IJCAI 2023 论文列表

Demand-Responsive Transport (DRT) has grown over the last decade as an ecological solution to both metropolitan and suburban areas. It provides a more efficient public transport service in metropolitan areas and satisfies the mobility needs in sparse and heterogeneous suburban areas. Traditionally, DRT operators build the plannings of their drivers by relying on myopic insertion heuristics that do not take into account the dynamic nature of such a service. We thus investigate in this work the potential of a Demand Prediction Framework used specifically to build more flexible routes within a Dynamic Dial-a-Ride Problem (DaRP) solver. We show how to obtain a Machine Learning forecasting model that is explicitly designed for optimization purposes. The prediction task is further complicated by the fact that the historical dataset is significantly sparse. We finally show how the predicted travel requests can be integrated within an optimization scheme in order to compute better plannings at the start of the day. Numerical results support the fact that, despite the data sparsity challenge as well as the optimization-driven constraints that result from the DaRP model, such a look-ahead approach can improve up to 3.5% the average insertion rate of an actual DRT service.

We study four NP-hard optimal seat arrangement problems which each have as input a set of n agents, where each agent has cardinal preferences over other agents, and an n-vertex undirected graph (called the seat graph). The task is to assign each agent to a distinct vertex in the seat graph such that either the sum of utilities or the minimum utility is maximized, or it is envy-free or exchange-stable. Aiming at identifying hard and easy cases, we extensively study the algorithmic complexity of the four problems by looking into natural graph classes for the seat graph (e.g., paths, cycles, stars, or matchings), problem-specific parameters (e.g., the number of non-isolated vertices in the seat graph or the maximum number of agents towards whom an agent has non-zero preferences), and preference structures (e.g., non-negative or symmetric preferences). For strict preferences and seat graphs with disjoint edges and isolated vertices, we correct an error in the literature and show that finding an envy-free arrangement remains NP-hard in this case.

Constrained clustering is a semi-supervised task that employs a limited amount of labelled data, formulated as constraints, to incorporate domain-specific knowledge and to significantly improve clustering accuracy. Previous work has considered exact optimization formulations that can guarantee optimal clustering while satisfying all constraints, however these approaches lack interpretability. Recently, decision trees have been used to produce inherently interpretable clustering solutions, however existing approaches do not support clustering constraints and do not provide strong theoretical guarantees on solution quality. In this work, we present a novel SAT-based framework for interpretable clustering that supports clustering constraints and that also provides strong theoretical guarantees on solution quality. We also present new insight into the trade-off between interpretability and satisfaction of such user-provided constraints. Our framework is the first approach for interpretable and constrained clustering. Experiments with a range of real-world and synthetic datasets demonstrate that our approach can produce high-quality and interpretable constrained clustering solutions.

Interpretability of reinforcement learning policies is essential for many real-world tasks but learning such interpretable policies is a hard problem. Particularly, rule-based policies such as decision trees and rules lists are difficult to optimize due to their non-differentiability. While existing techniques can learn verifiable decision tree policies, there is no guarantee that the learners generate a policy that performs optimally. In this work, we study the optimization of size-limited decision trees for Markov Decision Processes (MPDs) and propose OMDTs: Optimal MDP Decision Trees. Given a user-defined size limit and MDP formulation, OMDT directly maximizes the expected discounted return for the decision tree using Mixed-Integer Linear Programming. By training optimal tree policies for different MDPs we empirically study the optimality gap for existing imitation learning techniques and find that they perform sub-optimally. We show that this is due to an inherent shortcoming of imitation learning, namely that complex policies cannot be represented using size-limited trees. In such cases, it is better to directly optimize the tree for expected return. While there is generally a trade-off between the performance and interpretability of machine learning models, we find that on small MDPs, depth 3 OMDTs often perform close to optimally.

Coalition formation is a central approach for multiagent coordination. A crucial part of coalition formation that is extensively studied in AI is coalition structure generation: partitioning agents into coalitions to maximize overall value. In this paper, we propose a novel method for coalition structure generation by introducing a compact and efficient representation of coalition structures. Our representation partitions the solution space into smaller, more manageable subspaces that gather structures containing coalitions of specific sizes. Our proposed method combines two new algorithms, one which leverages our compact representation and a branch-and-bound technique to generate optimal coalition structures, and another that utilizes a preprocessing phase to identify the most promising sets of coalitions to evaluate. Additionally, we show how parts of the solution space can be gathered into groups to avoid their redundant evaluation and we investigate the computational gain that is achieved by avoiding that redundant processing. Through this approach, our algorithm is able to prune the solution space more efficiently. Our results show that the proposed algorithm is superior to prior state-of-the-art methods in generating optimal coalition structures under several value distributions.

Anomaly detection plays an increasingly important role in various fields for critical tasks such as intrusion detection in cybersecurity, financial risk detection, and human health monitoring. A variety of anomaly detection methods have been proposed, and a category based on the isolation forest mechanism stands out due to its simplicity, effectiveness, and efficiency, e.g., iForest is often employed as a state-of-the-art detector for real deployment. While the majority of isolation forests use the binary structure, a framework LSHiForest has demonstrated that the multi-fork isolation tree structure can lead to better detection performance. However, there is no theoretical work answering the fundamentally and practically important question on the optimal tree structure for an isolation forest with respect to the branching factor. In this paper, we establish a theory on isolation efficiency to answer the question and determine the optimal branching factor for an isolation tree. Based on the theoretical underpinning, we design a practical optimal isolation forest OptIForest incorporating clustering based learning to hash which enables more information to be learned from data for better isolation quality. The rationale of our approach relies on a better bias-variance trade-off achieved by bias reduction in OptIForest. Extensive experiments on a series of benchmarking datasets for comparative and ablation studies demonstrate that our approach can efficiently and robustly achieve better detection performance in general than the state-of-the-arts including the deep learning based methods.

Traditional semi-supervised learning tasks assume that both labeled and unlabeled data follow the same class distribution, but the realistic open-world scenarios are of more complexity with unknown novel classes mixed in the unlabeled set. Therefore, it is of great challenge to not only recognize samples from known classes but also discover the unknown number of novel classes within the unlabeled data. In this paper, we introduce a new open-world semi-supervised novel class discovery approach named OpenNCD, a progressive bi-level contrastive learning method over multiple prototypes. The proposed method is composed of two reciprocally enhanced parts. First, a bi-level contrastive learning method is introduced, which maintains the pair-wise similarity of the prototypes and the prototype group levels for better representation learning. Then, a reliable prototype similarity metric is proposed based on the common representing instances. Prototypes with high similarities will be grouped progressively for known class recognition and novel class discovery. Extensive experiments on three image datasets are conducted and the results show the effectiveness of the proposed method in open-world scenarios, especially with scarce known classes and labels.

Typically, trajectories considered anomalous are the ones deviating from usual (e.g., traffic-dictated) driving patterns. However, this closed-set context fails to recognize the unknown anomalous trajectories, resulting in an insufficient self-motivated learning paradigm. In this study, we investigate the novel Anomalous Trajectory Recognition problem in an Open-world scenario (ATRO) and introduce a novel probabilistic Metric learning model, namely ATROM, to address it. Specifically, ATROM can detect the presence of unknown anomalous behavior in addition to identifying known behavior. It has a Mutual Interaction Distillation that uses contrastive metric learning to explore the interactive semantics regarding the diverse behavioral intents and a Probabilistic Trajectory Embedding that forces the trajectories with distinct behaviors to follow different Gaussian priors. More importantly, ATROM offers a probabilistic metric rule to discriminate between known and unknown behavioral patterns by taking advantage of the approximation of multiple priors. Experimental results on two large-scale trajectory datasets demonstrate the superiority of ATROM in addressing both known and unknown anomalous patterns.

We study a new online assignment problem, called the Online Task Assignment with Controllable Processing Time. In a bipartite graph, a set of online vertices (tasks) should be assigned to a set of offline vertices (machines) under the known adversarial distribution (KAD) assumption. We are the first to study controllable processing time in this scenario: There are multiple processing levels for each task and higher level brings larger utility but also larger processing delay. A machine can reject an assignment at the cost of a rejection penalty, taken from a pre-determined rejection budget. Different processing levels cause different penalties. We propose the Online Machine and Level Assignment (OMLA) Algorithm to simultaneously assign an offline machine and a processing level to each online task. We prove that OMLA achieves 1/2-competitive ratio if each machine has unlimited rejection budget and Δ/(3Δ-1)- competitive ratio if each machine has an initial rejection budget up to Δ. Interestingly, the competitive ratios do not change under different settings on the controllable processing time and we can conclude that OMLA is "insensitive" to the controllable processing time.

Many real-world streaming data are sequentially collected over time and with skew-distributed classes. In this situation, online learning models may tend to favor samples from majority classes, making the wrong decisions for those from minority classes. Previous methods try to balance the instance number of different classes or assign asymmetric cost values. They usually require data-buffers to store streaming data or pre-defined cost parameters. This study alternatively shows that the imbalance of instances can be implied by the imbalance of gradients. Then, we propose the Online Harmonizing Gradient Descent (OHGD) for one-pass online classification. By harmonizing the gradient magnitude occurred by different classes, the method avoids the bias of the proposed method in favor of the majority class. Specifically, OHGD requires no data-buffer, extra parameters, or prior knowledge. It also handles imbalanced data streams the same way that it would handle balanced data streams, which facilitates its easy implementation. On top of a few common and mild assumptions, the theoretical analysis proves that OHGD enjoys a satisfying sub-linear regret bound. Extensive experimental results demonstrate the high efficiency and effectiveness in handling imbalanced data streams.

Recommender systems are facing scrutiny because of their growing impact on the opportunities we have access to. Current audits for fairness are limited to coarse-grained parity assessments at the level of sensitive groups. We propose to audit for envy-freeness, a more granular criterion aligned with individual preferences: every user should prefer their recommendations to those of other users. Since auditing for envy requires to estimate the preferences of users beyond their existing recommendations, we cast the audit as a new pure exploration problem in multi-armed bandits. We propose a sample-efficient algorithm with theoretical guarantees that it does not deteriorate user experience. We also study the trade-offs achieved on real-world recommendation datasets.

We propose a universal Graph Neural Network architecture which can be trained as an end-2-end search heuristic for any Constraint Satisfaction Problem (CSP). Our architecture can be trained unsupervised with policy gradient descent to generate problem specific heuristics for any CSP in a purely data driven manner. The approach is based on a novel graph representation for CSPs that is both generic and compact and enables us to process every possible CSP instance with one GNN, regardless of constraint arity, relations or domain size. Unlike previous RL-based methods, we operate on a global search action space and allow our GNN to modify any number of variables in every step of the stochastic search. This enables our method to properly leverage the inherent parallelism of GNNs. We perform a thorough empirical evaluation where we learn heuristics for well known and important CSPs, both decision and optimisation problems, from random data, including graph coloring, MAXCUT, and MAX-k-SAT, and the general RB model. Our approach significantly outperforms prior end-2-end approaches for neural combinatorial optimization. It can compete with conventional heuristics and solvers on test instances that are several orders of magnitude larger and structurally more complex than those seen during training.

Cross-domain NER is a challenging task to address the low-resource problem in practical scenarios. Previous typical solutions mainly obtain a NER model by pre-trained language models (PLMs) with data from a rich-resource domain and adapt it to the target domain. Owing to the mismatch issue among entity types in different domains, previous approaches normally tune all parameters of PLMs, ending up with an entirely new NER model for each domain. Moreover, current models only focus on leveraging knowledge in one general source domain while failing to successfully transfer knowledge from multiple sources to the target. To address these issues, we introduce Collaborative Domain-Prefix Tuning for cross-domain NER (CP-NER) based on text-to-text generative PLMs. Specifically, we present text-to-text generation grounding domain-related instructors to transfer knowledge to new domain NER tasks without structural modifications. We utilize frozen PLMs and conduct collaborative domain-prefix tuning to stimulate the potential of PLMs to handle NER tasks across various domains. Experimental results on the Cross-NER benchmark show that the proposed approach has flexible transfer ability and performs better on both one-source and multiple-source cross-domain NER tasks.

This paper studies the use of Curriculum Learning on Reinforcement Learning (RL) to improve the performance of the dispatching policies learned on the Job-shop Scheduling Problem (JSP). Current works in the literature present a large optimality gap when learning end-to-end solutions on this problem. In this regard, we identify the difficulty for RL to learn directly on large instances as part of the issue and use Curriculum Learning (CL) to mitigate this effect. Particularly, CL sequences the learning process in a curriculum of increasing complexity tasks, which allows learning on large instances that otherwise would be impossible to learn from scratch. In this paper, we present a size-agnostic model that enables us to demonstrate that current curriculum strategies have a major impact on the quality of the solution inferred. In addition, we introduce a novel Reinforced Adaptive Staircase Curriculum Learning (RASCL) strategy, which adjusts the difficulty level during the learning process by revisiting the worst-performing instances. Conducted experiments on Taillard’s and Demirkol’s datasets show that the presented approach significantly improves the current stateof-the-art models on the JSP. It reduces the average optimality gap from 19.35% to 10.46% on Taillard’s instances and from 38.43% to 18.85% on Demirkol’s instances.

We investigate opinion dynamics in a fully-connected system, consisting of n agents, where one of the opinions, called correct, represents a piece of information to disseminate. One source agent initially holds the correct opinion and remains with this opinion throughout the execution. The goal of the remaining agents is to quickly agree on this correct opinion. At each round, one agent chosen uniformly at random is activated: unless it is the source, the agent pulls the opinions of l random agents and then updates its opinion according to some rule. We consider a restricted setting, in which agents have no memory and they only revise their opinions on the basis of those of the agents they currently sample. This setting encompasses very popular opinion dynamics, such as the voter model and best-of-k majority rules. Qualitatively speaking, we show that lack of memory prevents efficient convergence. Specifically, we prove that any dynamics requires Omega(n^2) expected time, even under a strong version of the model in which activated agents have complete access to the current configuration of the entire system, i.e., the case l=n. Conversely, we prove that the simple voter model (in which l=1) correctly solves the problem, while almost matching the aforementioned lower bound. These results suggest that, in contrast to symmetric consensus problems (that do not involve a notion of correct opinion), fast convergence on the correct opinion using stochastic opinion dynamics may require the use of memory.

Importance sampling (IS) is a popular technique in off-policy evaluation, which re-weights the return of trajectories in the replay buffer to boost sample efficiency. However, training with IS can be unstable and previous attempts to address this issue mainly focus on analyzing the variance of IS. In this paper, we reveal that the instability is also related to a new notion of Reuse Bias of IS --- the bias in off-policy evaluation caused by the reuse of the replay buffer for evaluation and optimization. We theoretically show that the off-policy evaluation and optimization of the current policy with the data from the replay buffer result in an overestimation of the objective, which may cause an erroneous gradient update and degenerate the performance. We further provide a high-probability upper bound of the Reuse Bias and show that controlling one term of the upper bound can control the Reuse Bias by introducing the concept of stability for off-policy algorithms. Based on these analyses, we present a novel yet simple Bias-Regularized Importance Sampling (BIRIS) framework along with practical algorithms, which can alleviate the negative impact of the Reuse Bias, and show that our BIRIS can significantly reduce the Reuse Bias empirically. Moreover, extensive experimental results show that our BIRIS-based methods can significantly improve the sample efficiency on a series of continuous control tasks in MuJoCo.

Logical reasoning is needed in a wide range of NLP tasks. Can a BERT model be trained end-to-end to solve logical reasoning problems presented in natural language? We attempt to answer this question in a confined problem space where there exists a set of parameters that perfectly simulates logical reasoning. We make observations that seem to contradict each other: BERT attains near-perfect accuracy on in-distribution test examples while failing to generalize to other data distributions over the exact same problem space. Our study provides an explanation for this paradox: instead of learning to emulate the correct reasoning function, BERT has, in fact, learned statistical features that inherently exist in logical reasoning problems. We also show that it is infeasible to jointly remove statistical features from data, illustrating the difficulty of learning to reason in general. Our result naturally extends to other neural models (e.g. T5) and unveils the fundamental difference between learning to reason and learning to achieve high performance on NLP benchmarks using statistical features.

The Private Aggregation of Teacher Ensembles (PATE) is a machine learning framework that enables the creation of private models through the combination of multiple "teacher" models and a "student" model. The student model learns to predict an output based on the voting of the teachers, and the resulting model satisfies differential privacy. PATE has been shown to be effective in creating private models in semi-supervised settings or when protecting data labels is a priority. This paper explores whether the use of PATE can result in unfairness, and demonstrates that it can lead to accuracy disparities among groups of individuals. The paper also analyzes the algorithmic and data properties that contribute to these disproportionate impacts, why these aspects are affecting different groups disproportionately, and offers recommendations for mitigating these effects.

We study the computational complexity of counterfactual reasoning in relation to the complexity of associational and interventional reasoning on structural causal models (SCMs). We show that counterfactual reasoning is no harder than associational or interventional reasoning on fully specified SCMs in the context of two computational frameworks. The first framework is based on the notion of treewidth and includes the classical variable elimination and jointree algorithms. The second framework is based on the more recent and refined notion of causal treewidth which is directed towards models with functional dependencies such as SCMs. Our results are constructive and based on bounding the (causal) treewidth of twin networks---used in standard counterfactual reasoning that contemplates two worlds, real and imaginary---to the (causal) treewidth of the underlying SCM structure. In particular, we show that the latter (causal) treewidth is no more than twice the former plus one. Hence, if associational or interventional reasoning is tractable on a fully specified SCM then counterfactual reasoning is tractable too. We extend our results to general counterfactual reasoning that requires contemplating more than two worlds and discuss applications of our results to counterfactual reasoning with partially specified SCMs that are coupled with data. We finally present empirical results that measure the gap between the complexities of counterfactual reasoning and associational/interventional reasoning on random SCMs.