Meta: information retrieval from Tweets

I pick significant problems randomly sometimes and enjoy solving them, or at least attempt designing api . Here’s one such problem!

Problem:

How’d you go about finding meta information about a person’s tweets?

NOTE: a) Tweet == Twitter updates b) Meta information –> Loosely defined. You can interpret it anyway you want –> Frequency, topics, followers whatever appeals to you.

Solution:

Meta information about a person’s tweets largely includes information that’s based on one or more of:
a) Content of the Tweet (Video, Text, Image, URL’s etc)
b) User Attribute (Age, Gender, Location, Followers, Following, Topics)
c) User Action (Retweet, Reply, Expand, Tag, Favorite, Follow, View Conversation, Browse Content (Click on URL and tags in a tweet)).

My approach is presented in terms of Design Considerations, Notes; Domain Model, Object Model, Persistence Model and Implementation Notes in the sections below.

Design Considerations:
The schema  design decisions are made to accommodate high scalability needs of twitter data that’s in (usually in the range GB’s or TB’s). For the purposes of analytics, the database operations are mostly read-only. The data is a live one. Also, for frequent reads and in general, large scale needs, the tables aren’t normalized and the redundant data can be easily spread across different tables.

Caching is implemented on the server side, for frequently accessed read operations. I’ve used Object Oriented approach using Java in my API design.

Design Notes:
The source of twitter data, can be :
a) Twitter dataset file in ZIP/XLS format (GB’s or TB’s of data provided by Twitter)
b) Twitter data stored in database tables on the public cloud.
c) Twitter data stored in our database tables (assuming we’re providing Twitter Type Infrastructure and Data Analytics).
d) Twitter response in XML format.

In case of (a), the dataset file can be read and mapped to our Java Object Model by parsing the files, tokenizing and using cache service for retrieval of the XLS data. In case of both (b) &(c), we can map our Object Model using an ORM framework to the data model. Implementation choices include Spring, Hibernate, JPA etc.

In case of (d), marshalling and unmarshalling of XML (using JAXB), is done to extract meta data.

I’m dealing with case (c) in my approach for the design of the persistence model. It’s assumed that the content of data, the capture of user actions(clicks) and user tweet profile attributes can be easily passed (using web 2.0 technologies like Jquery, HTML5, Ext-Js, DWR etc, to Java API which then talks to persistence layer to persist data in tables).

Domain Model:
Below is the domain model to represent business entities or real-world objects.

Object Model:
The object model for the domain model above is represented below using Class Diagrams.
Class Diagrams (without arrrow marks showcasing OO relationship though –  explained later)

Class <User>

-userId: String -name: String -gender: char -age: int -location: String -tweets: List -followers: List -following: List -lists: Set+addFollower(User user): boolean +getFollowers(): List +addFollowing(User user): boolean +getFollowing(): List +addToList(String listName): boolean +getLists(): Set

+tweet(String content): boolean
+tweets(): List<Tweet>

+retweet(String content): boolean
+favorite(String content): boolean
+reply(String content): boolean
+delete(String content): boolean
+expand(String content): boolean
+viewConversation(String convn): void

+getUserId(): String
+setUserId(String userId): void
+getName(): String
+setName(String userId): void
+getGender(): char
+setGender(char gender): void
+getAge(): int
+setAge(int age): void
+getLocation(): String
+setLocation(String location): void

Class <Tweet>

-tweetId: String -maxLength: int -contentType: ContentType -content: String -tags: List -postedBy: String -favoriteFreq: int -replyFreq: int -expandFreq: int -createDateTime: TimeStamp

+setRetweetFreq(): void +getRetweetFreq(): int+setPostedBy(String userId): void +getPostedBy(): String

+setContentType(ContentType ContentType): void
+getContentType(): ContentType

+setFavoriteFreq(): void
+getFavoriteFreq(): int
+setReplyFreq(): void
+getReplyFreq(): int
+setExpandFreq(): void
+getExpandFreq(): int

+setViewConversationFreq(): void
+getViewConversationFreq(): int

+addTag(Tag tag): boolean
+getTags(): List

Class <UserTweetCache>

-cacheInstance: UserTweetCache -userTweetMap: Map<string, set<tweet=”">> +lookUpTweet(String userId, String twtStr): Tweet +getUserTweetMap(): Map<string, set<tweet=”">> +getCacheInstance(): UserTweetCache

 

Class <ContentType>

-TweetContentType: enum{ TEXT, VIDEO, IMAGE, URL, COMPOSITE} +getValue(TweetContentType): String

 

Class <Tag>

-id: String -name: String -source: String+setTagId(String tagid): void +getTagId(): String +setSource(String source): void +getSource(): String +setName(String source): void +getName(): String

 

Class <UserListHelper>

-name: String +addUserToList(String userId, String listName): boolean +lists: Set<String>

 

Class <UserList>

-name: String -userListMap: Map<string, set<user=”">> -addUser(String userId): User +setName(String listName): void +getName(): String

 

interface <MetaData>

+extractMetaData(Set tweets): TweetAnalysisMetaData +MetaDataType: enum {USERACTION, CONTENT,USER_ATTRIBUTE }

 

UserAttributeBasedMetaData

+extractMetaData(Set<User> users): String

 

ContentBasedMetaData

+extractMetaData(Set<Tweet> tweets): String

 

UserActionBasedMetaData

+extractMetaData(Set<Tweet> tweets): String

 

class <TextAnalyzer>

+computeTermFreq: void +applyIDF(): void +tokenize(String tweetContent): List ..

 

class TweetContentMagnitudeVectorImpl

+normalize(TweetContentMagnitudeVector vec): void ..

 

class TweetContentMagnitudeVector

+getTFIDFValue(): double ..

 

class TextClusterer

+clusterTweets(Set<Tweet>): void +applyKMeansSetting(KMeansSetting kmeanSetting): void +computeSimilarityMatrix(): String[][] +printClusteStatistics(): void

 

class UserTweetPersistenceManager

+persistUser(User user): boolean +persistTweet(Tweet user): boolean +persistTag(Tag tag): boolean -retrieveUser(String tweetId): boolean -retrieveTweet(String userId): boolean

 

class TweetQueryResult

+getQueryCount(): int +getRelevantTweets(): Set

 

class TweetQueryParameter

+QueryParameter.Parameter: enum {KEY, APIUSER, START, INDEX, LIMIT, SORTBY} +QueryType.Typer: enum {UPDATE, SELECT, DELETE}+setParameter(QueryParameter queryParam): void +getParameter(QueryParameter queryParam): void +setType(QueryType queryType): void +getType(): QueryType

 

class TweetAnalysisMetaData

+displayMetaDataProfile(MetadataType): void+displayUserAttributeBasedMetaDataProfile(MetadataType.USER_ATTRIBUTE): void +displayContentBasedMetaDataProfile(MetadataType.CONTENT):void +displayUserActionBasedMetaDataProfile(MetadataType.USER_ACTION): void

 

Implementation Notes:
To avoid frequent calls to DB, Cache (UserTweetCache) is implemented to retrieve tweets and users from a Map which is updated for every new tweet or user and whose reference is got through UserTweetCache singleton instance. If using Map could be a concern for Memory Leaks, size can be specified for Map. Also, LinkedList can be a viable option with LRU type of Cache implementation (discard the LRU LinkedList objects from the cache).

To ensure data integrity for much of the data (we’re not really concerned if some of the unimportant data from the perspective of Cache, gets updated, which is not reflected in Cache), only after the records are successfully deleted or updated in DB, is the UserTweetCache Tweet / User object  updated accordingly.

 

Persistence Model:
Schema Design:
1. User

userid            varchar2(50) gender           varchar2(2) age                int(3) location         varchar2(30)

2. Tweet

tweetid           varchar2(50) content           varchar2(2) createdatetime TimeStamp(19) favFreq           int replFreq          int retweetFreq    int expandFreq     int viewConvFreq int

3. Tag

tagid           varchar2(20) tagtext        varchar2(50)

4. UserAction

actionId           int(2) description       varchar2(20) details             varchar2(40)

5. User_Tweet_Tag

userId                 varchar2(20) tweetId               varchar2(20) tagId                   varchar2(40) createDate         TimeStamp(19)

6. User_Tweet_UserAction

userId                 varchar2(20) tweetId               varchar2(20) actionId               int

7.Tweet_Tag

sourceid             varchar2(20) tweetId               varchar2(20) tagId                   varchar2(20) weight                 double(22) createDate        TimeStamp(19)

 

Saving Data:
To persist data from the update based on user action (tweet, retweet, reply, view, expand, favorite etc), UserTweetPersistenceManager encapsulates TweetQueryParameter, QueryType and TweetQueryResult objects that interact with ORM API’s to persist data in DB.

If update is successful, the UserTweetCache is updated with the modified Tweet object.
If update is a failure, the UserTweetCache is not updated with the modified Tweet object.
All the error conditions are handled to ensure data integrity of the Cache.

 

Retrieving Data:
Set and Set are populated with corresponding ResultSet objects retrieved using UserTweetPersistenceManager – retrieveUsers(tweetId), retrieveTweets(userId) etc.

To retrieve user-attribute based metadata, the Set<User> is passed to the MetaData extractor infrastructure.
To retrieve content-based metadata, Set<Tweet> is passed to the MetaData extractor infrastructure.

To retrieve user action based metadata, Set<Tweet> is passed to the MetaData extractor infrastructure.

The Tweet data for a user, or the User data for a tweet can be fetched by doing a join operation amongst User_Tweet_Tag, User, Tweet, User_Tweet_UserAction tables.

TweetAnalysisMetaData displays the metadata profile or statistics based on the type of MetaData (MetaDataType.USERACTION, MetaDataType.CONTENT, MetaDataType.USERATTRIBUTE or aggregating all of them).

TweetAnalysisMetaData also displays cluster statistics or profile for predictive modeling or data mining tasks. The mechanism used for Text Clustering and Text Mining is similar to that explained on my blog http://sanstechbytes.wordpress.com/2012/04/28/my-masters-dissertation-thesis-revisited-series-part-1/.

To be Continued…

Understanding Big Data

It’s been a while, since I last posted!

To keep this rolling (I’m hardly getting any time to post my own articles or stuff about my experiments these days  ), I just wanted to share this ebook on Big Data titled  Understanding Big Data: Analytics for Enterprise Class and Streaming Data.

Cheers!

Java Design Patterns: Template Pattern in the Real World

If you’re declaring some abstract methods in your class and making the sub classes provide concrete implementation for them, you’re indeed using Template pattern. Template pattern formalizes the idea of defining an algorithm in a class, but leaving some of the details to be implemented in sub classes.

In Java, Arrays.sort(Object[] o) internally invokes mergesort, that in turn, does a .compareTo() check for object comparison. .compareTo() implementation for the object type being compared, will be invoked.  AbstractCollection is one example, too.

Suppose, you’re designing a Sort API that you intend to ship as a jar file to your customer. Your customer would like the flexibility of adding newer or better sort implementations (specialized sorts are needed in a large scale system, applications) in their code. Often, you wouldn’t want to invoke default sort implementation of Java for large data. For instance, your customer wants to sort a very large array of 0′s and 1′s. They would write their own sort logic yielding in-place, O(n) time-complexity, rather than use default Arrays.sort(..) with O( n log n) in time. Obviously, they don’t want you to modify code for them every time. You want details of sort to be implemented in the sub classes that your customer writes. You would come up with something like this below:

public abstract class Sorter {
 public void sort(Object[] o) {
 sort1(o);
 }

 abstract void sort1(Object[] o);
 ...
}

class CustomSorter1 extends Sorter {

 public CustomSorter1() {

 }

 public void sort(Object[] o) {

 //Your sorting logic goes here

 }

}

Suppose, you’re developing a Graphics project for school. One of the tasks is to, drawing different triangles based on the input points. You would have subclasses like IsosCelesTriangle, EquilateralTriangle,  an abstract class Triangle and Graphics and Point classes.  Graphics and Point will be composite objects of Triangle.  drawLine() is common to all Triangle types. draw2ndLine() methods will be different for each Traingle type and implemented by sub classes.

Suppose, you’re building a Tourist Trip Reservation application. In every package trip, the activities are defined for each day of a Trip. Depending on the package type, the activities might vary. If you apply Template Pattern here: PackageTrip class is the abstract class. PackageA, PackageB etc are sub classes. performTrip() will be called by PackageA, PackageB etc. It would internally invoke specific Package implementations.

Template patterns is a one of the very useful and frequently used design patterns. Think about scenarios where you could apply in your project, now!

A Complete Singleton in Java

Singleton pattern is a design pattern that restricts the instantiation of a class to one object. In Java, the implemention should be done to ensure that only one instance of the class exists per class loader, per JVM. The objective in this post is not to explain much theory for scenarios where you apply Singleton etc or it’s implementation variants (read this for those details) but to explain what it takes to make a Singleton, a Complete Singleton. We all know we can use many variants of the so called Singleton implementations in our code. I wonder if all of us really know if our implementation will still be a ‘Singleton’ in situations where we try to invoke clone() and/or make the class Serializable. The class will no longer be a Singleton.

Let’s consider first the clone() scenario and later, serialization scenario. Let’s say your Singleton implementation is as below:

public final class Singleton implements Serializable {

 // static Singleton reference is initialised to refer to Singleton instance, when this class is first loaded
 // so that this instance can be accessible without creating an object and without having to worry about thread
 // synchronization in case of, lazy loading (where we do a null check inside getInstance()...

public static final Singleton instance= new Singleton(); // make it private as an optimization step.

 // Suppresses instantiation outside this class
 private Singleton() {

 }

 // One instance per class loader of this class, per JVM
 public static Singleton getInstance(){
return instance;
 }

}

Suppose the client API invokes method like below:

public class Client {
 ...
 Singleton instance = Singleton.getInstance();
 Singleton clone = (Singleton) instance.clone();
 ...
}

So,  clone() call creates another instance of Singleton in heap!!! But, we wanted to ensure ONLY one instance at any time!!! Singleton loses Singleton-ness property!!!!

To prevent this, you should override clone() like below:

// Prevents Singleton.clone() call by the client
public Object clone() throws CloneNotSupportedException {
  throw new CloneNotSupportedException();
}

When we do so, the client code will give a compile-time error (CloneNotSupportedException is a checked exception). Half Done!

Now, let’s say our Singleton implements Serializable(in above class). The above class would no longer be a singleton if the words “implements Serializable” were added to its declaration. It doesn’t matter whether the class uses the default serialized form or a custom serialized form, nor does it matter whether the class provides an explicit readObject() method. Any readObject method, whether explicit or default, returns a newly created instance, which will not be the same instance that was created at class initialization time. As Joshua Bloch mentions in his Effective Java: Prior to the 1.2 release, it was impossible to write a
serializable singleton class. In the 1.2 release, the readResolve() feature was added to the serialization facility. If the class of an object being deserialized defines a readResolve method with the proper declaration, this method is invoked on the newly created object after it is deserialized. The object reference returned by this method is then returned in lieu of the newly created object. In most uses of this feature, no reference to the newly created object is retained; the object is effectively stillborn, immediately becoming eligible for garbage collection.

private Object readResolve() throws ObjectStreamException {
 // Return the one true Singleton and let the garbage collector
 // take care of the Singleton impersonator.
 return instance;
}

Note that since the class is final, the contract says that the readResolve() has to be a private method. This method ignores the deserialized object, simply returning the distinguished Singleton instance created when the class was initialized. Therefore the serialized form of a Singleton instance need not contain any real data; all instance fields should be marked transient. This applies not only to this Singleton, but to all singletons.

So here’s the complete singleton code, a complete Singleton Code!!!

public final class Singleton implements Serializable {
 
<pre>
 // static Singleton reference is initialised to refer to Singleton instance, when this class is first loaded
 // so that this instance can be accessible without creating an object and without having to worry about thread
 // synchronization in case of, lazy loading (where we do a null check inside getInstance()...
 public static final Singleton instance= new Singleton(); // make it private as an optimization step.</p>
 // Suppresses instantiation outside this class
 private Singleton() {

 }

 // One instance per class loader of this class, per JVM
 public static Singleton getInstance(){
  return instance;
 }
 // Prevents Singleton.clone() call by the client
 public Object clone() throws CloneNotSupportedException {
  throw new CloneNotSupportedException();
 }

 private Object readResolve() throws ObjectStreamException {
  // Return the one true Singleton and let the garbage collector
  // take care of the Singleton impersonator.
  return instance;
 }

Here’s the Lazy Loading variant of the Singleton implementaion using Double-Checked Locking. Based on the Java Memory Model issues, there could be a Just-In-Time (JIT) run-time compiler Threat with this kind of implementation, so explains, Peter Haggar, of Practical Java™ Programming Language Guide fame, here.

</p>
public final class Singleton implements Serializable {

public static Singleton instance=null; // make it private as an  optimization step.

// Suppresses instantiation outside this class
private Singleton() {

}

//One instance per class loader of this class, per JVM
public static Singleton getInstance(){
if(instance == null) {
synchronized(Singleton.class) {
if(instance == null)
return instance;
}
}
 }

// Prevents Singleton.clone() call by the client
public Object clone() throws CloneNotSupportedException {
throw new CloneNotSupportedException();
}

private Object readResolve() throws ObjectStreamException {
// Return the one true Singleton and let the garbage collector
// take care of the Singleton impersonator.
return instance;
}

}

My Master’s Dissertation, Revisited Series – Part 3

In this final part of the series, let us look at the choices of implementation that I made and the results of my work. I thought I would rather share that as a stuff, part of the download.

Although I liked world-leading data mining open-source software like WEKA, Rapid Miner, I don’t recall why I stuck to exploring more of a trial version of commercial Oracle Data Mining Suite. We used Oracle 10g at work; we use SAS Enterprise Miner for data mining needs. Intuitively, it should have been with the intention of seeing how Oracle Data Mining fitted the bill in that context. It turns out that last statement was too much of a speculative future then.

The repository for all my work along with other useful artifacts can be downloaded here(around 93 MB, I think!!!).

When you’ve downloaded RAR file -

For my explanation of how I’ve used ODM API’s, refer to dissertationrepository\results\My Dissertation POC Demo.docx.

For detailed steps on installation of database, creating user, sample schemas etc.:  refer to Oracle Data Mining Administrator’s Guide can be located in RAR  file at: dissertationrepository\pdf\datamining_admin_guide.pdf.

For installing softwares like odminer and importing java project: refer to dissertationrepository\software. Perhaps, oracle database can be downloaded from here.

For screenshots and results of my POC work,  refer to:  dissertationrepository\results\My Dissertation POC Demo.docx

For other sample projects, books, ideas, papers etc: refer to dissertationrepository\extra

REFERENCES

1. Campos, M.M., Milenova, B.L., “O-Cluster: Scalable Clustering of Large High

Dimensional Data Sets”, Oracle Data Mining Technologies, 10 Van De Graaff

Drive, Burlington, MA 01803.

2. Oracle Data Miner –  http://www.oracle.com/technology/products/bi/odm/

3. JSR Specifications - http://jcp.org/en/jsr/detail?id_247

4. Introduction to Data Mining by Vipin Kumar, Michael Steinbach, Pang-ning Tan, Addison Wesley, 2006.

5. Java Data Mining: Standard, Theory and Practice: A Practical Guide for Architecture, Design and Implementation by Mark Hornick, Erik Marcade, and Sunil Venkayala, Morgan Kauffman Publishers, 2007

6. Collective Intelligence in Action by Satnam Alag, Manning Publications, 2007

7. Using text mining to understand the call center customers’ claims http://library.witpress.com/pages/PaperInfo.asp?PaperID=16715 by G. M. Caputo, V. M. Bastos & N. F. F. Ebecken.

8. Oracle Database 11g / Data Mining API Documentation

9.  KD Nuggets www.kdnuggets.com

10. ACM www.acm.org

My Master’s Dissertation, Revisited Series – Part 2:

In part 1, we had a quick overview of data mining and text (data) mining and some terminology used. Let’s explore clustering in this part, some of the clustering algorithms and text clustering using text mining.

In clustering, the objective is to partition unstructured set of objects to clusters or groups. We often want that the objects to be as similar as possible to objects in the same cluster and as dissimilar to objects from other clusters as possible.

Clustering vs. Categorization

By automatic categorization, we mean to let a machine decide to which of a set of predefined categories a text belongs. In clustering, the machine decides how a given text set should be partitioned. Categorization is suitable when one wants to categorize new texts according to a known categorization, clustering when one wants to discover new structures not previously known. Both methods may give interesting results on an unknown text set; categorization sorts them according to a well-known structure, clustering displays the structure of the particular set. This thesis deals with clustering of texts.

Let’s look at an example for visualization of Clusters for Customer groups based on their age and income, for simplicity. I think the diagrams (Courtesy: Java Data Mining : Strategy, Standard and Practice book by Hornick, Venkayala, Marcade) are pretty much self-explanatory.

Customer Clusters for above set of data:

Cluster Model – Histograms

 

Clustering is an unsupervised learning method. The result (the clustering, the partition) is based solely on the  object representation, the similarity measure and the clustering algorithm. If these correspond to the users understanding the result might well be an intuitive and useful clustering. One must keep in mind, though, that clustering algorithms always produce clusterings, even when this is not justified, and that there in most cases exist many relevant clusterings of a set of complex objects.

There are several different approaches to the computation of clusters. Clustering algorithms may be characterized as:

• Hierarchical — Groups data objects into a hierarchy of clusters. The hierarchy can be formed top-down or bottom-up. Hierarchical methods rely on a distance function to measure the similarity between clusters.

• Partitioning — Partitions data objects into a given number of clusters. The clusters are formed in order to optimize an objective criterion such as distance.

• Locality-based — Groups neighboring data objects into clusters based on local conditions.

• Grid-based — Divides the input space into hyper-rectangular cells, discards the low-density cells, and then combines adjacent high-density cells to form clusters.

Let’s look at some algorithms. Let’s explore basic and popular k-Means algorithm for clustering.

k-Means Algorithm:

The k-Means algorithm is a distance-based clustering algorithm that partitions the data into a predetermined number of clusters provided there are enough distinct cases. In the basic version of the k-means, we first choose K initial centroids, where K is a cluster specified parameter, viz., the number of clusters desired. Each point is then assigned to the closest centroid, and each collections of points assigned to a centroid is a cluster. The centroid of each cluster is then updated based on the points assigned to the the cluster. We repeat the assignment and update steps until no point changes clusters, or equivalently, until the centroids remain the same.

Distance-based algorithms rely on distance metric (function) to measure the similarity between data points. The distance metric is either Euclidean, Cosine, or Fast Cosine distance. Data points are assigned to the nearest cluster according to the distance metric used.

Basic k-means is formally described as below:

1. Select K points as initial Centroids

2. repeat:

3.    Form K clusters by assigining each point to it’s closest centroid..

4.    Recompute the centroid of each cluster.

5. until: Centroids do not change

The operation of above k-means algorithm is illustrated in below figure, which shows, how starting from three centroids, the final clusters are found in four assignment-update steps. In these and other figures displaying k-means clustering, each subfigure shows (1) the centroids at the start of the iteration and (2) the assignment of the points to those centroids. The centroids are indicated by the “+” symbol; all points belonging to the same cluster have the same marker shape.

(Courtesy: Introduction to Data Mining: Tan, Vipin Kumar, Steinbach)

Now, let’s look at enhanced k-Means algorithm supported by ODM for clustering.

Enhanced k-means -

• Builds models in a hierarchical manner. The algorithm builds a model top down using binary splits and refinement of all nodes at the end. In this sense, the algorithm is similar to the bisecting k-means algorithm. The centroid of the inner nodes in the hierarchy is updated to reflect changes as the tree evolves. The whole tree is returned.

• Grows the tree one node at a time (unbalanced approach). Based on a user setting available in either of the programming interfaces, the node with the largest variance is split to increase the size of the tree until the desired number of clusters is reached. The maximum number of clusters is specified in the build setting for clustering models, CLUS_NUM_CLUSTERS.

• Provides probabilistic scoring and assignment of data to clusters.

• Returns, for each cluster, a centroid (cluster prototype), histograms (one for each attribute), and a  rule describing the hyperbox that encloses the majority of the data assigned to the cluster. The centroid reports the mode for categorical attributes or the mean and variance for numerical attributes.

This approach to k-means avoids the need for building multiple k-means models and provides clustering results that are consistently superior to the traditional k-means.

I used Oracle Data Mining (ODM) API’s that use enhanced k-means algorithm described above, for k-means clustering. O-Cluster algorithm implementation is also provided in ODM for clustering.

Time and Space Complexity:

Since only the centroids and data points are stored, the storage required is O((m + K)n), where m is the number of points and n is the number of attributes. The time required (time complexity) is modest too, and can be determined to be: O(I * K * m * n), where I is the number of iterations required for convergence.

‘I’ is often small and can usually be safely bounded, as most changes typically occur in the first few iterations. Hence, K-means is linear in m, the number of points, provided K, no of clusters, is significantly less than m.

Text Clustering 

Text Clustering is a specific task in text mining that refers to the process of deriving high-quality information from text and is used for unsupervised document organization, automatic topic extraction and fast information retrieval or filtering.

The main applications of clustering in text mining are:
■  Simple clustering. This refers to the creation of clusters of text features .  For example: grouping the hits returned by a search engine.

■  Taxonomy generation. This refers to the generation of hierarchical groupings. For example: a cluster that includes text about car manufacturers is the parent of child clusters that include text about car models.

■  Topic extraction. This refers to the extraction of the most typical features of a group. For example: the most typical characteristics of documents in each document topic.

Let’s talk about Simple Clustering a bit. Suppose you’ve a call center web application and your call center reps enter the comments about a particular offering (product) or in general about the call center experience on your site, and those comments are recorded in your database tables. To make sense of this unstructured data,  i.e. to know which group your customers belong to – HAPPY, SATISFIED, UNHAPPY etc or customer’s level of acceptance about a particular product, and come up with some marketing strategy to improve  customer call center experience or product usability, you want to mine the comments (text) and apply k-means algorithm to group comments. Remember, we’re talking about hundreds of thousands of customers and comments and add to that unstructured data (Comments can contain alphabets, numbers, special chars, image etc). Any traditional smart code to retrieve data and group them, is not apt here and is out of contention, for such a scale.

To apply clustering algorithm for above scenario, say k-means, you need to have the text undergo a special pre-processing step known as Term Extraction or Feature Extraction. This process breaks the text down into units (terms) that can be mined. Text terms may be keywords or other document-derived features . The parameters like Semantic similarity, similar words stop words  and weighting schemes like TF-iDF may be considered while applying a text extraction algorithm like NMF (refer to part 1). You may want to refer to Text Mining section in part 1 of this series, for some theory.

I used ODM API’s. Oracle enhanced k-means algorithm (see below for this) supports text mining. Oracle Data Miner graphical tool performs term extraction transparently when you create or apply a text mining model. The Oracle Data Mining Java API provides term extraction classes for use in Java applications. If you are using the PL/SQL API, you can use a set of Oracle Text table functions to extract the terms. For screenshots of clustering model and Term Extraction model, please watch out for part 3 in this series.

Suppose you want to predict if customers will increase spending with an affinity card. You want to include the comments field from a customer survey as one of the predictors. Before building the classification model, you want to create a reduced set of text features to improve the predictive power of the comments. To do this, you will create a feature extraction model.

In the next part, we’ll look at ODM API’s and my dissertation repository for code, results, links and pdfs.

My Master’s Dissertation, Revisited Series – Part 1:

As I had mentioned in one of my previous posts that I would talk about my Master’s dissertation in my future posts, I’m happy that now is the time. The dissertation was titled “Text Clustering in a Call Centre Application using Data Mining”. My primary motivation was to explore data mining concepts and in particular clustering and see how it can connect to real-world problems in CRM domain. Being a java developer, I found it easier to quickly use Oracle Data Mining (ODM) API built on top of Java Data Mining (JSR Spec – http://www.jcp.org/en/jsr/detail?id=247) for my proof-of-concept.

In this three-part series of “My Master’s Dissertation, Revisited”, I’ll give a quick overview of data mining, text mining(in particular) and some terminology used in this series, in the first part. In part two of this series, I’ll talk about Text Clustering, enhanced k-means algorithm for clustering, and the sample that I used. I’ll wrap up this series by explaining as to how I applied k-means model, the code that I customized, and the results with a link to the sample code zip file.

I’ll be focusing on data mining concepts and functions relevant to my dissertation in this series. For info on other mining functions or more details, please refer to my previous post on the topic.

What is Data Mining? Why is it used?

Data mining is the practice of automatically searching large stores of data to discover patterns and trends that go beyond simple analysis and simple querying.  Data Mining is, in plain terms, making sense of the data. The key properties of data mining are:

i) Automatic discovery of patterns : 

Data mining is accomplished by building models. A model uses an algorithm to act on a set of data. The notion of automatic discovery refers to the execution of the data mining models.

ii) Prediction of likely outcomes:

A model might predict income based on education and other demographic factors. A data mining technique can generate a rule might specify that a person who has a bachelor’s degree and lives in a certain neighborhood is likely to have an income greater than the regional average.

iii) Creation of actionable information:

A town planner might use a model that predicts income based on demographics to develop a plan for low-income housing. A car leasing agency might a use model that identifies customer segments to design a promotion targeting high-value customers.

iv) Focus on large data sets and databases:

When you want to retrieve data from thousands or millions of records that have tens or hundreds of columns (attributes or dimensions) like transactional data (eg: customer data) from table, data mining models make lot of sense.

What it can do and what it cannot?

Data mining discovers hidden information in your data, but it cannot tell you the value of the information to your organization.  You might already be aware of important patterns as a result of working with your data over time. Data mining can confirm or qualify such empirical observations in addition to finding new patterns that may not be immediately discernible through simple observation.

Steps in Data Mining Process:

The steps in the process of data mining can diagrammatically be represented as:

1. Problem Definition:

This initial phase of a data mining project focuses on understanding the project objectives and requirements.

For example, your business problem might be: “How can I sell more of my product to customers?” You might translate this into a data mining problem such as: “Which customers are most likely to purchase the product?” A model that predicts who is most likely to purchase the product must be built on data that describes the customers who have purchased the product in the past.

2. Data Gathering and Preparation:

The data understanding phase involves data collection and exploration. As you take a closer look at the data, you can determine how well it addresses the business problem. You might decide to remove some of the data or add additional data. This is also the time to identify data quality problems and to scan for patterns in the data. For example, you might transform a DATE_OF_BIRTH column to AGE; you might insert the average income in cases where the INCOME column is null.

3. Model Building and Evaluation:

In this phase, you select and apply various modeling techniques and calibrate the parameters to optimal values. If the algorithm requires data transformations, you will need to step back to the previous phase to implement them.

In preliminary model building, it often makes sense to work with a reduced set of data (fewer rows in the case table), since the final case table might contain thousands or millions of cases. Based on the extent to which the model has to be improved, we can increase the number of attributes.

4. Knowledge Deployment:

Knowledge deployment is the use of data mining within a target environment. In the deployment phase, insight and actionable information can be derived from data.

Deployment can involve scoring (the application of models to new data), the extraction of model details (for example the rules of a decision tree), or the integration of data mining models within applications, data warehouse infrastructure, or query and reporting tools. For example, a sales representative could run a model that predicts the likelihood of fraud within the context of an online sales transaction.

Data Mining Functions

Data mining functions can be broadly classified into Supervised and Unsupervised mining functions. Each data mining function uses one or more algorithms to build a model.

The building of a supervised model involves training, a process whereby the software analyzes many cases where the target value is already known. In the training process, the model “learns” the logic for making the prediction. For example, a model that seeks to identify the customers who are likely to respond to a promotion must be trained by mining functions analyzing the characteristics of many customers who are known to have responded or not responded to a promotion in the past.

The building of an unsupervised model doesn’t involve training, as there is no previously-known result to guide the algorithm in building the model. For example, clustering models use descriptive data mining techniques, but they can be applied to classify cases according to their cluster assignments. Anomaly detection, although unsupervised, is typically used to predict whether a data point is typical among a set of cases.

Let’s have a quick look at how can these functions help a data miner achieve and the business context in which they’re used.

Association:

The goal of the Associations mining function is to find items that are consistently associated with each other in a meaningful way. For example, you can analyze purchase transactions to discover combinations of goods that are often purchased together. The Associations mining function answers the question: If certain items are present in a transaction, what other item or items are likely to be present in the same transaction?

Classification:

The process of automatically creating a model of classes from a set of records that contain class labels. The classification technique analyzes records that are already known to belong to a certain class, and creates a profile for a member of that class from the common characteristics of the records. You can then use a data mining scoring tool to apply this Classification model to new records, that is, records that have not yet been classified. This enables you to predict if the new records belong to that particular class.

Commercial applications of this mining function include credit card scoring, ranking of customers for directed mailing, the prediction of credit risk in new customers, and attrition prediction.

Regression:

Regression is similar to classification except for the type of the predicted value. Classification predicts a class label, regression predicts a numeric value. Regression also can determine the input fields that are most relevant to predict the target field values. The predicted value might not be identical to any value contained in the data that is used to build the model. An example application is customer ranking by expected profit.

Clustering:

The Clustering mining function includes algorithms like k-means, O-cluster etc. It groups similar customers together. At the same time it maximizes the differences between the different customer groups that are formed in this way.

The groups that are found are known as clusters. Each cluster tells a specific story about customer identity or behavior, for example, about their demographic background, or about their preferred products or product combinations. In this way, customers that are similar are grouped together in homogeneous groups that are then available for marketing or for other business processes.

In the commercial environment, clustering is used in the areas like: Cross-marketing, Cross-selling, Customizing marketing plans for different customer types, Deciding which media approach to use, Understanding shopping goals, and Many other areas.

Clustering can also be used for anomaly detection. Once the data has been segmented into clusters, you might find that some cases do not fit well into any clusters. These cases are anomalies or outliers.

The prerequisites to understand text clustering and in particular k-means algorithm as well as some other mining functions are: Understanding of Numerical, Categorical Attributes, Data Transformation techniques, Euclidean Distance, Document Matrix, Vector Space Model (VSM), Text Mining. See section ‘Some Terminology’ below.

Some Terminology:

Numerical Attribute: An attribute whose values are numbers. The numeric value can be either an integer or a real number.

Categorical Attribute: An attribute where the values correspond to discrete categories. For example, state is a categorical attribute with discrete values (CA, NY, MA, etc.). Categorical attributes are either non-ordered (nominal) like state, gender, and so on, or ordered (ordinal) such as high, medium, or low temperatures.

Binning:  This transformation maps a set of input values to a smaller set of bins. The input values may be discrete or continuous.

Normalization: It maps numerical values to a particular numerical range, typically [0…1]. There are several types of normalization (e.g., z-score, min-max, and shift-scale).

Explosion: This transformation applies only to discrete data such as strings or numbers representing individual categories. The goal is to transform such attributes into numerical attributes.

Support: The ratio of the number of occurrences of R, given all occurrences of all rules.

Confidence: The confidence of a rule X->Y, is the ratio of the number of occurrences of Y given X, among all other occurrences given X.

Euclidean Distance: The Euclidean distance or Euclidean metric is the “ordinary” distance between two points that one would measure with a ruler, and is given by the Pythagorean formula. In Cartesian coordinates, if p = (p1, p2,…, pn) and q = (q1, q2,…, qn) are two points in Euclidean n-space, then the distance from p to q, or from q to p is given by:

d(p, q) = d(q, p) = √ (q1 – p1)2 + (q2 – p2)2  + ….. + (qn – pn)2  =  √∑( qi – pi)2

For N dimensions

In general, for an n-dimensional space, the distance is:

d(p, q) = √ (p1 – q1)2 + (p2 – q2)2  + … + (pi – qi)2  + … + (pn – qn)2  =  √∑( qi – pi)2

Vector Space Model: In the vector space model each text in a set of texts is represented by a vector in a high-dimensional space, with as many dimensions as the number of different words in the set. Each text gets weights (values) in the indices (dimensions) based on what words appear in them. These weights model how important the corresponding word is deemed to be to explain the content of the text. They are dependent on whether (and how often) the word appears in the document and in the entire set. Texts whose vectors are close to each other in this space are considered being similar in content.

Text Mining

Text mining, sometimes alternately referred to as text data mining, roughly equivalent to text analytics, refers to the process of deriving high-quality information from text. ‘High quality’ in text mining usually refers to some combination of relevance, novelty, and interestingness. Typical text mining tasks include text categorization, text clustering, concept/entity extraction, production of granular taxonomies, sentiment  analysis, document summarization, and entity relation modeling (i.e., learning relations between named entities).

Unstructured data like ppt, audio, images, call center notes etc can be converted into numerical attribute using Text Mining. The term extraction process might involve application of NMF (Non-Negative Matrix Factorization) algorithms and treats the text in each row of the original table as a separate document. Each document is transformed to a set of terms that have a numeric value and a text label.

Most of the text-mining methods are grounded in the term-based VSM to calculate the distance or similarity between documents. Let’s look at some definitions for using document representation of vector.

A document is commonly represented as a vector of terms in a VSM. The basis of the vector collection. In a simple way, we can use a word to be a term. Yet, morphological variants like ’actor’, ’action’ and ’acting’ are so closely related that they are usually conflated into a single word stem, e.g., ’act’ by stemming [4], [5]. After stemming, two word stems are treated as unrelated if they are different. For example, the stem of ’suggestion’ and ’advice’ are usually considered unrelated despite of their apparent relationship. Since different word stems are considered unrelated, i.e., independent, the base vectors in VSM are orthogonal to each other.

A set of documents X in traditional term-based VSM is represented by X = {X1, X2…… Xn}. One document is represented by Xj = {xj1, xj2,…..xj,m}. Terms in the vocabulary extracted from the corresponding document collection are represented by {t1, t2…tm}.

Where xjii is the weight of the term tii in document Xj and is usually determined by the number of times ti appears in Xj (known as the term frequency). Other weighting scheme like TF-iDF can also be applied.

Based on the definition of VSM, distance between two text documents X1 and X2 can be easily computed. There are many methods to measure this distance, such as: Cosine similarity and Minkowski distance including Euclidean distance, Manhattan distance and Maximum distance. Here, we give the definition of Euclidean distance which is effective and frequently used in text clustering.

Euclidean distance of two documents X1 and X2  in the VSM, is defined as:

D(X1, X2) = √( X1 – X2 ) ( X1 – X2 )   T

     = √i=1∑m( x1i-x2i)2

The smaller the value of d(X1, X2) is, the more similar the two documents are. From above equation, we can see that this distance definition does not take into account any patterns of term correlation that exist in  the real-world data.

Number to Text Representation Problem

Problem:

Given a number (up to a billion), display the string (text) representation of it. For eg.: 123456789 should be represented as “One Hundred and Twenty Three Million Four Hundred and Fifty Six thousand Seven Hundred and Eighty Nine. ‘and’ should only be used after hundred.

Solution:

Approach 1: In a OO approach, a number can be represented by Number class. Each Number object has a list of Triplet objects (composition). Triplet represents a group of max 3 digits and can’t exist without a Number.  With this approach, it’s easy to accommodate different text representation systems like British English International Unit (million etc), British English Indian Units (lac, crore).

I wrote the classes (see below) for both approaches including JUnit Tests. I’ve captured the outputs for sample runs in the project folder that you can download here.

Approach 2: Assume Text representation in British English. Think of converting 123456789 into “One Hundred and Twenty Three Million Four Hundred and Fifty Six thousand Seven Hundred and Eighty Nine.

So, convert(123456789) = convert(123) + “million” + convert(456) + “thousand” + convert(789). [Courtesy: Gayle Laakmann of careercup.com]. This makes use of the technique of recursion. A lot of new code has to be written to extend this behavior though. There’re too many special cases to handle.

Code: Approach 1:

package numbertextoop;

import java.util.LinkedList;
import java.util.List;

import utils.NumberTextUtils;

/**
* This class represents a given number. Can be used to display in British
* English phrase the number Eg: 1234 in British English phrase: One Thousand
* Two Hundred and Thirty Four This class can accommodate text representation
* types like British, Indian etc with the constructor provided 04/14/2012 @author
* sanjeev 04/19/2012 @author sanjeev
*
* @version 1.1
*/

public class Number {

private int num;
private List<Triplet> triplets;

/**
* Assume default text representation type as British
*/
public Number(int i) {

this.num = i;

/**
* Handle negative number
*/
if (i < 0) {
i = -1 * i;
}

int length = Integer.toString(i).length();

if (length % NumberTextUtils.THREE == 0) {
length = length / NumberTextUtils.THREE;
} else if (length % NumberTextUtils.THREE > 0) {
length = length / NumberTextUtils.THREE + 1;
}

this.triplets = new LinkedList<Triplet>();
buildTriplets(i, length);
}

/**
* Allows for specifying text representation system, British, Indian etc
* text representationType could be enum - TextRepresentationType.BRITISH,
* TextRepresentationType.INDIAN. In Case of 'Indian', constants like Lac,
* Crore, should be added appropriately in NumberTextUtils.java
*/
public Number(int i, String representationType) {

}

/**
* @param num
* @param length
*            Builds triplet objects and arranges the position (hundredth,
*            tenth, unit) of the digits forming a triplet
*/
private void buildTriplets(int num, int length) {
while (length != 0) {
int tripletValue = num % 1000;

Triplet triplet = new Triplet(tripletValue);
triplet = triplet.matchTripletPlaces(tripletValue);

System.out.println("triplet" + triplet.getHundredthPlace() + "."
+ triplet.getTenthPlace() + ":" + triplet.getUnitPlace());
((LinkedList<Triplet>) this.triplets).addFirst(triplet);

num = num / NumberTextUtils.THOUSAND;
length--;
}
}

/**
* String representation of this number
*
*/
public String toString() {

/**
* When num = 0, returns "Zero" or if accessed thro
*
*/
if (this.num == 0) {
return NumberTextUtils.ZERO_TEXT;
}

StringBuilder numTextBuilder = new StringBuilder();

/**
* Negative number
*
*/
if (this.num < 0) {
numTextBuilder.append("Negative ");
}

return constructEnglishText(this.triplets, numTextBuilder).toString();
}

/**
* Constructs the British English text from this group of triplets formed
* from
*
* @param num
* @param triplets
*            List
*/
private StringBuilder constructEnglishText(List<Triplet> triplets,
StringBuilder numTextBuilder) {
int mapIndex = triplets.size();

for (Triplet triplet : triplets) {
numTextBuilder = numTextBuilder.append(triplet.toString()).append(
" ");

/**
* When num = 1,000,000: print only one million
*/
if (triplet.getHundredthPlace() == 0
&& triplet.getTenthPlace() == 0
&& triplet.getUnitPlace() == 0
&& mapIndex < this.triplets.size()) {
numTextBuilder = numTextBuilder.append("");
} else {
numTextBuilder = numTextBuilder.append(
NumberTextUtils.tripletMap().get(mapIndex)).append(" ");
}

mapIndex--;
}

return numTextBuilder;
}

public int getNum() {
return num;
}

public String getText() {
return this.toString();
}

public List<Triplet> getTriplets() {
return triplets;
}

}

 

package numbertextoop;

import utils.NumberTextUtils;

/**
* This class represents a group of digits up to a maximum of 3 that form a
* given number. 04/14/12 @author sanjeev 04/19/12 @author sanjeev
*
* @version 1.1
*/
public class Triplet {

private int hundredthPlace;
private int tenthPlace;
private int unitPlace;
private int value;

public Triplet(int tripletGroupNum) {
this.value = tripletGroupNum;
}

/**
* Arranges the place of digits for hundredth, tenth and unit positions in
* this triplet
*
* @param tripletValue
* @return Triplet
*/
public Triplet matchTripletPlaces(int tripletValue) {
this.setHundredthPlace(tripletValue / NumberTextUtils.HUNDRED);

tripletValue = tripletValue % NumberTextUtils.HUNDRED;
this.setTenthPlace(tripletValue / NumberTextUtils.TEN);

tripletValue = tripletValue % NumberTextUtils.TEN;
this.setUnitPlace(tripletValue);

return this;
}

/**
* Returns the string representation of this triplet
*
* @see java.lang.Object#toString()
*/
public String toString() {
return constructTripletText().toString();
}

/**
* @return tripletText StringBuilder
*/
private StringBuilder constructTripletText() {

StringBuilder tripletText = new StringBuilder();

tripletText = appendHundredthPlaceText(tripletText);

/** 'Teen' candidate check: 13 -> Thirteen, 11 -> Eleven */
StringBuilder teenCandidateBuilder = new StringBuilder();
String teenCandidate = teenCandidateBuilder.append(
String.valueOf(this.tenthPlace)).append(
String.valueOf(this.unitPlace)).toString();
int teenCandidateValue = Integer.parseInt(teenCandidate);

tripletText = appendTenthPlaceText(tripletText, teenCandidateValue);

tripletText = appendUnitPlaceText(tripletText, teenCandidateValue);

return tripletText;
}

/**
* Unit place text candidates Unit place : Five, Zero, Two etc
*/
private StringBuilder appendUnitPlaceText(StringBuilder tripletText,
int teenCandidate) {
/** Print unit place text: 1 -> one, 9 -> nine etc */
if (this.unitPlace != 0) {
if (isATeenNumber(teenCandidate)) {
return tripletText;
}

tripletText = appendDigitText(tripletText, this.unitPlace);
}

return tripletText;
}

/**
* Tenth place and 'Teen' text candidates Tenth place : Five, Eleven,
* Thirteen etc.
*/
private StringBuilder appendTenthPlaceText(StringBuilder tripletText,
int teenCandidate) {
if (this.tenthPlace != 0) {
if (isATeenNumber(teenCandidate)) {
tripletText = tripletText
.append(NumberTextUtils.teens[this.unitPlace]);
} else {
tripletText = tripletText
.append(NumberTextUtils.tens[this.tenthPlace - 1]);
}
tripletText.append(" ");
}

return tripletText;
}

/**
* Hundredth place text candidates Hundredth place : Two Hundred, Two
* Hundred and etc.
*/
private StringBuilder appendHundredthPlaceText(StringBuilder tripletText) {
if (this.hundredthPlace != 0) {
if (this.tenthPlace == 0 && this.unitPlace == 0) {
tripletText = appendDigitText(tripletText, this.hundredthPlace)
.append(" Hundred ");
} else {
tripletText = appendDigitText(tripletText, this.hundredthPlace)
.append(" Hundred and ");
}
}

return tripletText;
}

/**
* @param tripletText
* @param digit
* @return
*/
public StringBuilder appendDigitText(StringBuilder tripletText, int digit) {
tripletText.append(NumberTextUtils.digits[digit]);
return tripletText;
}

/**
* @param teenCandidate
* @return true if the number belongs to [11, 12, 13, 14, 15, 16, 17, 18,
*         19]
*/
public boolean isATeenNumber(int teenCandidate) {
return (teenCandidate >= NumberTextUtils.ELEVEN && teenCandidate < NumberTextUtils.TWENTY);
}

public int getValue() {
return value;
}

public int getHundredthPlace() {
return hundredthPlace;
}

public void setHundredthPlace(int hundredthPlace) {
this.hundredthPlace = hundredthPlace;
}

public int getTenthPlace() {
return tenthPlace;
}

public void setTenthPlace(int tenthPlace) {
this.tenthPlace = tenthPlace;
}

public int getUnitPlace() {
return unitPlace;
}

public void setUnitPlace(int unitPlace) {
this.unitPlace = unitPlace;
}

}

Code: Approach 2:

package numbertextalgo;

import utils.NumberTextUtils;

/**
* This class represents a given number in British English phrase Eg: 1234 in
* British English phrase: One Thousand Two Hundred and Thirty Four
*
* @author sanjeev, 14-Apr-2012
* @version 1.0
*/

public class NumberText {

public NumberText() {

}

public static String numberToString(int number) {
if (number == 0) {
return NumberTextUtils.ZERO_TEXT;
} else if (number < 0) {
return NumberTextUtils.NEGATIVE_TEXT + numberToString(-1 * number);
}

int count = 0;
String str = "";

while (number > 0) {
if (number % NumberTextUtils.THOUSAND != 0) {
str = numberToString100(number % NumberTextUtils.THOUSAND)
+ NumberTextUtils.bigs[count] + " " + str;
number /= NumberTextUtils.THOUSAND;
count++;
}
}

return str;
}

private static String numberToString100(int number) {
String str = "";

/** Convert hundredth place */
if (number >= NumberTextUtils.HUNDRED) {
str += NumberTextUtils.digits[number / NumberTextUtils.HUNDRED - 1]
+ " Hundred and ";
number %= NumberTextUtils.HUNDRED;
}

/** Convert NumberTextUtils.tens place */
if (number >= NumberTextUtils.ELEVEN
&& number <= NumberTextUtils.NINETEEN) {
return str + NumberTextUtils.teens[number - NumberTextUtils.ELEVEN]
+ " ";
} else if (number == NumberTextUtils.TEN
|| number >= NumberTextUtils.TWENTY) {
str += NumberTextUtils.tens[number / NumberTextUtils.TEN - 1] + " ";
number %= NumberTextUtils.TEN;
}

/** Convert ones place */
if (number >= NumberTextUtils.ONE && number <= NumberTextUtils.NINE) {
str += NumberTextUtils.digits[number - 1] + " ";
}

return str;
}
}

 

package utils;

import java.util.HashMap;
import java.util.Map;

/**
* This class is used as a utility class for representing a number in text
* 04/14/12 @author sanjeev 04/19/12 @author sanjeev
*
* @version 1.1
*/
public class NumberTextUtils {

public static String[] digits = new String[] { "Zero", "One", "Two",
"Three", "Four", "Five", "Six", "Seven", "Eight", "Nine" };
public static String[] teens = new String[] { "", "Eleven", "Twelve",
"Thirteen", "Fourteen", "Fifteen", "Sixteen", "Seventeen",
"Eighteen", "Nineteen" };
public static String[] tens = new String[] { "Ten", "Twenty", "Thirty",
"Forty", "Fifty", "Sixty", "Seventy", "Eighty", "Ninety" };
public static String[] bigs = new String[] { "", "Thousand", "Million",
"Billion" };

public static final String NEGATIVE_TEXT = "Negative";
public static final String ZERO_TEXT = "Zero";

public static final int ONE = 1;
public static final int THREE = 3;
public static final int NINE = 9;
public static final int TEN = 10;
public static final int TWENTY = 20;
public static final int ELEVEN = 11;
public static final int NINETEEN = 19;
public static final int HUNDRED = 100;
public static final int THOUSAND = 1000;

public static Map<Integer, String> tripletGroupMap = new HashMap<Integer, String>();

static {
tripletGroupMap.put(1, "");
tripletGroupMap.put(2, "Thousand");
tripletGroupMap.put(3, "Million");
tripletGroupMap.put(4, "Billion");
}

public static Map<Integer, String> tripletMap() {
return tripletGroupMap;
}
} 

 

package numbertextoop;

/**
* Test Class with main() for Number.java
*
* @author sanjeev
*
*/
public class TestNumber {

public static void main(String[] args) {
int num = 1000000000;
Number number = new Number(num);

sop(num + " in British English Phrase using OOP: " + number.getText());
}

private static void sop(Object o) {
System.out.println(o);
}

}

Researching Research…

In a world of rapidly developing technologies, the citizens that shape up the technology landscape require a combination of at least a two or more skills:

a)  Hard (Technical) Skills
b)  Soft Skills (Leadership, Motivation, Emoti. Maturity, Communication etc.)
c)  Ability to acquire new skills.

I strongly believe that the third skill  above, is significantly enhanced by research and have experienced it myself. Given the complexity of nature of interaction of components, different subsystems and speed with which the technology landscape evolves or changes, the ability to acquire new skills to relate to that landscape, is a key attribute for anyone.

What is research after all? We often hear the word ‘research’ being used in R&D organizations and academia, and also not so often in Service Organizations. In a very literal sense, in R&D organizations and academia, research mostly refers to the first-class research that’s driven by the the sheer new idea, and the efforts converging to a well-defined charter and eventually, resulting in filing of a patent, in publications, in research journals etc. It’s mostly used when there’s a talk about the work based on some innovative idea. As Mihaly Csikszentmihalyi puts it, In a very literal sense, when the work done during research draws attention of or is submitted to, the body constituted by a group of people registered as experts in the that field, that have the authority to approve or disapprove that work. In Service Organizations, it’s used in the context of prototyping task with lot of unknowns to start with, in the context of troubleshooting an issue or improving performance of a system. One of the myths, is that, research is an activity with no realistic goal or objective.

As Uday Khedkar of IIT Bombay, puts it in one of his presentations on the topic: “Research could involve any of the below or none of the below activities in our world like – Building a new device, Writing a new software, Repairing a device or Debugging a software, Drawing a conclusion from a lot of data, Proving a theorem, Formulating a theorem”. With different meanings attributed to the word ‘research’,  research, in essence, is a game of innovative ideas that are significant. The significance of ideas could lie in any of – Beauty, Utility, Enhancement of Knowledge.

The ingredients of good research are:
a) Innovation
- An idea that changes one’s behavior contrary to invention that doesn’t change one’s behavior. Idea is the most generated entity in the world. But, if doesn’t change one’s behavior, although it
embarks a new way of thinking and highlights a new way of building things, it remains as an invention.

b) Aesthetics
As Henry Poincare puts it, “Scientists study science not because it is useful, but because it is
beautiful. Here I do not talk about the beauty of appearance or beauty of qualities . . . Here I talk about that profound beauty which comes from a harmonious order of parts . . .”. Take painting or any design activity.
c) Other important aspects viz., Completeness
- Rigour
- Empirical demonstration
- Effective communication

Rigour removes imprecision and adds concreteness. It makes an idea very immune to personal interpretation. Empirical demonstration is through an evidence acquired by observation and experimentation. It opens up new channels for innovation. At an abstract level, it’s about capturing the interaction of  systems with each other; and drawing conclusions based on consistency in the type and nature of interactions. Humans beings can be considered to be made up of some ‘systems’ as well!  When research is not effectively communicated, the innovative ideas, research is not termed as real research by the field.

In Service Organizations, the rigour is often lacking in the way people deliver software, when they give excuse of client deadlines, to hide their inability to learn to work with rigour for delivering a mediocre quality code or software.

When we talk about the process of research, we talk about parameters like The Spirit of Inquiry, Breadth vs Depth, Ability to Abstract and Modularize, Moving from Confusion Phase to start with to Conviction Phase in the end, Research in Academia vs. Research in Industry,

The Ten Commandments of Creativity in Research
1. Seek extension of an earlier known solution
2. If you have a solution, find a problem
3. Find out the right questions to ask
4. Seek generality by removing specificities
5. Seek symmetry by testing for duality
6. Observe patterns in ideas
7. Distill the essence, refine your ideas
8. Distinguish the relevant from the irrelevant
9. Build levels of abstractions and migrate between them
10. Mix deep thinking with routine mechanical work.

Research is often seen as a major source of ideas and by implication, of innovation. One can  internalize ‘research’ nature with practice. Research makes a researcher, a better learner. Although grooming this nature is a leadership attribute for one, to build efficient and quality teams and build robust and long-lasting systems, it’s also of paramount criticality to develop the skill or the ability to switch to different mindset in order to thrive in a dynamic nature of business, especially in service organizations.

Recommended Reading:
1. An Article to clear up some Misconceptions about the nature of Research
2. Richard Hamming on Research
3. Mihaly Csikszentmihalyi on Creativity and Flow

Performance Engineering

I’m invested in MIT OpenCourseWare. That reminds me to go back to MIT site to look for stuff that I like to learn with a first-principles based approach! .

Below are some of the videos of the lectures taught under 6.172 – Performance Engineering of Software Systems and I find ‘em  very useful. Check out this MIT Course Page playlist for more related lectures.

Distributed Systems:

Basic Performance Engineering: