opportunity to experiment, but it offered limited evidence for a system intended for Mauritian school-leavers.

Self-reported aptitude also needs careful interpretation. It describes how someone sees their abilities; it is not an independently measured assessment of those abilities.

These limitations helped shape the weighting: 60% for subjects and grades, 30% for interests and 10% for the experimental classifier.

Academic information received the largest share, interests retained a substantial role, and the experimental component had a limited contribution. These weights remain design choices that need further validation.

A simplified example shows how the combination works. Suppose one career field receives the following component scores on a common scale from 0 to 100:

These are invented values for explanation. The combined score helps rank fields within the framework. It has no established interpretation as a probability of career success.

A small weight can still affect the outcome. When two fields have similar scores, a modest contribution may change which one appears first.

That is why I also examined the experimental component's influence through an audit using 5,000 generated valid inputs. Adding it changed the leading field in 8.9% of those cases.

This result describes behaviour under the generated inputs. It does not establish how often the model improves guidance for real students.

Using a hybrid design also creates responsibilities. The rules need scrutiny, the model needs evaluation, and the combined system needs to be checked as a whole. A low model weight does not automatically make the recommendations fair or reliable.

For a stronger version, I would want a larger, locally relevant dataset and feedback from students and career guidance professionals. I would also compare the full system with the transparent baseline to investigate whether the experimental component adds useful information.

The project taught me to think about the role a model should play before giving it more influence. In this prototype, combining rules and machine learning let me explore that question while keeping most of the scoring open to inspection.