Not sure which algorithm to use? Tell us your problem type and we’ll recommend the best machine learning algorithms for your use case.
[01]Select your problem type
Select the type of machine learning problem you’re trying to solve
STEP 02 / 03 — CONSTRAIN THE SHORTLIST (OPTIONAL)
How much data do you have?
Do decisions need to be explainable?
STEP 03 / 03 — RUN RECOMMENDATION [○]
[03]The method
The selector above gives you a shortlist. This is the reasoning that turns a shortlist into a decision - the six steps that come up in every serious model review.
STEP / 01
Write the sentence "given X, predict Y, measured by Z" before touching a library. Whether Y is a category, a quantity, a ranking, or a structure decides 80% of the shortlist - and Z (the metric the business actually feels) decides the rest. Teams that skip this step tune the wrong loss beautifully.
STEP / 02
Tabular rows with mixed types → gradient-boosted trees first, linear models for the baseline. Images, audio, raw text → pretrained deep models, fine-tuned, never from scratch. A single time series → statistical models before deep ones. Under ~1,000 samples → simpler beats cleverer, almost without exception.
STEP / 03
Interpretability requirements (regulated decisions), latency budgets (real-time serving), training cadence (nightly retrains), and memory ceilings (on-device) eliminate more candidates than any benchmark. An algorithm you cannot deploy, explain, or retrain is not an option - it is a demo.
STEP / 04
Majority class. Yesterday’s value. The mean. A 10-line logistic regression. Every serious team measures lift over a baseline a child could describe, because half of all "ML wins" evaporate against one. Your first real model has one job: beat the dumb baseline honestly.
STEP / 05
Once a sensible model is in place, an hour spent on features, label quality, and leakage checks buys more accuracy than an hour of hyperparameter search - a ratio practitioners put at 10:1. Swap algorithms when the current one plateaus with clean data, not before.
STEP / 06
Random splits lie whenever time, users, or groups structure your data. Split by time for forecasting, by user for personalisation, by hospital for medical models. The gap between naive CV and honest CV is where production surprises live.
[04]Reference matrix
21 ALGORITHMS INDEXED
| Algorithm | Family | Interpretability | Training speed | Data needed | Reach for it when |
|---|---|---|---|---|---|
| Linear Regression | Regression | High | Fast | Small | Simple prediction problems |
| Logistic Regression | Classification | High | Fast | Small | Binary classification |
| Decision Tree | Classification/Regression | High | Fast | Small | Interpretable models |
| Random Forest | Classification/Regression | Medium | Medium | Medium | General-purpose classification |
| Gradient Boosting (XGBoost/LightGBM) | Classification/Regression | Medium | Medium | Medium | Kaggle competitions |
| Support Vector Machine (SVM) | Classification/Regression | Low | Slow | Medium | High-dimensional data |
| K-Nearest Neighbors (KNN) | Classification/Regression | High | Fast | Small | Recommendation systems |
| Naive Bayes | Classification | High | Fast | Small | Text classification |
| K-Means Clustering | Clustering | High | Fast | Medium | Customer segmentation |
| DBSCAN | Clustering | Medium | Medium | Medium | Spatial data |
| Hierarchical Clustering | Clustering | High | Slow | Small | Taxonomy creation |
| Principal Component Analysis (PCA) | Dimensionality Reduction | Medium | Fast | Medium | Data visualization |
| t-SNE | Dimensionality Reduction | Low | Slow | Medium | 2D/3D visualization |
| Neural Network (MLP) | Deep Learning | Low | Slow | Large | Complex patterns |
| Convolutional Neural Network (CNN) | Deep Learning | Low | Slow | Large | Image classification |
| RNN/LSTM | Deep Learning | Low | Slow | Large | Time series |
| Transformer | Deep Learning | Low | Slow | Large | NLP tasks |
| Isolation Forest | Anomaly Detection | Medium | Fast | Medium | Fraud detection |
| Collaborative Filtering | Recommendation | Medium | Medium | Large | E-commerce |
| ARIMA | Time Series | High | Fast | Medium | Economic forecasting |
| Prophet | Time Series | High | Fast | Medium | Business forecasting |
[06]Data profiler
Drop a CSV and get a column-by-column profile plus a recommendation. Everything runs in your browser - the file never leaves your machine.
[05]Honest answers
Q.01 — Q.07
Linear or logistic regression, seriously. They train in seconds, their coefficients teach you what the data is doing, and they set an honest baseline. Then learn decision trees for non-linear intuition, then random forests and gradient boosting - the pair that wins most real tabular problems.
When the input is unstructured - images, audio, video, free text - or when you can fine-tune a large pretrained model. On ordinary tabular business data, gradient-boosted trees match or beat neural networks in most published benchmarks while training faster and needing less tuning.
Under a few thousand rows: regularised linear/logistic regression, naive Bayes for text, or a shallow tree. Strong regularisation plus cross-validation matters more than the algorithm. Deep networks and unconstrained boosting will mostly memorise noise at this scale.
Boosting (XGBoost/LightGBM/CatBoost) is usually a few points more accurate when tuned with early stopping; forests are nearly tuning-free, parallel, and degrade more gracefully. Rule of thumb: forest for a robust v1 you barely babysit, boosting when accuracy is worth tuning care.
Rough orders of magnitude: linear models and naive Bayes work from hundreds of rows; trees and forests want thousands; gradient boosting shines from thousands to millions; training deep networks from scratch wants tens of thousands to millions - though fine-tuning a pretrained model can need only hundreds.
For distance- and gradient-based methods - kNN, SVM, k-means, PCA, neural networks, regularised regression - yes, always. Tree-based models (decision trees, random forests, gradient boosting) are scale-invariant: they split on thresholds, so scaling changes nothing.
Label quality, leakage prevention, honest validation splits, and features that encode domain knowledge. Practitioners consistently report bigger gains from an hour of data work than an hour of model swapping. The algorithm decides your ceiling; the data decides whether you ever reach it.