Search

This paper proposes a new algorithm for learning accurate tree-based models while ensuring the existence of recourse actions. Algorithmic Recourse (AR) aims to provide a recourse action for altering the undesired prediction result given by a model. Typical AR methods provide a reasonable action by solving an optimization task of minimizing the required effort among executable actions. In practice, however, such actions do not always exist for models optimized only for predictive performance. To alleviate this issue, we formulate the task of learning an accurate classification tree under the constraint of ensuring the existence of reasonable actions for as many instances as possible. Then, we propose an efficient top-down greedy algorithm by leveraging the adversarial training techniques. We also show that our proposed algorithm can be applied to the random forest, which is known as a popular framework for learning tree ensembles. Experimental results demonstrated that our method successfully provided reasonable actions to more instances than the baselines without significantly degrading accuracy and computational efficiency., Comment: 27 pages, 10 figures, to appear in the 41st International Conference on Machine Learning (ICML 2024)

Machine learning models need to be continually updated or corrected to ensure that the prediction accuracy remains consistently high. In this study, we consider scenarios where developers should be careful to change the prediction results by the model correction, such as when the model is part of a complex system or software. In such scenarios, the developers want to control the specification of the corrections. To achieve this, the developers need to understand which subpopulations of the inputs get inaccurate predictions by the model. Therefore, we propose correction rule mining to acquire a comprehensive list of rules that describe inaccurate subpopulations and how to correct them. We also develop an efficient correction rule mining algorithm that is a combination of frequent itemset mining and a unique pruning technique for correction rules. We observed that the proposed algorithm found various rules which help to collect data insufficiently learned, directly correct model outputs, and analyze concept drift.

This paper proposes a new framework of algorithmic recourse (AR) that works even in the presence of missing values. AR aims to provide a recourse action for altering the undesired prediction result given by a classifier. Existing AR methods assume that we can access complete information on the features of an input instance. However, we often encounter missing values in a given instance (e.g., due to privacy concerns), and previous studies have not discussed such a practical situation. In this paper, we first empirically and theoretically show the risk that a naive approach with a single imputation technique fails to obtain good actions regarding their validity, cost, and features to be changed. To alleviate this risk, we formulate the task of obtaining a valid and low-cost action for a given incomplete instance by incorporating the idea of multiple imputation. Then, we provide some theoretical analyses of our task and propose a practical solution based on mixed-integer linear optimization. Experimental results demonstrated the efficacy of our method in the presence of missing values compared to the baselines., Comment: 30 pages, 15 figures

The suffix trees are fundamental data structures for various kinds of string processing. The suffix tree of a text string $T$ of length $n$ has $O(n)$ nodes and edges, and the string label of each edge is encoded by a pair of positions in $T$. Thus, even after the tree is built, the input string $T$ needs to be kept stored and random access to $T$ is still needed. The \emph{linear-size suffix tries} (\emph{LSTs}), proposed by Crochemore et al. [Linear-size suffix tries, TCS 638:171-178, 2016], are a "stand-alone" alternative to the suffix trees. Namely, the LST of an input text string $T$ of length $n$ occupies $O(n)$ total space, and supports pattern matching and other tasks with the same efficiency as the suffix tree without the need to store the input text string $T$. Crochemore et al. proposed an \emph{offline} algorithm which transforms the suffix tree of $T$ into the LST of $T$ in $O(n \log \sigma)$ time and $O(n)$ space, where $\sigma$ is the alphabet size. In this paper, we present two types of \emph{online} algorithms which "directly" construct the LST, from right to left, and from left to right, without constructing the suffix tree as an intermediate structure. Both algorithms construct the LST incrementally when a new symbol is read, and do not access the previously read symbols. Both of the right-to-left construction algorithm and the left-to-right construction algorithm work in $O(n \log \sigma)$ time and $O(n)$ space. The main feature of our algorithms is that the input text string does not need to be stored., Comment: 27 pages, 14 figures. arXiv admin note: text overlap with arXiv:1901.10045

Given thousands of equally accurate machine learning (ML) models, how can users choose among them? A recent ML technique enables domain experts and data scientists to generate a complete Rashomon set for sparse decision trees--a huge set of almost-optimal interpretable ML models. To help ML practitioners identify models with desirable properties from this Rashomon set, we develop TimberTrek, the first interactive visualization system that summarizes thousands of sparse decision trees at scale. Two usage scenarios highlight how TimberTrek can empower users to easily explore, compare, and curate models that align with their domain knowledge and values. Our open-source tool runs directly in users' computational notebooks and web browsers, lowering the barrier to creating more responsible ML models. TimberTrek is available at the following public demo link: https://poloclub.github.io/timbertrek., Comment: Accepted at IEEE VIS 2022. 5 pages, 6 figures. For a demo video, see https://youtu.be/3eGqTmsStJM. For a live demo, visit https://poloclub.github.io/timbertrek

In any given machine learning problem, there may be many models that could explain the data almost equally well. However, most learning algorithms return only one of these models, leaving practitioners with no practical way to explore alternative models that might have desirable properties beyond what could be expressed within a loss function. The Rashomon set is the set of these all almost-optimal models. Rashomon sets can be extremely complicated, particularly for highly nonlinear function classes that allow complex interaction terms, such as decision trees. We provide the first technique for completely enumerating the Rashomon set for sparse decision trees; in fact, our work provides the first complete enumeration of any Rashomon set for a non-trivial problem with a highly nonlinear discrete function class. This allows the user an unprecedented level of control over model choice among all models that are approximately equally good. We represent the Rashomon set in a specialized data structure that supports efficient querying and sampling. We show three applications of the Rashomon set: 1) it can be used to study variable importance for the set of almost-optimal trees (as opposed to a single tree), 2) the Rashomon set for accuracy enables enumeration of the Rashomon sets for balanced accuracy and F1-score, and 3) the Rashomon set for a full dataset can be used to produce Rashomon sets constructed with only subsets of the data set. Thus, we are able to examine Rashomon sets across problems with a new lens, enabling users to choose models rather than be at the mercy of an algorithm that produces only a single model., Comment: NeurIPS 2022 (Oral)

Post-hoc explanation methods for machine learning models have been widely used to support decision-making. One of the popular methods is Counterfactual Explanation (CE), also known as Actionable Recourse, which provides a user with a perturbation vector of features that alters the prediction result. Given a perturbation vector, a user can interpret it as an "action" for obtaining one's desired decision result. In practice, however, showing only a perturbation vector is often insufficient for users to execute the action. The reason is that if there is an asymmetric interaction among features, such as causality, the total cost of the action is expected to depend on the order of changing features. Therefore, practical CE methods are required to provide an appropriate order of changing features in addition to a perturbation vector. For this purpose, we propose a new framework called Ordered Counterfactual Explanation (OrdCE). We introduce a new objective function that evaluates a pair of an action and an order based on feature interaction. To extract an optimal pair, we propose a mixed-integer linear optimization approach with our objective function. Numerical experiments on real datasets demonstrated the effectiveness of our OrdCE in comparison with unordered CE methods., Comment: 20 pages, 5 figures, to appear in the 35th AAAI Conference on Artificial Intelligence (AAAI 2021)

Learning of interpretable classification models has been attracting much attention for the last few years. Discovery of succinct and contrasting patterns that can highlight the differences between the two classes is very important. Such patterns are useful for human experts, and can be used to construct powerful classifiers. In this paper, we consider mining of minimal emerging patterns from high-dimensional data sets under a variety of constraints in a supervised setting. We focus on an extension in which patterns can contain negative items that designate the absence of an item. In such a case, a database becomes highly dense, and it makes mining more challenging since popular pattern mining techniques such as fp-tree and occurrence deliver do not efficiently work. To cope with this difficulty, we present an efficient algorithm for mining minimal emerging patterns by combining two techniques: dynamic variable-ordering during pattern search for enhancing pruning effect, and the use of a pointer-based dynamic data structure, called dancing links, for efficiently maintaining occurrence lists. Experiments on benchmark data sets showed that our algorithm achieves significant speed-ups over emerging pattern mining approach based on LCM, a very fast depth-first frequent itemset miner using static variable-ordering.

The suffix trees are fundamental data structures for various kinds of string processing. The suffix tree of a string $T$ of length $n$ has $O(n)$ nodes and edges, and the string label of each edge is encoded by a pair of positions in $T$. Thus, even after the tree is built, the input text $T$ needs to be kept stored and random access to $T$ is still needed. The linear-size suffix tries (LSTs), proposed by Crochemore et al. [Linear-size suffix tries, TCS 638:171-178, 2016], are a `stand-alone' alternative to the suffix trees. Namely, the LST of a string $T$ of length $n$ occupies $O(n)$ total space, and supports pattern matching and other tasks in the same efficiency as the suffix tree without the need to store the input text $T$. Crochemore et al. proposed an offline algorithm which transforms the suffix tree of $T$ into the LST of $T$ in $O(n \log \sigma)$ time and $O(n)$ space, where $\sigma$ is the alphabet size. In this paper, we present two types of online algorithms which `directly' construct the LST, from right to left, and from left to right, without constructing the suffix tree as an intermediate structure. Both algorithms construct the LST incrementally when a new symbol is read, and do not access to the previously read symbols. The right-to-left construction algorithm works in $O(n \log \sigma)$ time and $O(n)$ space and the left-to-right construction algorithm works in $O(n (\log \sigma + \log n / \log \log n))$ time and $O(n)$ space. The main feature of our algorithms is that the input text does not need to be stored., Comment: 20 pages, 9 figures

In the context of Multi Instance Learning, we analyze the Single Instance (SI) learning objective. We show that when the data is unbalanced and the family of classifiers is sufficiently rich, the SI method is a useful learning algorithm. In particular, we show that larger data imbalance, a quality that is typically perceived as negative, in fact implies a better resilience of the algorithm to the statistical dependencies of the objects in bags. In addition, our results shed new light on some known issues with the SI method in the setting of linear classifiers, and we show that these issues are significantly less likely to occur in the setting of neural networks. We demonstrate our results on a synthetic dataset, and on the COCO dataset for the problem of patch classification with weak image level labels derived from captions.

We analyze the grammar generation algorithm of the RePair compression algorithm and show the relation between a grammar generated by RePair and maximal repeats. We reveal that RePair replaces step by step the most frequent pairs within the corresponding most frequent maximal repeats. Then, we design a novel variant of RePair, called MR-RePair, which substitutes the most frequent maximal repeats at once instead of substituting the most frequent pairs consecutively. We implemented MR-RePair and compared the size of the grammar generated by MR-RePair to that by RePair on several text corpus. Our experiments show that MR-RePair generates more compact grammars than RePair does, especially for highly repetitive texts.

In this paper, we propose a novel approach to combine \emph{compact directed acyclic word graphs} (CDAWGs) and grammar-based compression. This leads us to an efficient self-index, called Linear-size CDAWGs (L-CDAWGs), which can be represented with $O(\tilde e_T \log n)$ bits of space allowing for $O(\log n)$-time random and $O(1)$-time sequential accesses to edge labels, and $O(m \log \sigma + occ)$-time pattern matching. Here, $\tilde e_T$ is the number of all extensions of maximal repeats in $T$, $n$ and $m$ are respectively the lengths of the text $T$ and a given pattern, $\sigma$ is the alphabet size, and $occ$ is the number of occurrences of the pattern in $T$. The repetitiveness measure $\tilde e_T$ is known to be much smaller than the text length $n$ for highly repetitive text. For constant alphabets, our L-CDAWGs achieve $O(m + occ)$ pattern matching time with $O(e_T^r \log n)$ bits of space, which improves the pattern matching time of Belazzougui et al.'s run-length BWT-CDAWGs by a factor of $\log \log n$, with the same space complexity. Here, $e_T^r$ is the number of right extensions of maximal repeats in $T$. As a byproduct, our result gives a way of constructing an SLP of size $O(\tilde e_T)$ for a given text $T$ in $O(n + \tilde e_T \log \sigma)$ time., Comment: 12 pages, 2 figures

In this paper, we present a new data structure called the packed compact trie (packed c-trie) which stores a set $S$ of $k$ strings of total length $n$ in $n \log\sigma + O(k \log n)$ bits of space and supports fast pattern matching queries and updates, where $\sigma$ is the size of an alphabet. Assume that $\alpha = \log_\sigma n$ letters are packed in a single machine word on the standard word RAM model, and let $f(k,n)$ denote the query and update times of the dynamic predecessor/successor data structure of our choice which stores $k$ integers from universe $[1,n]$ in $O(k \log n)$ bits of space. Then, given a string of length $m$, our packed c-tries support pattern matching queries and insert/delete operations in $O(\frac{m}{\alpha} f(k,n))$ worst-case time and in $O(\frac{m}{\alpha} + f(k,n))$ expected time. Our experiments show that our packed c-tries are faster than the standard compact tries (a.k.a. Patricia trees) on real data sets. As an application of our packed c-trie, we show that the sparse suffix tree for a string of length $n$ over prefix codes with $k$ sampled positions, such as evenly-spaced and word delimited sparse suffix trees, can be constructed online in $O((\frac{n}{\alpha} + k) f(k,n))$ worst-case time and $O(\frac{n}{\alpha} + k f(k,n))$ expected time with $n \log \sigma + O(k \log n)$ bits of space. When $k = O(\frac{n}{\alpha})$, by using the state-of-the-art dynamic predecessor/successor data structures, we obtain sub-linear time construction algorithms using only $O(\frac{n}{\alpha})$ bits of space in both cases. We also discuss an application of our packed c-tries to online LZD factorization., Comment: 10 pages, 2 figures

We consider construction of the suffix tree and the directed acyclic word graph (DAWG) indexing data structures for a collection $\mathcal{T}$ of texts, where a new symbol may be appended to any text in $\mathcal{T} = \{T_1, \ldots, T_K\}$, at any time. This fully-online scenario, which arises in dynamically indexing multi-sensor data, is a natural generalization of the long solved semi-online text indexing problem, where texts $T_1, \ldots, T_{k}$ are permanently fixed before the next text $T_{k+1}$ is processed for each $1 \leq k < K$. We present fully-online algorithms that construct the suffix tree and the DAWG for $\mathcal{T}$ in $O(N \log \sigma)$ time and $O(N)$ space, where $N$ is the total lengths of the strings in $\mathcal{T}$ and $\sigma$ is their alphabet size. The standard explicit representation of the suffix tree leaf edges and some DAWG edges must be relaxed in our fully-online scenario, since too many updates on these edges are required in the worst case. Instead, we provide access to the updated suffix tree leaf edge labels and the DAWG edges to be redirected via auxiliary data structures, in $O(\log \sigma)$ time per added character., Comment: 28 pages, 6 figures, LaTeX