Machine Learned Resume-Job Matching

Solution

Yiou Lin, Hang Lei, Prince Clement Addo, and Xiaoyu Li

University of Electronic Science and Technology of

China, Chengdu, 610054, China
lyoshiwo@gmail.com hlei@uestc.edu.cn pricetheboy@gmail.com
arXiv:1607.07657v1 [cs.CL] 26 Jul 2016

xiaoyuuestcg@uestc.edu.cn

Abstract. Job search through online matching engines nowadays are

very prominent and beneficial to both job seekers and employers. But
the solutions of traditional engines without understanding the semantic
meanings of different resumes have not kept pace with the incredible
changes in machine learning techniques and computing capability. These
solutions are usually driven by manual rules and predefined weights of
keywords which lead to an inefficient and frustrating search experience.
To this end, we present a machine learned solution with rich features and
deep learning methods. Our solution includes three configurable modules
that can be plugged with little restrictions. Namely, unsupervised fea-
ture extraction, base classifiers training and ensemble method learning.
In our solution, rather than using manual rules, machine learned methods
to automatically detect the semantic similarity of positions are proposed.
Then four competitive “shallow” estimators and “deep” estimators are
selected. Finally, ensemble methods to bag these estimators and aggre-
gate their individual predictions to form a final prediction are verified.
Experimental results of over 47 thousand resumes show that our solu-
tion can significantly improve the predication precision current position,
salary, educational background and company scale.

Keywords: job matching, online resume, semantic similarity, deep learn-

ing, ensemble method

1 Introduction

Following the effects of the global financial crisis in 2008, large financial insti-
tutions have collapsed. Knowledge workers, even in wealthiest countries have to
worry about losing their well-paid, full-time jobs, and cannot easily find similar
ones elsewhere. An effective e-recruiting engine can help job seekers easily access
recruitment opportunities and reduces the recruitment labor by providing suit-
able items which match their personal interests and qualifications. It also frees
companies from information overload and advertisement cost. The key module
for a dynamic e-recruiting engine is job matching system which makes an effort
to engage the unemployed who are well suited to the vacancies to be filled.
II

In this work, we evaluate the job matching problem as a classification prob-

lem. This is to identify a job seeker’s current employment detail (the last position
in the resume) by their previous employment history. The framework of our so-
lution is constructed by several modules based on keras[3] and sklearn[9], and
can be practically deployed and easily verified. Through empirical evaluation,
we show step by step how to intensify the solution and get a better performance
than a baseline manual rule-based job matching system.
The rest of the paper is organized as follows. In Section 2 we survey the
related literature to give an overview of the research background. In Section
3, we introduce the dataset description. Then we propose the feature extract
methods and the machine learned models in Sections 4 and 5 respectively. In
Section 6, we report two ensemble methods and analyze the empirical results.
Finally, the paper is concluded in Section 7 with our future work.

2 Literature Review

Job matching system is a kind of recommender system. Recommender system

was first introduced by Resnick and Varian [10] who pointed out that in a typi-
cal recommender system people provide recommendations as inputs, which the
system then aggregates and directs to appropriate recipients. After that recom-
mender systems are being highly accepted in various industries and academic
areas and are gaining momentum over the years. In general, recommender sys-
tems are applied in various domains (such as books, digital products, movies,
music, TV programs, and web sites) and help users to find content, products,
or services by aggregating and analyzing suggestions and behaviors from other
users[2][6]. In a detailed survey paper[6] provided researchers with the state-
of-the-art knowledge on recommender system including real-world applications,
recommendation methods, real-world application domains and application plat-
forms.
For job matching area, many researches have been conducted to discuss dif-
ferent recommender system related to the recruiting problem as well[1]. Among
them, Malinowski et al.[7] discussed a bilateral matching recommendation sys-
tems to bring people together with jobs using an Expectation Maximization
(EM) algorithm, while Golec and Kahya[4] delineated a fuzzy model for competency-
based employee evaluation and selection with fuzzy rules. Paparrizos et al.[8]
used Decision Table/Naive Bayes (DTNB) as hybrid classifier. Though these
system used many manual attributes and various information retrieval tech-
niques, compared to our work which employed deep learning methods to ac-
celerate the process of finding the most appropriate jobs, they still failed in
keeping with the rapid changes in computing capability and machine intelli-
gence. Similar to the work of Zhang et al.[11], our work also tries to optimize
the knowledge worker-position matching, considering various characteristics of
knowledge workers. Compared to the work of Guo et al.[5], Our solution is com-
pletely data-driven, without using exterior semantic tool (NLTK and DBpedia)
 III

like they did. Besides, our resume data are widely collected from various areas
which makes our solution more universal and robust.

3 Dataset description

The dataset used was tapped from a job recommend game1 and can be freely
downloaded2 . The original dataset contains 70,000 resumes with 34,090 different
positions. After cleaning and filtering, 47,346 resumes whose last jobs belong to
a particular predication list of most frequent 32 positions (e.g. software engineer,
cashier and project manager) were used. Even though, there are 18,736 different
positions in the dataset. The most frequent positions are shown in Figure 1.

Fig. 1: Most Frequent Positions which Appear in the Dataset

Table 1 is an example of resume with job seeker’s personal information and

work history. In particular, salary ∈ [0, 6], degree∈ [0, 2] while size means the
company’s scale. Our problem description can be shown as identifying a job
seeker’s current position (the last position in the resume) by their previous em-
ployment history.

4 Information Extraction from Resumes

There are three main kinds of features made up of 95 foundational manual

features, 72 cluster features and 380 semantic features. In all, there are 551
features for each resume.
1
http://www.pkbigdata.com/common/cmpt list/all all time 1.html
2
http://www.pkbigdata.com/common/cmptData/147.html
IV

Table 1: A Job Seeker’s Resume

The Json Stucture One of the workExperienceList
id 558d...761 size 3
major 通信工程 salary 4
degree 1 end date 今
id Object{...} start date 2014-8
gender 男 industry 计算机/互联网
age 31 position name 软件测试
wokExperimentList Array[3] department 硬件测试

4.1 Manual Features

In the beginning, a lot of manual features were designed. Some features are char-
acter type, some boolean values, others are numerical. For the example shown in
Table 1, the features include but not limited to gender, age, major, the details
and changes of previous jobs, the age when first employed, the highest salary
and so on. The values of character features were inputted into a dictionary, and
numerical keys were used instead of the values when training.

4.2 Cluster Features

The design features can be transformed from characters to numerical keys, but
similar phrases (such as software engineer and Android engineer) would never be
identified. Thus, we train a Chinese Word2Vec model using all the resumes’ text.
And the average word embeddings of a phrases now can represent its semantic
meaning. A simple K-mean method was used to classify phrases into 64 and 128
clusters. The last 5 job experiences in a resume were kept, each job experience
consists of 7 phrases, including department, industry, position name, salary, size,
type (is empty usually) and quarter (if a seeker works 4 quarters for a company
will be marked as “quarter 4”). Some clusters are shown in Table 2.

Table 2: Phrases in Different Clusters

cluster0 cluster1 cluster2 cluster3
商务部门经理 钳工/钣工 营运/人力资源 供应链管理
工程经理 机械/模具设计/制造 人事科员 采购/外贸
信审经理 工地电工助理 行政人事办主任 销售渠道
高速项目经理 建筑环境/设备工程 培训-培训策划 跟单员/品质专员

The rest cluster features are document features. We use LDA to classify
resumes into 32 and 64 topics respectively. In all, there are 72 cluster features.
 V

4.3 Semantic Features

Differently from cluster features, in this we try to find the semantic meaning
of phrases and the potential relationship of the employment history directly.
According to the common sense in NLP, the meaning of a word is decided by its
context and similar words have similar contexts. Thus, we transformed the work
experiences of a resume into a ordered list of (n*7+3) phrases (n experiences in
a resume with 3 extra phrases including age, major and gender). Assume each
Chinese phrase as a word and each list of phrases as a sentence, after word2vec
training, each Chinese phrase would be presented by a vector of 10 dimension
and an example of similar semantic meaning is shown in Table 3.

Table 3: An Example of Similar Semantic Meaning

similarity similarity
软件工程师 .938 高级软件工程师 电子商务 .900 艺术设计
软件工程师 .932 开发工程师 电子商务 .876 国际经济/贸易
软件工程师 .890 软件测试 电子商务 .864 英语
软件工程师 .864 技术支持 电子商务 .844 广告学

5 Machine Learned Models

5.1 Shallow Estimators

(a) (b)

Fig. 2: The Performance of RF and XGB with Different Parameters

Random Forests (RF) is an ensemble learning method using the general tech-
nique of random decision trees. Each tree in the ensemble is built from a sample
VI

drawn from the training set and the best split is picked among a random subset
of the features. XGBoost (XGB), a short form for ”Extreme Gradient Boosting”,
is an optimized distributed gradient boosting library designed to be highly effi-
cient, flexible and portable. The grid search results of Results of RF and XGB
are shown in Figure 2. After analysis, we find that 473 features are used in XGB
while only 163 features are selected by RF.

5.2 Deep Estimators

LSTM is a recurrent neural network which is well-suited to learn from experience

to classify, process and predict time series. CNN is formed by a stack of distinct
layers that transform the input volume into an output volume through Convo-
lutional layers and Pooling layers. The architecture of our CNN is as shown in
Figure 3.

Manuel

Cluster
Convolution
Pooling Softmax

Semantic

Flatten

Full-
connect

Embedding
Flatten

Fig. 3: The Architecture of CNN Model

6 Ensemble Method and Results Analysis

Bagging is one of the earliest and simplest ensemble based algorithms. Usu-
ally, individual classifiers will be combined by taking a simple majority vote
of their decisions. Assume that there are three classifiers to make a positive or
negative predication. We improve the bagging method (named IBagging) by vot-
ing according to the sum of decision probabilities and can easily be extended to
multi-class ensemble. Without any information retrieval techniques and machine
 VII

learning methods, the basal manual rule will recommend the most frequent label
as the recommend item. Then, we can measure our resume-job matching solu-
tion in two ways. One is precision, whose goal is to cover as many of correct
positions. The results are as shown in Table 4. By analyzing the experiments,
we can find that XGB performs best among four basal estimators with longest
training time, while CNN model convergences in shortest time with acceptable
precision. In the meantime, our solution benefits from both bagging methods
and our semantic unsupervised feature extract method.

Table 4: The Precision of Different Classifiers

based on degree salary size position training time
XGB manual features .676 .509 .392 .460 20m 6s
XGB semantic features .685 .498 .391 .458 41m 24s
XGB all features .704 .511 .396 .467 53m 19s
RF all features .666 .511 .394 .453 7m 58s
CNN all features .695 .508 .391 .465 1m 14s
LSTM all features .696 .507 .390 .454 5m 52s
Bagging all features .699 .517 .396 .476 -
IBagging all features .710 .516 .397 .477 -
Manual Rule frequent item .484 .254 .256 .141 -

As we know, lots of resume-position pairs may not appear in testing data, but
they are reasonable and often quit similar to those correct pairs over the training
dataset. Thus, the other evaluation method, recall for Top-N recommendations
is used to evaluate different matching solutions. In this case, recall is the pro-
portion of the correct position from the testing dataset. There are 32 possible
position for a resume, with their probabilities given by classifiers, the solution
recommends top N positions to a given resume, and reports recall for various
values of N . The results of Top-N are as shown in Table 5. The results shows
a significant improvement in recall for Top-N using IBagging method compared
to the baseline method.

Table 5: Recall for Top-N Recommendations

Manual Rule IBagging
N =2 N =3 N =4 N =2 N =3 N =4
size .467 .664 .784 .629 .783 .898
degree .929 1.00 1.00 .965 1.00 1.00
salary .394 .573 .712 .800 .920 .971
position .223 .299 .373 .647 .726 .780
VIII

7 Conclusion and Feature Work

In this paper, we have considered the resume-job matching problem and pro-
posed a solution by using unsupervised feature extraction, surprised machine
learning methods and ensemble methods. Our solution is completely date-driven
and can detect similar position without extra semantic tools. Besides, our so-
lution is modularized and can rapidly run on GPU or simultaneously run on
CPU. Compared to a manual rule-based solution, our method shows better per-
formance in both precision and Top-N recall. Our code is now public and can
be tapped from Github3 . In the future, with more information to be snatched
from website, our solution could be extended by including location information,
professional skills and description of requirements from both job seekers and
employers.

8 Acknowledge
This work is supported by the National Science Foundation of China (Grant Nos.
61502082) and the Fundamental Research Funds for the Central Universities
(ZYGX2014J065).

References
1. Al-Otaibi, S.T., Ykhlef, M.: A survey of job recommender systems. International
Journal of the Physical Sciences 7(29), 5127–5142 (2012)
2. Carrer-Neto, W., Hernández-Alcaraz, M.L., Valencia-Garcı́a, R., Garcı́a-Sánchez,
F.: Social knowledge-based recommender system. application to the movies do-
main. Expert Systems with Applications 39(12), 10990–11000 (2012)
3. Chollet, F.: Keras. https://github.com/fchollet/keras (2015)
4. Golec, A., Kahya, E.: A fuzzy model for competency-based employee evaluation
and selection. Computers & Industrial Engineering 52(1), 143–161 (2007)
5. Guo, S., Alamudun, F., Hammond, T.: Résumatcher: A personalized résumé-job
matching system. Expert Systems with Applications 60, 169–182 (2016)
6. Lu, J., Wu, D., Mao, M., Wang, W., Zhang, G.: Recommender system application
developments: A survey. Decision Support Systems 74(C), 12–32 (2015)
7. Malinowski, J., Keim, T., Wendt, O., Weitzel, T.: Matching people and jobs: A
bilateral recommendation approach. In: null. p. 137c (2006)
8. Paparrizos, I., Cambazoglu, B.B., Gionis, A.: Machine learned job recommenda-
tion. In: Proceedings of the fifth ACM Conference on Recommender Systems. pp.
325–328. ACM (2011)
9. Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O.,
Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A.,
Cournapeau, D., Brucher, M., Perrot, M., Duchesnay, E.: Scikit-learn: Machine
learning in Python. Journal of Machine Learning Research 12, 2825–2830 (2011)
10. Resnick, P., Varian, H.R.: Recommender systems. Communications of the Acm
40(3), 56–58 (1997)
11. Zhang, L., Fei, W., Wang, L.: P-j matching model of knowledge workers. Procedia
Computer Science 60(1), 1128–1137 (2015)

3
https://github.com/lyoshiwo/resume job matching