Skills-Engine

What is the problem?

Given a candidate resume and a query of desired role, identify the skill gap between the candidate and the desired role, then generate a unique learning path for the candidate.

What is this project for?

This project is a testing playground that was used to record all progress for the project and how to solve the problems accordingly.

Instructions to run the notebook

Place all the files in ./app/datasets/ and ./app/pkl_files/ in the same directory level as ./app/Skills Engine.ipynb. Sections 1 to 5 are the walkthrough sections while Section 6 is the final compilation of all the functions.

Process

Step 1 Skills Extraction

In order to efficiently extract skills out from the resume, use free online EMSI Skills api to get the list of available skills.

• EMSI Skills API

  Example: calling the API endpoint using Python. The detailed documentation is given [Here](Link text Here)

  url = "https://auth.emsicloud.com/connect/token"
  
  payload = "client_id=<CLIENT_ID>&client_secret=<CLIENT_SECRET>&grant_type=client_credentials&scope=emsi_open"
  headers = {'Content-Type': 'application/x-www-form-urlencoded'}
  
  response = requests.request("POST", url, data=payload, headers=headers)
  

• Skills Extraction from Resume

  Example: extract all the skills from a given resume. The output will be in a Python set.

  extract_skills(resume_text=resume_string, skills_api=skills_api, clean=True)
  
  # Output:
  # {
  #     'ada',
  #     'business administration',
  #     'coaching',
  #     'conflict resolution',
  #     'environmental quality',
  #     'labor law',
  #     'management',
  #     'mediation',
  #     'organizational development',
  #     'policy development',
  #     ...
  # }
  

Step 2 Identification of Year of Experience

The output shown above is implicitly indicating that all the skills have the same experience level. This is undesirable as there is much more information in the resume that can be used to weigh the skills differently, e.g. project period.

In order to weigh each skill differently based on the project period, we assumed that:

• All the experience levels are using year(s) as unit.
• All the skills extracted from the resume have at least 1 year of experience.
• The project description shall come after the project period.
• The year of experience is computed by taking the ceiling of the year difference, i.e. $\text{difference in time} // \text{365 days} + 1$
• When there are multiple projects of different periods having the same skill, the experience level will be computed by $max(\text{years of project 1}, \text{years of project 2}...)$.
• The title of a section is an uppercased Noun

We will first break down the resume text into sections using title, for each section, we then break down the paragraph into subsections using date range

Example: extract all the sections and subsections out using title and date range as guidance from a resume.

experience_tagging(date_list)

# Output:
# key: year of experience
# value: list of tuples, where each tuple is (section start index, section end index)

# {10: [(981, 1066)], 4: [(1996, 3608)], 3: [(3635, -1)]}

Step 3 Experience Level Tagging

We looped through each section and subsection to extract all the skills out, then tag all these skills with the corresponding years of experience. For all the other skills that are not found within any sections, we just tag the year of experience to be 1.

Example: extract all the skills with corresponding years from a resume. The output is in a Python dictionary.

skills_experience_level_identification(resume_string)

# Output:
# {
#   'reduction': 4,
#   'environmental quality': 4,
#   'teamwork': 4,
#   'management': 4,
#   'poultry': 4,
#   'strong work ethic': 4,
#   'business administration': 3,
#   'coaching': 3,
#   'conflict resolution': 3,
#   'workplace safety': 3,
#   'mediation': 3,
#   'organizational development': 3,
#   'session': 3,
#   'ada': 3,
#   'labor law': 1,
#   'policy development': 1
# }

Step 4: Identification of Significant Skills

There are 2 sets of documents that significant skills can be identified from, job-description documents and resume documents.

• From job-description documents, we extract the required skills to meet current industry standards.
• From resume documents, we extract the required skills to standout among peers in the same industry. Assumption: all candidates applying for the specific role must be qualified, i.e. they have all the skills required for the particular role.

Aggregation is performed on all documents that belong to the same category to find out the frequency of documents that have each specified skill. High frequency indicates that they are the common skills required for the particular role.

We then set the significance level $\alpha$ as .95, hence all the skills below 95 percentile will be treated as not so significant. The remaining frequencies are scaled to the range of [0, 1] to help us decide on the competency level to be given to each skill.

Example: We use the percentile method to set the threshold for classifying the skills significance level. We save all these remaining skills with the corresponding competency level required into a dictionary.

# Output: where the value is the competency level based on the client's business settings.
# {'ADVOCATE': 
#  {'active listening': 3,
#   'administrative support': 3,
#   'advocacy': 3,
#   'agenda': 3,
#   'auditing': 3,
#   'authorization': 3,
#   'banking': 3,
#   'billing': 3,
#   'budgeting': 3,
#   'business administration': 3,
#   'business development': 3,
#   ...
#  }, 
#  ...
# }

Step 5: Identification of Skill Gap

With the significant skills identified for each industry, we can now make comparison between your skillset and the skillset required for a particular industry. We can do that by first mapping the years of experience into pre-defined competency levels: 3 for entrant level, 4 for specialist level and 5 for expert level. Then a simple intersection is performed to get the skills gap, i.e. $\text{Required Level 5} - \text{User Level 3} = \text{Skills Gap [4,5]}$.

Example: The output will be a dictionary with skills as the keys and the skills gap as the values

skill_gap_identification_peers(skills[31605080], skills_required['AVIATION'])

# Output:
# {'aircraft maintenance': [3],
#  'aviation': [4, 5],
#  'b': [3],
#  'business administration': [3],
#  'c': [3],
#  'construction': [3],
#  'critical thinking': [3],
#  'drawing': [3],
#  'filing': [3],
#   ...
#  'supervision': [3],
#  'test equipment': [3],
#  'time management': [3],
#  'track': [3, 4],
#  'tracking': [3]
# }

Step 6: Generating The Learning Path

Once we successfully find out the skills gap between an applicant and the job and peer standards, we can suggest the applicant to job and continuous learning programmes which help the applicant to fulfill the skills gap and even outperforming their peers in the specific industry.

Example: We infer skills required in a vector form using Doc2Vec and find the top 5 unique most similar courses

# Output:
# {'job': 
#        {
#            3: [(7203,'Innovation and Entrepreneurship Capstone','3 - Entrant Level'), ...],
#            4: [(4602, 'Express Data Base Administrator', '4 - Specialist Level'), ...],
#            5: [(12502, 'Cyber Security Management Capstone Project','5 - Expert Level'), ...]
#        },
#  'continuous learning':
#        {
#            3: [(7203,'Innovation and Entrepreneurship Capstone','3 - Entrant Level'), ...],
#            4: [(4602, 'Express Data Base Administrator', '4 - Specialist Level'), ...],
#            5: [(12502, 'Cyber Security Management Capstone Project','5 - Expert Level'), ...]
#        }
# }

Final Function

The Final Function is a single function that takes in a user's resume and desired industry/job name to output a learning pathway to fulfill their skills gap idenitified. Refer to Section 6 of Skills Engine.ipynb for the following information.

6.1 contains the Python Libraries required to run this final function.

6.2 contains the Helper Functions that make up this final function.

6.3 contains the Pre-Training Codes that will create the pickle files required for the final function to run. Can be used to reproduce this resources whenever there is any change to their sources.

• skills_api.pkl used in skills extraction, created from all_skills_emsi.xlsx.
• resume_skills_required.pkl used in skills gap identification, created from Resume.csv.
• job_skills_required.pkl used in skills gap identification, created from Job Description.csv.
• d2v_model.pkl used in learning pathway generation, trained from Courses.xlsx.

6.4 contains the Final Function with detailed comments.

final_course_suggestion_d2v(resume_text, 'INFORMATION-TECHNOLOGY', 'Software Developers, Applications',
                            skills_api, resume_skills_required_pickle, job_skills_required_pickle, 
                            courses_dataset, doc2vec_model, True)

How to Build and Tag a Docker Image?

Due to the time remaining for devlopment is limited and we need to start compiling and documenting all the source codes, the current limitation is:

• we do not have the access to directly call PUT request to save a pdf file to S3 bucket via API Gateway
• we do not have the access to read pdf files from S3 bucket via Lambda (errorMessage: when calling GetObject operation: Access Denied)

Hence, we decided to use GET request with the key to be doc and the value to be resume_string. This can be easily modified to only take the pdf filename once the access to read files from S3 bucket is given.

Follow the instructions below to build the Docker image:

1. Make sure that you have the latest version of the AWS CLI and Docker installed.
2. Type aws configure in cmd and specify the correct Access Key ID and Secret Access Key. They can be found from AWS IAM.
3. Type cd "<path_to_folder>/bt4103team13/app/SkillsExtract Docker"
4. Retrieve an authentication token and authenticate your Docker client to your registry. Further details can be referred from AWS ECR repository -> view push commands
5. The current directory structure shall be:

SkillsExtract Docker
|__ /app
|__ Dockerfile
|__ requirements.txt

6. Type docker build -t skills-engine .
7. Further details can be referred from AWS ECR repository -> view push commands

The current Docker Image only includes the function to extract skills and corresponding years of experience for the demonstration purpose. To include more lambda_handler to show skill gap and recommended courses, follow the instructions below:

1. Replace ./app/SkillsExtract Docker/requirements.txt with ./app/requirements.txt
2. Replace ./app/SkillsExtract Docker/app/functions.py with ./app/functions.py
3. Create other .py files that contain the lambda_handler in the same directory level as ./app/SkillsExtract Docker/app/lambda_function.py
4. Add all the other necessary .pkl files to the current directory level from ./app/pkl_files
5. Modify ./app/SkillsExtract Docker/Dockerfile correspondingly to be able to call different lambda_handlers

Dashboard

Code found in update_dashboard_data/main.py

Aim and rationale

The dashboard found on QuickSight consists of data from two sources: 1. the current model output as represented through top 10 skills (and in the future, top 10 courses), and 2. the metric data from CloudWatch logs that indicates model performance like the number of invocations and the model runtime required to output results.

The aim of this file and its function are to automatically track skills output from the model everytime it is run to update skills_freq.json saved in S3, and to pull the latest data from CloudWatch within a time period to update invocations.json, avg_duration.json, highest_duration.json and errors.json in the S3 Bucket. The code in this file also reformats the data ingested into a format usable by QuickSight to display the given dashboard with the latest data.

Originally, the function was meant to use AWS Invoke to be run every time the model Lambda was run and update the corresponding json files saved in S3. S3 was to be connected to QuickSight so that QuickSight could update the dashboard to the latest results on demand.

However, limitations included:

• Inadequate permission access to link S3 Bucket to QuickSight for automatic update
• Unable to deploy API Gateway to test the entire functionality pipeline with integrated dashboard code as Lambda function, and thus unable to use AWS Invoke to gain access.

Instructions to utilise current code to obtain data

Currently, code must be manually run to update to latest data for dashboard

1. After running initial model code to get skills output as json, copy-paste the json output to variable skills_json in lambda_handler function
2. Navigate to ./update_dashboard_data/main.py and run python main.py.
3. This will update all json files in the S3 Bucket
4. Manually download invocations.json, avg_duration.json, highest_duration.json, errors.json and skills_freq.json into local folder
5. Manually upload files into QuickSight dashboard for updated data.

Suggestions for later integration into proper AWS pipeline

1. Link nus-sambaash/skills-engine-dashboard Bucket to QuickSight
2. Convert functions into suitable Lambda function format by removing name == main portion
3. Use AWS Invoke function to link current function to model Lambda function to be triggered together and to also receive skills json output from model function.