Code with SAP Labs India - Discover, Design, Deliver

6701 Registered Allowed team size: 1 - 4

Winners are announced.

POC Submission
starts on:
Mar 16, 2021, 03:30 AM
ends on:
May 16, 2021, 06:25 PM


Explainable AI

Bringing Explainable AI into Business Applications with SAP


Addressable markets for AI-centric applications are projected to increase significantly from 2019 to 2024. SAP is building more AI driven business applications by infusing AI into the workflows or by using AI to provide additional values via add-on use cases. This new paradigm of building AI based applications brings in more challenges in terms of explaining the ML outcomes from the business context. Hence, explaining the results of the algorithms with local and global explanations that can be leveraged across the business applications (for eg: transaction systems) is becoming increasingly important.

Challenge: The target is to develop a Model agnostic Explainable AI (XAI) framework/ library that can explain the classification of different kinds of datasets. The team that comes up with the best algorithm for end-user understandable explanation framework wins.


[Minimum]: The XAI model should:

  • Work for the different kind of classification models given (Model Agnostic) and different kinds of datasets. The datasets can contain text, numerical and categorical fields. If the algorithm has various flavours or add-ons, depending on dataset type or classification algorithm, that is okay.
  • Generate explanations which are understandable for end-users that can increase trust of the generated outcomes and bridge gap between predictions and expectations.
  • Generate both global and local explanations.
  • The solution should be able to handle text embedding techniques such as: Bag of Words.


  • Explanations need to be context-specific to the use case.
  • Understand the problem deep enough to generate explanations that are as close to explanations by a human. The explanations should be in natural language and should be understandable by business users.
  • Performance considerations for volume of data – Time taken to generate explanations should add least overhead on performance.
  • Multi-level explanations - granular level of explanations that is well represented in a usable form. Eg: flow from high level explanations to low level explanations which users can easily navigate and understand.
  • A content specific explainable framework/library that can be integrated easily with business applications.
  • Technical approach using the open sources, cloud, k8s etc that could help build the solution faster.
  • The solution should be able to handle text embedding techniques such as: Bag of Words + Word Embeddings (eg: Word2Vec).


  • Model agnostic way of generating explanations for all supervised learning problems (Image Classification, LSTM models etc).
  • Visualization framework which can help depict the explanations and can be integrated with cloud solutions.
  • Start with predictions -> Generate explanations -> Assess generated explanations -> Convert into meaningful explanations -> Visual representation of explanations.
  • The solution should be able to handle text embedding techniques such as: Bag of Words + Word Embeddings + Sentence Embeddings (eg: Universal Sentence Encoder).

Dataset: Training Data

The dataset given consists of categorical, numerical and text fields. The XAI framework should be able to explain the classification by these models : a Linear SVC Model, a Deep Neural Network with 3+ hidden layers and a CNN. Ability to explain more complex models such as LSTM and RNN is a bonus. 

Note: Your solution will be evaluated using different classification datasets and classification algorithms. The solution should be robust enough to generate business explanations for any dataset (comprising numeric, categorical and text features, and images [advanced]) that is given as an input.

Submission Template

Overview Session Recording

Click here to view the PPT for understanding this problem statement

Social Share

View All Notifications