Journal of Information Processing
Online ISSN : 1882-6652
ISSN-L : 1882-6652
Current issue
Showing 1-21 articles out of 21 articles from the selected issue
  • Shunsuke Inenaga
    Type: 60th Anniversary Best Paper
    Subject area: 60th Anniversary Best Paper
    2021 Volume 29 Pages 1-13
    Published: 2021
    Released: January 15, 2021
    JOURNALS FREE ACCESS

    A labeled tree (or a trie) is a natural generalization of a string which can also be seen as a compact representation of a set of strings. This paper considers the labeled tree indexing problem, and provides a number of new results on space bound analysis and on algorithms for efficient construction and pattern matching queries. Kosaraju [FOCS 1989] was the first to consider the labeled tree indexing problem and he proposed the suffix tree for a backward trie, where the strings in the trie are read in the leaf-to-root direction. In contrast to a backward trie, we call an ordinary trie as a forward trie. Despite a few follow-up works after Kosaraju's paper, indexing forward/backward tries is not well understood yet. In this paper, we show a full perspective on the sizes of indexing structures such as suffix trees, DAWGs, CDAWGs, suffix arrays, affix trees, affix arrays for forward and backward tries. Some of them take O(n) space in the size n of the input trie, while the others can occupy O(n2) space in the worst case. In particular, we show that the size of the DAWG for a forward trie with n nodes is Ω(σn), where σ is the number of distinct characters in the trie. This becomes Ω(n2) for an alphabet of size σ =Θ(n). Still, we show that there is a compact O(n)-space implicit representation of the DAWG for a forward trie, whose space requirement is independent of the alphabet size. This compact representation allows for simulating each DAWG edge traversal in O(log σ) time, and can be constructed in O(n) time and space over any integer alphabet of size O(n). In addition, this readily extends to the first indexing structure that permits bidirectional pattern searches over a trie within linear space in the input trie size. We also discuss the size of the DAWG built on a labeled DAG or on an acyclic DFA, and present a quadratic lower bound for its size.

    Download PDF (764K)
  • Akifumi Inoue
    Type: Special Issue of Collaboration technologies and network services toward the sustainable society
    2021 Volume 29 Pages 14-15
    Published: 2021
    Released: January 15, 2021
    JOURNALS FREE ACCESS
    Download PDF (38K)
  • Abu Nowshed Chy, Umme Aymun Siddiqua, Masaki Aono
    Type: Special Issue of Collaboration technologies and network services toward the sustainable society
    Subject area: Information Retrieval
    2021 Volume 29 Pages 16-29
    Published: 2021
    Released: January 15, 2021
    JOURNALS FREE ACCESS

    During emergencies and disaster situations, microblogging sites, especially twitter, can be used as a source of providing situational information needs. Monitoring and identifying informative tweets from tweet streams provide enormous opportunities for public safety personnel in coordinating aid operations as well as conducting the post-incident analysis. However, the brevity of tweets and noisy tweet contents makes it challenging to extract the situational information effectively and identify the tweets based on different information types. In this paper, we propose a neural network model with a naive rule-based classifier for actionable informative tweets classification. In our proposed neural architecture, we exploit the transfer learning features from a pre-trained sentence embeddings model along with a rich set of hand-crafted features to train a multilayer perceptron (MLP) network. In addition, we employ the state-of-the-art LSTM variants, nested LSTMs (NLSTMs) to capture the long-term dependency effectively. On top of nested LSTMs, we perform the convolution using multiple kernels (CMK) to obtain the higher-level representation of tweets. Experiments on the 2018 TREC incident streams (TREC-IS) dataset show that our proposed neural model learns the contextual information effectively and achieves the overall best result compared to the state-of-the-art methods.

    Download PDF (838K)
  • Ryo Ishii, Ryuichiro Higashinaka, Koh Mitsuda, Taichi Katayama, Masahi ...
    Type: Special Issue of Collaboration technologies and network services toward the sustainable society
    Subject area: User Interfaces and Interactive Systems
    2021 Volume 29 Pages 30-44
    Published: 2021
    Released: January 15, 2021
    JOURNALS FREE ACCESS

    Starting from their early years, many persons dream of being able to chat with their favorite anime characters. To make such a dream possible, we propose an efficient method for constructing a system that enables users to text chat with existing anime characters. We tackled two research problems to generate verbal and nonverbal behaviors for a text-chat agent system utilizing an existing character. A major issue in creating verbal behavior is generating utterance text that reflects the personality of existing characters in response to any user questions. To cope with this problem we propose use of role play-based question-answering to efficiently collect high-quality paired data of user questions and system answers reflecting the personality of an anime character. We also propose a new utterance generation method that uses a neural translation model with the collected data. Rich and natural expressions of nonverbal behavior greatly enhance the appeal of agent systems. However, not all existing anime characters move as naturally and as diversely as humans. Therefore, we propose a method that can automatically generate whole-body motion from spoken text in order to give the anime characters natural, human-like movements. In addition to these movements, we try to add a small amount of characteristic movement on a rule basis to reflect personality. We created a text-dialogue agent system of a popular existing anime character using our proposed generation methods. As a result of a subjective evaluation of the implemented system, our methods for generating verbal and nonverbal behavior improved the impression of the agent's responsiveness and reflected the personality of the character. Since generating characteristic motions with a small amount of characteristic movement on the basis of heuristic rules was not effective, our proposed motion generation method which can generate the average motion of many people, is useful for generating motion for existing anime characters. Therefore, our proposed methods for generating verbal and nonverbal behaviors and the system-construction method are likely to prove a powerful tool for achieving text-dialogue agent systems for existing characters.

    Download PDF (5475K)
  • Ken Ohta
    Type: Special Issue of Mobile and Intelligent Transportation Systems Evolving Society in the 5G Era
    2021 Volume 29 Pages 45
    Published: 2021
    Released: January 15, 2021
    JOURNALS FREE ACCESS
    Download PDF (35K)
  • Akira Uchiyama, Shunsuke Saruwatari, Takuya Maekawa, Kazuya Ohara, Ter ...
    Type: Special Issue of Mobile and Intelligent Transportation Systems Evolving Society in the 5G Era
    Subject area: Invited Papers
    2021 Volume 29 Pages 46-57
    Published: 2021
    Released: January 15, 2021
    JOURNALS FREE ACCESS

    Context recognition is a topic that has garnered considerable interest in the ubiquitous and pervasive computing research community. A wide variety of Internet-of-things devices with micro-electromechanical system (MEMS) sensors are used to obtain sensor data (e.g., acceleration, vibration, and sound) related to target contexts. However, devices for context recognition also have limitations such as deployment cost, battery maintenance cost, and the requirement for wearing/carrying the devices. To solve this problem, wireless sensing has attracted the attention of many researchers because it enables device-free and/or maintenance-free context recognition. In this study, we will comprehensively review studies on context recognition by wireless sensing, focusing on WiFi channel state information (CSI), radio-frequency identification (RFID), and backscatter. We will also discuss the design choices of wireless sensing with their pros and cons through a review of the state-of-the-art.

    Download PDF (786K)
  • Kenta Urano, Kei Hiroi, Takuro Yonezawa, Nobuo Kawaguchi
    Type: Special Issue of Mobile and Intelligent Transportation Systems Evolving Society in the 5G Era
    Subject area: Mobile Computing
    2021 Volume 29 Pages 58-69
    Published: 2021
    Released: January 15, 2021
    JOURNALS FREE ACCESS

    This paper proposes an indoor localization method for Bluetooth Low Energy (BLE) devices using an end-to-end LSTM neural network. We focus on a large-scale indoor space where there is a tough environment for wireless indoor localization due to signal instability. Our proposed method adopts end-to-end localization, which means input is a time-series of signal strength and output is the estimated location at the latest time in the input. The neural network in our proposed method consists of fully-connected and LSTM layers. We use a custom-made loss function with 3 error components: MSE, the direction of travel, and the leap of the estimated location. Considering the difficulty of data collection in a short preparation term, the data generated by a simple signal simulation is used in the training phase, before training with a small amount of real data. As a result, the estimation accuracy achieves an average of 1.92m, using the data collected in GEXPO exhibition in Miraikan, Tokyo. This paper also evaluates the estimation accuracy assuming the troubles in a real operation.

    Download PDF (5190K)
  • Ryosuke Hasegawa, Akira Uchiyama, Takuya Magome, Juri Tatsumi, Teruo H ...
    Type: Special Issue of Mobile and Intelligent Transportation Systems Evolving Society in the 5G Era
    Subject area: Mobile Computing
    2021 Volume 29 Pages 70-80
    Published: 2021
    Released: January 15, 2021
    JOURNALS FREE ACCESS

    In wheelchair basketball (WB), players are constantly trying to improve their wheelchair maneuvering techniques since these are the most basic and important actions in all situations. However, assessing maneuvering quality is difficult due to the lack of quantitative metrics. In this paper, we propose two classification methods for maneuvering actions and turns by focusing on the specific wheelchair movement. For this purpose, inertial sensors are fixed to the left and right wheels of the wheelchair. In maneuver classification, the occurrence of maneuvers is detected using the angular velocity. Major maneuver activities in WB are classified into 2 types: PUSH and PULL. First, our method segments candidates of maneuver periods by the local maximum/minimum of the angular velocity since the rotation of the wheel generated by maneuvering that leads to sharp changes in the angular velocity. We then classify maneuvering actions based on thresholds. As for the turn classification, we first detect turns by calculating the amount of wheelchair rotation from the angular velocities of both wheels. We then classify the detected turns into PIVOT and TURN by using thresholds based on the typical movement of both wheels during each turn. To evaluate the performance of the proposed maneuver classification method, we collected real data from 6 players. From the result, we confirmed our method achieves an average recall and precision of 91.9% and 84.6% for maneuver classification, respectively. The results also show that our turn classification achieves an average recall and precision of 99.7% and 99.7%, respectively. Furthermore, we confirmed the effectiveness of the classification results for the assessment of maneuver quality.

    Download PDF (4468K)
  • Yuta Ishizaki, Yurie Koyama, Toshinori Takayama, Nozomu Togawa
    Type: Special Issue of Mobile and Intelligent Transportation Systems Evolving Society in the 5G Era
    Subject area: ITS
    2021 Volume 29 Pages 81-92
    Published: 2021
    Released: January 15, 2021
    JOURNALS FREE ACCESS

    As smartphones and tablets are widely spread and used, route recommendation and guidance services have become commonplace. Conventional services in route recommendation and guidance try to give best routes in terms of route length, time required, and train/bus fares, whereas even different users are given the same route when inputting the same parameters. However, each user has various preferences from the aspect of safety and comfort. It is strongly desirable to reflect the user's preferences in route recommendation and recommend the most preferable route to every user. Since user's preferences are extremely vague and complicated, how to evaluate them in route recommendation is one of the key problems there. In this paper, we propose a route recommendation method, called P-UCT method, considering individual user's preferences utilizing Monte-Carlo tree search. In the proposed method, we firstly extract route features based on the route recommendation history of every user and construct a route evaluator based on Support Vector Machine (SVM). After that, the method generates a random route from a start point to an end point by Monte-Carlo tree search. The route evaluator determines how well every generated route matches the user's preferences. By repeating the evaluation, the method obtains the route, which must be closest to the user's preferences. Experimental results demonstrate that the proposed method outperforms the existing method from the viewpoint of the average evaluation scores. They also demonstrate that the proposed method provides the recommended route reflecting the user's individual preferences even if it learns the recommended route history of areas in different situations.

    Download PDF (5352K)
  • Naoki Fukushi, Daiki Chiba, Mitsuaki Akiyama, Masato Uchida
    Type: Regular Papers
    Subject area: Network Security
    2021 Volume 29 Pages 93-102
    Published: 2021
    Released: January 15, 2021
    JOURNALS FREE ACCESS

    Cloud services are maliciously used as an infrastructure for cyber-attacks. In a cloud service, the assigned Internet Protocol (IP) address for a server is owned by the cloud service provider. When the server is shut down, the assigned IP address is freed for reuse and assigned to another server in the same cloud service. Cyber-attackers abusing cloud services in this way therefore pose a serious risk since legitimate service providers, developers, and end users may be mistakenly blacklisted which lowers the image and hurts the reputation of the service. In this study, we conducted a large-scale measurement of cloud service abuse by using blacklisted IP addresses. Our analysis of four cloud services over 154 days using 39 blacklists revealed that a total of 61,060 IP addresses from these cloud service providers were blacklisted, approximately 14,000 IP addresses continue to be blacklisted, and approximately 5% are replaced daily. Moreover, our study revealed trends in attacks that abuse cloud services with respect to attack type, region, duration, and anti-abuse countermeasures. Finally, we discuss recommendations for cloud service users, cloud service providers, and blacklist providers.

    Download PDF (991K)
  • Hideaki Goto
    Type: Regular Papers
    Subject area: Network Security
    2021 Volume 29 Pages 103-112
    Published: 2021
    Released: January 15, 2021
    JOURNALS FREE ACCESS

    Wireless LAN Roaming enables users to be authenticated and authorized for access to networks at various places they visit. The largest roaming federation today is probably eduroam developed for research and education institutions. There have been some projects for developing roaming federations for public wireless LAN services as well. Since the introduction of the Next Generation Hotspot concept, the need to interconnect these roaming federations has emerged. However, some technical problems exist and these need to be addressed simultaneously. In this paper, we analyze existing roaming systems and develop an inter-federation roaming architecture with the aim of realizing a worldwide Wireless LAN roaming system that would accommodate eduroam. To realize realm-based routing of authentication requests across different federations and to achieve system scalability, the roaming architecture is designed to use regional hubs instead of a single exchange hub. We also aim to develop an actual roaming platform for secure public wireless LAN services based on the developed architecture. The developed platform was implemented and tested in the City Wi-Fi Roaming trials conducted by the Wireless Broadband Alliance.

    Download PDF (1005K)
  • Takuo Suganuma
    Type: Special Issue of Network Services and Distributed Processing
    2021 Volume 29 Pages 113-114
    Published: 2021
    Released: February 15, 2021
    JOURNALS FREE ACCESS
    Download PDF (38K)
  • Nguyen Minh Tri, Masahiro Shibata, Masato Tsuru
    Type: Special Issue of Network Services and Distributed Processing
    Subject area: Distributed System Operation and Management
    2021 Volume 29 Pages 115-123
    Published: 2021
    Released: February 15, 2021
    JOURNALS FREE ACCESS

    The prevalence of cloud computing and contents delivery networking has led to demand for OpenFlow-based centrally-managed networks with dynamic and flexible traffic engineering. Maintaining a high level of network service quality requires detecting and locating high-loss links. Therefore, in this paper, a measurement framework is proposed to promptly locate all high-loss links with a minimized load on both data-plane and control-plane incurred by the measurement, which assumes only standard OpenFlow functions. It combines an active measurement by probing multicast packets along a designed route and a passive measurement by collecting flow-stats of the probing flow at selected switch ports in an appropriate sequential order to access switches. In particular, by designing the measurement route based on the backbone-and-branch tree (BBT) route scheme, the measurement accuracy and the measurement overhead (the number of accesses to switches until locating all high-loss links) can be balanced. The numerical simulation demonstrates the effectiveness of our proposal.

    Download PDF (1175K)
  • Arnan Maipradit, Tomoya Kawakami, Ying Liu, Juntao Gao, Minuro Ito
    Type: Special Issue of Network Services and Distributed Processing
    Subject area: Distributed System Operation and Management
    2021 Volume 29 Pages 124-131
    Published: 2021
    Released: February 15, 2021
    JOURNALS FREE ACCESS

    Nowadays traffic congestion has increasingly been a significant problem, which results in a longer travel time and aggravates air pollution. Available works showed that back-pressure based traffic control algorithms can effectively reduce traffic congestion. However, those works control traffic based on either inaccurate traffic information or local traffic information, which causes inefficient traffic scheduling. In this paper, we propose an adaptive traffic control algorithm based on back-pressure and Q-learning, which can efficiently reduce congestion. Our algorithm controls traffic based on accurate real-time traffic information and global traffic information learned by Q-learning. As verified by simulation, our algorithm significantly decreases average vehicle traveling time from 17% to 38% when compared with a state-of-the-art algorithm under tested scenarios.

    Download PDF (959K)
  • Ushio Yamamoto
    Type: Special Issue of Network Services and Distributed Processing
    Subject area: Wireless/Mobile Networks
    2021 Volume 29 Pages 132-139
    Published: 2021
    Released: February 15, 2021
    JOURNALS FREE ACCESS

    Many multichannel MAC protocols for wireless ad-hoc networks have been proposed to enhance the network performance by avoiding contention among adjacent nodes which want to transmit data frames. However, multichannel MAC protocols need channel switching behavior with some time period, therefore frequent channel switching may deteriorate the network performance. In this paper, we propose a new multi-channel MAC protocol called RPNT-MMAC (Receiver-Prioritized Next Transmission in Multichannel MAC) protocol. In RPNT-MMAC protocol, the transmitter and the receiver nodes negotiate the data channel for data frame transmission by exchanging RTS/CTS frame, and if the receiver node already has another data frame to be forwarded to its neighbor node, then it can acquire the right of the next transmission on the same data channel of ongoing transmission procedure and notifies the next transmission's receiver before switching channel. When the node received the notification for the next transmission, it switches to the same data channel at the end of first data frame transmission and receives the data frame. Our proposed method can decrease the frequency of switching channels and improve the network performance such as the packet arrival rate and the end-to-end delay. Moreover, the fairness among data flows can be improved. Network simulation results showed a better performance of our proposed method than the traditional methods.

    Download PDF (486K)
  • Eiji Nii, Shoma Nishigami, Takamasa Kitanouma, Hiroyuki Yomo, Yasuhisa ...
    Type: Special Issue of Network Services and Distributed Processing
    Subject area: Wireless/Mobile Networks
    2021 Volume 29 Pages 140-148
    Published: 2021
    Released: February 15, 2021
    JOURNALS FREE ACCESS

    Autonomous mobile devices, such as robots and unmanned aerial vehicles, as alternatives to humans, are expected to be applied to searching for and manipulating a variety of emergent events of which the location and number of occurrences are unknown. When an autonomous mobile device searches for an event, it needs to sense a physical signal emitted by an event, such as radio waves, smell or temperature. After a device finds an event, it must manipulate the event. We previously proposed Mobile Sensing Cluster (MSC), which applies swarm intelligence to multiple autonomous mobile devices to quickly search for and manipulate multiple events using dynamically formed multiple swarms of mobile devices. However, in an environment that the physical signal emitted by an event and sensed by a device includes some random noises, the behavior of swarms in MSC becomes unstable. As a result, MSC requires a long time to search and manipulate. In this paper, we propose a dynamic swarm spatial scaling MSC for improving the tolerance of MSC against such random noises, and show its effectiveness.

    Download PDF (915K)
  • Miya Fukumoto, Takuya Yoshihiro
    Type: Special Issue of Network Services and Distributed Processing
    Subject area: Mobile Computing
    2021 Volume 29 Pages 149-156
    Published: 2021
    Released: February 15, 2021
    JOURNALS FREE ACCESS

    For IoT applications LPWA is a useful communication choice that enables us to connect tiny devices spread over the land to the Internet. Since many low-price IoT devices usually need to work with limited power budget, this kind of low-power long-range communication technique is a strong tool to populate IoT deployment. Since LPWA devices are less functional, localization of devices are addressed as one of the important practical problems. UNB (Ultra Narrow Band)-based LPWA networks such as Sigfox are one of the major LPWA services for IoT applications, which have a long communication range more than 10km. However, due to the long-range communications and the property of UNB-based modulation, it is not possible to use state-of-the-art localization techniques with high-accuracy; UNB-based LPWA should use simple methods based on RSSI (Radio Signal Strength Indicator) that involves large position estimation errors. In this paper, we propose a method to improve accuracy of device localization in UNB-based LPWA networks by utilizing portable Access Points (APs). By introducing a distance-based weighting technique, we improve the localization accuracy in combination with stationary and portable APs. We demonstrated that the portable AP and the new weighting technique effectively works in UNB-based LPWA networks.

    Download PDF (3928K)
  • Hiroko Nagashima, Yuka Kato
    Type: Special Issue of Network Services and Distributed Processing
    Subject area: Mobile Computing
    2021 Volume 29 Pages 157-165
    Published: 2021
    Released: February 15, 2021
    JOURNALS FREE ACCESS

    The quantity of data available for analysis, including data collected by sensors and wearable devices, has been increasing hugely. However, to obtain accurate analysis results, data pre-processing such as outlier detection, handling of missing data, and preparing data recorded by different measuring instruments in different units, is essential. Considering that the pre-processing task consumes 80% of analyst resources, we previously proposed a method to address this problem. The method integrates machine learning based on Bayesian inference with human knowledge by using programming by example approach. However, in situations in which the process of generating the model and the process of updating the model are executed at different sites, the previous method is problematic in two ways: 1) all sites have to use the same features defined when the model is generated, and 2) a helpful process to generate new training data from features without using inference data when updating the model, is not available. This prompted us to propose APREP-S, which has flexible feature processes and a process for updating the model using a clustering method. We evaluate the accuracy of the imputation and the similarity of the trends by comparing APREP-S with the original data and other existing methods. The results show that APREP-S can return the most optimal methods with both accuracy and similarity.

    Download PDF (1935K)
  • Yuki Matsuyama, Shuichi Miyazaki
    Type: Regular Papers
    Subject area: Algorithm Theory
    2021 Volume 29 Pages 166-173
    Published: 2021
    Released: February 15, 2021
    JOURNALS FREE ACCESS

    In a variant of the stable marriage problem where ties and incomplete lists are allowed, finding a stable matching of maximum cardinality is known to be NP-hard. There are a lot of experimental studies for evaluating the performance of approximation algorithms or heuristics, using randomly generated or artificial instances. One of standard evaluation methods is to compare an algorithm's solution with an optimal solution, but finding an optimal solution itself is already hard. In this paper, we investigate the possibility of generating instances with known optimal solutions. We propose three instance generators based on a known random generation algorithm, but unfortunately show that none of them meet our requirements, implying a difficulty of instance generation in this approach.

    Download PDF (1295K)
  • Masaki Umeda, Atusi Maeda
    Type: Regular Papers
    Subject area: Special Section on Programming
    2021 Volume 29 Pages 174-187
    Published: 2021
    Released: February 15, 2021
    JOURNALS FREE ACCESS

    One of the common problems with the recursive descent parsing method is that when parsing with a left-recursive grammar, the parsing does not terminate because the same parsing function is recursively invoked indefinitely without consuming the input string. Packrat parsing, which is a variant of recursive descent parsing method that handles grammars described in parsing expression grammars (PEGs) by backtracking, is also affected by the above problem. Although naive backtracking parsers may exhibit an exponential execution time, packrat parsers achieve a linear time complexity (for grammars that are not left-recursive) by memoizing the result of each call to the parsing functions. Some methods have been proposed to solve the problem of left recursion in packrat parsers. In these methods, memoization tables in packrat parsers are modified to limit the depth of the recursive calls. By calling the same parsing function repeatedly while increasing the limit, the parsed range in the input string is expanded gradually. These methods have problems in that multiple occurences of left-recursive calls at the same input position cannot be handled correctly, and some of the grammars that does not include left recursion cannot be handled. In this research, we propose and implement a new packrat parser to address these problems. This packrat parser can handle multiple occurences of left-recursive calls at the same position in the input by giving priority to the most recently used rule when gradually increasing the parsed range of the recursion. In the evaluation of the proposed method, in addition to the grammars including left recursion manageable by the methods proposed in existing studies, we confirmed that our approach supports the grammars that cannot be handled by those existing methods.

    Download PDF (1730K)
  • Shohei Matsugu, Hiroaki Shiokawa, Hiroyuki Kitagawa
    Type: Regular Papers
    Subject area: Special Section on Databases
    2021 Volume 29 Pages 188-196
    Published: 2021
    Released: February 15, 2021
    JOURNALS FREE ACCESS

    Searching communities on attributed graphs has attracted much attention in recent years. The community search algorithm is currently an essential graph data management tool to find a community suited to a user-specified query node. Although community search algorithms are useful in various web-based applications and services, they have trouble handling attributed graphs due to the strict topological constraints of traditional algorithms. In this paper, we propose an accurate community search algorithm for attributed graphs. To relax the topological constraints, we proposed a new model of the community. And we defined the problem of finding them in an attributed graph class called the Flexible Attributed Truss Community (F-ATC). The F-ATC problem has the advantage of being applicable in many situations because it can explore diverse communities. Consequently, the community search accuracy is enhanced compared to traditional community search algorithms. Additionally, we present a novel heuristic algorithm to solve the F-ATC problem. This effective algorithm detects more accurate communities from attributed graphs than the traditional algorithms. For further optimization, we pre-processed the query response to make it faster. Finally, we conducted extensive experiments with real-world attributed graphs to demonstrate that our approach outperforms state-of-the-art methods.

    Download PDF (970K)
feedback
Top