With the rapid development of Artificial Intelligence （AI） in recent years， the demand for Deep Neural Network （DNN） from devices with limited resources such as embedded devices and mobile devices has increased sharply. The problem of how to compress neural networks without affecting the effect of DNNs has great theoretical and practical significance， and is a hot research topic in deep learning now. Firstly， aiming at the problem that DNN is difficult to be ported to resource-limited devices such as mobile devices due to their large models and large computational cost， the experimental performance of existing DNN compression algorithms in terms of memory usage， running speed， and compression effect was deeply analyzed， so that the influence factors of the DNN compression algorithm were explored. Then， the knowledge transfer structure composed of student network and teacher network was designed， the knowledge distillation， structural design， network pruning， and parameter quantization mechanisms were fused together， and a DNN optimization and compression model based on hybrid mechanism was proposed. Experimental comparison and analysis were conducted on mini-ImageNet dataset using AlexNet as the Benchmark. Experimental results show that the capacity of compressed AlexNet is reduced by 98.5% with 6.3% loss of accuracy， which verify the effectiveness of the proposed algorithm.
For the problems that the performance of the nearest neighbor classification algorithm is greatly affected by the adopted similarity or distance measuring method， and it is difficult to select the optimal similarity or distance measuring method， with multi-similarity method adopted， a K-Nearest Neighbor algorithm with Ordered Pairs of Normalized real numbers （OPNs-KNN） was proposed. Firstly， the new mathematical theory of Ordered Pair of Normalized real numbers （OPN） was introduced in machine learning. And all the samples in the training and test sets were converted into OPNs by multiple similarity or distance measuring methods， so that different similarity information was included in each OPN. Then， the improved nearest neighbor algorithm was used to classify the OPNs， so that different similarity or distance measuring methods were able to be mixed and complemented to improve the classification performance. Experimental results show that compared with 6 improved nearest neighbor classification algorithms， such as distance-Weighted K-Nearest-Neighbor rule （WKNN） rule on Iris， seeds， and other datasets， OPNs-KNN has the classification accuracy improved by 0.29 to 15.28 percentage points， which proves that the performance of classification can be improved greatly by the proposed algorithm.
The existing similarity-based moving target trajectory prediction algorithms are generally classified according to the spatial-temporal characteristics of the data， and the characteristics of the algorithms themselves cannot be reflected. Therefore， a classification method based on algorithm characteristics was proposed. The calculation of the distances between two points is required for the trajectory similarity algorithms to carry out the subsequent calculations， however， the commonly used Euclidean Distance （ED） is only applicable to the problem of moving targets in a small region. A method of similarity calculation using geodetic distance instead of ED was proposed for the trajectory prediction of sea targets moving in a large region. Firstly， the trajectory data were preprocessed and segmented. Then， the discrete Fréchet Distance （FD） was adopted as similarity measure. Finally， synthetic and real data were used to test. Experimental results indicate that when sea targets move in a large region， the ED-based algorithm may gain incorrect prediction results， while the geodetic distance-based algorithm can output correct trajectory prediction.
The performance of the recommendation systems is restricted by data sparsity， and the accuracy of prediction can be effectively improved by reasonably filling the missing values in the rating matrix. Therefore， a new algorithm named Recommendation Rating Prediction based on Concept Lattice Reduction （RRP-CLR） was proposed. RRP-CLR algorithm was composed of nearest neighbor selection module and rating prediction module， which were respectively responsible for generating reduced nearest neighbor set and realizing rating prediction and recommendation. In the nearest neighbor selection module， the user rating matrix was transformed into a binary matrix， which was regarded as the user interest formal background. Then the formal background reduction rules and concept lattice redundancy concept deletion rules were proposed to improve the efficiency of generating reduced nearest neighbors. In the rating prediction module， a new user similarity calculation method was proposed to eliminate the impact of rating deviations caused by user’s subjective factors on similarity calculation. When the number of common rating items of two users was less than a specific threshold， the similarity was scaled appropriately to make the similarity between users more consistent with the real situation. Experimental results show that compared with PC?UCF （User-based Collaborative Filtering recommendation algorithm based on Pearson Coefficient） and RRP-UICL （Recommendation Rating Prediction method based on User Interest Concept Lattice）， RRP-CLR algorithm has smaller Mean Absolute Error （MAE） and Root Mean Squared Error （RMSE）， and better rating prediction accuracy and stability.
Aiming at the problem that the existing microblog event extraction methods are based on the content characteristics of events and ignore the relationship between the social attributes and time characteristics of events， so that they cannot identify the key events in the propagation process of microblog hot spots， a key event extraction method from microblog by integrating social influence and temporal distribution was proposed. Firstly， the social influence was modeled to present importance of microblog events. Secondly， the temporal characteristics of microblog events during evolution were considered to capture the differences of events under different temporal distributions. Finally， the key microblog events were extracted under different temporal distributions. Experimental results on real datasets show that the proposed method can effectively extract key events in microblog hot spots. Compared with four methods of random selection， Term Frequency-Inverse Document Frequency （TF-IDF）， minimum-weight connected dominating set and degree and clustering coefficient information， the proposed method has the event set integrity index improved by 21%， 18%， 26% and 30% on dataset 1 respectively， and 14%， 2%， 21% and 23% on dataset 2 respectively. The extraction effect of the proposed method is better than those of the traditional methods.
The detection precision is low when the diatom training sample size is small， so a Multi-scale Multi-head Self-attention （MMS） and Online Hard Example Mining （OHEM） based few-shot diatom detection model， namely MMSOFDD was proposed based on the few-shot object detection model Two-stage Fine-tuning Approach （TFA）. Firstly， a Transformer-based feature extraction network Bottleneck Transformer Network-101 （BoTNet-101） was constructed by combining ResNet-101 with a multi-head self-attention mechanism to make full use of the local and global information of diatom images. Then， multi-head self-attention was improved to MMS， which eliminated the limitation of processing single object scale of the original multi-head self-attention. Finally， OHEM was introduced to the model predictor， and the diatoms were identified and localized. Ablation and comparison experiments between the proposed model and other few-shot object detection models were conducted on a self-constructed diatom dataset. Experiment results show that the mean Average Precision （mAP） of MMSOFDD is 69.60%， which is improved by 5.89 percentage points compared with 63.71% of TFA； and compared with 61.60% and 60.90% the few-shot object detection models Meta R-CNN and Few-Shot In Wild （FSIW）， the proposed model has the mAP improved by 8.00 percentage points and 8.70 percentage points respectively. Moreover， MMSOFDD can effectively improve the detection precision of the detection model for diatoms with small size of diatom training samples.
As a key platform for people to acquire and disseminate news events， Weibo hides rich event information. Extracting storylines from Weibo data provides users with an intuitive way to accurately understand event evolution. However， the data sparseness and lack of context make it difficult to extract storylines from Weibo data. Therefore， two consecutive tasks for extracting storylines automatically from Weibo data were introduced： 1） events were modeled by propagation impact of Weibo， and the primary events were extracted； 2） the heterogeneous event graph was built based on the event features， and an Event Graph Convolution Network （E-GCN） model was proposed to improve the learning ability of implicit relations between events， so as to predict story branches of the events and link the events. The proposed method was evaluated from the perspectives of story branch and storyline on real datasets. In story branch generation evaluation， the results show that compared with Bayesian model， Steiner tree and Story forest， the proposed method has the F1 value higher by 28 percentage points， 20 percentage points and 27 percentage points on Dataset1 respectively， and higher by 19 percentage points， 12 percentage points and 22 percentage points on Dataset2 respectively. In storyline extraction evaluation， the results show that compared with Story timeline， Steiner tree and Story forest， the proposed method has the correct edge accuracy higher by 33 percentage points， 23 percentage points and 17 percentage points on Dataset1 respectively， and higher by 12 percentage points， 3 percentage points and 9 percentage points on Dataset2 respectively.
Traditional methods of generating digital camouflages cannot generate digital camouflages based on the background information in real time. In order to cope with this problem， a digital camouflage generation method based on cycle-consistent adversarial network was proposed. Firstly， the image features were extracted by using densely connected convolutional network， and the learned digital camouflage features were mapped into the background image. Secondly， the color retention loss was added to improve the quality of generated digital camouflages， ensuring that the generated digital camouflages were consistent with the surrounding background colors. Finally， a self-normalized neural network was added to the discriminator to improve the robustness of the model against noise. For the lack of objective evaluation criteria for digital camouflages， the edge detection algorithm and the Structural SIMilarity （SSIM） algorithm were used to evaluate the camouflage effects of the generated digital camouflages. Experimental results show that the SSIM score of the digital camouflage generated by the proposed method on the self-made datasets is reduced by more than 30% compared with the existing algorithms， verifying the effectiveness of the proposed method in the digital camouflage generation task.
