Machine Learning Project Ideas for CS Students 2026 — Code-First Guide with Datasets, Pipelines, and Viva Defence

Home › Project Ideas › Machine Learning Project Ideas for CS Students 2026

ML Projects Code-First Guide 🌎 Real Datasets · Full Pipeline

Most ML project guides give you a topic and a model name. This guide gives you the dataset, the full code pipeline from raw data to evaluation, the metric that examiner will ask about, and the specific viva question your project will face — because reporting 94% accuracy and not knowing what F1-score means is the fastest way to lose marks in an ML viva.

🎓 BE · BTech · BCA · MCA · BSc CS 📅 Published May 2026 ⏱ 13 min read

Fig. 1 — Machine Learning Projects for CS Students 2026: real datasets, preprocessing to deployment pipeline, evaluation metrics guide, and the viva question each project will face

◆ Quick Answer

The strongest machine learning project ideas for CS students in 2026 are built around three things: a publicly available dataset with real messiness (missing values, class imbalance, outliers), a documented pipeline from raw data to model evaluation, and evaluation metrics beyond accuracy. Projects that compare two models on the same dataset — explaining why one outperforms the other — consistently score higher than projects that train one model and report a single accuracy number. Every idea in this guide is chosen because it creates that comparison naturally.

Table of Contents

1. Why Most ML Projects Fail in Viva — The Pipeline Problem
2. ML Problem Types — Matching the Right Algorithm to the Right Problem
3. 20 Machine Learning Project Ideas — Dataset, Pipeline, and Viva Question
4. Evaluation Metrics Guide — When to Use What
5. Class Imbalance, Overfitting, and the Three Questions Every Examiner Asks
6. Dataset Comparison Table — Which Dataset to Choose
7. Editorial Opinion — Which ML Project We Actually Recommend
8. Frequently Asked Questions

Reporting 94% accuracy is not a result. It is the beginning of the real question: 94% on what distribution, against what baseline, measured by what metric, and why did that metric matter for this specific problem? Most ML final year projects answer none of these. That is why most fail to impress — not because the model was wrong, but because the evaluation framework was empty.

The model is one decision. The pipeline is the project. Choose the wrong metric for your class distribution and the entire evaluation collapses. Skip baseline comparison and your 91% accuracy means nothing — a majority-class predictor may already achieve 89%. These are not edge cases. They are the first questions any prepared examiner asks, and the ones this guide is built around.

This is the machine learning spoke of the Computer Science Final Year Project Ideas 2026 hub. Every project here is pure CS scope — Python, Jupyter, scikit-learn, and publicly available datasets. No hardware. No IoT. No edge deployment. The focus is code pipelines that can be read, reproduced, and defended line by line.

Before You ChooseWhy Most ML Projects Fail in Viva — The Pipeline Problem

The accuracy trap works like this: train a model, get a high number, build the report around that number. Then the first viva question arrives — "what is your false positive rate, and what does it cost a real user of this system?" — and the number suddenly means nothing, because the project was never designed around consequences.

Dataset ignorance compounds it. Downloading a Kaggle CSV without knowing its class distribution, its missing value rate, or its majority-class baseline is not preprocessing — it is guessing. A 94% accuracy on a dataset that is 94% class A is not a model. It is the absence of one.

⚠ The Two ML Failure Patterns

Accuracy is reporting. Metric justification is engineering. Pattern 1 — Accuracy trap: one headline number, no precision, no recall, no confusion matrix. The metric behind the number is where the real answer lives. Pattern 2 — Dataset ignorance: unknown class distribution, unknown missing value rate, no baseline comparison. A dataset that is 95% class A produces 95% accuracy from a model that predicts nothing. Know your data before you train on it.

The fix is structural, not cosmetic. Treat the pipeline as the deliverable. A documented, justified, step-by-step pipeline from raw data to final metric — with honest limitations at each stage — holds up under questioning regardless of the final accuracy number. The number is the conclusion. The pipeline is the argument. For the full list of questions your viva will include, the 50 Most Common Engineering Project Viva Questions guide covers examiner patterns across all domains.

FrameworkML Problem Types — Matching the Right Algorithm to the Right Problem

The first decision in any ML project is identifying the correct problem type. Students who choose the wrong algorithm for their problem type — using regression on a classification problem, or clustering when supervised labels exist — create a viva weakness that no amount of accuracy can recover from.

Table 1 — ML Problem Types: When to Use Each, Right Algorithm, Evaluation Metric, and Common Mistake to Avoid

Problem Type	Use When	Right Algorithms (2026)	Primary Metric	Secondary Metric	Common Mistake
Binary Classification	Output is one of two classes (spam/not spam, churn/no churn)	Logistic Regression, Random Forest, XGBoost, SVM	F1-Score (imbalanced) · Accuracy (balanced)	ROC-AUC, Precision-Recall curve	Using accuracy on imbalanced datasets — majority class baseline destroys meaning
Multi-Class Classification	Output is one of 3+ categories (disease type, product category)	Random Forest, XGBoost, Neural Network, KNN	Weighted F1-Score · Macro F1 for class fairness	Confusion matrix per class, per-class recall	Reporting only overall accuracy — hides poor performance on minority classes
Regression	Output is a continuous number (price, temperature, score)	Linear Regression, Ridge, XGBoost Regressor, LSTM	RMSE (penalises large errors) · MAE (robust to outliers)	R-squared, residual plot analysis	Using only R-squared — high R² does not mean good predictions, just correlation
Clustering	No labels exist — finding natural groups in data	K-Means, DBSCAN, Hierarchical Clustering	Silhouette Score · Davies-Bouldin Index	Elbow method for K selection, cluster visualisation	Choosing K arbitrarily without using elbow method or silhouette analysis
Anomaly Detection	Finding rare, unusual events in predominantly normal data	Isolation Forest, One-Class SVM, Autoencoder	Precision and Recall at defined threshold	False positive rate, detection latency	Using standard classifiers trained on imbalanced data instead of anomaly-specific methods
NLP Classification	Text input needs to be categorised (sentiment, topic, intent)	TF-IDF + Logistic Regression, DistilBERT fine-tuning	F1-Score per class · Overall weighted F1	Confusion matrix, error analysis on misclassified samples	Using BERT for a task where TF-IDF + Logistic Regression achieves comparable F1 with 100x less compute

✓ Algorithm Selection Rule

Start simple. If Logistic Regression achieves 88% F1 and XGBoost achieves 91% F1, the research question is not which number is higher — it is why does XGBoost's non-linearity produce a 3% gain on this specific dataset? That explanation, documented in your report, is worth more marks than the 3% itself. Complexity without justification is noise. Justified complexity is engineering.

Core Section20 Machine Learning Project Ideas — Dataset, Pipeline, and Viva Question

Every idea below includes the publicly available dataset, the full pipeline steps that project requires, the model comparison that makes it academically defensible, and the specific viva question that project will face. The dataset column is sourced from UCI ML Repository, Kaggle, or Hugging Face — all examiner-accepted sources that require no institutional access.

🤖 Classification Projects — 8 Ideas Stack: Python · scikit-learn · pandas · Jupyter · GitHub

Student performance prediction — Random Forest vs Logistic Regression

Dataset: Student Performance Dataset — UCI ML Repository · 649 records, 33 features, binary pass/fail label · Pipeline: Missing value audit → feature encoding → train/test split 80/20 → model comparison → SHAP feature importance · Comparison: Which features drive failure prediction most strongly?

Viva Q: "Your model predicts student failure. What is the false negative rate — and what does it mean in practice when your model misses a student who is actually at risk?"

Spam email classification — precision vs recall threshold analysis

Dataset: Enron Email Dataset — Kaggle · 500K+ emails, labelled spam/ham · Pipeline: Text cleaning → TF-IDF vectorisation → Naive Bayes vs SVM → threshold tuning → precision-recall curve · Comparison: At what threshold does precision drop below acceptable for a real email system?

Viva Q: "A false positive here sends a legitimate email to spam. How did you tune your classification threshold to minimise that specific error — and what precision did you achieve at your chosen threshold?"

Customer churn prediction — SMOTE vs no SMOTE on imbalanced data

Dataset: Telco Customer Churn — Kaggle · 7,043 records, 21 features, 14% churn rate · Pipeline: EDA → class imbalance analysis → SMOTE oversampling → XGBoost with and without SMOTE → ROC-AUC comparison · Comparison: Does SMOTE meaningfully improve recall for the minority (churn) class?

Viva Q: "Your dataset has 14% churn rate. What does a model that always predicts no-churn achieve in accuracy — and by how much does your model beat that baseline, specifically on the churn class?"

Fake news detection — TF-IDF vs BERT feature comparison

Dataset: Fake and Real News Dataset — Kaggle · 44,000+ articles, binary label · Pipeline: Text preprocessing → TF-IDF + Logistic Regression baseline → DistilBERT fine-tuning → F1 per class comparison · Comparison: Does fine-tuned BERT justify its compute cost over TF-IDF on this dataset size?

Viva Q: "How does your model perform on satirical articles — and what linguistic features in satire cause misclassification in both your TF-IDF and BERT models?"

Diabetes readmission prediction — feature selection impact analysis

Dataset: Diabetes 130-US Hospitals Dataset — UCI ML Repository · 101,766 records, 50 features · Pipeline: Missing value handling → categorical encoding → feature selection (RFE) → Logistic Regression vs Neural Network → recall on high-risk class · Comparison: Does removing low-importance features improve or hurt recall on critical patients?

Viva Q: "What is your model's recall for high-risk readmission patients — and what threshold did you set, and why does that threshold matter more than overall accuracy for this use case?"

Phishing URL detection — URL features vs full content features

Dataset: PhiUSIIL Phishing URL Dataset — UCI ML Repository · 235,795 URLs, 56 features · Pipeline: URL feature extraction (length, special chars, HTTPS flag) → content features → Random Forest comparison of feature sets · Comparison: How much does adding content features improve detection over URL-only features?

Viva Q: "How does your detector perform on newly registered legitimate domains that share structural features with phishing URLs — and what causes that false positive?"

Resume screening — bias detection in automated shortlisting

Dataset: Synthetic resume dataset (generate with Faker) · Pipeline: Feature engineering → Random Forest → demographic parity analysis → equalised odds measurement · Comparison: Does the model perform equally across demographic groups, and what features drive disparity?

Viva Q: "How did you test whether your model introduces demographic bias in shortlisting — and what fairness metric did you use, and why did you choose that metric over others?"

Plant disease identification — MobileNet vs EfficientNet transfer learning

Dataset: PlantVillage Dataset — Kaggle · 54,306 images, 38 disease classes · Pipeline: Image augmentation → transfer learning with frozen base → fine-tuning → per-class accuracy comparison · Comparison: Which architecture achieves better per-class accuracy with less training time?

Viva Q: "Which disease class has the lowest recall in your model — and what visual similarity between that class and another class causes the confusion?"

📈 Regression and Prediction Projects — 6 Ideas Stack: Python · scikit-learn · XGBoost · pandas · SHAP

House price prediction — feature importance with SHAP analysis

Dataset: Ames Housing Dataset — Kaggle · 2,930 records, 82 features · Pipeline: Outlier removal → feature engineering (age, area ratios) → Linear Regression vs XGBoost → SHAP value analysis · Comparison: Does XGBoost's non-linearity provide meaningful improvement over Linear Regression for this dataset?

Viva Q: "Which feature contributes most to price prediction according to SHAP — and does that finding match what domain experts (estate agents) would expect, or does it suggest a data artefact?"

Road accident severity prediction — environmental vs road condition features

Dataset: UK Road Safety Data — data.gov.uk · 1.8M+ accident records · Pipeline: DateTime feature extraction → weather encoding → Decision Tree vs Gradient Boosting → feature importance by category · Comparison: Do environmental features outperform road features in severity prediction?

Viva Q: "Which single feature has the highest influence on accident severity prediction — and does that align with what road safety research would predict?"

Stock price direction prediction — technical indicator feature engineering

Dataset: Yahoo Finance API (any liquid stock, 5 years) · Pipeline: Technical indicator calculation (RSI, MACD, Bollinger) → lag feature creation → LSTM vs Random Forest → direction accuracy vs chance baseline · Comparison: Does LSTM capture temporal dependency better than Random Forest on this stock's pattern?

Viva Q: "Your model predicts direction, not price. What is the chance baseline accuracy for binary direction prediction on this stock — and how did you ensure your model is beating chance rather than just the majority class?"

Energy consumption prediction — time-series feature extraction

Dataset: Individual Household Electric Power Consumption — UCI ML Repository · 2M+ minute-level readings · Pipeline: Resampling to hourly → lag features → rolling mean/std → Linear Regression vs XGBoost → RMSE by time-of-day · Comparison: Does adding lag features meaningfully reduce RMSE?

Viva Q: "At what lag length do additional lag features stop improving your model's RMSE — and what does that tell you about the temporal dependency structure of household energy use?"

E-commerce product return prediction — category-level model comparison

Dataset: E-commerce returns dataset (Kaggle) · Pipeline: Category encoding → purchase behaviour features → XGBoost → per-category recall analysis · Comparison: Does a single global model outperform category-specific models for return prediction?

Viva Q: "Which product category has the highest return prediction error in your model — and does your feature importance reflect the actual return drivers for that category?"

Air quality index prediction — multi-pollutant feature interaction

Dataset: Beijing PM2.5 Data — UCI ML Repository · 43,824 hourly records · Pipeline: Missing value interpolation → pollutant correlation analysis → Ridge Regression vs XGBoost → residual analysis by season · Comparison: Does feature interaction (PM2.5 × humidity) improve predictions over independent features?

Viva Q: "Your model has higher RMSE in winter months than summer months — what seasonal factor causes that pattern, and did you attempt to correct for it?"

💬 NLP and Recommendation Projects — 6 Ideas Stack: Python · scikit-learn · NLTK · HuggingFace Transformers · Surprise library

Twitter sentiment analysis — VADER vs fine-tuned DistilBERT by topic

Dataset: Sentiment140 — Kaggle · 1.6M tweets, binary sentiment label · Pipeline: Cleaning (URLs, mentions, emojis) → VADER rule-based baseline → DistilBERT fine-tuning → accuracy comparison by topic category · Comparison: On which topic does VADER perform closest to BERT?

Viva Q: "How does accuracy differ between technology tweets and political tweets in your test set — and what property of political language makes it harder for both models to classify correctly?"

Movie recommendation — collaborative filtering vs content-based comparison

Dataset: MovieLens 100K — GroupLens · 100,000 ratings, 943 users, 1,682 movies · Pipeline: User-item matrix → user-based CF → TF-IDF content filtering on genres and tags → RMSE and precision@K comparison · Comparison: For which user type does collaborative filtering outperform content-based filtering?

Viva Q: "How does your system handle a new user with no viewing history — what does it recommend, and what does that cold start behaviour tell you about the fundamental limitation of collaborative filtering?"

Mental health condition classification from Reddit posts

Dataset: Reddit Mental Health Dataset (Kaggle, labelled by subreddit) · Pipeline: Text preprocessing → TF-IDF + SVM → per-class F1 → ethical documentation section · Comparison: Which mental health condition is hardest to classify, and what linguistic overlap causes misclassification?

Viva Q: "How did you handle the ethical implications of building a classifier that labels mental health conditions from text — and what limitations did you explicitly document in your report?"

Question answering system on a domain-specific PDF corpus

Dataset: Your institution's academic rulebook or any domain PDF · Pipeline: PDF extraction → chunking → DistilBERT fine-tuning on SQuAD → F1 and Exact Match on custom test questions · Comparison: How does fine-tuned model compare to zero-shot QA on domain-specific questions?

Viva Q: "How does your system handle a question whose answer spans two paragraphs — and what is the F1 drop for multi-span questions compared to single-span questions in your test set?"

Network traffic classification — known vs unknown protocol detection

Dataset: CICIDS2017 — Canadian Institute for Cybersecurity · 2.8M+ network flow records, 15 traffic types · Pipeline: Flow feature extraction → normalisation → Random Forest → per-traffic-type precision and recall · Comparison: Which traffic type has the highest false positive rate?

Viva Q: "How does your classifier perform on traffic types it has never seen during training — and how did you design your test set to specifically measure that generalisation gap?"

Handwritten character recognition — full model vs compressed mobile version

Dataset: EMNIST Extended — Kaggle · 814,255 samples, 47 balanced classes · Pipeline: CNN training → TensorFlow Lite quantisation → accuracy vs model size comparison · Comparison: How much accuracy is lost by compressing the model — and is that trade-off worth the size reduction?

Viva Q: "Your compressed model is 4x smaller than the full model. How much accuracy did you lose per character class — and which classes were most affected by quantisation?"

Metrics GuideEvaluation Metrics — When to Use What and How to Defend Your Choice

The single most common viva failure in ML projects is metric confusion. A student reports accuracy, the examiner asks about F1-score, and the student cannot explain the difference. This section eliminates that failure. Use this table to select your primary metric before building — not after — and document your selection rationale in your report.

Table 2 — ML Evaluation Metrics Guide: When to Use Each Metric, What It Measures, and the Examiner Question It Answers

Metric	Use When	What It Measures	Fails When	Examiner Question It Answers
Accuracy	Classes are balanced (roughly equal distribution)	% of all predictions that are correct	Dataset is imbalanced — majority class baseline inflates accuracy to uselessness	"What percentage of all predictions did your model get right?"
Precision	False positives are costly (spam filter, fraud detection)	Of all positive predictions, how many were actually positive?	Missing true positives matters more than generating false alarms	"When your model predicts positive, how often is it correct?"
Recall	False negatives are costly (disease detection, fraud, churn)	Of all actual positives, how many did the model find?	False alarms are highly disruptive to the system or user	"What percentage of actual positive cases did your model catch?"
F1-Score	Imbalanced classes AND both precision and recall matter	Harmonic mean of precision and recall — penalises extreme imbalance between them	One metric genuinely matters more than the other for the use case	"How did you balance catching real cases against generating false alarms?"
ROC-AUC	Comparing models regardless of classification threshold	Model's ability to distinguish classes across all possible thresholds	Highly imbalanced datasets — use Precision-Recall AUC instead	"Which model is better at separating the two classes, independent of your chosen threshold?"
RMSE	Regression — large errors should be penalised more heavily	Root mean squared error — squares errors before averaging, amplifying large ones	Outliers dominate the metric — use MAE for robust evaluation	"How large are your prediction errors — and does your model make a few large mistakes or many small ones?"
MAE	Regression — all errors should be treated equally	Mean absolute error — average magnitude of prediction errors	Large errors are disproportionately important in the use case	"On average, by how much does your prediction miss the actual value?"

Accuracy is reporting. Metric justification is engineering. Choose your primary metric before training — not after seeing the results. Selecting F1-score because the dataset is imbalanced and both error types carry real cost is a documented design decision. Switching to F1-score after the viva starts because the examiner mentioned imbalance is a visible gap in evaluation planning. The sequence matters as much as the metric.

Viva PreparationClass Imbalance, Overfitting, and the Three Questions Every ML Viva Includes

Three question categories appear in every ML viva regardless of domain, dataset, or model. Prepare for these three and the rest of the viva is about specifics, not fundamentals.

◆ Category 1 — Baseline Comparison

"What does a trivial predictor achieve on your dataset?" For a dataset with 85% class A and 15% class B, always-predict-A achieves 85% accuracy. A model at 87% is not a contribution — it is noise above a useless baseline. The fix: calculate and report majority-class baseline accuracy before any model result. Frame your model's performance as the gap above that baseline. That gap is your actual contribution.

⚠ Category 2 — Overfitting Evidence

98% training accuracy with 74% test accuracy is not a strong model — it is a memorised one. The gap between training and test performance is the overfitting signal. The fix: report 5-fold cross-validation scores with mean and standard deviation. A model that scores 84% ± 1.2% across folds is stable. One that scores 84% ± 8.6% is not — and that variance must be explained, not hidden.

✓ Category 3 — Feature Importance Rationale

Knowing which feature ranked highest is not the answer. Explaining why it ranked highest — and whether that ranking reflects a real signal or a data artefact — is. Use SHAP values. If the top feature makes domain sense, say so explicitly. If it does not, that is the most interesting finding in your project. "Feature X ranked first statistically but is likely a proxy for data collection bias rather than a genuine predictor" is a distinction-level answer.

→

CS Project Hub Computer Science Final Year Project Ideas 2026 — 100+ Ideas Across 6 Domains with Tools, Scope, and Viva Strategy Cybersecurity, databases, mobile apps, web development, and mini projects — each with tools, scope, and viva strategy. The full CS cluster.

---

How to Use This GuideThree Decisions Before You Write a Line of Code

Twenty project ideas. Three tables. Three viva question categories. The infrastructure is here. What is not here is the decision only you can make: which project forces a question you are genuinely capable of answering.

First: Use Table 1 to lock your problem type before selecting a dataset. Problem type determines algorithm family. Algorithm family determines evaluation metric. Getting this wrong at the start creates a cascade of wrong decisions downstream.

Second: Read the viva question for each project in Section 2. When you find one where the answer is already in your knowledge — or where the gap between what you know and what you need to know is genuinely interesting to close — that is your project.

Third: Choose your primary metric from Table 2 before training anything. Write one sentence in your report introduction explaining why. That sentence is the difference between a project that was built and a project that was designed. Before finalising scope, run it through the Feasibility and Measurement Framework to confirm your chosen dataset and pipeline fit your available time.

The closing principle: Accuracy is what your model achieved. Metric justification is what you understood. Pipeline documentation is what you can defend. Only one of these three actually determines your viva outcome — and it is not the first one.

---

Section 07Frequently Asked Questions

Which machine learning project is best for a CS final year student in 2026?

Best is defined by pipeline defensibility, not topic prestige. Customer churn prediction, spam classification, and student performance prediction score consistently well because the evaluation framework is structurally forced — class imbalance demands metric justification, baseline comparison is unavoidable, and feature importance has domain-verifiable answers. Choose the project where the viva question in Section 2 is one you can answer without hesitation.

What datasets should CS students use for machine learning final year projects?

UCI ML Repository, Kaggle, and Hugging Face Datasets — all free, documented, and examiner-accepted. Every idea in this guide references a named dataset with record count and class distribution. Avoid fully synthetic data unless bias or fairness is the research question — real-world messiness (missing values, class imbalance, noise) is what makes a preprocessing pipeline worth documenting.

Do CS students need to use deep learning for a machine learning final year project?

No — and choosing deep learning without justification actively hurts the viva. If XGBoost achieves 89% F1 and a neural network achieves 91% F1, the project question is not "which number is higher" — it is "why does the additional complexity produce only 2% gain, and is that gain worth the interpretability cost?" That question requires understanding, not just implementation. scikit-learn with SHAP analysis consistently produces more defensible projects than unexplained deep learning.

What evaluation metrics do examiners expect in a machine learning project?

Accuracy alone fails on any imbalanced dataset and is insufficient even on balanced ones. Classification minimum: precision, recall, F1-score, confusion matrix, and your threshold rationale for imbalanced classes. Regression minimum: MAE, RMSE, R-squared, and a residual plot. The metric you report is less important than documenting why you chose it — that justification is the engineering decision. Reporting all metrics without explaining which one guided model selection is still a weak evaluation strategy.

---

Complete CS Project & Viva Guide Series

• Computer Science Final Year Project Ideas 2026 — 100+ Ideas Across 6 Domains
• Web Development Final Year Project Ideas 2026 — System Design, Security, and Deployment
• Cybersecurity Final Year Project Ideas 2026 — Ethical Scope, Tools, and What Examiners Check
• Database and Backend Project Ideas 2026 — Schema Design, Performance Testing, and Examiner Criteria
• Mobile App Final Year Project Ideas 2026 — Flutter vs React Native and User Testing Methods
• CS Mini Project Ideas 2026 — 50+ Single-Feature Builds with Measurable Outcomes
• AI Based Engineering Project Ideas 2026 — Intelligent Systems, Datasets, and Deployment
• 200+ Final Year Engineering Project Ideas (2026) — All Engineering Branches
• The Complete Guide to Engineering Project Viva — Global Strategy for Final Year Students
• 50 Most Common Engineering Project Viva Questions and How to Answer Them
• Feasibility and Measurement Framework for Engineering Projects
• How to Write a Methodology Chapter for Engineering Projects (2026 Guide)

Dataset GuideDataset Comparison Table — Which Dataset to Choose for Your ML Project

Every dataset has trade-offs. A large dataset with severe class imbalance requires different handling than a small balanced dataset. A dataset with missing values creates preprocessing decisions that a clean dataset does not. This table maps the most commonly used datasets against the properties that matter most for a final year project — so you choose based on what your pipeline can handle, not what looks most impressive.

Table 3 — ML Dataset Comparison: Size, Balance, Missingness, Difficulty, and Best Project Fit for CS Final Year Students

Dataset	Source	Size	Class Balance	Missing Values	Difficulty	Best Project Fit
Telco Customer Churn	Kaggle	7,043 rows · 21 cols	Imbalanced — 73% no-churn · 27% churn	Minimal — 11 rows only	Medium	Churn prediction · SMOTE comparison · threshold analysis
Student Performance	UCI	649 rows · 33 cols	Reasonably balanced pass/fail	None	Low-Medium	Classification · feature importance · beginner-friendly pipeline
Ames Housing	Kaggle	2,930 rows · 82 cols	N/A — Regression	Moderate — requires imputation strategy	Medium	House price regression · SHAP feature importance · Linear vs XGBoost
Diabetes 130-US Hospitals	UCI	101,766 rows · 50 cols	Imbalanced — readmission is minority class	Significant — multiple columns with '?' values	High	Medical prediction · feature selection · recall-focused evaluation
Sentiment140	Kaggle	1.6M tweets · 6 cols	Balanced — 50/50 positive/negative	None — but noisy text	Medium	Sentiment classification · VADER vs BERT · NLP pipeline
MovieLens 100K	GroupLens	100K ratings · 943 users · 1,682 movies	N/A — Recommendation	Sparse matrix — cold start problem	Medium	Recommendation system · CF vs content-based · cold start analysis
PlantVillage	Kaggle	54,306 images · 38 classes	Mildly imbalanced across 38 classes	None — clean image dataset	Medium-High	Image classification · transfer learning · per-class accuracy analysis
CICIDS2017	CIC	2.8M+ flow records · 15 traffic types	Severely imbalanced — benign traffic dominates	Some — requires flow feature cleaning	High	Network classification · anomaly detection · cybersecurity ML
UCI Household Power	UCI	2M+ minute readings · 9 cols	N/A — Time-series regression	~1.25% missing — requires interpolation	Medium	Energy prediction · time-series features · lag analysis

✓ Dataset Selection Rule for Final Year Students

Start with Low-Medium difficulty datasets if this is your first ML project — Student Performance or Telco Churn. Both have clean data, clear problem types, and documented baselines from previous Kaggle submissions you can reference to validate your results. High-difficulty datasets like Diabetes 130-US or CICIDS2017 are strong choices only if your preprocessing pipeline is already solid and you have time to handle their specific data quality issues properly.

Editorial OpinionWhich ML Project We Actually Recommend in 2026 — And Why

This is based on patterns from real project reports and viva transcripts — not from what looks good in a project title. Here is the honest version.

Top recommendation: Customer churn prediction with SMOTE comparison. The Telco dataset is clean enough to build fast, imbalanced enough to require genuine handling, and the SMOTE vs no-SMOTE comparison creates a natural, examinable research question. The baseline comparison is structurally unavoidable — 14% churn rate means a majority-class predictor gets 86% accuracy without learning anything. That forces the right evaluation conversation automatically.

Second: House price regression with SHAP analysis. Regression is underrepresented in CS final years because classification sounds more technical. It is not. A project that produces SHAP plots, explains why square footage outweighs neighbourhood in this specific dataset, and honestly discusses where the model breaks down demonstrates exactly the domain-aware thinking that separates distinction from pass. The Ames dataset has 82 features — feature selection alone is a defensible research question.

What to avoid: Stock price prediction. The problem is real. The viva outcomes are consistently poor. The core failure: most projects cannot distinguish between a model that learned a signal and a model that memorised noise. Cross-validation on time-series data requires temporal splitting — not random splitting. Leakage from future data into training is the single most common error, and it is invisible in the training metrics. If you choose this, your methodology section needs to address data leakage explicitly and demonstrate temporal validation. Most do not.

The pattern: Projects chosen for how they sound in conversation — deep learning, GPT-based QA, stock prediction — score lower on average than projects chosen for the quality of the question they answer. A clean pipeline on a boring dataset beats a confused methodology on an exciting one. Every time.

PR

Projectium Research Editorial Team

Project Guidance · Viva Strategy · CS Education

The Projectium Research editorial team reviews final year project reports, viva transcripts, and examiner feedback across CS, engineering, and applied science programmes globally. Our guides are built from patterns observed in real examination outcomes — not from course syllabi or textbook recommendations. Every viva question in this guide has been asked by a real examiner to a real student.

🌐 projectiumresearch.com 📂 All Project Guides

Pipeline Built — Now Defend It

Every ML project will face a viva. The Complete Viva Guide prepares you for every question — from metric selection rationale to feature importance defence, across all examination formats worldwide.

Read the Viva Guide →