NeurIPS 2020

Explaining Naive Bayes and Other Linear Classifiers with Polynomial Time and Delay


Review 1

Summary and Contributions: This paper shows proposes an algorithm that enumerates "PI-explanations" from a naive Bayes classifier, where the 1st best explanation can be found in polynomial time, and where each next-best explanation can be found with some additional polynomial time. Initially, I would have guessed that enumerating such explanations would be NP-hard, but after reading this paper, it is now obviously easy.

Strengths: This result was initially surprising to me, but now in retrospect it is obvious---this will really shift the way I think about this problem in the future, and likely some others as well.

Weaknesses: I expect some more practically-minded people will dismiss this work since it only (yet) deals with linear classifiers.

Correctness: The presentation is very thorough and clear, to the point where I believe the results are "obviously true".

Clarity: Yes

Relation to Prior Work: Yes

Reproducibility: Yes

Additional Feedback: Given that it is NP-hard to compile the decision function of a linear classifier [29], and given that finding the smallest prime implicant is Sigma_2^p hard, that enumerating smallest prime implicants of a linear classifier would still be NP-hard. But, now, it is obviously easy. Another, possibly-related surprise about this problem (explaining linear classifiers) is discussed in this recent paper: On Tractable Representations of Binary Neural Networks Weijia Shi and Andy Shih and Adnan Darwiche and Arthur Choi in KR'20 In particular, they note a folk result from learning theory, that implies that a linear classifier with integer weights can be compiled into an OBDD in pseudo-polynomial time. This was a surprise initially because again, it is NP-hard in general [29], but also now obvious in retrospect. For me, this paper provides an interesting result, but as it is, I admit that it may currently be of somewhat narrow interest. If there were broader implications, say in terms of explaining neural networks, I believe I could more strongly argue for this paper. For example, the above paper by Shi et al. applied this algorithm for compiling linear classifiers towards the compilation of some non-trivial neural networks, with some compelling analyses of some neural networks learned from data, provided through case studies.


Review 2

Summary and Contributions: The paper studies linear classifiers, in particular Naive Bayes. The focus is on explanations that can be computed in poly-time.

Strengths: 1. Bayesian networks, including Naive Bayes, is significant and has obvious relevance to the NeurIPS community. 2. The paper contains interesting theoretical and experimental results.

Weaknesses: 1. While the experimental section in the paper is nice, it could be improved with some more details, please see below. 2. The paper has both a quite broad focus (on explanations in AI, for black boxes, etc.) and narrow focus (on explanations for Naive Bayes and linear classifiers). There is substantial related work in the area of explanations in Bayesian networks that is not considered. Please see below for further information about this.

Correctness: In general, the claims and method appear correct. But I have a few questions about the description of «Reducing linear classifiers to XLCs» in lines 101 to 107: Where is this entering into the overall picture of explanations in this paper? For example, do you use this reduction in the theoretical part of the paper, in the experimental part, or both? Giving some more details on the role of this description would be helpful. The experimental evaluation in section 4 is reasonable, using the 37 datasets and comparing with the STEP compilation-based approach. It would be helpful to see some more details (number of features and instances, type of data,…) for the 37 datasets in the paper, instead of being referred to an online repository for all specifics.

Clarity: In general, the paper is well-written. This is a more detailed comment: the integration of equations and mathematical expressions into the running text is a little strange. For example, with respect to equation (9), you would want to: have a «,» right after the «)» of the equation and before «(9)», get rid of the comma in «Let,», add a word before «\sigma» and say for example «where \sigma». A similar comment applies to many of the other equations, where you want to put a period or a comma after the mathematical expression or equation, not before. And then start the next line after the mathematical expression or equation accordingly, with either a new sentence or a connecting word (like the «where» above) respectively.

Relation to Prior Work: There is some room for improved discussion of prior work, please see below. Different concepts related to explanations and Bayesian networks have been studied for several decades in the literature, see below for some example references: M. P. Wellman and M. Henrion. 1993. Explaining “Explaining Away.” IEEE Trans. Pattern Anal. Mach. Intell. 15, 3 (March 1993), 287–292. C. Yuan, H. Lim, and T. C. Lu. Most relevant explanation in Bayesian networks. Journal of Artificial Intelligence Research 42, 309-352, 2011. O. J. Mengshoel, D. Roth, and D. C. Wilkins. Portfolios in Stochastic Local Search: Efficiently Computing Most Probable Explanations in Bayesian Networks. Journal of Automated Reasoning, volume 46, 103–160, 2011. J. Kwisthout. Most frugal explanations in Bayesian networks. Artificial Intelligence, Volume 218, 56-73, 2015. C. S. Vlek, H. Prakken, S. Renooij, and B. Verheij. A method for explaining Bayesian networks for legal evidence with scenarios. Artificial Intelligence and Law volume 24, 285–324, 2016. With exception of Park's 2002 paper (reference [22]) there is no reference to the "explanations in Bayesian network" literature as sampled above. Clearly, the Naive Bayes setting considered in this paper is a restricted class of Bayesian networks. However, it would strengthen the present paper if there was an improved positioning of the work within the Bayesian network context, including how the paper differs from previous contributions such as the ones above (and related references).

Reproducibility: Yes

Additional Feedback: === Feedback on Aug 17 after opportunity for response from authors === I have read the response from the authors, and I think it is good. I believe the authors will be able to update the paper, considering the comments by myself and other reviewers, within the required timeline. All things considered, I remain positive about the paper, and in fact have bumped up its score one level.


Review 3

Summary and Contributions: This paper proposes a log-linear time method to compute the smallest PI-explanation for linear classifiers including Naive Bayes classifiers, improving the worst-case exponential time complexity of the compilation approach. It also shows that PI-explanations for a given instance can be enumerated with a log-linear delay.

Strengths: The proposed algorithm for PI-explanation significantly improves the time and space complexity of the state-of-the-art compilation based algorithm which has worst-case exponential complexity (although it is not limited to Naive Bayes or linear classifiers). Empirical evaluation also demonstrates this speed-up. The approach is also simple and natural: computing the smallest PI-explanation reduces to sorting the variables by their weights and choosing the first k, whether the classifier uses categorical, real-valued, or mixed features. This makes it easy to adopt and use this work.

Weaknesses: A major weakness of this work is that it is limited to linear classifiers, which are often already regarded as interpretable models in the explainable AI community. Moreover, the paper does not convincingly motivate and argue the benefits of explaining linear classifiers using PI-explanations. I think this is important because PI-explanations are computed directly from the ranking based on feature weights, which is how features of linear classifiers are often interpreted. On a similar note, as the experiments showed, there will be many PI-explanations (thousands to even millions) per instance, which will be hard to interpret.

Correctness: The algorithm for finding PI-explanation appears to be correct; the backtrack search approach for enumeration makes sense although I did not check the proof for log-linear delay in detail. The experimental comparisons are sound and thorough.

Clarity: The paper is fairly readable, but there is some room for improvement. After introducing extended linear classifiers which admit categorical or real-valued features (or both), restricting back to categorical case to describe the algorithm was awkward. The algorithm could be more cleanly described in the general XLC formulation. Also, the backtrack step of the enumeration was not very clear, in particular how it guarantees whether more PI-explanations exist. An example of the search procedure may help.

Relation to Prior Work: The paper describes PI-explanations in contrast with other ways of explaining ML models, and makes its algorithmic contribution clear against prior work on computing PI-explanations.

Reproducibility: Yes

Additional Feedback: In a sense, the feature weights are also a heuristic. How does ranking of features based on weights compared to the heuristics approaches (Anchor and SHAP)? In section 3.2, I would suggest changing the notation for value \Gamma to depend on instance e=a either in the subscript or as an input, to make it clear its dependence on a particular instance unlike the worst case value \Gamma^{\omega}. Also, additional notations like \Phi for -\Gamma feel superfluous and may even hurt readability. --------------------------------------Post-Rebuttal-------------------------------------- I have read the author response, and most of my concerns were addressed; I have increased my score by 1.


Review 4

Summary and Contributions: The authors describe how to compute cardinality-minimal / PI-explanations for the special case of linear classifiers. They show how to compute the shortest explanation, and how to enumerate explanations with polynomial time delay. Prior works rely on an offline compilation phase which may take exponential time in the worst case, or rely on calls to a SAT-solver (although can handle more general classifiers)

Strengths: The paper presents an enumeration algorithm for explanations of naive bayes classifier. The backtracking mechanism is tweaked so that the search will always find the next explanation with polynomial delay. It is interesting to see it laid out in detail.

Weaknesses: I'm not sure if the computation of a single explanation for naive bayes classifier is that entirely novel or interesting. It seems rather straightforward to greedily flip the least impactful variables. The main weakness of the paper is that the scope of the contribution is quite small. Linear classifiers are far from opaque, and I can't see a great need to generate explanations for their decisions. Even if practitioners do have the need to identify the minimal subset of features to guarantee a decision, I imagine most of them can immediately recognize the obvious approach of keeping the most impactful features. As such, I am doubtful that this contribution will be interesting to the NeurIPS community.

Correctness: Yes to the best of my judgment.

Clarity: Yes.

Relation to Prior Work: Yes.

Reproducibility: Yes

Additional Feedback: It would be interesting to see a discussion of how this work lies in comparison to classes of knowledge bases that enable tractable abductive reasoning [1]. For example, is this result a special case of some known class/language? [1] What makes propositional abduction tractable. Gustav Nordh, Bruno Zanuttini. ********************************************* Post-Rebuttal ********************************************* I have read the rebuttal and the other reviews, and may reconsider my score. I just wanted to address the author's request for specific references "that might cast doubt on the novelty of our work". Sorry for not being more concrete, but here are some specific references. [1] Near-Shortest and K-Shortest Simple Paths. W. Matthew Carlyle, R. Kevin Wood [2] Finding the k Shortest Paths. David Eppstein The polynomial time enumeration algorithm proposed for Eq 16 is basically subset sum where we enumerate all subsets that sum less than some threshold. This is a special case of the problem solved in [1] that enumerates all shortest s.t. paths in a DAG less than some length threshold, using polynomial delay and space. Let d_1,...,d_n be the weights from Eq16 and define d_{n+1}=0. The the reduction can be seen by constructing a DAG with nodes [0,1,2,3,...,n-1,n,n+1] and having edges from i to j for all i < j, where the weight of edge (i,j) is d_j. Then all simple paths from 0 to n+1 with length < k are subset sums with sum < k. [2] may also relevant where the same reduction can be used to enumerate subset sums in increasing order of sums, but I think it is not polynomial space. Edit: I've decided to bump by score up, since I realized the shortest paths algorithm does not satisfy the subset minimiality constraint that the author's method guarantees, so the author's enumeration algorithm is more novel than I originally thought.