Recommender System for Data Science Learning and Research

Published in IJAITL (2025), this paper presents a sequential recommendation system for data science learners and researchers. The system models how learner interests evolve over time and surfaces relevant papers, tutorials, and tools at each stage of a learning trajectory.

Problem

Data science education has a cold-start and progression problem. Recommending content to a learner who just started is very different from recommending to someone six months in. Static content libraries and keyword-based search don't capture this progression.

We model learner behavior as a sequential process — each interaction (paper read, tutorial completed, search query) updates the learner's state, which drives the next recommendation.

Approach

Sequential learner model: We use a transformer-based sequential recommendation architecture (similar to SASRec) that attends over a learner's history of interactions to produce a current learner state embedding.

Content representation: Papers, tutorials, and tools are encoded using a fine-tuned sentence transformer, capturing semantic content relationships that keyword matching misses.

Curriculum-aware objectives: We add a curriculum signal to the training objective — items that appear later in typical learning trajectories receive a bonus, nudging the recommender toward progressive complexity.

Evaluation

We evaluate on a dataset of learner interaction sequences from an online data science platform, measuring recall and NDCG at multiple list lengths. The sequential model outperforms non-sequential baselines, with the curriculum-aware variant showing additional improvement on longitudinal retention metrics.

The implementation is available at pagand/AITutor_SeqModeling. This work is part of the broader DaTu research program on AI-assisted data science education at SFU.