Updating Readme m4

.flake8

@@ -0,0 +1,4 @@
+[flake8]
+max-line-length = 88
+extend-ignore = E203, W503, E231, E128
+exclude = .git, __pycache__, .ipynb_checkpoints

.github/workflows/lint.yml

@@ -0,0 +1,27 @@
+name: Lint Python
+
+on: [push, pull_request]
+
+jobs:
+  lint:
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v3
+
+      - name: Set up Python
+        uses: actions/setup-python@v4
+        with:
+          python-version: '3.10'
+
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip
+          pip install flake8 nbqa
+
+      - name: Lint .py files
+        run: flake8 --exclude=.venv,__pycache__ .
+
+      - name: Lint .ipynb notebooks
+        run: nbqa flake8 .

.gitignore

@@ -5,3 +5,7 @@ venv/
 *.db
 *.idea
 *.ruff_cache
+.env
+.venv_datacleaning/
+*.ipynb_checkpoints
+.env

.ls-lint.yml

@@ -1,17 +1,27 @@
+version: 1
+
 ls:
   .dir: snake_case
   .*: snake_case
   .md: snake_case | regex:[0-9A-Z\-]+
   .txt: snake_case | kebab-case
   .yml: snake_case | kebab-case
 
+rules:
+  filename:
+    regex: '^[a-z0-9_]+$'
+    match: true
+
 ignore:
   - .git
   - .github
   - .vscode
+  - "**/.vscode/**"
   - venv
   - .ruff_cache
   - .pytest_cache
   - __pycache__
   - .ls-lint.yml
   - .markdownlint.yml
+  - "**/node_modules"
+  - README.md

.vscode/settings.json

@@ -122,5 +122,9 @@
       "source.fixAll.ruff": "explicit",
       "source.organizeImports.ruff": "explicit"
     }
-  }
+  },
+  "cSpell.words": [
+    "Alaa",
+    "stopwords"
+  ]
 }

0_domain_study/README.md

@@ -1 +1,123 @@
 # Domain Research
+
+## Event
+
+In the modern time, the evolution of technology has transformed the way people
+navigate their professional lives. Nowadays, online job platforms serve as a
+shared environment between recruiters and job seekers.
+The influence of internet has become central to recruitment strategies, hence, the
+recruitment process has increasingly moved online. This shift has also brought
+new challenges, as although the number of global job postings is increasing
+with this shift, not all of them are legitimate.
+
+Online job scams surged by 84% in 2023 worldwide. The rise of remote work,
+online recruiting, and AI tools has made it easier for scammers to create
+convincing fake job ads. This concerning trend continues to escalate, blurring
+the line between legitimate opportunities and fraudulent schemes. It has become
+crucial to understand what specific features in job postings and application
+processes influence individuals’ engagement, and more importantly, their ability
+to recognize and avoid scams.
+
+This research aims to explore the role of visual and textual elements,
+application process, along with LLMs in shaping user perceptions and actions
+in the online job-seeking environment.
+
+---
+
+## Pattern
+
+One emerging pattern is the use of highly attractive job descriptions and offers,
+often mimicking real postings. These ads typically highlight flexible hours,
+remote work, and unusually high salaries, features that immediately catch the
+attention of job seekers, especially those urgently looking for income or remote
+opportunities.
+
+Simultaneously, scammers have become more sophisticated in how they design and
+execute fake job listings. Leveraging LLMs,
+they craft professional-sounding job descriptions and even imitate the branding
+of real companies. These tools allow scammers to avoid earlier red flags such
+as poor grammar or suspicious formatting, making the scams appear more credible
+and trustworthy.
+
+A major red flag in many scams involves the request for upfront payments. Whether
+it's for background checks, training materials, or equipment, fraudulent
+postings often ask for money early in the process with a promise of being paid back.
+Patterns show that the
+decision to proceed or back out often depends on how the request is framed. If the
+request is made after a series of
+seemingly normal interactions, people are more likely to comply, especially if
+they’ve already invested time and emotional energy in the application.
+
+---
+
+## Structure
+
+At a structural level, one of the key reasons online job scams are so effective
+is that fraudulent processes often mimic the legitimate structure of real hiring
+systems. Across industries and regions, job applications tend to follow a
+standardized format. Most job seekers are familiar with being asked for personal
+details such as **_full name_**, **_address, date of birth, contact information_**,
+and even **_bank details_** in some cases. In many cases,
+**_background checks_** and **_document verification_** are also expected steps.
+These practices are so deeply rooted in professional norms that applicants rarely
+question them, which makes it easier for scammers to
+blend in without raising suspicion.
+
+---
+
+## Mental Mode
+
+Underlying beliefs and social pressures play a significant role in sustaining
+the system in which online job scams thrive. One core belief is rooted in the
+reality that the global job market is increasingly competitive. With more
+individuals seeking employment than there are available positions, people often
+feel immense pressure to secure a job as quickly as possible. This urgency is
+further amplified by the growing concerns over financial stability and societal
+expectations. Therefore, job seekers may be more willing
+to overlook potential red flags if it means landing a promising opportunity.
+
+Additionally, feelings of shame and embarrassment further enforce the system.
+Victims of job scams often blame themselves, fearing judgment if
+they share their experience. This silence prevents others from learning about
+common scam tactics and creates an environment where fraudulent practices can
+continue unchecked.
+Along with the fact that this issue is viewed
+as individual problems, rather than systemic ones. Society often emphasizes
+personal responsibility, encouraging individuals to “try harder” or
+“be more careful”, instead of addressing broader structural issues.
+
+These internalized beliefs about urgency, personal failure,
+and individual responsibility, act as invisible forces that keep the scam
+ecosystem alive and difficult to disrupt.
+
+---
+
+## Conclusion
+
+By mapping the user journey, we can better understand how people navigate scam
+encounters. When users say, “ _**What would I lose? I’ll apply and see how this
+goes,**_ ” it reflects a sense of low-risk experimentation driven by desperation
+or hope. Internally, they may think,
+“ _**I can’t help but set high hopes incase I get the offer,**_ ”
+revealing the emotional vulnerability that scammers exploit. Behaviorally, users
+often proceed to apply despite some awareness of scams. Emotionally, the journey
+is intense, initial excitement when
+they’re contacted, followed by feelings of disappointment, panic, and
+discouragement once they realize the offer was fake.
+
+This research aims to deeply understand the
+online job-seeking experience from the user's perspective, empathizing with their
+motivations, behaviors, and emotional journey. By this research we're planing to
+go beyond just raising awareness, and actually identify patterns to build
+preventative strategies, platform-level solutions, and behavioral cues that
+empower users earlier in the process.
+
+---
+
+## Sub Domains
+
+- Feature engineering.
+- NLP ( Natural Language Processing ).
+- Human-Machine interaction.
+- Machine-Machine interaction.
+- LLMs.

0_domain_study/brainstorming.md

@@ -0,0 +1,7 @@
+# Brainstorming
+
+To access our **_group_** brainstorming:
+[**Group Brainstorm**](https://docs.google.com/document/d/1GRCxNJAbXSTDNgDnKzDF3IzyhiA_VJNgygS9rDqrbrI/edit?usp=sharing).
+
+To access our **_initial presentations_** and brainstorming:
+[**The Hypatia Circle: Individual Brainstorm**](https://drive.google.com/drive/folders/1pBaj0Guxkd12CB6fYI9RAk0tbH2KZCLa?usp=sharing).

0_domain_study/sub_questions_and_refrences.md

@@ -0,0 +1,18 @@
+# Sub-questions
+
+- What type of features or details are included in job postings that influence
+people to interact and apply?
+- Given the fact that postings could be created by LLMs, how does this affect
+people's ability to distinguish legitimate postings from fraud ones?
+- What type of features or details are included in the application process that
+influence people to continue with the process?
+- What type of features or details are included in the request for upfront
+payment that influence people's decision whether to pay and proceed or to
+recognize this as a scam?
+- If a ‘scam percentage score’ appears on the post, what is the impact of it?
+Could it be helpful?
+
+## References
+
+- [BBC News](https://www.bbc.com/news/business-66592219).
+- [User Cases](https://assets.publishing.service.gov.uk/media/61981a328fa8f5037e8ccacb/Job_Scams_-_Case_Studies.pdf).

1_datasets/README.md

@@ -1 +1,162 @@
 # Datasets
+
+Only one dataset was used for the analysis process of this project, that is the
+**Employment Scam Aegean Dataset**. All data processing, analysis, and
+observations are mainly based on this dataset, along with 500 real job posts
+which were extracted from **Indeed.**
+
+## Aegean Raw Data
+
+This dataset can be found in **Kaggle**, referred to as **_'Recruitment Scam
+Dataset'_**, or in
+**EMSCAD** project **website**, referred to as **_'Employment Scam Aegean Dataset'_**.
+
+This dataset was collected and curated by the **_Laboratory of Information and Communication
+Systems Security_** at the **_University of the Aegean_** in **_Greece_**.
+It contains **18,000** job postings, with around **800** labeled as fraudulent.
+
+This dataset
+was mainly designed to  provide a realistic and comprehensive resource for research
+on employment scams. It was collected from real online job ads between **2012**
+and **2014**.
+The job postings were gathered from multiple **_online_** sources, including **_job
+portals_** and **_corporate websites_**.
+
+Each entry in the dataset includes a
+variety of features such as **_job title, location, department, salary range, company_**
+**_profile, job description, requirements, benefits, telecommuting status, company
+logo_**, **_presence, employment type, required experience, required education, industry,
+function,_** and a **_class label_** indicating whether the posting is fraudulent
+or not.
+
+The dataset is publicly available and has been widely used in academic research
+for developing and testing machine learning models to detect fraudulent job postings.
+
+The main goal of this data is to clean it based on specific features that we aim
+to use in the job board, organize it, refine it by Gemini, then use it to test
+humans' and machines' ability to detect if the job provided is legitimate or
+fraudulent when it's written by AI, since it mimics real job posts while
+maintaining fraudulent posts main features.
+
+[**File Path to Aegean Raw Data**](https://github.com/MIT-Emerging-Talent/ET6-CDSP-group-21-repo/blob/28fb2c5be79be0883c8366fb2b4bacbbec9c6809/1_datasets/aegean_raw_data)
+
+---
+
+## Cleaned Data
+
+There are three files in the [`../1_datasets/cleaned_data`](https://github.com/MIT-Emerging-Talent/ET6-CDSP-group-21-repo/blob/1559fd4f70f49837b9626a46db57799e8c5a39da/1_datasets/cleaned_data)
+folder:
+
+### [Fake Job Posts](https://github.com/MIT-Emerging-Talent/ET6-CDSP-group-21-repo/blob/14894562ec2b519501aaed5b0525f54313fdfb0f/1_datasets/cleaned_data/fake_jobs.csv)
+
+- Includes 866 fake job posts which were extracted from the Aegean raw dataset. Shape
+of the dataset is (866, 11) after cleaning.
+
+### [Real Job Posts](https://github.com/MIT-Emerging-Talent/ET6-CDSP-group-21-repo/blob/14894562ec2b519501aaed5b0525f54313fdfb0f/1_datasets/cleaned_data/real_jobs.csv)
+
+- Includes 17014 real job posts which were extracted from the Aegean raw dataset.
+Shape of the dataset is (17014, 11) after cleaning.
+
+### [LLM-Refined Fake Job Posts](https://github.com/MIT-Emerging-Talent/ET6-CDSP-group-21-repo/blob/1559fd4f70f49837b9626a46db57799e8c5a39da/1_datasets/cleaned_data/llm_refined_fake_posts2.csv)
+
+- Includes 866 fake job posts which were all LLM-refined. Shape of the dataset
+is (866, 15) after cleaning.
+
+---
+
+## Aegean500_Hypatia500 Datasets Folder
+
+In addition to the primary analysis focusing on the Aegean dataset, a
+complementary analyses was done to explore the detection capabilities of four
+traditional Machine Learning Models when given fake jobs refined by AI and real
+jobs posted in the AI era.
+
+### A note on the datasets that were used for the analyses
+
+There are three files in the [`../1_datasets/aegean500_vs_hypatia500_datasets`](https://github.com/MIT-Emerging-Talent/ET6-CDSP-group-21-repo/blob/e46c53bf17c3d608c8e67b607300d9faf4b6043e/1_datasets/aegean500_vs_hypatia500_datasets)
+folder:
+
+### [**Fake Job Posts**](https://github.com/MIT-Emerging-Talent/ET6-CDSP-group-21-repo/blob/e46c53bf17c3d608c8e67b607300d9faf4b6043e/1_datasets/aegean500_vs_hypatia500_datasets/aegean500_fakejobs.csv)
+
+- The (866, 17) fake jobs in the Aegean dataset contains lots of missing values
+and needed to be cleaned for the analysis. A research of modern day job posts
+structure revealed the following as core features - **job_id, job_title, location,
+job_description, benefits** (with benefits often lumped into the job description).
+
+- These features were retained and used to remove missing values from the dataset,
+achieving a (500, 6) shape of fake jobs.
+
+- Hence the difference between this file and the one in [`../1_datasets/cleaned_data/fake_jobs.csv`](https://github.com/MIT-Emerging-Talent/ET6-CDSP-group-21-repo/blob/14894562ec2b519501aaed5b0525f54313fdfb0f/1_datasets/cleaned_data/fake_jobs.csv)
+is that they both followed different data cleaning processes.
+
+### [**Real Job Posts**](https://github.com/MIT-Emerging-Talent/ET6-CDSP-group-21-repo/blob/e46c53bf17c3d608c8e67b607300d9faf4b6043e/1_datasets/aegean500_vs_hypatia500_datasets/hypatia500_realjobs.csv)
+
+- contains 500 real jobs scraped from the job board **Indeed** (_Jobs were
+sorted by 'recently posted' on the board, and retrieved between 7/19/2025 - 7/20/2025_).
+The chrome extention "webscraper.io" was utilized for the scrapping and these
+key features - **Job_title, location, job description and job link** were
+scrapped, with source-date retained as metadata.
+
+- What informed the 'type' of real jobs to scrape?
+  - The fake jobs word cloud showed that certain job titles - Data Entry,
+  Engineer, Customer service, and Entry clerk, were dominant. These became the
+  key jobs that were searched and scrapped from Indeed.
+
+- Hence, the difference between this file and the one in [`../1_datasets/cleaned_data/real_jobs.csv`](https://github.com/MIT-Emerging-Talent/ET6-CDSP-group-21-repo/blob/14894562ec2b519501aaed5b0525f54313fdfb0f/1_datasets/cleaned_data/real_jobs.csv)
+is the fact that they are real jobs of different times,
+(aegean (2012 - 2014), hypatia (2025))
+
+### [**LLM-Refined Fake Job Posts**](https://github.com/MIT-Emerging-Talent/ET6-CDSP-group-21-repo/blob/e46c53bf17c3d608c8e67b607300d9faf4b6043e/1_datasets/aegean500_vs_hypatia500_datasets/aegean500_fakejobs_llmrefined.csv)
+
+- The 500 job descriptions were fed to Gemini 2.5 flash with a robust prompt
+aimed at modernizing the fake jobs. Specifically, this **"Add appealing but
+potentially exaggerated benefits/responsibilities."** was part of the prompt to
+bring the job descriptions up to modern standard.
+
+- The refinement task was split between team members in this file:
+  - [`/1_datasets/fakejobs_to_refine`](https://github.com/MIT-Emerging-Talent/ET6-CDSP-group-21-repo/blob/08990371387dcddd06fb6f3361478bf4c33d45fb/1_datasets/fakejobs_to_refine).
+- The refined version is seen here:
+  - [`../1_datasets/fakejobs_refined`](https://github.com/MIT-Emerging-Talent/ET6-CDSP-group-21-repo/blob/8251dfa7db2ae2e0b35c3e619dd3c7f6e52af037/1_datasets/fakejobs_refined).
+- The 6 refined batches were recombined to get the aegean500_fakejobs_llmrefined.csv
+- Hence, the difference between this file and [`../1_datasets/cleaned_data/llm_refined_fake_posts2.csv`](https://github.com/MIT-Emerging-Talent/ET6-CDSP-group-21-repo/blob/1559fd4f70f49837b9626a46db57799e8c5a39da/1_datasets/cleaned_data/llm_refined_fake_posts2.csv)
+is that they both read and refined files that followed different cleaning processes.
+
+### All scripts related to this complementary analyses
+
+- [fake jobs extraction script](https://github.com/MIT-Emerging-Talent/ET6-CDSP-group-21-repo/blob/main/2_data_preparation/fake_jobs_extraction_script.ipynb)
+- [fake jobs refinement script](https://github.com/MIT-Emerging-Talent/ET6-CDSP-group-21-repo/blob/main/2_data_preparation/fake_jobs_ai_refinement_script.ipynb)
+- [batch recombination script](https://github.com/MIT-Emerging-Talent/ET6-CDSP-group-21-repo/blob/main/2_data_preparation/recombine_aegean500_batches.py)
+- [data exploration script](https://github.com/MIT-Emerging-Talent/ET6-CDSP-group-21-repo/blob/main/3_data_exploration/aegean_hypatia_datasets_exploration.ipynb)
+- [data analysis script](https://github.com/MIT-Emerging-Talent/ET6-CDSP-group-21-repo/blob/main/4_data_analysis/aegean_hypatia_datasets_analysis.ipynb)
+
+---
+
+## Past Relevant Studies
+
+The research, **_Assessing AI vs Human-Authored Spear Phishing SMS Attacks: An
+Empirical Study_**, a **2025** study, was conducted by researchers at **_Brigham
+Young University._**
+
+To collect data, participants were recruited and asked to provide personal
+information that could be used to personalize phishing messages.
+
+Meanwhile, on the other hand, both human and AI authors used this personal information
+to craft spear phishing SMS messages tailored to each participant.
+
+There were mainly **_four_** steps for the data collection procedure. **_Firstly_**
+participants were shown a set of personalized SMS messages (some human-authored,
+some AI-generated).
+**_Secondly_**, they were asked to rank the messages by how convincing they found
+them. **_Thirdly_**, participants also provided qualitative feedback on each message.
+**_Lastly_**, they were then asked to guess whether each message was written by a
+human or by AI.
+
+This experiment measured how convincing the messages were (both written by humans
+and AI) and whether participants could distinguish between human and AI authorship.
+
+There was **_no study_** that specifically tested humans' ability to differentiate
+if a job post is legitimate or fraudulent when the post is AI-generated that we
+know of, hence we're planning to use this study as a foundation for the process
+of building and analyzing the job board.
+
+[**Past Relevant Studies File Path.**](https://github.com/MIT-Emerging-Talent/ET6-CDSP-group-21-repo/blob/549942e9039edafdae73dff7d904ec97fa432148/1_datasets/past_relevant_studies.md)