20c

RandomForest Machine Learning – Oracle Machine Learning (OML)

Posted on

Oracle Machine Learning has 30+ different machine learning algorithms built into the database. This means you can use SQL to create machine learning models and then use these models to score or label new data stored in the database or as the data is being created dynamically in the applications.

One of the most commonly used machine learning algorithms, over the past few years, is can RandomForest. This post will take a closer look at this algorithm and how you can build & use a RandomForest model.

Random Forest is known as an ensemble machine learning technique that involves the creation of hundreds of decision tree models. These hundreds of models are used to label or score new data by evaluating each of the decision trees and then determining the outcome based on the majority result from all the decision trees. Just like in the game show. The combining of a number of different ways of making a decision can result in a more accurate result or prediction.

Random Forest models can be used for classification and regression types of problems, which form the majority of machine learning systems and solutions. For classification problems, this is where the target variable has either a binary value or a small number of defined values. For classification problems the Random Forest model will evaluate the predicted value for each of the decision trees in the model. The final predicted outcome will be the majority vote for all the decision trees. For regression problems the predicted value is numeric and on some range or scale. For example, we might want to predict a customer’s lifetime value (LTV), or the potential value of an insurance claim, etc. With Random Forest, each decision tree will make a prediction of this numeric value. The algorithm will then average these values for the final predicted outcome.

 

Under the hood, Random Forest is a collection of decision trees. Although decision trees are a popular algorithm for machine learning, they can have a tendency to over fit the model. This can lead higher than expected errors when predicting unseen data. It also gives just one possible way of representing the data and being able to derive a possible predicted outcome.

Random Forest on the other hand relies of the predicted outcomes from many different decision trees, each of which is built in a slightly different way. It is an ensemble technique that combines the predicted outcomes from each decision tree to give one answer. Typically, the number of trees created by the Random Forest algorithm is defined by a parameter setting, and in most languages this can default to 100+ or 200+ trees.

The Random Forest algorithm has three main features:

  • It uses a method called bagging to create different subsets of the original training data
  • It will randomly section different subsets of the features/attributes and build the decision tree based on this subset
  • By creating many different decision trees, based on different subsets of the training data and different subsets of the features, it will increase the probability of capturing all possible ways of modeling the data

For each decision tree produced, the algorithm will use a measure, such as the Gini Index, to select the attributes to split on at each node of the decision tree.

To create a RandomForest model using Oracle Data Mining, you will follow the same process as with any of the other algorithms, the core of these are:

  1. define the parameter settings
  2. create the model
  3. score/label new data

Let’s start with the first step, defining the parameters. As with all the classification algorithms the same or similar parameters are set. With RandomForest we can set an additional parameter which tells the algorithm how many decision trees to create as part of the model. By default, 20 decision trees will be created. But if you want to change this number you can use the RFOR_NUM_TREES parameter. Remember the larger the value the longer it will take to create the model. But will have better accuracy. On the other hand with a small number of trees the quicker the model build will be, but might night be as accurate. This is something you will need to explore and determine. In the following example I change the number of trees to created to ten.

CREATE TABLE BANKING_RF_SETTINGS (
    SETTING_NAME VARCHAR2(50),
    SETTING_VALUE VARCHAR2(50)
);

BEGIN
  DELETE FROM BANKING_RF_SETTINGS;

  INSERT INTO banking_RF_settings (setting_name, setting_value)
  VALUES (dbms_data_mining.algo_name, dbms_data_mining.algo_random_forest);

  INSERT INTO banking_RF_settings (setting_name, setting_value)
  VALUES (dbms_data_mining.prep_auto, dbms_data_mining.prep_auto_on);

  INSERT INTO banking_RF_settings (setting_name, setting_value)
  VALUES (dbms_data_mining.RFOR_NUM_TREES, 10);

  COMMIT;
END;

Other default parameters used include, for creating each decision tree, use random 50% selection of columns and 50% sample of training data.

Now for step 2, create the model.

DECLARE
   v_start_time  TIMESTAMP;
BEGIN
   DBMS_DATA_MINING.DROP_MODEL('BANKING_RF_72K_1');

   v_start_time := current_timestamp;

   DBMS_DATA_MINING.CREATE_MODEL(
      model_name          => 'BANKING_RF_72K_1',
      mining_function     => dbms_data_mining.classification,
      data_table_name     => 'BANKING_72K',
      case_id_column_name => 'ID',
      target_column_name  => 'TARGET',
      settings_table_name => 'BANKING_RF_SETTINGS');

   dbms_output.put_line('Time take to create model = ' || to_char(extract(second from (current_timestamp-v_start_time))) || ' seconds.');
END;

The above code measures how long it takes to create the model.

I’ve run this same parameters and create models for different training data set sizes. I’ve also changed the number of decision trees to create. The following table shows the timings.

You can see it took 5.23 seconds to create a RandomForest model using the default settings for a data set of 72K records. This increase to just over one minute for a data set of 2 million records. Yo can also see the effect of reducing the number of decision trees on how long it takes the create model to run.

For step 3, on using the model on new data, this is just the same as with any of the classification models. Here is an example:

SELECT cust_id, target,
       prediction(BANKING_RF_72K_1 USING *)  predicted_value,
       prediction_probability(BANKING_RF_72K_1 USING *) probability
FROM   bank_test_v;

 

That’s it. That’s all there is to creating a RandomForest machine learning model using Oracle Machine Learning.

It’s quick and easy 🙂

 

XGBoost in Oracle 20c

Posted on Updated on

Another of the new machine learning algorithms in Oracle 20c Database is called XGBoost. Most people will have come across this algorithm due to its recent popularity with winners of Kaggle competitions and other similar events.

XGBoost is an open source software library providing a gradient boosting framework in most of the commonly used data science, machine learning and software development languages. It has it’s origins back in 2014, but the first official academic publication on the algorithm was published in 2016 by Tianqi Chen and Carlos Guestrin, from the University of Washington.

The algorithm builds upon the previous work on Decision Trees, Bagging, Random Forest, Boosting and Gradient Boosting. The benefits of using these various approaches are well know, researched, developed and proven over many years. XGBoost can be used for the typical use cases of Classification including classification, regression and ranking problems. Check out the original research paper for more details of the inner workings of the algorithm.

Regular machine learning models, like Decision Trees, simply train a single model using a training data set, and only this model is used for predictions. Although a Decision Tree is very simple to create (and very very quick to do so) its predictive power may not be as good as most other algorithms, despite providing model explainability. To overcome this limitation ensemble approaches can be used to create multiple Decision Trees and combines these for predictive purposes. Bagging is an approach where the predictions from multiple DT models are combined using majority voting. Building upon the bagging approach Random Forest uses different subsets of features and subsets of the training data, combining these in different ways to create a collection of DT models and presented as one model to the user. Boosting takes a more iterative approach to refining the models by building sequential models with each subsequent model is focused on minimizing the errors of the previous model. Gradient Boosting uses gradient descent algorithm to minimize errors in subsequent models. Finally with XGBoost builds upon these previous steps enabling parallel processing, tree pruning, missing data treatment, regularization and better cache, memory and hardware optimization. It’s commonly referred to as gradient boosting on steroids.

The following three images illustrates the differences between Decision Trees, Random Forest and XGBoost.

The XGBoost algorithm in Oracle 20c has over 40 different parameter settings, and with most scenarios the default settings with be fine for most scenarios. Only after creating a baseline model with the details will you look to explore making changes to these. Some of the typical settings include:

  • Booster =  gbtree
  • #rounds for boosting = 10
  • max_depth = 6
  • num_parallel_tree = 1
  • eval_metric = Classification error rate  or  RMSE for regression

 

As with most of the Oracle in-database machine learning algorithms, the setup and defining the parameters is really simple. Here is an example of minimum of parameter settings that needs to be defined.

BEGIN
   -- delete previous setttings
   DELETE FROM banking_xgb_settings;

   INSERT INTO BANKING_XGB_SETTINGS (setting_name, setting_value)
   VALUES (dbms_data_mining.algo_name, dbms_data_mining.algo_xgboost);

   -- For 0/1 target, choose binary:logistic as the objective.
   INSERT INTO BANKING_XGB_SETTINGS (setting_name, setting_value)
   VALUES (dbms_data_mining.xgboost_objective, 'binary:logistic);

   commit;
END;

 

To create an XGBoost model run the following.


BEGIN
   DBMS_DATA_MINING.CREATE_MODEL (
      model_name          => 'BANKING_XGB_MODEL',
      mining_function     => dbms_data_mining.classification,
      data_table_name     => 'BANKING_72K',
      case_id_column_name => 'ID',
      target_column_name  => 'TARGET',
      settings_table_name => 'BANKING_XGB_SETTINGS');
END;

That’s all nice and simple, as it should be, and the new model can be called in the same manner as any of the other in-database machine learning models using functions like PREDICTION, PREDICTION_PROBABILITY, etc.

One of the interesting things I found when experimenting with XGBoost was the time it took to create the completed model. Using the default settings the following table gives the time taken, in seconds to create the model.

As you can see it is VERY quick even for large data sets and gives greater predictive accuracy.

 

MSET (Multivariate State Estimation Technique) in Oracle 20c

Posted on

Oracle 20c Database comes with some new in-database Machine Learning algorithms.

The short name for one of these is called MSET or Multivariate State Estimation Technique. That’s the simple short name. The more complete name is Multivariate State Estimation Technique – Sequential Probability Ratio Test.  That is a long name, and the reason is it consists of two algorithms. The first part looks at creating a model of the training data, and the second part looks at how new data is statistical different to the training data.

 

What are the use cases for this algorithm?  This algorithm can be used for anomaly detection.

Anomaly Detection, using algorithms, is able identifying unexpected items or events in data that differ to the norm. It can be easy to perform some simple calculations and graphics to examine and present data to see if there are any patterns in the data set. When the data sets grow it is difficult for humans to identify anomalies and we need the help of algorithms.

The images shown here are easy to analyze to spot the anomalies and it can be relatively easy to build some automated processing to identify these. Most of these solutions can be considered AI (Artificial Intelligence) solutions as they mimic human behaviors to identify the anomalies, and these example don’t need deep learning, neural networks or anything like that.

Other types of anomalies can be easily spotted in charts or graphics, such as the chart below.

There are many different algorithms available for anomaly detection, and the Oracle Database already has an algorithm called the One-Class Support Vector Machine. This is a variant of the main Support Vector Machine (SVD) algorithm, which maps or transforms the data, using a Kernel function, into space such that the data belonging to the class values are transformed by different amounts. This creates a Hyperplane between the mapped/transformed values and hopefully gives a large margin between the mapped/transformed points. This is what makes SVD very accurate, although it does have some scaling limitations. For a One-Class SVD, a similar process is followed. The aim is for anomalous data to be mapped differently to common or non-anomalous data, as shown in the following diagram.

 

Getting back to the MSET algorithm. Remember it is a 2-part algorithm abbreviated to MSET. The first part is a non-linear, nonparametric anomaly detection algorithm that calibrates the expected behavior of a system based on historical data from the normal sequence of monitored signals. Using data in time series format (DATE, Value) the training data set contains data consisting of “normal” behavior of the data. The algorithm creates a model to represent this “normal”/stationary data/behavior. The second part of the algorithm compares new or live data and calculates the differences between the estimated and actual signal values (residuals). It uses Sequential Probability Ratio Test (SPRT) calculations to determine whether any of the signals have become degraded. As you can imagine the creation of the training data set is vital and may consist of many iterations before determining the optimal training data set to use.

MSET has its origins in computer hardware failures monitoring. Sun Microsystems have been were using it back in the late 1990’s-early 2000’s to monitor and detect for component failures in their servers. Since then MSET has been widely used in power generation plants, airplanes, space travel, Disney uses it for equipment failures, and in more recent times has been extensively used in IOT environments with the anomaly detection focused on signal anomalies.

How does MSET work in Oracle 20c?

An important point to note before we start is, you can use MSET on your typical business data and other data stored in the database. It isn’t just for sensor, IOT, etc data mentioned above and can be used in many different business scenarios.

The first step you need to do is to create the time series data. This can be easily done using a view, but a Very important component is the Time attribute needs to be a DATE format. Additional attributes can be numeric data and these will be used as input to the algorithm for model creation.

-- Create training data set for MSET
CREATE OR REPLACE VIEW mset_train_data
AS SELECT time_id, 
          sum(quantity_sold) quantity,
          sum(amount_sold) amount 
FROM (SELECT * FROM sh.sales WHERE time_id <= '30-DEC-99’)
GROUP BY time_id 
ORDER BY time_id;

The example code above uses the SH schema data, and aggregates the data based on the TIME_ID attribute. This attribute is a DATE data type. The second import part of preparing and formatting the data is Ordering of the data. The ORDER BY is necessary to ensure the data is fed into or processed by the algorithm in the correct time series order.

The next step involves defining the parameters/hyper-parameters for the algorithm. All algorithms come with a set of default values, and in most cases these are suffice for your needs. In that case, you only need to define the Algorithm Name and to turn on Automatic Data Preparation. The following example illustrates this and also includes examples of setting some of the typical parameters for the algorithm.

BEGIN
  DELETE FROM mset_settings;

  -- Select MSET-SPRT as the algorithm
  INSERT  INTO mset_sh_settings (setting_name, setting_value)
  VALUES(dbms_data_mining.algo_name, dbms_data_mining.algo_mset_sprt);

  -- Turn on automatic data preparation
  INSERT INTO mset_sh_settings (setting_name, setting_value)
  VALUES(dbms_data_mining.prep_auto, dbms_data_mining.prep_auto_on);

  -- Set alert count
  INSERT INTO mset_sh_settings (setting_name, setting_value)
  VALUES(dbms_data_mining.MSET_ALERT_COUNT, 3);

  -- Set alert window
  INSERT INTO mset_sh_settings (setting_name, setting_value)
  VALUES(dbms_data_mining.MSET_ALERT_WINDOW, 5);

  -- Set alpha
  INSERT INTO mset_sh_settings (setting_name, setting_value)
  VALUES(dbms_data_mining.MSET_ALPHA_PROB, 0.1);

  COMMIT;
END;

To create the MSET model using the MST_TRAIN_DATA view created above, we can run:

BEGIN
--   DBMS_DATA_MINING.DROP_MODEL(MSET_MODEL');

   DBMS_DATA_MINING.CREATE_MODEL (
      model_name          => 'MSET_MODEL',
      mining_function     => dbms_data_mining.classification,
      data_table_name     => 'MSET_TRAIN_DATA',
      case_id_column_name => 'TIME_ID',
      target_column_name  => '',
      settings_table_name => 'MSET_SETTINGS');
END;

The SELECT statement below is an example of how to call and run the MSET model to label the data to find anomalies. The PREDICTION function will return a values of 0 (zero) or 1 (one) to indicate the predicted values. If the predicted values is 0 (zero) the MSET model has predicted the input record to be anomalous, where as a predicted values of 1 (one) indicates the value is typical. This can be used to filter out the records/data you will want to investigate in more detail.

-- display all dates with Anomalies
SELECT time_id, pred
FROM (SELECT time_id, prediction(mset_sh_model using *) over (ORDER BY time_id) pred 
      FROM mset_test_data)
WHERE pred = 0;