OUG Norway Agenda is now live

The OUG Norway spring conference (17th April – 19th April) agenda is now live and is open for registrations.

Click here for the Conference Agenda

Click here for the Conference Registration

This is a 3 day conference. The first day (17th April) will be held in the Radisson BLU Scandinavia ( Holbergsplass ) and the next two (and a bit) days will be on the Color Magic boat that will be travelling between Oslo and Kiel in Germany and back to Oslo. The boat will be arriving back in Oslo on the Saturday morning (20th April).

There will be some presentations in Norwegian, but it looks like most of the presentations will be in English. There will also be some well known names from the Oracle world presenting at this conference.

In addition to these people, I will be giving two presentations on using Predictive Analytics in Oracle using the Oracle Data Miner tool and in-database functionality.

My first presentation will be an overview of the advanced analytics option and a demonstration of what you can do using the Oracle Data Miner tool (part of SQL Developer). This presentation is currently scheduled for Thursday (18th April) at 5pm.

My second presentation will be at 9:30am on the Friday morning (19th April). In this presentation we will look at the in-database features, what can we do in SQL and PL/SQL, and we will look at what you need to do deploy you Oracle Data Mining models in a production environment.

If possible we might be able to review some new 12c new features for Oracle Data Miner

The ‘Oh No You Don’t’ of (Oracle) Data Science

Over the past couple of weeks I’ve had conversations with a large number of people about Data Science in the Oracle arena.

A few things have stood out. The first and perhaps the most important of these is that there is confusion of what Data Science actually means. Some think it is just another name for Statistics or Advanced Statistics, some Predictive Analytics or Data Mining, or Data Analysis, Data Architecture, etc.. The reality is it is not. It is more than what these terms mean and this is a topic for discussion for another day.

During these conversations the same questions or topics keep coming up and the simplest answer to all of these is taken from a Pantomime (Panto).

We need to have lots of statisticians
       'Oh No You Don't !'
We can only do Data Science if we have Big Data
        'Oh No You Don't !'
We can only do data mining/data science if we have 10’s or 100’s of Million of records
        'Oh No You Don't !'
We need to have an Exadata machine
        'Oh No You Don't !'
We need to have an Exalytics machine
        'Oh No You Don't !'
We need extra servers to process the data
        'Oh No You Don't !'
We need to buy lots of Statistical and Predictive Analytics software
        'Oh No You Don't !'
We need to spend weeks statistically analysing a predictive model
        'Oh No You Don't !'
We need to have unstructured data to do Data Science
        'Oh No You Don't !'
Data Science is only for large companies
        'Oh No You Don't !'
Data Science is very complex, I can not do it
        'Oh No You Don't !'

Let us all say it together for one last time ‘Oh No You Don’t’

In its simplest form, performing Data Science using the Oracle stack, just involves learning and using some simple SQL and PL/SQL functions in the database.

Maybe we (in the Oracle Data Science world and those looking to get into it) need to adopt a phrase that is used by Barrack Obama of ‘Yes We Can’, or as he said it in Irish when he visited Ireland back in 2011, ‘Is Feidir Linn’.

Remember it is just SQL.

My Blog Stats for 2012

Here are the stats from my blog for 2012.

In total I’ve had almost 28,000 blog post views. This is a 7 fold increase on the number of blog post views I had in 2011.

I had 92 blog posts in 2012 and the most popular blog posts were

• Celtic Knot Mirror Wood Carving
• Oracle Magazine Volume 1 Number 1
• Oracle Database next release (12c) new features
• Oracle Advanced Analytics Option in Oracle 12c
• Exalytics: How much will it cost me?
• Update on Exalytics Pricing
• Data Science is Multidisciplinary

Top search keywords used to find my blog

• exalytics pricing
• oracle data mining
• oracle data miner
• data science
• brendan tierney

Top Countries

• United States  52%
• Ireland  8%
• United Kingdom  8%
• India  4%
• Russia  4%
• Germany  3%
• France  3%
• Netherlands  1%
• Canada  1%
• Turkey  1%

Top OS

• Windows  59%
• Macintosh  28%
• Linux  5%
• iPhone  2%
• iPad  1%

Top Browsers

• Firefox  47%
• Internet Explorer  26%
• Chrome  15%
• Safari  4%

OUG Norway April 2013 - New Year’s News

I received an email at 23:24 on the 1st January from the OUG in Norway telling me that I’ve had two presentations accepted for the Annual OUG Norway seminar event. This will be on during the 17th-19th April.

The first day of this event (17th April) will be held in a hotel in Oslo. Then on the morning of 18th April we board the Color Magic cruise for the next two days of the conference. The ferry/cruise will go from Oslo to Kiel in Germany and then back again to Oslo, returning around 10am on Saturday 20th April.

I will be giving two presentations on the Oracle Advanced Analytics Option. The first presentation, ‘Using Predictive Analytics in Oracle’, will give an overview of the Oracle Advanced Analytics Option and will then focus on the Oracle Data Miner work-flow tool. This will presentation will include a live demo of using Oracle Data Miner to create some data mining models.

The second presentation, ‘How to Deploy and Use your Oracle Data Miner Models in Production’, builds on the examples given in the first presentation and will show how you can migrate, user and update your Oracle Data Miner models using the features available in SQL and PL/SQL. Again a demo will be given.

Association Rules in ODM-Part 4

This is a the final part of a four part blog post on building and using Association Rules in the Oracle Database using Oracle Data Miner. The following outlines the contents of each post in the series on Association Rules

1. This first part will focus on how to building an Association Rule model
2. The second post will be on examining the Association Rules produced by ODM – This blog post
3. The third post will focus on using the Association Rules on your data.
4. The final post will look at how you can do some of the above steps using the ODM SQL and PL/SQL functions.

In my previous posts I showed how you can go about setting up for Association Rule analysis in Oracle Data Miner and how to examine the rules that are generated.

This post will focus on how we build and use association rules using the functionality that is available in SQL and PL/SQL.

Step 1 – Build the Settings Table

As with all Oracle Data Mining functions in SQL and PL/SQL you will need to setup or build a settings table. This table contains all the settings necessary to run the model build functions. It is a good idea to create a separate settings table for each model build that you complete.

CREATE TABLE assoc_sample_settings (
setting_name VARCHAR2(30),
setting_value VARCHAR2(4000));

Step 2 – Define the Settings for the Model

Before you go to generate your model you need to set some of the parameters for the algorithm. To start with you need to defined that we are going to generate an Association Rules model, turn off the Automatic Data Preparation.

We can also set 3 additional settings for Association Rules.

The ASSO_MIN_SUPPORT has a default of 0.1 or 10%. That means that only rules that exist in 10% or more of the cases will be generated. This is really a figure that is too high. In the code below we will set this to a 1%. This matches the settings that we used in SQL Developer in my previous posts.

BEGIN

INSERT INTO assoc_sample_settings (setting_name, setting_value) VALUES

(dbms_data_mining.algo_name, dbms_data_mining.ALGO_APRIORI_ASSOCIATION_RULES);

INSERT into assoc_sample_settings (setting_name, setting_value) VALUES

(dbms_data_mining.prep_auto, dbms_data_mining.prep_auto_off);

INSERT into assoc_sample_settings (setting_name, setting_value) VALUES

(dbms_data_mining.ODMS_ITEM_ID_COLUMN_NAME, ‘PROD_ID’);

INSERT into assoc_sample_settings (setting_name, setting_value) VALUES

(dbms_data_mining.ASSO_MIN_SUPPORT, 0.01);

COMMIT;

END;

/

Step 3 – Prepare the Data

In our example scenario we are using the SALE data that is part of the SH schema. The CREATE_MODEL function needs to have an attribute (CASE_ID) that identifies the key of the shopping basket. In our case we have two attributes, so we will need to use a combined key. This combined key consists of the CUST_ID and the TIME_ID. This links all the transaction records related to the one shopping event together.

We also just need the attribute that has the information that we need. In our Association Rules (Market Basket Analysis) scenario, we will need to include the PROD_ID attribute. This contains the product key of each product that was included in the basket

CREATE VIEW ASSOC_DATA_V AS (

SELECT RANK() OVER (ORDER BY CUST_ID, TIME_ID) CASE_ID,


t.PROD_ID


FROM SH.SALES t );

Step 4 – Create the Model

We will need to use the DBMS_DATA_MINING.CREATE_MODEL function. This will use the settings in our ASSOC_SAMPLE_SETTINGS table. We will use the view created in Step 3 above and use the CASE_ID attribute we created as the Case ID in the function all.

BEGIN 
   DBMS_DATA_MINING.CREATE_MODEL( 
     model_name          => 'ASSOC_MODEL_2', 
     mining_function     => DBMS_DATA_MINING.ASSOCIATION, 
     data_table_name     => 'ASSOC_DATA_V', 
     case_id_column_name => ‘CASE_ID’, 
     target_column_name  => null, 
     settings_table_name => 'assoc_sample_settings');


END;

On my laptop this took approximately 5 second to run on just over 918K records involving just over 143K cases or baskets.

Now that is quick!!!

Step 5 – View the Model Outputs

There are a couple of functions that can be used to extract the rules produced in our previous step. These include:

GET_ASSOCIATION_RULES : This returns the rules from an association model.

SELECT rule_id, 
       antecedent, 
       consequent, 
       rule_support,


       rule_confidence


FROM TABLE(DBMS_DATA_MINING.GET_ASSOCIATION_RULES('assoc_model_2', 10));

The 10 here returns the top 10 records or rules. GET_FREQUENT_ITEMSETS : returns a set of rows that represent the frequent item sets from an association model. In the following code we want the top 30 item sets to be returned, but filtered to only display item sets where there are 2 or more rules.

SELECT itemset_id,

       items,


       support,


       number_of_items


FROM TABLE(DBMS_DATA_MINING.GET_FREQUENT_ITEMSETS('assoc_model_2', 30))


WHERE number_of_items >= 2;

Association Rules in ODM-Part 3

This is a the third part of a four part blog post on building and using Association Rules in Oracle Data Miner. The following outlines the contents of each post in the series on Association Rules

1. This first part will focus on how to building an Association Rule model
2. The second post will be on examining the Association Rules produced by ODM – This blog post
3. The third post will focus on using the Association Rules on your data.
4. The final post will look at how you can do some of the above steps using the ODM SQL and PL/SQL functions.

In my previous posts I showed how you can go about setting up for Association Rule analysis in Oracle Data Miner and how to examine the rules that are generated.

This post will focus on how we can extract and use these rules in Oracle Data Miner.

Step 1 – Model Details

Association Rules are an unsupervised method of data mining. In Oracle Data Miner we cannot use the Apply node to to score new data. What we have to do is to generate the Model Details. These in turn can then be used.

The Model Details node is used when we do unsupervised learning to extract the rules that are generated.

To do this we need to click on the Model Details node in the Models section of the Component Palette and then click on our workspace, just to the right of the Association Rule node.

The Edit Model Selection window will open. Connect the Association Rule node to the Model Details node. Then Run the node. This will then generate the Association Rules in a format what we can reuse.

When you get the small green tick on the Model Details node you can then view what was generated.

Right click on the Model Details node and click on View Details from the menu.

The output is similar to what we would have seen under the Association Rule node with the addition of a few more attributes that include the schema name and model name.

We can order the rules based on the Confidence level by double clicking on the Confidence column header. You might need to do this twice to get the rule appearing based on a descending confidence value.

At this point we can no look at persisting the Association Rules. See step 2 below.

We can also view the SQL that was used to generate the Association Rules that we see in the Model Details node. While still viewing the rules, click on the SQL tab.

Step 2 – Persisting the Association Rules

To make the rules persist and be useable outside of ODM we can persist the Association Rules in a table. The first step to do this is to create a new Table Node. This can be found under the Data section of the Component Palette. Click this Create Table or View node in the component palette and then click on the workspace, just to the right of the Model Details node.

Connect the Model Details node to the Output node, by right clicking on the Model Details node, select Connect from the menu and then click on the Output Node.

We can now edit the format of the Output i.e. specify what attributes are to be in our Output table. Double click on the Output node or right click and select Edit from the menu. We now get the Edit Create Table or View Node.

We can give the output a meaningful name e.g. AR_OUTPUT_RULES. We can also specify what rule properties we can to export to attributes in out table.

We will need to un-tick the Auto Input Columns Selection tick box before we can remove any of the output attributes. In my case I only want to have ANTECENDENT_ITEMS, CONSEQUENT_ITEMS, ID, LENGTH, CONFIDENCE and SUPPORT in my out put. So I need to select and highlight all the other attributes (holding the control button). After selecting all the attributes I do not want included in the final output table, I need to click on the red X icon.

When complete click on the OK button to go back to the workflow.

To generate the table right click on the AR_OUTPUT_RULES node and select Run from the menu. When you get the green tick mark on the AR_OUTPUT_RULES node the table has been created with records containing the details of each rules.

To view the contents of the AR_OUTPUT_RULES table we can right click on this node and select view data from the menu.

We can now use these rules in our applications.

 

Check out the next post in the series (Part 4) where we will look at the functionality available in the ODM SQL & PL/SQL functions to perform Association Rule analysis.

Association Rules in ODM–Part 2

This is a the second part of a four part blog post on building and using Association Rules in Oracle Data Miner.  The following outlines the contents of each post in the series on Association Rules

1. This first part will focus on how to building an Association Rule model
2. The second post will be on examining the Association Rules produced by ODM – This blog post
3. The third post will focus on using the Association Rules on your data.
4. The final post will look at how you can do some of the above steps using the ODM SQL and PL/SQL functions. 

In the previous post I looked at the steps needed to setup a data source and to setup the Association Rule node. When everything was setup we ran the workflow.

Step 1 – Viewing the Model

We the workflow has finished running we will have the green tick marks on each node. This is where we left thing at the end of the previous post (Part 1). To view the model details, right click on the Association Role Node and select View Models from the menu.

There are 3 main concepts that are important in relation to Association Rules:

• Support: is the proportion of transactions in the data set that contain the item set i.e. the number of times the rule occurs
• Confidence: is the proportion of the occurrences of the antecedent that result in the consequent e.g. how many times do we get C when we have A and B  {A, B} => C
• Lift: indicates the strength of a rule over the random co-occurrence of the antecedent and the consequent

Support and Confidence are the primary measures that are used to access the usefulness of an association rule.

In our example we can see that the the antecedent and the consequent has numbers separated by the word AND. These numbers correspond to the product numbers.

Step 2 – Examining the Model Rules

To read the antecedent and the consequent for the first rule in our example we have:

Antecedent: 137 AND 143 AND 128

Consequent: 144

To read this association rule we would say that if a Customer bought product 137 and product 143 and product 128, then we have a Confidence value of almost 71%. This is a strong association.

We can check the ordering of the rules by changing the Sort By criteria. As Confidence and Support are the main ways to evaluate the rules, we can change the Sort By criteria to be Confidence. Then click on the Query button to refresh the rules section.

Here get a list of the strongest rules listed in descending order.

Below the section of the screen that has the Rules, we have the Rule Details section.

 

Here we can see that the rule gets formatted into an IF statement. The first rule in the list has a confidence of almost 97%. As it is a simple IF statement it can be easily implemented in our applications.

We want use the information that these rules provides in a number of ways. One such consequence of these rules is that we can look at improving the ordering and distribution of these products to ensure that we have sufficient numbers of each. Another consequence is that we can enhance the front end selling mechanism to make sure that if a customer is buying product 114, 118 and 115 then we can remind the customer of product 119. We can also ensure that all these products are not located beside each other, so that the customer will have to walk past many other products in order to find them. That is why we never see milk and bread beside each other in a grocery store.

Step 3 – Applying Filters to the Model Rules

In the previous step we were able to sort our rules based on some of the measures of our Association Rules and to see how these rules are structured.

Association Rule Analysis can generate many thousands of possible rules for a small data set. In some cases the similar rules can appear and we can have lots of rules that occur so infrequently that they are perhaps meaningless.

ODM provides us with a number of filters that we can apply to the rules that enables use to look for the rules that are of must interest to use. We can access these filters by clicking on the More button, that is located just under the Query button.

We can refine our query on the rules based on the various measures and the number if items in the rule. In addition to this we can also filter based on the values of the items. This is particularly useful if we want to concentrate on specific items (in our example Products). To illustrate this use focus on the rules that involve Product 115. Click on the green + symbol on the right hand side of the window. Select 115 from the list provided. Next we need to decide if we want Product 115 involved in the Antecedent or the Consequent. In our example select the Consequent. This is located to the bottom right of the window. Then click the OK button and then click on the Query button to update the list of rules that correspond with the new filter.

We can see that we only have rules that have Product 115 in the Consequent column.

We can also see that we have 134 rules for this scenarios out of a total of 20,988 (your results might differ slightly to mine and that’s OK. It really depends on what version of the sample data you are using)

 

Check out the next post in the series (Part 3) where we will look at how you can use the Association Rules produced by ODM.

Association Rules in ODM–Part 1

This is a the first part of a four part blog post on building and using Association Rules in Oracle Data Miner. The following outlines the contents of each post in the series on Association Rules

1. This first part will focus on how to building an Association Rule model
2. The second post will be on examining the Association Rules produced by ODM – This blog post
3. The third post will focus on using the Association Rules on your data.
4. The final post will look at how you can do some of the above steps using the ODM SQL and PL/SQL functions.

The data set we will be using for Association Rule Analysis will be the sample data that comes with the SH schema in the database. Access to this schema and it’s data was setup when we created our data mining schema and ODM Repository.

Step 1 – Getting setup

As with all data mining projects you will need a workspace that will contain your workflows. Based on my previous ODM blog posts you will have already created a Project and some workflows. You can either reuse an existing workflow you have used for one of the other ODM modeling algorithms or you can create a new Workflow called Association Rules.

Step 2 – Define your Data Set

Assuming that your database has been setup to have the Sample schemas and their corresponding data, we will be using the data that is in the SH schema. In a previous post, I gave some instructions on setting up your database to use ODM and part of that involved a step to give your ODM schema access to the sample schema data.

We will start off by creating a Data Source Node. Click on the Data Source Node under the Component Palette. Then move your mouse to your your workspace area and click. A Data Source Node will be created and a window will open. Scroll down the list of Available Tables until you find the SH.SALES table. Click on this table and then click on the Next button. We want to include all the data so we can now click the Finish Button.

Our Data Source Node will now be renamed to SALES.

Step 3 – Setup the Association Build Node

Under the Model section of the Component Palette select Association. Move the mouse to your work area (and perhaps just the to right of the SALES node) click. Our Association Node will be created.

For the next step we need to join the our data source (SALES) with the Association Build Node. Right click on the SALES data node and select Connect from the drop down menu. Then move the mouse to the Association Build node and click. You should now have the two nodes connected.

We will now get the Edit Association Build Node property window opening for us. We will need to enter the following information:

• Transaction ID: This is the attribute(s) that can be used to uniquely identify each transaction. In our example the Customer ID and the Time ID of the transaction allows us to identify what we want to analyse by i.e. the basket. This will group all the related transactions together
• Item ID: What is the attribute of the thing you want to analyse. In our case we want to analyse the Products purchased, so select PROD_ID in this case
• Value: This is an identifier used to specify another column with the transaction data to combine with the Item ID. <Existence> means that you want to see if there are any type of common bundling among all values of the selected Item ID. Use this.

Like all data mining products, Oracle has just one Algorithm to use for Association Rule Analysis, the Apriori Algorithm.

Click the OK button. You are now ready to run the Association Build Node. Right click on the node and select Run from the menu. After a short time everything should finish and we will have the little green tick makes on each of the nodes.

 

Check out the next post in the series (Part 2) where we will look at how you can examine the rules produced by our model in ODM.

Events for Oracle Users in Ireland-November 2012

November (2012) is going to be a busy month for Oracle users in Ireland. There is a mixture of Oracle User Group events, with Oracle Day and the OTN Developer Days. To round off the year we have the UKOUG Conference during the first week in December.

Here are the dates and web links for each event.

Oracle User Group

The BI & EPM SIG will be having their next meeting on the Tuesday 20th November. This is almost a full day event, with presentations from End Users, Partners and Oracle product management. The main focus of the day will be on EPM, but will also be of interest to BI people.

As with all SIG meetings, this SIG will be held in the Oracle office in East Point (Block H). Things kick off at 9am and are due to finish around 4pm with plenty of tea/coffee and a free lunch too.

Remember to follow OUG Ireland on twitter using  #oug_ire

Oracle Day

Oracle will be having their Oracle Day 2012, on Thursday 15th, in Croke Park. Here is some of the blurb about the event,  “…to learn how Oracle simplifies IT, whether it’s by engineering hardware and software to work together or making new technologies work for the modern enterprise. Sessions and keynotes feature an elite roster of Oracle solutions experts, partners and business associates, as well as fascinating user case studies and live demos.”

This is a full day event from 9am to 5pm with 3 parallel streams focusing on Big Data, Enterprise Applications and the Cloud.

Click here to register for this event.

Click here for the full details and agenda.

OTN Developer Days

Oracle run their developer days about 3 times a year in Dublin. These events are run like a Hands-on Lab. So most of the work during the day is by yourself. You are provided with a workbook, a laptop and a virtual machine configured for the hands-on lab. This November we have the following developers days in the Oracle office in East Point, Dublin.

Tuesday 27th November (9:45-15:00) : Real Application Testing

Wednesday 28th November (9:00-14:00) : Partitioning/Advanced Compression

Thursday 29th November (9:15-13:30) : Database Security

Friday 30th November (9:45-16:00) : Business Process Management Using BPM Suite 11g

As you can see we have almost a full week of FREE training from Oracle. So there is no reason not to sign up for these days.

UKOUG Conference – in Birmingham

In December we have the annual UKOUG Conference. This is the largest Oracle User Group conference in Europe and the largest outside of the USA. At this conference you will have some of the main speakers and presentations from Oracle Open World, along with a range of speakers from all over the work.

In keeping with previous years there will be the OakTable Sunday and new this year there will be a Middleware Sunday. You need to register separately for these events. Here are the links

OakTable Sunday

Middleware Sunday

The main conference kicks off on the Monday morning with a very full agenda for Monday, Tuesday and Wednesday. There are a number of social events on the Monday and Tuesday, so come well rested.

On the Monday evening there is the focus pubs. This year it seems to have an Irish Pub theme. At the focus pub event there will be table for each of the user group SIGs. 

Come and join me at the Ireland table on the Monday evening.

The full agenda in now live and you can get all the details here.

I will be giving a presentation on the Tuesday afternoon titled Getting Real Business Value from Predictive Analytics (OBIEE and Oracle Data Mining). This is a joint presentation with Antony Heljula of Peak Indicators.

Oracle Advanced Analytics Option in Oracle 12c

At Oracle Open World a few weeks ago there was a large number of presentations on Big Data and Analytics.  Most of these were marketing type presentations, with a couple of presentations on using R and how it can not be integrated into the Oracle Database 11.2.

In addition this these there was one presentation that focused on the Oracle Advanced Analytics (OAA) Option.

The Oracle Advanced Analytics Option covers the Oracle Data Mining features and the Oracle R Enterprise features in the Database.

The purpose of this blog post is to outline and summarise what was mentioned at these presentations, and will include what changes are/may be coming in the “Next Release” of the database i.e. Oracle 12c.

Health Warning: As with all the presentations at OOW that talked about what may be in or may be in the next release, there is not guarantee that the features will actually be in the release version of the database. Here is the slide that gives the Safe Harbor statement.

• 12c will come with R embedded into it. So there will be no need for any configurations.
• Oracle R client will come as part of the server install.
• Oracle R client will be able to use the Analytics functions that exist in the database.
• Will be able to run R code in the database.
• The database (12c) will be able to spawn multiple R engines.
• Will be able to emulate map-reduce style algorithms.
• There will be new PREDICTION function, replacing the existing (11g) functionality. This will combine a number of steps of building a model and applying it to the data to be scored into one function.  But we will still need the functionality of the existing PREDICTION function that is in 11g. So it will be interesting to see how this functionality will be kept in addition to the new functionality being proposed in 12c.
• Although the Oracle Data Miner tool will still exits and will have many new features. It was also referred to as the ‘OAA Workflow’.  So those this indicate a potential name change?  We will have to wait and see.
• Oracle Data Miner will come with a new additional graphing feature. This will be in addition to the Explore Node and will allow us to produce more typical attribute related graphs. From what I could see these would be similar to the type of box plot, scatter, bar chart, etc. graphs that you can get from R.
• There will be a number of new algorithms too, including a useful One Class Support Vector Machine. This can be used when we have a data set with just one class value. This algorithm will work out what records/cases are more important and others.
• There will be a new SQL node. This will allow us to write our own data transformation code.
• There will be a new node to allow the calling of R code.
• The tool also comes with a slightly modified layout and colour scheme.

Again, the points that I have given above are just my observations. They may or may not appear in 12c, or maybe I misunderstood what was being said.

It certainly looks like we will have a integrate analytics environment in 12c with full integration of R and the ODM in-database features.

Extracting the rules from an ODM Decision Tree model

One of the most interesting of important aspects of a Decision Model is that we as a user can get to see what rules the machine learning algorithm has generated for our data.

I’ve give a number of examples in various blog posts over the past few years on how to generate a number of classification models. An example of the workflow is below.

In the Class Build node we get four models being generated. These include a Generalised Linear Model, Support Vector Machine, Naive Bayes and a Decision Tree model.

We can explore the Decision Tree model by right clicking on the Class Build Node, selecting View Models and then the Decision Tree model, which will be labelled with a ‘DT’ in the name.

As we explore the nodes and branches of the Decision Tree we can see the rule that was generated for a node in the lower pane of the applications. So by clicking on each node we get a different rule appearing in this pane

Sometimes there is a need to extract this rules so that they can be presented to a number of different types of users, to explain to them what is going on.

How can we extract the Decision Tree rules?

To do this, you will need to complete the following steps:

• From the Models section of the Component Palette select the Model Details node.
• Click on the Workflow pane and the Model Details node will be created
• Connect the Class Build node to the Model Details node. To do this right click on the Class Build node and select Connect. Then move the mouse to the Model Details node and click. The two nodes should now be connected.
• Edit the Model Details node, uncheck the Auto Settings, select Model Type to be Decision Tree, Output to be Full Tree and all the columns.

• Run the Model Details node. Right click on the node and select run. When complete you you will have the little green box with a tick mark, on the top right hand corner.
• To view the details produced, right click on the Model Details node and select View Data
• The rules for each node will now be displayed. You will need to scroll to the right of this pane to get to the rules and you will need to expand the columns for the rules to see the full details

My Presentations on Oracle Advanced Analytics Option

I’ve recently compiled my list of presentation on the Oracle Analytics Option. All these presentations are for a 45 minute period.

I have two versions of the presentation ‘How to do Data Mining in SQL & PL/SQL’, one is for 45 minutes and the second version is for 2 hour.

I have given most of these presentations at conferences or SIGS.

Let me know if you are interesting in having one of these presentations at your SIG or conference.

• Oracle Analytics Option - 12c New Features - available 2013
• Real-time prediction in SQL & Oracle Analytics Option - Using the 12c PREDICTION function - available 2013
• How to do Data Mining in SQL & PL/SQL
• From BIG Data to Small Data and Everything in Between
• Oracle R Enterprise : How to get started
• Oracle Analytics Option : R vs Oracle Data Mining
• Building Predictive Analysts into your Forms Applications
• Getting Real Business Value from OBIEE and Oracle Data Mining  (This is a cut down and merged version of the follow two presentations)
• Getting Real Business Value from OBIEE and Oracle Data Mining - Part 1 : The Oracle Data Miner part
• Getting Real Business Value from OBIEE and Oracle Data Mining - Part 2 : The OBIEE part
• How to Deploying and Using your Oracle Data Miner Models in Production
• Oracle Analytics Option 101
• From SQL Programmer to Data Scientist: evolving roles of an Oracle programmer
• Using an Oracle Oracle Data Mining Model in SQL & PL/SQL
• Getting Started with Oracle Data Mining
• You don't need a PhD to do Data Mining

Check out the ‘My Presentations’ page for updates on new presentations.

Using ODM Regression for the Leaning Tower of Pisa tilt problem

This blog post will look at how you can use the Regression feature in Oracle Data Miner (ODM) to predict the lean/tilt of the Leaning Tower of Pisa in the future.

This is a well know regression exercise, and it typically comes with a set of know values and the year for these values. There are lots of websites that contain the details of the problem. A summary of it is:

The following table gives measurements for the years 1975-1985 of the "lean" of the Leaning Tower of Pisa. The variable "lean" represents the difference between where a point on the tower would be if the tower were straight and where it actually is. The data is coded as tenths of a millimetre in excess of 2.9 meters, so that the 1975 lean, which was 2.9642.

Given the lean for the years 1975 to 1985, can you calculate the lean for a future date like 200, 2009, 2012.

Step 1 – Create the table

Connect to a schema that you have setup for use with Oracle Data Miner. Create a table (PISA) with 2 attributes, YEAR_MEASURED and TILT. Both of these attributes need to have the datatype of NUMBER, as ODM will ignore any of the attributes if they are a VARCHAR or you might get an error.

CREATE TABLE PISA
  (
    YEAR_MEASURED NUMBER(4,0),
    TILT          NUMBER(9,4)
);

Step 2 – Insert the data

There are 2 sets of data that need to be inserted into this table. The first is the data from 1975 to 1985 with the known values of the lean/tilt of the tower. The second set of data is the future years where we do not know the lean/tilt and we want ODM to calculate the value based on the Regression model we want to create.

Insert into DMUSER.PISA (YEAR_MEASURED,TILT) values (1975,2.9642);
Insert into DMUSER.PISA (YEAR_MEASURED,TILT) values (1976,2.9644);
Insert into DMUSER.PISA (YEAR_MEASURED,TILT) values (1977,2.9656);
Insert into DMUSER.PISA (YEAR_MEASURED,TILT) values (1978,2.9667);
Insert into DMUSER.PISA (YEAR_MEASURED,TILT) values (1979,2.9673);
Insert into DMUSER.PISA (YEAR_MEASURED,TILT) values (1980,2.9688);
Insert into DMUSER.PISA (YEAR_MEASURED,TILT) values (1981,2.9696);
Insert into DMUSER.PISA (YEAR_MEASURED,TILT) values (1982,2.9698);
Insert into DMUSER.PISA (YEAR_MEASURED,TILT) values (1983,2.9713);
Insert into DMUSER.PISA (YEAR_MEASURED,TILT) values (1984,2.9717);
Insert into DMUSER.PISA (YEAR_MEASURED,TILT) values (1985,2.9725);
Insert into DMUSER.PISA (YEAR_MEASURED,TILT) values (1986,2.9742);
Insert into DMUSER.PISA (YEAR_MEASURED,TILT) values (1987,2.9757);
Insert into DMUSER.PISA (YEAR_MEASURED,TILT) values (1988,null);
Insert into DMUSER.PISA (YEAR_MEASURED,TILT) values (1989,null);
Insert into DMUSER.PISA (YEAR_MEASURED,TILT) values (1990,null);
Insert into DMUSER.PISA (YEAR_MEASURED,TILT) values (1995,null);
Insert into DMUSER.PISA (YEAR_MEASURED,TILT) values (2000,null);
Insert into DMUSER.PISA (YEAR_MEASURED,TILT) values (2005,null);
Insert into DMUSER.PISA (YEAR_MEASURED,TILT) values (2010,null);
Insert into DMUSER.PISA (YEAR_MEASURED,TILT) values (2009,null);

Step 3 – Start ODM and Prepare the data

Open SQL Developer and open the ODM Connections tab. Connect to the schema that you have created the PISA table in. Create a new Project or use an existing one and create a new Workflow for your PISA ODM work.

Create a Data Source node in the workspace and assign the PISA table to it. You can select all the attributes..

The table contains the data that we need to build our regression model (our training data set) and the data that we will use for predicting the future lean/tilt (our apply data set).

We need to apply a filter to the PISA data source to only look at the training data set. Select the Filter Rows node and drag it to the workspace. Connect the PISA data source to the Filter Rows note. Double click on the Filter Row node and select the Expression Builder icon. Create the where clause to select only the rows where we know the lean/tilt.

Step 4 – Create the Regression model

Select the Regression Node from the Models component palette and drop it onto your workspace. Connect the Filter Rows node to the Regression Build Node.

Double click on the Regression Build node and set the Target to the TILT variable. You can leave the Case ID at <None>.  You can also select if you want to build a GLM or SVM regression model or both of them. Set the AUTO check box to unchecked. By doing this Oracle will not try to do any data processing or attribute elimination.

You are now ready to create your regression models.

To do this right click the Regression Build node and select Run. When everything is finished you will get a little green tick on the top right hand corner of each node.

Step 5 – Predict the Lean/Tilt for future years

The PISA table that we used above, also contains our apply data set

We need to create a new Filter Rows node on our workspace. This will be used to only look at the rows in PISA where TILT is null.  Connect the PISA data source node to the new filter node and edit the expression builder.

Next we need to create the Apply Node. This allows us to run the Regression model(s) against our Apply data set. Connect the second Filter Rows node to the Apply Node and the Regression Build node to the Apply Node.

Double click on the Apply Node.  Under the Apply Columns we can see that we will have 4 attributes created in the output. 3 of these attributes will be for the GLM model and 1 will be for the SVM model.

Click on the Data Columns tab and edit the data columns so that we get the YEAR_MEASURED attribute to appear in the final output.

Now run the Apply node by right clicking on it and selecting Run.

Step 6 – Viewing the results

Where we get the little green tick on the Apply node we know that everything has run and completed successfully.

To view the predictions right click on the Apply Node and select View Data from the menu.

We can see the the GLM mode gives the results we would expect but the SVM does not.

Data Science Is Multidisciplinary

[Update :October 2016.  There appears to be some discussion about the Venn diagram I've proposed below. The central part of this diagram is not anything I can up with. It was a commonly used Venn diagram for Data Mining. Thanks to Polly Michell-Guthrie for providing the original reference for the Venn. I just added the outer ring of additional skills needed for the new area of Data Science. This was just my view of things back in 2012. Things have moved on a bit since then]

A few weeks ago I had a blog post called Domain Knowledge + Data Skills = Data Miner.
In that blog post I was saying that to be a Data Scientist all you needed was Domain Knowledge and some Data Skills, which included Data Mining.
The reality is that the skill set of a Data Scientist will be much larger. There is a saying ‘A jack of all trades and a master of none’. When it comes to being a data scientist you need to be a bit like this but perhaps a better saying would be ‘A jack of all trades and a master of some’.
I’ve put together the following diagram, which includes most of the skills with an out circle of more fundamental skills. It is this outer ring of skills that are fundamental in becoming a data scientist. The skills in the inner part of the diagram are skills that most people will have some experience in one or more of them. The other skills can be developed and learned over time, all depending on the type of person you are.

Can we train someone to become a data scientist or are they born to be a data scientist. It is a little bit of both really but you need to have some of the fundamental skills and the right type of personality. The learning of the other skills should be easy(ish)
What do you think?  Are their Skill that I’m missing?

Domain Knowledge + Data Skills = Data Miner

Over the past few weeks I have been talking to a lot of people who are looking at how data mining can be used in their organisation, for their projects and to people who have been doing data mining for a log time.

What comes across from talking to the experienced people, and these people are not tied to a particular product, is that you need to concentrate on the business problem. Once you have this well defined then you can drill down to the deeper levels of the project. Some of these levels will include what data is needed (not what data you have), tools, algorithms, etc.

Statistics is only a very small part of a data mining project. Some people who have PhDs in statistics who work in data mining say you do not use or very rarely use their statistics skills.

Some quotes that I like are:

"Focus hard on Business Question and the relevant target variable that captures the essence of the question." Dean Abbott PAW Conf April 2012

"Find me something interesting in my data is a question from hell. Analysis should be guided by business goals." Colin Shearer PAW Conf Oct 2011

There has need a lot of blog posting and articles on what are the key skills for a Data Miner and the more popular Data Scientist. What is very clear from all of these is that you will spend most of your time looking at, examining, integrating, manipulating, preparing, standardising and formatting the data. It has been quoted that all of these tasks can take up to 70% to 85% of a Data Mining/Data Scientist time. All of these tasks are commonly performed by database developers and in particular the developers and architects involved in Data Warehousing projects. The rest of the time for the running of the data mining algorithms, examining the results, and yes some stats too.

Every little time is spent developing algorithms!!! Why is this ? Would it be that the algorithms are already developed (for a long time now and are well turned) and available in all the data mining tools. We can almost treat these algorithms as a black box. So one of the key abilities of a data miner/data scientist would be to know what the algorithms can do, what kind of problems they can be used for, know what kind of outputs they produce, etc.

Domain knowledge is important, no matter how little it is, in preparing for and being involved in a data mining project. As we define our business problem the domain expert can bring their knowledge to the problem and allows us separate the domain related problems from the data related problems. So the domain expertise is critical at that start of a project, but the domain expertise is also critical when we have the outputs from the data mining algorithms. We can use the domain knowledge to tied the outputs from the data mining algorithms back to the original problem to bring real meaning to the original business problem we are working on.

So what is the formula of skill sets for a data mining or data scientist. Well it is a little like the title of this blog;

Domain Knowledge + Data Skills + Data Mining Skills + a little bit of Machine Learning + a little bit of Stats = a Data Miner / Data Scientist

Oracle Advanced Analytics Video by Charlie Berger

Charlie Berger (Sr. Director Product Management, Data Mining & Advanced Analytics) as produced a video based on a recent presentation called ‘Oracle Advanced Analytics: Oracle R Enterprise & Oracle Data Mining’.

This is a 1 hour video, including some demos, of product background, product features, recent developments and new additions, examples of how Oracle is including Oracle Data Mining into their fusion applications, etc.

Oracle has 2 data mining products, with main in-database Oracle Data Mining and the more recent extensions to R to give us Oracle R Enterprise.

Check out the video – Click here.

Check out Charlie’s blog at https://blogs.oracle.com/datamining/

Oracle University : 2 Day Oracle Data Mining training course

ODM–Attribute Importance using PL/SQL API

In a previous blog post I explained what attribute importance is and how it can be used in the Oracle Data Miner tool (click here to see blog post).

In this post I want to show you how to perform the same task using the ODM PL/SQL API.

The ODM tool makes extensive use of the Automatic Data Preparation (ADP) function. ADP performs some data transformations such as binning, normalization and outlier treatment of the data based on the requirements of each of the data mining algorithms. In addition to these transformations we can specify our own transformations.  We do this by creating a setting tables which will contain the settings and transformations we can the data mining algorithm to perform on the data.

ADP is automatically turned on when using the ODM tool in SQL Developer. This is not the case when using the ODM PL/SQL API. So before we can run the Attribute Importance function we need to turn on ADP.

Step 1 – Create the setting table

CREATE TABLE Att_Import_Mode_Settings (
  setting_name  VARCHAR2(30),
  setting_value VARCHAR2(30));

Step 2 – Turn on Automatic Data Preparation

BEGIN
   INSERT INTO Att_Import_Mode_Settings (setting_name, setting_value)
   VALUES (dbms_data_mining.prep_auto,dbms_data_mining.prep_auto_on);
  COMMIT;
END;

^{Step 3 – Run Attribute Importance}

BEGIN
  DBMS_DATA_MINING.CREATE_MODEL(
    model_name => 'Attribute_Importance_Test',
    mining_function  => DBMS_DATA_MINING.ATTRIBUTE_IMPORTANCE,
    data_table_name  > 'mining_data_build_v',
    case_id_column_name => 'cust_id',
    target_column_name  => 'affinity_card',
    settings_table_name => 'Att_Import_Mode_Settings');
END;

Step 4 – Select Attribute Importance results

SELECT *
FROM TABLE(DBMS_DATA_MINING.GET_MODEL_DETAILS_AI('Attribute_Importance_Test'))
ORDER BY RANK;

ATTRIBUTE_NAME       IMPORTANCE_VALUE       RANK
-------------------- ---------------- ----------
HOUSEHOLD_SIZE             .158945397          1
CUST_MARITAL_STATUS        .158165841          2
YRS_RESIDENCE              .094052102          3
EDUCATION                  .086260794          4
AGE                        .084903512          5
OCCUPATION                 .075209339          6
Y_BOX_GAMES                .063039952          7
HOME_THEATER_PACKAGE       .056458722          8
CUST_GENDER                .035264741          9
BOOKKEEPING_APPLICAT       .019204751         10
ION

CUST_INCOME_LEVEL                   0         11
BULK_PACK_DISKETTES                 0         11
OS_DOC_SET_KANJI                    0         11
PRINTER_SUPPLIES                    0         11
COUNTRY_NAME                        0         11
FLAT_PANEL_MONITOR                  0         11

What has Oracle done to R to give us ORE

Oracle R Enterprise (ORE) was officially launched over the past couple of days and it has been receiving a lot of interest in the press.

We now have the Oracle Advanced Analytics (OAA) option which comprises, the already existing, Oracle Data Mining and now Oracle R Enterprise. In addition to the Oracle Advanced Analytics option we also 2 free set of tools available to use to use. The first of these free tools are the statistical functions which are available in all versions of the Oracle Database and the second free tool is the Oracle Data Miner tool that is part of the newly released SQL Developer 3.1 (7th Feb).

What has Oracle done to Oracle to make Oracle R Enterprise ?

The one of the main challenges with using R is that it is memory constrained, resulting in the amount of data that it can process. So the ORE development team have worked ensuring R can work transparently with data within the database. This removes the need extract the data from the database before it can be used by R. We still get all the advanced on in-Database Data Mining.

They have also embedded R functions within the database, so we an run R code on data within the database. By having these functions with the database, this allows R to use the database parallelism and so we get quicker execution of our code. Most R implementation are constrained to being able to process dataset containing 100Ks of records. With ORE we can now process 10M+ records

In addition to the ORE functions and algorithms that are embedded in the database we can also use the R code to call the suite of data mining algorithms that already exist as part of Oracle Data Miner.

For more details of what Oracle R Enterprise is all about check out the following links.

Oracle Advanced Analytics Options website

ORE Webpage

ORE Blog

ORE Download

ORE Forum

ODM 11gR2–Attribute Importance

I had a previous blog post on Data Exploration using Oracle Data Miner 11gR2. This blog post builds on the steps illustrated in that blog post.

After we have explored the data we can identity some attributes/features that have just one value or mainly one value, etc.  In most of these cases we know that these attributes will not contribute to the model build process.

In our example data set we have a small number of attributes. So it is easy to work through the data and get a good understanding of some of the underlying information that exists in the data. Some of these were pointed out in my previous blog post.

The reality is that our data sets can have a large number of attributes/features. So it will be very difficult or nearly impossible to work through all of these to get a good understanding of what is a good attribute to use, and keep in our data set, or what attribute does not contribute and should be removed from the data set.

Plus as our data evolves over time, the importance of the attributes will evolve with some becoming less important and some becoming more important.

The Attribute Importance node in Oracle Data Miner allows use to automate this work for us and can save us many hours or even days, in our work on this task.

The Attribute Importance node using the Minimum Description Length algorithm.

The following steps, builds on our work in my previous post, and shows how we can perform Attribute Importance on our data.

1. In the Component Palette, select Filter Columns from the Transforms list

2. Click on the workflow beside the data node.

3. Link the Data Node to the Filter Columns node. Righ-click on the data node, select Connect, move the mouse to the Filter Columns node and click. the link will be created

4. Now we can configure the Attribute Importance settings.Click on the Filter Columns node. In the Property Inspector, click on the Filters tab.

- Click on the Attribute Importance Checkbox

- Set the Target Attribute from the drop down list. In our data set this is Affinity Card

5. Right click the Filter Columns node and select Run from the menu

After everything has run, we get the little green box with the tick mark on the Filter Column node. To view the results we right clicking on the Filter Columns node and select View Data from the menu. We get the list of attributes listed in order of importance and their Importance measure.

We see that there are a number of attributes that have a zero value. It algorithm has worked out that these attributes would not be used in the model build step. If we look back to the previous blog post, some of the attributes we identified in it have also been listed here with a zero value.

ODM 11gR2–Real-time scoring of data

In my previous posts I gave sample code of how you can use your ODM model to score new data.

Applying an ODM Model to new data in Oracle – Part 2

Applying an ODM Model to new data in Oracle – Part 1

The examples given in this previous post were based on the new data being in a table.

In some scenarios you may not have the data you want to score in table. For example you want to score data as it is being recorded and before it gets committed to the database.

The format of the command to use is

prediction(ODM_MODEL_NAME USING <list of values to be used and what the mode attribute they map to>)

prediction_probability(ODM_Model_Name, Target Value, USING <list of values to be used and what model attribute they map to>)

So we can list the model attributes we want to use instead of using the USING *  as we did in the previous blog posts

Using the same sample data that I used in my previous posts the command would be:

Select prediction(clas_decision_tree
USING
20 as age,
'NeverM' as cust_marital_status,
'HS-grad' as education,
1 as household_size,
2 as yrs_residence,
1 as y_box_games) as scored_value
from dual;

SCORED_VALUE
------------
           0

Select prediction_probability(clas_decision_tree, 0
USING
20 as age,
'NeverM' as cust_marital_status,
'HS-grad' as education,
1 as household_size,
2 as yrs_residence,
1 as y_box_games) as probability_value
from dual;

PROBABILITY_VALUE
-----------------
                1

So we get the same result as we got in our previous examples.

Depending of what data we have gathered we may or may not have all the values for each of the attributes used in the model. In this case we can submit a subset of the values to the function and still get a result.

Select prediction(clas_decision_tree
USING
20 as age,
'NeverM' as cust_marital_status,
'HS-grad' as education) as scored_value2
from dual;

SCORED_VALUE2
-------------
            0

Select prediction_probability(clas_decision_tree, 0
USING
20 as age,
'NeverM' as cust_marital_status,
'HS-grad' as education) as probability_value2
from dual;

PROBABILITY_VALUE2
------------------
                 1

Again we get the same results.