Personalized Mentor Recommendation System - Approach Documentation

Requirements

1. Python 3.8 or above
2. pip (Python package manager)
3. pandas
4. scikit-learn

---

Approach Used: Content-Based Filtering (CBF)

1. Data Loading & Profile Construction

• The system reads mentor data from a CSV file containing fields such as:

  • subjects
  • college
  • languages_spoken
  • mentorship_style
  • available_times
  • mode_of_mentorship
  • years_of_experience

• A custom profile string is generated for each mentor by concatenating key fields.

• Weighting is applied to prioritize more relevant attributes:

  • subjects × 3
  • college × 2
  • languages_spoken × 2
  • Remaining fields × 1

2. User Input / Query

• The user (aspirant) provides preferences via command-line input for:

  • preferred subjects
  • target college
  • preparation level
  • preferred learning style
  • languages spoken

• A profile string is generated from these responses.

• Note: No custom weights are currently applied to aspirant input fields.

3. TF-IDF Transformation

• All mentor profiles and the aspirant profile are converted into TF-IDF vectors.
• This creates a document-term matrix where each term's importance is determined based on frequency and uniqueness.

4. Similarity Calculation

• Cosine similarity is computed between the aspirant profile vector and each mentor vector.
• This measures how closely the aspirant's interests align with mentor profiles.

5. Recommendation

• Mentors are ranked based on similarity scores.
• The top 3 mentors with the highest cosine similarity values are selected and recommended to the user.

---

Limitations & Future Scope

1. No collaborative filtering — system does not learn from other users' behavior.
2. No semantic understanding — relies only on keyword matching; does not use BERT or sentence embeddings.
3. No handling of:
   • Synonyms (e.g., "legal GK" vs "legal general knowledge")
   • Spelling mistakes or typos
4. No personalization loop — lacks feedback-based learning or user rating system.
5. Fixed weights — all mentors are evaluated with static profile generation logic.