CSCI 447
Machine Learning
Spring 2019
|
CSCI 447 |
| Algorithm | Input Data | Output Type | Problem Type | Benefits | Drawbacks |
|---|---|---|---|---|---|
| Linear Regression | Numeric | Numeric continuous | Regression / curve fitting / prediction | Small, fast, easy simple | Assumes residuals fit normal distribution, data are iid, variance constant |
| Logistic Regression | Numeric | Categorical | Curve fitting / prediction | Small, fast, easy simple | Assumes data are iid |
| Clustering | Unlabeled anything | Categorical / buckets | Unsupervised | Flexibe; can be used to label data | Will always output clusters, but they might not be meaningful; sensitive to initial random placement of centroids; assume some categorization exists |
| Nearest Neighbor | Labeled anything | Categorical / labeled | Heuristic prediction | No learning needed; straightforward | Doesn't scale well, curse of dimensionality |
| Deep Networks | Labeled Numeric | Prediction - categorical, probability, numeric prediction | Supervised, high dimensional | Few assumptions; can fit complex anything | Prone to overfitting; doesn't react to change in operation; not as simple to implement (hyperparameters); hard to explain; solution not optimal |
| Convolutional Networks | High dimensional data (images, etc.) | Prediction - categorical, probability, numeric prediction | Classification; supervised high dimensional | Learns important features / preprocessing; don't need to handcraft features | Prone to overfitting; doesn't react to change in operation; not as simple to implement (hyperparameters); hard to explain; solution not optimal |
| Recurrent Networks | Sequence data | Prediction - categorical, probability, numeric prediction | Time/space sensitive | No independence assumptions; memory of previous input | Prone to overfitting; doesn't react to change in operation; not as simple to implement (hyperparameters); hard to explain; solution not optimal; recurrence difficult to program |
| Bayesian Networks | Mostly categorical, but can be numeric | 1) Model, 2) Probabilities for many variable combinations, 3) Conditional independencies | Many | Generative; easy to understand | Could be expensive based on number of variables |
| Genetic Algorithms | Data | Optimization / design solutions | Optimization / design | Generative; doesn't use gradient; novel solutions | Very expensive |
| Decision Trees | Labeled anything | Prediction / clasification, numeric | Simple | Easy to build, use, understand | Prone to under- and over-fitting |
Page last updated: April 08, 2019
