Data

Data Overview

1. NUSMods

Course data was retrieved from the NUSMods API and compiled into modules.csv (7,015 rows, 14 columns).

Core Fields	Description
moduleCode	Unique identifier
title	Course title
description	Course description
faculty	Offering faculty
prerequisite	Required modules
moduleCredit	Credit units

---

2. MyCareersFuture Job Ads

The dataset consists of 22,720 job postings, flattened from JSON into structured format.

Core Fields	Description
title	Job title
skills	Required skills
categories	Job categories
minimum_years_experience	Experience required
salary_min / salary_max	Salary range
posted_date / expiry_date	Posting dates
position_levels	Seniority level

---

Exploratory Data Analysis (EDA)

EDA identifies structural properties that inform modelling choices and potential sources of bias.

---

NUSMods

Distribution by Faculty

Figure 1: Module representation is uneven, with FASS, CDE, and Science dominating the corpus.

Implications for Framework:
- Risk of representation bias in similarity matching
Mitigation:
- Construct degree-specific module baskets (≈15 core + 8 common modules) - Use length-normalised embeddings for fair comparison

---

Description Length

Figure 2: Most descriptions fall within 60–100 words, with few long outliers (>250 words).

Implications for Framework:
- Descriptions provide sufficient semantic signal for embeddings.
- Text length is bounded during profile construction to control computational cost.

---

MyCareersFuture Job Ads

Market Breadth and Skills

Figure 3: Soft skills (e.g., teamwork, communication) dominate job postings.

Implications for Framework:
- These skills are non-discriminative and introduce noise.
- They are removed during preprocessing.

---

Category Co-occurrence

Figure 4: Job categories frequently co-occur, reflecting overlapping roles.

Implications for Framework:
- Categories are retained as structured features to enrich representations.

---

Seniority and Experience

Figure 5: Entry-level roles dominate, though senior roles exist.

Implications for Framework:
- Aligns with graduate outcomes - Senior roles are excluded to maintain relevance

---

Data Cleaning and Preprocessing

NUSMods

1. Standardise text: Remove HTML and normalise whitespace

2. Handle missing descriptions: Use title as fallback if informative and exclude generic modules (e.g., internship, UROPS)

3. Construct module text: Combine cleaned title and description

4. Filter by relevance: Derive module level from code and retain undergraduate modules

5. Remove low-quality entries: Drop modules without sufficient text

---

MyCareersFuture Job Ads

1. Parse and structure data: Convert JSON into structured format and preserve multi-label fields

2. Clean text fields: Remove HTML, URLs, and boilerplate, then filter low-information descriptions

3. Clean skills: Standardise text (lowercase), remove generic soft skills and retain meaningful technical terms

4. Filter by scope: Exclude internships, academic, and senior roles

5. Deduplicate postings: Remove exact and near-duplicates

6. Construct job text: Combine title, categories, skills and truncated description