Branch-wise Analysis and Prediction of Student Placement Using Academic and Skill Factors

---

Table of Contents

1. Problem Statement
2. Project Objectives
3. Dataset Description
4. Project Structure
5. Data Cleaning Pipeline
6. Feature Engineering
7. Imputation Strategy
8. Encoding and Preprocessing
9. Final Dataset Summary
10. Tech Stack
11. How to Run
12. Results and Output
13. Key Decisions and Rationale
14. Limitations
15. Authors

---

Problem Statement

The objective of this study is to predict whether a college student will be placed or not, while simultaneously analyzing placement patterns across different academic branches. By examining factors such as CGPA, communication skills, projects completed, internship experience, and academic branch, this project develops a predictive model to estimate placement probability and identifies branches with higher placement success rates.

---

Project Objectives

• Predict whether a student will receive a placement offer based on academic and skill-related features.
• Analyze placement patterns across different engineering and technology branches.
• Handle severely corrupted, noisy, and inconsistently formatted real-world data through a robust cleaning and imputation pipeline.
• Produce a fully numeric, model-ready dataset for downstream classification tasks.

---

Dataset Description

Property	Details
Source File	Raw_College_student_placement_dataset.csv
Raw Records	43,500 rows
Raw Columns	21 columns
Final Records	40,307 rows
Final Columns	10 columns
Target Variable	Placement (1 = Placed, 0 = Not Placed)

Raw Columns (Original)

Column	Description
College_ID	Unique college identifier
IQ	Student IQ score
Prev_Sem_Result	Previous semester result
CGPA	Cumulative Grade Point Average
Academic_Performance	Academic performance score (0–10)
Internship_Experience	Whether the student had internship experience (Yes/No)
Extra_Curricular_Score	Score for extracurricular activities (0–10)
Communication_Skills	Communication skills rating (0–10)
Projects_Completed	Number of projects completed
Placement	Target — whether student was placed
college_name	Name of the college
sem_1_result to sem_8_result	Semester-wise results
branch	Academic branch of the student
KT	Number of backlogs (KT = Keep Term)

Final Columns (After Processing)

Column	Type	Description
IQ	Float	Student IQ score, cleaned and clipped to [50, 200]
Academic_Performance	Float	Academic performance score (0–10)
Internship_Experience	Integer	Binary — 1 = Yes, 0 = No
Extra_Curricular_Score	Float	Extracurricular score (0–10)
Communication_Skills	Float	Communication rating (0–10)
Projects_Completed	Integer	Count of projects completed
branch	String / Encoded	Standardized branch name or encoded integer
KT	Integer	Number of backlogs
CGPA_Final	Float	Reconstructed or original CGPA (0–10)
Placement	Integer	Target variable — 1 = Placed, 0 = Not Placed

---

Project Structure

Placement-Prediction/
│
├── Raw_College_student_placement_dataset.csv   # Original raw dataset
├── cleaned_placement_data.csv                  # Cleaned and imputed dataset
├── scaled_placement_data.csv                   # Feature-scaled dataset for modelling
├── data_modified.ipynb                         # Data cleaning and preprocessing pipeline
├── eda.ipynb                                   # Exploratory Data Analysis
├── placement_model.ipynb                       # Model training and evaluation
└── README.md                                   # Project documentation

---

Data Cleaning Pipeline

The raw dataset contained severely corrupted values far beyond standard missing data. Each column required a tailored cleaning strategy.

Issues Identified Per Column

Column	Issues Found
IQ	Scientific notation (1e309), text values ("one hundred"), salary-like values ("5k"), extreme outliers (999999999, -1e308)
CGPA	Text junk ("zero.zero", "ten_thousand"), separator characters ("-"), values exceeding valid range
sem_1 to sem_8	Text entries ("absent", "fail/pass", "zero.zero"), values outside [0, 10]
Internship_Experience	Mixed case and spacing ("Yes", "y", "YES", "Ye s"), HTML-injected strings
Placement	Inconsistent labels ("placed", "Placde", "NO", "N", "YES", empty string)
branch	Abbreviations ("C.S.E", "cse"), typos ("Compter Scince"), pipe-delimited junk

Cleaning Functions Applied

clean_cgpa(value)

• Converts value to lowercase string.
• Returns NaN for known junk tokens: "zero.zero", "ten_thousand", "?", "nan", "none", "".
• Returns NaN for values ending with "k" (salary-format entries).
• Returns NaN for values outside the range [0, 10] or negative values.

Applied to: CGPA, Extra_Curricular_Score, Communication_Skills, Academic_Performance, Projects_Completed, KT.

clean_cgpa_value(val)

• Extended version with explicit text mapping (e.g., "zero.zero" → 0.0).
• Clips result to [0, 10] range.
• Returns NaN for non-parseable strings.

Applied to: sem_1_result through sem_8_result.

clean_iq(value)

• Converts to lowercase string.
• Returns NaN for text tokens and "k"-suffixed values.
• Converts to float, then clips to valid IQ range [50, 200].
• Rejects infinity-like scientific notation values.

Applied to: IQ.

clean_binary(value)

• Standardizes binary Yes/No columns.
• Maps any string starting with "y" or containing "yes" → 1.
• Maps any string starting with "n", containing "no", or equal to "0" → 0.
• Maps strings starting with "p" (e.g., "placed", "Placde") → 1.
• Returns NaN for all unrecognized values (HTML junk, random strings).

Applied to: Internship_Experience, Placement.

clean_branch(value)

• Keyword-based categorization using lowercase string matching.
• Maps to five standardized branch names.
• Returns NaN for unrecognized or junk values.

Keywords Detected	Standardized Label
"cs", "computer"	Computer Science
"civil"	Civil Engineering
"mech"	Mechanical Engineering
"elect", "ece", "eee"	Electronics Engineering
"it" (without "cs")	Information Technology

Target Column Handling

Rows where Placement was NaN after the initial read were dropped before any further processing, as the target variable is mandatory for supervised learning.

---

Feature Engineering

CGPA Reconstruction

A significant portion of the CGPA column was missing (13.26%). However, many of these students had partial or complete semester result data available. A row-wise mean was computed across all eight semester columns to derive a proxy CGPA.

CGPA_Final = CGPA (if present) else mean(sem_1 to sem_8)

Metric	Before	After
CGPA missing rate	13.26%	0.23%
Coverage	86.74%	99.77%

This domain-aware reconstruction significantly reduced missingness in the most important continuous predictor without introducing external assumptions.

---

Imputation Strategy

After cleaning, the remaining missing rates were as follows:

Column	Missing Rate
IQ	62.41%
Academic_Performance	55.71%
Extra_Curricular_Score	54.54%
Communication_Skills	54.99%
Projects_Completed	56.98%
KT	43.03%
branch	8.93%
Placement	9.75%
CGPA_Final	0.23%
Internship_Experience	8.91%

Method: MICE (Multiple Imputation by Chained Equations)

MICE was selected over simpler alternatives (mean, median, mode) for the following reasons:

• Mean/median imputation on columns with 50–60% missingness would collapse the variance of those features, introducing severe bias.
• MICE treats each column with missing values as a dependent variable and regresses it against all other columns iteratively, producing statistically consistent estimates.
• The iterative process (max_iter=10) ensures convergence across all columns simultaneously.

Estimator Used: BayesianRidge (scikit-learn default for IterativeImputer)

BayesianRidge was preferred over RandomForest for the following reasons:

• Features in this dataset (scores, IQ, CGPA) are approximately linearly correlated.
• RandomForest with n_estimators=100 would provide only marginal accuracy gains while increasing computation time by a factor of 10 to 30 on a 40,000-row dataset.
• BayesianRidge is robust, fast, and well-suited for continuous features with linear relationships.

Configuration:

IterativeImputer(
    max_iter     = 10,
    random_state = 42,
    min_value    = 0
)

Post-Imputation Corrections

Column	Correction Applied
Internship_Experience	Rounded to nearest integer, cast to 0 or 1
Placement	Rounded to nearest integer, cast to 0 or 1
KT	Rounded to nearest integer
Projects_Completed	Rounded to nearest integer
IQ	Clipped to [50, 200], rounded
Academic_Performance	Clipped to [0, 10]
Extra_Curricular_Score	Clipped to [0, 10]
Communication_Skills	Clipped to [0, 10]
CGPA_Final	Clipped to [0, 10]
branch	Rounded encoded float to nearest class index, decoded back to string

---

Encoding and Preprocessing

Branch Encoding

The branch column was encoded using LabelEncoder after standardization. The final mapping is:

Branch Label	Encoded Value
Civil Engineering	0
Computer Science	1
Electronics Engineering	2
Information Technology	3
Mechanical Engineering	4

Encoding choice for downstream modelling should follow these guidelines:

Model Type	Recommended Encoding
Random Forest, XGBoost, Decision Tree	Label Encoding
Logistic Regression, SVM, Neural Networks	One-Hot Encoding

Binary Columns

Both Internship_Experience and Placement were standardized to integer binary values: 1 (Yes / Placed) and 0 (No / Not Placed).

---

Final Dataset Summary

Property	Value
Total rows	40,307
Total columns	10
Null values remaining	0
Target variable	Placement
Output file	cleaned_placement_data.csv

---

Tech Stack

Library	Version	Purpose
Python	3.13	Core language
pandas	Latest	Data loading, manipulation, and export
numpy	Latest	Numerical operations and NaN handling
scikit-learn	Latest	IterativeImputer (MICE), LabelEncoder, model training
matplotlib	Latest	Distribution visualization
seaborn	Latest	Statistical plotting

---

How to Run

Prerequisites

pip install pandas numpy scikit-learn matplotlib seaborn

Steps

1. Clone the repository and navigate to the project directory.
2. Ensure Raw_College_student_placement_dataset.csv is present in the root directory.
3. Run the notebooks in the following order:
   • data_modified.ipynb — Data cleaning and preprocessing
   • eda.ipynb — Exploratory Data Analysis
   • placement_model.ipynb — Model training and evaluation
4. The cleaned output cleaned_placement_data.csv and scaled output scaled_placement_data.csv will be generated in the working directory.

Verify Output

import pandas as pd

df = pd.read_csv('cleaned_placement_data.csv')
print(df.shape)
print(df.isnull().sum())
print(df.head())

---

Results and Output

Upon successful execution of the full pipeline, the following is produced:

• A fully cleaned and imputed DataFrame with zero null values.
• All features in numeric format, ready for direct use in classification models.
• Branch labels standardized to five canonical engineering disciplines.
• Binary columns (Placement, Internship_Experience) encoded as integer 0/1.
• A feature-scaled dataset (scaled_placement_data.csv) for use with distance-sensitive models.
• Exported CSV: cleaned_placement_data.csv

Recommended Next Steps

• Analyze placement rates by branch using groupby aggregations.
• Train and compare classification models: Logistic Regression, Random Forest, XGBoost.
• Evaluate models using: Accuracy, Precision, Recall, F1-Score, ROC-AUC.
• Conduct feature importance analysis to identify the top predictors of placement.

---

Key Decisions and Rationale

Decision	Rationale
Dropped rows where Placement is NaN	Target variable is mandatory for supervised learning
Used CGPA row-mean from semesters	Domain knowledge — CGPA is the average of semester grades
Chose BayesianRidge over RandomForest for MICE	Linear features, large dataset, negligible accuracy difference
Clipped IQ to [50, 200]	Standard psychometric IQ range; values outside are data errors
Clipped scores to [0, 10]	All academic scores in this dataset are on a 10-point scale
Keyword-based branch cleaning	Handles abbreviations, typos, and junk uniformly without manual enumeration
random_state=42	Ensures reproducibility across all runs

---

Limitations

• Columns with 50–60% missing data (IQ, Academic_Performance, etc.) are largely imputed. While MICE is statistically sound, imputed values at this scale should be interpreted with caution.
• The branch cleaning function uses keyword matching. Rare or highly corrupted branch names that do not contain recognizable keywords will be assigned NaN and subsequently imputed.
• The binary encoding of Internship_Experience and Placement using string-start matching may misclassify edge cases where junk strings coincidentally begin with "y" or "n".
• This pipeline assumes the raw data schema remains consistent. A different version of the source CSV may require adjustments to the cleaning functions.

---

Authors

Prepared as part of an academic data science project on student placement prediction.

• Dataset: Raw College Student Placement Dataset
• Notebooks: data_modified.ipynb, eda.ipynb, placement_model.ipynb