ML Commons supports various algorithms to help train and predict machine learning (ML) models or test data-driven predictions without a model. This page outlines the algorithms supported by the ML Commons plugin and the API operations they support.
K-means is a simple and popular unsupervised clustering ML algorithm built on top of [Tribuo](https://tribuo.org/) library. K-means will randomly choose centroids, then calculate iteratively to optimize the position of the centroids until each observation belongs to the cluster with the nearest mean.
`centroids` | integer | The number of clusters in which to group the generated data | `2`
`iterations` | integer | The number of iterations to perform against the data until a mean generates | `10`
`distance_type` | enum, such as `EUCLIDEAN`, `COSINE`, or `L1` | The type of measurement from which to measure the distance between centroids | `EUCLIDEAN`
Linear regression maps the linear relationship between inputs and outputs. In ML Commons, the linear regression algorithm is adopted from the public machine learning library [Tribuo](https://tribuo.org/), which offers multidimensional linear regression models. The model supports the linear optimizer in training, including popular approaches like Linear Decay, SQRT_DECAY, [ADA](https://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf), [ADAM](https://tribuo.org/learn/4.1/javadoc/org/tribuo/math/optimisers/Adam.html), and [RMS_DROP](https://tribuo.org/learn/4.1/javadoc/org/tribuo/math/optimisers/RMSProp.html).
`learningRate` | Double | The initial step size used in an iterative optimization algorithm. | `0.01`
`momentumFactor` | Double | The extra weight factors that accelerate the rate at which the weight is adjusted. This helps move the minimization routine out of local minima. | `0`
`epsilon` | Double | The value for stabilizing gradient inversion. | `1.00E-06`
`beta1` | Double | The exponential decay rates for the moment estimates. | `0.9`
`beta2` | Double | The exponential decay rates for the moment estimates. | `0.99`
`decayRate` | Double | The Root Mean Squared Propagation (RMSProp). | `0.9`
`momentumType` | String | The defined Stochastic Gradient Descent (SGD) momentum type that helps accelerate gradient vectors in the right directions, leading to a fast convergence.| `STANDARD`
`optimizerType` | String | The optimizer used in the model. | `SIMPLE_SGD`
ML Commons only supports the linear Stochastic gradient trainer or optimizer, which cannot effectively map the non-linear relationships in trained data. When used with complicated datasets, the linear Stochastic trainer might cause some convergence problems and inaccurate results.
[Random Cut Forest](https://github.com/aws/random-cut-forest-by-aws) (RCF) is a probabilistic data structure used primarily for unsupervised anomaly detection. Its use also extends to density estimation and forecasting. OpenSearch leverages RCF for anomaly detection. ML Commons supports two new variants of RCF for different use cases:
For FIT RCF, you can train the model with historical data and store the trained model in your index. The model will be deserialized and predict new data points when using the Predict API. However, the model in the index will not be refreshed with new data, because the model is fixed in time.
RCF Summarize is a clustering algorithm based on the Clustering Using Representatives (CURE) algorithm. Compared to [k-means](#k-means), which uses random iterations to cluster, RCF Summarize uses a hierarchical clustering technique. The algorithm starts, with a set of randomly selected centroids larger than the centroids' ground truth distribution. During iteration, centroid pairs too close to each other automatically merge. Therefore, the number of centroids (`max_k`) converge to a rational number of clusters that fits ground truth, as opposed to a fixed `k` number of clusters.
| `max_k` | Integer | The max allowed number of centroids. | 2 |
| `distance_type` | String. Valid values are `EUCLIDEAN`, `L1`, `L2`, and `LInfinity` | The type of measurement used to measure the distance between centroids. | `EUCLIDEAN` |
The `rows` parameter within the prediction result has been modified for length. In your response, expect more rows and columns to be contained within the response body.
The Localization algorithm finds subset-level information for aggregate data (for example, aggregated over time) that demonstrates the activity of interest, such as spikes, drops, changes, or anomalies. Localization can be applied in different scenarios, such as data exploration or root cause analysis, to expose the contributors driving the activity of interest in the aggregate data.
The Localization algorithm can only be executed directly. Therefore, it cannot be used with the ML Commons Train and Predict APIs.
## Logistic regression
A classification algorithm, logistic regression models the probability of a discrete outcome given an input variable. In ML Commons, these classifications include both binary and multi-class. The most common is the binary classification, which takes two values, such as "true/false" or "yes/no", and predicts the outcome based on the values specified. Alternatively, a multi-class output can categorize different inputs based on type. This makes logistic regression most useful for situations where you are trying to determine how your inputs fit best into a specified category.
| `learningRate` | Double | The initial step size used in an iterative optimization algorithm. | `1` |
| `momentumFactor` | Double | The extra weight factors that accelerate the rate at which the weight is adjusted. This helps move the minimization routine out of local minima. | `0` |
| `epsilon` | Double | The value for stabilizing gradient inversion. | `0.1` |
| `beta1` | Double | The exponential decay rates for the moment estimates. | `0.9` |
| `beta2` | Double | The exponential decay rates for the moment estimates. | `0.99` |
| `decayRate` | Double | The Root Mean Squared Propagation (RMSProp). | `0.9` |
| `momentumType` | String | The Stochastic Gradient Descent (SGD) momentum that helps accelerate gradient vectors in the right direction, leading to faster convergence between vectors. | `STANDARD` |
| `optimizerType` | String | The optimizer used in the model. | `AdaGrad` |
| `target` | String | The target field. | null |
| `objectiveType` | String | The objective function type. | `LogMulticlass` |
| `epochs` | Integer | The number of iterations. | `5` |
| `batchSize` | Integer | The size of min batches. | `1` |
| `loggingInterval` | Integer | The interval of logs lost after many iterations. The interval is `1` if the algorithm contains no logs. | `1000` |
The following example creates an index in OpenSearch with the [Iris dataset](https://en.wikipedia.org/wiki/Iris_flower_data_set), then trains the data using logistic regression. Lastly, it uses the trained model to predict Iris types separated by row.
This example uses a multi-class logistic regression categorization methodology. Here, the inputs of sepal and petal length and width are used to train the model to categorize centroids based on the `class`, as indicated by the `target` parameter.
Convergence metrics are not built into Tribuo's trainers. Therefore, ML Commons cannot indicate the convergence status through the ML Commons API.
## Metrics correlation
The metrics correlation feature is an experimental feature released in OpenSearch 2.7. It can't be used in a production environment. To leave feedback on improving the feature, create an issue in the [ML Commons repository](https://github.com/opensearch-project/ml-commons).
{: .warning }
The metrics correlation algorithm finds events in a set of metrics data. The algorithm defines events as a window in time in which multiple metrics simultaneously display anomalous behavior. When given a set of metrics, the algorithm counts the number of events that occurred, when each event occurred, and determines which metrics were involved in each event.
To enable the metrics correlation algorithm, update the following cluster setting:
The metrics correlation input is an M x T array of metrics data, where M is the number of metrics and T is the length of each individual sequence of metric values.
-`event_pattern`: The intensity score across the time window and the overall severity of the event
-`suspected_metrics`: The set of metrics involved
In the following example response, each item corresponds to an event discovered in the metrics data. The algorithm finds one event in the input data of the request, as indicated by the output in `event_pattern` having a length of `1`. `event_window` shows that the event occurred between time point $t$ = 52 and $t$ = 72. Lastly, `suspected_metrics` shows that the event involved all three metrics.
