BPSC (Ministry of Food), Network/Website Manager, Exam: 21.05.2025 (ICT)প্রশ্ন ও সমাধান

a) SSD (Solid State Drive): a storage device that stores data electronically in NAND flash memory cells instead of on spinning magnetic platters (as in an HDD).

Working principle:

• No moving parts - data is stored/retrieved purely electronically via a memory controller, unlike an HDD's mechanical read/write head.
• Reading: the controller locates the requested data in flash cells and returns it almost instantly.
• Writing: data is written in fixed-size blocks/pages; if the target block already holds data, it must first be erased before being rewritten (unlike an HDD, which overwrites in place).
• Wear-leveling: the controller spreads writes evenly across all cells, since flash cells degrade after a limited number of write/erase cycles.

Advantages over HDD: much faster boot/access time, silent (no moving parts), lower power draw, and higher shock resistance.

b) CPU vs GPU:

Basis	CPU	GPU
Full form	Central Processing Unit	Graphics Processing Unit
Purpose	General-purpose computation	Specialized parallel computation (graphics, ML)
Core count	Few, powerful cores	Thousands of simpler cores
Task style	Sequential, complex, branch-heavy tasks	Highly parallel, data-uniform tasks (SIMD)
Best for	OS control, single-threaded logic	Rendering, matrix/vector math, deep learning
Memory	Large cache, low latency	High bandwidth, optimized for throughput

Exam takeaway: SSD trades mechanical delay for electronic access (fast, no moving parts, needs wear-leveling); CPU optimizes for latency on complex sequential tasks while GPU optimizes for throughput on massively parallel tasks.

a) Data Mining is the process of discovering meaningful patterns, correlations, and trends in large datasets using techniques from statistics, machine learning, and database systems, turning raw data into actionable knowledge (e.g. market-basket analysis, fraud detection).

Supervised classification: the model is trained on data with known class labels (e.g. emails labeled "spam"/"not spam") and learns to assign labels to new, unseen data.
Unsupervised classification (clustering): the model groups data by similarity without predefined labels (e.g. segmenting customers into spending clusters without knowing group names in advance).

b) Validating a machine learning model's reliability:

• Train-test split - evaluate on data the model never saw during training.
• K-fold cross-validation - rotate the held-out fold across the dataset for a more stable performance estimate.
• Appropriate metrics - Accuracy/Precision/Recall/F1/ROC-AUC for classification; RMSE/MAE/R-squared for regression.
• Overfitting/underfitting check - compare training vs test performance (a large gap indicates overfitting).
• Baseline comparison - ensure the model beats a trivial baseline (e.g. majority-class predictor).
• Hyperparameter tuning - grid/random search to find optimal settings without leaking test data.
• Avoid data leakage - ensure test data never influences training (feature scaling/encoding fit only on the train set).
• Real-world validation - test on fresh, real production data before full deployment.

Exam takeaway: Data mining extracts patterns from data; validation must always be done on unseen data (cross-validation) using metrics matched to the task, with a deliberate check for overfitting.