Little Bits of Artificial Intelligence

1. Do you have a classification problem?

• SVM is mainly used for binary classification (e.g., spam vs not spam, fraud vs not fraud).
• Can be extended to multi-class but is naturally binary.

Use SVM if: To classify data into two groups clearly.

2. Is our data labeled?

• SVM is a supervised learning algorithm.
• We must have known output labels for training.

Use SVM if: We have a dataset with clearly labeled examples.

3. Is our data linearly separable or nearly separable?

• SVM shines when a clear boundary exists between groups.
• If data is not linearly separable, kernel trick can help.

Use SVM if: We think a line (or a curved boundary with a kernel) can separate the classes well.

4. Do we have a small to medium-sized dataset?

• SVM is computationally intensive for large datasets.
• Not ideal for huge datasets with millions of records.

Use SVM if: The dataset is not too large (e.g., < 100,000 samples typically).

5. Are features scaled or normalized?

• SVM is sensitive to feature scales (e.g., height in cm vs weight in kg).
• Works best when features are on similar scales.

Use SVM if: We can scale/normalize our features before training.

6. Do we care more about accuracy than interpretability?

• SVM gives high accuracy, especially with clear margins.
• But it’s not easily interpretable (unlike decision trees).

Use SVM if: We want high performance, and we don’t need to explain the model easily.

7. Do we have more features than samples?

• SVM handles high-dimensional spaces very well.
• Great for text classification, image recognition, etc.

Use SVM if: We have problems like text, where features > samples (e.g., 10000 words vs 200 emails).