previous day
all days

View: session overviewtalk overview

08:30-09:15 Session 5: KEYNOTE III

Invited speaker: Prof. João Manuel Portela da Gama (University of Porto, Portugal)

Data Mining for the XXI Century: Real-Time Data Mining

Location: Non Nuoc Room
09:15-09:30COFFEE BREAK
09:30-11:30 Session 6A: AI, Computational Inteligence and Data Analytics I
Location: My Khe 1 Room
Using Individual Accuracy to Create Context for Non-Contextual Multi-Armed Bandit Problems

ABSTRACT. Contextual Multi-Armed Bandit (CMAB) algorithms suppose a linear dependency between the expected reward of an action and its context. If given a complete and relevant context they eventually provide full personalization. However, when context is sparse or missing entirely they fail to solve the problem efficiently. In this article, we give an overview of the Individual Context Enrichment (ICE) method, initially designed to enrich sparse context. Our main contribution is the use of the ICE method with CMAB problems when no context information is available. We propose to measure the levels of performance of a MAB algorithm for its different users and utilize it as context information. This way we transform a hidden context Multi-Armed Bandit problem into an equivalent contextual problem that can be worked on by CMAB algorithms. To be effective, our method requires regular and identifiable users, thus it is particularly interesting in the case of applications having subscribers e.g., recommender systems, clinical trials or mobile health. Our method has been experimented on several datasets and shows better results in terms of accuracy and cumulative regrets than Thompson Sampling, a competitive MAB method.

Aspect Extraction with Bidirectional GRU and CRF

ABSTRACT. Opinion mining or sentiment analysis used to understand the community’s opinions on a particular product. Sentiment analysis involves building the opinion collection and classification system. One of the most crucial tasks of sentiment analysis is the ability to extract aspects or features that opinions expressed on. There are many approaches and techniques used to explore these features from unstructured comments. We proposed a different approach to the above mentioned aspect extraction task in sentiment analysis using a deep learning model combining Bidirectional Gated Recurrent Unit (BiGRU) and Conditional Random Field (CRF). This model is trained on labeled data to extract and classify feature sets in comments Our model uses a BiGRU neural network with word embeddings achieved by training GloVe on the SemEval 2014 dataset. The SemEval 2014 dataset include 7,686 reviews on two domains, Laptop and Restaurant. Experimental results showed that our aspect extraction model in sentiment analysis using BiGRU-CRF achieved significantly better accuracy than the state-of-the-art methods.

A New Location-based Topic Model For Event Attendees Recommendation

ABSTRACT. Event-based social networks (EBSNs) have gained increasing popularity and growing rapidly, which provide services for users to create events and make plan to attend. Developing and creating recommendation models are important and hot issues in EBSNs in recent years, such as event recommendation to users. Although several recommendation models have been proposed, event attendees recommendation models are not fully studied. In this paper, we study the event attendees recommendation problem through empirical experiments. Because of the nature of new events and severe data sparsity in EBSNs, traditional recommender systems work less efficient for event attendees recommendation problem. To solve this problem, we propose a new location-based topic model that bases on scores of users computed from three major factors extracted from previously attended events, namely content, location and time. The proposed model includes two phases. The first phase uses the topic modeling Latent Dirichlet Allocation (LDA) and Jensen Shannon divergence to compute the similarity of events based on their contents. The spatial and temporal factors are also calculated. The scores of previous events with an upcoming event are computed from a combination of these three factors. Previous events with higher scores are selected, then users who are extracted from the selected events are scored by temporal factors of these events in the second phase. Finally we recommend users with top scores to the upcoming event. A series of experiments were conducted on real data collected from Meetup Event and the results have demonstrated the improvement of our model over baseline methods.

Towards a business value framework for Linked Enterprise Data

ABSTRACT. In light of new industry trends and ICT-developments enterprises face an unprecedented amount of data, produced on a daily basis by countless in-house and external ICT-applications and stakeholders. Gaining valuable information and business-insights out of these masses of data has become a true challenge. At the same time, traditional data warehouse and content management systems, in-volving data silos, relational databases and data warehouses are reaching their limits. Linked Data and Semantic Web Technologies offer a new approach to Enterprise Content Management. Based on graph-based data models and seman-tic models to establish logical connections between data sources, enterprises can build Linked Enterprise Data Networks (LEDN), incorporating both internal and/or external data sources and allowing users to retrieve documents, infor-mation and datasets across entire organizations in a fast and easy manner. To outline the potentials of Linked Enterprise Data Networks this paper presents a framework of relevant business values and aspects to be considered by entrepre-neurs, when implementing Linked Data technologies in their business.

A Predictive Control Solution for Contingency Motion Planning for Autonomous Vehicle

ABSTRACT. With a model-based prediction control approach to generate optimal trajectory and avoid collision in many situations that can not perform emergency braking, this paper proposes a solution to optimal path planning and safety assurance to avoid collision for autonomous vehicle by identifying the possible motion trajectories of the vehicles ahead when participating in traffic. The safety factor to avoid collision is performed in each motion phase of the vehicle, with the development of a nonlinear constraint between the brake and the steering angle for the vehicle's contingency motion planning. The efficiency of this solution is evaluated through assumpted simulations with different scenarios, then there are orientations toward applied research on the problem of autonomous vehicles in reality.

Constructing Metrics for Evaluating Multi-Relational Association Rules in the Semantic Web from Metrics for Scoring Association Rules

ABSTRACT. We propose a method to construct asymmetric metrics for evaluating the quality of multi-relational association rules coded in the form of SWRL rules. These metrics are derived from metrics for scoring association rules. We use each constructed metric as a fitness function for evolutionary inductive programming employed to discover hidden knowledge patterns (represented in SWRL) from assertional data of ontological knowledge bases. This new knowledge can be integrated easily within the ontology to enrich it. In addition, we also carry out a search for the best metric to score candidate multi-relational association rules in the evolutionary approach by experiment. We performed experiments on three publicly available ontologies validating the performances of our approach and comparing them with the main state-of-the-art systems.

09:30-11:30 Session 6B: Image, Language and Speech Processing I
Location: My Khe 2 Room
Direction-free Person Identification with Distance Transform and Random Forest

ABSTRACT. The problem of person identification is usually dealt with by explicitly extracting gait-related characteristics from input silhouettes. This paper describes another approach that does not perform such pre-defined and/or hand-crafted features to highlight useful properties. Unlike related works where the task of gait cycle extraction is required, our method directly employs a sliding window to process the temporal factor. In addition, we also introduce a scheme of combining identification results obtained under different camera directions in order to improve the system confidence. Our experiments on the CASIA-A gait dataset demonstrated the potential of the proposed approach and the obtained results were promising compared with related methods.

Attentive Neural Network for Named Entity Recognition in Vietnamese

ABSTRACT. We propose an attentive neural network for the task of named entity recognition in Vietnamese. The proposed attentive neural model makes use of character-based language models and word embeddings to encode words as vector representations. A neural network architecture of encoder, attention, and decoder layers is then utilized to encode knowledge of input sentences and to label entity tags. The experimental results show that the proposed attentive neural network achieves the state-of-the-art results on the benchmark named entity recognition datasets in Vietnamese in comparison to both hand-crafted features based models and neural models.

Object counting based on density using perspective transformation

ABSTRACT. This paper presents a framework to count objects in images with perspective issues, it extends the work of Victor Lempitsky and Andrew Zisserman. The counting framework based on density estimation of the two authors [1] produces excellent results in counting objects that are severely overlapped. However, the framework does not address the perspective issues, which frequently happen in many situations. Therefore, we propose a perspective model and integrate it into the existing density estimation framework to reduce the error caused by perspective distortions. First, the heights of some objects are collected in addition to their locations by marking the top and bottom points of the objects. At that time, a perspective map is learned to estimate the heights of the objects based on their positions in the images. The estimated heights are used to customize the size of the normalized 2D Gaussian kernel and the size of Gaussian window when computing the Scale-Invariant Feature Transform (SIFT) features. The adapted framework is then applied to count the number of motorbikes in Hanoi streets. The results are later compared with the ones produced by the original framework to see the effective of our proposed solution. Our proposed framework performs better in high-density images while maintaining competitive accuracy in low-density images.

Disease Prediction using Synthetic Image Representations of Metagenomic data and Convolutional Neural Networks

ABSTRACT. Information from metagenomic data from human microbiome may improve diagnosis and prognosis for multiple human diseases. However, to achieve a prediction based on bacterial abundance information remains a challenge. Indeed, the number of features being much higher than the number of samples, we face difficulties related to high dimensional data processing, as well as overfitting. In this study, we investigate several convolutional neural network architectures for synthetic images and some experimental techniques to generate and train these synthetic images. We also explore supervised learning for visualizing high dimensional data that use data on genus, species and higher taxonomic level information. In addition, some dimensionality reduction approaches are examined on very very-high-dimensional data such as gene families abundance. We evaluated our approach on six different metagenomic datasets including five types of diseases with more than 1000 samples. Our method displays promising results and can be used in different omics data settings, including integrative ones.

Using Matrix Decomposition and Frequency Transforms to Detect Forgeries in Digital Images

ABSTRACT. Digital images are very popular, but they are quite easy to manipulate. This may have bad consequences, especially when the editing changes important content in the images. Therefore, detecting traces of modifications in digital images is an urgent need. There are various techniques have been proposed, mainly based on finding specific features on the spatial domain or on the frequency domain of the images. In this paper, we study some matrix decomposition methods, which apply for detection of image manipulations. Next, we employ QR matrix factorization method for design a new image forgery detection scheme. We also compare the proposed scheme with some other widely-used schemes. The experimental results on a set of hundreds of images show that the matrix decomposition based schemes accurately detect the copy-paste operations with rather low false positive rates and robust against some attacks. Besides, the QR based scheme is much faster than the others. This characteristic is very useful when working with a large scale image dataset in practice.

A Novel Public Key Robust Watermarking Method for Still Images Based on Intentional Permutation based on DCT and DWT

ABSTRACT. This article proposes a novel asymmetric watermarking method as an alternative to classical direct sequence spread spectrum for still images. In this new technique, the public key is a real sequence chosen according to a normal distribution with mean = 0 and variance = 1, the secret key is created by permuting intentionally elements in subsets of the public key. Watermark is combined linearly to public and secret key, is embedded into embedding location, it is extracted from host image after applying incorporating Discrete Cosine Transform (DCT) with discrete wavelet transform transformations (DWT). The experimental results prove that the proposed method is extremely more robust than the related work.

13:00-13:45 Session 7: KEYNOTE IV

Invited speaker: Prof. Fabien Ferrero (Polytech' Nice Sophia, France)

Fostering Innovation in Education through Smart Campus IoT Contest

Location: Non Nuoc Room
13:45-14:00COFFEE BREAK
14:00-16:20 Session 8A: AI, Computational Inteligence and Data Analytics II
Location: My Khe 1 Room
A Fast Filtering Algorithm for Continuous Constraint Satisfaction Problems

ABSTRACT. We propose an efficient filtering algorithm to solve for continuous constraint satisfaction systems. The algorithm uses an approximate method, named Landmark-based Approximate Inference (LAI) to find landmark-based values of variables in constraint system. The method used the landmark-based values for finding not only local consistency but also global consistency of continuous constraint system. The key idea of this approach is to generate candidates of landmark-based values to be utilized during the constraint propagation to facilitate a quiescent condition. This approach solved separate subintervals in solutions problem to get global consistency, handles complicated problems with a low cost and captures the solutions on various types of constraints. The algorithm is also compared with other methods in real-word applications.

Incremental Clustering for Time Series Data based on an Improved Leader Algorithm

ABSTRACT. Incremental clustering is one of the most important tasks in streaming data mining. As for streaming time series, incremental clustering can be used in some advanced time series data mining tasks such as anomaly detection and motif discovery. In this paper, we propose a novel incremental clustering method for time series. The proposed method is an improved variant of the Leader algorithm, called I-Leader. Leader is a well-known incremental clustering method which is a single-pass distance-based partitional clustering method. However, Leader suffers from some weaknesses such as producing not good clustering results with a large number of clusters with different data distribution in each cluster and requiring a lot of time for clustering. To overcome these weaknesses of the Leader algorithm, in our I-Leader algorithm, we employ a data summarization technique and use this data summarization to maintain the high quality for the clusters. Besides, we use an incremental formula to speed up the calculation of cluster centers. We have conducted some experiments on several datasets to compare the performance of the I-Leader algorithm with that of the original Leader one in time series clustering. At the same time, we compare I-Leader with k-Means for the popularity of k-Means. The experimental results show that our I-Leader algorithm runs faster than both Leader and k-Means algorithms. It is also more effective than both of those algorithms with better clustering results.

A Scalable Network-on-Chip based Neural Network Implementation on FPGAs

ABSTRACT. This paper presents a feed-forward neural network implementation using a network on chip (NoC) on a field-programmable gate array (FPGA). The design is intended for large neural networks with hundreds of neurons in each layer. It is one of the first publications that fully realizes a large neural network using NoC-based communication. A dedicated NoC is implemented to handle many interconnections among neurons. The design uses multiple processing elements (PE) to exploit the parallelism available in neural network layers. Different PE configurations are implemented and evaluated for area, power consumption and performance. Implementation and experimental results for a handwritten digit recognition system show that using a large neural network with NoC interconnect allows system scalability with scalable performance. The design achieves the maximum performance of 1339 FPOPs/cycle on a Xilinx Virtex-7 XC7VX485T FPGA. Its latency is 18-times faster than an Intel i7-3.4 GHz CPU running Matlab for this application.

An Effective Implementation of Motif-based Time Series Classification

ABSTRACT. Time series classification is one of the crucial tasks in time series data mining. Due to the unique characteristics of time series, most classic classification algorithms in data mining do not work well for time series. So far, empirical evidence has shown that the nearest neighbor method is very effective to classify directly time series data without any feature extraction step. Recently, there have been two research works that proposed a framework for time series classification which uses motif information to convert time series to feature vectors. This framework facilitates the use of classical classification algorithms for time series classification. In this paper, we present an effective implementation of the framework. In our method we exploit the motif discovery algorithm which combines EP-C, a segmentation-based motif discovery method and MK, an exact motif discovery method. Furthermore, we devise a segmentation-based method that helps in transforming efficiently a set of time series to a set of feature-vectors. Experimental results on benchmark datasets reveal that our proposed implementation method for motif-based time series classification improves the accuracy of not only 1-nearest neighbor algorithm but also SVMs and ANNs for time series classification.

Enriching user representation in Neural Matrix Factorization

ABSTRACT. Personalized recommendation is a crucial problem in the era of information overload. One of the most prominent methods addressing this problem is collaborative filtering (CF). Recently, NeuMF \cite{he2017neural} has employed Deep Neural Network to facilitate traditional CF methods. Although NeuMF achieved the state-of-the-art performance, it is expected that predictive accuracy can be improved by enriching user representation. In this paper, we propose a neural network incorporating latent features and personalized information of users, namely Feature based Neural Matrix Factorization (FeaNMF). In FeaNMF, personalized information can be considered as auxiliary features which are generated by associating past behavior of users (e.g. purchase history, browsing activity, watching habits) with item categories. By explicitly modeling personalized information in the metric of item categories, the major advantages of FeaNMF are: (1) it enriches user representation and thus enhances the predictive capability, (2) it inherits the advantages of NeuMF to model the latent features in a more comprehensive approach, rather than applying a simple linear function as existing methods (e.g. Matrix Factorization), and (3) it is easy to extend FeaNMF for employing a large amount of external sources to construct user preferences. Experiments show that our model significantly improves predictive accuracy compared to the three previous approaches including NeuMF.

An adaptive hash-based text deduplication for ADS-B data-dependent trajectory clustering problem

ABSTRACT. The Automatic Dependent Surveillance-Broadcast (ADS-B) protocol is equipped in aircraft as an alternative to secondary radar. This emerging technology produces such a prospective type of data to effectively broadcast the aircraft's status (location, velocity, etc.,) in a specific area, which is very useful in air traffic management (ATM). However, there is still a limited number of advanced studies from machine learning/data mining perspectives relying on this kind of data in ATM research. On the other hand, Locality Sensitive Hashing (LSH) is a data mining technique often used to find similar items in the data with high-dimension properties. It is thus relatively suitable for handling with trajectories data to group similar flight paths. From these factors, we reveal in this paper an adaptive LSH-based algorithm, used in near-duplicated documents detection, for the problem of clustering the nearest trajectories by representing the trajectories as a bag-of-words used popularly in text mining. To illustrate our proposed method, an experiment is designed and carried out in thirty successive days, employing the raw ADS-B data collected from FlightAware for the case of Changi International Airport, Singapore. The evaluation based on Silhouette score shows quite good results of measuring the clustering performance.

A Simplified Framework for Air Route Clustering Based on ADS-B Data

ABSTRACT. The volume of flight traffic gets increasing over the time, which makes the strategic traffic flow management become one of the challenging problems since it requires a lot of computational resources to model entire traffic data. On the other hand, Automatic Dependent Surveillance - Broadcast (ADS-B) technology has been considered as a promising data technology to provide both flight crews and ground control staff the necessary information safely and efficiently about the position and velocity of the airplanes in a specific area. In the attempt to tackle this problem, we presented in this paper a simplified framework that can support to detect the typical air routes between airports based on ADS-B data. Specifically, the flight traffic will be classified into major groups based on similarity measures, which helps to reduce the number of flight paths between airports. As a matter of fact, our framework can be taken into account to reduce practically the computational cost for air flow optimization and evaluate the operational performance. Finally, in order to illustrate the potential applications of our proposed framework, an experiment was performed using ADS-B traffic flight data of three different pairs of airports. The detected typical routes between each couple of airports show quite good results by virtue of combining two indices for measuring the clustering performance and incorporating human judgment into the visual inspection.

14:00-15:40 Session 8B: Image, Language and Speech Processing II
Location: My Khe 2 Room
An End-to-End model for Vietnamese speech recognition

ABSTRACT. This paper presents an approach of End-to-End model based on Long Short-Term Memory (LSTM) and Time Delay Deep Neural Network (TDNN) models for Vietnamese speech recognition. Two Vietnamese End-to-End architectures using Connectionist Temporal Classification (CTC) as the loss function are proposed. The paper also presents the method to construct phonesets based on Vietnamese characters or tonemes to produce the label sequence for any given transcription when applying CTC model. The experimental results showed that CTC based End-To-End models are competitive to traditional models with only 5% of WER worse for Vietnamese speech recognition (SR), but the advantage is no requirement of forced alignment for training acoustic models. In addition, it is similar to proposed studies using traditional models, the tone information and the toneme set are solutions to optimize the performance for Vietnamese SR.

Long Sentence Preprocessing in Neural Machine Translation

ABSTRACT. Machine translation based on neural networks has been shown to produce superior results, compared with other approaches. To build an efficient neural machine translation (NMT) system, it is essential to have an accurate and massive bilingual corpus for training, and ensure the continuous improvement of the methods and techniques used in the translation system. Despite multiple advantages, one challenging issue for current neural network translation system is long sentence processing (Koehn and Knowles,2017). In this paper, we propose a method to extract bilingual phrases to build a phrase-aligned bilingual corpus, and the implementation of a long sentence preprocessing technique to be used in the neural machine translation model. Experimental training of the neural machine translation system to translate Vietnamese into English using our proposed technique shows an improvement in BLEU scores.

Enhancing the quality of inserted 3D object into a video

ABSTRACT. Inserting 3d object into a real video is one of major task in an augmented reality system. To enhance the real quality, inserted object usually added with the environment effects. In this paper, we proposed a technique to enhance the quality of inserted object by adding its shadow. Experiments show the effectiveness of our proposal method.

Deep Laplacian Pyramid Network for Text Images Super-Resolution

ABSTRACT. Convolutional neural networks have recently demonstrated interesting results for single image super-resolution. However, these networks were trained to deal with super-resolution problem on natural images. In this paper, we adapt a deep network, which was proposed for natural images super-resolution, to single text image super-resolution. To evaluate the network, we present our database for single text image super-resolution. Moreover, we propose to combine Gradient Difference Loss (GDL) with L1/L2 loss to enhance edges in super-resolution image. Quantitative and qualitative evaluations on our dataset show that adding the GDL improves the super-resolution results.

Neural sequence labeling for VietnamesePOS Tagging and NER

ABSTRACT. This paper presents a neural architecture for Viet-namese sequence labeling tasks including part-of-speech (POS)tagging and named entity recognition (NER). We applied themodel described in [1] that is a combination of bidirectionalLong-Short Term Memory and Conditional Random Fields,which rely on two sources of information about words: character-based word representations learned from the supervised corpusand pre-trained word embeddings learned from other unanno-tated corpora. Experiments on benchmark datasets show thatthis work achieves state-of-the-art performances on both tasks -93.52% accuracy for POS tagging and 94.88% F1 for NER.Our sourcecode is available at here.

16:20-16:30 Session 9: CLOSING SESSION
  • Program Committee Chairs
  • All the participants
Location: Non Nuoc Room