The biggest beginner mistake in machine learning is not a hard algorithm. It is testing a model in a way that makes it look better than it really is. So before you pick machine learning classes near you, check that they teach you to catch the five mistakes below.
Training a model takes three lines of Python. Knowing whether it works takes much longer. Most of the mistakes here come from that gap.
1. Starting with the model, not the data
Many beginners want to train a model in the first week. But a model can only be as good as the table you give it. Before any model runs, someone has to join sales from two files, remove duplicate rows and fix the dates.
In Nepal, that often means turning Bikram Sambat dates into normal dates before Python can even sort them. No algorithm does that for you.
Do this instead: learn pandas and SQL first. In our Data Science and Machine Learning course, machine learning starts in Week 9. The weeks before it go to Python, pandas, SQL, statistics and Power BI. We chose that order because Excel and SQL are the skills Kathmandu data jobs name most.
2. Trusting accuracy on its own
Accuracy is the share of predictions that were right. It sounds like the only number you need. It is not.
Google's Machine Learning Crash Course gives the classic example. If only 1 percent of cases are positive, a model that always says "no" scores 99 percent accuracy. It is also useless, because it never finds a single positive case.
Think of a model that flags fraud in mobile wallet payments. Fraud is rare, so a model that flags nothing still looks great on accuracy.
Do this instead: read the confusion matrix, the table of what the model got right and wrong for each class. Then pick the number that fits the cost. Google's guide says to use recall when missing a real case costs more, and precision when a positive prediction must be right. Week 9 of our course drills the confusion matrix, precision, recall and F1. Week 10 adds class weights and threshold tuning for rare classes.
3. Having no baseline to beat
A score means nothing alone. Is 82 percent good? It depends on what a model with no skill at all would score.
scikit-learn, the main Python library for classic machine learning, has a tool for exactly this. DummyClassifier ignores the input features completely. With strategy="most_frequent", it always predicts the most common class. The scikit-learn docs call it a simple baseline to compare against real classifiers.
Do this instead: train the dummy first, then your model. If your model cannot beat the dummy, it has learned nothing useful. For a sales forecast, the baseline is "same month last year". Our Week 9 forecast starts from that baseline, and the Week 11 project starts by writing down a do-nothing baseline to beat.
4. Letting test data leak into training
Data leakage means the model sees information it would not have when making a real prediction. scikit-learn's common pitfalls guide explains the most common form. You scale the data or pick features on the full dataset, and only then split it into train and test.
Their example is worth knowing. They use random features and random answers, so an honest model should score about 0.5. Picking features before the split pushed the score to 0.76. The model "learned" from rows it was later tested on.
Do this instead: split first, before any cleaning step that learns from the data. Use fit_transform on the training set and only transform on the test set. Better still, put every step inside a scikit-learn Pipeline, so nothing can learn from the test rows. Week 9 of our course covers leakage and pipelines together.
5. Shuffling data that has a date in it
Sales, prices and bookings all happen over time. A normal random split mixes the months. Then the model trains on next month and gets tested on last month, which can never happen in real work.
scikit-learn's TimeSeriesSplit exists for this. It always trains on the earlier part and tests on the part that comes after it.
Do this instead: split by date. Train on the older months and test on the newest ones. For Nepali business data, think in the fiscal year, Shrawan to Ashadh, because that is how the business reads its numbers. Our Week 9 forecast uses a time-based split so the model never sees the future.
What should you ask machine learning classes near you?
These questions tell you more than any brochure:
- Which week does machine learning start, and what is taught before it?
- Do you teach baselines and data leakage, or only algorithms?
- Who reviews my project, and does anyone re-run it from scratch?
- How many students sit in one batch?
Here is how we answer them. Saarathi Academy in Old Baneshwor, Kathmandu, runs this course for 12 weeks, in person and online. A batch has up to 10 students in the classroom, and a few online. Our mentors still do the job they teach. In Week 12, a classmate re-runs your project from a clean copy, to prove it works on another laptop. The syllabus is public before anyone pays, and the course ends with a certification exam taken online at home.
What to try tonight
Open Google's Machine Learning Crash Course. Google describes it as a free, 15-hour self-study course. If pandas and NumPy are new to you, its prework includes short tutorials on both.
Then take any CSV file, train one model, and train a DummyClassifier next to it. If the two scores are close, you have just caught mistake 3 on your own.
The course page has batch details, and the syllabus shows every week before you decide.
Top comments (0)