Proactive Career Path Management (Talent Management)

Proactive Career Path AnalysisSAP Predictive Analysis for HR

Henner Schliebs, August 2013

© 2013 SAP AG. All rights reserved. 2Customer

How can we analyze the flow of people withSAP Predictive Analysis?

1

Visually explore HR data on employee moves through the company

Apply statistical algorithm to identify groups of employees with similar patterns for changing positions

Derive a model that can predict the success of a change of position based on employee attributes

Use this model to predict the probability of success for some upcoming moves

2 3 4

© 2013 SAP AG. All rights reserved. 3

Our Source data is based on Excel but could also come via HANA or BO

Universe




Here we can preview the data that is to be

imported


The imported data contains information on position changes as well as master data before

and after the move

List of dimensions inside the dataset

Since Excel does not distinguish between dimensions and measures, we need to manually define our key

figures for the analysis


Our newly created measures

Aggregation behavior


Switch aggregation to “average” to calculate average success

ratio of a position change


Let’s rename the measures


Now we will filter out all employees who did not have a performance rating after

their change of position



We repeat the previous steps to also exclude employees who did not have a performance rating before their move

(This is skipped here)


The two filters we have just applied

Our source data has some geographical

information that can be leveraged for analysis



If you have additional data like Region or City you can

create a navigation hierarchy that can be

browsed visually


PA is quite smart and can identify both ISO-coded

and explicitly named geographical data


We have finished preparation of the data and can now proceed to visual exploration

We will start with an analysis of the success ratio of position changes along the country hierarchy we just created


It seems that there are large differences in the success ratio between

countries


Now, let’s look at how the success probability might be linked with moves

across job functions

For this analysis, we will use a Heatmap with the job functions before and after as

axes


Job functions that are quite different seem to go

along with a low probability of success…

…while moves inside a job function work quite

well.


The heat map tells us something about how well certain combinations of job

functions work together, but…

…it does not take into account how frequent certain combinations appear in

the first place.

To analyze this we will look at a Treemap which is basically a Heatmap with a

weighting factor.


The size of the rectangle is now proportional to the total number of moves for

this combination.These combinations are quite frequent and rarely

go well.


Let’s switch to a different view and look at the question: How are success probability and performance rating before the move

tied together?


For employees with low performance ratings a

change of position has a very high probability of

success.

Employees with high performance ratings on the other hand have it

much harder to perform equally well in their new

role.


We see that there are some anomalies but so far we have only explored the data

visually: Nothing has yet been proven.

Therefore we will now apply some statistical algorithms to see which

patterns emerge scientifically.



1





2 3 4


Here you can see the available algorithms that can be applied to your

data.In this area you can combine individual analyses to form a

comprehensive model for understanding relationships.


First we want to apply some filters to the data



Let’s filter out the employees without

performance ratings


We rename the filter so we can keep track once

the model becomes more complex


We configure the second filter to filter out all dimensions that we are not going to

use in our first analyses – just to be more convenient for us.



Here we will have only the employees and only the dimensions we are

interested in

Now we add a classification algorithm

that will try to find clusters of position

changes with similar values in the dimensions


We’ve added the algorithm twice to calculate two

different models on the same data and compare them



We will use all the available information to see what kind of clusters the algorithm can

find

This first analysis will try to categorize the available data

into five clusters…



…while our second analysis will try to find ten clusters.


Let’s run the analysis!



Here we can see to which cluster a record was assigned by the

model…

The different components of our

model can be viewed along with their

intermediate results

…but we want to see a summary of the models

first.


Here we see the number of records that were assigned to each

cluster in the 5-K analysis

This chart shows how homogenous (“dense”)

the clusters are and how different from one

another

In this chart we can look at individual dimensions and

check which dimension values were how common in

each cluster

This chart shows a profile diagram for each cluster

(the axes are the dimensions that were put

into the analysis)


Example: In cluster 1 the average time in position

before the move was 1.53 years

We can see that most employees were assigned to

this cluster and that this cluster is very heterogenous

– this is a strong indicator that 5 clusters are not enough to sufficiently

describe our data


Is the model with ten clusters better than the one with five?

Let’s check…


Two of the ten clusters are quite heterogenous – but not

as bad as in the previous analysis.

Also: They are not as big as before – that’s a big

improvement.


So we would prefer to use ten clusters. But what describes these clusters?

Ordinarily one would use the charts on this page or a custom visualization to find

out how the clusters are comprised.

We are not going to pursue this here but are going to enhance our model with a

second type of analysis.



1





2 3 4


To answer the question: “When is a move going to be

successful?” we will use a decision tree.


We connect the decision tree to the first filter since we want to use additional columns that were not

necessary for our previous analyses.


From our visual exploration we have formed the

hypothesis that the following parameters affect the probability of success:

Job Function beforeJob Function after

Performance Rating beforeCountry AfterFrom our clustering we have

found that Time in Position is important.

We will additionally add:

Job Level beforeJob Level after

Change of Career Path FlagTotal Tenure

And we are going to model whether a move will be

successful


We will save the decision tree as a custom model so

we can apply it to a different dataset later.


Run the analysis!



Here we can see whether our models thinks that a

certain move will be successful

But let’s look at the decision tree directly


“Yes” = Position Change will be

successful Green = relative share with

successful moves

Dimension on which decision for

categorization is made

Result 1: If low or medium performance rating, very high probability that any move will

lead to improvement


Result 2: For very good ratings the move has a much higher chance of being successful if

employee has been in his previous position for a long

time (>3.7 years)


Result 3: Promotions for good employees rarely go well when they go along with a change of job function – especially if the

employee was in R&D or support before

Result 4: For more customer oriented Job Functions

Promotions can go very well in certain countries

Complex decision tree! Let’s see how well it fits the data

overall.


This chart compares the relative frequency of actually successful moves against the predicted success of a move

We can see that the model has some trouble predicting

negative success where it has an accuracy of about 55%

But if the model says a move is going to be successful,

chances are quite high that it will be so in reality as well


We have now created a prediction model for success of a position change based on

historical data.

Now we will apply this model to some upcoming moves and see what the model

predicts for these employees.



1





2 3 4


We add a new dataset into our analysis that contains data on

the upcoming position changes for seven employees




In the new dataset we have all information that we need to

apply our decision tree







Here we have the predicted results after applying our model

along with the predicted probabilities for positive and

negative success

Based on the prediction – Mr. Gonzales’ move will probably lead to a lower performance

after the move

Mrs. Adams’ move on the other hand will most probably go

well!


Here we can use visualizations to compare the different

success probabilities for our seven employees.

Last but not least, we can share our models and

predictions with our colleagues – directly from the application.


Select what you would like to share – e. g. your datasets,

results, visualizations,…

And select how you would like to share it.

Henner SchliebsAnalytics Product Marketing

@hschliebshenner

[email protected]

hschliebsFlow of people blog: http://bit.ly/SAP-TA-blog Talent Analytics Video: http://bit.ly/TA-YT

https://twitter.com/hschliebs

https://twitter.com/hschliebs

http://www.linkedin.com/in/henner

mailto:[email protected]%20?subject=nice%20meeting%20you

http://bit.ly/HS-YT

http://bit.ly/SAP-TA-blog

http://bit.ly/SAP-TA-blog

http://bit.ly/TA-YT

http://bit.ly/TA-YT

Business

Proactive Career Path Management (Talent Management)