IJCAI 2023 Papers

This paper presents a job recommendation algorithm designed and validated in the context of the French Public Employment Service. The challenges, owing to the confidential data policy, are related with the extreme sparsity of the interaction matrix and the mandatory scalability of the algorithm, aimed to deliver recommendations to millions of job seekers in quasi real-time, considering hundreds of thousands of job ads. The experimental validation of the approach shows similar or better performances than the state of the art in terms of recall, with a gain in inference time of 2 orders of magnitude. The study includes some fairness analysis of the recommendation algorithm. The gender-related gap is shown to be statistically similar in the true data and in the counter-factual data built from the recommendations.

The deep clustering paradigm has shown great potential for discovering complex patterns that can reveal cell heterogeneity in single-cell RNA sequencing data. This paradigm involves two training phases: pretraining based on a pretext task and fine-tuning using pseudo-labels. Although current models yield promising results, they overlook the geometric distortions that regularly occur during the training process. More precisely, the transition between the two phases results in a coarse flattening of the latent structures, which can deteriorate the clustering performance. In this context, existing methods perform euclidean-based embedding clustering without ensuring the flatness and convexity of the latent manifolds. To address this problem, we incorporate two mechanisms. First, we introduce an overclustering loss to flatten the local curves. Second, we propose an adversarial mechanism to adjust the global geometric configuration. The second mechanism gradually transforms the latent structures into convex ones. Empirical results on a variety of gene expression datasets show that our model outperforms state-of-the-art methods.

Recent dynamic hypergraph neural networks (DHGNNs) are designed to adaptively optimize the hypergraph structure to avoid the dependence on the initial hypergraph structure, thus capturing more hidden information for representation learning. However, most existing DHGNNs cannot adjust the hyperedge number and thus fail to fully explore the underlying hypergraph structure. This paper proposes a new method, namely, totally hypergraph neural network (TDHNN), to adjust the hyperedge number for optimizing the hypergraph structure. Specifically, the proposed method first captures hyperedge feature distribution to obtain dynamical hyperedge features rather than fixed ones, by conducting the sampling from the learned distribution. The hypergraph is then constructed based on the attention coefficients of both sampled hyperedges and nodes. The node features are dynamically updated by designing a simple hypergraph convolution algorithm. Experimental results on real datasets demonstrate the effectiveness of the proposed method, compared to SOTA methods. The source code can be accessed via https://github.com/HHW-zhou/TDHNN.

We are interested in realistic planning problems to model the behavior of Non-Playable Characters (NPCs) in video games. Search-based action planning, introduced by the game F.E.A.R. in 2005, has an exponential time complexity allowing to control only a dozen NPCs between two frames. A close study of the plans generated in first-person shooters shows that: (1) actions are unary, (2) actions are contextually post-unique and (3) there is no two instances of the same action in an NPC’s plan. By considering (1), (2) and (3) as restrictions, we introduce new classes of problems with the Simplified Action Structure formalism which indeed allow to model realistic problems and whose instances are solvable by a linear-time algorithm. We also experimentally show that our algorithm is capable of managing millions of NPCs per frame.

Video action segmentation under timestamp supervision has recently received much attention due to lower annotation costs. Most existing methods generate pseudo-labels for all frames in each video to train the segmentation model. However, these methods suffer from incorrect pseudo-labels, especially for the semantically unclear frames in the transition region between two consecutive actions, which we call ambiguous intervals. To address this issue, we propose a novel framework from the perspective of clustering, which includes the following two parts. First, pseudo-label ensembling generates incomplete but high-quality pseudo-label sequences, where the frames in ambiguous intervals have no pseudo-labels. Second, iterative clustering iteratively propagates the pseudo-labels to the ambiguous intervals by clustering, and thus updates the pseudo-label sequences to train the model. We further introduce a clustering loss, which encourages the features of frames within the same action segment more compact. Extensive experiments show the effectiveness of our method.

To combat poor health and living conditions, policymakers in Africa require temporally and geographically granular data measuring economic well-being. Machine learning (ML) offers a promising alternative to expensive and time-consuming survey measurements by training models to predict economic conditions from freely available satellite imagery. However, previous efforts have failed to utilize the temporal information available in earth observation (EO) data, which may capture developments important to standards of living. In this work, we develop an EO-ML method for inferring neighborhood-level material-asset wealth using multi-temporal imagery and recurrent convolutional neural networks. Our model outperforms state-of-the-art models in several aspects of generalization, explaining 72% of the variance in wealth across held-out countries and 75% held-out time spans. Using our geographically and temporally aware models, we created spatio-temporal material-asset data maps covering the entire continent of Africa from 1990 to 2019, making our data product the largest dataset of its kind. We showcase these results by analyzing which neighborhoods are likely to escape poverty by the year 2030, which is the deadline for when the Sustainable Development Goals (SDG) are evaluated.

We study the complexity of deciding if there is a tie in a given approval-based multiwinner election, as well as the complexity of counting tied winning committees. We consider a family of Thiele rules, their greedy variants, Phragmen's sequential rule, and Method of Equal Shares. For most cases, our problems are computationally hard, but for sequential rules we find an FPT algorithm for discovering ties (parameterized by the committee size). We also show experimentally that in elections of moderate size ties are quite frequent.

We consider the computational problem of finding a smallest local explanation (anchor) for classifying a given feature vector (example) by a black-box model. After showing that the problem is NP-hard in general, we study various natural restrictions of the problem in terms of problem parameters to see whether these restrictions make the problem fixed-parameter tractable or not. We draw a detailed and systematic complexity landscape for combinations of parameters, including the size of the anchor, the size of the anchor's coverage, and parameters that capture structural aspects of the problem instance, including rank-width, twin-width, and maximum difference.

The advent of large language models (LLMs) has brought about a revolution in the development of tailored machine learning models and sparked debates on redefining data requirements. The automation facilitated by the training and implementation of LLMs has led to discussions and aspirations that human-level labeling interventions may no longer hold the same level of importance as in the era of supervised learning. This paper presents compelling arguments supporting the ongoing relevance of human-labeled data in the era of LLMs.

We study formal languages which are capable of fully expressing quantitative probabilistic reasoning and do-calculus reasoning for causal effects, from a computational complexity perspective. We focus on satisfiability problems whose instance formulas allow expressing many tasks in probabilistic and causal inference. The main contribution of this work is establishing the exact computational complexity of these satisfiability problems. We introduce a new natural complexity class, named succ∃R, which can be viewed as a succinct variant of the well-studied class ∃R, and show that these problems are complete for succ∃R. Our results imply even stronger limitations on the use of algorithmic methods for reasoning about probabilities and causality than previous state-of-the-art results that rely only on the NP- or ∃R-completeness of the satisfiability problems for some restricted languages.

The Non-dominated Sorting Genetic Algorithm-II (NSGA-II) is one of the most prominent algorithms to solve multi-objective optimization problems. Recently, the first mathematical runtime guarantees have been obtained for this algorithm, however only for synthetic benchmark problems. In this work, we give the first proven performance guarantees for a classic optimization problem, the NP-complete bi-objective minimum spanning tree problem. More specifically, we show that the NSGA-II with population size N >= 4((n-1) wmax + 1) computes all extremal points of the Pareto front in an expected number of O(m^2 n wmax log(n wmax)) iterations, where n is the number of vertices, m the number of edges, and wmax is the maximum edge weight in the problem instance. This result confirms, via mathematical means, the good performance of the NSGA-II observed empirically. It also shows that mathematical analyses of this algorithm are not only possible for synthetic benchmark problems, but also for more complex combinatorial optimization problems. As a side result, we also obtain a new analysis of the performance of the global SEMO algorithm on the bi-objective minimum spanning tree problem, which improves the previous best result by a factor of |F|, the number of extremal points of the Pareto front, a set that can be as large as n wmax. The main reason for this improvement is our observation that both multi-objective evolutionary algorithms find the different extremal points in parallel rather than sequentially, as assumed in the previous proofs.

The use of AI-based decision aids in diverse domains has inspired many empirical investigations into how AI models’ decision recommendations impact humans’ decision accuracy in AI-assisted decision making, while explorations on the impacts on humans’ decision fairness are largely lacking despite their clear importance. In this paper, using a real-world business decision making scenario—bidding in rental housing markets—as our testbed, we present an experimental study on understanding how the bias level of the AI-based decision aid as well as the provision of AI explanations affect the fairness level of humans’ decisions, both during and after their usage of the decision aid. Our results suggest that when people are assisted by an AI-based decision aid, both the higher level of racial biases the decision aid exhibits and surprisingly, the presence of AI explanations, result in more unfair human decisions across racial groups. Moreover, these impacts are partly made through triggering humans’ “disparate interactions” with AI. However, regardless of the AI bias level and the presence of AI explanations, when people return to make independent decisions after their usage of the AI-based decision aid, their decisions no longer exhibit significant unfairness across racial groups.

This survey overviews a new research agenda on the economics of machine learning, pursued at the Strategic IntelliGence for Machine Agent (SIGMA) Lab at UChicago. This overall research agenda has two themes: machine learning for economics and, conversely, economics for machine learning (ML). The first theme focuses on designing and analyzing ML algorithms for economic problems, ranging from foundational economic models to influential real-world applications such as recommender systems and national security. The second theme employs economic principles to study machine learning itself, such as the valuation and pricing of data, information and ML models, and designing incentive mechanisms to improve large-scale ML research peer reviews. While our research focuses primarily on developing methodologies, in each theme we also highlight some real-world impacts of these works, including ongoing large-scale live experiments and potential deployments for various applications.

We study single-player extensive-form games with imperfect recall, such as the Sleeping Beauty problem or the Absentminded Driver game. For such games, two natural equilibrium concepts have been proposed as alternative solution concepts to ex-ante optimality. One equilibrium concept uses generalized double halving (GDH) as a belief system and evidential decision theory (EDT), and another one uses generalized thirding (GT) as a belief system and causal decision theory (CDT). Our findings relate those three solution concepts of a game to solution concepts of a polynomial maximization problem: global optima, optimal points with respect to subsets of variables and Karush–Kuhn–Tucker (KKT) points. Based on these correspondences, we are able to settle various complexity-theoretic questions on the computation of such strategies. For ex-ante optimality and (EDT,GDH)-equilibria, we obtain NP-hardness and inapproximability, and for (CDT,GT)-equilibria we obtain CLS-completeness results.

This paper introduces the ACCompanion, an expressive accompaniment system. Similarly to a musician who accompanies a soloist playing a given musical piece, our system can produce a human-like rendition of the accompaniment part that follows the soloist's choices in terms of tempo, dynamics, and articulation. The ACCompanion works in the symbolic domain, i.e., it needs a musical instrument capable of producing and playing MIDI data, with explicitly encoded onset, offset, and pitch for each played note. We describe the components that go into such a system, from real-time score following and prediction to expressive performance generation and online adaptation to the expressive choices of the human player. Based on our experience with repeated live demonstrations in front of various audiences, we offer an analysis of the challenges of combining these components into a system that is highly reactive and precise, while still a reliable musical partner, robust to possible performance errors and responsive to expressive variations.

Deep Neural Networks are increasingly adopted in critical tasks that require a high level of safety, e.g., autonomous driving. While state-of-the-art verifiers can be employed to check whether a DNN is unsafe w.r.t. some given property (i.e., whether there is at least one unsafe input configuration), their yes/no output is not informative enough for other purposes, such as shielding, model selection, or training improvements. In this paper, we introduce the #DNN-Verification problem, which involves counting the number of input configurations of a DNN that result in a violation of a particular safety property. We analyze the complexity of this problem and propose a novel approach that returns the exact count of violations. Due to the #P-completeness of the problem, we also propose a randomized, approximate method that provides a provable probabilistic bound of the correct count while significantly reducing computational requirements. We present experimental results on a set of safety-critical benchmarks that demonstrate the effectiveness of our approximate method and evaluate the tightness of the bound.

Text-video retrieval is a challenging cross-modal task, which aims to align visual entities with natural language descriptions. Current methods either fail to leverage the local details or are computationally expensive. What's worse, they fail to leverage the heterogeneous concepts in data. In this paper, we propose the Disentangled Conceptualization and Set-to-set Alignment (DiCoSA) to simulate the conceptualizing and reasoning process of human beings. For disentangled conceptualization, we divide the coarse feature into multiple latent factors related to semantic concepts. For set-to-set alignment, where a set of visual concepts correspond to a set of textual concepts, we propose an adaptive pooling method to aggregate semantic concepts to address the partial matching. In particular, since we encode concepts independently in only a few dimensions, DiCoSA is superior at efficiency and granularity, ensuring fine-grained interactions using a similar computational complexity as coarse-grained alignment. Extensive experiments on five datasets, including MSR-VTT, LSMDC, MSVD, ActivityNet, and DiDeMo, demonstrate that our method outperforms the existing state-of-the-art methods.

The problem of audio-to-text alignment has seen significant amount of research using complete supervision during training. However, this is typically not in the context of long audio recordings wherein the text being queried does not appear verbatim within the audio file. This work is a collaboration with a non-governmental organization called CARE India that collects long audio health surveys from young mothers residing in rural parts of Bihar, India. Given a question drawn from a questionnaire that is used to guide these surveys, we aim to locate where the question is asked within a long audio recording. This is of great value to African and Asian organizations that would otherwise have to painstakingly go through long and noisy audio recordings to locate questions (and answers) of interest. Our proposed framework, INDENT, uses a cross-attention-based model and prior information on the temporal ordering of sentences to learn speech embeddings that capture the semantics of the underlying spoken text. These learnt embeddings are used to retrieve the corresponding audio segment based on text queries at inference time. We empirically demonstrate the significant effectiveness (improvement in R-avg of about 3%) of our model over those obtained using text-based heuristics. We also show how noisy ASR, generated using state-of-the-art ASR models for Indian languages, yields better results when used in place of speech. INDENT, trained only on Hindi data is able to cater to all languages supported by the (semantically) shared text space. We illustrate this empirically on 11 Indic languages.

Most networks are not static objects, but instead they change over time. This observation has sparked rigorous research on temporal graphs within the last years. In temporal graphs, we have a fixed set of nodes and the connections between them are only available at certain time steps. This gives rise to a plethora of algorithmic problems on such graphs, most prominently the problem of finding temporal spanners, i.e., the computation of subgraphs that guarantee all pairs reachability via temporal paths. To the best of our knowledge, only centralized approaches for the solution of this problem are known. However, many real-world networks are not shaped by a central designer but instead they emerge and evolve by the interaction of many strategic agents. This observation is the driving force of the recent intensive research on game-theoretic network formation models. In this work we bring together these two recent research directions: temporal graphs and game-theoretic network formation. As a first step into this new realm, we focus on a simplified setting where a complete temporal host graph is given and the agents, corresponding to its nodes, selfishly create incident edges to ensure that they can reach all other nodes via temporal paths in the created network. This yields temporal spanners as equilibria of our game. We prove results on the convergence to and the existence of equilibrium networks, on the complexity of finding best agent strategies, and on the quality of the equilibria. By taking these first important steps, we uncover challenging open problems that call for an in-depth exploration of the creation of temporal graphs by strategic agents.