Vol. 48, No. 11,
Nov. 2021
Digital Library
An Automatic Parameter Optimizing Scheme for RocksDB
Jiwon Kim, Hyeonmyeong Lee, Sungmin Jung, Heeseung Jo
http://doi.org/10.5626/JOK.2021.48.11.1167
For users with low understanding of application, it is very difficult to optimize a complex application. Leading studies that optimize application using one or two parameters can enhance the performance of an application. However, it is difficult to consider the relationship between various parameters using a single parameter optimization. In this paper, we proposed two techniques, LDH-Force and PF-LDH, that could optimize several parameters at the same time. The LDH-Force technique could efficiently reduce the number of searches by adding an LDH process, while simultaneously finding the optimal parameter combination for several parameters. The PF-LDH technique could further reduce the search cost by adding a filtering process and confirming that the degree to which the parameter affects the performance is different. Evaluation results confirmed that the proposed scheme had performance improvement of up to 42.55 times. The proposed scheme was able to find the optimal parameter combination at the lowest search cost without user intervention under various workloads.
Fair Feature Distillation Using Teacher Models of Larger Architecture
http://doi.org/10.5626/JOK.2021.48.11.1176
Achieving algorithmic fairness is becoming increasingly essential for various vision applications. Although a state-of-the-art fairness method, dubbed as MMD-based Fair feature Distillation (MFD), significantly improved accuracy and fairness via feature distillation based on Maximum Mean Discrepancy (MMD) compared to previous works, MFD could be limitedly applied into when a teacher model has the same architecture as a student model. In this paper, based on MFD, we propose a systematic approach that mitigates unfair biases via feature distillation of a teacher model of larger architecture, dubbed as MMD-based Fair feature Distillation with a regressor (MFD-R). Throughout the extensive experiments, we showed that our MFD-R benefits from the use of the larger teacher compared to MFD as well as other baseline methods.
Ensemble of Sentence Interaction and Graph Based Models for Document Pair Similarity Estimation
Seonghwan Choi, Donghyun Son, Hochang Lee
http://doi.org/10.5626/JOK.2021.48.11.1184
Deriving the similarity between two documents, such as, news articles, is one of the most important factors of clustering documents. Sequence similarity models, one of the existing deep-learning based approaches to document clustering, do not reflect the entire context of documents. To address this issue, this paper uses interaction-based and graph-based approaches to construct document pair similarity models suitable for news clustering. This paper proposes four interaction-based models that measures the similarity between two documents through the aggregation of similarity information in the interaction of sentences. The experimental results demonstrated that two out of these four proposed models outperformed SVM and HAN. Ablation studies were conducted on the graph-based model through experiments on the depth of the model’s neural network and its input features. Through error analysis and ensemble of models with an interaction and graph-based approach, this paper showed that these two approaches could be complementarity due to the differences in their prediction tendencies.
Improving Super-Resolution GAN Performance through Discriminator using U-Net Structure and Auxiliary Classifier
Dong Min Cheon, Younghwan Jeong, Wonsik Lee, Sounghyouk Wi, Sangjin Nam
http://doi.org/10.5626/JOK.2021.48.11.1194
In this paper, We propose a new super resolution method using a Generative Adversarial Network(GAN). Several super resolution techniques, including interpolation, CNN(Convolutional Neural Network), and GAN, have been proposed. Among them, GAN is the most preferred because of its good performance in image synthesis. Consequently, there have been many attempts to improve the super resolution quality by changing the network structure and loss function of GAN’s Generator, but the focus of improvement was not focused on the discriminator. The findings of the present study confirmed that the U-Net structure and the auxiliary classifier structure for image rotation, which were presented in other papers, had a positive effect on super-resolution network.
Knowledge Completion System using Neuro-Symbolic-based Rule Induction and Inference Engine
Won-Chul Shin, Hyun-Kyu Park, Young-Tack Park
http://doi.org/10.5626/JOK.2021.48.11.1202
Recently, there have been several studies on knowledge completion methods aimed to solve the incomplete knowledge graphs problem. Methods such as Neural Theorem Prover (NTP), which combines the advantages of deep learning methods and logic systems, have performed well over existing methods. However, NTP faces challenges in processing large-scale knowledge graphs because all the triples of the knowledge graph are involved in the computation to obtain prediction results for one input. In this paper, we propose an integrated system of deep learning and logic inference methods that can learn vector representations of symbols from improved models of computational complexity of NTP to rule induction, and perform knowledge inference from induced rules using inference engines. In this paper, for rule-induction performance verification of the rule generation model, we compared test data inference ability with NTP using induced rules on Nations, Kinship, and UMLS data set. Experiments with Kdata and WiseKB knowledge inference through inference engines resulted in a 30% increase in Kdata and a 95% increase in WiseKB compared to the knowledge graphs used in experiments.
Improving Performance of Recurrent Neural Network based Recommendations by Utilizing Personal Preferences
Dong Shin Lim, Yong Jun Yang, Shin Cho
http://doi.org/10.5626/JOK.2021.48.11.1211
As the amount of content provided on the platform surged, a recommendation system became an essential element of the platform. The collaborative filtering technique is a widely used recommendation system in academia and industry, but it also has a limitation because it relies on quantitative information from consumers such as ratings and purchase history. To overcome this shortcoming, various studies have been done in a bid to improve its performance by collecting qualitative information such as review data in a model. Recently, some studies that applied recurrent neural networks showed better performance than the existing recommendation system by using time-series behavioral data only, but studies that reflect each customer"s preference in the recommendation model have not yet been made. In this paper, an improved recommendation model was presented by calculating a preference matrix based on customer log data and learning it in a recurrent neural network through an embedding vector. It was confirmed that the prediction performance was improved compared to the existing recurrent neural network recommendation model.
Double-Averaging Acceleration with Synchronization Barrier Repositioning and Pipelining in Deep Learning
http://doi.org/10.5626/JOK.2021.48.11.1221
In deep learning using distributed computing, synchronization is one of the most important factors. While Local SGD is a low-frequency synchronization method that enables fast training, it is limited by high convergence difficulties. And Double-Averaging and SlowMo have been proposed to reduce the convergence difficulties of Local SGD. Double-Averaging improves the convergence difficulties by adding momentum buffer synchronization. However, the training time also increases due to the increased data synchronization. On the other hand, SlowMo adds a Two-layer momentum structure to the Local SGD resulting in reduced convergence difficulties without additional synchronization. However, this requires finding the appropriate SlowMo hyper-parameters. Therefore, in this paper, we proposed accelerated Double-Averaging via synchronization barrier repositioning and pipelining. The proposed method significantly reduced the convergence difficulties and accelerated performance.
Knowledge Graph Completion using Hyper-class Information and Pre-trained Language Model
http://doi.org/10.5626/JOK.2021.48.11.1228
Link prediction is a task that aims to predict missing links in knowledge graphs. Recently, several link prediction models have been proposed to complete the knowledge graphs and have achieved meaningful results. However, the previous models used only the triples" internal information in the training data, which may lead to an overfitting problem. To address this problem, we propose Hyper-class Information and Pre-trained Language Model (HIP) that performs hyper-class prediction and link prediction through a multi-task learning. HIP learns not only contextual relationship of triples but also abstractive meanings of entities. As a result, it learns general information of the entities and forces the entities connected to the same hyper-class to have similar embeddings. Experimental results show significant improvement in Hits@10 and Mean Rank (MR) compared to KG-BERT and MTL-KGC.
Instagram User Embedding and Fashion Photo Recommendation Using "likes" of Fashion Photos
http://doi.org/10.5626/JOK.2021.48.11.1235
As individual preference of fashion styles diversifies, demands for research recommending personalized fashion are increasing. Recently, with the development of deep learning technology, many studies have been conducted using deep learning to extract features from fashion photos and use them for recommendations. In this work, we exploit social network data to consider users and fashion styles in recommending fashion photos. Since social network users tend to post fashion photos in their preferred style and tag them with “Like“, social network data are very important for understanding relationship between users and fashion photos. We propose a technique to map users and fashion photos into the same vector space using social network data structure which consists of users and fashion photos. Especially, it is possible to use our method to recommend fashion photos that a user might prefer by mapping users and fashion photos not used for learning into a vector space without additional learning.
Learning-based QoS Path Prediction Method in SDN Environment
Seunghoon Jeong, Seondong Heo, Hosang Yun
http://doi.org/10.5626/JOK.2021.48.11.1241
When Quality of Service (QoS) is supported by flow path control in Software-Defined Networking (SDN) environment, the current simple least cost path finding method can cause inefficient rerouting problems. The measured performance of the flow path derived based on the link quality may differ from the predicted performance. In particular, in the case of sequential QoS condition search for candidate paths, the effectiveness of path-based QoS support may decrease due to repeatedly searching for the same path previously identified as the final path. In this paper, we propose a learning-based QoS path search model. The model learns the path that finally satisfies the QoS conditions according to the network state, and predicts the QoS path for the network state when rerouting is required. The experiment shows that this learning model can reduce unnecessary path iteration search costs given the similar network conditions, and is more effective than other learning-based models in a service environment that requires rapid QoS quality restoration.
Search
Journal of KIISE
- ISSN : 2383-630X(Print)
- ISSN : 2383-6296(Electronic)
- KCI Accredited Journal
Editorial Office
- Tel. +82-2-588-9240
- Fax. +82-2-521-1352
- E-mail. chwoo@kiise.or.kr