Skip to Content

Source: SAP

This was an ASUG BITI webcast given by SAP

Figure 2: Source: SAP
Recommend modeling iwth PAL – algorithms are implemented in server, where data resides, see much better performance

Figure 3: Source: SAP

PAL is for data scientists

Credit card data, which customer is credit worthy (classification)

Predicting house prices based on characteristics such as # of rooms, using regression

Cluster similar customers to do targeted marketing campaigns

 

Time dependent models, sequential pattern modeling to issue coupons

Figure 4: Source: SAP
Color coding indicates investments

Figure 5: Source: SAP
Training the model for random forest modeling

Figure 6: Source: SAP
Model scoring PAL code with confidence level

Figure 7: Source: SAP
Other options include integration with R
Connect from R studio, connect from ODBC

Figure 8: Source: SAP

R integration looks like a stored procedure

Figure 9: Source: SAP
Build models in TensorFlow and call in HANA – like a stored procedure

On the HANA side, you have the Application Function Library

With SPS02 – create the EML/AFL – interfaces between HANA and TensorFlow server

On the right side, build and train TensorFlow and upload to TensorFlow server, and then consume from HANA side

Connect through a Google Remote Function call

TensorFlow Serving Server can run in same box in HANA in development; should be separate in production

Scope is for scoring

Figure 10: Source: SAP

Train the model in TensorFlow

Figure 11: Source: SAP
Step 1 create remote source, host and port of TensorFlow server

Next map the model to the remote source; insert test model

Next – any config changes get applied immediately

Then check all connections are working before start using in the application

Figure 12: Source: SAP

Generate the EML in HANA and then call procedure using the input and output table

Figure 13: Source: SAP

 

Machine learning in HANA end to end; depends on type of use case

Machine learning is not just developing models, but how do these models get optimized and in a scalable real-time way

Figure 14: Source: SAP
Customer churn prediction with PAL to build a model

You can grow decision trees, output is the class

In PAL, you have fine grained control

Figure 15: Source: SAP

First step is to train the model, populate the parameter table

Figure 16: Source: SAP

Create table to store model

Capture variable importance with a table

Store the out of bag erorr

Store confusion matrix of the model

Figure 17: Source: SAP
Train the model, call the function

Decision trees are stored in PMML format

Out of bag error, the variable importance

Confusion Matrix is also output

Only 3 out of 14 cases did the model predict inaccurately

Figure 18: Source: SAP

Now built the model, want to predict the churn, and will customer to be retained

Create the parameter table
Create the results table for the results of scoring

Prediction and confidence

Used PAL to train the model and then use it to predict scenarios

Figure 19: Source: SAP

Could also build using Web IDE

Figure 20: Source: SAP

Push execution close to data

Figure 21: Source: SAP

 

Think of performance in terms of batch and real-time

SP02 enhancements

Real time prediction with SPS01

Figure 22: Source: SAP

 

Decisions need to happen in real time

 

Figure 23: Source: SAP

The model can remain in memory

Figure 24: Source: SAP

Partitioning – score in parallel

Ability to do large batch style processing in parallel

Figure 25: Source: SAP

Streaming analytics engine to take input from a variety of sources

Figure 26: Source: SAP

Train data as they arrive
Predict in real time

Figure 27: Source: SAP
A summary of a jam-packed webcast

Upcoming ASUG webcasts:

 

To report this post you need to login first.

Be the first to leave a comment

You must be Logged on to comment or reply to a post.

Leave a Reply