doslim / job-recommendation-pjfnn

Several baseline models and PJFNN on Job Recommendation Challenge

Job Recommendation PJFNN

This repository contains several baseline models and PJFNN (Person-Job Fit Neural Network from the paper: Person-Job Fit: Adapting the Right Talent for the Right Job with Joint Representation Learning) on the datasets from Job Recommendation Challenge.

Requirements and Structures

The following packages are required.

• torch-1.11.0
• nltk-3.6.5
• gensim-3.8.3

The structure of our project is as follows.

• code: contain all the code in the form of jupyter notebooks.
  • build_dataset.ipynb: a set of functions to build a small dataset from the original dataset.
  • baseline.ipynb: build and evaluate baseline models on the person-job fit task (binary classification).
  • ranking.ipynb: build and evaluate baseline models on the job recommendation task.
  • pjfnn.ipynb: build and evaluate modified PJFNN on the two tasks.
• report.pdf: a brief report (in Chinese) of the details of our implementations and implementation results.

Usage

To run our codes, you should first download the datasets from Kaggle. And rewrite the data path in our codes (mostly in build_dataset.ipynb).

Results

We conduct our experiments on a subset of the original dataset, where we set the WindowID=6. Since the data only have the application records (positive samples), we carry out randomly negative sampling to build a dataset such that the ratio of postive and negative samples are 1:1. More details can be found in the build_dataset.ipynb.

We consider the following two tasks:

• Person-Job Fit: the inputs are in the forms like (job, person), and the outputs are binary labels indicating that whether the people are suitable for the jobs.
• Job Recommendation: the models are required to rank 20 randomly selected jobs for the given person.

We use classic metrics for binary classification in the person-job fit task and hit rate in the recommendation task.

Person-Job Fit

	Accuracy	Precision	Recall	F1-score	AUC
Linear Regression	0.544	0.546	0.522	0.534	0.549
Logistic Regression	0.534	0.536	0.510	0.523	0.550
Naive Bayes	0.514	0.516	0.446	0.479	0.531
Decision Tree	0.605	0.609	0.590	0.599	0.631
Random Forest	0.637	0.634	0.647	0.640	0.702
AdaBoost	0.526	0.529	0.465	0.495	0.535
GBDT	0.629	0.630	0.624	0.628	0.663
XGBoost	0.607	0.606	0.615	0.610	0.614
PJFNN-m	0.660	0.654	0.679	0.667	0.715

Hit Rate@N

	N=1	N=5	N=10	N=20
Linear Regression	0.008	0.100	0.158	0.288
Logistic Regression	0.008	0.092	0.142	0.300
Naive Bayes	0.015	0.085	0.208	0.369
Decision Tree	0.031	0.081	0.162	0.304
Random Forest	0.027	0.123	0.227	0.431
AdaBoost	0.019	0.123	0.212	0.327
GBDT	0.038	0.150	0.235	0.404
XGBoost	0.027	0.104	0.196	0.377
PJFNN-m	0.042	0.150	0.262	0.446