Generative Adversarial Network (GAN) can automatically generate target images, and is of great significance to the generation of building arrangement of similar blocks. However, there are problems in the existing process of model training such as the low accuracy of generated images, the mode collapse, and the too low efficiency of model training. To solve these problems, a Conditional-Wassertein Generative Adversarial Network (C-WGAN) model for image generation was proposed. First, the feature correspondence between the real sample and the target sample was needed to be identified by this model, and then the target sample was generated according to the identified feature correspondence. The Wassertein distance was used to measure the distance between the distributions of two image features in the model, the GAN training environment was stablized, and mode collapse was avoided during model training, so as to improve the accuracy of the generated images and the training efficiency. Experimental results show that compared with the original Conditional Generative Adversarial Network (CGAN) and the pix2pix models, the proposed model has the Peak Signal-to-Noise Ratio (PSNR) increased by 6.82% and 2.19% at most respectively; in the case of the same number of training rounds, the proposed model reaches the convergence state faster. It can be seen that the proposed model can not only effectively improve the accuracy of image generation, but also increase the convergence speed of the network.

The current research on the activity semantic recognition only extracts the sequence features and periodic features on the time dimension, and lacks deep mining of spatial information. To solve these problems, an activity semantic recognition method based on joint features and eXtreme Gradient Boosting (XGBoost) was proposed. Firstly, the activity periodic features in the temporal information as well as the latitude and longitude features in the spatial information were extracted. Then the latitude and longitude information was used to extract the heat features of the spatial region based on the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm. The user activity semantics was represented by the feature vectors combined with these features. Finally, the activity semantic recognition model was established through the XGBoost algorithm in the integrated learning method. On two public check-in datasets of FourSquare, the model based on joint features has a 28 percentage points improvement in recognition accuracy compared to the model with only temporal features, and compared with the Context-Aware Hybrid (CAH) method and the Spatial Temporal Activity Preference (STAP) method, the proposed method has the recognition accuracy increased by 30 percentage points and 5 percentage points respectively. Experimental results show that the proposed method is more accurate and effective on the problem of activity semantic recognition compared to the the comparison methods.

In order to solve the link prediction problem of knowledge graph, a shared variable network model named LCPE (Local Combination Projection Embedding) was proposed, which realized the prediction of links by embedding entities and relationships into vector space. By analyzing the Unstructured Model, it was derived that the distance between related entities' embedding was shorter in the vector space, in other words, similar entities were more likely to be related. In LCPE model, ProjE model was used based on similarity between two entities to judge whether the two entities were related and the relation type between them. The experiment shows that with the same number of parameters, the LCPE improves Mean Rank by 11 and lifts Hit@10 0.2 percentage points in dataset WN18 while improves Mean Rank 7.5 and lifts Hit@10 3.05 percentage points in dataset FB15k, which proves that the similarity between entities, as auxiliary information, can improve predictive capability of the ProjE model.

To overcome the limitation of the assumptions of topic model-word independence and topic independence, a topic model which inosculated the spatial relationship of visual words was proposed, namely Markov Topic Random Field (MTRF). In addition, it was discussed that the "topic" of topic model represented the part of object in image processing. There is a high probability of the neighbor visual words generated from the same topic, and whether the visual words were generated from the same topic determined the topic was generated from Markov Random Field (MRF) or multinomial distribution of topic model. Meanwhile, both theoretical analysis and experimental results prove that "topic" of topic model appeared as mid-level feature to represent the parts of objects rather than the instances of objects. In experiments of image classification, the average accuracy of MTRF was 3.91% higher than that of Latent Dirichlet Allocation (LDA) on Caltech101 dataset, and the mean Average Precision (mAP) of MTRF was 2.03% higher than that of LDA on VOC2007 dataset. Furthermore, MTRF assigned topics to visual words more accurately and got the mid-level features which represented the parts of objects more effectively than LDA. The experimental results show that MTRF makes use of the spatial information effectively and improves the accuracy of the model.