ACML 2023 Papers

Multi-behavior session-based recommendation (MBSBR) is a critical task in e-commerce and online advertising. By modeling these multiple behaviors, models can better capture the user intent and make more effective recommendations. However, existing models face the challenge of incompletely differentiating between different behavior types, which hinders their ability to fully capture the different tendencies exhibited by each behavior. In addition, most existing multi-behavior methods focus only on predicting a single target behavior and fail to achieve a unified model for predicting the next user-item interaction across multiple behavior types. To address these limitations, we introduce reinforcement learning to the multi-behavior session-based recommendation task and propose a novel approach called the multi-behavior graph reinforcement learning network (MB-GRL). Specifically, we use a graph neural network to encode item transition information from the session graph. Then, we use an attention network to obtain a session representation and generate recommendations based on it. At the same time, we also apply Deep Q-Network (DQN) as a regularizer to improve the recommendation performance for certain behavior types. Experiments on various public benchmark datasets show that MB-GRL outperforms other models for multi-behavior session-based recommendation.

Deep learning models have demonstrated remarkable capabilities in learning complex patterns and concepts from training data. However, recent findings indicate that these models tend to rely heavily on simple and easily discernible features present in the background of images rather than the main concepts or objects they are intended to classify. This phenomenon poses a challenge to image classifiers as the crucial elements of interest in images may be overshadowed. In this paper, we propose a novel approach to address this issue and improve the learning of main concepts by image classifiers. Our central idea revolves around concurrently guiding the model’s attention toward the foreground during the classification task. By emphasizing the foreground, which encapsulates the primary objects of interest, we aim to shift the focus of the model away from the dominant influence of the background. To accomplish this, we introduce a mechanism that encourages the model to allocate sufficient attention to the foreground. We investigate various strategies, including modifying the loss function or incorporating additional architectural components, to enable the classifier to effectively capture the primary concept within an image. Additionally, we explore the impact of different foreground attention mechanisms on model performance and provide insights into their effectiveness. Through extensive experimentation on benchmark datasets, we demonstrate the efficacy of our proposed approach in improving the classification accuracy of image classifiers. Our findings highlight the importance of foreground attention in enhancing model understanding and representation of the main concepts within images. The results of this study contribute to advancing the field of image classification and provide valuable insights for developing more robust and accurate deep-learning models.

Domain generalization is a popular machine learning technique that enables models to perform well on the unseen target domain, by learning from multiple source domains. Domain generalization is useful in cases where data is limited, difficult, or expensive to collect, such as in object recognition and biomedicine. In this paper, we propose a novel domain generalization algorithm called "meta-forests", which builds upon the basic random forests model by incorporating the meta-learning strategy and maximum mean discrepancy measure. The aim of meta-forests is to enhance the generalization ability of classifiers by reducing the correlation among trees and increasing their strength. More specifically, meta-forests conducts meta-learning optimization during each meta-task, while also utilizing the maximum mean discrepancy as a regularization term to penalize poor generalization performance in the meta-test process. To evaluate the effectiveness of our algorithm, we test it on two publicly object recognition datasets and a glucose monitoring dataset that we have used in a previous study. Our results show that meta-forests outperforms state-of-the-art approaches in terms of generalization performance on both object recognition and glucose monitoring datasets.

Deep learning-based recommendation systems (e.g., DLRMs) are widely used AI models to provide high-quality personalized recommendations. Training data used for modern recommendation systems commonly includes categorical features taking on tens-of-millions of possible distinct values. These categorical tokens are typically assigned learned vector representations, that are stored in large embedding tables, on the order of 100s of GB. Storing and accessing these tables represent a substantial performance burden. Our work proposes MEM-REC, a novel alternative representation approach for embedding tables. MEM-REC leverages Bloom filters and hashing methods to encode categorical features using two cache-friendly embedding tables. The first table (token embedding) contains raw embeddings (i.e. learned vector representation), and the second table (weight embedding), which is much smaller, contains weights to scale these raw embeddings to provide better discriminative capability to each data point. We provide a detailed architecture, design and analysis of MEM-REC addressing trade-offs in accuracy and computation requirements. In comparison with state-of-the-art techniques MEM-REC can not only maintain the recommendation quality and significantly reduce the memory footprint for commercial scale recommendation models but can also improve the embedding latency. In particular, based on our results, MEM-REC compresses the MLPerf CriteoTB benchmark DLRM model size by $2900\times$ and performs up to $3.4\times$ faster embeddings while achieving the same AUC as that of the full uncompressed model.

\emph{Finding diverse solutions} to combinatorial optimization problems is beneficial for a deeper understanding of complicated real-world problems and for simpler and more practical mathematical modeling. For this purpose, it is desirable that every solution is far away from one another and that solutions can be found in time not depending polynomially on the size of the family of feasible solutions. In this paper, we investigate the problem of finding diverse sets in the sense of maximizing the \emph{minimum} of \emph{weighted Hamming distance} between \emph{set pairs}. Under a particular assumption, we provide an algorithm that gives diverse sets of almost $\mu /2$-approximation in expectation in the sense of maximization of the minimum of the expected value, where $\mu \in [0, 1]$ is a parameter on a subroutine. We further give a hardness result that any approximation ratio better than $2/3$ is impossible in polynomial time under the assumption of $\mathrm{P} \neq \mathrm{NP}$.

GPS coordinates are a fundamental aspect of location-based applications, yet prior methods for representing them do not fully capture the intricate relationships between different locations. In this paper, we propose a novel map-based approach to embedding GPS coordinates using self-supervised learning. Unlike most prior studies that directly embed GPS coordinates to a latent space, we leverage a map-based approach, allowing embeddings to capture geographical and economic features. Namely, we use a student-teacher architecture, where a student network is trained to mimic the outputs of the teacher, using two different augmented versions of the same input. To capture the rich underlying semantics of GPS coordinates, we further leverage auxiliary tasks including \textit{geo} prediction, high-level reconstruction, and intermediate clustering. The intermediate clustering loss facilitates learning features at different levels of granularity, while the high-level reconstruction loss encourages “local-to-global” correspondences. We evaluate our method on a large-scale dataset of GPS coordinates and demonstrate that it outperforms several baseline methods in terms of the quality of the learned embeddings. Moreover, we show the usefulness of our embeddings in various downstream tasks, such as predicting land price, land cover type, or water quality indice.

Stroke classification is a crucial step for applications with online handwritten input. It is a challenging task due to the variations in writing style, complex structure, long contextual semantic dependence of written content and etc. In this work, we propose a method called Long-Range Graph U-Nets, which involves using a novel node and edge clustering graph pooling layer in the encoder block and a multi-level feature fusion strategy. Such operations guide the model to leverage both temporal and spatial contextual information, establish long-range semantic dependencies, and effectively reduce redundant information caused by local instances of the same category. Extensive experiments conducted on publicly available online handwritten document datasets, demonstrate that our proposed method outperforms previous methods by a significant margin, particularly in the List category, and achieves state-of-the-art performance.

We study regret minimization in an average-reward and communicating Markov Decision Process (MDP) with known dynamics, but unknown reward function. Although learning in such MDPs is a priori easier than in fully unknown ones, they are still largely challenging as they include as special cases large classes of problems such as combinatorial semi-bandits. Leveraging the knowledge on transition function in regret minimization, in a statistically efficient way, appears largely unexplored. As it is conjectured that achieving exact optimality in generic MDPs is NP-hard, even with known transitions, we focus on a computationally efficient relaxation, at the cost of achieving order-optimal logarithmic regret instead of exact optimality. We contribute to filling this gap by introducing a novel algorithm based on the popular Indexed Minimum Empirical Divergence strategy for bandits. A key component of the proposed algorithm is a carefully designed stopping criterion leveraging the recurrent classes induced by stationary policies. We derive a non-asymptotic, problem-dependent, and logarithmic regret bound for this algorithm, which relies on a novel regret decomposition leveraging the structure. We further provide an efficient implementation and experiments illustrating its promising empirical performance.

This paper tackles the critical challenge of object navigation in autonomous navigation systems, particularly focusing on the problem of target approach and episode termination in environments with long optimal episode length in Deep Reinforcement Learning (DRL) based methods. While effective in environment exploration and object localization, conventional DRL methods often struggle with optimal path planning and termination recognition due to a lack of depth information. To overcome these limitations, we propose a novel approach, namely the Depth-Inference Termination Agent (DITA), which incorporates a supervised model called the Judge Model to implicitly infer object-wise depth and decide termination jointly with reinforcement learning. We train our judge model along with reinforcement learning in parallel and supervise the former efficiently by reward signal. Our evaluation shows the method is demonstrating superior performance, we achieve a 9.3% gain on success rate than our baseline method across all room types and gain 51.2% improvements on long episodes environment while maintaining slightly better Success Weighted by Path Length (SPL). Code and resources, visualization are available at: \url{https://github.com/HuskyKingdom/DITA_acml2023}

Recent studies have shown that active object tracking algorithms based on deep reinforcement learning have the difficulty of model training while achieving favorable tracking outcomes. In addition, current active object tracking methods are not suitable for air-to-ground object tracking scenarios in high-altitude environments, such as air search and rescue. Therefore, we proposed a Knowledge-gUided Reinforcement learning (KURL) model for active object tracking, which includes two embedded knowledge-guided models (i.e., the state recognition model and the world model), together with a reinforcement learning module. The state recognition model utilizes the correlation between the observed states and image quality (as measured by object recognition probability) as prior knowledge to guide reinforcement learning algorithm to improve the observed image quality. The reinforcement learning module actively controls the Pan-Tilt-Zoom (PTZ) camera to achieve stable tracking. Additionally, a world model is proposed to replace the traditional Unreal Engine (UE) simulator for model training, which significantly enhancing the training efficiency (about ten times). The results indicate that the KURL model can significantly enhance the image quality, stability and robustness of tracking, compared with other methods in similar tasks.

Learning latent representations of nodes in graphs is important for many real-world applications, such as recommender systems, traffic prediction and fraud detection. Most of the existing research on graph representation learning has focused on static graphs. However, many real-world graphs are dynamic and their structures change over time, which makes learning dynamic node representations challenging. We propose a novel k-truss based temporal graph convolutional network named TTGCN to learn potential node representations on dynamic graphs. Specifically, TTGCN utilizes a novel truss-based graph convolutional layer named TrussGCN to capture the topology and hierarchical structure information of graphs, and combines it with a temporal evolution module to capture complex temporal dependencies. We conduct link prediction experiments on five different dynamic graph datasets. Experimental results demonstrate the superiority of TTGCN for dynamic graph embedding, as it consistently outperforms several state-of-the-art baselines in the link prediction task. In addition, our ablation experiments demonstrate the effectiveness of adopting TrussGCN in a dynamic graph embedding method.

The discrete logarithm problem is a fundamental challenge in number theory with significant implications for cryptographic protocols. In this paper, we investigate the limitations of gradient-based methods for learning the parity bit of the discrete logarithm in finite cyclic groups of prime order. Our main result, supported by theoretical analysis and empirical verification, reveals the concentration of the gradient of the loss function around a fixed point, independent of the logarithm’s base used. This concentration property leads to a restricted ability to learn the parity bit efficiently using gradient-based methods, irrespective of the complexity of the network architecture being trained. Our proof relies on Boas-Bellman inequality in inner product spaces and it involves establishing approximate orthogonality of discrete logarithm’s parity bit functions through the spectral norm of certain matrices. Empirical experiments using a neural network-based approach further verify the limitations of gradient-based learning, demonstrating the decreasing success rate in predicting the parity bit as the group order increases.

This paper introduces two key contributions aimed at improving the speed and quality of images generated through inverse diffusion processes. The first contribution involves reparameterizing the diffusion process in terms of the angle on a quarter-circular arc between the image and noise, specifically setting the conventional $\sqrt{\bar{\alpha}}=\cos(\eta)$. This reparameterization eliminates two singularities and allows for the expression of diffusion evolution as a well-behaved ordinary differential equation (ODE). In turn, this allows higher order ODE solvers such as Runge-Kutta methods to be used effectively. The second contribution is to directly estimate both the image ($\mathbf{x}_0$) and noise ($\mathbf{\epsilon}$) using our network, which enables more stable calculations of the update step in the inverse diffusion steps, as accurate estimation of both the image and noise are crucial at different stages of the process. Together with these changes, our model achieves faster generation, with the ability to converge on high-quality images more quickly, and higher quality of the generated images, as measured by metrics such as Fr{é}chet Inception Distance (FID), spatial Fr{é}chet Inception Distance (sFID), precision, and recall.

In the task of Hierarchical Multi-label Text Classification (HTMC), there exist multiple multivariate relations between labels, particularly the semantic dependencies within label branches of the hierarchy. However, existing methods struggle to fully exploit these potential multivariate dependencies since they can only model binary relationships at best. In this paper, we address this limitation by focusing on leveraging semantic dependencies among labels within branches and propose a Hyper-Label-Graph Model (HLGM). Specifically, we first construct a label hypergraph based on the taxonomy hierarchy and utilize a hypergraph attention mechanism to learn branch-level multivariate dependencies among labels. Furthermore, the model employs a label-text fusion module to generate label-level text representations, facilitating the comprehensive integration of semantic features between text and labels. Additionally, we introduce a hierarchical triplet loss to enhance the ability to distinguish labels within the hyperedge structure. We validate the effectiveness of the proposed model on three benchmark datasets, and the experimental results demonstrate that HLGM outperforms competitive GNN-based baselines.

Graph classification is a crucial task in the field of graph learning with numerous practical applications. Typically, the first step is to construct vertex features by the statistical information of the graph. Existing graph neural networks often adopt the one-hot degree encoding strategy to construct vertex features. Then, these features are fed into a linear layer, which outputs a low-dimensional real vector serving as the initial vertex representation for the graph model. However, the conventional approach of constructing vertex features may not be optimal. Intuitively, the method of constructing vertex features can have significant impact on the effectiveness of model. Hence, the construction of informative vertex features from the graph and the design of an efficient graph model to process these features pose great challenges. In this paper, we propose a novel method for constructing hierarchical topology vertex features and designing a hybrid convolution method to handle these features. Experimental results on public graph datasets of Social Networks, Small Molecules, and Bioinformatics demonstrate the superior performance of our method compared to baselines.

In this work, we propose a fast and accurate method to reconstruct activations of classification and semantic segmentation networks by stitching them with a GAN generator utilizing a 1x1 convolution. We test our approach on images of animals from the AFHQ wild dataset, ImageNet1K, and real-world digital pathology scans of stained tissue samples. Our results show comparable performance to established gradient descent methods but with a processing time that is two orders of magnitude faster, making this approach promising for practical applications.

The recent surge of interest surrounding Multimodal Neural Networks (MM-NN) is attributed to their ability to effectively process and integrate multiscale information from diverse data sources. MM-NNs extract and fuse features from multiple modalities using adequate unimodal backbones and specific fusion networks. Although this helps strengthen the multimodal information representation, designing such networks is labor-intensive. It requires tuning the architectural parameters of the unimodal backbones, choosing the fusing point, and selecting the operations for fusion. Furthermore, multimodality AI is emerging as a cutting-edge option in Internet of Things (IoT) systems where inference latency and energy consumption are critical metrics in addition to accuracy. In this paper, we propose \textit{Harmonic-NAS}, a framework for the joint optimization of unimodal backbones and multimodal fusion networks with hardware awareness on resource-constrained devices. \textit{Harmonic-NAS} involves a two-tier optimization approach for the unimodal backbone architectures and fusion strategy and operators. By incorporating the hardware dimension into the optimization, evaluation results on various devices and multimodal datasets have demonstrated the superiority of \textit{Harmonic-NAS} over state-of-the-art approaches achieving up to $\sim \textbf{10.9%} accuracy improvement, $\sim$\textbf{1.91x} latency reduction, and $\sim$\textbf{2.14x} energy efficiency gain.

Graph Neural Networks (GNNs) are commonly applied to analyze real-world graph-structured data. However, GNNs are sensitive to the given graph structure, which cast importance on graph structure learning to find optimal graph structures and representations. Previous methods have been restricted from large graphs due to high computational complexity. Lottery ticket hypothesis suggests that there exists a subnetwork that has comparable or better performance with proto-networks, which has been transferred to suit for pruning GNNs recently. There are few studies that address lottery ticket hypothesis’s performance on defense in graphs. In this paper, we propose a scalable graph structure learning method leveraging lottery (ticket) hypothesis : GSL-LH. Our experiments show that GSL-LH can outperform its backbone model without attack and show better robustness against attack, achieving state-of-the-art performances in regular-size graphs compared to other graph structure learning methods without feature augmentation. In large graphs, GSL-LH can have comparable results with state-of-the-art defense methods other than graph structure learning, while bringing some insights into explanation of robustness.

Graph neural networks (GNNs) have demonstrated their great power in learning graph-structured data. Due to the limitations of expensive labeled data, contrastive learning has been applied in graph domain. We propose GWGCL, a graph contrastive learning method based on feature group whitening to achieve two key properties of contrastive learning: alignment and uniformity. GWGCL achieves the alignment by ensuring consistency between positive samples. There is no need for negative samples to participate, but rather to achieve the uniformity between samples through whitening. Because whitening has the effect of feature divergence, it avoids the collapse of all sample representations to a single point, which is called dimensional collapse. Moreover, GWGCL can achieve better results and higher efficiency without the need for asymmetric networks, projection layers, stopping gradients and complex loss function. Through extensive experiments, GWGCL performs competitively on node classification and graph classification tasks across ten common graph datasets.The code is in: https://github.com/MR9812/GWGCL.