5 March 2024 in IEEE Transactions on Consumer Electronics
Online reviews on E-commerce platforms contain users’ opinions and emotions about the purchased products. How to enable machines to understand the semantics of review texts and judge their emotional tendencies has become a hot topic of common concern in academia and industry. However, review texts often come from various domains in practical applications, thus there are differences in the description and emotional vocabulary. In this case, traditional classification models trained on review texts from a single domain often have poor classification results. This paper proposes a cross-domain discriminative subspace classification (CDDSC) algorithm based on domain adaptation and subspace learning. First, we adopt the progressive data projection strategy from different domains sequentially into the low-rank subspace and label space, which can reduce the interference of useless information on classification task while mining the global structure of data in each domain. Second, to reduce domain differences, we adopt the dual metric techniques, i.e., maximum mean discrepancy (MMD) based on marginal distribution and MMD based on conditional distribution to approximate the inter-domain overall probability distributions and the inter-domain conditional probability distributions, respectively. Thirdly, we use the slack label and low-rank constraint technique to promote inter-class interval and intra-class tightness. We conduct experiments with several review text corpus oriented Chinese E-commerce platforms. Experiments results show that the CDDSC algorithm obtains good sentiment classification accuracy on Chinese E-commerce platforms.
1 March 2024 in IEEE Transactions on Consumer Electronics
With the rapid development of smart healthcare, digital histopathological images are playing an increasingly important role in disease diagnosis. Due to factors such as diverse pathological information and individual patient differences in histopathological images, the differences of intra-class sample in may be greater than the differences in inter-class sample. Additionally, collecting a large number of labeled images is very difficult. This study proposes semi-supervised dictionary pair learning (SSDPL) algorithm with local graph embedding and discriminant coding for lung histopathological image classification. All samples are used to train the synthesis-analysis dictionary pair in semi-supervised learning. Based on the graph structure of the neighborhood, a local graph embedding term is constructed. Considering the geometric neighborhood topological relationships between samples for each class, the local structural information can be accurately represented to the analysis dictionary, so the coding coefficients corresponding to samples with high similarity can maintain a high degree of similarity. Based on mean coefficient matrices, the discriminant coding term is introduced to improve the separability of each class. Additionally, the sparse l2,1-norms on the reconstruction error and analysis dictionary ensures the robust dictionary pair. The experiment on the lung histopathological image dataset shows that SSDPL obtains satisfactory performance.
1 January 2024 in IET Software
In the rapidly evolving landscape of social media, the demand for precise sentiment analysis (SA) on multimodal data has become increasingly pivotal. This paper introduces a sophisticated data lake architecture tailored for efficient multimodal emotion feature extraction, addressing the challenges posed by diverse data types. The proposed framework encompasses a robust storage solution and an innovative SA model, multilevel spatial attention fusion (MLSAF), adept at handling text and visual data concurrently. The data lake architecture comprises five layers, facilitating real‐time and offline data collection, storage, processing, standardized interface services, and data mining analysis. The MLSAF model, integrated into the data lake architecture, utilizes a novel approach to SA. It employs a text‐guided spatial attention mechanism, fusing textual and visual features to discern subtle emotional interplays. The model’s end‐to‐end learning approach and attention modules contribute to its efficacy in capturing nuanced sentiment expressions. Empirical evaluations on established multimodal sentiment datasets, MVSA‐Single and MVSA‐Multi, validate the proposed methodology’s effectiveness. Comparative analyses with state‐of‐the‐art models showcase the superior performance of our approach, with an accuracy improvement of 6% on MVSA‐Single and 1.6% on MVSA‐Multi. This research significantly contributes to optimizing SA in social media data by offering a versatile and potent framework for data management and analysis. The integration of MLSAF with a scalable data lake architecture presents a strategic innovation poised to navigate the evolving complexities of social media data analytics.
13 January 2022 in Advances in Multimedia
With the continuous development of social economy, the expansion of cities often leads to the disorderly utilization of land resources and even waste. In view of these limitations and requirements, this paper introduces the automatic extraction algorithm of closed area boundary, combs the requirements of urban boundary extraction involved in urban planning and design, and uses the technology of geospatial analysis to carry out spatial analysis practice from three angles, so as to realize the expansion of functional analysis of urban planning and design and improve the efficiency and rationality of urban planning. The simulation results show that the automatic extraction algorithm of closed area boundary is effective and can support the functional analysis of urban planning and design expansion.