Cold-start item recommendation is a significant challenge in recommendation systems, particularly when new items are introduced without any historical interaction data. While existing methods leverage multi-modal content to alleviate the cold-start issue, they often neglect the inherent multi-view structure of modalities, namely the distinction between shared and modality-specific features. In this paper, we propose Multi-Modal Multi-View Variational AutoEncoder (M²VAE), a generative model that addresses the challenges of modeling common and unique views in attribute and multi-modal features, as well as user preferences over single-typed item features. Specifically, we generate type-specific latent variables for item IDs, categorical attributes, and image features, and use Product-of-Experts (PoE) to derive a common representation. A disentangled contrastive loss decouples the common view from unique views while preserving feature informativeness. To model user inclinations, we employ a user-aware hierarchical Mixture-of-Experts (MoE) to adaptively fuse representations. We further incorporate co-occurrence signals via contrastive learning, eliminating the need for pretraining. Extensive experiments on real-world datasets validate the effectiveness of our approach.

In advertising recommendation systems, click-through rate (CTR) prediction is a critical task. Capturing users' interests from their rich historical behaviors is key to improving prediction results. Although traditional deep learning methods can capture users' interests to some extent, they fail to account for the local and global interests reflected in users' historical behaviors and the dynamic relationships between them. In this paper, we propose a novel architecture - Dual-Interest Adaptive Network (DIAN), which adaptively extracts both local and global interests of users. Specifically, to better explore users' interests in depth, we propose an Adaptive Interest Extraction Block applied to users' historical behavior sequences. By introducing an attention mechanism, this module can flexibly allocate weights to users' local and global interests after decoupling user behaviors. Additionally, to capture complex feature interactions, our model introduces two feature extractors: one combines a Multi-Layer Perceptron (MLP) with a Cross Network for high-order feature extraction, and the other incorporates an Attention Factorization Machine (AFM) for low-order feature extraction. We conducted extensive experiments on the Movielens-1M and Amazon Electronics datasets, validating the effectiveness of DIAN.

Root cause analysis (RCA) is crucial for the stability and reliability of large-scale microservice architectures. Existing multi-source RCA methods primarily rely on logs, traces, and metrics data to detect anomalies and identify abnormal services and root causes. And most multi-source methods focus only on service-level operations and inter-service dependencies, neglecting transactions and their dynamic changes. Furthermore, these approaches typically address only a subset of the RCA tasks, such as anomaly detection, root cause service localization, or root cause type determination. To address these limitations, we propose MicroTR, a transaction reproduction fault diagnosis framework for multi-source RCA in microservice environments. MicroTR deeply analyzes transaction execution logic and service states, utilizing multi-source data to reproduce the dynamic changes of transaction execution states in knowledge graph. This approach enables efficient and synchronized anomaly detection, root cause service localization, and root cause type determination. Experimental evaluations on two widely-adopted open-source microservice platforms demonstrate that MicroTR outperforms state-of-the-art multi-source methods, achieving an average F1 of 97.3% for anomaly detection, an average Hit@l of 90.8% for root cause service localization, and an average Hit@1 of 89.2% for root cause type determination. These results highlight the effectiveness of reproducing transaction execution states for RCA.

Sequential recommendation (SR) aims to predict the next purchasing item according to users' dynamic preference learned from their historical user-item interactions. To improve the performance of recommendation, learning dynamic heterogeneous cross-type behavior dependencies is indispensable for recommender system. However, there still exists some challenges in Multi-Behavior Sequential Recommendation (MBSR). On the one hand, existing methods only model heterogeneous multi-behavior dependencies at behavior-level or item-level, and modeling interaction-level dependencies is still a challenge. On the other hand, the dynamic multi-grained behavior-aware preference is hard to capture in interaction sequences, which reflects interaction-aware sequential pattern. To tackle these challenges, we propose a Multi-Grained Preference enhanced Transformer framework (M-GPT). First, M-GPT constructs an interaction-level graph of historical cross-typed interactions in a sequence. Then graph convolution is performed to derive interaction-level multi-behavior dependency representation repeatedly, in which the complex correlation between historical cross-typed interactions at specific orders can be well learned. Secondly, a novel multifaceted transformer architecture equipped with multi-grained user preference extraction is proposed to encode the interaction-aware sequential pattern enhanced by capturing temporal behavior-aware multi-grained preference . Experiments on the real-world datasets indicate that our method M-GPT consistently outperforms various state-of-the-art recommendation methods. Our code is available at: https://github.com/hchchchchchchc/MGPT.

Lane detection plays a crucial role in autonomous driving. Though modern anchor-based deep lane detection methods have demonstrated remarkable performance on standard benchmarks, they continue to struggle with complex topological variations. In this work, we propose the Dual Weighting Attention Feature Fusion Network(DW_AFFNet), which enhances detection performance through two key innovations: (1) hierarchical feature fusion for improved location and (2) anchor’s IoU classification score consistency optimisation. In computer vision, shallow low-level features provide precise spatial localization while deep high-level features capture essential global information. Therefore, complementary global-local representations for enhanced lane detection accuracy are established via Iterative Coordinate Attention Feature Fusion (ICAFF) module, which systematically combines these hierarchical features through coordinate-sensitive attention mechanisms. Furthermore, we introduce the Dual Weighting Label Assignment Scheme (DW Scheme) to align IoU and classification scores through importance-aware dynamic weighting, significantly improving sample discrimination capability. We evaluate our method on two benchmarks of lane detection and the results demonstrate its effectiveness. Our method surpasses baseline on CULane. On CULane, it obtains 56.45 mF1 with 64.21/55.76/22.16 F1@75/80/90 scores, outperforming CLRNet by 1.52%/2.27%/3.06%/7.4% respectively. Significant improvements are observed across most scenarios in complex road conditions.

In the field of object detection, Feature Pyramid Network have gained widespread adoption in object detection algorithms due to its simplicity, efficiency, and robust feature generation capabilities. Despite its merits, the Feature Pyramid Network exhibits certain limitations in its architectural design. Within the scope of this paper, we aim to dissect the structural limitations inherent in the Feature Pyramid Network and introduce a new network architecture, termed Hybrid-Residual Feature Pyramid Network(HR-FPN), which is designed to effectively mitigate these identified issues. The HR-FPN is primarily composed of two key modules: the Hybrid-Operation Module and the Residual Feature Augmentation Module. The Hybrid-Operation Module effectively integrates semantic information from high-level features into low-level features, while the Residual Feature Augmentation Module mitigates information loss in the highest pyramid layer feature maps by extracting scale-invariant contextual information. Then, We have designed an enhanced version of Sparse R-CNN, termed Specific Proposal Feature R-CNN(SPF R-CNN), which incorporates Learnable Proposal Classification Features and Learnable Proposal Regression Features to effectively mitigate the issue of the discrepancy between classification and localization tasks. Extensive experiments have demonstrated the effectiveness of our approach. Compared to the baseline model, Sparse R-CNN, our method achieved a 1.9 average precision (AP) improvement on the MS-COCO dataset.

Few-shot object detection usually faces the challenge of imbalanced data distribution. The limited training data for novel classes not only leads to insufficient representation of support features but also biases the detector toward base classes. To address these problems, we propose a novel dynamic latent feature guidance method. First, the latent feature reconstruction module utilizes a variational autoencoder to reconstruct query and support features, extracting additional information representations from the latent space, thereby enriching feature representation and compensating for the information deficiencies caused by limited samples. Second, we design the dynamic multiscale similarity guidance module, which highlights information relevant to query images and suppresses background noise and occlusion interference through global, regional, and local similarities. Extensive experimental results demonstrate that our proposed method significantly improves detection accuracy on the PASCAL VOC and MS COCO datasets, outperforming existing state-of-the-art methods.

Mobile Crowd Sensing (MCS) represents an emerging paradigm for collecting sensory data, leveraging the extensive sensing capabilities of widely used mobile devices to execute sensing tasks. Among the array of challenges facing current MCS systems, the incentive mechanism for data requesters and participants consistently stands out as a paramount concern. Existing incentive mechanisms often rely on model-based approaches, assuming a certain degree of prior knowledge about the MCS system, such as expected pricing for data requesters and participants. However, these assumptions are impractical in real-world scenarios. To address this challenge, we endeavor to explore a wholly model-free incentive mechanism. Specifically, we propose a Simulated Annealing Deep Q-learning (SADQ-learning) algorithm to dynamically generate the pricing policy for the sensing platform. Furthermore, to accommodate diverse incentive needs, we devise three distinct incentive modes: one focuses on maximizing the profit of the sensing platform, another dedicates to maximizing the successful matching amount of sensing tasks, and an equilibrium mode seeks a balance between the aforementioned objectives. Finally, numerical results demonstrate the superiority of SADQ-learning through comparisons with baseline algorithms.

The sensing coverage of roadside sensing system usually does not cover the entire road network, resulting in block missing values in traffic data. Traditional methods either adopted simple hints or graph neural networks to capture the speed variation. However, these methods fail to consider that block missing values have less surrounding values and are less susceptible to the influence of adjacent data. The paper proposes Dynamic Style Transfer-based Generative Adversarial Imputation Network (ST-GAIN) for block missing traffic speed imputation. The core idea is to adopt temporal clustering to abstract a large volume of traffic speed into a series of style data. Subsequently, a dynamic encoding network of latent style codes is performed based on the similarity between the missing speed data and the style data. These style codes are then fed into a style transfer network to guide the imputation of the missing values. Additionally, a style discriminator is used to guide style transfer during training. The experimental results demonstrate that the proposed model outperforms state-of-the-art methods by an average of more than 15% in accuracy.

Traffic flow prediction, as one of the problems of spatial correlation analysis of time series, has been extensively studied. The extraction and fusion of effective spatio-temporal features are crucial for achieving high-precision traffic flow prediction. Traditionally, the adjacency graph designed based on the neighboring nodes of real-world road networks has been indispensable for learning spatial features. However, this single connected component graph structure is prone to the phenomenon of over-smoothing, leading to homogenization of the learned spatial feature. Addressing this challenge, this paper proposes a novel Semantic Multi-connected Graph Neural Network (SMGNN) aimed at mitigating the homogeneity of spatial features and effectively modeling spatio-temporal interactions. Firstly, considering the existence of several nodes in large-scale road networks with similar traffic flow variation patterns, we semantically connect these nodes to construct multi-connected semantic spatial graphs (MSSG), replacing the traditionally used neighboring node graph in conventional graph neural networks. Correspondingly, we design a novel graph neural network architecture that cyclically fuses dynamic scale spatio-temporal features from MSSG using an improved Dynamic Spatial Graph Attention (DSGA) module. Secondly, to achieve a more effective representation, we design a Inverted Temporal Attention (ITA) module to supplement static scale temporal features. Furthermore, we introduce a Multi-dimensional and Multi-scale Feature Extraction (MMFE) module to fuse spatio-temporal features at various scales within different receptive fields. Extensive experiments conducted on real-world datasets have verified the effectiveness of our proposed method, significantly outperforming various baseline models.

Mobile Crowd Sensing (MCS) represents a novel paradigm which utilizes intelligent devices carried by mobile users to collect and transmit data. Appropriate incentives are essential to recruit enough participants for sensing tasks. Existing works have designed some incentive mechanisms for MCS, which are not suitable for scenarios when the participants increase significantly as a result of the booming cost. To solve the above cost problem, a Utility-Maximization Incentive Mechanism (UMIM) is proposed in this paper by leveraging the influence propagation on the social network. Participants in the same social network can benefit from the data shared by others, which shows the utility of sensing data and can be regarded as a non-monetary incentive and make the participants stay positive under relative low payoff. Therefore, by improving the utility of sensing data, the incentive cost can be effectively reduced. To maximize the data utility, we further design a tree-based structure to improve the priority experience replay mechanism of Proximal Policy Optimization (PPO) in UMIM. This improvement makes the high priority experience to be sampled more quickly and efficiently, as a result, the network can learn more effectively. Numerical results show that UMIM can further improve the data utility and have better convergence.

Learning effective image representation and constructing a suitable metric space are two main challenges in few-shot image classification. Existing methods normally consider the joint characteristic distribution of the image to improve the image representation ability, but it also brings high computational cost and produces high-dimensional embedding vectors, which limit their real-world applicability. In this paper, in order to reduce computational complexity and embedding dimension, we propose an effectively low-dimensional feature representation module (LDFR), which introduces a window mechanism to make the model focus on the correlation between local channels by using Brownian Distance Covariance. Furthermore, we combine LDFR with Hybrid Attention (LDFR-HA) to solve the problem of unbalanced sample features in few-shot image classification. Specifically, weighted average is adopted in the Hybrid Attention to construct the metric space and perform classification. Numerous experiments are conducted on two standard few-shot image classification benchmarks, i.e., general object recognition and fine-grained categorization. Extensive evaluations demonstrate that LDFR-HA significantly outperforms existing approaches. On popular datasets miniImageNet and CUB, LDFR-HA achieves 3.52 percentage point(pp)/2.29pp and 2.1pp/1.45pp gains over the state-of-the-art method on 5-way 1-shot/5-shot tasks, respectively.

The popularity of the microservices architecture has led to an evident trend of integration in system design, enabling systems to integrate numerous services to meet the diverse requirements of users. However, with extensive integration, it is difficult to quickly identify the root cause of the issues. In this paper, we propose MicroKGCL, a knowledge graph for root cause localization of feedback issues in microservices. Through the knowledge graph constructed with historical cases, MicroKGCL explores the potential relationship between user feedback and the root cause of system issues, it analyzes possible issues and ranks candidate root causes. In detail, in order to provide a more accurate representation of the feedback profile, we design a multi-modal embedding block, which utilizes a contrastive learning model and BERT model to extract feedback features from visual and content modalities. Experimental results demonstrate the superiority of the proposed MicroKGCL in terms of both MRR and Hits@n by comparing with the baselines. Moreover, to further verify the MicroKGCL, it has been deployed in the real production environment of Ant Group, achieving a top-3 hit ratio of 75% and a coverage ratio of 81.8%.

In recent years, advanced machine learning techniques have demonstrated remarkable achievements in many areas. Despite the great success, one of the bottlenecks in applying machine learning techniques in real world applications lies in the lack of a large amount of high-quality training data from diverse domains. Meanwhile, massive personal data is being generated by mobile devices and is often underutilized. To bridge the gap, we propose a general dataset purchasing framework, named CrowdBuy and CrowdBuy++, based on crowdsourcing, with which a buyer can efficiently buy desired data from available mobile users with quality guarantee in a way respecting users' data ownership and privacy. We present a complete set of tools including privacy-preserving image dataset quality measurements and image selection mechanisms, which are budget feasible, truthful and highly efficient for mobile users. We conducted extensive evaluations of our framework on large-scale images and demonstrate that the system is capable of crowdsourcing high quality datasets while preserving image privacy with little computation and communication overhead.

A large number of recommender systems have been developed to serve users with interesting news, ads, products or other contents. One main limitation with the existing work is that they do not take into account the inventory size of of items to be recommended. As a result, popular items are likely to be out of stock soon as they have been recommended and sold to many users, significantly affecting the impact of recommendation and user experience. This observation motivates us to develop a novel aware recommender system. It jointly optimizes the recommended items for all users based on both user preference and inventory sizes of different items. It requires solving a non-smooth optimization involved estimating a matrix of nxn, where n is the number of items. With the proliferation of items, this approach can quickly become computationally infeasible. We address this challenge by developing a dual method that reduces the number of variables from n^2 to n, significantly improving the computational efficiency. We also extend this approach to the online setting, which is particularly important for big promotion events. Our empirical studies based on a real benchmark data with 100 millions of user visits from Tmall verify the effectiveness of the proposed approach.