Knowledge Management and Acquisition for Intelligent Systems

Backpropagation (BP) is the most widely used algorithm for the training of deep neural networks (DNN) and is also considered a de facto standard algorithm. However, the BP algorithm often requires a lot of computation time, which remains a major challenge. Thus, to reduce the time complexity of the BP algorithm, several methods have been proposed so far, but few do not apply to the BP algorithm. In the meantime, a new DNN algorithm based on nonnegative matrix factorization (NMF) has been proposed, and the algorithm has different convergence characteristics from the BP algorithm. We found that the NMF-based method could lead to rapid performance improvement in DNNs training, and we developed a technique to accelerate the training time of the BP algorithm. In this paper, we propose a novel training method for accelerating the BP algorithm by using an NMF-based algorithm. Furthermore, we present a technique to boost the efficiency of our proposed method by concurrently training DNNs with the BP and NMF-based algorithms. The experimental results indicate that our method significantly improves the training time of the BP algorithm.

Intuitively, Automated Planning systems capable of learning from previous experiences should be able to achieve better performance. One way to build on past experiences is to augment domains with macro-operators (i.e. frequent operator sequences). In most existing works, macros are generated from chunks of adjacent operators extracted from a set of plans. Although they provide some interesting results this type of analysis may provide incomplete results. In this paper, we propose ERA, an automatic extraction method for macro-operators from a set of solution plans. Our algorithm is domain and planner independent and can find all macro-operator occurrences even if the operators are non-adjacent. Our method has proven to successfully find macro-operators of different lengths for six different benchmark domains. Also, our experiments highlighted the capital role of considering non-adjacent occurrences in the extraction of macro-operators.

In fashion electronic commerce services, two item-based recommendation approaches, image similarity-based and click likelihood-based, are used to improve the revenue of a website. To improve accuracy, in this paper, we propose a hybrid model, a deep neural network (DNN) that predicts click probability of a target fashion item by incorporating both image similarity and click likelihood. To create an image similarity feature, we acquire a latent image feature through a CNN-based classification of fashion color, type and pattern. To create a click likelihood feature, we calculate matrix factorization (MF) and use decomposed item features as latent click log feature. To solve a cold-start problem (recommendation of new items), we complement the latent log features of new items with those of existing ones. An offline evaluation shows that the accuracy of proposed model (both log and image) improved by 14% compared with matrix factorization (log only) and 56% the image-only model. Moreover, the complement of latent log features changes the new item ratio to six times.

Predicting health risk from electronic health records (EHRs) is increasingly demanded in the medical and health fields. Many studies have pursued prediction accuracy while ignoring the interpretability of their developed models. To encourage lifestyle changes by patients and employees, an appropriate explanation of why the model outputs high risk is as important as accurately predicting the health risk. In this study, we construct 33 predictive models (11 health-checkup items checked after one, two, and three years). We also clarify a problem in the existing Local Interpretable Model-agnostic Explanations (LIME), namely, inconsistency among the health-risk predictions of the three target years. To resolve this problem, we find and exclude an anomalous sample that deteriorate the interpretation, and output a consistent interpretation of the health-risk predictions over the three years. We evaluate proposed method using more than 10,000 medical examination data. Accuracy was improved by 16% at the maximum compared to the baseline that output the risk at year Y + 1,2,3 equaling to that at year Y. Also, proposed LIME called C-LIME improve number of employees whom we can provide consistent lifestyle advice over the years three times compared to LIME. We have released a health-risk prediction and lifestyle recommendation service using proposed method for employees of the Nippon Telegraph and Telephone Group from April of 2019.

Visual inspection of electrocardiograms (ECGs) is a common clinical practice to diagnose heart diseases (HDs), which are still responsible for millions of deaths globally every year. In particular, myocardial infarction (MI) is the leading cause of mortality among HDs. ECGs reflect the electrical activity of the heart and provide a quicker process of diagnosis compared to laboratory blood tests. However, still it requires trained clinicians to interpret ECG waveforms, which poses a challenge in low-resourced healthcare systems, such as poor doctor-to-patient ratios. Previous works in this space have shown the use of data-driven approaches to predict HDs from ECG signals but focused on domain-specific features that are less generalizable across patient and device variations. Moreover, limited work has been conducted on the use of longitudinal information and fusion of multiple ECG leads. In contrast, we propose an end-to-end trainable solution for MI diagnosis, which (1) uses 12 ECG leads; (2) fuses the leads at data-level by stacking their spectrograms; (3) employs transfer learning to encode features rather than learning representations from scratch; and (4) uses a recurrent neural network to encode temporal dependency in long duration ECGs. Our approach is validated using multiple datasets, including tens of thousands of subjects, and encouraging performance is achieved.

Several studies have demonstrated a good predictive performance of ear emergence in rice crops. However, significant regional variations in performance have been discovered and they remain unsolved. In this study, we aim to realize a stable predictive performance for ear emergence in rice crops regardless of its regional variations. Although a variety of data that represents regional characteristics have been adopted as the variables for prediction in related work, stability of the predictive performance has not been attained. These results imply that explicit regional data is insufficient for stabilizing the regional variances of the prediction. This study proposes to use engineered variables that uncover hidden regional characteristics behind the explicit regional data. Pre-examinations of the regional data indicate distinctive patterns of time dependency according to each region. Based on the findings, hidden Markov models are applied to the micro climate data to create engineered variables that represent the implicit time dependent regional characteristics. The efficiency of these variables is empirically studied, and the results show a significant improvement in the regional predictive variance.

In recent years, the expectation that new businesses and economic value can be created by combining/exchanging data from different fields has risen. However, value creation by data exchange involves not only data, but also technologies and a variety of stakeholders that are integrated and in competition with one another. This makes the data exchange ecosystem a challenging subject to study. In this paper, we propose a framework for describing the stakeholder-centric value chain (SVC) of data by focusing on the relationships among stakeholders in data businesses and discussing creative ways to use them. The SVC framework enables the analysis and understanding of the structural characteristics of the data exchange ecosystem. We identified stakeholders who carry potential risk, those who play central roles in the ecosystem, and the distribution of profit among them using business models collected by the SVC.

Network embedding aims to embed the network into a low-dimensional vector space wherein the structural characteristic of the network and the attribute information of nodes are preserved as much as possible. Many existing network embedding works focused on the homogeneous or heterogeneous plain networks. However, networks in the real world are usually not plain since the nodes in the networks have rich attributes, and these attributes play important roles for encoding nodes’ vector representations. Although some works took into account the attribute information, they could not handle the homogeneous and heterogeneous structure information of the network simultaneously. In order to solve this problem, a new network embedding method that considers both the network’s homogeneous and heterogeneous structure information and nodes attribute information simultaneously is proposed in this paper. The proposed method first obtains nodes attribute information, homogeneous and heterogeneous structure information as three views of the network and learns network embeddings of the three views through different technologies respectively. Then, an attention mechanism is utilized to fuse the embedding results learned from the three views to obtain the final vector representations of nodes. We verify the performance of the proposed model through link prediction tasks on four real-world datasets, and extensive experimental results show that the proposed model outperforms the advanced baselines.

Due to the increase in spam email and other anti-social behavior (such as the bot net command and control server) on the Internet, Domain Name System (DNS) blacklists (DNSBLs) have been created. These are “blacklists” of malicious hosts on the Internet that are reputed to send spam email and other anti-social behavior. Most email servers can be configured to reject messages that are sent from the hosts on these lists. Because it is difficult to keep up with new malicious hosts created every day, research is required to automate DNSBL generation. To address this problem, the application of machine learning is being studied thoroughly. Deep learning is considered to be a promising approach to classify the malicious host names. This study explores the risks of these approaches by showing a simple domain generation algorithm (DGA). This report shows the importance of attributes that are used rather than machine learning techniques. Researchers in machine learning and knowledge acquisition should focus on attributes that are more important in application fields than techniques when considering the practical applications.

This paper proposes a method to analyze the temporal change in the quality of LoRa communication, which is a low-power wide-area (LPWA) wireless technology for the IoT, using massive measurement data from a bus location management system. To analyze this temporal change in LoRa communication quality, many measurement data at the same position must be compared. However, buses do not always follow the same route every day. Even if buses follow the same route on different days, the measurement data are not acquired from exactly the same position. To solve these problems, we use locality-sensitive hashing (LSH) to extract comparable data from massive measurement data. Furthermore, we treat the extracted data with different distances on the same day as a series and compare regression formulas determined respectively from the plural series. As a result, it was confirmed that the received signal strength indicator (RSSI) of LoRa communication had almost no effect even in heavy rainfall of 30 mm/h. In addition, we confirmed that the radio attenuation was smaller about 2 dB than usual when there was over 7 cm of snow cover.

Prior work in standardized science exams requires support from large text corpus, such as targeted science corpus from Wikipedia or SimpleWikipedia. However, retrieving knowledge from the large corpus is time-consuming and questions embedded in complex semantic representation may interfere with retrieval. Inspired by the dual process theory in cognitive science, we propose a MetaQA framework, where system 1 is an intuitive meta-classifier and system 2 is a reasoning module. Specifically, our method based on meta-learning method and large language model BERT, which can efficiently solve science problems by learning from related example questions without relying on external knowledge bases. We evaluate our method on AI2 Reasoning Challenge (ARC), and the experimental results show that meta-classifier yields considerable classification performance on emerging question types. The information provided by meta-classifier significantly improves the accuracy of reasoning module from \(46.6\%\) to \(64.2\%\), which has a competitive advantage over retrieval-based QA methods.

Recently, the business-to-business (B2B) corporate brands have garnered attention. Studies using large-scale B2B company brand surveys across companies and industries have not been conducted because it is generally difficult to identify people who possess some familiarity with B2B companies externally. In this study, we use a B2B corporate brand survey data using a large business card exchange network in Japan as a proxy variable for brand power. We use the survey data to investigate the relationship between stakeholders’ impressions of B2B corporate brands and corporate performance. The results show that firms with high brand have high performance. We also identified the B2B brand components using supervised topic models, and we clarified the relationship with performance. These results are not only new findings for B2B brand research but also useful for brand strategies of B2B companies.

Readability formulas consider word familiarity as one of the factors for predicting the readability of children’s books. Word familiarity is dependent on the frequency in which the words are encountered in daily reading. Often referred to as “sight words”, developing effective recognition of these high-frequency words can assist young readers to develop their reading fluency and comprehension. In this paper, we describe our work in building a dictionary of sight words for Filipino with the use of a corpus of Filipino literary materials written for children. We expanded the dictionary to a total of 664 words with the use of pre-trained word embedding model. The availability of such dictionary can facilitate the development of a readability formula for Filipino text, especially in the context of its lexical complexity.

The recent outbreak of the coronavirus disease (COVID-19) rapidly spreads across most of the countries. To alleviate the panics and prevent any potential social crisis, it is essential to effectively detect public concerns through social media. Twitter, a popular online social network, allows people to share their thoughts, views and opinions towards the latest events and news. In this study, we propose a deep learning-based framework to explore public concerns for COVID-19 automatically, where Twitter has been utilised as the key source of information. We extract and analyse public concerns towards the pandemic. Furthermore, as part of the proposed framework, a knowledge graph of the extracted public concern has been constructed to investigate the interconnections.