RE4ES- Holistic Explainability Requirements for End-to-end ML in IoT Cloud Systems
- 1. Holistic Explainability Requirements for End-to-end ML in IoT Cloud Systems My-Linh Nguyen, Thao Phung, Duong-Hai Ly, Hong-Linh Truong {linh.m.nguyen, thao.phungduc, duong.ly, linh.truong}@aalto.fi Aalto University https://rdsea.github.io International Workshop on Requirements Engineering for Explainable Systems (RE4ES- IEEE RE 2021)
- 2. Motivating example: ML for Base Transceiver Stations (BTS) • ML solution in IoT Cloud systems for predicting behaviors of BTS equipment and infrastructures • Dynamic inferences of near real-time IoT data • Challenges: • Multiple stakeholders, each stakeholder is related to only a part of the ML development. - How to do we identify and capture the requirements for explainability 2 BTS owner (Telco) Third-party providers: Equipment manufacturers, Electricity/power providers, Equipment suppliers, etc. ML development team (software and telco professionals) 21/09/2021 My-Linh Nguyen, Thao Phung, Duong-Hai Ly, Hong-Linh Truong - RE4ES’21 (Predictive) Maintenance Company Multiple relevant stakeholders in predictive maintenance
- 3. Methodology - Holistic approach • Explainability requirements covering: • multi-stakeholders • multiple explainability aspects • for an end-to-end ML system 3 21/09/2021 My-Linh Nguyen, Thao Phung, Duong-Hai Ly, Hong-Linh Truong - RE4ES’21
- 4. Scoping stakeholders 4 Goal: • Who is the explaination for? • What is their relationship? Direct: explanation triggered through “explain” task, e.g., between developers for feature engineering and model training Indirect: following a chain of direct dependencies, e.g., BTS owner and MLOps team 21/09/2021 My-Linh Nguyen, Thao Phung, Duong-Hai Ly, Hong-Linh Truong - RE4ES’21
- 5. Scoping ML Processes • Identify processes/ tasks stakeholders responsible for or interested in • Map stakeholders to relevant phases, covering ML requirement elicitation, service design, and development. • Explainability for end-to-end ML preserves informative connection 5 Stakeholders Aspects ML processes 21/09/2021 My-Linh Nguyen, Thao Phung, Duong-Hai Ly, Hong-Linh Truong - RE4ES’21
- 6. Scoping ML explainability aspects • Each stakeholder works/ supervises directly specific entities (data, ML models, etc.). • Overall, a wide range of entities, and associated constraints (metrics and aspects) • Analyzing their role in each process/task => requirements and constraints 21/09/2021 My-Linh Nguyen, Thao Phung, Duong-Hai Ly, Hong-Linh Truong - RE4ES’21 6
- 7. Example in the BTS case 21/09/2021 My-Linh Nguyen, Thao Phung, Duong-Hai Ly, Hong-Linh Truong - RE4ES’21 7 Stakeholders: BTS owner, predictive maintenance software company, technical project managers of the project, ML developers. Scope of BTS example
- 8. Explainability requirements in ML processes/tasks Feature Engineering (FE): • Why certain feature engineering techniques (e.g.: feature split, scaling) are used? • In BTS, we may use some FE techniques: feature split, grouping operations, scaling. Training: • Who did the training? (manually or by AutoML tools). Serving: • Which service provider is used? • Where is the solution deployed? 21/09/2021 My-Linh Nguyen, Thao Phung, Duong-Hai Ly, Hong-Linh Truong - RE4ES’21 8 Data & Model Collection: • Who collects or provides data? • E.g.: BTS data is provided by BTS monitoring system/company. Data Preprocessing: • Who did the pre-processing? • Why did the stakeholder choose to perform specific techniques? Model Development: • Who decides machine learning methods?
- 9. Requirements in explainability aspects 21/09/2021 My-Linh Nguyen, Thao Phung, Duong-Hai Ly, Hong-Linh Truong - RE4ES’21 9 Data-related aspects are crucial in ML solutions in IoT Cloud Systems
- 10. Examples: capture data related requirements 21/09/2021 My-Linh Nguyen, Thao Phung, Duong-Hai Ly, Hong-Linh Truong - RE4ES’21 10 Data Drift Impact IoT Data – important data quality metrics An example of comparison between BTS data distribution regarding feature sensor reading temperature of air-conditioner in one station within one hour tested using Kolmogorov Smirnov (KS) Test.
- 11. Concrete explainability requirements elicitation 21/09/2021 My-Linh Nguyen, Thao Phung, Duong-Hai Ly, Hong-Linh Truong - RE4ES’21 11 Transform surveyed data and ingest the data into data services Interview concrete stakeholders Continuous requirements update
- 12. BTS Explainability Requirement Example 21/09/2021 My-Linh Nguyen, Thao Phung, Duong-Hai Ly, Hong-Linh Truong - RE4ES’21 12 Example of requirements based on our proposed methods
- 13. How to utilize our result? • Integrate explainability requirements into cloud-native DevOps ML in IoT cloud data • Provide input for identifying and managing diverse types of trails for explanation tools - Experiment data, metadata about data, metadata about models etc • Identify, recommend and configure explainer tools - Combine suitable explainer for a particular case 21/09/2021 My-Linh Nguyen, Thao Phung, Duong-Hai Ly, Hong-Linh Truong - RE4ES’21 13
- 14. Conclusion and future works Conclusion • We identify and classify explainability requirements engineering through: • Involvement of relevant stakeholders • End-to-end data, model, and service engineering processes • Multiple explainability aspects. Future work • Tools and services for collecting different type data • Composition of explainer for data drift in end-to-end 21/09/2021 My-Linh Nguyen, Thao Phung, Duong-Hai Ly, Hong-Linh Truong - RE4ES’21 14