Information Extraction (IE) focuses on transforming unstructured data into structured knowledge, of which Named Entity Recognition (NER) is a fundamental component. In the realm of Information Retrieval (IR), effectively recognizing entities can substantially enhance the precision of search and recommendation systems. Existing methods frame NER as a sequence labeling task, which requires extra data and, therefore may be limited in terms of sustainability. One promising solution is to employ a Machine Reading Comprehension (MRC) approach for NER tasks, thereby eliminating the dependence on additional data. This process encounters key challenges, including: (1) Unconventional predictions; (2) Inefficient multi-stream processing; (3) Absence of a proficient reasoning strategy. To this end, we present the Single-Stream Reasoner (SSR), a solution utilizing a reasoning strategy and standardized inputs. This yields a type-agnostic solution for both flat and nested NER tasks, without the need for additional data. On ten NER benchmarks, SSR achieved state-of-the-art results, highlighting its robustness. Furthermore, we illustrated its efficiency through convergence, inference speed, and low-resource scenario performance comparisons. Our architecture displays adaptability and can effortlessly merge with various foundational models and reasoning strategies, fostering advancements in both the IR and IE fields.

Dark patterns are deceptive user interface designs for online services that make users behave in unintended ways. Dark patterns, such as privacy invasion, financial loss, and emotional distress, can harm users. These issues have been the subject of considerable debate in recent years. In this paper, we study interpretable dark pattern auto-detection, that is, why a particular user interface is detected as having dark patterns. First, we trained a model using transformer-based pre-trained language models, BERT, on a text-based dataset for the automatic detection of dark patterns in e-commerce. Then, we applied post-hoc explanation techniques, including local interpretable model agnostic explanation (LIME) and Shapley additive explanations (SHAP), to the trained model, which revealed which terms influence each prediction as a dark pattern. In addition, we extracted and analyzed terms that affected the dark patterns. Our findings may prevent users from being manipulated by dark patterns, and aid in the construction of more equitable internet services. Our code is available at https://github.com/yamanalab/whydarkpattern.

This study addresses the privacy problems of data provided by multiple data owners for range query processing on the cloud. Although existing methods preserve data privacy against the cloud or data analysts who receive query responses, protecting data privacy from both remains a challenge. Combining differential privacy (DP) and homomorphic encryption (HE) to construct differentially private outputs over encrypted raw data is a promising way to avoid data privacy leakage to the cloud with encryption while protecting data privacy from data analysts with DP. Although DP adopts several partitioning algorithms to achieve small noise, partitioning cannot be executed once the data is encrypted. In this paper, we propose a new HE-friendly privacy-preserving partitioning algorithm satisfying DP. Although HE enables operations over encrypted data, the execution time of such primitive arithmetic operations is approximately 109times slower than without encryption. Therefore, it is mandatory to reduce the calculation complexity. The proposed partitioning method, which only compares every next-to-each-other data to merge, achieves $\mathcal{O}\left( n \right)$ calculation complexity, where n is the domain size of the input histograms, whereas the greedy algorithm requires $\mathcal{O}\left( {{2^n}} \right)$. The experimental evaluation showed that the execution time of the proposed algorithm for 4,096-domain-size data was approximately 4 h and 35 min, which was acceptable when creating a data summary for the range query processing system and not targeting on-the-fly adoption of DP. Additionally, we confirmed that the accuracy of the proposed algorithm was equivalent to that of the state-of-the-art partitioning algorithm.

Dark patterns, which are user interface designs in online services, induce users to take unintended actions. Recently, dark patterns have been raised as an issue of privacy and fairness. Thus, a wide range of research on detecting dark patterns is eagerly awaited. In this work, we constructed a dataset for dark pattern detection and prepared its baseline detection performance with state-of-the-art machine learning methods. The original dataset was obtained from Mathur et al.’s study in 2019 [1], which consists of 1,818 dark pattern texts from shopping sites. Then, we added negative samples, i.e., non-dark pattern texts, by retrieving texts from the same websites as Mathur et al.’s dataset. We also applied state-of-the-art machine learning methods to show the automatic detection accuracy as baselines, including BERT, RoBERTa, ALBERT, and XLNet. As a result of 5-fold cross-validation, we achieved the highest accuracy of 0.975 with RoBERTa. The dataset and baseline source codes are available at https://github.com/yamanalab/ec-darkpattern.

Overfitting reduces the generalizability of deep neural networks (DNNs). Overfitting is generally detected by comparing the accuracies and losses of training and validation data; however, the detection method requires vast amounts of training data and is not always effective for forthcoming data due to the heterogeneity between training and forthcoming data. The dropout technique has been employed to prevent DNNs from overfitting, where the neurons in DNNs are invalidated randomly during their training. It has been hypothesized that this technique prevents DNNs from overfitting by restraining the co-adaptions among neurons. This hypothesis implies that overfitting of a DNN is a result of the co-adaptions among neurons and can be detected by investigating the inner representation of DNNs. Thus, we propose a method to detect overfitting of DNNs using no training and test data. The proposed method measures the degree of co-adaptions among neurons using persistent homology (PH). The proposed PH-based overfitting measure (PHOM) method constructs clique complexes on DNNs using the trained parameters of DNNs, and the one-dimensional PH investigates the co-adaptions among neurons. Thus, PHOM requires no training and test data to measure overfitting. We applied PHOM to convolutional neural networks trained for the classification problems of the CIFAR-10, SVHN, and Tiny ImageNet data sets. The experimental results demonstrate that PHOM reveals the degree of overfitting of DNNs to the training data, which suggests that PHOM enables us to filter overfitted DNNs without requiring the training and test data.

In the big data era, cloud-based machine learning as a service (MLaaS) has attracted considerable attention. However, when handling sensitive data, such as financial and medical data, a privacy issue emerges, because the cloud server can access clients' raw data. A common method of handling sensitive data in the cloud uses homomorphic encryption, which allows computation over encrypted data without decryption. Previous research adopted a low-degree polynomial mapping function, such as the square function, for data classification. However, this technique results in low classification accuracy. This study seeks to improve the classification accuracy for inference processing in a convolutional neural network (CNN) while using homomorphic encryption. We apply various orders of the polynomial approximations of Google's Swish and ReLU activation functions. We also adopt batch normalization to normalize the inputs for the approximated activation functions to fit the input range to minimize the error. We implemented CNN inference labeling over homomorphic encryption using the Microsoft's Simple Encrypted Arithmetic Library (SEAL) for the Cheon-Kim-Kim-Song (CKKS) scheme. The experimental evaluations confirmed classification accuracies of 99.29% and 81.06% for MNIST and CIFAR-10, respectively, which entails 0.11% and 4.69% improvements, respectively, over previous methods.

Online advertising is widely used by commercial companies to attract customers. Tuning advertisement delivery to achieve a high conversion rate (CVR) is crucial for improving advertising effectiveness. Because advertisers require demandside platforms (DSPs) to deliver a certain number of ads within a fixed period, it is challenging to maximize CVR while satisfying ads delivery constraints. Such a combinatorial optimization problem is NP-hard when we have a considerable number of both ads and users. In this paper, we adopt Digital Annealer (DA), a quantum-inspired Ising computer, to solve the combinatorial optimization problem. The experimental evaluation result shows that the proposed method increases accuracy from 0.176 to 0.326 and achieves 20.8 times speed-up compared to baseline.

In the personalized Point-Of-Interest (POI) (or venue) recommendation, the diversity of recommended POIs is an important aspect. Diversity is especially important when POIs are recommended in the target users’ frequently visited areas, because users are likely to revisit such areas. In addition to the (POI) category diversity that is a popular diversification objective in recommendation domains, diversification of recommended POI locations is an interesting subject itself. Despite its importance, existing POI recommender studies generally focus on and evaluate prediction accuracy. In this article, geographical diversification ( geo-diversification ), a novel diversification concept that aims to increase recommendation coverage for a target users’ geographic areas of interest, is introduced, from which a method that improves geo-diversity as an addition to existing state-of-the-art POI recommenders is proposed. In experiments with the datasets from two real Location Based Social Networks (LSBNs), we first analyze the performance of four state-of-the-art POI recommenders from various evaluation perspectives including category diversity and geo-diversity that have not been examined previously. The proposed method consistently improves geo-diversity (CPR(geo)@20) by 5 to 12% when combined with four state-of-the-art POI recommenders with negligible prediction accuracy (Recall@20) loss and provides 6 to 18% geo-diversity improvement with tolerable prediction accuracy loss (up to 2.4%).

Unreliable web pages, such as fake news, have become an unavoidable problem. To tackle this problem, recent researches have adopted both content and social features to predict the credibility of the web pages; however, the accuracy is almost saturated. In this paper, we propose the adoption of Google Lighthouse features to predict web page credibility. Our experimental results show that the proposed method achieves an increased accuracy of 7.9% in comparison with state-of-the-art methods.

Phishing is a type of personal information theft in which phishers lure users to steal sensitive information. Phishing detection mechanisms using various techniques have been developed. Our hypothesis is that phishers create fake websites with as little information as possible in a webpage, which makes it difficult for content- and visual similarity-based detections by analyzing the webpage content. To overcome this, we focus on the use of Uniform Resource Locators (URLs) to detect phishing. Since previous work extracts specific special-character features, we assume that non-alphanumeric (NAN) character distributions highly impact the performance of URL-based detection. We hence propose a new feature called the entropy of NAN characters for URL-based phishing detection. Experimental evaluation with balanced and imbalanced datasets shows 96% ROC AUC on the balanced dataset and 89% ROC AUC on the imbalanced dataset, which increases the ROC AUC as 5 to 6% from without adopting our proposed feature.

The preservation of privacy during a search has become a serious problem in recent years. There is an increasing requirement to ensure that user queries are not abused by a third party, including the search provider. Fully homomorphic encryption (FHE) can conduct addition and multiplication directly over ciphertext. Using FHE, privacy, concerning both the user queries and the database of the search provider, can be protected. In this paper, we propose a privacy-preserving query system model. We implemented the proposed model in a real-world medicine side-effect query system. We applied a filtering technique, prior to the query deployment, to reduce the size of the database and used multi-threading to accelerate the search. The system was tested 10,000 times with a random query, using a database comprising 40,000 records of simulation data, and completed 99.84% of the queries within 60 seconds (s), proving the real-world applicability of our system.

In the era of big data and cloud computing, outsourcing data storage to the cloud poses the risk of its abuse or leakage. Thus, we address the problem of delegating computation on outsourced private datasets while maintaining privacy. In this study, we consider a scenario involving two data owners outsourcing their datasets to a cloud service. The cloud performs a set union computation, after which the querier sends a query to obtain information from both datasets. We propose a protocol that uses fully homomorphic encryption (FHE) and Cartesian-join of Bloom filters (CBF) as proposed by Wang et al. The protocol obtains information on the existence of a particular set of elements without learning about the residing source. To the best of our knowledge, our protocol, by using the FHE and CBF matrix, is a novel approach to ensuring the security of outsourced set union operations.

Machine learning classification has a wide range of applications. In the big data era, a client may want to outsource classification tasks to reduce the computational burden at the client. Meanwhile, an entity may want to provide a classification model and classification services to such clients. However, applications such as medical diagnosis require sensitive data that both parties may not want to reveal. Fully homomorphic encryption (FHE) enables secure computation over encrypted data without decryption. By applying FHE, classification can be outsourced to a cloud without revealing any data. However, existing studies on classification over FHE do not achieve the scenario of outsourcing classification to a cloud while preserving the privacy of the classification model, client's data and result. In this work, we apply FHE to a naïve Bayes classifier and, to the best of our knowledge, propose the first concrete secure classification protocol that satisfies the above scenario.

Recommender systems have been used for analyzing users' preference through their past activities and recommend items in which they might be interested in. There are numerous research on improving the accuracy of recommendation being conducted, so the recommender system reads user preference more accurately. However, it is important to consider the recommendation diversity, because lacking diversity will lead to recommendation being repetitive and obvious. In this paper, we propose a method to re-rank the recommendation list by appearance frequency of items to recommend more range of items. The experimental result shows that our method consistently performs better than a related work to improve recommendation diversity.

Unreliable web pages such as fake news has become a global problem in big data era. The motivation to publish fake news is often for profit; for example, earning advertisement income by putting ads on their web pages. In this paper, we focus on different usage of HTML source tags between reliable and unreliable web pages, then propose new features for predicting their credibility. The experimental result shows that our proposed features increase accuracy when used together with previously proposed Contents features.

In the big data era, multithreaded processing on a many-core machine, whose core number is still increasing, has become essential to parallelize the execution of big data applications, besides distributed computing. In such a machine, malloc-intensive applications cannot scale due to lock contentions among threads, which becomes worse as the number of threads increases. To solve the problem, we propose a new method to reduce lock contentions by batch malloc, pseudo free, and fine-grained data-locking. Experimental result shows 4.72 times speed-up in comparison with JEmalloc which is the fastest memory allocator among previous ones.

In the big data era, many users upload data to cloud while security concerns are growing. By using attribute-based encryption (ABE), users can securely store data in cloud while exerting access control over it. Revocation is necessary for real-world applications of ABE so that revoked users can no longer decrypt data. In actual implementations, however, revocation requires re-encryption of data in client side through download, decrypt, encrypt, and upload, which results in huge communication cost between the client and the cloud depending on the data size. In this paper, we propose a new method where the data can be re-encrypted in cloud without downloading any data. The experimental result showed that our method reduces the communication cost by one quarter in comparison with the trivial solution where re-encryption is performed in client side.

Discovery of repeated patterns, known as motifs, from long time series is essential for providing hidden knowledge to real-world applications like medical, financial and weather analysis. Motifs can be discovered on raw time series directly or on their transformed abstract representation alternatively. Most of time series motif discovery methods require predefined motif length, which results in long execution time because we have to vary the length to discover motifs with different lengths. To solve the problem, we propose an efficient method for discovering variable length motifs in combination of approximate method with exact verification. First, symbolic representation is adopted to discover motifs roughly followed by exact examination of the found motifs with original real-valued data to achieve fast and exact discovery. The experiments show that our proposed method successfully discovered significant motifs efficiently in comparison with state-of-the-art methods: MK and SBF.

In recommending POIs(Point-Of-Interests), factors such as the diversity of the recommended POIs are as important as accuracy for providing a satisfactory recommendation. Although existing diversification methods can help POI recommender systems suggest more diverse POIs, they lack "geographical diversification," which results in the concentration of the supposedly "diverse" recommended POIs on "a small portion" in areas where the target-user is most active. This is caused by the neglect of POI locations in the diversification, i.e., existing diversification methods try to diversify the categories of recommended items. However, geographical diversification is essential for users whose activity interests comprise many sub-areas and who require a variety of recommended POIs encompassing all their activity interests. In this paper, we propose a novel proportional geographical diversification method that recommends a variety of POIs located in the activity district of a user such that the variety of sub-areas in the district is proportional to the frequency of his/her activity in each sub-area. We compare the performance of the proposed method with existing diversification methods using real datasets. The evaluation result shows that no method except the proposed one can significantly increase geographical diversity at the expense of tolerable accuracy loss.

In the big data era, security and privacy concerns are growing. One of the big challenges is secure Frequent Pattern Mining (FPM) over Fully Homomorphic Encryption (FHE). There exist some research efforts aimed at speeding-up, however, we have a big room so as to decrease time and space complexity. Apriori over FHE, in particular, generates a large number of ciphertexts during the support calculation, which results in both large time and space complexity. To solve it, we proposed a speedup technique, around 430 times faster and 18.9 times smaller memory usage than the state-of-the-art method, by adopting both packing and caching mechanism. In this paper, we further propose to decrease the memory space used for caching. Our goal is to discard redundant cached ciphertexts without increasing the execution time. Our experimental results show that our method decreases the memory usage by 6.09% at most in comparison with our previous method without increasing the execution time.

Privacy-preserving string search is a crucial task for analyzing genomics-driven big data. In this work, we propose a cryptographic protocol that uses Fully Homomorphic Encryption (FHE) to enable a client to search on a genome sequence database without leaking his/her query to the server. Though FHE supports both addition and multiplication over encrypted data, random noise inside ciphertexts grows with every arithmetic operation especially multiplication, which results in incorrect decryption when the noise amount exceeds its threshold called level. There are two approaches to avoid the incorrect decryption: one is setting the sufficient level that assures correct decryption within the limited number of operations, and the other is resetting the noise by the method called bootstrapping. It is important to find an optimal balance between overhead caused by the level and overhead caused by the bootstrapping, since using higher level deteriorates the performance of all the arithmetic operations, while the more number of bootstrappings causes more expensive overhead. In this study, we propose an efficient approach to minimize the number of bootstrappings while reducing the level as much as possible. Our experimental result shows that it runs at most 10 times faster than a naïve approach.

The challenges in efficient data representation and similarity measures on massive amounts of time series have enormous impact on many applications. This paper addresses an improvement on Symbolic Aggregate approXimation (SAX), is one of the efficient representations for time series mining. Because SAX represents its symbols by the average (mean) value of a segment with the assumption of Gaussian distribution, it is insufficient to serve the entire deterministic information and causes sometimes incorrect results in time series classification. In this work, SAX representation and distance measure is improved with the addition of another moment of the prior distribution, standard deviation; SAX_SD is proposed. We provide comprehensive analysis for the proposed SAX_SD and confirm both the highest classification accuracy and the highest dimensionality reduction ratio on University of California, Riverside (UCR) datasets in comparison to state of the art methods such as SAX, Extended SAX (ESAX) and SAX Trend Distance (SAX_TD).

Predicting various types of user-attributes in social networks has become indispensable for personalizing applications since there are many non-disclosed attributes in social networks. However, extracted attributes in existing works are limited to pre-defined types of attributes, which results in no extraction of unexpected-types of attributes. In this paper, we therefore propose a novel method that extracts various, i.e., unlimited, types of attributes by adopting personalized PageRank to a large social network. The experimental results using over 7.9 million of Japanese Twitter-users show that our proposed method successfully extracts four types of attributes per-user in average with 0.841 of MAP@20.

Human mobility analysis with Location-Based Social Network (LBSN) data is the basis of personalized point-of-interest (POI) recommendations or location-aware advertisements. In addition to personal preference and spatiotemporal factors such as time and distance, personal context has a strong influence on mobility. An individual's familiarity with an area is an interesting context because it can bias the influence of certain factors. For example, the mobility patterns of two persons who have similar preferences are different when their familiarity with the area is different, even in the same area. In this paper, we analyze familiarity's effect on mobility patterns by using over 1.4 million check-ins gathered from Foursquare. The analysis indicates that there is a skewness of the visit time and visited venue distribution in unfamiliar areas. For instance, people go to unfamiliar areas on weekends; and venues for cultural experiences, such as museums, strongly contribute to the motivation of visit.

Internet security issues require authorship identification for all kinds of internet contents; however, authorship identification for microblog users is much harder than other documents because microblog texts are too short. Moreover, when the number of candidates becomes large, i.e., big data, it will take long time to identify. Our proposed method solves these problems. The experimental results show that our method successfully identifies the authorship with 53.2% of precision out of 10,000 microblog users in the almost half execution time of previous method.

In this paper, we propose a non-blocking buffer management scheme based on a lock-free variant of the GCLOCK page replacement algorithm. Concurrent access to the buffer management module is a major factor that prevents database scalability to processors. Therefore, we propose a non-blocking scheme for bufferfix operations that fix buffer frames for requested pages without locks by combining Nb-GCLOCK and a non-blocking hash table. Our experimental results revealed that our scheme can obtain nearly linear scalability to processors up to 64 processors, although the existing locking-based schemes do not scale beyond 16 processors.

In this paper, we propose a new system extracting potentially copyright infringement texts from the Web, called EPCI. EPCI extracts them in the following way: (1) generating a set of queries based on a given copyright reserved seed-text, (2) putting every query to search engine API, (3) gathering the search result Web pages from high ranking until the similarity between the given seed-text and the search result pages becomes less than a given threshold value, and (4) merging all the gathered pages, then re-ranking them in the order of their similarity. Our experimental result using 40 seed-texts shows that EPCI is able to extract 132 potentially copyright infringement Web pages per a given copyright reserved seed-text with 94% precision in average.

No abstract available.

No abstract available.