DEEP LEARNING & MACHINE LEARNING

• Minkyu Kim and Suan Lee
• Big Data and Smart Computing (BigComp), 2021 IEEE International Conference on. IEEE, 2021.
• January 17-20, 2021

Abstract

Those who want to start their own businesses must decide a location and service to start. In order to make the decision, they must know characteristics of the location and service, such as average revenues and floating population. However, it is usually very difficult to collect and analyze these characteristics. Therefore, we propose a novel deep learning model named Neural Tensor Factorization (NeuralTF) that automatically analyzes the characteristics for predicting revenues, and a method for recommending appropriate location or service to start their businesses based on the predicted revenues. NeuralTF is a combination of Tensor Factorization(TF) and Deep Neural Network(DNN). We compare NeuralTF with other machine learning models using Seoul Commercial Alley dataset. In addition, we compare performances of NeuralTF when TF and DNN components share the embedding space and when they do not.

Keywords

Recommender System, Tensor Factorization, Neural Network, Deep Learning, Neural Tensor Factorization

• Arousha Haghighian Roudsari, Jafar Afshar, Suan Lee, and Wookey Lee
• Big Data and Smart Computing (BigComp), 2021 IEEE International Conference on. IEEE, 2021.
• January 17-20, 2021

Abstract

Patent text mining is an important task that requires domain knowledge. The patent text is sometimes not clear and contains many ambiguous and technical words. Traditional text mining approaches are not satisfactory enough for patent text mining. In this paper, we consider various embedding techniques for patent documents and try to find how to represent the patent text for other downstream tasks such as patent classification, patent recommendation, finding similar patents, knowledge mining, etc. We compare several embedding approaches with the patent classification task. The experimental results demonstrate that using contextual word embeddings can perform better than the conventional static word embedding approaches.

Keywords

• Minkyu Kim, Suan Lee, and Jinho Kim
• 2020 IEEE International Conference on Big Data (IEEE BigData 2020)
• December 10-13, 2020

Abstract

Regression analysis is one of the most widely used data analysis methods, and it is increasingly important to obtain accurate results from it. To obtain accurate prediction results of regression analysis through machine learning, we must select appropriate features and train various feature interactions. The combinatorial model consists of a combination of various subordinate components and is used for automatic training of various feature interactions. However, existing combinatorial models are inefficient b ecause t hey c an t rain o nly limited feature interactions and must combine several components. To overcome these limitations, this study proposes a new model called eXtreme Interaction Network (XIN). XIN can automatically learn various explicit interactions, various levels of implicit higher-order interactions, and polynomial features. We compared the proposed XIN with existing models using four datasets with different characteristics to demonstrate that the proposed model has higher performance and lower or comparable time and space complexities. Furthermore, we conducted experiments while changing the various hyper-parameters of the XIN and demonstrated the improved performance of the proposed method in various environments.

Keywords

Neural Networks, Deep Learning, Cross Network, Combinatorial Model, Feature Interactions

• Dylan Myungchul Kang, Charles Cheolgi Lee, Suan Lee, and Wookey Lee
• 24th International Database Engineering & Applications Symposium (IDEAS 2020)
• August 2020

Abstract

A patent is one of the essential indicators of new technologies and business processes, which becomes the main driving force of the companies and even the national competitiveness as well, that has recently been submitted and exploited in a large scale of quantities of information sources. Since the number of patent processing personnel, however, can hardly keep up with the increasing number of patents, and thus may have been worried about from deteriorating the quality of examinations. In this regard, the advancement of deep learning for the language processing capabilities has been developed significantly so that the prior art search by the deep learning models also can be accomplished for the labor-intensive and expensive patent document search tasks. The prior art search requires differentiation tasks, usually with the sheer volume of relevant documents; thus, the recall is much more important than the precision, which is the primary difference from the conventional search engines. This paper addressed a method to effectively handle the patent documents using BERT, one of the major deep learning-based language models. We proved through experiments that our model had outperformed the conventional approaches and the combinations of the key components with the recall value of up to '94.29%' from the real patent dataset.

Keywords

Prior Art Search, Patent Document Classification, Language Model

• Minkyu Kim, Suan Lee, and Jinho Kim
• Big Data and Smart Computing (BigComp)
• February 19-22, 2020

Abstract

Wide & Deep model is deep learning model that jointed between of wide component and deep component for regression, recommendation, and classification. However, there is no study of regression analysis using Wide & Deep model. Also, Wide component of Wide & Deep model only deal with categorical variables and that need hand-crafted variables for efficient training. Therefore, this paper propose Lattice Wide & Deep Architecture which improve Wide & Deep model. Furthermore we show that the Lattice Wide & Deep model has better performance than Wide & Deep model in regression analysis.

Keywords

lattice wide & deep, wide & deep learning, regression analysis

• SungJin Park, Suan Lee, and Jinho Kim
• Big Data and Smart Computing (BigComp)
• February 19-22, 2020

Abstract

As for people who wish to start their own businesses, their concerns are whether they could survive during their operations because most stores or services in Seoul are not able to survive for more than 5 year possibly due to the poor decision of location/service to start. In order to solve this problem, using big data could be helpful to increase the survival rate. Singular Value Decomposition (SVD) has been widely used in finding the similarity between all pairs of alleys and obtaining predictions from unknown relevance scores. Since tensor decomposition is the extension of SVD for multi-dimensional data, using this method to find the similarity between all pair of alleys could be the solution of increasing survival rate. This paper aims to generate good prediction tensor, TENSORCABS, that is able to recommend users appropriate alley location to start their businesses or the appropriate services to start at the user’s desired location. Both CP and Tucker decompositions are used and compared to evaluate which method has better performance. Also, r-square for regression problem and precision & recall for top-k recommendation performance are used to evaluate the TENSORCABS. As results, actual and predicted values are good-fitted, and prediction tensor performs well on the top-k recommendation. In addition, Tucker outperforms CP for this situation. Therefore, the proposed method has advantages that can handle high-dimensional data and can use decomposition for recommendations of various perspectives. With this method, users are able to obtain recommendations of the appropriate alleys with predicted revenues for opening any business service or the appropriate services with predicted revenues on a user’s desired alley location.

Keywords

Recommendation System, Tucker Decomposition, Canonical Polyadic, Collaborative Filtering, Top-k Recommendation, Tensor, Commercial Alley

• Toshpulatov Mukhiddin, WooKey Lee, Suan Lee, Tojiboev Rashid
• The Second International Workshop on Big Data, Cloud, and IoT Technologies for Smart Cities (IWBigDataCity2020)
• February 19, 2020

Abstract

Keywords

• Myungchul Kang, Suan Lee, and Wookey Lee
• The 3rd International Workshop on Dialog Systems (IWDS2020) 2020.
• February 19, 2020

Abstract

Keywords