Human-oriented data science has emerged as a novel perspective to adapt the learning process to the needs of the data consumer [Ara16]. Putting the human “in the loop” seeks to overcome some limitations of traditional learning processes, such as the difficulty to capture expert knowledge, imprecise goals, lack of context and mistrust on automatic results [Vid14]. Within this area, interactive machine learning (iML) proposes a new approach in which the data consumer takes part in some steps of the learning process [Dud18]. In fact, iML has already succeeded in enhancing some steps like feature selection [Cor19], sample labelling [Kim18], algorithm training [Li19] and data visualization [Sac17] thanks to the human intervention.
Since 2019, iML is being explored with the goal of making the learning process more explainable. Explainable artificial intelligence (XAI) is a general area that studies how AI systems work internally, under the assumption that more transparent decisions will increase the trust in the results, among other positive effects [Bar20]. Explainable methods can be classified depending on what they explain or how they infer knowledge [Gui18] but are not designed with the target user (data analyst or data consumer) and the application context in mind [Tom18]. Interactive approaches can help to achieve this, adapting the generation of explanations based on human’s feedback. Some recent iML proposals related to XAI let the data consumer comment on the algorithm recommendations looking for a more personalized experience [Koh19] or interact with a classifier to edit its rules [Voj20]. However, current explainable methods do not allow data consumers directly guide the generation of explanations, adapting them to their preferences or level of expertise.
Therefore, we plan to work on:
- The design and development of contextualized explainable methods by using highly configurable techniques like grammar-based genetic programming. This idea will be explored for the two main categories of explainable methods, those that build simplified decision models (global model explanation) and those that analyse the influence of features in particular predictions (local outcome explanation).
- The adaptation of the proposed methods to interactively learn from the human’s feedback while keeping the fidelity of the learning process. The data analyst will participate in building simplified methods, adapting their characteristics and parameters. For feature analysis, the data consumer will be allowed to dynamically select the desired features, thus generating user-oriented explanations.
- The proposed methods will be validated in areas like software analytics and predictive maintenance in which the team has worked in the past. Both areas can benefit from explainable methods to understand the causes of predicted failures.
In this field, our team is already experienced in the design and evaluation of highly configurable and interactive computational intelligence techniques [Lun13, Ram18], its application to iML and XAI being a new challenge.
[Ara16] C. Aragon, et al. “Developing a research agenda for human-centered data science”. Proc. ACM Conference on Computer Supported Cooperative Work and Social Computing, vol. 26, pp. 529 535. 2016.
[Bar20] A. Barredo Arrieta et al. “Explainable Artificial Intelligence (XAI): Concepts, Taxonomies, Opportunities and Challenges toward Responsible AI”. Information Fusion, vol. 58, pp. 82-115. 2020.
[Cor19] A. Correira, F. Lecue.“Human in the loop Feature Selection”. 33rd Conference of the Association for the Advancement of Artificial Intelligence. 2019.
[Dud18] J. Dudley, P.O. Kristensson. “A Review of User Interface Design for Interactive Machine Learning”. ACM Transactions on Interactive Intelligent Systems, vol. 8(2), article 18. 2018.
[Gui18] R. Guidotti et al. “A Survey Of Methods For Explaining Black Box Models”. ACM Computing Surveys, vol. 51(5). Art. No. 93. 2018.
[Kim18] B. Kim, B. Pardo. “A human in the loop system for sound event detection and annotation”. ACM Trans. Interactive Intelligent Systems, vol. 8(2). 2018.
[Koh19] S. Koh et al. “Personalizing the Prediction: Interactive and Interpretable machine learning”. Proc. 16th International Conference on Ubiquitous Robots, pp. 354 359. 2019.
[Li19] G. Li, R. Gomez, K. Nakamura, B. He. “Human Centered Reinforcement Learning: A Survey”. IEEE Trans. Human Machine Systems, vol. 49(4). 2019.
[Lun13] J.M. Luna, J.R. Romero, S. Ventura. “Grammar-Based Multi-Objective Algorithms for Mining Association Rules”. Data and Knowledge Engineering, vol. 86, pp. 19-37, 2013.
[Ram18] A. Ramírez, J.R. Romero, S. Ventura. “Interactive Multi-Objective Evolutionary Optimization of Software Architectures”. Inf. Sciences, vol. 463-464, pp. 92-109. 2018.
[Sac17] D. Sacha et al. “What you see is what you can change: Human centered machine learning by interactive visualization”. Neurocomputing, vol. 268, pp. 164 175. 2017.
[Tom18] R. Tomsett et al. “Interpretable to Whom? A Role-based Model for Analyzing Interpretable Machine Learning Systems”. Workshop on Human Interpretability in Machine Learning at Int. Conference on Machine Learning, pp. 8 14. 2018.
[Vid14] V. Vidulin, M. Bohanex, M. Gams. “Combining human analysis and machine data mining to obtain credible data relations”. Inf. Sciences, vol. 288, pp. 254 278. 2014.
[Voj20] S. Vojir. T. Kliegr. “Editable machine learning models? A rule-based framework for user studies of explainability”. Advances in Data Analysis and Classification. 2020.