Scholarlink

<p>Identification of the geolocation of users for various applications, including social networking, public health, safety, and marketing. Researchers have presented several approaches using diverse data sources such as textual data, social connections, and context. However, some challenges exist due to the large and diverse nature of data, which may result in lower prediction performance. Therefore, motivated by this, the paper presented a multi-layered attention-based learning model for location prediction, termed the spatio-temporal-friendship feature transition matrix attention model (STFFTMAM). This model is designed to effectively capture and integrate temporal, semantic, spatial, and social interaction features. For generating attention-weighted features, transition matrices are created using temporal, semantic, spatial, and friendship features to enhance the efficacy of location prediction. For the generation of these features, trajectory and friendship data are used that can identify users’ location patterns. To implement these, the entire model is divided into four layers: data aggregation layer (DAL), feature extraction transition matrix generation layer (FETMGL), weighted feature matrix generation layer (WFMGL), and learning layers (LL). DAL extracts the trajectory information and friendship information of users and pre-processes them. FETMGL extracts temporal, semantic, spatial, and friendship features and extracts transition matrices. WFMGL generates the weighted feature vector. The attention mechanism dynamically prioritizes features based on their relevance for accurate location forecasting. The LL layer used the bidirectional long short-term memory (BiLSTM) networks for more precise learning. The analysis of results was performed on the Foursquare dataset. The analysis of the proposed model’s results was compared with three baseline models. The proposed model achieved an accuracy of approximately 99% execution time of approx. 0.03&#xa0;s. As compared to state-of-the-art models such as graph convolutional networks (GCNNs) and genetic programming (GP), the proposed model has achieved approximately the same performance. About 3% of improvement in prediction performance while maintaining low execution time (≈0.03&#xa0;s). To support city-wide inference at high data rates, the STFFTMAM is engineered to benefit from HPC environments. Tasks such as distributed transition matrix computation, parallel graph embedding, and GPU-accelerated BiLSTM learning make STFFTMAM highly suitable for real-time deployment.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

User location prediction using a multi-layered learning model with weighted feature context transition matrix attention mechanism

  • 期刊论文

  • 发表日期：

    2025年10月29日

查看全文
  • Madhur Arora,
  • Sanjay Agrawal,
  • Ravindra Patel

摘要

Identification of the geolocation of users for various applications, including social networking, public health, safety, and marketing. Researchers have presented several approaches using diverse data sources such as textual data, social connections, and context. However, some challenges exist due to the large and diverse nature of data, which may result in lower prediction performance. Therefore, motivated by this, the paper presented a multi-layered attention-based learning model for location prediction, termed the spatio-temporal-friendship feature transition matrix attention model (STFFTMAM). This model is designed to effectively capture and integrate temporal, semantic, spatial, and social interaction features. For generating attention-weighted features, transition matrices are created using temporal, semantic, spatial, and friendship features to enhance the efficacy of location prediction. For the generation of these features, trajectory and friendship data are used that can identify users’ location patterns. To implement these, the entire model is divided into four layers: data aggregation layer (DAL), feature extraction transition matrix generation layer (FETMGL), weighted feature matrix generation layer (WFMGL), and learning layers (LL). DAL extracts the trajectory information and friendship information of users and pre-processes them. FETMGL extracts temporal, semantic, spatial, and friendship features and extracts transition matrices. WFMGL generates the weighted feature vector. The attention mechanism dynamically prioritizes features based on their relevance for accurate location forecasting. The LL layer used the bidirectional long short-term memory (BiLSTM) networks for more precise learning. The analysis of results was performed on the Foursquare dataset. The analysis of the proposed model’s results was compared with three baseline models. The proposed model achieved an accuracy of approximately 99% execution time of approx. 0.03 s. As compared to state-of-the-art models such as graph convolutional networks (GCNNs) and genetic programming (GP), the proposed model has achieved approximately the same performance. About 3% of improvement in prediction performance while maintaining low execution time (≈0.03 s). To support city-wide inference at high data rates, the STFFTMAM is engineered to benefit from HPC environments. Tasks such as distributed transition matrix computation, parallel graph embedding, and GPU-accelerated BiLSTM learning make STFFTMAM highly suitable for real-time deployment.