The quick delivery of emergency road services (ERS) to rescue disabled vehicles is a significant task for maintaining the meticulously planned logistics and transportation network in modern society. The location optimization of service base shops of ERS is an effective strategy for minimizing the waiting time of ERS users. For the optimization in real ERS situation, we must consider the effects of various uncertain factors such as trouble situations, shop characteristics, and backup service. Additionally, straightforward incorporation of such variables to location optimization problems causes computational complexity due to the calculation of higher-order nonlinear discrete optimization. The study presents a novel and simple analytical framework to incorporate the various factors in real ERS situations to reduce the waiting time of ERS users. In an experiment using the actual records of Bridgestone Corporation's emergency road service in Japan, we demonstrate that the proposed method successfully yields an equivalent result to the solution by the strict expected value optimization without high computational costs and reduces the waiting time by up to 27.0% relative to the conventional method.

Towards explainable affective computing (XAC), researchers have invested considerable effort into post hoc approaches and reverse engineering to seek explanations for deep learning models. However, alternative, intrinsic approaches that aim to build inherently interpretable models by restricting their complexity are yet to be widely explored. In this study, we integrate an explanatory polytomous item response model that provides a well-established psychological interpretation for ordinal scales with deep neural networks to realize high prediction performance and good result interpretability. We conducted an experiment on a growing task (i.e., predicting the idiosyncratic perception of emotional faces of an individual); as expected theoretically, the topmost parameters of our model demonstrated strong correlations with those of the corresponding ordinal item response model: r = 0.928 to 1.00. Our proposed intrinsic approach can used as a complementary framework for post-hoc methods in XAC to coach and support human social interactions.

Carrying abundant side information, knowledge graph (KG) has shown its great potential in enriching the sparsity of collaborative filtering (CF) for recommendation. Although graph neural networks (GNNs) have been successfully employed to learn user preferences from KG and CF signals simultaneously, most models suffer from inferior performance due to their deficient designs, i.e., 1) formulating no distinction between users, items and KG entities, 2) confounding KG signals with CF signals and 3) completely neglecting the effects of edges, which is vital for graph information propagation. In this paper, we propose a quad-channel graph model (X-2ch) to tackle these problems. First, rather than lodging KG entities on graph as nodes, X-2ch distills KG information and embeds them as edge attributes in a bi-directional manner to model the natural user-item interaction process. Second, X-2ch introduces a novel quad-channel learning scheme, including a collaborative user-item update and a CF-KG attentive propagation, to holistically capture the interconnectivity of users and items while preserving their distinct properties. Experiments on two real-world benchmarks show substantial improvement over the state-of-the-art baselines.

Multi-domain recommender systems can solve cold-start problems and can support cross-selling of products and services. We propose a model to address these difficulties by extracting homogeneous and divergent features from domains. Our Domain-to-Domain Translation Model (D2D-TM), which is based on generative adversarial networks (GANs) and variational autoencoders (VAEs), uses the user interaction history. Domain cycle consistency (CC) constrains the inter-domain relations. Results obtained from experimentation demonstrate the great effectiveness of the proposed system when compared to several state-of-the-art systems.

Most works of recommender system seek to provide highly accurate item prediction while having potentially great bias to popular items. Both users and items' providers will suffer if their system has strong preference for monotonous popular items. A better system should consider also item novelty. Previous works of novel recommendation focus mainly on re-ranking a top-N list generated by an accuracy-focused base model. As a result, these frameworks are 2-stage and essentially limited to the base model. In addition, when training the base model, the common BRP loss function treats all pairs in the same manner, consistently suppresses interesting negative items which should have been recommended. In this work, we propose a personalized pairwise novelty weighting for BPR loss function, which covers the limitations of BPR and effectively improves novelty with marginal loss in accuracy. Base model will be guided by the loss weights to learn user preference and to generate novel suggestion list in 1 stage. Comprehensive experiments on 3 public datasets show that our approach effectively promotes novelty with almost no decrease in accuracy.

Balancing accuracy and privacy is an important tradeoff problem for information systems, including recommender systems. To achieve high performance, modern recommender systems tend to use as much information as possible. This trend is borne out by the increasing number of studies of hybrid methods that combine rating and auxiliary information. However, because of privacy concerns, in many cases, service providers can not require users to give their personal information. Therefore, numerous earlier reported methods only use item attributes for auxiliary information. To address these shortcomings, our manuscript provides a method to extract user profiles without using demographic data. Our model learns user and item latent variables through two separate deep neural networks and also learns implicit relations between users and items using the information and their ratings. Experiments verified that our model is a more effective recommender system than state-of- the-art baselines.

Latent Dirichlet allocation (LDA) based topic inference is a data classification method, that is used efficiently for extremely large data sets. However, the processing time is very large due to the serial computational behavior of the Markov Chain Monte Carlo method used for the topic inference. We propose a pipelined hardware architecture and memory allocation scheme to accelerate LDA using parallel processing. The proposed architecture is implemented on a reconfigurable hardware called FPGA (field programmable gate array), using OpenCL design environment. According to the experimental results, we achieved maximum speed-up of 2.38 times, while maintaining the same quality compared to the conventional CPU-based implementation.

In the framework of personalization or micromarketing of services, an effective strategy is to examine customers or items of a specific category. This paper describes an actual service support system using discovery of category-based customer behavior knowledge. The method is realized by modeling a customers' purchase behavior with some purchase situations or conditions using massive point of sales data with a customer ID (ID-POS data) in a department store chain. We automatically generate categories of customers and items based on a purchase patterns identified in ID-POS data using probabilistic latent semantics indexing. We produce a Bayesian network model including the customer and item categories, situations and conditions of purchases, and the properties and demographic information of customers. Based on that network structure, we can systematically identify useful knowledge for use in furthering business intelligence or sustainable services. This method is applicable for marketing support, service modeling, and decision making in various business fields, including retail services.

This paper describes an appropriate category discovery method that simultaneously involves a customer's lifestyle category and item category for the sustainable management of retail services, designated as ``category mining''. Category mining is realized using a large-scale ID-POS data and customer's questionnaire responses with respect to their lifestyle. For the heterogeneous data fusion, we propose a probabilistic double-latent semantic indexing (PdLSI) model that is an extension of PLSI model. In the PdLSI model, customers and items are classified probabilistically into some latent lifestyle categories and latent item category. Then, understanding of relation between the latent categories and various purchased situations is realized using Bayesian network modeling. This method provides useful knowledge based on a large-scale data for efficient customer relationship management and category management, and can be applicable for other service industries.

By understanding the behavior, satisfaction level, or values of the customer, the productivity and level of customer satisfaction of a service industry can be improved. Such customer-based considerations are estimated from questionnaire data in a general manner. The useful estimation of such considerations requires effective methods for modeling the cognitive structures of customers based on such data. However, it is difficult to model the behavior or decision making process of the customer, which involves nonlinear or non-Gaussian variables, using conventional statistical modeling techniques, which assume linear or Gaussian models. The present paper describes a method of constructing a probabilistic model of the cognitive structure of the customer, which clarifies the satisfaction level and decision making process of the customer of a retail service through statistical graphical modeling. The proposed method constructs a probabilistic cognitive structure model by integrating questionnaire data and a Bayesian network, which can handle nonlinear and non-Gaussian variables as conditional probabilities. The model structure can be constructed automatically based on information criteria and can embed some of the experiences of the model designer and/or physical or social rules in advance. The proposed method is applied to an analysis of the requested function from customers regarding the continued use of an item of interest. We obtained useful knowledge for function design and marketing from the model constructed by a simulation and sensitivity analysis. The proposed method can be applied to various services that use a variety of data.

In the kernel method, the appropriate selection or design of the kernel function is important for the construction of a high-performance classifier. The present paper describes a dynamic spectrum classification method using kernel classifiers with the divergence-based kernel and its application to the detection of worn-out banknotes. We introduce the divergence-based kernel that was proposed as a measure between two probability distributions into the dynamic spectrum classification. The present method is applied to the detection of worn-out banknotes by using acoustic signals for the facilitation of identifying counterfeit banknotes. As a result, the classification performance using the divergence-based kernel is shown to have better performance than those using common kernels such as the Gaussian kernel or the polynomial kernel.

Conventional sensors have been designed to minimize noise effects. Any sensor that is designed to detect a certain physical variable is influenced to a certain degree by other physical variables. This suggests that any sensor is potentially tap able of detecting multiple physical variables. In the present study, we consider sensing devices that are easily influenced by several physical variables and make full use of their multi-sensing characteristics through statistical signal processing and machine learning techniques with a wide variety of prior information. The proposed sensor design approach is completely different from the conventional approach with respect to system design and has advantages in terms of cost and system simplification compared to existing approaches. This new idea can be realized by developing a novel multiple-input/single-output sensor that can detect various variables such as pressure, acceleration, temperature and light emission by a single device. The sensor is applied to monitor the symptoms of tire, earthquake and break-in for the purpose of home security. The proposed security system consists of the following three steps: (1) detection of disaster by a probabilistic outlier detection procedure using an auto-regressive model, (2) disaster feature extraction by Kalman filter on a state space model, and (3) disaster classification by multiclass support vector machine.