Extended abstract: 𝑓𝐢𝐴𝑆𝑃 - A forgetting technique for XAI based on goal-directed constraint ASP models Luciana Fidilio-Allende, Joaquin Arias CETINIA, Universidad Rey Juan Carlos, MΓ³stoles, Spain Abstract This paper is an extended abstract of: L. Fidilio-Allende, J. Arias, 𝑓𝐢𝐴𝑆𝑃 : A forgetting technique for XAI based on goal-directed constraint ASP models, in: XXIII Jornadas sobre ProgramaciΓ³n y Lenguajes (PROLE), 2024. URL: https://hdl.handle.net/11705/PROLE/2024/13. [1]. Keywords Privacy, Value Awareness, XAI, ASP, s(CASP), Forgetting The automation of all sorts of processes through Artificial Intelligence (AI) systems has made significant progress. More recently, whether through self-regulation and soft law such as guidelines or through legal regulation (e.g., the General Data Protection Regulation (GDPR) or the Regulation on AI, both by the EU), it has become apparent that this development needs to be accompanied by measures that safeguard the fundamental rights and safety of people affected by AI systems. In this sense, Explainable Artificial Intelligence (XAI) [2] is of foremost importance to design trustworthy systems. Proposals such as s(LAW) [3], which are based on Answer Set Programming, have shown their ability to model values and explain the reasons for their decisions, thanks to their rule-based models. But these explanations could lead to the disclosure of sensitive information, such as details about victims of gender-based violence. This could violate the right to privacy and confidentiality, or even cause legal issues, among other concerns. Although explanations can be adjusted to prevent leaks, e.g., using the s(CASP) framework to control which elements are shown and/or hidden [4], adapting the models requires the application of techniques such as forgetting (variable elimination) to avoid revealing sensitive information during an audit. However, current forgetting techniques are mostly only applied in propositional ASP programs, and they have limitations dealing with even loops. In this work, we present 𝑓𝐢𝐴𝑆𝑃 , a new forgetting technique that supports the presence of non- stratified negations in Constraint Answer Set Programs. 𝑓𝐢𝐴𝑆𝑃 is based on the dual rules of s(CASP), a goal-directed CASP reasoner, and therefore, we believe that it can be applied to generic CASP programs without grounding. We have validated our proposal by solving flagship examples from the literature, and we plan to use this technique in the context of school places allocation while preserving the privacy of victims of gender-based violence. Table 1 Comparison of the more relevant forgetting operators vs. 𝑓𝐢𝐴𝑆𝑃 (UP) (SP) Loops Commutative Predicates Constraints π‘“π‘†π‘ˆ [5] Yes No Yes No No No 𝑓𝑆𝑃 [6] Yes Maybe No No No No * 𝑓𝑆𝑃 [7] Yes Maybe Yes Yes No No 𝑓𝐴𝐢 [8] Yes Yes Yes Yes No No 𝑓𝐢𝐴𝑆𝑃 Yes Yes Yes Yes Maybe Maybe 4th Workshop on Goal-directed Execution of Answer Set Programs (GDE’24), October 12, 2024 $ luciana.fidilio@urjc.es (L. Fidilio-Allende); joaquin.arias@urjc.es (J. Arias)  0009-0004-7779-8265 (L. Fidilio-Allende); 0000-0003-4148-311X (J. Arias) Β© 2024 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). CEUR ceur-ws.org Workshop ISSN 1613-0073 Proceedings Figure 1: Sketch of the implementation of the predicate f_casp/4 1 f_casp (Flag, [Pred|Preds], P_0, P_Forgetting) :- 2 add_fact (Pred, P_0, P_11), 3 add_missing (Pred, P_11, P_12), 4 add_even_loop (Pred, P_12, P_13), % Step 1 5 gen_dual (Pred, P_13, Dual_Pred), 6 add_dual (P_13, Dual_Pred, P_2), % Step 2 7 forget_pred (Pred, P_2, P_31), 8 restore_even_loop (P_31, P_32), % Step 3 9 f_casp (Flag, Preds, P_32, P_Forgetting). % Repeat 1,2,3 10 f_casp (0, [], P_Forgetting, P_Forgetting). % Skip Step 4 11 f_casp (1, [], P_Forgetting, P_Scasp) :- 12 scasp_even_loop(P_Forgetting, P_Scasp). % Step 4 In Table 1, we compare fCASP with other relevant operators such as those described in [9] and [10], evaluating its preliminary performance and properties. Through validation using literature examples, we believe that 𝑓𝐢𝐴𝑆𝑃 can generate programs with the same answer sets, even when additional facts (when complying with UP) or propositional rules (when complying with SP) are added to both programs, avoiding the removed or auxiliary predicates. Additionally, we have tested the operator with programs involving even loops, and it can generate equivalent programs regardless of the order in which predicates are removed, that is, it is commutative. In the future, there is potential to extend it to programs with variables and constraints, thanks to the use of s(CASP) dual rules. The preliminary design of the algorithm involves three steps repeated iteratively for each predicate marked to be forgotten, plus a final optional step. Figure 1 shows an implementation sketch of the operator’s steps: β€’ The first step (lines 2 βˆ’ 4) involves adding auxiliary predicates (neg_x) and clauses when the predicate to be forgotten is part of an even loop, is a fact, or is a missing predicate. β€’ The second step (lines 5 and 6) is generating the dual rule of the predicate to forget, being the dual rule the negated version of all the predicate’s clauses. β€’ The third step (lines 7 and 8) is forgetting the predicate, replacing its appearances with the content of its clauses and its negation with the content of its dual rule. β€’ The final step (line 12) is transforming the double negations (not not) into even loops, as s(CASP) does not explicitly support them. This step is optional. We have performed an evaluation of fCASP using examples from the literature. In the first example, we forget predicates in even loops. To conserve the symmetry in answer sets [11], (preserving the predicates not forgotten in the answer sets even when adding additional rules), it is necessary to add additional predicates (neg_x) as strong persistence (SP) cannot be achieved (in some cases) without them [12, 8]. Below we can see the result of forgetting p and q, and the answer sets of both the original program and the generated one. As they have the same answer sets (ignoring the forgotten and auxiliary predicates), the programs are equivalent. 𝑃1 = Example 3 from [12] 𝑓𝐢𝐴𝑆𝑃 (𝑃1 , {𝑝, π‘ž}) {a, p}, {b, q} {a, neg_2}, {b, neg_1} 1 a :- p. 1 a :- not not neg_2. 2 b :- q. 2 b :- not not neg_1. 3 p :- not q. 3 neg_1 :- not not neg_1. 4 q :- not p. 4 neg_2 :- not neg_1. In the second example, we forget predicates present in double negations. This case cannot be resolved without the use of additional predicates. As we can see on the next example, both the original program and the one generated with 𝑓𝐢𝐴𝑆𝑃 forgetting p are equivalent. 𝑃2 = Example 4 from [11] 𝑓𝐢𝐴𝑆𝑃 (𝑃2 , {𝑝}) {p, q}, {r} {q}, {r, neg_1} 1 p :- not not p. 1 q :- not neg_1. 2 q :- p. 2 r :- not not neg_1. 3 r :- not p. 3 neg_1 :- not not neg_1. In the third example we forget multiple predicates regardless of the order. 𝑃3 = Example 1 from [7] 𝑓𝐢𝐴𝑆𝑃 (𝑃3 , {𝑝, π‘ž}) and 𝑓𝐢𝐴𝑆𝑃 (𝑃2 , {π‘ž, 𝑝}) {p}, {q} {}, {neg_1} 1 a :- p, q. 1 a :- not neg_1, not not neg_1. 2 q :- not p. 2 neg_1 :- not not neg_1. 3 p :- not not p. In the final example, we compare the performance of 𝑓𝐢𝐴𝑆𝑃 with 𝑓𝐴𝐢 , the operator that is closer to the desired properties. As we can see, 𝑓𝐴𝐢 may generate a program that outputs redundant answers, while the one generated with 𝑓𝐢𝐴𝑆𝑃 preserves the original count. 𝑃4 = Example 5 from [8] 𝑓𝐴𝐢 (𝑃4 , {π‘ž}). 𝑓𝐢𝐴𝑆𝑃 (𝑃4 , {π‘ž}). {c} {c}, {c, π›Ώπ‘ž } {c, neg_1} 1 q :- not not q, b. 1 a :- b, π›Ώπ‘ž. 1 a :- not neg_1, b. 2 a :- q. 2 c :- not π›Ώπ‘ž. 2 c :- not not neg_1. 3 c :- not q. 3 c :- not b. 3 c :- not b. 4 π›Ώπ‘ž :- not not π›Ώπ‘ž. 4 neg_1 :- not not neg_1. 5 neg_1 :- not b. To evaluate the practicality of 𝑓𝐢𝐴𝑆𝑃 , we have defined two (real) use cases in which we use the operator to remove private and confidential information. In the first use case, we model the Spanish Organic Law 2/2004, May 3, Articles 116 and 117 and the Spanish Constitution Articles 27 and 149.1.30. This legislation establishes the criteria used for assigning public school places in the Comunidad de Madrid, Spain, when the number of applications for a given center is greater than the offer. After applying 𝑓𝐢𝐴𝑆𝑃 , we can successfully remove the students’ medical, socio-economical and gender-based violence related private information while conserving the original model decisions and explainability, complying with the need for transparency required by the applicable regulations such as the recent European AI Act. In the second use case [13], we propose an automated decision-making system for energy assignment in agricultural cooperatives. In this case, the energy is assigned based on how fairly the workers are paid, pondering their salary and productivity, which are considered confidential and, in some cases, private. After applying forgetting, we can preserve the confidentiality of the stakeholders without affecting the decisions and justifications of the model, crucial to make these answers trustworthy. In conclusion, we have presented a new forgetting operator, 𝑓𝐢𝐴𝑆𝑃 , designed to work with goal- directed Answer Set Programs and support dual rules and double negations while being commutative. As potential lines of work for the future, we have identified the extension of the algorithm to support variables and constraints, formally determining and proving 𝑓𝐢𝐴𝑆𝑃 ’s properties, and applying the operator to real use cases. References [1] L. Fidilio-Allende, J. Arias, 𝑓𝐢𝐴𝑆𝑃 : A forgetting technique for XAI based on goal-directed constraint ASP models, in: XXIII Jornadas sobre ProgramaciΓ³n y Lenguajes (PROLE), 2024. URL: https: //hdl.handle.net/11705/PROLE/2024/13. [2] D. Gunning, D. Aha, DARPA’s Explainable Artificial Intelligence (XAI) Program, AI Magazine 40 (2019) 44–58. doi:10.1609/aimag.v40i2.2850. [3] J. Arias, M. Moreno-Rebato, J. A. Rodriguez-GarcΓ­a, S. Ossowski, Automated legal reasoning with discretion to act using s(LAW), 2023. doi:10.1007/s10506-023-09376-5. [4] J. Arias, M. Carro, Z. Chen, G. Gupta, Justifications for goal-directed constraint answer set pro- gramming, 2020. doi:10.4204/EPTCS.325.12. [5] R. GonΓ§alves, T. Janhunen, M. Knorr, J. Leite, On Syntactic Forgetting under Uniform Equivalence, in: European Conference on Logics in Artificial Intelligence, Springer, 2021, pp. 297–312. [6] R. GonΓ§alves, M. Knorr, J. Leite, S. Woltran, When you must forget: Beyond Strong Persistence when Forgetting in Answer Set Programming, 2017. [7] M. Berthold, On Syntactic Forgetting with Strong Persistence, in: Proceedings of the Int. Conf. on Principles of Knowledge Representation and Reasoning, volume 19, 2022, pp. 43–52. [8] M. Berthold, R. GonΓ§alves, M. Knorr, J. Leite, Forgetting in Answer Set Programming with Anonymous Cycles, in: EPIA, Springer, 2019, pp. 552–565. doi:10.1007/978-3-030-30244-3\ _46. [9] T. Eiter, G. Kern-Isberner, A Brief Survey on Forgetting from a Knowledge Representation and Reasoning Perspective, 2019. [10] R. GonΓ§alves, M. Knorr, J. Leite, Forgetting in Answer Set Programming–A Survey, 2023. [11] M. Knorr, J. J. Alferes, Preserving Strong Equivalence while Forgetting, in: Logics in Artificial Intelligence: 14th European Conference, JELIA 2014, Springer, 2014, pp. 412–425. doi:10.1007/ 978-3-319-11558-0\_29. [12] R. GonΓ§alves, M. Knorr, J. Leite, You can’t always forget what you want: on the limits of forgetting in Answer Set Programming, in: Proceedings of the Twenty-second European Conference on Artificial Intelligence, 2016, pp. 957–965. [13] L. Fidilio-Allende, J. Arias, Private-safe (logic-based) decision systems for energy assignment in agricultural cooperatives, in: Highlights in Practical Applications of Agents, Multi-Agent Systems, and Cognitive Mimetics. The PAAMS Collection: International Workshops of PAAMS 2024, Salamanca, Spain, June 26β€”28, 2024, Proceedings, Springer, 2024.