Abstract
Two types of explanations have been receiving increased attention in the literature when analyzing the decisions made by classifiers. The first type explains why a decision was made and is known as a sufficient reason for the decision, also an abductive explanation or a PI-explanation. The second type explains why some other decision was not made and is known as a necessary reason for the decision, also a contrastive or counterfactual explanation. These explanations were defined for classifiers with binary, discrete and, in some cases, continuous features. We show that these explanations can be significantly improved in the presence of non-binary features, leading to a new class of explanations that relay more information about decisions and the underlying classifiers. Necessary and sufficient reasons were also shown to be the prime implicates and implicants of the complete reason for a decision, which can be obtained using a quantification operator. We show that our improved notions of necessary and sufficient reasons are also prime implicates and implicants but for an improved notion of complete reason obtained by a new quantification operator that we also define and study.
Access this chapter
Tax calculation will be finalised at checkout
Purchases are for personal use only
Notes
- 1.
We will use sufficient reasons and PI/abductive explanations interchangeably.
- 2.
- 3.
We will use necessary reasons and contrastive explanations interchangeably in this paper. Counterfactual explanations are related but have alternate definitions in the literature. For example, as defined in [5], they correspond to length-minimal necessary reasons; see [18]. But according to some other definitions, they include contrastive explanations (necessary reasons) as a special case; see Sect. 5.2 in [34]. See also [1] for counterfactual explanations that are directed towards Bayesian network classifiers and [2] for a relevant recent study and survey.
- 4.
Interestingly, the axiomatic study of explanations in [3] allows non-binary features, yet Axiom 4 (feasibility) implies that explanations must be simple.
- 5.
For example, we can use it to provide forgetting semantics for the dual operator \(\overline{\exists }\,x_i \cdot \varDelta = \overline{\overline{\forall }\,x_i \cdot \overline{\varDelta }}\). Using Definition 2, we get \(\overline{\exists }\,x_i \cdot \varDelta = \varDelta +\varDelta | x_i\). Using Proposition 4, we get \(\overline{\exists }\,x_i \cdot \varDelta = \varDelta | x_i +\sum _{j \not = i} (x_j \cdot \varDelta | x_j)\). We can now easily show that (1) \(\varDelta \models \overline{\exists }\,x_i \cdot \varDelta \) and (2) \(\overline{\exists }\,x_i \cdot \varDelta \) is equivalent to an NNF whose X-literals do not mention state \(x_i\). That is, \(\overline{\exists }\,x_i\) can be understood as forgetting the information about state \(x_i\) from \(\varDelta \). This is similar to the dual operator \(\exists x_i \cdot \varDelta = \overline{\forall x_i \cdot \overline{\varDelta }}\) studied in [19, 32] except that \(\overline{\exists }\,x_i\) erases less information from \(\varDelta \) since one can show that \(\varDelta \models \overline{\exists }\,x_i \cdot \varDelta \models \exists x_i \cdot \varDelta \).
- 6.
Unlike SRs, two GSRs may mention the same set of variables. Consider the class formula \(\varDelta = (x_1 \cdot y_{12}) +(x_{12} \cdot y_1)\) and instance \(\mathcal{I}= x_1 \cdot y_1\). There are two GSRs for the decision on \(\mathcal{I}\), \(x_1 \cdot y_{12}\) and \(x_{12} \cdot y_1\), and both mention the same variables X, Y.
- 7.
A dual notion, contrastive path explanation (CPXp), was also proposed in [27].
- 8.
An NNF circuit is a DAG whose leaves are labeled with \(\bot , \top \), or literals; and whose internal nodes are labelled with \(\cdot \) or \(+\).
- 9.
The condition \(V \cap (vars(\tau ) \setminus vars(\tau ')) \ne \emptyset \) is trivially satisfied when \(\varDelta \) is the root of the NNF circuit since V will include all circuit variables in this case.
- 10.
The number of clauses in this CNF will be no more than the number of NNF nodes if the NNF is the general reason of a decision tree (i.e., the NNF has a tree structure).
References
Albini, E., Rago, A., Baroni, P., Toni, F.: Relation-based counterfactual explanations for Bayesian network classifiers. In: IJCAI, pp. 451–457 (2020). https://www.ijcai.org/
Amgoud, L.: Explaining black-box classifiers: properties and functions. Int. J. Approx. Reason. 155, 40–65 (2023)
Amgoud, L., Ben-Naim, J.: Axiomatic foundations of explainability. In: IJCAI, pp. 636–642 (2022). https://www.ijcai.org/
Audemard, G., Bellart, S., Bounia, L., Koriche, F., Lagniez, J., Marquis, P.: On the explanatory power of Boolean decision trees. Data Knowl. Eng. 142, 102088 (2022)
Audemard, G., Koriche, F., Marquis, P.: On tractable XAI queries based on compiled representations. In: KR, pp. 838–849 (2020)
Audemard, G., Lagniez, J., Marquis, P., Szczepanski, N.: Computing abductive explanations for boosted trees. CoRR abs/2209.07740 (2022)
Belson, W.A.: Matching and prediction on the principle of biological classification. J. R. Stat. Soc. Ser. C (Appl. Stat.) 8(2), 65–75 (1959). https://www.jstor.org/stable/2985543
Boumazouza, R., Alili, F.C., Mazure, B., Tabia, K.: ASTERYX: a model-agnostic sat-based approach for symbolic and score-based explanations. In: CIKM, pp. 120–129. ACM (2021)
Breiman, L., Friedman, J.H., Olshen, R.A., Stone, C.J.: Classification and Regression Trees. Wadsworth (1984)
Bryant, R.E.: Graph-based algorithms for Boolean function manipulation. IEEE Trans. Comput. 35(8), 677–691 (1986)
Chan, H., Darwiche, A.: Reasoning about Bayesian network classifiers. In: UAI, pp. 107–115. Morgan Kaufmann (2003)
Choi, A., Shih, A., Goyanka, A., Darwiche, A.: On symbolically encoding the behavior of random forests. CoRR abs/2007.01493 (2020)
Choi, A., Xue, Y., Darwiche, A.: Same-decision probability: a confidence measure for threshold-based decisions. Int. J. Approx. Reason. 53(9), 1415–1428 (2012)
de Colnet, A., Marquis, P.: On the complexity of enumerating prime implicants from decision-DNNF circuits. In: IJCAI, pp. 2583–2590 (2022). https://www.ijcai.org/
Crama, Y., Hammer, P.L.: Boolean functions - theory, algorithms, and applications. In: Encyclopedia of Mathematics and Its Applications (2011)
Darwiche, A.: Logic for explainable AI. In: 38th Annual ACM/IEEE Symposium on Logic in Computer Science, LICS, pp. 1–11. IEEE (2023). CoRR abs/2305.05172
Darwiche, A., Hirth, A.: On the reasons behind decisions. In: ECAI. Frontiers in Artificial Intelligence and Applications, vol. 325, pp. 712–720. IOS Press (2020)
Darwiche, A., Ji, C.: On the computation of necessary and sufficient explanations. In: AAAI, pp. 5582–5591. AAAI Press (2022)
Darwiche, A., Marquis, P.: On quantifying literals in Boolean logic and its applications to explainable AI. J. Artif. Intell. Res. 72, 285–328 (2021)
Gurvich, V., Khachiyan, L.: On generating the irredundant conjunctive and disjunctive normal forms of monotone Boolean functions. Discrete Appl. Math. 96, 363–373 (1999)
Huang, X., Izza, Y., Ignatiev, A., Cooper, M.C., Asher, N., Marques-Silva, J.: Efficient explanations for knowledge compilation languages. CoRR abs/2107.01654 (2021)
Huang, X., Izza, Y., Ignatiev, A., Marques-Silva, J.: On efficiently explaining graph-based classifiers. In: KR, pp. 356–367 (2021)
Ignatiev, A., Izza, Y., Stuckey, P.J., Marques-Silva, J.: Using MaxSAT for efficient explanations of tree ensembles. In: AAAI, pp. 3776–3785. AAAI Press (2022)
Ignatiev, A., Narodytska, N., Asher, N., Marques-Silva, J.: From contrastive to abductive explanations and back again. In: Baldoni, M., Bandini, S. (eds.) AIxIA 2020. LNCS (LNAI), vol. 12414, pp. 335–355. Springer, Cham (2021). https://doi.org/10.1007/978-3-030-77091-4_21
Ignatiev, A., Narodytska, N., Marques-Silva, J.: Abduction-based explanations for machine learning models. In: Proceedings of the Thirty-Third Conference on Artificial Intelligence (AAAI), pp. 1511–1519 (2019)
Ignatiev, A., Narodytska, N., Marques-Silva, J.: On validating, repairing and refining heuristic ML explanations. CoRR abs/1907.02509 (2019)
Izza, Y., Ignatiev, A., Marques-Silva, J.: On tackling explanation redundancy in decision trees. J. Artif. Intell. Res. 75, 261–321 (2022)
Izza, Y., Marques-Silva, J.: On explaining random forests with SAT. In: IJCAI, pp. 2584–2591 (2021). https://www.ijcai.org/
Jackson, P.: Computing prime implicates. In: Proceedings of the 1992 ACM Annual Conference on Communications, CSC 1992, pp. 65–72. Association for Computing Machinery, New York, NY, USA (1992). https://doi.org/10.1145/131214.131223
Ji, C., Darwiche, A.: A new class of explanations for classifiers with non-binary features. CoRR abs/2304.14760 (2023)
Kean, A., Tsiknis, G.: An incremental method for generating prime implicants/implicates. J. Symbolic Comput. 9(2), 185–206 (1990)
Lang, J., Liberatore, P., Marquis, P.: Propositional independence: formula-variable independence and forgetting. J. Artif. Intell. Res. 18, 391–443 (2003)
Lipton, P.: Contrastive explanation. Roy. Inst. Philos. Suppl. 27, 247–266 (1990). https://doi.org/10.1017/S1358246100005130
Liu, X., Lorini, E.: A unified logical framework for explanations in classifier systems. J. Log. Comput. 33(2), 485–515 (2023)
Luo, W., Want, H., Zhong, H., Wei, O., Fang, B., Song, X.: An efficient two-phase method for prime compilation of non-clausal Boolean formulae. In: 2021 IEEE/ACM International Conference On Computer Aided Design (ICCAD), pp. 1–9 (2021). https://doi.org/10.1109/ICCAD51958.2021.9643520
Marques-Silva, J., Gerspacher, T., Cooper, M.C., Ignatiev, A., Narodytska, N.: Explanations for monotonic classifiers. In: ICML. Proceedings of Machine Learning Research, vol. 139, pp. 7469–7479. PMLR (2021)
Marques-Silva, J., Ignatiev, A.: Delivering trustworthy AI through formal XAI. In: AAAI, pp. 12342–12350. AAAI Press (2022)
Marquis, P.: Consequence finding algorithms. In: Kohlas, J., Moral, S. (eds.) Handbook of defeasible reasoning and uncertainty management systems, pp. 41–145. Springer, Cham (2000). https://doi.org/10.1007/978-94-017-1737-3_3
Miller, T.: Explanation in artificial intelligence: insights from the social sciences. Artif. Intell. 267, 1–38 (2019)
Narodytska, N., Kasiviswanathan, S.P., Ryzhyk, L., Sagiv, M., Walsh, T.: Verifying properties of binarized deep neural networks. In: Proceedings of AAAI 2018, pp. 6615–6624 (2018)
Previti, A., Ignatiev, A., Morgado, A., Marques-Silva, J.: Prime compilation of non-clausal formulae. In: IJCAI, pp. 1980–1988. AAAI Press (2015)
Ramesh, A., Becker, G., Murray, N.V.: CNF and DNF considered harmful for computing prime implicants/implicates. J. Autom. Reason. 18(3), 337–356 (1997)
Ribeiro, M.T., Singh, S., Guestrin, C.: “Why should I trust you?”: explaining the predictions of any classifier. In: KDD, pp. 1135–1144. ACM (2016)
Ribeiro, M.T., Singh, S., Guestrin, C.: Anchors: high-precision model-agnostic explanations. In: AAAI, pp. 1527–1535. AAAI Press (2018)
Shi, W., Shih, A., Darwiche, A., Choi, A.: On tractable representations of binary neural networks. In: KR, pp. 882–892 (2020)
Shih, A., Choi, A., Darwiche, A.: A symbolic approach to explaining Bayesian network classifiers. In: IJCAI, pp. 5103–5111 (2018). https://www.ijcai.org/
Shih, A., Choi, A., Darwiche, A.: Compiling Bayesian network classifiers into decision graphs. In: AAAI, pp. 7966–7974. AAAI Press (2019)
Slagle, J., Chang, C.L., Lee, R.: A new algorithm for generating prime implicants. IEEE Trans. Comput. C- 19(4), 304–310 (1970). https://doi.org/10.1109/T-C.1970.222917
Wang, E., Khosravi, P., den Broeck, G.V.: Probabilistic sufficient explanations. In: IJCAI, pp. 3082–3088 (2021). https://www.ijcai.org/
Acknowledgments
This work has been partially supported by NSF grant ISS-1910317.
Author information
Authors and Affiliations
Corresponding author
Editor information
Editors and Affiliations
Rights and permissions
Copyright information
© 2023 The Author(s), under exclusive license to Springer Nature Switzerland AG
About this paper
Cite this paper
Ji, C., Darwiche, A. (2023). A New Class of Explanations for Classifiers with Non-binary Features. In: Gaggl, S., Martinez, M.V., Ortiz, M. (eds) Logics in Artificial Intelligence. JELIA 2023. Lecture Notes in Computer Science(), vol 14281. Springer, Cham. https://doi.org/10.1007/978-3-031-43619-2_8
Download citation
DOI: https://doi.org/10.1007/978-3-031-43619-2_8
Published:
Publisher Name: Springer, Cham
Print ISBN: 978-3-031-43618-5
Online ISBN: 978-3-031-43619-2
eBook Packages: Computer ScienceComputer Science (R0)