Search

Attributes are important auxiliary information for representing user and item features, especially in data-sparse scenario, and can serve as their main source. However, different attributes have different importance for representing users or items, and users have different preferences for different attributes of an item, which is often ignored in current research. With the rapid development of social network platforms, the use of users' social relationships to improve and enhance the performance of recommender systems has also received considerable attention. However, most studies have treated the influence of different social relationships in the same way. Different social friends have different influences on users' preferences. To solve these problems, this study proposes a recommendation model based on attributes and social relationships. In the model, we use attributes to learn the feature representations of users and items, while we use social relationships to augment the feature representations of users. We developed two attention mechanisms: attribute-level and social-level. The former was used to distinguish the relative importance of different attributes, and the latter was used to model the influence of different social relationships. The novel recommendation model in this study innovatively combines attribute information and social relationships to improve feature representation learning of users and items, and to alleviate the problem of data sparsity. Extensive experiments on public datasets have confirmed the effectiveness of the proposed model, and comparisons with different baselines showed significant improvement in recommendation accuracy. [ABSTRACT FROM AUTHOR]

Using trust relationships can improve the accuracy of recommendation systems; however, it is affected by data sparsity. Random walks can harvest the behavioral relationships between users and compensate for data sparsity. However, existing random walk methods may generate information of little value or even interference, which affects recommendation accuracy. A collaborative filter recommendation based on multi-factor random walk was proposed to address these problems. In this method, the comprehensive trust values of the current user over other users based on the rating time, user attribute preference, and number of mutual friends were computed more accurately to determine the trust neighborhood of the current user. Thus, trustworthy users with preferences similar to those of the current user were accurately and conveniently obtained to act as neighboring users, and the sparsity of the trust relationship could be alleviated. The final predicted rating of the target item was obtained using the ratings of multiple neighboring users for the target item or similar items to improve recommendation accuracy. Moreover, to avoid generating ratings that affect the prediction accuracy during the walk, a decision to stop the walk was made based on the comprehensive trust value, item similarity, and depth of the current walk, further improving the recommendation accuracy. An evaluation was performed on two datasets, and the proposed method achieved superior prediction accuracy and coverage rate even with sparse data and exhibited a high recommendation accuracy even with cold-start users. [ABSTRACT FROM AUTHOR]

Most standard classification algorithms are difficult to effectively learn and predict from imbalanced network traffic data, which usually leads to lower classification accuracy. To analyze the influence of imbalanced network traffic data on the performance of standard classification algorithms, the imbalanced data augmentation algorithms are first designed to obtain the imbalanced network traffic data set with gradually varying Imbalance Ratio (IR) and belonging to the same distribution. Then, to obtain more objective classification result and simplify the evaluation process, the evaluation metric AFG is used to evaluate the classification performance of standard classification algorithms based on area under the receiver operating characteristic curve (AUC), F-measure and G-mean. Finally, based on AFG and coefficient of variation (CV), performance stability of standard classification algorithms on imbalanced network traffic data is obtained. Experiments of eight widely used standard classification algorithms on 25 different imbalanced network traffic data demonstrate that the classification performance of GNB, RF and DT is unstable, while BNB, KNN, LR, GBDT, and SVC are relatively stable and not susceptible to imbalanced data. Especially, the KNN has the most stable classification performance. Also, the results are statistically confirmed by Friedman and Nemenyi post hoc statistical tests. [ABSTRACT FROM AUTHOR]

Point alignment is an important topic in computer vision. In order to implement cross-stitch embroidery in an automatic way, the grid points in a raster like pattern are desired to be recognized at first. In this paper, we design an algorithm to align the grid points in a raster like pattern in case of checkerboard pattern, and the proposed method was used to recognize the weave holes in a cross-stitch pattern. The method implemented in four steps. First, vertices on the boundary were extracted via removing the inner points. Next, the four corner points were detected in the extracted boundary points. Then the left-lower and left-upper points were positioned. Finally, the points alignment was implemented starting at the left-lower point from the bottom to top, left to right. The comparison experiments demonstrated that our method is robust to geometrical distortion and pose change. The method was applied to align the weave holes in a cross-stitch pattern. The results showed that the proposed method addressed its potential application to the machine weave of the cross-stitch. [ABSTRACT FROM AUTHOR]