Oracle Data Mining Primer (2 of 3): Viewing Models

Now that we have a mining model in the database, we can interrogate the database to understand what has been discovered during the training process.

Three catalog views contain the high-level information for mining models: ALL_MINING_MODELS, ALL_MINING_MODEL_ATTRIBUTES, and ALL_MINING_MODEL_SETTINGS.

By querying ALL_MINING_MODELS, we retrieve the list of all mining models available to the current user.

select model_name, mining_function, algorithm
from all_mining_models;

PREDICT_INCOME    CLASSIFICATION    SUPPORT_VECTOR_MACHINES

By querying ALL_MINING_MODEL_ATTRIBUTES, we retrieve the list of column names that were relevant when training the model, as well as other per-attribute information.

select attribute_name, attribute_type, target

from all_mining_model_attributes
where model_name = 'PREDICT_INCOME';
AGE               NUMERICAL      NO
WORKCLASS         CATEGORICAL    NO
FNLWGT            NUMERICAL      NO
EDUCATION         CATEGORICAL    NO
EDUCATION_NUM     NUMERICAL      NO
MARITAL_STATUS    CATEGORICAL    NO
OCCUPATION        CATEGORICAL    NO
RELATIONSHIP      CATEGORICAL    NO
RACE              CATEGORICAL    NO
SEX               CATEGORICAL    NO
CAPITAL_GAIN      NUMERICAL      NO
CAPITAL_LOSS      NUMERICAL      NO
HOURS_PER_WEEK    NUMERICAL      NO
NATIVE_COUNTRY    CATEGORICAL    NO
INCOME            CATEGORICAL    YES

By querying ALL_MINING_MODEL_SETTINGS, we retrieve the list of model settings used during training.  Some of these settings may have been specified by the user, others automatically computed while training.  For the support vector machine algorithm, Oracle supports two kernel functions: gaussian and linear.  In this instance, based on the shape and content of data, Oracle chose to use the linear kernel as evidenced by the setting value SVMS_LINEAR.

select setting_name, setting_value
from all_mining_model_settings
where model_name = 'PREDICT_INCOME';
ALGO_NAME               ALGO_SUPPORT_VECTOR_MACHINES
SVMS_ACTIVE_LEARNING    SVMS_AL_ENABLE
PREP_AUTO               ON
SVMS_COMPLEXITY_FACTOR  0.495891
SVMS_KERNEL_FUNCTION    SVMS_LINEAR
SVMS_CONV_TOLERANCE     .001

The above three catalog views provide high-level information that is relevant to most Oracle Data Mining algorithms.  Additional, deeper insight is often available by querying the details of the model.  Each algorithm tends to use a different structure to represent what was learned while training, and therefore the structure of this returned information will vary from one algorithm to the next (possibly from one flavor of a given algorithm to the next).  For SVM with linear kernel, the model will contain a set of coefficients, as shown below for the PREDICT_INCOME model.

select class, attribute_name, attribute_value, coefficient 
from table(dbms_data_mining.get_model_details_svm('PREDICT_INCOME')) a, table(a.attribute_set) b
order by abs(coefficient) desc;
 >50K    CAPITAL_GAIN      (null)          8.1179930161383904
 >50K    (null)            (null)         -4.1469740381933802
 >50K    EDUCATION_NUM     (null)          1.85498650687918
 >50K    HOURS_PER_WEEK    (null)          1.80588516494733
 >50K    CAPITAL_LOSS      (null)          1.28361583304225
 >50K    AGE               (null)          1.20889883984869
 >50K    EDUCATION         Doctorate       1.1139153328993401
 >50K    NATIVE_COUNTRY    Nicaragua      -1.0957665557355201
 >50K    WORKCLASS         Without-pay    -0.99178110036931799
 >50K    NATIVE_COUNTRY    Columbia       -0.99178110036931699
 >50K    RELATIONSHIP      Wife            0.99046458006739702
 >50K    NATIVE_COUNTRY    Hungary        -0.973898034330827

 ...

In order to understand the information from the above query, it is necessary to scratch the surface of the Support Vector Machine algorithm.  As a result of training the model, we generate an equation which includes the following snippet:

a₁x₁ + a₂x₂ + a₃x₃ + ... + a_nx_n

where a_i is a coefficient from the above query result and x_i is the corresponding value for the attribute in the dataset.  For example, for a record that has HOURS_PER_WEEK of 10, the resulting contribution of HOURS_PER_WEEK in determining income level is (10*1.80588516494733).   For a categorical attribute, we assign 1 for the value of x for the attribute_value that is present in the incoming record, and 0 for all the others.  In this case, if the incoming record represents a person from Nicaragua, then we will include -1.0957665557355201 in the equation (1*-1.0957665557355201), but will not include anything related to people from Columbia, Hungary, etc. as their respective contributions are 0.

So, what does all this mean?

This means that, according to the trained model, larger values of HOURS_PER_WEEK will tend to make it more likely that the resulting income level is >50K (the class value identified in the above query result).  This also means that residents of Nicaragua are slightly less likely to have an income level >50K than residents of Columbia (the contribution away from the high income class is larger for Nicaragua since the magnitude of the coefficient is larger and both are negative).

By taking information from all of the attributes together, the model is able to yield a prediction as to whether or not a particular individual is likely to earn more than 50K.  The final part of this ODM primer will demonstrate how the database provides these resulting predictions.