Many real-world applications require precise and fast time-series forecasting. Recent trends in time-series forecasting models are shifting from LSTM-based models to Transformer-based models. However, the Transformer-based model has a limited ability to represent sequential relationships in time-series data. In addition, the transformer-based model suffers from slow training and inference speed due to the bottleneck incurred by a deep encoder and step-by-step decoder inference. To address these problems, we propose a time-series forecasting optimized Transformer model, called TS-Fastformer. TS-Fastformer introduces three new optimizations: First, we propose a Sub Window Tokenizer for compressing input in a simple manner. The Sub Window Tokenizer reduces the length of input sequences to mitigate the complexity of self-attention and enables both single and multi-sequence learning. Second, we propose Time-series Pre-trained Encoder to extract effective representations through pre-training. This optimization enables TS-Fastformer to capture both seasonal and trend representations as well as to mitigate bottlenecks of conventional transformer models. Third, we propose the Past Attention Decoder to forecast target by incorporating past long short-term dependency patterns. Furthermore, Past Attention Decoder achieves high performance improvement by removing a trend distribution that changes over a long period. We evaluate the efficiency of our model with extensive experiments using seven real-world datasets and compare our model to six representative time-series forecasting approaches. The results show that the proposed TS-Fastformer reduces MSE by 10.1% compared to state-of-the-art model and demonstrates 21.6% faster training time compared to the existing fastest transformer, respectively.

Numerous research efforts have been proposed for efficient processing of range queries in high-dimensional space by either redesigning R-tree access structure for exploring massive parallelism on single GPU or exploring distributed framework of R-tree. However, none of the existing efforts explores the integration of the parallelization of the R-tree on a single GPU with a distributed framework for the R-tree. The problem of designing an efficient multi-GPU indexing method, which can effectively combine the parallelism maximization with distributed processing of the R-tree, remains an open challenge. In this article, we present a novel multi-GPU efficient parallel and distributed indexing method, called LBPG-tree. The rationale of the LBPG-tree is to combine the advantages of an instruction pipeline in CPU with the massive parallel processing potential of multiple GPUs by introducing two new optimization strategies: First, we exploit the GPU L2 cache for accelerating both index search and index node access on GPUs. Second, we further improve utilization of L2 cache on GPUs by compacting and sorting candidate nodes called Compact-and-Sort. Our experimental results show that the LBPG-tree outperforms G-tree, the previous representative GPU index method and effectively support multiple GPUs for providing efficient high dimensional range query service.

The problem of the curse of dimensionality for processing large high-dimensional datasets has been an open challenge. Numerous research efforts have been proposed for improving query performance in high-dimensional space through hierarchical indexing using the R-tree or its variants and exploring parallel processing of the R-tree on GPUs. Despite these existing efforts, the curse of dimensionality remains to be a grand challenge since the existing methods deteriorate drastically as the dimensionality of datasets increases. To cope with this problem, we present a novel GPU-aware parallel indexing method called G-tree, which offers consistent and stable performance in high-dimensional space. The rationale of the G-tree is to combine the efficiency of the R-tree in low-dimensional space with the massive parallel processing potential of GPUs by introducing a new data structure and three new optimization techniques to better utilize the GPU memory structure for accelerating both index search and index node access on GPUs. The first two optimizations promote effective parallelism utilization in GPU memory access. We dedicate the third optimization to further speed up the G-tree index by conducting progressive filtering using our dimension filters. We evaluate the validity of the G-tree approach by extensive experiments on high-dimensional datasets, showing that the G-tree outperforms the existing state-of-the-art techniques.

This paper presents a cloud-based positioning method that leverages visualized signal images through visual analytics and deep learning. At a mobile client, such as a smart phone, this approach transforms multidimensional signals captured at a known location and at a given time into a signal image and transmits such visual signal images to the Cloud. By collecting and storing many such visual signal images as fingerprints, we can build a visual signal image cloud and produce a signal image map for the geographical region of interest and utilize such signal image map to serve the positioning requests of mobile clients on the move. When a user Alice wants to know her current position, her mobile client will generate a signal image from the multiple signals it receives with timestamp and send this query image to a Cloud server. The server searches the existing signal images stored in the cloud to find those that are similar to this query signal image and estimates the positioning of Alice based on the locations of those similar signal images collected by the server. We evaluate our visual signal images based positioning system on the entire two floors of a large department store and on the street and shops outside the department store, with the signal images collected over 30 minutes before serving positioning queries. The mean error of up to 4 meters is observed. To further verify the applicability of the proposed method, extensive experiments were conducted to distinguish whether a user is indoor or outdoor by applying a deep learning algorithm with 60% of signal images collected for training and 40% signal images for testing. This experiment shows that the proposed method is able to distinguish indoor and outdoor with accuracy of about 95%.

Recently, various research efforts have been conducted to develop strategies for accelerating multi-dimensional query processing using the Graphics Processing Units (GPUs). Well-known multidimensional access methods such as R-tree, B-tree and their variants, however, are hardly applicable to GPUs in practice, mainly due to the characteristics of a hierarchical structure. More specifically, the structure not only causes frequent transfers of small volume of data but also gives limited opportunities to exploit GPUs' advanced data parallelism. To address the problems, we propose an approach which uses GPUs as a buffer. Main idea is that object entries in recently visited leaf nodes are buffered in the global memory of the GPU and are processed by massive threads of GPUs in parallel. In this paper, we presented both a single buffer strategy and a multiple buffer strategy since a single buffer strategy has problems deteriorating query performance seriously. Through the extensive performance studies, we observed that the proposed approach achieved up to 5 times higher query performance than the original CPU-based R-trees.