Using AI to evaluate cassudy's screening results

AI is a well established method in evaluating data patterns.

The time when Artifial Intelligence was an exotic and extremely innovative „tool“ is obviously over. Nowadays AI is an integral part of many IT applications. The most well known example for successful AI application is image recognition or more specifically face recognition.

There is a whole bunch of tools [1] available to apply AI without the need to go into the (more complex) details of the underlying mathematical or statistical methods. The tools hide this complexity from the programmer and the programmer mainly describes and designs the business problem.

How AI is used to evaluate results of a cassudy case study

Cassudy uses a neural network [2] of three layers – an input layer, that receives the data points captured during the processing of a cassudy case study – an output layer, that provides the result of the evaluation - and a single layer in between (hidden layer) that is needed to generate reliable and meaningful results.

In a cassudy case study several hundred data points may be collected.

These data points are, for the sake of simplicity, consolidated and thus reduced to a set of 25-30 data points, which finally are allocated to the input layer neurons.

The neural network will use the input data to assign such data set to an output class.

Cassudy typically uses 4 output classes: „needs improvement“, „partially meets expectations“, „fully meets expectations“, „exceeds expectations“. The classes look like a performance rating of an individual and in fact this is what they are.

Using a properly „trained“ network (more on training a neural network see below) the network predicts, based on the measured input data points, to which class the processor of the case study will most likely belong. Meaning the neural network delivers a prediction of his or her most likely future performance.

Important steps to set up a properly functioning neural network

For the neural network to function properly it is vital to train the network. In order to do this, we need a set of training data.

More explicitly this means we need data sets that relate input data to an output class – and we need many of such data sets.

How can these training data be obtained?

The input data can only be generated by running the cassudy case study many times with different individuals.

Then these same individuals need to be allocated to an output class. This allocation could be done using other rating mechanisms (like a rating by the superior, results from other tests or objectively measured success of this individual in the role under consideration).

If reliable data sets are available these sets can be used to train the neural network.

Training in this case means, that the parameters of the neural network are set such that the difference between the output class allocation of the neural network (i.e. the network´s prediction) and the allocations as per original training data set is minimal.

In the cassudy environment, an accuracy rate of >80% is achievable. The more reliable the training data is, the more reliable the predictions of the neural network will be. [3]

Practical implications of using AI support in evaluating cassudy case study data sets

The first implication is, that we need a larger pool of training data to make the neural network „more accurate“.

Several hundred data sets will typically be needed to achieve the above mentioned accuracy. Less training data simply means less accuracy.

So it makes sense for companies to apply the exact same case study many times with many different individuals and to capture the results properly in (new) training data sets which can then be used in turn to fine-tune the neural network.

Secondly it means that altering the case studies frequently is not a good idea as this would dilute the accuracy of the predictions of the network again.

Thirdly it may be recommendable to use cassudy pre-configured case studies (not customized) because for those case studies a more reliable AI based prediction may be available.

If I do not meet the requirements for a meaningful AI based evaluation – is cassudy then useless for me?

Certainly not!

Cassudy case studies always generate a meaningful report on how the case study was processed by the individual, providing information about personality traits of the individual as well as providing insights on decision making behavior.

Such behavioral indications can still be evaluated and interpreted by an experienced expert (People Development Manager, Recruiting Expert, Superior Line Manager - just to name a few) and used in respective staff decisions.

This would, by the way, happen in any evaluation scenario as no serious organisation or HR department would blindly rely on what AI tells them.

The AI component within cassudy can be seen as the „icing on the cake“ – but the cake is delicious in any case!

Interested in more details?

Dr. Werner Sohn

CEO and Founder of cassudy

werner.sohn@cassudy.com

[1] For a quick introduction into neural networks see for example https://www.youtube.com/watch?v=VB1ZLvgHlYs&t=396s

[2] Cassudy uses the Pytorch library for result evaluation. Pytorch tutorials are available on youtube.com covering complexity levels from simple to advanced.

[3] Reliability here means that statistically 4 out of 5 predictions will be correct. So there is still a non neglectable remaining error potential