Predicting classes with classification

Classification is a machine learning process that predicts the class or category of a data point in a data set. For a simple example, consider how the shapes in the following graph can be differentiated and classified as "circles" and "triangles":

In reality, classification problems are more complex, such as classifying malicious and benign domains to detect DGA activities for security reasons or predicting customer churn based on customer calling data. Classification is for predicting discrete, categorical values.

When you create a classification job, you must specify which field contains the classes that you want to predict. This field is known as the dependent variable. It can contain maximum 100 classes. By default, all other supported fields are included in the analysis and are known as feature variables. You can optionally include or exclude fields. For more information about field selection, refer to the explain data frame analytics API.

Classification analysis uses an ensemble algorithm that is similar to extreme gradient boosting (XGBoost) which combines multiple weak models into a composite one. It uses decision trees to learn to predict the probability that a data point belongs to a certain class. XGBoost trains a sequence of decision trees and every decision tree learns from the mistakes of the forest so far. In each iteration, the trees added to the forest improve the decision quality of the combined decision forest. The classification algorithm optimizes for a loss function called cross-entropy loss.

Classification can be useful in cases where discrete, categorical values needs to be predicted. If your use case requires predicting such values, then classification might be the suitable choice for you.

Before you can use the Elastic Stack machine learning features, there are some configuration requirements (such as security privileges) that must be addressed. Refer to Setup and security.

Classification is a supervised machine learning method, which means you need to supply a labeled training data set. This data set must have values for the feature variables and the dependent variable which are used to train the model. The training process uses this information to learn the relationships between the classes and the feature variables. This labeled data set also plays a critical role in model evaluation.

If possible, prepare your input data such that it has less classes. A classification analysis with many classes takes more time to run than a binary classification job. The relationship between the number of classes and the runtime is roughly linear.

You might also need to transform your data to create a data frame which can be used as the source for classification.

To learn more about how to prepare your data, refer to the relevant section of the supervised learning overview.

Data frame analytics jobs contain the configuration information and metadata necessary to perform an analytics task. You can create data frame analytics jobs via Kibana or using the create data frame analytics jobs API.

Select classification as the analytics type, then select the field that you want to predict (the dependent variable). You can also include and exclude fields.

To improve performance, consider using a small training_percent value to train the model more quickly. It is a good strategy to make progress iteratively: run the analysis with a small training percentage, then evaluate the performance. Based on the results, you can decide if it is necessary to increase the training_percent value.

You can start the job via Kibana or using the start data frame analytics jobs API. A classification job has the following phases:

After the last phase is finished, the job stops and the results are ready for evaluation.

Using the data frame analytics features to gain insights from a data set is an iterative process. After you defined the problem you want to solve, and chose the analytics type that can help you to do so, you need to produce a high-quality data set and create the appropriate data frame analytics job. You might need to experiment with different configurations, parameters, and ways to transform data before you arrive at a result that satisfies your use case. A valuable companion to this process is the evaluate data frame analytics API, which enables you to evaluate the data frame analytics performance. It helps you understand error distributions and identifies the points where the data frame analytics model performs well or less trustworthily.

To evaluate the analysis with this API, you need to annotate your index that contains the results of the analysis with a field that marks each document with the ground truth. The evaluate data frame analytics API evaluates the performance of the data frame analytics against this manually provided ground truth.

You can measure how well the model has performed on your training data set by using the classification evaluation type of the evaluate data frame analytics API or by viewing the job results in Kibana. The classification evaluation offers the following metrics to evaluate the model performance:

The following metrics helps you interpret the analysis results:

The multiclass confusion matrix provides a summary of the performance of the classification analysis. It contains the number of occurrences where the analysis classified data points correctly with their actual class as well as the number of occurrences where it misclassified them.

This is an example of a confusion matrix for a binary problem:

It is a two by two matrix because there are only two classes (true and false). It shows the proportion of data points that is correctly identified as members of a each class and the proportion that is misidentified.

As the number of classes increases, the confusion matrix becomes more complex:

This matrix contains the actual labels on the left side while the predicted labels are on the top. The proportion of correct and incorrect predictions is broken down for each class. This enables you to examine how the classification analysis confused the different classes while it made its predictions.

The receiver operating characteristic (ROC) curve is a plot that represents the performance of the classification process at different predicted probability thresholds. It compares the true positive rate for a specific class against the rate of all the other classes combined ("one versus all" strategy) at the different threshold levels to create the curve.

For example, there are three classes: A, B, and C, and AUC ROC is calculated for A. In this case, the number of correctly classified As (true positives) are compared to the number of Bs and Cs that are misclassified as As (false positives).

From this plot, you can compute the area under the curve (AUC) value, which is a number between 0 and 1. The higher the AUC, the better the model is at predicting As as As, in this case.

Feature importance provides further information about the results of an analysis and helps to interpret the results in a more subtle way. If you want to learn more about feature importance, refer to Feature importance.

The class_probability is a value between 0 and 1, which indicates how likely it is that a given data point belongs to a certain class. The higher the number, the higher the probability that the data point belongs to the named class. This information is stored in the top_classes array for each document in the destination index.

The class_score is a function of the class_probability and has a value that is greater than or equal to zero. It takes into consideration your objective (as defined in the class_assignment_objective job configuration option): accuracy or recall.

If your objective is to maximize accuracy, the scores are weighted to maximize the proportion of correct predictions in the training data set.

By default, classification analysis jobs accept a slight degradation of the overall accuracy in return for greater sensitivity to classes that are predicted incorrectly. That is to say, their objective is to maximize the minimum recall. For example, in the context of a multi-class confusion matrix, the predictions of interest are in each row:

For each class, the recall is calculated as the number of correct predictions divided by the sum of all the other predicted labels in that row. This value is represented as a percentage in each cell of the confusion matrix. The class scores are then weighted to favor predictions that result in the highest recall values across the training data. This objective typically performs better than accuracy when you have highly imbalanced data.

To learn more about choosing the class assignment objective that fits your goal, refer to this Jupyter notebook.

The model that you created is stored as Elasticsearch documents in internal indices. In other words, the characteristics of your trained model are saved and ready to be deployed and used as functions.

Inference is a machine learning feature that enables you to use supervised machine learning processes – like regression or classification – not only as a batch analysis but in a continuous fashion. This means that inference makes it possible to use trained machine learning models against incoming data.

For instance, suppose you have an online service and you would like to predict whether a customer is likely to churn. You have an index with historical data – information on the customer behavior throughout the years in your business – and a classification model that is trained on this data. The new information comes into a destination index of a continuous transform. With inference, you can perform the classification analysis against the new data with the same input fields that you’ve trained the model on, and get a prediction.

Inference can be used as a processor specified in an ingest pipeline. It uses a trained model to infer against the data that is being ingested in the pipeline. The model is used on the ingest node. Inference pre-processes the data by using the model and provides a prediction. After the process, the pipeline continues executing (if there is any other processor in the pipeline), finally the new data together with the results are indexed into the destination index.

Check the inference processor and the machine learning data frame analytics API documentation to learn more about the feature.

Inference can also be used as a pipeline aggregation. You can reference a trained model in the aggregation to infer on the result field of the parent bucket aggregation. The inference aggregation uses the model on the results to provide a prediction. This aggregation enables you to run classification or regression analysis at search time. If you want to perform the analysis on a small set of data, this aggregation enables you to generate predictions without the need to set up a processor in the ingest pipeline.

Check the inference bucket aggregation and the machine learning data frame analytics API documentation to learn more about the feature.

Let’s try to predict whether a flight will be delayed or not by using the sample flight data. The data set contains information such as weather conditions, carrier, flight distance, origin, destination, and whether or not the flight was delayed. The classification model learns the relationships between the fields in your data to predict the value of the dependent variable, which in this case is the boolean FlightDelay field.

Each document in the sample flight data set contains details for a single flight, so the data is ready for analysis; it is already in a two-dimensional entity-based data structure. In general, you often need to transform the data into an entity-centric index before you can analyze it.

In order to be analyzed, a document must contain at least one field with a supported data type (numeric, boolean, text, keyword or ip) and must not contain arrays with more than one item. If your source data consists of some documents that contain the dependent variable and some that do not, the model is trained on the subset of documents that contain it.

{
  "_index": "kibana_sample_data_flights",
  "_type": "_doc",
  "_id": "S-JS1W0BJ7wufFIaPAHe",
  "_version": 1,
  "_seq_no": 3356,
  "_primary_term": 1,
  "found": true,
  "_source": {
    "FlightNum": "N32FE9T",
    "DestCountry": "JP",
    "OriginWeather": "Thunder & Lightning",
    "OriginCityName": "Adelaide",
    "AvgTicketPrice": 499.08518599798685,
    "DistanceMiles": 4802.864932998549,
    "FlightDelay": false,
    "DestWeather": "Sunny",
    "Dest": "Chubu Centrair International Airport",
    "FlightDelayType": "No Delay",
    "OriginCountry": "AU",
    "dayOfWeek": 3,
    "DistanceKilometers": 7729.461862731618,
    "timestamp": "2019-10-17T11:12:29",
    "DestLocation": {
      "lat": "34.85839844",
      "lon": "136.8049927"
    },
    "DestAirportID": "NGO",
    "Carrier": "ES-Air",
    "Cancelled": false,
    "FlightTimeMin": 454.6742272195069,
    "Origin": "Adelaide International Airport",
    "OriginLocation": {
      "lat": "-34.945",
      "lon": "138.531006"
    },
    "DestRegion": "SE-BD",
    "OriginAirportID": "ADL",
    "OriginRegion": "SE-BD",
    "DestCityName": "Tokoname",
    "FlightTimeHour": 7.577903786991782,
    "FlightDelayMin": 0
  }
}

To predict whether a specific flight is delayed:

Now you have a new index that contains a copy of your source data with predictions for your dependent variable.

When you view the classification results in Kibana, it shows the contents of the destination index in a tabular format. It also provides information about the analysis details, model evaluation metrics, total feature importance values, and a scatterplot matrix.

The table shows a column for the dependent variable (FlightDelay), which contains the ground truth values that you are trying to predict. It also shows a column for the predicted values (ml.FlightDelay_prediction), which were generated by the classification analysis. The ml.is_training column indicates whether the document was used in the training or testing data set. You can use the Training and Testing filter options to refine the contents of the results table. You can also enable histogram charts to get a better understanding of the distribution of values.

If you want to understand how certain the model is about each prediction, you can examine its probability and score (ml.prediction_probability and ml.prediction_score). The higher these values are, the more confident the model is that the data point belongs to the named class. If you examine the destination index more closely in the Discover app in Kibana or use the standard Elasticsearch search command, you can see that the analysis predicts the probability of all possible classes for the dependent variable. The top_classes object contains the predicted classes with the highest scores.

GET model-flight-delays-classification/_search

          ...
          "FlightDelay" : false,
          ...
          "ml" : {
            "FlightDelay_prediction" : false,
            "top_classes" : [ 
              {
                "class_name" : false,
                "class_probability" : 0.9427605087816684,
                "class_score" : 0.3462468700158476
              },
              {
                "class_name" : true,
                "class_probability" : 0.057239491218331606,
                "class_score" : 0.057239491218331606
              }
            ],
            "prediction_probability" : 0.9427605087816684,
            "prediction_score" : 0.3462468700158476,
            ...

If you chose to calculate feature importance, the destination index also contains ml.feature_importance objects. Every field that is included in the analysis (known as a feature of the data point) is assigned a feature importance value. It has both a magnitude and a direction (positive or negative), which indicates how each field affects a particular prediction. Only the most significant values (in this case, the top 10) are stored in the index. However, the trained model metadata also contains the average magnitude of the feature importance values for each field across all the training data. You can view this summarized information in Kibana:

You can also see the feature importance values for each individual prediction in the form of a decision plot:

In Kibana, the decision path shows the relative impact of each feature on the probability of the prediction. The features with the most significant positive or negative impact appear at the top of the decision plot. Thus in this example, the features related to flight time and distance had the most significant influence on the probability value for this prediction. This type of information can help you to understand how models arrive at their predictions. It can also indicate which aspects of your data set are most influential or least useful when you are training and tuning your model.

If you do not use Kibana, you can see the summarized feature importance values by using the get trained model API and the individual values by searching the destination index.

GET _ml/trained_models/model-flight-delays-classification*?include=total_feature_importance

{
  "count" : 1,
  "trained_model_configs" : [
    {
      "model_id" : "model-flight-delays-classification-1601405047985",
      ...
      "metadata" : {
        ...
        "feature_importance_baseline" : { 
          "classes" : [
            {
              "class_name" : true,
              "baseline" : -1.5869016940485443
            },
            {
              "class_name" : false,
              "baseline" : 1.5869016940485443
            }
          ]
        },
        "total_feature_importance" : [
          {
            "feature_name" : "dayOfWeek",
            "classes" : [
              {
                "class_name" : false,
                "importance" : {
                  "mean_magnitude" : 0.037513174351966404, 
                  "min" : -0.20132653028125566, 
                  "max" : 0.20132653028125566 
                }
              },
              {
                "class_name" : true,
                "importance" : {
                  "mean_magnitude" : 0.037513174351966404,
                  "min" : -0.20132653028125566,
                  "max" : 0.20132653028125566
                }
              }
            ]
          },
          {
            "feature_name" : "OriginWeather",
            "classes" : [
              {
                "class_name" : false,
                "importance" : {
                  "mean_magnitude" : 0.05486662317369895,
                  "min" : -0.3337477336556598,
                  "max" : 0.3337477336556598
                }
              },
              {
                "class_name" : true,
                "importance" : {
                  "mean_magnitude" : 0.05486662317369895,
                  "min" : -0.3337477336556598,
                  "max" : 0.3337477336556598
                }
              }
            ]
          },
          ...

GET model-flight-delays-classification/_search

          ...
          "FlightDelay" : false,
          ...
          "ml" : {
            "FlightDelay_prediction" : false,
            ...
            "prediction_probability" : 0.9427605087816684,
            "prediction_score" : 0.3462468700158476,
            "feature_importance" : [
              {
                "feature_name" : "DistanceMiles",
                "classes" : [
                  {
                    "class_name" : false,
                    "importance" : -1.4766536146534828
                  },
                  {
                    "class_name" : true,
                    "importance" : 1.4766536146534828
                  }
                ]
              },
              {
                "feature_name" : "FlightTimeMin",
                "classes" : [
                  {
                    "class_name" : false,
                    "importance" : 1.0919201754729184
                  },
                  {
                    "class_name" : true,
                    "importance" : -1.0919201754729184
                  }
                ]
              },
              ...

Lastly, Kibana provides a scatterplot matrix in the results. It has the same functionality as the matrix that you saw in the job wizard. Its purpose is to help you visualize and explore the relationships between the numeric fields and the dependent variable.

Though you can look at individual results and compare the predicted value (ml.FlightDelay_prediction) to the actual value (FlightDelay), you typically need to evaluate the success of your classification model as a whole.

Kibana provides a normalized confusion matrix that contains the percentage of occurrences where the analysis classified data points correctly with their actual class and the percentage of occurrences where it misclassified them.

If you want to see the exact number of occurrences, select a quadrant in the matrix. You can also use the Training and Testing filter options to refine the contents of the matrix. Thus you can see how well the model performs on previously unseen data. You can check how many documents are true in the testing data, how many of them are identified correctly (true positives) and how many of them are identified incorrectly as false (false negatives).

Likewise if you select other quadrants in the matrix, it shows the number of documents that have the false class as their actual value in the testing data. The matrix shows the number of documents that are correctly identified as false (true negatives) and the number of documents that are incorrectly predicted as true (false positives). When you perform classification analysis on your own data, it might take multiple iterations before you are satisfied with the results and ready to deploy the model.

Kibana also provides the receiver operating characteristic (ROC) curve as part of the model evaluation. The plot compares the true positive rate (y-axis) to the false positive rate (x-axis) for each class; in this example, true and false. From this plot, the area under the curve (AUC) value is computed. It is a number between 0 and 1. The higher the AUC, the better the model is at predicting the classes correctly.

You can also generate these metrics with the data frame analytics evaluate API. For more information about interpreting the evaluation metrics, see 6. Evaluate and interpret the result.

POST _ml/data_frame/_evaluate
{
 "index": "model-flight-delays-classification",
   "query": {
    "term": {
      "ml.is_training": {
        "value": true  
      }
    }
  },
 "evaluation": {
   "classification": {
     "actual_field": "FlightDelay",
     "predicted_field": "ml.FlightDelay_prediction",
     "metrics": {
       "multiclass_confusion_matrix" : {}
     }
   }
 }
}

POST _ml/data_frame/_evaluate
{
 "index": "model-flight-delays-classification",
   "query": {
    "term": {
      "ml.is_training": {
        "value": false  
      }
    }
  },
 "evaluation": {
   "classification": {
     "actual_field": "FlightDelay",
     "predicted_field": "ml.FlightDelay_prediction",
     "metrics": {
       "multiclass_confusion_matrix" : {}
     }
   }
 }
}

{
  "classification" : {
    "multiclass_confusion_matrix" : {
      "confusion_matrix" : [
        {
          "actual_class" : "false", 
          "actual_class_doc_count" : 8802, 
          "predicted_classes" : [
            {
              "predicted_class" : "false", 
              "count" : 7262 
            },
            {
              "predicted_class" : "true",
              "count" : 1540
            }
          ],
          "other_predicted_class_doc_count" : 0
        },
        {
          "actual_class" : "true",
          "actual_class_doc_count" : 2952,
          "predicted_classes" : [
            {
              "predicted_class" : "false",
              "count" : 794
            },
            {
              "predicted_class" : "true",
              "count" : 2158
            }
          ],
          "other_predicted_class_doc_count" : 0
        }
      ],
      "other_actual_class_count" : 0
    }
  }
}

If you don’t want to keep the data frame analytics job, you can delete it in Kibana or by using the delete data frame analytics job API. When you delete data frame analytics jobs in Kibana, you have the option to also remove the destination indices and data views.