Why AI/ML models fail in Production

In my myriads consulting assignments, I have barely seen an AI/ML model in production. With renewed interest in AI/ML ( Artificial intelligence & Machine learning) and rightly so, enterprises are embracing for smart applications powered by models but their frustration is visible when models could not make it to production to work in tandem with applications.

Applications and models have to work together and in realtime to feed each other. Which means the models has to increasing learn with near real-time smaller data set and provide output to applications which is already in production.

Here is a summary of issues I see :-

Dirty and Stale Data : Typically the data given to data-scientists are months old and from older application where all the attributes are not properly mapped or missing. To add, data are given from raw systems which are incomplete and inconsistent.

Models are generated in silos : Data Engineers, Dev Ops, Service engineering are not involved in model planning. They do not understand model life-cycle and production need. Data-scientists hide their model under the hood, and when it is exposed to production system it falls flat.

Data at scale not considered : Many models are recursive and iterative. When you run or train with smaller dataset ( < 1M rows) it seems like a good candidate for production. When it is exposed to production level data, time and resource consumption exponentially increases. I think, it is largest cause of no-go to production.

Data Scientists not understanding production environment : Creating a model in silos with couple of thousand of data points is different, but to have in production , you need to understand integration points with other systems (applications), CI/CD, backup & HA, hardware topology and most importantly data pipeline and workflow. A model should retrain and refresh at right time and right dataset and give the output to right system.

Selection of code and technology : Data Scientists are comfortable with R/Matlab/Python while most systems in production uses C/C++ , Java, Go, MS technologies. Data Scientist’s language of choice has limitations in production.

Conclusion: The line between application and models will be getting blurred in coming days. Models will be developed as applications and applications will be developed at models. Models will copy CI/CD, Atscale, HA, Integration , real time interactions etc from applications and applications can be retrained , refreshed , self learning, reenforce learning like models.

Feel free to share your thoughts.

About author : Vivek is creator of osDQ – world’s first open source data quality and prep tool. https://sourceforge.net/projects/dataquality/

He has also open sourced apache spark based data pipeline framework.https://sourceforge.net/projects/apache-spark-osdq/

Advertisements

Sampling using apache Spark 2.1.1

There has been a debate in big data community “do we need sampling in big data world” ? Now big data platform has storage and processing power to use all dataset for analysis so sampling is not required – is the argument.

This arguments holds true for data discovery where doing analysis on full dataset give more confidence and every anomaly and pattern is captured. However, sampling still saves considerable amount of time in dimension reduction, correlation and model generation. You have to go through hundreds of attributes (with all permutation and combination) to find dependent, independent , correlated variables where representative dataset saves you hour with almost same accuracy.

We have open sourced data sampling code using apache spark 2.1.1 at https://sourceforge.net/projects/apache-spark-osdq/

Random Sampling : on dataset<row> random sampling is provided by apache spark where user can provide the fraction he needs for sampling.

Dataset<Row> org.apache.spark.sql.Dataset.sample(boolean withReplacement, double fraction)

Stratified Sampling : dataset<row> does not provide stratified sampling so dataset is converted into PairedRDD with key column which need to be stratified and then use samplebyKeyExact. It does many pass to find the exact fraction

for(Row uniqueKey:uniqueKeys){

fractionsMap.merge(uniqueKey.mkString(),fraction, (V1,V2) -> V1);

}

		

JavaPairRDD<String, Row> dataPairRDD = SparkHelper.dfToPairRDD(keyColumn, df);
JavaRDD<Row> sampledRDD = dataPairRDD.sampleByKeyExact(false, fractionsMap).values();
Dataset<Row> sampledDF = df.sqlContext().createDataFrame(sampledRDD, df.schema());

return sampledDF;

Keylist Sampling : This is like Stratified sampling but only make one pass to meet the fraction value.

JavaRDD<Row> sampledRDD = dataPairRDD.sampleByKey(false, fractionsMap).values();

Command Line Option :

 -c,--keyColumn <arg>          Key Column for stratified/keylist sampling

 -f,--fraction <arg>           Sample fraction size

 -fm,--fractionMapping <arg>   comma seperated pairs of key,fraction size

 -h,--help                     show this help.

 -i,--input <arg>              Input Folder/File path

 -if,--inputFormat <arg>       input file format

 -o,--output <arg>             Output Folder path

 -of,--outFormat <arg>         output file format

 -t,--type <arg>               Sampling type  - ran

                               dom/stratified/keylist

example : ” -f 0.2 -i ./testfile -if csv -o ./outputFile -of csv -t stratified -c key1 -fm ./keymapping”

testfile:

key1,key2

vivek,1

vivek,1

vivek,1

vivek,1

vivek,1

vivek,1

vivek,1

vivek,1

singh,2

singh,2

singh,2

singh,2

singh,2

singh,2

singh,2

singh,2

singh,2

keymapping:

vivek,0.1

singh,0.1

about Vivek Singh – data architect, open source evangelist , chief contributor of Open Source Data Quality project http://sourceforge.net/projects/dataquality/

Author of fiction book “The Reverse Journey” http://www.amazon.com/Reverse-Journey-Vivek-Kumar-Singh/dp/9381115354/

Big Data QA – Why it is so difficult ?

Recently Big Data QA profile was needed for a project which processes terabyte of machine generated data everyday. Below is the cryptic job description that was sent to recruitment team —

” We are looking for Big data developer, who can write code on Hadoop cluster to validate structured and semi structured data.

Skill sets – apache spark, Hive, Hadoop, data quality ”

Recruitment team was amused.

Are you looking for a developer or QA ? Is it manual testing or automated testing ? Where we should look for these kind of profile ?

Big Data QA is an emerging area and it is different from traditional product or enterprise QA. Big data QA is big data developer who develops code for validating data which is beyond human scale. BDQ (Big data QA) does not fit into traditional QA profiles – automated vs manual ; backend vs frontend ; feature vs non feature etc.

BDQ is required by data products to ensure their processes which manage data lifecycle, are ingesting and emitting right data. The volumes is so huge; it is beyond human to load the data into spreadsheet and validate row by row ; column by column.

As per business requirement, BDQ may write very complex code to validate complex business rules. However some day to day activity of BDQ requires :

  • Profile the data to validate dataset — pre and post business rules
  • Find out data hole to ensure all data are coming
  • Statistical modeling to compare and contrast two or more datasets
  • Implement unified formatting and unified data types
  • Implement multiple type of joins on dataset including fuzzy joins
  • Implement UDF ( user defined functions) for reusable validation logic
  • Find outlier data and how to manage them
  • Find Null and Empty data and how to manage them
  • Validate the naming conventions of datasets
  • Validate the file format ( csv, avro, parquet etc) of datasets
  • Monitor incoming and outgoing dataset and redistribute if it fails
  • Validate the data models for analytics
  • Create sample data and make sure sample is not skewed
  • Create training and test dataset
  • Encrypt and DeCrypt (anonymize) data
  • Implement and monitor compliance rule

By no means, above is an exhaustive list but it is more of a indicative list. BDQ is a prolific programmer who understands data ( like business analyst) well and code for Data Quality domain. These bold combination makes BDQ so hard to find.

It is best to train data engineers to become BDQ by providing courses and exposures on data quality and data management.

Vivek Singh – data architect, open source evangelist , chief contributor of Open Source Data Quality project http://sourceforge.net/projects/dataquality/

Author of fiction book “The Reverse Journey” http://www.amazon.com/Reverse-Journey-Vivek-Kumar-Singh/dp/9381115354/

How to do Diff of Spark dataframe

Apache spark does not provide diff or subtract method for Dataframes. However, it is common requirement to do diff of dataframes – especially where data engineers have to find out what changes from previous values ( dataframe).

Requirements has generally following use cases:

a.) Find out diff (subtract) with complete dataframes

b.) Find out diff (subtract) with primary keys (Single column)

c.) Find out diff (subtract) with composite keys (Mupltiple columns)

Since dataframe does not have substract method here is the following step you need to do

i) First convert dataframe to RDD keeping the schema of dataframe safe.

ii) Create a pairedRDD for key value pair for step b and c

iii.) Use the substract method of RDD and apply the schema on RDD

iv.) Get back your dataframe

	// find the diff between two data sets A -B
	public DataFrame findDiff ( DataFrame left, DataFrame right) {
		if (left == null || right == null ) {
			return null;
		}
		StructType schema = left.schema();
		JavaRDD<Row> leftRDD = left.toJavaRDD();
		JavaRDD<Row> rightRDD = right.toJavaRDD();
		
		// diff which is there in right but not in left deleted value
		JavaRDD<Row> diffRDD = rightRDD.subtract(leftRDD);
		DataFrame newdf = sqlContext.createDataFrame(diffRDD, schema);
		
		return newdf;
		
	}
	
	// find the diff between two data sets A -B using colname
	public DataFrame findDiff ( DataFrame left, String leftCol, DataFrame right,  String rightCol) {
		if (left == null || right == null ) {
			return null;
		}
		StructType schema = right.schema();
		JavaRDD<Row> leftRDD = left.toJavaRDD();
		JavaRDD<Row> rightRDD = right.toJavaRDD();
		String[] leftColName = left.columns();
		String[] rightColName = right.columns();
		int leftI=0; int rightI=0;
		for (int i=0 ; i < leftColName.length; i++)
			if (leftCol.equals(leftColName[i])) {
				leftI = i; break;
			}
		for (int i=0 ; i < rightColName.length; i++)
			if (rightCol.equals(rightColName[i])) {
				rightI = i; break;
			}
		final int leftIf = leftI;
		final int rightIf = rightI;
				
						
		
		// Now creare paired RDD for substract
		JavaPairRDD<String, Row> leftPair = leftRDD.mapToPair(new PairFunction<Row, String, Row>() {
		            /**
			 * 
			 */
			private static final long serialVersionUID = 1L;

					public Tuple2<String, Row> call(Row row) throws Exception {
		            	
		                return new Tuple2<String, Row>(row.get(leftIf).toString(), row);
		            }
		 }).cache();
		
		JavaPairRDD<String, Row> rightPair = rightRDD.mapToPair(new PairFunction<Row, String, Row>() {
            /**
	 * 
	 */
			private static final long serialVersionUID = 1L;

			public Tuple2<String, Row> call(Row row) throws Exception {
            	
                return new Tuple2<String, Row>(row.get(rightIf).toString(), row);
            }
		}).cache();
		
		// diff which is there in right but not in left deleted value
		// apply schema of right
		JavaPairRDD<String, Row> diffRDD = rightPair.subtractByKey(leftPair);
		JavaRDD<Row> newdataframe= diffRDD.values();
		DataFrame newdf = sqlContext.createDataFrame(newdataframe, schema);
		
		return newdf;
		
	}

osDQ releases apache spark based data quality

World’s first open source data quality and data preparation project (osDQ – https://sourceforge.net/projects/dataquality/ ) releases apache spark based data quality and data preparation modules for big data.

Apache Spark based APIs can be downloaded from here : https://sourceforge.net/projects/apache-spark-osdq/

This beta release has following features :

Normalization:

  • ZScore

Functional input – mean and std dev

Return type : dataframe

  • ZeroScore (between 0 and 1)

Functional input – min and max

Return type : dataframe

  • RatioScore (num/denum)

Functional input – ratio number

Return type : dataframe

  • Subtraction Score (a –b)

Functional input – Subtraction number

Return type : dataframe

Replacement:

  • Replacement with key-value pairs

Functional input – hashtable and columns type

Return type : dataframe

  • Replacement Null with default value

Functional input – value

Return type : dataframe

  • Replacement using regression value (linear and multi-linear)

Functional input – No of iterations

Return type : dataframe

Remove:

  • Removing Null Rows

Functional input – all or any

Return type : dataframe

  • Removing Duplicate Rows

Functional input – all or any

Return type : dataframe

Profiling:

Functional input – DataFrame

Return type: Hashtable<Colname,Hashtatable<Key,Value>>

HashKeys – “count”,”unique”,”nullcount”,”pattern”,”min”,”max”

Fuzzy Join and Replacement :

Function Input – two strings

Return type – cosine similarity ( between -1 to 1)

Summary : osDQ will enhance the project to provide more core APIs for data quality , data preparation and data science. It will save community time to write those functions for big data environment.

 

 

Apache Spark ML for Data Quality

Apache Spark is becoming de-facto standard for data processing. Spark platform is over-arching to all aspects of data lifecycle – Ingestion, Discovery, Preparation and Data Science with easy to use, developers friendly APIs.

Availability of large set of statistical and machine leaning based, scalable algorithm in Spark will bring a new perspective to data quality and validation where these algorithms will be used to automatic and machine based data anomaly detection and correction. Though these algorithms are not new, but bring them into data engineering and data architect domain is new. R, SAS, MATLAB etc were confined to data scientists and not popular with data engineering.

Algorithms like Principal Component Analysis, Support Vector Machine, Pairwise comparison, Regression , Edit Distance, K-Mean have to play critical role in automation of data quality and data correction rules. In the following code, I have used Spark Mlib linear regression model to replace null from column A based on regression values from column B ( linear regression model), and a set of columns (multilinear regression model)

This code snippet is for educational purpose only.

1.) Create a dataframe on which you want to apply the rules – inputBean is that object

2.) Create and Train Model – Linear

public LinearRegressionModel doLinearReg(DataFrameProperty inputBean, int numIterations) {

DataFrame df = inputBean.getDataFrame();
String labelCol = inputBean.getLabelCol(); // replace null from this column
String regCol = inputBean.getRegCol(); // use this column for regression
DataFrame newdf = df.select(labelCol, regCol);

JavaRDD<LabeledPoint> parseddata = newdf.javaRDD().map(new Function<Row, LabeledPoint>() {
private static final long serialVersionUID = 1L;
public LabeledPoint call(Row r) throws Exception {
Object labVObj = r.get(0);
Object regVarObj = r.get(1);
if (labVObj != null && regVarObj != null) {
double[] regv = new double[] { (Double) regVarObj };

Vector regV = new DenseVector(regv);
return new LabeledPoint((Double) labVObj, regV);
} else {

double[] regv = new double[] { 0.0 };
Vector regV = new DenseVector(regv);
return new LabeledPoint(0.0, regV);
} }
});

// Building the model
return LinearRegressionWithSGD.train(parseddata.rdd(), numIterations);
}

3.) Use model to replace Null Value

//LinearRegressionModel model = dqu.doLinearReg(inputBean,20);

//System.out.println(“\n Intercept: ” + model.intercept());

// System.out.println(“Weight :” + model.weights().toString());

public DataFrame replaceNull(DataFrameProperty inputBean, final double intercept, final double weight) {

DataFrame df = inputBean.getDataFrame();
String labelCol = inputBean.getLabelCol();
String regCol = inputBean.getRegCol();
String uniqCol = inputBean.getUniqColName();
SQLContext sqlContext = inputBean.getSqlContext();
DataFrame newdf = df.select(labelCol, regCol, uniqCol);

JavaRDD<Row> parseddata = newdf.toJavaRDD().map(new FunctionMap(intercept,weight));

// Generate the schema based on the string of schema
StructField[] fields = new StructField[3];
fields[0] = DataTypes.createStructField(labelCol, DataTypes.DoubleType, true);
fields[1] = DataTypes.createStructField(regCol, DataTypes.DoubleType, true);
fields[2] = DataTypes.createStructField(uniqCol, DataTypes.DoubleType, true);
StructType schema = DataTypes.createStructType(fields);
DataFrame newdf1 = sqlContext.createDataFrame(parseddata, schema);
df.join(newdf1, uniqCol); // After replace join to main dataframe based on unique column 
returndf;
}

4.) Create and Train Model – Multi Linear

public LinearRegressionModel doMultiLinearReg(DataFrameProperty inputBean, int numIterations) {

DataFrame df = inputBean.getDataFrame();
String labelCol = inputBean.getLabelCol();
String [] inputCols = inputBean.getInputCols(); // multiple columns for regression
DataFrame newdf = df.select(labelCol, inputCols);

JavaRDD<LabeledPoint> parseddata = newdf.javaRDD().map(new Function<Row, LabeledPoint>() {
private static final long serialVersionUID = 1L;
public LabeledPoint call(Row r) throws Exception {
Object labVObj = r.get(0);
int colC = r.size();
double[] regv = new double[colC -1]; // -1 for first index

for (int i =1; i < colC; i++) {
Object regVarObj = r.get(i);
if (regVarObj != null)
regv[i-1] = (Double)regVarObj;
else
regv[i-1] = 0.0D; // Null replaced with 0.0
}
Vector regV = new DenseVector(regv);
if (labVObj != null ) {
return new LabeledPoint((Double) labVObj, regV);
} else {
return new LabeledPoint(0.0D, regV);
}}
});

// Building the model
return LinearRegressionWithSGD.train(parseddata.rdd(), numIterations);
}

5.) Use for null replacement

DataFrame newdf = dqu.replaceNull(inputBean,model.intercept(),model.weights().toArray()[0]);

6.) Function Map class

public class FunctionMap  implements java.io.Serializable, Function<Row,Row> {

private static final long serialVersionUID = 1L;
double _intercept, _weight;
public FunctionMap(double intercept, double weight) {
_intercept=intercept;
_weight = weight;
}

public Row call(Row r) throws Exception {
Double regv = r.getDouble(1);
if (r.get(0) == null && regv != null) {
double newVal = _intercept + _weight * regv;
return RowFactory.create(newVal, regv, r.getDouble(2));
} else
return r;
}
}

Biggest Problem of Big Data – Entity Resolution

Big Data has gone past PoC phase. Different companies are at different stages of implementation. Data Ingestion, Data Storage ( Data Lake and EDW), Data Processing and Data Visualization processes have been quite mature and there are many open source and proprietary software to solve these problems.

One major hurdle Big Data faces today is – Entity Resolution ( defining a business entity form multitude of data sources). In EDW (Enterprise data warehouse) world, data were structured and sources were limited. Also keys of sources were pre and well defined . So RDBMS joins ( inner, outer, left outer, right outer, semi join etc) were enough to merge data from two different systems and tables.

In Big Data world, there is hardly any key or attribute that runs across the sources. Also keys of one system is useless as other systems are completely independent of each other. So business have to define their own logic for merging data from different sources which defines one entity. To make it worse, RDBMS kind of exact match joins should be replaced by fuzzy joins as referential integrity across systems can be ensured.

Following is a practical approach for resolving Entity Resolution:-

1.) Pre-define your entities and super set of  attributes ( coming from all data sources)

2.) Attributes may have multiple related values that should to mapped to same attributes. Plan your storage like this  ( Graph databases work fine for this relationship storage)

3.) Merge data from multiple sources using merge business logic to create a virtual entity with sizable attributes ( we used Apache spark for this )

Mapping attributes

Location

  • Zip, County, State, Lat/long
  • Nearby locations (+/- area)  – high propensity area
  • IP location

Time

  • Date/time stamp
  • Nearby time stamp (+/-) – Event happening before or after a period

Attributes

  • String match (Fuzzy) – Name, Address, Cause
  • Cardinal Match  – events sharing same or similar key
  • IP Correlation (Primary IP, Secondary IP, IP from same ZIP code)
  • Other business logic related merging rules

You can make you merge model Machine Learning based so it will be leaning over time to do a relevant merge.

4.) Right Sizing Merges Columns

  • Remove transaction columns
  • Remove  database columns
  • Remove technical identifiable columns
  • Remove duplicate Columns

5.) Search this entity in your relational graph database to find ranks of similar entities more than threshold. If the outcome is less than threshold then make a new entry in your entity table with the attributes of the searching entity.

 6.) Take the highest ranked entity and mapped the missing attributes from highest ranked entities. This entity is you final entity for business

How to enhance entities with changing attributes:

Like any practical entity, values of attributes keep changing. You map the attributes values of searching entity to highest ranked entity and see the different of values. let’s say the IP values of an entity is matching to its secondary value over time. Then the secondary IP value becomes primary and primary becomes secondary. Or it is new IP then add one more relationship node with new values.

Ranking algorithm:

Business can assign different weight-age to must have, critical, important and good to have attributes and their matching threshold. This model also matches with secondary or related values so make it more accurate. Let’s say, address is a must match attributes. A customer other attributes are matching but address not matching so other models will reject it but in this model if matches with his or her office address or other address, it will boost the record that is right.