Job-market Oriented Course Recommendation System

Repository of experiments for a job market oriented course recommender system. Ongoing project, more work is required.

Installation

Requires python 3.10

Install required packages via pip

pip install -r requirements.txt

Usage

python src/pipeline.py --config config/run.yaml

Below you will find a detailed description of the parameters of the config file

taxonomy_path: data/taxonomy.csv # Path to the taxonomy file
course_path: data/courses.json # Path to the courses file
cv_path: data/resumes.json # Path to the resumes file
job_path: data/jobs.json # Path to the jobs file
mastery_levels_path: data/mastery_levels.json # Path to the mastery levels file
results_path: results # Path to the results directory where results are saved
level_3: true # Whether to use only the third level of the taxonomy and not the fourth  (if true: less skills)
nb_courses: 100 # Number of courses to use (set to -1 to use all)
nb_cvs: -1 # Number of resumes to use (set to -1 to use all)
max_cv_skills: 15 # Maximum number of skills per resume
nb_jobs: 100 # Number of jobs to use (set to -1 to use all)
threshold: 0.8 # Threshold for the similarities
k: 2 # Number of courses to recommend
model: greedy # Model to use (greedy, optimal, dqn, ppo, a2c)
total_steps: 50000 # Total number of steps for the training of the agent
eval_freq: 5000 # Frequency of the evaluation of the agent
nb_runs: 1 # Number of runs (set to 1 for greedy and optimal since they are deterministic)
seed: 42 # Seed for the random number generator

Renforcement learning framework for Job-Market Oriented Course Recommender

Skill-based Modeling

• User $u$: vector of skills where each dimension corresponds to a skill and the value is the user's proficiency level.
• Job $j$: vector of skills where each dimension corresponds to a skill and the value is the job's required proficiency level.
• Course $c = (c_r, c_p)$: two vectors of skills where each dimension corresponds to a skill and the value is the course's required (resp. provided) proficiency level.

MDP

• State user $u$
• Action space $\mathcal{C}$: the set of courses. One action being to recommend a course.
• Reward $r$: number of jobs the learner can apply to.
• Transition probabilities: binary, for now our MDP is deterministic. For a given action $c$ and state $u$, the outcome will always be the same: we update the state $u$ with the skills provided by the course: $c_p$.

Relavance and similarity functions

TBD

Description of src files

• pipeline.py: Trains and evaluate the agents, or heurisitc apporoaches.
• Greedy.py: Classe that implements the greedy recommendation strategy.
• Optimal.py: Classe that implements the optimal recommendation.
• Reinforce.py: Classe that implements the training and evaluation of the Reinfrocement-based recommendation using agents from stable_baselines3.
• CourseRecEnv.py: Classe that implements the evironment for training the agents using the gymnasium library.
• Dataset.py: Class that implments the dataset using resumes, courses and jobs.
• matchings.py: Contains various matching, similarity, and relevance functions.