9 results about "Feature structure" patented technology

The invention relates to a water eutrophication prediction method based on a semantic-driven GPT2 model. The water eutrophication prediction method is applied to intelligent prediction of multivariable water quality monitoring time sequence data. The method comprises the following steps: firstly, designing a description template, extracting statistical characteristic information of water quality indexes, converting the statistical characteristic information into natural language cues, and enhancing the perception ability of a model through semantic guidance; dynamic fusion of related vocabularies and time series data is realized by adopting a sparse attention mechanism; a graph attention network is introduced to model the dependency relationship among multiple variables, and the feature structure expression ability is improved; performing context feature extraction in combination with a GPT2 model, and capturing long-term dependency and global dynamic features; and finally, output prediction is carried out through a bilinear gating residual module. The method can accurately capture the complex interaction relationship between the water quality variables, improves the prediction precision of the water quality index trend, and has the advantages of high adaptability, strong robustness and good popularization and application values.

The invention discloses a recommendation system user behavior sequence feature enhancement method based on a big language model, which comprises the following steps: acquiring a recommendation data set, performing feature structure analysis according to the use of the recommendation data set in a recommendation system and a conventional field to construct a first text, and inputting the big language model to generate a data set description; constructing a second text according to the target task and the sequence feature type, and inputting the second text and the data set description into a large language model to generate feature semantics; constructing a third text based on the feature semantics, and inputting the large language model to generate feature codes; feature codes are embedded into a prediction model training environment, and automatic generation and loading of new features in training data are achieved; and carrying out a new feature experiment in the prediction model framework, and evaluating and generating a new round of sequence features. The method has higher expressive power, context perceptibility and task adaptability in a behavior data modeling-oriented recommendation system and a personalized system, and can effectively solve the technical bottlenecks of an existing method in the aspects of structure, expression, generalization and the like.

The invention provides an image subtitle algorithm based on structured semantic extraction and geometric feature fusion. The image subtitle algorithm comprises a visual semantic feature extraction module and a subtitle generation module, the visual semantic feature extraction module comprises a regional feature, a network feature, a structured semantic feature and the subtitle generation module, firstly, the most similar text sentence is retrieved from an image through a CLIP model, concept semantics and attribute features are extracted, and after Top-K text features are retrieved through cosine similarity, multi-level clustering is carried out, and finally, the subtitle generation module is used for generating subtitles; and the isolation problem of semantic information extraction is fundamentally solved. And secondly, gradually fusing the grid features, the regional features and the structured semantic features by designing a geometric perception semantic enhancement encoder, and introducing geometric coordinate information in the fusion process, thereby enhancing the expression ability of the structured semantic features. According to the process, the modeling capability of the model for the image space relationship is remarkably enhanced, and the understanding of global semantics is improved.

The application discloses a self-supervised heterogeneous knowledge graph learning method based on feature structure homogeneity and long-range heterogeneity. First, a self-expression solver is proposed, which can capture the complementary homogeneity between meta-path and node features, and then obtain homogeneity representation. At the same time, the application designs a path encoder to model various interaction relationships. Through adaptive fusion, the cross-type interaction relationship is explicitly included while reducing noise interference, and long-range heterogeneity is captured. Theoretical analysis verifies that the homogeneity representation has a high-order grouping effect and can effectively capture complementary homogeneity; the path encoder has adaptive smoothing capability to filter noise; and cross-type interaction modeling integrates homogeneity and heterogeneity to introduce more task-related information. Experiments on various data sets including large-scale data sets fully verify the superiority of the method.

The invention discloses a method for relieving catastrophic forgetting of incremental target detection based on incremental relationship consistency, and relates to the technical field of catastrophic forgetting. According to the method, disastrous forgetting is relieved, and under the condition that old task data is not repeatedly used, the old category detection performance is stably reserved when a new category is learned; knowledge offset is eliminated, dependency of the model on task specific semantics is reduced by learning cross-task relation consistency, and semantic differences of different tasks are adapted; the feature structure is stabilized, the category boundary is kept clear through feature distribution constraint, and new category features are prevented from damaging the old task feature space; the cross-task adaptability is improved, intra-class and inter-class relation learning is enhanced, and the generalization ability of the model in single-step and multi-step increment scenes is enhanced; moreover, the training process is simplified, a collaborative optimization framework is designed, the training efficiency and performance are considered, and the actual application of the incremental detection technology is promoted.

The invention provides a combinatorial optimization problem-oriented structure-perceptible code generation method and system. The method comprises the following steps of S1, extracting a feature structure for solving a problem by using a trained graph neural network; s2, the feature structure and a natural language prompt jointly generate an optimization code so as to generate a structure-perceptible code; and S3, carrying out iterative optimization on the code. According to the method, the optimization effect is remarkably improved; according to the method, codes better than those of a previous framework can be found, compared with code generation without structure perception, the effect of optimizing the combination optimization problem of the found codes can be remarkably improved, and higher interpretability is achieved.

The invention relates to the technical field of security development, and provides a code defect analysis method, which is used for improving the code defect analysis efficiency and reducing the workload of manual review. The method comprises the following steps: acquiring a code defect report and appeal data; extracting a first feature structure from the code snippet, and generating a composite fingerprint based on the first feature structure; when it is determined that no fingerprint matched with the composite fingerprint exists in a pre-stored composite fingerprint library, in a pre-stored feature structure library, determining a similar feature structure of which the similarity with the first feature structure meets a preset condition, and determining a first false alarm probability of the code snippet; analyzing the code defect report and the appeal data by using a pre-trained large language model, and determining a second false alarm probability of the code snippets and a rationality score of the appeal data; and determining a defect misinformation evaluation value of the code snippet, and determining that the code snippet does not have defects under the condition that the defect misinformation evaluation value is greater than or equal to a first preset threshold value.

The application relates to the technical field of text processing, in particular to an advertisement copy automatic generation method based on artificial intelligence, which comprises the following steps: collecting advertisement copies in a traffic delivery platform, identifying high-quality copies in the platform, determining the copy feature structure of the high-quality copies, respectively analyzing user keywords of access users and transaction users, determining user demand labels in different platforms, extracting key information words from function information of a target product based on the key demand labels of the users, identifying corresponding matching information words in the key information words of the target product based on the text structure of the high-quality copies, combining the identified matching information words according to the text structure of the high-quality copies, and automatically generating advertisement copies. The application automatically collects and analyzes the feature structure of high-quality copies in the platform through artificial intelligence, improves the production efficiency of the copies, and further improves the adaptability to the reading habits of the users of the platform based on the copies generated according to the feature structure of the platform.

The present invention discloses a generative zero-shot learning method based on dual-state collaborative decoupling and semantic refinement, which belongs to the field of generative zero-shot learning. This method realizes the complementary expression of static and dynamic features by decoupling the static features such as background, structure and details of the image, and extracting dynamic common features by a cross-modal label generation module; dynamically adjusts the focus of features according to different confusion types, enhances feature discriminability, and alleviates cross-category interference; at the semantic level, by constructing a visual-semantic mirror cross-attention mechanism, it realizes bidirectional alignment between semantic features and visual features, further improving the multi-granularity ability and adaptability of semantic representation. The present invention takes feature structure decoupling, confusion adaptive regulation and dynamic semantic alignment as the core, effectively improving the cross-category generalization ability and the quality of generated samples, breaking through the performance bottleneck of traditional generative zero-shot learning, and has strong theoretical value and broad application prospects.