A tour of the PlacementPredict codebase and course scope

Intuition

When you join a team, you inherit a running system and learn it by reading it. A good repository makes this easy: each file has one job, and its name tells you what that is.

How it works

placement-predict/
  data/placement_predict_50k.csv raw records (never edited by hand)
  src/
    data_loader.py               read and validate the data
    feature_pipeline.py          turn records into features
    trainer.py                   fit models and save the best
    evaluate.py                  metrics and reports
    app.py                       FastAPI service: /health, /predict
    batch_predict.py             score a whole file
    monitor.py                   drift checks
  models/model.joblib            the trained pipeline
  tests/                         unit and end-to-end tests
  configs/train.yaml             settings, kept out of the code

Where this course stops.

Classical ML is not out of date. On tables of numbers and categories, like ours, tree ensembles are still very hard to beat, as Topic 12 shows.

The first thing to do with any dataset is look at it. In the lab, you write a reusable summary function.

Run it: summarise the dataset

import pandas as pd
import matplotlib.pyplot as plt
from placement_predict import NUMERIC, load_data

def summarise_dataset(df, target=None):
    """Print the shape, type groups, missing values and class balance of any DataFrame."""
    print("Shape:", df.shape)
    for kind, columns in df.columns.groupby(df.dtypes.astype(str)).items():
        print(f"  {kind:8}: {list(columns)}")
    missing = df.isna().sum()
    print("Missing values:", missing[missing > 0].to_dict())
    if target:
        print(f"Class balance of {target}:", df[target].value_counts(normalize=True).round(3).to_dict())

df = load_data()
summarise_dataset(df, target="PlacementStatus")

for column in ["Stream", "CollegeTier", "HistoryOfBacklogs"]:
    print(f"\nPlacement rate by {column}")
    print(df.groupby(column)["PlacementStatus"].mean().sort_values(ascending=False).round(3))
print("\nCorrelation with PlacementStatus")
print(df[NUMERIC + ["PlacementStatus"]].corr()["PlacementStatus"].drop("PlacementStatus").sort_values(ascending=False).round(2))

fig, ax = plt.subplots(1, 3, figsize=(11, 3.1))
ax[0].hist(df["CGPA"], bins=25, color="#0f766e"); ax[0].set(title="CGPA", xlabel="CGPA")
df.groupby("CollegeTier")["PlacementStatus"].mean().sort_values().plot.barh(ax=ax[1], color="#2563eb")
ax[1].set(title="Placement rate by college tier", ylabel="")
df.boxplot(column="CodingTestScore", by="PlacementStatus", ax=ax[2], grid=False)
ax[2].set(title="CodingTestScore by outcome", xlabel="PlacementStatus"); fig.suptitle("")
plt.tight_layout()
plt.show()

Stream makes little difference here, but college tier and backlog history make a large one. Finding this kind of thing is why you look before you model.

Try this: run summarise_dataset with target="CGPA_Tier". Then compare df.groupby("CGPA_Tier")["CGPA"].describe(). Is the tier always what the CGPA says it should be?

Common mistakes

• Editing raw data by hand. Keep it untouched, so results can be repeated.
• Hard-coding settings. Put the learning rate and paths in a config file (Topic 21).
• Thinking a correlation of 0 means no relationship. A curved or threshold effect can hide from it.

Problems

1. Which module changes if the college adds a new attribute, "Language certification"? List all that apply.
2. Lab 1: extend summarise_dataset(df) so that it also lists columns with more than 5% missing values.
3. Lab 2: export a 5-page EDA report as a PDF with PdfPages: missing values, 6 histograms, placement rate by stream, correlation heatmap and box plots. Do this on your own computer.
4. Why is Salary Package left out when calculating correlation with placement?
5. Name a real problem suited to each of the four courses in the table above.