Thursday, 30 May 2024

MARKET BASKET ANALYSIS-TURBODATA

 










 

Market Basket  Analysis is a modelling  technique based upon the  theory that if we buy  “ ITEM A” then there is a probability of buying “ITEM B”.
Purpose of Market Basket Analysis  through Turbodata
The purpose of market basket analysis is to determine what products customer purchase together .i.e. if a customer purchases milk  then there is possibility that the customer will also buy  sugar.MBA(Market basket analysis) is used to analyze the customer behavior.
Why Market Basket Analysis in Turbodata?
·       Data has been cleansed and consolidated in Turbodata.
·       Maximum data redundancies have been removed and joins have been minimized. All the business logic from the ERPs has been pre coded. This process helps us in the following ways:
o   The reporting team can use parameterized queries to get the required outputs. There is no need to store all the data in pre aggregated form. This is important since there are a large number of item combinations that are possible
o   Turbodata’s market basket analysis solution offers analysis at the reporting layer without the use of expensive reporting tools.
o   The solution is available in both Tableau and PowerBI.
o   Multiple users can access  the report. Turbodata offers analysis at any point in time, across any number of locations or any combination of items.

o   Facility for large scale deployment of the code because of the usage of table variables and functions.
STEPS:
STEP1  : A  master procedure with two input parameters was created with date.
STEP 2:  Five variables was created for the  calculation of confidence ,lift and support.
·       @n_total :total  number of order
·       @n_dual: number of invoices having more than ONE item
·       @n_item1:number of invoices having ITEM1
·       @n_item2: number of invoices having ITEM2
·       @n_item1_item2: number of invoices having both ITEM1 and ITEM2.
STEP 3: Calculate Confidence , Support and Lift value.




So for this we have been created a function instead of procedure because:
·       Function is compiled and executed every time whenever it is called .It is just like a formula we use.
·       We can add the parameter according to the requirement.
·       When we need only input parameter.
There are two types of function we use  in our code
1.    Scalar valued  function
2.    Table valued function

*   Scalar valued function
Scalar valued function are used when we have to returns  a single value(integer,decimal or any).it might or might not be related to tables in our database.
*   Table valued function
Table valued function returns the specified columns for row in our table meeting the given criteria.

So for the above process  we use the following solutions:
*   Calculation of @n_total
N_total is the sum of total orders.Total number of orders should be come from  item_supplier_order at a particular date.so we join item_supplier_order table with dim_date and we take only that date which is less than or equal to the @date.


Here is the snapshot for total order.

*   Calculation  for@n_dual
 This is the function which we have been created for invoices having more than one order.Here we need partyledgername from item_supplier_order,item  from item_sales_order and date from dim_date.so we join item_supplier_order  with item_sales_order on the basis of composite key ,the result will further join with dim date.
For more than one item we take the count of dim_item_id having count greater than one.


Attached is the snapshot from the Turbodata:

*   Calculation for @n_item1
This is the function we created for the item1.
We took order_no from item_supplier_order,dim_item_id from item_sales_order  and item from dim_item having the condition that count of dim_item_id is greater than one.




*   Calculation for @n_item2
The similar process is used for the calculation for item2 as we have been done for item1

*      Calculation for both the items @n_item1_item2
                      This is the function created for both the ITEM1 and ITEM2.



*   Confidence is defined as division of the frequency of both the item to item1.
i.e., Confidence=@n_item1_item2/@n_item1





*   Support  is defined as the division of occurrence of both the items to total.
i.e  support=2n_item1_item2/@n_total




*   Lift is defined as the division of the support of both the item to muliplication of support A and support B.
i.e  Lift =@support/@support_x * @support_y


How the Market Basket Analysis code has been developed over compressed database.
How Turbodata helped lower the costs of developing a datawarehouse and helped the end clients do predictive analytics with quickness and ease-applicable for retail sales and inventory(Website: www.mnnbi.com)

Purpose of the development of the product: The Turbodata team intends to reduce the costs of the analytic solutions by creating a single platform for ETL, Reporting, Reporting development and predictive analytics. The team also intends to provide the best in class analytics on the same machine on which the ERP is running or with the addition of minimum hardware requirements for the end client. This has been done to develop scalable systems that can be deployed over a large number of customers(with limited budgets) with ease(deployment, delivery and usage) and convenience(maintenance).
The end goal is to increase derisking and predictability for the end clients at lower costs.




Methodology for achieving the required ends for the end client:
·         Turbodata adopted the Inmon methodology for datawarehouse development so that multiple data sources could be added onto the same datawarehouse. That is the change from one data source to another was done with ease. More details on the attached web page link: http://mnnbi.com/mnnbi_tallydataconsolidation.html


o   The benefits of the normalization of data were as follows:
§  The incremental data load took minimum time and had minimum impact on the source system. The ETL team was able to commit the incremental data load to a maximum of 2GB RAM from multiple source systems. The source systems did not hang with the incremental data load working.
§  Error handling was done with ease at staging layer.
§  Massive data compression took place due to reduced data redundancies.
§  The business logic was coded between staging and the ODS layers thereby reducing the length of the final sql code.
The attached video shows a more detailed description of the benefits listed above:
The joins were reduced in the data mart layer(over which a reporting layer was built).

The ETL team was able to develop extremely complex reports using the datawarehouse as in the attached sample:

Due to the data compression for most projects the ETL team are able to bring the data within 1 GB. Hence the desktop version of Microsoft Power BI could be used free of cost for the end client.

Reducing the cost of predictive analytics solutions
 Most of the end clients use high end predictive tools over the datawarehouse/ over the direct data extract from various source databases. With large datasets predictive analytics using in memory solutions entails high usage of RAM. The ETL team has gone around this issue in the following manner:
o   A seamless environment was created for ETL, reporting and thereafter predictive analytics on SQL/C# and .Net. The reasons for the same are attached herewith:
§  Maintenance becomes easier since there is a single platform for all aspects.
§  The cost comes down since the resources to be used for ETL can also be used for predictive analytics
§  Error handling becomes very easy since errors can be captured before in the


Hypothesis testing
Based on the hypothesis testing, the ETL team developed ARIMA analysis and Market Basket analysis in SQL using seamless integrated set of stored procedures. That is the ARIMA analysis flowed from the datawarehouse A,B,C categorization. The ETL team thus reduced the requirement for high end R and Python developers to code over the datawarehouse thereby presenting a seamless solution to the end client on a 8GB RAM machine.

Benefits to the end client:
·         The end client gets immediate and confirmed peace of mind and satisfaction through immediate deployment of predictive and forecasting analytics modules.
·         No additional hardware/software requirements need to be taken
·         The costs are way lower for the end client.
·         Large scale deployment is possible with the given set of solutions.
Please check the attached video for the same:
A more detailed video is attached herewith:


Example of predictive analytics with Turbodata: Example of predictive analytics-Turbodata


The PowerBI code has been developed to enable the following:·       Run the market basket analysis code based on 2 input item parameters across any filter date and location ranges.
·       Removes the dependence on SQL server for the implementation of Market basket analysis.
·       Faster execution times since the measures and columns are calculated on a columnar basis instead of row by row basis.
The attached is the methodology of Market basket analysis using PowerBI over Turbodata

Step 1:Import the  file from SQL server.


Step2:Three parameters are declared.i.e. item_1,item_2,date_1.This has been done through Power Query Editor.
For creating parameters:
·       Open power query editor
·       Go to manage parameter
·       Create a parameter

Set the parameter of date_1 in item_supplier_order

Step 3:we have to calculate support,confidence and lift value.
calculation of support :
Support=f(X,Y)/total
Where f(X,Y) is the instances of occurrance both the “ITEM A” and ”ITEM B” in the invoices for the given time period.
Total is the count of the total sales order nos. for the given time period

Attached is the snapshot from dax function for f(X,Y):

Snapshot for total:



Snapshot for the calculation of support:


 Snapshot for the calculation of item A only:
Calculation for the confidence:
 Confidence =F(X,Y)/F(X)
Snapshot for the calculation of confidence:
Where P_X_Y is the occurrence of both the item
And P_X is the order no where there is item A.
Snapshot for the calculation of P_X:

Calculation for lift:
Lift =support_1/(support_x*support_y)
Where  support_x  =f(X)/total
Support_y=f(Y)/total

Snapshot for the calculation of lift value:

Snapshot for the calculation of support_x:
Snapshot for the calculation of support_y:
Attached is the summary of final output for the given set of input items(CPU and ALMIRAH).

Market Basket Analysis through Tableau

Market Basket Analysis has been done through Tableau .
There are four data sources we have been taken for the calculation of market basket analysis.
1.    CPU
2.    ALMIRAH
3.    MARKET_BASKET

4.    ORDER_MASTER


We have to calculate the lift,confidence and support value in Market Basket Analysis.So we  have  take the two items ‘CPU’ and ‘ALMIRAH’.
Make three parameters.
1.    Date_1
2.    Item_1
3.    Item_2




We have take the order no. where the CPU has been purchased.
So we join the item_supplier_order with item_sales order on the basis of item_supplier_order_id , dim_date with dim_date_id and dim_item with dim_item_id.


We have pass the parameter through Custom SQL.

Attached is the snapshot from the Tableau for the date parameter:

Similarly attached is the snapshot from Custom SQL to pass the item_1(CPU) parameter:


Attached is the snapshot from Custom SQL to pass the item_2(ALMIRAH) parameter:

This is the another database for the calculation of order_No where almirah are purchased.


MARKET_BASKET is the database where we calculate the order no where both item_1 and item_2 both are purchased.i.e. “ALMIRAH”,”CPU”.


This database is used to calculate total number of invoices(order_no).

For the calculation of the order_no of item_1 we have taken the count distinct of item_1


For the calculation of the order_no of item_2 we have taken the count distinct of item_2


Attached is the snapshot for the calculation of the order_no for both the  item_1 and item_2.



Attached is the calculation for the total order_no.



Based on the above calculation we calculate the confidence ,lift and support value.
Attached is the snapshot for the  calculation of confidence:
Confidence is the value that if a customer purchase item “A” then what likelihood  he will buy item “B”.




Attached is the snapshot for the  calculation of support:
Support value calculation for how the strong is the base.




Attached is the snapshot for the  calculation of lift:



Attached is the snapshot of the resultant:
We take two items “CPU” and ”ALMIRAH” which cannot be a combo so its lift value is less than 1.
Note:
Lower the value of lift lower is the strength of the combo and higher the value of lift higher is the strength of the combo that customer like.

ARIMA model in R

Data is made of noise and signal .
In short time period the probability of noise in the data is greater than signal. For e.g. if we analyze the data on hourly basis ,the  probability of noise is 99.5% and that of signal is 0.5%.
In long time period , the proportion of noise reduces and the proportion of signal increases.For e.g. if we analyze the data on monthly basis, probability of noise could be 50% and that of signal is 50%.
The purpose of tike series forecasting is to segregate the noise from the signal for better management decision making.
ARIMA  stands for :
AR:Auto regressive(p)
Y(t)= c+alpha*y(t-1): Understand the time dependency
I:integration(d)
Y(t)= c+alpha*y(t-1)+beta*y(t-2)+error: understand behavior of error term. Check for stationarity(check whether error is hiding signal patterns).
MA :Moving Average(q)
Y(T)=c+alpha*e(t)+beta*e(t-1). MA indicates the dependence of y(t) on the errors in the previous time periods
Components of ARIMA Model:
}  Trend: A long-term increase or decrease in the data is referred to as a trend. It is not necessarily linear. It is the underlying pattern in the data over time.
}  Seasonal: When a series is influenced by seasonal factors i.e. quarter of the year, month or days of a week seasonality exists in the series. It is always of a fixed and known period. E.g. – A sudden rise in sales during Christmas, etc.
}  Cyclicity: When data exhibit rises and falls that are not of the fixed period we call it a cyclic pattern. For e.g. – duration of these fluctuations is usually of at least 2 years.

Assumption of ARIMA model
·        Data should be stationary: the error terms should not have time dependency
·        Data should be univariate: the regression model is based on only one variable.


Sample code:
data = read.csv('http://ucanalytics.com/blogs/wp-content/uploads/2015/06/sample_data.csv')
data = ts(data[,2],start = c(2003,1),frequency = 12)
plot(data, xlab='Years', ylab = 'sample)
The above code is from R. it reads the data which is csv type. ’ts’ is the time series function which reads the data starting from  first month of 2003 and read plot the data of every month.

Where x-axis is for the “years” and y-axis is for the “sample”.


plot(diff(data),ylab='Differenced sample Sales')
This is  for the difference of y(t)-y(t-1).



plot(log10(data),ylab='Log (sample Sales)')

This graph is for the log value of given sample sale.
plot(diff(log10(data)),ylab='Differenced Log (sample Sales)')              
This graph is for the log difference of sample sales.
ACF:ACF stands for auto-correlation function. It describes how well the present value of the series is related with its past values. A time series can have components like trend, seasonality, cyclic and residual. ACF considers all these components while finding correlations hence it’s a ‘complete auto-correlation plot’.
PACF:PACF stands  for partial auto correlation function. Basically instead of finding correlations of present with lags like ACF, it finds correlation of the residuals. So if there is any hidden information in the residual which can be modeled by the next lag, we might get a good correlation and we will keep that next lag as a feature while modeling.
Attached is the code :
par(mfrow = c(1,2))
acf(ts(diff(log10(data))),main='ACF sample Sales')
pacf(ts(diff(log10(data))),main='PACF sample Sales')


Epilogue
The ETL team has been inspired by the following management books:
·         ‘Profit Beyond Measure’ by Thomas Johnson and Anders Brohms
·         McKinsey Mind by Ethan Rasiel and Paul Friga.
·         Blue Ocean Strategy by W. Chan Kim and Renee Mauborgne
·         Better India, Better World by Narayana Murthy





In case of any requirements please contact the following:

Name: Apoorv Chaturvedi
Email: support@mndatasolutions.com;mndatasolutionsindia@gmail.com;support@turbodatatool.com
Phone: +91-8802466356

No comments:

Post a Comment

Initial and Incremental data Load Template by M&N Business Intelligence-SAP Data Services

  INCREMENTAL LOAD/CREATION OF DIMENSION TABLE LOGIC AT SAP DATA SERVICES END CLIENT In this particular document we shall be looking at the ...