data mining

Pre-build Machine Learning Models

Posted on Updated on

Machine learning has seen widespread adoption over the past few years. In more recent times we have seem examples of how the models, created by the machine learning algorithms, can be shared. There have been various approaches to sharing these models using different model interchange languages. Some of these have become more or less popular over time, for example a few years ago PMML was very popular, and in more recent times ONNX seems to popular. Who knows what it will be next year or in a couple of years time.

With the increased use of machine learning models and the ability to share them, we are now seeing other uses of them. Typically the sharing of models involved a company transferring a model developed by the data scientists in their lab environment, to DevOps teams who then deploy the model into the production environment. This has developed a new are of expertise of MLOps or AIOps.

The languages and tools used by the data scientists in the lab environment are different to the languages used to deploy applications in production. The model interchange languages can be used take the model parameters, algorithm type and data transformations, etc and map these into the interchange language. The production environment would read this interchange object and apply it to the production language. In such situations the models will use the algorithms already coded in the production language. For example, the lab environment could be using Python. But the product environment could be using C, Java, Go, etc.  Python is an interpretative language and in a lot of cases is not suitable for real-time use in a production environment, due to speed and scalability issues. In this case the underlying algorithm of the production language will be used and not algorithm used in the lab. In theory the algorithms should be the same. For example a decision tree algorithm using Gini Index in one language should function in the same way in another language. We all know there can be a small to a very large difference between what happens in theory and how it works in practice. Different language and different developers will do things slightly differently. This means there will be differences between the accuracy of the models developed in the lab versus the accuracy of the (same) model used in production. As long as everyone is aware of this, then everything will be ok. But it will be important task, for the data science team, to have some measurements of these differences.

Heres One I Made Earlier: 9780857835130: Amazon.com: Books

Moving on a little this a little, we are now seeing some other developments with the development and sharing of machine learning models, and the use of these open model interchange languages, like ONNX, makes this possible.

We are now seeing people making their machine learning models available to the wider community, instead of keeping them within their own team or organization.

Why would some one do this? why would they share their machine learning model?  It’s a bit like the picture to the left which comes from a very popular kids programme on the BBC called Blue Peter. They would regularly show some craft projects for kids to work on at home. They would never show all the steps needed to finish the project and would end up showing us “one I made earlier”. It always looked perfect and nothing like what they tried to make in the studio and nothing like my attempt.

But having pre-made machine learning models is now a thing. There ware lots of examples of these and for example the ONNX website has several pre-trained models ready for you to use. These cover various examples for image classification, object detection, machine translation and comprehension, language modeling, speech and audio processing, etc.  More are being added over time.

Most of these pre-trained models are based on defined data sets and problems and allows others to see what they have done, and start building upon their work without the need to go through the training and validating phase.

Could we have something like this in the commercial world? Could we have pre-trained machine learning models being standardized and shared across different organizations?  Again the in-theory versus in-practical terms apply. Many organizations within a domain use the same or similar applications for capturing, storing, processing and analyzing their data. In this case could the sharing of machine learning models help everyone be more competitive or have better insights and discoveries from their data? Again the difference between in-theory versus in-practice applies.

Some might remember in the early days of Data Warehousing we used to have some industry (dimensional) models, and vendors and consulting companies would offer their custom developed industry models and how to populate these. In theory these were supposed to help companies to speed up their time to data insights and save money. We have seem similar attempts at doing similar things over the decades. But the reality was most projects ended up being way more expensive and took way too long to deploy due to lots of technical difficulties and lots of differences in the business understand, interpretation and deployment of the underlying applications. The pre-built DW model was generic and didn’t really fit in with the business needs.

Although we are seeing more and more pre-trained machine learning models appearing on the market. Many vendors are offering pre-trained solutions. But can these really work. Some of these pre-trained models are based on certain data preparation, using one particular machine learning model and using only one particular evaluation matric. As with the custom DW models of twenty years ago, pre-trained ML models are of limited use.

Everyone is different, data is different, behavior is different, etc. the list goes on. Using the principle of the “No Free Lunch” theorem, although we might be using the same or similar applications for capturing, storing, processing and analysing their data, the underlying behavior of the data (and the transactions, customers etc that influence that), will be different, the marketing campaigns will be different, business semantics may be different, general operating models will be different, etc.  Based on “No Free Lunch” we need to explore the data using a variety of different algorithms, to determine what works for our data at this point in time. The behavior of the data (and business influences on it) keep on changing and evolving on a daily, weekly, monthly, etc basis.  A great example of this but in a more extreme and rapid rate of change happened during the COVID pandemic. Most of the machine learning models developed over the preceding period no longer worked, the models developed during the pandemic have a very short life span, and it will take some time before “normal” will return and newer models can be built to represent the “new normal”

Scottish Whisky Data Set – Updated

Posted on Updated on

The Scottish Whiskey data set consist of tasting notes and evaluations from 86 distilleries around Scotland. This data set has been around a long time andwas a promotional site for a book, Whisky Classified: Choosing Single Malts by Flavour. Written by David Wishart of the University of Saint Andrews, the book had its most recent printing in February 2012.

I’ve been using this data set in one of my conference presentations (Planning my Summer Vacation), but to use this data set I need to add 2 new attributes/features to the data set. Each of the attributes are listed below and the last 2 are the attributes I added. These were added to include the converted LAT and LONG comparable with Google Maps and other similar mapping technology.

Attributes include:

  • RowID
  • Distillery
  • Body
  • Sweetness
  • Smoky
  • Medicinal
  • Tobacco
  • Honey
  • Spicy
  • Winey
  • Nutty,
  • Malty,
  • Fruity,
  • Floral,
  • Postcode,
  • Latitude,
  • Longitude
  • lat  — newly added
  • long  — newly added

Here is the link to download and use this updated Scottish Whisky data set.

The original website is no longer available but if you have a look at the Internet Archive you will find the website.

Screenshot 2020-01-23 14.44.53

Demographics vs Psychographics for Machine Learning

Posted on Updated on

When preparing data for data science, data mining or machine learning projects you will create a data set that describes the various characteristics of the subject or case record. Each attribute will contain some descriptive information about the subject and is related to the target variable in some way.

In addition to these attributes, the data set will be enriched with various other internal/external data to complete the data set.

Some of the attributes in the data set can be grouped under the heading of Demographics. Demographic data contains attributes that explain or describe the person or event each case record is focused on. For example, if the subject of the case record is based on Customer data, this is the “Who” the demographic data (and features/attributes) will be about. Examples of demographic data include:

  • Age range
  • Marital status
  • Number of children
  • Household income
  • Occupation
  • Educational level

These features/attributes are typically readily available within your data sources and if they aren’t then these name be available from a purchased data set.

Additional feature engineering methods are used to generate new features/attributes that express meaning is different ways. This can be done by combining features in different ways, binning, dimensionality reduction, discretization, various data transformations, etc. The list can go on.

The aim of all of this is to enrich the data set to include more descriptive data about the subject. This enriched data set will then be used by the machine learning algorithms to find the hidden patterns in the data. The richer and descriptive the data set is the greater the likelihood of the algorithms in detecting the various relationships between the features and their values. These relationships will then be included in the created/generated model.

Another approach to consider when creating and enriching your data set is move beyond the descriptive features typically associated with Demographic data, to include Pyschographic data.

Psychographic data is a variation on demographic data where the feature are about describing the habits of the subject or customer.  Demographics focus on the “who” while psycographics focus on the “why”. For example, a common problem with data sets is that they describe subjects/people who have things in common. In such scenarios we want to understand them at a deeper level. Psycographics allows us to do this. Examples of Psycographics include:

  • Lifestyle activities
  • Evening activities
  • Purchasing interests – quality over economy,  how environmentally concerned are you
  • How happy are you with work, family, etc
  • Social activities and changes in these
  • What attitudes you have for certain topic areas
  • What are your principles and beliefs

The above gives a far deeper insight into the subject/person and helps to differentiate each subject/person from each other, when there is a high similarity between all subjects in the data set. For example, demographic information might tell you something about a person’s age, but psychographic information will tell you that the person is just starting a family and is in the market for baby products.

I’ll close with this. Consider the various types of data gathering that companies like Google, Facebook, etc perform. They gather lots of different types of data about individuals. This allows them to build up a complete and extensive profile of all activities for individuals. They can use this to deliver more accurate marketing and advertising. For example, Google gathers data about what places to visit throughout a data, they gather all your search results, and lots of other activities. They can do a lot with this data. but now they own Fitbit. Think about what they can do with that data and particularly when combined with all the other data they have about you. What if they had access to your medical records too!  Go Google this ! You will find articles about them now having access to your health records. Again combine all of the data from these different data sources. How valuable is that data?

 

Managing imbalanced Data Sets with SMOTE in Python

Posted on Updated on

When working with data sets for machine learning, lots of these data sets and examples we see have approximately the same number of case records for each of the possible predicted values. In this kind of scenario we are trying to perform some kind of classification, where the machine learning model looks to build a model based on the input data set against a target variable. It is this target variable that contains the value to be predicted. In most cases this target variable (or feature) will contain binary values or equivalent in categorical form such as Yes and No, or A and B, etc or may contain a small number of other possible values (e.g. A, B, C, D).

For the classification algorithm to perform optimally and be able to predict the possible value for a new case record, it will need to see enough case records for each of the possible values. What this means, it would be good to have approximately the same number of records for each value (there are many ways to overcome this and these are outside the score of this post). But most data sets, and those that you will encounter in real life work scenarios, are never balanced, as in having a 50-50 split. What we typically encounter might be a 90-10, 98-2, etc type of split. These data sets are said to be imbalanced.

Screenshot 2019-05-20 15.34.14

The image above gives examples of two approaches for creating a balanced data set. The first is under-sampling. This involves reducing the class that contains the majority of the case records and reducing it to match the number of case records in the minor class. The problems with this include, the resulting data set is too small to be meaningful, the case records removed could contain important records and scenarios that the model will need to know about.

The second example is creating a balanced data set by increasing the number of records in the minority class. There are a few approaches to creating this. The first approach is to create duplicate records, from the minor class, until such time as the number of case records are approximately the same for each class. This is the simplest approach. The second approach is to create synthetic records that are statistically equivalent of the original data set. A commonly technique used for this is called SMOTE, Synthetic Minority Oversampling Technique. SMOTE uses a nearest neighbors algorithm to generate new and synthetic data we can use for training our model. But one of the issues with SMOTE is that it will not create sample records outside the bounds of the original data set. As you can image this would be very difficult to do.

The following examples will illustrate how to perform Under-Sampling and Over-Sampling (duplication and using SMOTE) in Python using functions from Pandas, Imbalanced-Learn and Sci-Kit Learn libraries.

NOTE: The Imbalanced-Learn library (e.g. SMOTE)requires the data to be in numeric format, as it statistical calculations are performed on these. The python function get_dummies was used as a quick and simple to generate the numeric values. Although this is perhaps not the best method to use in a real project. With the other sampling functions can process data sets with a sting and numeric.

Data Set:  Is the Portuaguese Banking data set and is available on the UCI Data Set Repository, and many other sites. Here are some basics with that data set.

import warnings

import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
get_ipython().magic('matplotlib inline')

bank_file = ".../bank-additional-full.csv"

# import dataset
df = pd.read_csv(bank_file, sep=';',)

# get basic details of df (num records, num features)
df.shape
df['y'].value_counts() # dataset is imbalanced with majority of class label as "no".
no     36548
yes     4640
Name: y, dtype: int64
#print bar chart
df.y.value_counts().plot(kind='bar', title='Count (target)');

Screenshot 2019-05-27 09.38.36

Example 1a – Down/Under sampling the majority class y=1 (using random sampling)

count_class_0, count_class_1 = df.y.value_counts()

# Divide by class
df_class_0 = df[df['y'] == 0] #majority class
df_class_1 = df[df['y'] == 1] #minority class

# Sample Majority class (y=0, to have same number of records as minority calls (y=1)
df_class_0_under = df_class_0.sample(count_class_1)
# join the dataframes containing y=1 and y=0
df_test_under = pd.concat([df_class_0_under, df_class_1])

print('Random under-sampling:')
print(df_test_under.y.value_counts())
print("Num records = ", df_test_under.shape[0])

df_test_under.y.value_counts().plot(kind='bar', title='Count (target)');
Random under-sampling:
1    4640
0    4640
Name: y, dtype: int64
Num records =  9280

Screenshot 2019-05-27 09.41.06

Example 1b – Down/Under sampling the majority class y=1 using imblearn

from imblearn.under_sampling import RandomUnderSampler

X = df_new.drop('y', axis=1)
Y = df_new['y']

rus = RandomUnderSampler(random_state=42, replacement=True)
X_rus, Y_rus = rus.fit_resample(X, Y)

df_rus = pd.concat([pd.DataFrame(X_rus), pd.DataFrame(Y_rus, columns=['y'])], axis=1)

print('imblearn over-sampling:')
print(df_rus.y.value_counts())
print("Num records = ", df_rus.shape[0])

df_rus.y.value_counts().plot(kind='bar', title='Count (target)');

[same results as Example 1a]

Example 1c – Down/Under sampling the majority class y=1 using Sci-Kit Learn

from sklearn.utils import resample

print("Original Data distribution")
print(df['y'].value_counts())

# Down Sample Majority class
down_sample = resample(df[df['y']==0],
replace = True, # sample with replacement
n_samples = df[df['y']==1].shape[0], # to match minority class
random_state=42) # reproducible results

# Combine majority class with upsampled minority class
train_downsample = pd.concat([df[df['y']==1], down_sample])

# Display new class counts
print('Sci-Kit Learn : resample : Down Sampled data set')
print(train_downsample['y'].value_counts())
print("Num records = ", train_downsample.shape[0])
train_downsample.y.value_counts().plot(kind='bar', title='Count (target)');

[same results as Example 1a]

Example 2 a – Over sampling the minority call y=0 (using random sampling)

df_class_1_over = df_class_1.sample(count_class_0, replace=True)
df_test_over = pd.concat([df_class_0, df_class_1_over], axis=0)

print('Random over-sampling:')
print(df_test_over.y.value_counts())

df_test_over.y.value_counts().plot(kind='bar', title='Count (target)');
Random over-sampling:
1    36548
0    36548
Name: y, dtype: int64

Screenshot 2019-05-27 09.46.08

Example 2b – Over sampling the minority call y=0 using SMOTE

from imblearn.over_sampling import SMOTE

print(df_new.y.value_counts())
X = df_new.drop('y', axis=1)
Y = df_new['y']

sm = SMOTE(random_state=42)
X_res, Y_res = sm.fit_resample(X, Y)

df_smote_over = pd.concat([pd.DataFrame(X_res), pd.DataFrame(Y_res, columns=['y'])], axis=1)

print('SMOTE over-sampling:')
print(df_smote_over.y.value_counts())

df_smote_over.y.value_counts().plot(kind='bar', title='Count (target)');

[same results as Example 2a]

Example 2c – Over sampling the minority call y=0 using Sci-Kit Learn

from sklearn.utils import resample

print("Original Data distribution")
print(df['y'].value_counts())

# Upsample minority class
train_positive_upsample = resample(df[df['y']==1],
replace = True, # sample with replacement
n_samples = train_zero.shape[0], # to match majority class
random_state=42) # reproducible results

# Combine majority class with upsampled minority class
train_upsample = pd.concat([train_negative, train_positive_upsample])

# Display new class counts
print('Sci-Kit Learn : resample : Up Sampled data set')
print(train_upsample['y'].value_counts())
train_upsample.y.value_counts().plot(kind='bar', title='Count (target)');

[same results as Example 2a]

Cluster Distance using SQL with Oracle Data Mining – Part 4

Posted on Updated on

This is the fourth and last blog post in a series that looks at how you can examine the details of predicted clusters using Oracle Data Mining. In the previous blog posts I looked at how to use CLUSER_ID, CLUSTER_PROBABILITY and CLUSTER_SET.

In this blog post we will look at CLUSTER_DISTANCE. We can use the function to determine how close a record is to the centroid of the cluster. Perhaps we can use this to determine what customers etc we might want to focus on most. The customers who are closest to the centroid are one we want to focus on first. So we can use it as a way to prioritise our workflows, particularly when it is used in combination with the value for CLUSTER_PROBABILITY.

Here is an example of using CLUSTER_DISTANCE to list all the records that belong to Cluster 14 and the results are ordered based on closeness to the centroid of this cluster.

SELECT customer_id, 
       cluster_probability(clus_km_1_37 USING *) as cluster_Prob,
       cluster_distance(clus_km_1_37 USING *) as cluster_Distance
FROM   insur_cust_ltv_sample
WHERE   cluster_id(clus_km_1_37 USING *) = 14
order by cluster_Distance asc;

Here is a subset of the results from this query.

NewImage

When you examine the results you may notice that the records that is listed first and closest record to the centre of cluster 14 has a very low probability. You need to remember that we are working in a N-dimensional space here. Although this first record is closest to the centre of cluster 14 it has a really low probability and if we examine this record in more detail we will find that it is at an overlapping point between a number of clusters.

This is why we need to use the CLUSTER_DISTANCE and CLUSTER_PROBABILITY functions together in our workflows and applications to determine how we need to process records like these.

Cluster Sets using SQL with Oracle Data Mining – Part 3

Posted on Updated on

This is the third blog post on my series on examining the Clusters that were predicted by an Oracle Data Mining model. Check out the previous blog posts.

In the previous posts we were able to list the predicted cluster for each record in our data set. This is the cluster that the records belonged to the most. I also mentioned that a record could belong to many clusters.

So how can you list all the clusters that the a record belongs to?

You can use the CLUSTER_SET SQL function. This will list the Cluster Id and a probability measure for each cluster. This function returns a array consisting of the set of all clusters that the record belongs to.

The following example illustrates how to use the CLUSTER_SET function for a particular cluster model.

SELECT t.customer_id, s.cluster_id, s.probability
FROM   (select customer_id, cluster_set(clus_km_1_37 USING *) as Cluster_Set
        from   insur_cust_ltv_sample 
        WHERE  customer_id in ('CU13386', 'CU100')) T,
      TABLE(T.cluster_set) S
order by t.customer_id, s.probability desc; 

The output from this query will be an ordered data set based on the customer id and then the clusters listed in descending order of probability. The cluster with the highest probability is what would be returned by the CLUSTER_ID function. The output from the above query is shown below.

NewImage

If you would like to see the details of each of the clusters and to examine the differences between these clusters then you will need to use the CLUSTER_DETAILS function (see previous blog post).

You can specify topN and cutoff to limit the number of clusters returned by the function. By default, both topN and cutoff are null and all clusters are returned.

– topN is the N most probable clusters. If multiple clusters share the Nth probability, then the function chooses one of them.

– cutoff is a probability threshold. Only clusters with probability greater than or equal to cutoff are returned. To filter by cutoff only, specify NULL for topN.

You may want to use these individually or combined together if you have a large number of customers. To return up to the N most probable clusters that are greater than or equal to cutoff, specify both topN and cutoff.

The following example illustrates using the topN value to return the top 4 clusters.

SELECT t.customer_id, s.cluster_id, s.probability
FROM   (select customer_id, cluster_set(clus_km_1_37, 4, null USING *) as Cluster_Set
        from   insur_cust_ltv_sample 
        WHERE  customer_id in ('CU13386', 'CU100')) T,
      TABLE(T.cluster_set) S
order by t.customer_id, s.probability desc;

and the output from this query shows only 4 clusters displayed for each record.

NewImage

Alternatively you can select the clusters based on a cut off value for the probability. In the following example this is set to 0.05.

SELECT t.customer_id, s.cluster_id, s.probability
FROM   (select customer_id, cluster_set(clus_km_1_37, NULL, 0.05 USING *) as Cluster_Set
        from   insur_cust_ltv_sample 
        WHERE  customer_id in ('CU13386', 'CU100')) T,
      TABLE(T.cluster_set) S
order by t.customer_id, s.probability desc;

and the output this time looks a bit different.

NewImage

Finally, yes you can combine these two parameters to work together.

SELECT t.customer_id, s.cluster_id, s.probability
FROM (select customer_id, cluster_set(clus_km_1_37, 2, 0.05 USING *) as Cluster_Set
from insur_cust_ltv_sample
WHERE customer_id in (‘CU13386’, ‘CU100’)) T,
TABLE(T.cluster_set) S
order by t.customer_id, s.probability desc;