Job recommendation is a crucial part of the online job recruitment business. To match the right person with the right job, a good representation of job postings is required. Such representations should ideally recommend jobs with tting titles, aligned skill set, and reasonable commute. To address these aspects, we utilize three information graphs (job-job, skill-skill, job-skill) from historical job data to learn a joint representation for both job titles and skills in a shared latent space. This allows us to gain a representation of job postings/ resume using both elements, which subsequently is combined with location. In this paper, we rst present how the representation of each component is obtained, then we discuss how these di erent representations are combined together into one single space to acquire the nal representation. The comparison of the proposed methodology against base-line methods shows signi cant improvement in terms of relevancy.
Online recruiting and job portals like Careerbuilder.com, Linkedin. com, and Indeed.com, have become the norm in the talent acquisition business. Millions of jobs are posted and even more resumes are uploaded daily. Di erent machine learning and information retrieval models have been applied to analyze these resumes and job descriptions, and multiple e orts have been made to match the two parties of the recruiting process. A good job/resume representation helps to improve many downstream products that in turn support the company's goal of empowering employment and helping job seekers nd jobs and the training they need. Speci cally, it facilitates matching job seekers and employers by improving our search and recommendation products.
The content of a job posting incorporates all aspects of a position.
Entities like job title, required skills, experience, degrees, bene ts, company culture,
location, etc, can be extracted and normalized. Among these, title, skills, and
location are the top factors when de ning a job position. While a job title
determines the nature of a job, skills enrich the job title, and di erentiate jobs with
the same title by identifying their niceties. In order to o er a good job match,
an accurate representation consisting of both is needed. Rather than simply
using various word representations of job title and skills, recent works [
Our contributions in this paper are as follows:
The vector representation proposed is applicable for both job postings and resumes. It's not only a exible representation to obtain similar jobs or similar candidates, but also provides a direct mapping of jobs and resumes.
We generate a more holistic vector representation jointly learned for both titles and skills which can be used in the job recommendation system.
We incorporate location explicitly in our representation.
We employ retro tting to re ne the job vectors by using the skill vectors, detailed in Section 3.3.
Job Postings Resumes
P a r s e r Recommendation
Job Title Skills Location
R e ingnLea ittrsenape itJon r o n 3D Vectorization
The two major solutions of job recommendation in existing works are
contentbased representation learning and collaborative ltering (CF). CF builds a very
high dimension sparse matrix to maintain the relationship between user and
product [
Most recent works, including our proposed system, use hybrid approaches of
both pro le properties and interaction history [
Our work is similar to the hybrid approaches in terms of learning the embedding of job titles and skills based on both their characteristics and interactions. Meanwhile, it is di erent that, instead of simply pairing the related jobs and skills, our pro le representation models three speci c relationships simultaneously between jobs and required skills. Moreover, we employ retro tting to achieve better representation. 3
In this section, we describe the design and methodology of our tripartite vector representation. A job posting contains important information of the hiring position. Most importantly it includes job title, skills, and location. Here, we use a combined representation vectors of job titles and skills, and obtain a vector representation for location transformed from latitude and longitude. Figure 1 shows the design of our method. Job posting and resume data go to an in-house job parser, where job title/skills/location are extracted. While titles and skills are jointly trained by a representation learning framework (Section 3.1), location information is also vectorized and normalized (Section 3.2). In order to combine the tripartite, the title representation is rst retro tted by a list of skills associated with each job before added with the location vector (Section 3.3).
Title and skills are the de ning features of any job. Embeddings for both titles
and skills are learned in the same k-dimensional space [
For the job-job transition graph, we assume that the transition between similar jobs x and y is more likely to happen than non-similar jobs x and z. Let Ajxy = hwx; wyi, the dot product of the two embedding vectors, be the a nity score between job x and job y, and Djj (job-job) represents the transition relationship of job triplets (x; y; z). The objective is to learn representation W so that
Ojj = min
W
X ln (Ajxjy
(x;y;z)2Djj Where sigmoid function ( ) = 1+1e is used as the probability function which preserves the order Ajxy > Ajxz.
Similarly, for the skill-skill graph, coexisting skills x and y which appear on the same job posting or the same resume are closer to each other than noncoexisting skills such as x and z. Let Dss (skill-skill) be the set of training triplets of skills with coexisting relationship, our objective here is
Ojs = min
W W 0
X (xj;ys;zs)2Djs ln (Ajxsy
(1) (2) (3) (4) Oss = min W 0
X (x;y;z)2Dss ln (Asxsy
Moreover, for the job-skill graph, if skill ys appears on the advertisement of job xj , its embedding vector is more similar to xj than non-related skill zs. Given the set of training triplets Djs (job-skill), our desired vector representation of jobs W and skills W 0 are learned according to the objective
Finally, to achieve high quality job and skill embedding, we optimize this joint objective function
O(W; W 0) = min Ojj + Oss + Ojs +
W;W 0 (jj W jj2F + jj W 0 jj2F ) where is the coe cient of the l2 regularization term jj jj2F to avoid over- tting.
Vectors of dimension size 50 are obtained for 4325 unique job titles and 6214
skills, using joint Bayesian Personalized Ranking (BPR) [
Location is another key factor in a job posting. Commute time matters when people are choosing a position. The common practice when dealing with location is to pre- lter or post- lter the recommendations with a xed radius. This method has a few downsides: 1. Some jobs are less sensitive to distance than others. For example, people are more willing to commute longer with a highly compensated job than a minimum wage part-time one; 2. The system is more di cult to implement because of the extra layer of ltering. 3. Speci cally for our case where Faiss is used for similarity search, only one index le is needed if location is embedded in the vector.
Before latitude and longitude can be added to our embedding model, they need to be transformed since they are not on the same scale as the title and skill vector. Geo-locations are three dimensional in its nature, and to represent location in a similar fashion as title and skill vectors, we perform a transformation of latitude and longitude, as shown in Figure 2: x = cos( )
cos( ) y = cos( )
sin( ) z = sin( ) (5) (6) (7) Where represents the latitude and is the longitude.
Thus, location is represented as a normalized three dimensional vector, which can be later combined with the title and skill vector.
Longitude
Latitude Fig. 2. Latitude and longitude can be converted into a three-dimensional coordinates.
Job postings with the same title might require di erent skill sets at di erent companies in di erent industries. The next natural question is how to combine the title and skill vectors together to a personalized vector for a speci c job, given both title and skill vectors are trained in the same latent space.
To assemble the vectors back into one to represent the job posting, we apply
the retro tting method [
In Faruqui's work [
Pj:(i;j) ijqj +
q^i Pj:(i;j) ij + i (8) where qi is the modi ed vector, q^i is the initial vector, qj are the neighbors. This is derived by minimizing the distance between initial qi and neighbors. And for their case, the results converge after 10 iterations.
Intuitively, if we add two vectors together, the result vector will lie in between the initial two vectors. Adjusting vectors by adding the vectors of neighboring words will bring words similar in meanings closer to each other. For example, in Figure 3, Sm^ile and T e^ars are the initial vectors for two very di erent words, so they are further apart in direction (cosine similarity). After the tuning using a neighboring word: happy, the new vectors for smile and tears are closer than before, since the tears are in the context of "happy tears".
Similarly, we apply the above operation to our job posting data, and generate a combined representation of the job posting using title and the skills. Jobs with similar set of skills as well as the same title will have a higher rank than the ones with the same title but less overlapping skills. Rather than going through several iterations as in the original paper, we only apply the updates once, as skill vectors are not updated in the process. We can view this one time update as ne tuning of the title vector, or the calculation of the "job + title" vector, shown as below, where n is the number of skills in a job posting, which are given the same weight:
qjob = n qtitle2+n P qskill (9)
Location vector is then concatenated with the new vector. The weight of the location vector can be adjusted based on how the jobs are sensitive to distance. qjob+location = [qjob; x; y; z] (10) To test and evaluate our proposed system, we use a real dataset (jobs and users) via CareerBuilder.com. CareerBuilder operates the largest job posting board in the U.S. and has an extensive growing global presence, with millions of job postings, more than 60 million actively searchable resumes, over one billion searchable documents, and more than a million searches per hour.
Each job/ resume is parsed into job title, skill list, and location (Figure 1). We then translate these three parts into their vector representations, which are later combined to the nal job posting vector. We have also curated a list of top skills for each job title, in case skills are not available. The curation is based on millions of resumes and job postings, as well as human knowledge. We use this skill list to generate a vector if no skill is given. This provides a more accurate representation of the job than just using a job title. These vectors all go into our recommendation engine for di erent recommendation products.
We calculate the vectors on 300K actual job postings on the Careerbuilder
Website, using 6 di erent models (FastText, Word2Vec, Glove-6B-300 ,
Glove840B-300 [
Table 1 shows that by explicitly including latitude and longitude in our embedding, the average distance of recommended jobs is reduced by 90%. All top 50 jobs are within a reasonable proximity of the original input job.
Table 2 shows the number of jobs within 50 miles in the top 50 jobs increased dramatically for retro tter with location embedding. Both table 1 and table 2 show that the base-line methods are not location sensitive, even though location is included in the text le. Implicitly having location in the embedding is far from enough, and the advantage of explicitly including location in the embedding model is prevailing.
Table 3 shows title match rate and coverage. Our retro tter model with no location embedding has a much higher title match rate and coverage than the base-line methods as well as the retro tter model with location. However, the drop in performance in our location retro tter model purely comes from the limited relevant jobs in close proximity, which is a trade-o we have to make. A perfect job match from ten thousand miles away is simply not a perfect job for most people. Even so, our retro tter model still has a better results in both measurements than the base-line methods. In this paper, we discuss the representation model which can be used for a recommendation system and it is currently being utilized within CareerBuilder. Three facets of a job posting are considered: job title, job skills, and location. While job title carries the most weight in determining what a job is, skill set de nes the nuances which di ers from job to job. Most job seekers are also very sensitive to the location of a job. In our model, we encompass all three of these aspects, and are able to give location sensitive, highly related job recommendations to our users.
There are a number of improvements being worked on such as : Develop an inductive learning framework to accommodate newly emerged job titles and skills, as the current model is transductive, and representation vectors only exist if it is in the input graph; Incorporate more features in the job representation such as education and previous experience; Adjust the location embedding in a more quanti able way to control the radius of recommended jobs; Combine the current representation model with other models to provide a better results; Apply di erent weights to skills based on their importance to the job title.