Showing posts with label Oracle Advanced Analytics. Show all posts
Showing posts with label Oracle Advanced Analytics. Show all posts

Friday, July 12, 2013

Oracle 12c Advanced Analytics Option new features

With the release of Oracle 12c (finally) now have a lot of learning to do. Oracle 12c is a different beast to what we have been used to up to now.

As part of the 12c there are a number of new in-database Advanced Analytics features. These are separate to the Advanced Analytics new features that come as part of the Oracle Data Miner tool, that is part of SQL Developer.

This post will only look at the new features that are part of the 12c Database. The new in-Database Advanced Analytics features include:

  • Using Decisions Trees for Text analysis is now possible. Up to now (11.2g) when you wanted to do text classification you had to exclude Decision Trees from the process. This was because the Decision Trees algorithm could not support nested data.
  • Additionally for text mining some of the text processing has been moved from having a separate step, to being part of the some of the algorithms.
  • A number of additional features are available for Clustering. These include a cluster distance (from the centroid) and details functions.
  • There is a new clustering algorithm (in addition to the K-Means and O-Cluster algorithms), called Expectation Maximization algorithm. This creates a density model that can be give better results when data from different domains are combined for clustering. This algorithm will also determine the optimal number of clusters.
  • There are two new Feature Extraction methods that are scalable for high dimensional data, large number of records, for both structured and unstructured. This can be used to reduce the number of dimensions to use as input to the data mining algorithms. The first of these is called Singular Value Decomposition (SVD) and is widely used in text mining. The second method can be considered a special scoring method of SVD is called Principal Component Analysis (PCA). With this method it produces projections that are scaled with the data variance.
  • A new feature of the GLM algorithm is that it will perform a feature section step. This is used to reduce the number of predictors used by the algorithm and allow for faster builds. This will makes the outputs more understandable and model more transparent. This feature is not default so you will need to set this on if you want to use it with the GLM algorithm.
  • In previous versions of the database, there could be some performance issues that relate to the data types used. In 12c these has been addressed for BINARY_DOUBLE and BINARY_FLOAT. So if you are using these data types you should now see faster scoring of the data in 12c
  • There is new in-database feature called Predictive Queries. This allows on-the-fly models that are temporary models that are formed as part of an analytics clause. These models cannot be tuned and you cannot see the details of the model produced. They are formed for the query and do not exist afterwards.
  • SELECT cust_id, age, pred_age, age-pred_age age_diff, pred_det FROM
    (SELECT cust_id, age, pred_age, pred_det,
    RANK() OVER (ORDER BY ABS(age-pred_age) DESC) rnk FROM
    (SELECT cust_id, age,
    PREDICTION(FOR age USING *) OVER () pred_age,
    PREDICTION_DETAILS(FOR age ABS USING *) OVER () pred_det
    FROM mining_data_apply_v))
    WHERE rnk <= 5;


  • There is a new function called PREDICTION_DETAILS. This allows you to see what the algorithm used to make the prediction. For example if we want to score a customer to see if they will churn, we can use the PREDICTION and PREDITION_PROBABILITY functions to do this and to see how how strong this prediction is. With PREDICTION_DETAILS we can now see what attributes and values the algorithm used to make that particular prediction. The output is in XML format.


  • image



These are the new in-database Advanced Analytics (Data Mining) features. Apart from the new algorithms or changes to them, most of the other changes gives greater transparency into what the algorithms/models are doing. This is good as it allows us to better understand and see what is happening.



The rest of the new Advanced Analytics Option new features will be part of Oracle Data Miner tool in SQL Developer 4. My next blog post will cover the new features in SQL Developer 4.



I haven’t mentioned anything about ORE. The reason for that is that it comes as a separate install and its current version 1.3 works the same in 11.2.0.3g as well as 12c. I’ve had some previous blog posts on this and you can check out the ORE website on OTN.

Monday, July 8, 2013

12c Roundup so far and Events

I’m on vacation at the moment. As a result I’ve missed all the 12c launch and excitement that goes with it. I’ve managed to get a few minutes to put this post together. The aim of this post is to list some interesting blog posts (by other people over the past few days). I intend to expand the list when I get time.

I also wanted to highlight two 12c launch events. The first of these is the official Oracle 12c webcast. It is on Wednesday 10th July. Click on the following image to register etc. The webcast will have Mark Hurd, Andy Mendelsohn and Tom Kyte.

image

The second 12c launch event will be hosted by Oracle in Ireland. This will be on the 5th September in the Gibson Hotel (Dublin) between  13:00 and 17:30.  I believe their might be some 12c goodies available for the attendees. Again click on the image below to register and to check out the agenda.

image

The following are some articles and blog posts that have been published since 12c has been launched. This is not a complete list or and indication of quality, but I’ve noted them for me to come back to after my vacation to read. You might have come across others. If so let me know and I will add them to the list.

12.1c Download page

12.1c Documentation page

12.1c New Features Guide

Oracle Advanced Analytics Option 12c and SQL Dev 4 new features

Oracle Database 12c: Oracle Multitenant Option

Oracle website for Multitenent 

New DB12c feature involves invisibility

12c - SQL Text Expansion

Ever expanding SQL for 12c

Oracle 12c Magazine by @leight0nn in Flipboard

How long can you hold off on Oracle 12c

Oracle 12c Install articles by Tim Hall (oraclebase) on Linux5 and Linux6

 

Over the coming weeks (after my vacation) I will be posting some articles on the Advanced Analytics Option in 12c. There are a number of new features. Also when SQL Developer 4 comes out I will be including all the new functionality that is included in the updated ODM tool.

Friday, June 7, 2013

DBMS_PREDICTIVE_ANALYTICS & Predict

In this blog post I will look at the PREDICT procedure that is part of the DBMS_PREDICTIVE_ANALTYICS package. This package allows you to perform data mining in an automated way without having to go through the steps of building, testing and scoring data.

I had a previous blog post that showed how to use the EXPLAIN function to create an Attribute Importance model.

The predictive analytics procedures analyze and prepare the input data, create and test mining models using the input data, and then use the input data for scoring. The results of scoring are returned to the user. The models and supporting objects are not persisted and are removed from the database when the procedure is finished.

The PREDICT procedure should only be used for a Classification problem and data set.

The PREDICT procedure create a model based on the supplied data (out input table) and a target value,  and returns scored data set in a new table. When using PREDICT you do not get to select an algorithm to use.

The input data source should contain records that already have the target value populated.  It can also contain records where you do not have the target value. In this case the PREDICT function will use the records that have a target value to generate the model. This model will then score all records a the predicted target value

The syntax of the PREDICT procedure is:

DBMS_PREDICTIVE_ANALYTICS.PREDICT (
   accuracy OUT NUMBER,
   data_table_name IN VARCHAR2,
   case_id_column_name IN VARCHAR2,
   target_column_name IN VARCHAR2,
   result_table_name IN VARCHAR2,
   data_schema_name IN VARCHAR2 DEFAULT NULL);

Where

Parameter Name Description
accuracy This output parameter from the procedure. You do not pass anything into this parameter. The Accuracy value returned is the predictive confidence of the model generated/used by the PREDICT procedure
data_table_name The name of the table that contains the data you want to use
case_id_column_name The case id for each record. This is unique for each record/case.
target_column_name The name of the column that contains the target column to be predicted
result_table_name The name of the table that will contain the results. This table should not exist in your schema, otherwise an error will occur.
data_schema_name The name of the schema where the table containing the input data is located. This is probably in your current schema, so you can leave this parameter NULL.

The PREDICT procedure will produce an output tables (result_table_name parameter) and will contain 3 attributes.

CASE_ID This is the Case Id of the record from the original data_table_name. This will allow you to link up the data in the source table to the prediction in the result_table_name
PREDICTION This will be the predicted value of the target attribute
PROBABILITY This is the probability of the prediction being correct

Using the sample example data set that I have given in previous blog posts and in the blog post on the EXPLAIN procedure, the following code illustrates how to use the PREDICT procedure.

set serveroutput on

DECLARE
   v_accuracy NUMBER(10,9);
BEGIN
   DBMS_PREDICTIVE_ANALYTICS.PREDICT(
      accuracy => v_accuracy,
      data_table_name => 'mining_data_build_v',
      case_id_column_name => 'cust_id',
      target_column_name => 'affinity_card',
      result_table_name => 'PA_PREDICT');
   DBMS_OUTPUT.PUT_LINE('Accuracy of model = ' || v_accuracy);
END;

image

This took about 15 seconds to run on my laptop, which is surprisingly quick given all the work that is doing internally. To see the predictions and the results from the PREDICT procedure, you will need to query the PA_PREDICT table.

image

The final step that you might be interested in is to compare the original target value with the prediction value.

SELECT v.cust_id,
       v.affinity_card,
       p.prediction,
       p.probability
FROM   mining_data_build_v  v,
       pa_predict p
WHERE  v.cust_id = p.cust_id
AND    rownum <= 12;

image

Remember we do not get to see how or what Oracle did to generate these results. We do not get the opportunity to tune the process and the model.

So you have to be careful when you use the PREDICT function and on what data. Would you use this as a way to explore your data and to see if predictive analytics/data mining might be useful for your? Yes it would. Would you use it in a production scenario? the answer is maybe but it depends on the scenario. In reality if you want to do this in a production environment you will put some work into developing data mining models that best fit your data. To do this you will need to move onto the ODM tool and the DBMS_DATA_MINING package. But the PREDICT function is a quick way to get some small data scored (in some way) based on your existing data. If your marketing department says they want to start a tele marketing campaign in a couple of hours then PREDICT is what you need to use. It may not give you the most accurate of results, but it does give you results that you can start using quickly.

Monday, May 20, 2013

DBMS_PREDICTIVE_ANALYTICS & Explain

There are 2 PL/SQL packages for performing data mining/predictive analytics in Oracle. The main PL/SQL package is DBMS_DATA_MINING. This package allows you to build data mining models and to apply them to new data. But there is another PL/SQL package.

The DBMS_PREDICTIVE_ANALYTICS package is very different to the DBMS_DATA_MINING package. The DBMS_PREDICTIVE_ANALYTICS package includes routines for predictive analytics, an automated form of data mining. With predictive analytics, you do not need to be aware of model building or scoring. All mining activities are handled internally by the predictive analytics procedure.

Predictive analytics routines prepare the data, build a model, score the model, and return the results of model scoring. Before exiting, they delete the model and supporting objects.

The package comes with the following functions: EXPLAIN, PREDICT and PROFILE. To get some of details about these functions we can run the following in SQL.

image

This blog post will look at the EXPLAIN function.

EXPLAIN creates an attribute importance model. Attribute importance uses the Minimum Description Length algorithm to determine the relative importance of attributes in predicting a target value. EXPLAIN returns a list of attributes ranked in relative order of their impact on the prediction. This information is derived from the model details for the attribute importance model.

Attribute importance models are not scored against new data. They simply return information (model details) about the data you provide.

I’ve written two previous blog posts on Attribute Importance. One of these was on how to calculate Attribute Importance using the Oracle Data Miner tool. In the ODM tool it is now called Feature Selection and is part of the Filter Columns node and the Attribute Importance model is not persisted in the database.  The second blog post was how you can create the Attribute Importance using the DBMS_DATA_MINING package.

EXPLAIN ranks attributes in order of influence in explaining a target column.

The syntax of the function is

DBMS_PREDICTIVE_ANALYTICS.EXPLAIN (
data_table_name IN VARCHAR2,
explain_column_name IN VARCHAR2,
result_table_name IN VARCHAR2,
data_schema_name IN VARCHAR2 DEFAULT NULL);


where


data_table_name = Name of input table or view



explain_column_name = Name of column to be explained



result_table_name = Name of table where results are saved. It creates a new table in your schema.



data_schema_name = Name of schema where the input table or view resides. Default: the current schema.



So when calling the function you do not have to include the last parameter.



Using the same example what I have given in the previous blog posts (see about for the links to these) the following command can be run to generate the Attribute Importance.



BEGIN

    DBMS_PREDICTIVE_ANALYTICS.EXPLAIN(


        data_table_name      => 'mining_data_build_v',


        explain_column_name  => 'affinity_card',


        result_table_name    => 'PA_EXPLAIN');


END;



One thing that stands out is that it is a bit slower to run than the DBMS_DATA_MINING method. On my laptop it took approx. twice to three time longer to run. But in total it was less than a minute.



To display the results,



image



The results are ranked in a 0 to 1 range. Any attribute that had a negative value are set to zero.

Thursday, May 16, 2013

Outputting your data using inbuilt SQL Dev formatting

Oracle has build a number of formatting options into SQL Developer to allow you to output your data in some standard formats. This removes the need to use other tools or to write extra code or performs various follow up steps.
All you need to do is to add a comment and use the Scrip button
SELECT /*csv*/ * FROM scott.emp;
SELECT /*xml*/ * FROM scott.emp;
SELECT /*html*/ * FROM scott.emp;
SELECT /*delimited*/ * FROM scott.emp;
SELECT /*insert*/ * FROM SCOTT.EMP;
SELECT /*loader*/ * FROM scott.emp;
SELECT /*fixed*/ * FROM scott.emp;
SELECT /*text*/ * FROM scott.emp;

Hint: for some of these it is best to list the schema and table name in upper case
These are comments and not hints so they will not work in SQL*Plus.

Friday, May 10, 2013

Getting Real Business Value from Oracle Data Mining and OBIEE

Over the past 16 months (or so) I have give a join presentation with Anthony Heljula called ‘Getting Real Business Value from Oracle Data Mining and OBIEE’, at a number of conferences and OUG SIGs.

We have had a lot of very positive feedback on this presentation. The presentation is a busy 45 minutes (questions only at the end) that walks through a pilot data science project we did for a University in the UK.

We used Oracle Data Miner to build a predictive model that looks at student churn. We then integrated this Student Churn model into OBIEE Dashboards to illustrate how combining an Oracle Data Miner model into our data analysis we can gain a greater insight of our data.

image

We have submitted this presentation for Oracle Open World 2013 but we have renamed the title of the presentation to

“How UK Universities are using Oracle Data Science to protect their income”

If you are involved in presentation selection or know someone who is then maybe you might select this to be presented at OOW13 in September.

We submitted the presentation for OOW12 with not luck. So fingers crossed this time.

Tuesday, April 23, 2013

Oracle buys Darwin back in 1999

The following is an extract from 1999 September/October edition of Oracle Magazine, about Oracle buying Thinking Machines. Their data mining software Darwin was integrated into the Oracle Database and renamed Oracle Data Miner.

Oracle Corporation’s recent acquisition of Thinking Machines’ data mining business extends Oracle’s data warehouse platform and business intelligence solution to include enterprise reporting, ad hoc query, advanced analysis and data mining software based on a common internet platform.

Oracle plans to incorporate the data mining software as an integral feature of Oracle Applications Customer Relationship Management site, which will facilitate the implementation of the e0business solutions developed by Oracle customers. In addition o the software technology, Oracle will receive rights to the domains think.com and thinkingmachines.com.

About Thinking Machines

Originally founded in 1983, Thinking Machines Corporation revolutionized high performance computing with its massively parallel supercomputing technology. The company has since evolved to focus exclusively on its Darwin data mining software for database marketing in the financial services and telecommunications industries. Darwin analyzes massive volumes of customer transaction, demographic and psychographic data, which can often amount to hundreds of millions of customer data records.

These advanced analyses help companies profile and target customers with greater accuracy, which allows companies to reduce customer attrition, assess customer profitability, cross sell to existing customers and detect fraud.

Darwin puts powerful data mining techniques in the hands of general business users and experienced analysts alike. Each to use wizards automate data mining while providing advanced users with full control over all options and parameters. The Darwin software combines advanced analytics - including neural networks, decisions trees and memory based reasoning, with impressive power and performance.

The solution’s one button model code generation, powerful scripting language and robust software development kit bring prediction capabilities to sales, call center, marking and the web.

Platforms and Languages

Darwin runs on Sun Microsystems and Hewlett-Packard servers and exports data mining models in C, C++ and Java for execution within Oracle Databases. A Microsoft Windows NT release is planned for later this year.”

Friday, April 19, 2013

Part 2–Getting start with Statistics for Oracle Data Science projects

This is the second blog on getting started with Statistics for Oracle Data Science projects.

In this blog post I will look at 3 more useful statistical functions that are available in the Oracle database. Remember these come are standard with the database. The first function I will look at is the WIDTH_BUCKET function. This can be used to create some histograms of the data. A common task in analytics projects is to produce some cross tabs of the data. Oracle has the STATS_CROSSTAB. The last function I will look the different ways you an sample the data.

Histograms using WIDTH_BUCKET

When exploring your data it is useful to group values together into a number of buckets. Typically you might want to define the width of each bucket yourself before passing the data into your data mining tools, but before you can decide what these are you need to do some exploring using a variety of widths. A good way to do this is to use the WIDTH_BUCKET function. This takes the following inputs:

Expression: This is the expression or attribute on which the you want to build the histogram.

Min Value: This is the lower or starting value of the first bucket

Max Value: This is the last or highest value for the last bucket

Num Buckets: This is the number of buckets you want created.

Typically the Min Value and the Max Value can be calculated using the MIN and MAX functions. As a starting point you generally would select 10 for the number of buckets. This is the number you will change, downwards as well as upwards, to if a particular pattern exists in the attribute.

Using the example scenario that I used in the first blog post, let us start by calculating the MIN and MAX for the AGE attribute.

image

Lets say that we wanted to create 10 buckets. This would create a bucket width of 7.3 for each bucket, giving us the following.

Bucket 1 : 17-24.3
Bucket 2: 24.3-31.6
Bucket 3: 31.6-38.8
Bucket 4: 38.8-46.1
Bucket 5: 46.1-53.4
Bucket 6: 53.4-60.7
Bucket 7: 60.7-68
Bucket 8: 68-75.3
Bucket 9: 75.3-82.6
Bucket 10: 82.6-90

These are the buckets that the WIDTH_BUCKET function gives us in the following:

SELECT cust_id,
       age,
       width_bucket(age,
                    (SELECT min(age) from mining_data_build_v),
                    (select max(age)+1 from mining_data_build_v),
                    10)  bucket
from mining_data_build_v
where rownum <=12
group by cust_id, age

image

An additional level of detail that is needed to allow us to plot the histograms for AGE, we need to aggregate up for all the records by bucket.

select intvl, count(*) freq
from (select width_bucket(salary,
(select min(salary) from employees),
(select max(salary)+1 from employees), 10) intvl
from HR.employees)
group by intvl
order by intvl;

image

We can take this code and embed it into the GATHER_DATA_STATS procedure that I gave in my Part 1 blog post.

Cross Tabs using STATS_CROSSTAB

Typically cross tabulation (or crosstabs for short) is a statistical process that summarises categorical data to create a contingency table. They provide a basic picture of the interrelation between two variables and can help find interactions between them.

Because Crosstabs creates a row for each value in one variable and a column for each value in the other, the procedure is not suitable for continuous variables that assume many values.

In Oracle we can perform crosstabs using one of their reporting tools. But if you don’t have one of these we will need to use the in-database function STATS_CROSSTAB. This function takes three parameters, the first two of these are the attributes you want to compare and the third is what test we want to perform. The tests available include:

  • CHISQ_OBS: Observed value of chi-squared
  • CHISQ_SIG: Significance of observed chi-squared
  • CHISQ_DF: Degree of freedom for chi-squared
  • PHI_COEFFICIENT: Phi coefficient
  • CRAMERS_V: Cramer’s V statistic
  • CONT_COEFFICIENT: Contingency coefficient
  • COHENS_K: Cohen’s kappa

CHISQ_SIG is the default.

Now let us look at some examples using our same data set.

image

Sampling Data

When our datasets are of relatively small size consisting of a few hundred thousand records we can explore the data is a relatively short period of time. But if your data sets are larger that that you may need to explore the data by taking a sample of it. What sampling does is that it takes a “random” selection of records from our data set up to the new number of records we have specified in the sample.

In Oracle the SAMPLE function takes a percentage figure. This is the percentage of the entire data set you want to have in the Sampled result. 

image

There is also a variant called SAMPLE BLOCK and the figure given is the percentage of records to select from each block.

image

Each time you use the SAMPLE function Oracle will generate a random seed number that it will use as a Seed for the SAMPLE function. If you omit a Seed number (like in the above examples), you will get a different result set in each case and the result set will have a slightly different number of records. If you run the sample code above over and over again you will see that the number of records returned varies by a small amount.

If you would like to have the same Sample data set returned each time then you will need to specify a Seed value. The Seed much be an integer between 0 and 4294967295.

image

In this case because we have specified the Seed we get the same “random” records being returned with each execution.

Thursday, April 11, 2013

Part 1–Getting started with Statistics for Oracle Data Science projects

With all analytics or data science projects one of the first steps typically involves the extraction of data from various sources, merging the data and then performing various statistics.

The extraction and merging of the data is well covered by lots of other people blogging about how to use Oracle Data Integration (ODI), Oracle Warehouse Builder (OWB), among many others.

What I’m going to look at in this series of blog posts will be what statistical functions you might look at using in the Oracle and how to use them.

  • This the first blog post in the series will look at the DBMS_STAT_FUNCS PL/SQL package, what it can be used for and I give some sample code on how to use it in your data science projects. I also give some sample code that I typically run to gather some additional stats.
  • The second blog post will look at some of the other statistical functions that exist in SQL that you will/may use regularly in your data science projects.
  • The third blog post will provide a summary of the other statistical functions that exist in the database.

These statistical functions can be grouped into 2 main types. The first is the descriptive statistics that are available by using the DBMS_STAT_FUNCS PL/SQL package and then there is the extensive list of other SQL stats functions.  It is worth mentioning at this point that all these stats packages and functions come as standard in the database (i.e. they are FREE, you do not have to pay for an add on option for the database to use them). So once you have you Oracle database installed you can start using them. There is no need to spend money buying another stats package to do stats. All you need to know is some SQL and what the stats functions are.

DBMS_STAT_FUNCS

One of the stats package that I use a lot is the SUMMARY function. This is part of the DBMS_STAT_FUNC PL/SQL package. This package calculates a number of common statistics for an attribute in a table. Yes that’s correct, it only gather statistics for just one attribute. So you will have to run it for all the numeric attributes in the table.

For does people who are familiar with the Oracle Data Miner tool, the explore data node produces a lot of these statistics that the SUMMARY function produces. See below for details of how to produce the Histograms.

The SUMMARY function has the following parameters

image

Although you will probably be running this this function on the data in your schema you still have to give the schema name. The table name is the name of the table where the data exists, the column name is the name of the column that contains the actual data you want to analyse, and the ‘s’ is the record that will be returned by the SUMMARY function that contains all the summary information.

An example of the basic script to run the SUMMARY function is given below. It will use the data that is available in the sample schemas and the views that where setup for the Oracle Data Mining sample schemas. The table (or in this case the view) that we are going to use is the MINING_DATA_BUILD_V. What we are going to do is to replicate some of what the Explore Node does in the Oracle Data Miner tool.

set serveroutput on

declare
   s         DBMS_STAT_FUNCS.SummaryType;
begin
 
   DBMS_STAT_FUNCS.SUMMARY('DMUSER', 'MINING_DATA_BUILD_V', 'AGE', 3, s);

   dbms_output.put_line('SUMMARY STATISTICS');
   dbms_output.put_line('Count  : '||s.count);
   dbms_output.put_line('Min    : '||s.min);
   dbms_output.put_line('Max    : '||s.max);
   dbms_output.put_line('Range  : '||s.range);
   dbms_output.put_line('Mean   : '||round(s.mean));
   dbms_output.put_line('Mode Count : '||s.cmode.count);
   dbms_output.put_line('Mode        : '||s.cmode(1));
   dbms_output.put_line('Variance    : '||round(s.variance));
   dbms_output.put_line('Stddev      : '||round(s.stddev));
   dbms_output.put_line('Quantile 5  : '||s.quantile_5);
   dbms_output.put_line('Quantile 25 : '||s.quantile_25);
   dbms_output.put_line('Median      : '||s.median);
   dbms_output.put_line('Quantile 75 : '||s.quantile_75);
   dbms_output.put_line('Quantile 95 : '||s.quantile_95);
   dbms_output.put_line('Extreme Count : '||s.extreme_values.count);
   dbms_output.put_line('Extremes      : '||s.extreme_values(1));
   dbms_output.put_line('Top 5 : '||s.top_5_values(1)||','||
                                                s.top_5_values(2)||','||
                                                s.top_5_values(3)||','||
                                                s.top_5_values(4)||','||
                                                s.top_5_values(5));
   dbms_output.put_line('Bottom 5 : '||s.bottom_5_values(5)||','||
                                                     s.bottom_5_values(4)||','||
                                                     s.bottom_5_values(3)||','||
                                                     s.bottom_5_values(2)||','||
                                                     s.bottom_5_values(1));
end;
/

image

We can compare this to what is produced by the Explore Node in ODM

image

image

We can see that the Explore Node gives us more statistics to help us with understanding the data.

What Statistics does the Explore Node produce

We can see the actual SQL code that the Explore Node runs to get the statistics that are displayed in the Explore Node View Data window. To do this you will need to right-click on the Explore Node and move the mouse down to the Deploy option. The submenu will open and select ‘SQL to Clipboard’ from the list. Open a text editor and past the code. You  will need to tidy up some of this code to point it at the actual data source you want. You will get the following

SELECT /*+ inline */  ATTR, 
       DATA_TYPE, 
       NULL_PERCENT, 
       DISTINCT_CNT, 
       DISTINCT_PERCENT, 
       MODE_VALUE,
       AVG,
       MIN,
       MAX,
       STD,
       VAR,
       SKEWNESS,
       KURTOSIS,
       HISTOGRAMS
FROM OUTPUT_1_23;

Where OUTPUT_1_23 is a working table that ODM has created to store intermediate results from some of its processing. In this case the Explore Node. You will need to change this to the ODM working table in your schema.

This query does not perform any of the statistics gathering. It just presents the results.

Creating our own Statistics gathering script – Part 1

The attribute names in the above SQL query tells us what statistics functions within Oracle that were used. We can replicate this statistics gathering task using the following script. There are four parts to this script. The first part gathers most of the common statistics for the attribute. The second and third parts calculate the Skewness and Kurtosis for the attribute. The final (fourth) part combines the first three parts and lists the outputs.

The one statistic function that we are not including at this point is the Histogram information. I will cover this in the next (second) blog post on statistics.

The following script has the data source table/view name included (MINING_DATA_BUILD_V) and the attribute we are going to use (AGE).  You will need to modify this script to run it for each attribute.

WITH
    basic_statistics AS (select (sum(CASE WHEN age IS NULL THEN 1 ELSE 0 END)/COUNT(*))*100 null_percent,
          count(*)    num_value,
          count(distinct age)   distinct_count,
          (count(distinct age)/count(*))*100     distinct_percent,
          avg(age)      avg_value,
          min(age)      min_value,
          max(age)     max_value,
          stddev(age)  std_value,
          stats_mode(age)   mode_value,
          variance(age)       var_value
        from   mining_data_build_v),
    skewness AS (select avg(SV) S_value
                 from (select power((age - avg(age) over ())/stddev(age) over (), 3) SV
                       from mining_data_build_v) ),
    kurtosis AS (select avg(KV) K_value
                 from (select power((age - avg(age) over ())/stddev(age) over (), 4) KV
                       from mining_data_build_v) )
SELECT null_percent,
       num_value,
       distinct_percent,
       avg_value,
       min_value,
       max_value,
       std_value,
       mode_value,
       var_value,
       S_value,
       K_value
from basic_statistics,
     skewness,
     kurtosis;

image

Part 2 – Lets do it for all the attributes in a table

In the code above I’ve shown how you can gather the statistics for one particular attribute of one table.But in with an analytics project you will want to gather the statistics on all the attributes.

What we can do is to take the code above and put it into a procedure. This procedure accepts a table name as input, loops through the attributes for this table and calculates the various statistics. The statistics are saved in a table called DATA_STATS (see below).

drop table data_stats;

create table DATA_STATS (
table_name VARCHAR2(30) NOT NULL,
column_name VARCHAR2(30) NOT NULL,
data_type VARCHAR2(106) NOT NULL,
data_length NUMBER,
data_percision NUMBER,
data_scale NUMBER,
num_records NUMBER,
distinct_count NUMBER,
null_percent NUMBER,
distinct_percent NUMBER,
avg_value NUMBER,
min_value NUMBER,
max_value NUMBER,
std_value NUMBER,
mode_value VARCHAR2(1000),
var_value NUMBER,
s_value NUMBER,
k_value NUMBER,
PRIMARY KEY (table_name, column_name));

This is one of the first things that I do when I start on a new project. I create the DATA_STATS table and run my procedure GATHER_DATA_STATS for each table that we will be using. By doing this it allows me to have a permanent records of the stats for each attribute and saves me time in having to rerun various stats at different points of the project. I can also use these stats to produces some additional stats or to produce some graphs.

He is the code for the GATHER_DATA_STATS procedure.

CREATE OR REPLACE PROCEDURE gather_data_stats(p_table_name IN varchar2) AS

   cursor c_attributes (c_table_name varchar2)
                       is SELECT table_name,
                                 column_name,
                                 data_type,
                                 data_length,
                                 data_precision,
                                 data_scale
                          FROM user_tab_columns
                          WHERE table_name = upper(c_table_name);

   v_sql     NUMBER;
   v_rows    NUMBER;
BEGIN
   dbms_output.put_line('Starting to gather statistics for '||upper(p_table_name)||' at '||to_char(sysdate,'DD-MON-YY HH24:MI:SS'));

   FOR r_att in c_attributes(p_table_name) LOOP
      --
      -- remove any previously generated stats
      --
      v_sql := DBMS_SQL.OPEN_CURSOR;
      DBMS_SQL.PARSE(v_sql, 'delete from DATA_STATS where table_name = '''||r_att.table_name||''' and column_name = '''||r_att.column_name||'''', DBMS_SQL.NATIVE);
      v_rows := DBMS_SQL.EXECUTE(v_sql);
--dbms_output.put_line('delete from DATA_STATS where table_name = '''||r_att.table_name||''' and column_name = '''||r_att.column_name||'''');

      IF r_att.data_type = 'NUMBER' THEN
         dbms_output.put_line(r_att.table_name||' : '||r_att.column_name||' : '||r_att.data_type);

         --
         -- setup the insert statement and execute
         --
         v_sql := DBMS_SQL.OPEN_CURSOR;
         DBMS_SQL.PARSE(v_sql, 'insert into data_stats select '''||r_att.table_name||''', '''||r_att.column_name||''', '''||r_att.data_type||''', '||r_att.data_length||', '||nvl(r_att.data_precision,0)||', '||nvl(r_att.data_scale,0)||', count(*) num_value, (sum(CASE WHEN '||r_att.column_name||' IS NULL THEN 1 ELSE 0 END)/COUNT(*))*100 null_percent, count(distinct '||r_att.column_name||') distinct_count, (count(distinct '||r_att.column_name||')/count(*))*100 distinct_percent, avg('||r_att.column_name||') avg_value, min('||r_att.column_name||') min_value, max('||r_att.column_name||') max_value, stddev('||r_att.column_name||') std_value, stats_mode('||r_att.column_name||') mode_value, variance('||r_att.column_name||') var_value, null, null from '|| r_att.table_name, DBMS_SQL.NATIVE);
         v_rows := DBMS_SQL.EXECUTE(v_sql);

      ELSIF r_att.data_type IN ('CHAR', 'VARCHAR', 'VARCHAR2') THEN
         dbms_output.put_line(r_att.table_name||' : '||r_att.column_name||' : '||r_att.data_type);

         --
         -- We need to gather a smaller number of stats for the character attributes
         --
         v_sql := DBMS_SQL.OPEN_CURSOR;
         begin

         DBMS_SQL.PARSE(v_sql, 'insert into data_stats select '''||r_att.table_name||''', '''||r_att.column_name||''', '''||r_att.data_type||''', '||r_att.data_length||', '||nvl(r_att.data_precision,0)||', '||nvl(r_att.data_scale,0)||', count(*) num_value, (sum(CASE WHEN '||r_att.column_name||' IS NULL THEN 1 ELSE 0 END)/COUNT(*))*100 null_percent, count(distinct '||r_att.column_name||') distinct_count, (count(distinct '||r_att.column_name||')/count(*))*100 distinct_percent, null, null, null, null, stats_mode('||r_att.column_name||') mode_value, null, null, null from '|| r_att.table_name, DBMS_SQL.NATIVE);
         v_rows := DBMS_SQL.EXECUTE(v_sql);

-- dbms_output.put_line('insert into data_stats select '''||r_att.table_name||''', '''||r_att.column_name||''', '''||r_att.data_type||''', '||r_att.data_length||', '||nvl(r_att.data_precision,0)||', '||nvl(r_att.data_scale,0)||', count(*) num_value, (sum(CASE WHEN '||r_att.column_name||' IS NULL THEN 1 ELSE 0 END)/COUNT(*))*100 null_percent, count(distinct '||r_att.column_name||') distinct_count, (count(distinct '||r_att.column_name||')/count(*))*100 distinct_percent, null, null, null, null, stats_mode('||r_att.column_name||') mode_value, null, null, null from '|| r_att.table_name);
         exception
         when others then
            dbms_output.put_line(v_rows);
         end;

      ELSE
         dbms_output.put_line('Unable to gather statistics for '||r_att.column_name||' with data type of '||r_att.data_type||'.');
      END IF;
   END LOOP;

   dbms_output.put_line('Finished gathering statistics for '||upper(p_table_name)||' at '||to_char(sysdate,'DD-MON-YY HH24:MI:SS'));
  
   commit;
END;

Then to run it for a table:

exec gather_data_stats('mining_data_build_v');

We can view the contents of the DATA_STATS table by executing the following in SQL*Plus or SQL Developer

select * from DATA_STATS;

image

Tuesday, March 5, 2013

Clustering in Oracle Data Miner–Part 4

This is a the fourth part of a five (5) part blog post on building and using Clustering in Oracle Data Miner. The following outlines the contents of each post in this series on Clustering.
  1. The first part we looked at what clustering features exist in ODM and how to setup the data that we will be using in the examples
  2. The second part will focus on how to building Clusters in ODM .
  3. The third post will focus on examining the clusters produced by ODM and how to use the Clusters to apply to new data using ODM.
  4. The fourth post will look at how you can build and evaluate a Clustering model using the ODM SQL and PL/SQL functions.
  5. The fifth and final post will look at how you can apply your Clustering model to new data using the ODM SQL and PL/SQL functions.
With Clustering we are trying to find hidden patterns in our data. Unlike classification we a not directing the algorithms on what areas/attributes to focus on.
In my preceding post on Clustering in Oracle Data Miner I showed how you can use the Oracle Data Miner (ODM) tool, that is part of SQL Developer, to build a cluster model and to apply it to new data.
As I have shown in previous blog posts we can do all of the same steps using SQL and PL/SQL. In this blog post I will show you how to to these SQL and PL/SQL functions to generate a Clustering model and how you can evaluate the model produced.
Step 1 – What Cluster Models to we already have
Oracle Data Miner comes with a number of data dictionary tables. We can use these to see what already exists in our schema. In the example below we will have a look to see the Cluster models that were produced in the Part 2 blog post.
format model_name format a20
format mining_function format a20
format algorithm format a20

SELECT model_name,
       mining_function, 
       algorithm, 
       build_duration, 
       model_size
FROM ALL_MINING_MODELS
WHERE mining_function = 'CLUSTERING';

image
We can also look at the model settings that ODM produced. We can look at the settings of the K-Means model. This model (in my case) is called CLUS_KM_1_25.
column setting_value format a20
select setting_name,
         setting_value,
         setting_type
from  all_mining_model_settings
where model_name = 'CLUS_KM_1_25'

image
We can also look to see what attributes are used in the K-Mean model.
SELECT attribute_name, 
       attribute_type, 
       usage_type, 
       target
from  all_mining_model_attributes
where model_name = 'CLUS_KM_1_25'

image
I will show you how to use this Clustering model or the one that we are about to generate in my next blog post.
Step 2 – Creating the Setting table
When creating an Oracle Data Mining model in SQL you need to create Setting table. This will contain a record for each setting they you need to set for the algorithm you want to run. The settings table has a very basic structure and only consists of two columns. It is a good idea to create a separate Settings table for each algorithm or type of data mining you want to do. To create the settings table we run the following:
CREATE TABLE CLUSTER_SETTINGS (
Setting_Name  VARCHAR2(30),
Setting_Value VARCHAR2(4000));

The next step is to define the setting we want to use. The first of these will be the actual algorithm you want to run. Oracle has two Clustering algorithms: algo_kmeans and algo_o_cluster. In the examples we are working through we are using the K-Mean algorithm and we want to set the number of clusters to 10. We are also going to set the Automatic Data Preparation (ADP) on. Here is the code.
BEGIN
   INSERT INTO CLUSTER_SETTINGS (setting_name, setting_value)
   VALUES (dbms_data_mining.algo_name, dbms_data_mining.ALGO_KMEANS);
   INSERT INTO CLUSTER_SETTINGS (setting_name, setting_value)
   VALUES (dbms_data_mining.prep_auto, dbms_data_mining.PREP_AUTO_ON);
   INSERT INTO CLUSTER_SETTINGS (setting_name, setting_value)
   VALUES (dbms_data_mining.clus_num_clusters, 10);
  
   COMMIT;
END;

We can check to see that these records were inserted.
column setting_value format a25
select * from CLUSTER_SETTINGS;
image
The other settings will be dependent on what clustering algorithm is being used. Each one has their own setting. If you do not define any additional setting Oracle will use the in-build default settings.
To see what the default settings:
column setting_value format a35
select *
from table(dbms_data_mining.get_default_settings)
where setting_name like 'KM%';

image
If you want to use a different value to the default, just create a new record in the CLUSTER_SETTINGS table with the new value.

Step 3 – Create the new Cluster Model
We have the algorithm settings ready, so the next steps is to create the Cluster model. We will be using the DBMS_DATA_MINING.CREATE_MODEL function and pointing it to our CLUSTER_SETTINGS table.
BEGIN
   DBMS_DATA_MINING.CREATE_MODEL(
      model_name          => 'CLUSTER_KMEANS_MODEL',
      mining_function     => dbms_data_mining.clustering,
      data_table_name     => 'INSURANCE_CUST_LTV',
      case_id_column_name => 'CUSTOMER_ID',
      target_column_name  => null,
      settings_table_name => 'CLUSTER_SETTINGS');
END;

This takes a few seconds to run on my laptop. When finished we can uses queries like we used in Step 1 above to find out the details of what was produced.
Step 4 – Examining the Cluster model
The first thing that we will need to check is that the model was created.
format model_name format a20
format mining_function format a15
format algorithm format a15

SELECT model_name, 
       mining_function,
       algorithm,
       build_duration,
       model_size
FROM ALL_MINING_MODELS
WHERE mining_function = 'CLUSTERING';

image
One thing that is a bit curious is the mode size. We have basically created two K-means models using the same settings, but the final models have different sizes. See below for an explanation.
Now lets look at the algorithms settings for our new Cluster model (Cluster KMeans Model)
select setting_name,
       setting_value,
       setting_type
from  all_mining_model_settings
where model_name = 'CLUSTER_KMEANS_MODEL'

image
We can also look at the attributes used in the clusters.
SELECT attribute_name, 
       attribute_type, 
       usage_type, 
       target
from  all_mining_model_attributes
where model_name = 'CLUSTER_KMEANS_MODEL'
image
If we compare this list back to the list of attributes that were part of the model that we developed in the ODM tool, you will see that we have some extra attributes listed for the new K-means model.
The reason for these extra attributes and the bigger size (although still small) will be due to us having applied a sampling in the ODM tool. We took a subset of the data in the ODM tool and built the model based on this data. For the model we just created in SQL we took all the data (no sampling). That is why it would have looked at including more attributes because it was looking at a much bigger data set, etc.
Step 5 – Examining the Cluster Details
In this step we are going to look at some of the clusters produced by our new Cluster model.
We can produce the information relating to the clusters that were produced and what level in the hierarchy each cluster belongs. We can use this information to create/draw our own cluster hierarchy diagram.
set lines 90
column child format a40

select id, 
       record_count, 
       parent, 
       tree_level, 
       child
from table(dbms_data_mining.get_model_details_km('CLUSTER_KMEANS_MODEL'))

image
To look at the Centroid details for a particular cluster e.g. Cluster ID = 7, we can run:
column attribute_name format a25
column mode_value format a25

select t.id,
       c.attribute_name, 
       c.mean,
       c.mode_value,
       c.variance
from table (dbms_data_mining.get_model_details_KM('CLUSTER_KMEANS_MODEL')) t,
     table(t.centroid) c
where t.id = 7
order by c.attribute_name

The results will contain details of the centroid for cluster number 7. The centroid details will include the attributes that were used to form the centroid and the values for each attribute.
image
The last thing we can look at is the rule for a cluster. The following query will give you the full details of the cluster rule for Cluster 7. This will give for each attribute what the values or ranges of values that is used for that cluster. I’ll let you run this as it returns a large number of records. Run it and have a look at some of the output.
select t.id, 
       a.attribute_name, 
       a.conditional_operator,
       nvl(a.attribute_str_value,
       a.attribute_num_value) AS value,
       a.attribute_support,
       a.attribute_confidence
from TABLE(dbms_data_mining.get_model_details_km('CLUSTER_KMEANS_MODEL'))  t, 
     TABLE(t.rule.antecedent)  a
where t.id = 7
ORDER BY t.id, a.attribute_name, attribute_support, attribute_confidence desc, value;


My next blog post on Clustering in Oracle, will be the final one in the series. It will show you how to apply a Clustering model to new data in your database.

Wednesday, February 27, 2013

Clustering in Oracle Data Miner–Part 3

This is a the third part of a five (5) part blog post on building and using Clustering in Oracle Data Miner. The following outlines the contents of each post in this series on Clustering.

  1. The first part we looked at what clustering features exist in ODM and how to setup the data that we will be using in the examples
  2. The second part will focus on how to building Clusters in ODM .
  3. The third post will focus on examining the clusters produced by ODM and how to use the Clusters to apply to new data using ODM.
  4. The fourth post will look at how you can build and evaluate a Clustering model using the ODM SQL and PL/SQL functions.
  5. The fifth and final post will look at how you can apply your Clustering model to new data using the ODM SQL and PL/SQL functions.

In my previous posts on Clustering in ODM we have setup our data, we have explored it, we have taken a sample of the data and used this sample as input to the Cluster Build node. Oracle Data Miner has two clustering algorithms and our Cluster Build node created a clustering model for each.

In this post we will look at the next steps. The first of these is that we will look at examining what clustering models ODM produced. In the second part we will look at how we can use one of these clustering models to apply and label new data.

image

Step 1 – View the Cluster Models

To view the the cluster modes we need to right click the Cluster Build node and select View Models from the drop down list. We get an additional down down menu that gives the names of the two cluster models that were developed.

In my case these are called CLUS_KM_1_25 and CLUS_OC_1_25. You may get different numbers on your model names. These numbers are generated internally in ODM

image

The first mode that we will look at will be the K-Mean Cluster Model (CLUS_KM_1_25). Select this from the menu.

Step 2 – View the Cluster Rules

The hierarchical K-Mean cluster mode will be displayed. You might need to readjust/resize some of the worksheets/message panes etc in ODM to get the good portion of the diagram to display.

image

With ODM you cannot change, alter, merge, split, etc any the clusters that were generated. Oracle take the view of, this is what we have found it it up to you now to decide how you are going to use it.

To see that the cluster rules are for each cluster you can click on a cluster. When you do this you should get a pane (under the cluster diagram) that will contain two tabs, Centroid and Cluster Rule.

The Centroid tab provides a list of the attributes that best define the selected cluster, along with the average value for each attribute and some basic statistical information.

image

The Cluster Rules tab contains a set of rules that define the cluster in a IF/THEN statement format.

image

For each cluster in the tree we can see the number of cases in each cluster the percentage of overall cases for this cluster.

Work your way down the tree exploring each of the clusters produced.

The further down the tree you go the smaller the percentage of cases will fall into each cluster. In some tools you can merge these clusters. Not so in ODM. What you have to do is to use an IF statement in your code. Something like IF cluster_num IN (16, 17, 18, 19) THEN …..

Step 3 – Compare Clusters

In addition to the cluster tree, ODM also has two addition tabs to allow us to explore the clusters. These are Detail and Compare tabs.

image

Click on the Detail tab. We now get a detailed screen that contain various statistical information for each attribute. We can for each attribute get a histogram of the values within each attribute for this cluster.

We can use this important to start building up a picture of what each cluster might represent based on the values (and their distribution) for each cluster.

image

Try this out for a few clusters.

Step 4 – Multi-Cluster - Multi-variable Comparison of Clusters

The next level of comparison and evaluation of the clusters can be found under the Compare tab.

This lets us compare two clusters against each other at an attribute level. For example let us compare cluster 4 and 9. The attribute and graphics section gets updated to reflect the data for each of cluster. These are colour coded to distinguish the two clusters.

image 

We can work our way down through each attribute and again we can use this information to help us to understand what each cluster might represent.

An additional feature here is that we can do multi-variable (attribute) comparison. Holding down the control button select LTV_BIN, SEX and AGE. With each selection we get a new graph appearing at the bottom of the screen. This shows the distribution of the values by attribute for each cluster.  We can learn a lot from this.

image

So one possible conclusion we could draw from this data would be that Cluster 4 could be ‘Short Term Value Customers’ and Cluster 9 could be ‘Long Term Value Customer’

Step 5 – Renaming Clusters

When you have discovered a possible meaning for a Cluster, you can give it a meaningful name instead of it having a number. In our example, we would like to re-label Cluster 4 to ‘Short Term Value Customers’. To do this click on the Edit button that is beside the drop down that has cluster 4. Enter the new label and click OK.

SNAGHTML1bf2277

In the drop down we will now get the new label appearing instead of the cluster number.

Similarly we can do this for the other cluster e.g. ‘Long Term Value Customer’.

image

 

We have just looked at how to explore our K-Means model. You can do similar exploration of the O-Cluster model. I’ll leave that for you to do.

 

We have now explored our clusters and we have decided which of our Clustering Models best suits our needs. In our scenario we are going to select the K-Mean model to apply and label our new data.

Step 1 – Create the Apply Node

We have already setup our sample of data that we are going to use as our Apply Data Set. We did this when we setup the two different Sample node.

We are going to use the Sample node that was set to 40%.

The first step requires us to create an Apply Node. This can be found under the Component Palette and Evaluate and Apply tab. Click on the Apply node and move the mouse to the workflow worksheet and click near the Sample Apply node.

To connect the two nodes, move the mouse to the Sample Apply node and right click. Select Connect from the drop down menu and then move the mouse to the Apply node and click again. An connection arrow will be created joining these nodes.

Step 2 – Specify which Clustering Model to use & Output Data

Next we need to specify which of the clustering model we want to use to apply to our new data.

We need to connect the Cluster Build node to the Apply node. Move the mouse to the Cluster Build node, right click and select connect from the drop down menu. Move the mouse to the Apply node and click. We get the connection arrow between the two node.

We now have joined the Data and the Cluster Build node to the Apply node.

The final step is to specify what clustering mode we would like to use. In our scenario we are going to specify the K-Mean model.

(Single) Click the Cluster Build node. We now need to use the Property Inspector to select the K-Means model for the apply set. In the Models tab of the Property Inspector we should have our two cluster models listed. Under the Output column click in the box for the O-Cluster model. We should now get a little red X mark appearing. The K-Mean model should still have the green arrow under the Output column.

Step 3 – Run the Apply Node

We have one last data setup to do on the Apply node. We need to specify what data from the apply data set we want to include in the output from the Apply node.  For simplicity we want to just include the primary key, but you could include all the attributes.  In addition to including the attributes from the apply data source, the Apply Node will also create some attributes based on the Cluster model we selected. In our scenario, the K-Means model will create two additional attributes. One of these will contain the Cluster ID and the other attribute will be the probability of the that cluster being valid.

To include the attributes from the source data, double click on the Apply node. This will open the Edit Apply Node window. You will see that it already contains the two attributes that will be created by the K-Mean model.

image

To add the attributes from the source data, click on the Data Columns tab and then click on the green ‘+’ symbol. For simplicity we are going to just select the CUSTOMER_ID. Click the OK button to finish.

image

Now we are ready to run the Apply node. To do this right click on the Apply Node and select Run from the drop down menu. When everything is finished you will get the little green tick mark on the top right hand corner of the Apply node.

image

Step 4 – View the Results

To view the results and the output produced by the Apply node, right click on the Apply node and select View Data from the drop down menu.

We get a new tab opened in SQL Developer that will contain the data. This will consist of the CUSTOMER_ID, the K-means Cluster ID and the Cluster Probability. You will see that the some of the clusters assigned will have a number and some will have the cluster labels that we assigned in a previous step.

image

It is now up to you to decide how you are going to use this clustering information in an operational or strategic way in your organisation.

 

In my next (fourth) blog post in the series on Clustering in Oracle Data Miner, I will show how you can perform similar steps, of building and evaluating clustering models, using the SQL and PL/SQL functions in the database. So we will not be using the ODM tool. We will be doing everything in SQL and SQL/PLSQL.