Meet PlacementPredict: a working ML service

Intuition

Most courses begin with equations and reach a working system at the end. This one runs the other way. PlacementPredict already works on day one: give it a student's profile, and it returns the chance of that student being placed. All term, we open it up, one layer at a time, until you can build and run the whole thing yourself.

All the programs use the course dataset of 50,000 students, placement_predict_50k. To keep the page quick, most take a random sample of 5,000 of them. You can read the small file that loads the data.

How it works

A prediction service has a simple contract:

Request:   {"CGPA": 8.2, "CodingTestScore": 74, "Internships": 1, ...}
Response:  {"placed_probability": 0.81, "prediction": "Placed"}

Three kinds of learning differ in what the data looks like:

Supervised learning divides again:

What PlacementPredict is not. It does not recommend jobs, since that is a ranking task. It does not write résumés, which is generative AI. And it does not decide anyone's future: it is an estimate based on patterns in past data, and it can be wrong.

Run it: use the service

This builds a small version of the service and sends it requests. The first run loads several libraries, which can take a minute or two. Later runs on the same page take seconds.

import json
import pandas as pd
from sklearn.linear_model import LogisticRegression
from sklearn.pipeline import make_pipeline
from placement_predict import FEATURES, load_data, preprocessor, train_val_test

df = load_data()
X_train, X_val, X_test, y_train, y_val, y_test = train_val_test(df)
model = make_pipeline(preprocessor(), LogisticRegression(max_iter=1000)).fit(X_train, y_train)

def predict(request):
    """What the service does for each request."""
    row = pd.DataFrame([request]).reindex(columns=FEATURES)
    p = float(model.predict_proba(row)[0, 1])
    return {"placed_probability": round(p, 3), "prediction": "Placed" if p >= 0.5 else "Not placed"}

strong = {"CollegeTier": "Tier1", "Stream": "CS", "Specialisation": "AI", "Hostel": "Yes", "HistoryOfBacklogs": "No",
          "CGPA": 8.6, "AttendancePercent": 88, "Internships": 2, "Projects": 4, "Workshops": 2, "Certifications": 3,
          "Publications": 1, "AptitudeTestScore": 78, "SoftSkillsRating": 3.8, "CodingTestScore": 82,
          "MockInterviewScore": 72, "ExtraCurricular": 1}
weak = {"CollegeTier": "Tier3", "Stream": "Civil", "Specialisation": "Embedded", "Hostel": "No", "HistoryOfBacklogs": "Yes",
        "CGPA": 5.4, "AttendancePercent": 58, "Internships": 0, "Projects": 1, "Workshops": 0, "Certifications": 0,
        "Publications": 0, "AptitudeTestScore": 45, "SoftSkillsRating": 1.8, "CodingTestScore": 30,
        "MockInterviewScore": 30, "ExtraCurricular": 0}

print("Strong profile:", json.dumps(predict(strong)))
print("Weak profile:  ", json.dumps(predict(weak)))
print("Missing fields:", json.dumps(predict({"CGPA": 7.5, "Stream": "ECE", "CollegeTier": "Tier2"})))
print(f"\nOverall: {len(df)} students, {df['PlacementStatus'].mean():.0%} placed")
print(f"Accuracy on students the model has not seen: {model.score(X_val, y_val):.3f}")

Try this: press Edit and raise only the weak student's CodingTestScore. How much does the probability move?

Common mistakes

• Reading a probability as a verdict. 0.81 is an estimate from past patterns, not a guarantee.
• Assuming "AI" means deep learning. Much real-world tabular ML uses linear models and trees.
• Calling every prediction task classification. If the output is a number, it is regression.

Problems

1. Classify each as binary, multi-class, regression or clustering: (a) expected salary; (b) CGPA tier; (c) placed or not; (d) grouping students with no labels.
2. Change three fields of the strong profile so that it is predicted "Not placed".
3. The program handles a request with missing fields. Which part of the pipeline makes that possible? Look at Topic 8.
4. Write the request and response for a model that predicts next semester's CGPA.
5. Name one harm that could come from using this service to decide who may sit for a placement drive.