58results about How to "Alleviate data sparsity" patented technology

The invention discloses a collaborative filtering recommending method based on the characteristics and credibility of users. The collaborative filtering recommending method based on the characteristics and credibility of the users comprises the following steps: respectively computing similarity between users according to the historical evaluating scores of users to a project and characteristics of the users; selecting proper weight to combine two kinds of similarity to obtain the finial similarity among the users; computing the nearest neighbor of the user according to KNN (K Nearest Neighbors) method; using the quantity of historical evaluating scores of the users as the credibility, and computing a recommending result by using the nearest neighbor matrix of the user and a predictor formula added with the credibility.

The invention relates to a recommending method integrating user and project rating and characteristic factors, which mainly aims at solving the problems of cold starting and data sparsity in traditional collaborative filtering recommending algorithm. A selected experimental data set is classified into a training data set and a test data set; similarity based on users and user characteristics and similarity based on projects and project characteristics are respectively calculated on the training data set; the similarities based on the users and the user characteristics are combined by selecting appropriate weight; a user nearest neighboring matrix is calculated respectively according to a dynamic selection threshold value method and a K nearest neighbor (KNN) method, an optimum method is selected through the comparison, and user recommend results are calculated by utilizing a recommending formula; the similarities based on the projects and the project characteristics are combined by selecting an appropriate weight; a project nearest neighboring matrix is calculated respectively according to the dynamic selection threshold value method and the KNN method,, an optimum method is selected through the comparison, and project recommend results are calculated by utilizing a recommending formula; and the user recommend results are combined with the project recommend results by selecting an appropriate weight.

The invention relates to a cross-network behavior association method for individual application. The method comprises the steps that firstly specific heterogeneous information of different networks is determined, topic modeling is carried out on the networks where the heterogeneous information is located respectively, and topic distribution expression of the corresponding information is obtained. Then, behavior information of users in the corresponding networks is aggregated to obtain the topic distribution expression of the users in the different networks respectively through calculation, and a set of network-associated user attribute factor vectors is learnt for the different networks through the one-to-one corresponding relation of topic distribution of the inter-network association users in the different networks. Finally, the topic distribution of the information of the different networks can be conversed mutually through the user attribute factor vectors of the different networks, and the aim of inter-network behavior information association is achieved. A topic model and user perception are introduced in, the association breaks through the limitation of semantic association, and perception is carried out under thinner granularity.

The invention discloses a POI recommendation method combining travel interest and social preference, and the method comprises the steps: learning a user travel behavior according to the historical POIdata distribution of a user in an LBSN, and predicting a POI accessed by the user in the future travel according to the current position; Constructing social contact associated interest similarity byextracting theme vectors; Constructing a heterogeneous travel information network, and establishing interest similarity of travel behaviors; Determining a similar group by integrating social interestsimilarity analysis and travel behavior similarity analysis; Generating a candidate POI set by combining the predicted POI of the future travel access of the user and the similar groups of the user,and discovering TOP-N POI that users are most likely to go to by the calculation. According to the method, the similar groups of the user are discovered by utilizing social interest and travel preferences while position prediction is considered, more proper interest point recommendation can be comprehensively provided for the user by utilizing the similar groups instead of friend users, and the problem of data sparsity in the LBSN is relieved, so that the recommendation effect can be better improved.

The invention discloses a hybrid recommendation method based on a graph convolutional neural network. The method comprises the following steps of collecting the behavior data of a user to an article and the attribute information of the article; modeling the collected data into a heterogeneous information graph, and learning the feature vector representation of each node by utilizing the graph convolutional neural network; carrying out hybrid recommendation based on the user and the item feature vectors. According to the method, multiple data including the behavior data of the user to the articles and the article attribute information are mainly utilized to obtain the feature vector representation of the user and the articles, and the hybrid recommendation based on the feature vectors is implemented, so that the influence of a data sparsity problem and a heavy feature engineering problem is relieved, then the recommendation effect is improved, and the user satisfaction is improved.

The invention discloses a social image recommendation method based on the hierarchical attention mechanism, which comprises the following steps: 1. Constructing heterogeneous data: a scoring matrix ofan image by a user, an uploading information matrix of the image by the user, and a social relationship matrix between the user and the user; 2. Processing heterogeneous data and image set to obtainsocial embedding matrix, content embedding matrix and style embedding matrix; 3. Input that embedding matrix into the bottom attention network to obtain the social semantic information; 4. The weightsof the three aspects of the social semantic information obtained from the top-level attention network when the user chooses the image; 5. According to the feature matrix, three kinds of social semantic information matrices and their weights, we can get the scoring prediction value, and then we can recommend the image. The invention not only effectively solves the problem of data sparsity by usingthe uploading information of the user and the social information among the users, but also well explains the user preference through the hierarchical attention mechanism, and realizes the accurate social image recommendation.

The invention provides a probability matrix decomposition recommendation method. The method provided by the invention is based on explicit and implicit feedback information and combined with a probability matrix decomposition technology at the same time so as to help a user to score the commodity. The method comprises the steps of firstly carrying out probability matrix decomposition on a user trust relationship matrix, then carrying out probability matrix decomposition on a commodity score matrix of the user and the implicit feedback information, solving an implicit user characteristic matrix, an implicit commodity characteristic matrix and an implicit feedback information characteristic vector by integrating results of the two times of decomposition, and finally calculating a recommendation score of the user. The method provided by the invention can help the user to recommend a score of the commodity by well utilizing an online social network, relieves a data sparse problem and a cold start problem in the social network, and has an excellent recommendation effect. Meanwhile, the probability matrix decomposition recommendation method can be applied to a recommendation system with a large-scale data set.

The invention relates to a real-time individuation video recommending method based on a text stream network. The method includes the steps: S1 real-time user documents are built by using current tweets published and forwarded by users on the text stream network, a hot spot event space is built for a plurality of real-time user documents, and a plurality of current hot spot events followed by the users are obtained; S2 user documents are built by using all information of the users on the text stream network, a long-term interest topic space is built for the plurality of user documents, and long-term interest distribution vectors of the users on the space are obtained; S3 the plurality of hot spot events are ranked by using the long-term interest distribution vectors of the users, and the most interested current hot spot event of the users is obtained; S4 a plurality of videos relevant to the most interested current hot spot event of the users are retrieved on a video application platform; S5 long-term interest feature vectors of the users are obtained by using information of the users on the video application platform; S6 the plurality of videos are ranked again by using the long-term interest feature vectors of the users, and top N videos are recommended to the users.

The invention discloses a cold-start recommendation method based on user preferences and trust. The method comprises the steps of S1, measuring comprehensive trust values between users according to social information of the users, and constructing a trust relation matrix; S2, calculating preference similarity degrees of the users according to user scoring data, and constructing a preference relation matrix; S3, utilizing a calculation method of comprehensive similarity degrees to fuse preference relations and trust relations, and using a bee colony algorithm to iteratively update weights in the comprehensive similarity degrees, carrying out multi-objective optimization to enable the weights to become optimal in a self-adaptive manner, and constructing a preference trust relation matrix; S4, selecting a most-trusted neighbour set of the target user to predict scoring values of corresponding items for the target user on the basis of the preference trust relation matrix; and S5, recommending the items with high prediction scores to the target user. According to the method, the precision of user trust measuring is improved, the user behavior preferences are more accurately constructed, and the quality of recommendation for the cold-start user is improved.

The invention discloses a K neighbor-based Bayesian personalized recommendation method. The K neighbor-based Bayesian personalized recommendation method comprises the following steps: 1) through behavior data of a user, seeking K neighbors of the user; 2) according to observed positive feedback items of the user and observed positive feedback items of a user group consisting of k neighbor users of the user, dividing an item set; 3) determining an item level preference relation of the user; 4) maximizing the probabilities of all the users on the item set to obtain an objective function, wherein item prediction of the user is realized by adopting a matrix decomposition model; parameters in the objective function are solved by adopting a stochastic gradient descent method. The invention further discloses a K neighbor-based Bayesian personalized recommendation device. Through the K neighbor-based Bayesian personalized recommendation method and the K neighbor-based Bayesian personalized recommendation device, mutual impact between the users is taken into account, and through the impact, the item set is divided, so that the number of unobserved items is reduced, and an adverse impact caused by data imbalance and data sparseness in the recommendation process is effectively relieved.

The invention discloses a sorting prediction-oriented social recommendation method. The sorting prediction precision is further improved by utilizing concealed features of a user and a project. The method comprises the steps of firstly modeling the potential features of the project and the potential features of the user extracted by using LFM through utilizing a PL model; secondly building a multidimensional trust model by utilizing the user as features of a truster and a trustee; and finally performing joint modeling to build a sorting prediction model, and performing optimization processingon the model to obtain optimal first N recommendation lists. According to the method, while social network structure information is considered, two different roles of the user serve as the truster andthe trustee to perform modeling, so that the social information is combined in recommendation, and the purpose of optimizing a recommendation result can be achieved under the condition of data sparseness.

The invention discloses an automatic Chinese real word error proofreading method. The method comprises the following steps: firstly, generating a Chinese real word confusion set by utilizing a correct word dictionary and a Chinese character confusion set; secondly, verifying a current word by utilizing statistical knowledge; thirdly, generalizing a data sparseness problem of context feature easing corpora through synonyms; finally, judging whether the current word is a real word error by estimating the probability of occurrence of the current word in the text by utilizing a Bayesian model, marking the real word error, and giving a modification suggestion list. According to the automatic Chinese real word error proofreading method disclosed by the invention, the problems of data sparseness, low correct word judging and proofreading efficiency and the like in the prior art are solved; the method has relatively high efficiency and accuracy.

The invention discloses an attribute reasoning and product recommendation method based on an adaptive graph convolutional network. The method comprises the following steps: 1, constructing heterogeneous data: a scoring matrix of a user to a product, a user attribute matrix, a product attribute matrix, a user attribute index matrix and a product attribute index matrix; 2, performing missing value filling preprocessing on the user attribute matrix and the product attribute matrix; 3, obtaining a cooperative matrix through one-hot coding; 4, constructing a feature fusion layer according to the attribute matrix and the collaborative matrix; 5, carrying out feature propagation through a graph convolution layer; 6, constructing a prediction layer to perform attribute reasoning and product recommendation; 7, updating the node attribute matrix according to the output result of the prediction layer; and 8, repeating the steps 4-7 until the attribute reasoning and product recommendation effectsare optimal. According to the method, the high-order structure information of the graph, the internal interaction between the node attributes and the potential association between the node attributesand the link relationship can be fully mined, so that more accurate attribute reasoning and product recommendation are realized.

The invention discloses a commodity recommendation model for relieving data sparsity and commodity cold start, relates to the technical field of interest mining, and provides a CTR prediction model GEARec for personalized recommendation of commodities. By introducing a graph embedding technology, the problems of sparsity and cold start are solved, pre-trained Embedding is input into an upper-layer deep neural network, the convergence speed of the network can be increased, and the model performance is further improved. The GEARec model is constructed by a pre-trained Embedding layer, a multi-head self-attention network layer, an AUGRU and an MLP. After the input Embedding vector representation of the user behavior sequence passes through the multi-head self-attention network layer and the AUGRU, the calculation result and other Embedding are spliced and input into the MLP to automatically learn the nonlinear relation between the features, and finally the probability that the user purchases candidate commodities is output.

The invention discloses a short text classification method based on multiple weak supervision integration, and the method comprises the steps: obtaining an original data set and a knowledge base, andcarrying out the data preprocessing; carrying out knowledge extraction on the preprocessed data; representing the extracted knowledge as an annotation function, and using the annotation function for data annotation; carrying out label integration through a conditional independent model; training a classification model based on a full-connection neural network; evaluating and optimizing the classification model to obtain an optimal model; and performing short text classification by utilizing the optimal model. According to the short text classification method based on multiple weak supervisionintegration, explicit knowledge and implicit knowledge are completely expressed in a mode of combining keyword matching, regular expression and remote supervision clustering; by means of probability labels generated by a label integration mechanism, automatic labeling of label-free data is achieved, the problem of data sparsity of short texts is relieved, and the problem of unbalanced classification of the short texts is effectively solved.

The invention discloses a graph neural network recommendation method based on multi-aspect enhancement. According to the invention, an aspect-enhanced graph neural network framework is defined, and the framework mainly comprises three modules: a feature learning module, an aspect sensing graph module and a routing-based fusion module. The feature learning module learns aspect emotion features and interaction features by using BERT and an embedded layer. The aspect sensing graph module captures fine-grained user preferences and item attributes by constructing a plurality of aspect sensing graphs in parallel. The routing-based fusion module realizes dynamic fusion of aspect preferences by learning distribution of user preferences in different aspects. According to the method, fine-grained user preferences are learned by means of multiple aspect sensing graphs, and unified preference representation is obtained by dynamically fusing multi-aspect preference features by means of a routing mechanism, so that the recommendation performance and interpretability are improved.

The invention discloses a recommendation method based on heterogeneous context perception. The method is characterized by comprising the steps that (1) a user-item score matrix, interactive context information and property context information are acquired; (2) an interactive context score prediction function and an interactive context perception model are constructed based on a tensor decomposition technology; (3) a property context score prediction function and a property context perception model are constructed based on a matrix decomposition technology; (4) semi-supervised cooperative training is performed on the interactive context perception model and the property context perception model; and (5) weight fusion scoring is performed on the interactive context perception model and the property context perception model, and then recommendation is performed according to scores. Through the method, influences of property context information and interactive context information on recommendation can be perceived, the problem of data sparsity of a recommendation system can be relieved, and the accuracy of recommendation is improved.

The invention discloses a network service recommendation method based on positions and a trust relationship. Through the method, mobile user position context information, user preferences and the trust relationship among users are fused, and personalized network service recommendation with real-time performance and interactivity is realized. According to the method, on the basis of a network service recommendation method framework based on the positions and the trust relationship, a mobile user preference similarity calculation method based on the mobile user position context information is proposed while the generality of neighbor similarity measurement is met; then a trust value calculation method is proposed in combination with basic user communication dataset information; and the mobile user preference similarity calculation method and the trust value calculation method are fused and applied to a network service recommendation process based on the mobile user positions and the trust relationship, and therefore the personalized network service recommendation method based on the mobile user positions is formed. By use of the method, the accuracy and reliability of network servicerecommendation are effectively improved, and meanwhile the problems of data sparsity and cold start possibly existing in the recommendation process are relieved.

The invention provides a score prediction method based on multi-source user comments, and belongs to the field of recommendation systems. The method includes: performing data preprocessing on the historical consumption record of the user; calculating the similarity between different users according to the user-article scoring matrix and a similarity formula, and supplementing the user comments byusing the related comments written by the similar users with the highest similarity; extracting comment features; and finally performing comment feature fusion processing. According to the invention,related comments written by similar users can be screened out based on historical consumption records of the users; according to the method, the user comments are supplemented, and the user comment supplementing document is constructed for each user, so that the problem of data sparsity of the user comments can be relieved, user portraits are enriched, the accuracy of score prediction is improved,and the satisfaction degree of the users on a recommendation system is further improved. Besides, the user comment supplementary document is composed of related comments written by similar users, andhas a certain difference from the comments written by the users themselves, so that the recommendation diversity can be improved.

The invention belongs to the technical field of artificial intelligence and information retrieval, and particularly relates to a session recommendation model integrating user microcosmic behaviors anda knowledge graph, called MKMSR for short. According to the model, firstly, different deep learning models are used for encoding an interactive operation sequence and an interactive article sequenceof a user respectively, and embedded representation vectors of operations and articles are combined into microscopic behavior representation (vectors) of the user; a comprehensive representation (vector) of the session is generated based on the microscopic behavior representation of the user by adopting an attention mechanism, thereby realizing accurate prediction of the next interactive article.Besides, the model adjusts and optimizes item representation based on a TransH knowledge representation model, and integrates the item representation and the session recommendation task into a multi-task learning framework together, so that the session recommendation effect is further improved, and the problem caused by sparse historical interaction data of the user is effectively relieved.

The invention discloses a time-aware adaptive interest point recommendation method based on K-means clustering. The method comprises the following steps: 1, converting a sign-in data set into a three-dimensional scoring matrix; 2, counting the number of sign-in users, the number of accessed interest points and sign-in times in each time slot, and constructing a three-dimensional sign-in feature vector of each time slot; 3, performing K-means clustering on the time slots, and calculating the time similarity between the time slots in the same cluster; 4, calculating the user similarity at the current time by using the score information in other time slots in the same time cluster; step 5, improving a traditional user-based collaborative filtering method by using a time clustering result and time similarity in a cluster, so that the traditional user-based collaborative filtering method can adaptively generate interest point prediction scores according to current recommendation time; and 6, comparing the recommendation precision of the recommendation system provided by the invention with the recommendation precision of other classic recommendation systems, and evaluating the accuracy and effectiveness of the provided technology.

The invention discloses an event detection method and a device based on a hybrid attention network, and the method comprises the steps: constructing a hybrid attention network model which comprises amulti-language representation layer, a hybrid attention layer and a classification layer; performing translation of a source text and acquisition of a target text of multiple languages on the multi-language representation layer, performing text alignment, and converting the target text of multiple languages into vector representation of a sentence sequence; using the hybrid attention layer for carrying out context attention learning on texts of multiple languages in parallel and carrying out information fusion of cross-source languages and multiple target languages through a multi-language attention mechanism; and performing prediction classification of event types in the classification layer.

The invention discloses a cross-dynamic filling method for relieving a data sparsity problem in a recommendation system. The method comprises the steps of (1) preprocessing e-commerce user behavior data and obtaining a user information table and a product information table, (2) establishing a user-commodity-score matrix, (3) calculating a user similarity and a product similarity according to the established user information table and product information table, and (4) carrying out data sparsity processing: gradually filling the user-commodity-score matrix through a user similarity and productsimilarity cross-dynamic filling method. Compared with a single user and commodity similarity information filling method, the method has the advantages that the user similarity and the product similarity are skillfully combined, the user-commodity-score matrix can be filled well, and thus the data sparsity problem in recommendation can be relieved. An accurate recommendation can be carried out foreach user, a user personalized experience is improved, merchant sales are promoted, and merchant revenue is increased.

The invention discloses a probability matrix decomposition recommendation method based on user context coupling similarity. The method comprises the steps of firstly, constructing a user relationshipmatrix in combination with the coupling similarity of user context information and user credibility; combining the user trust relationship matrix and the user project matrix to carry out probability decomposition to learn a feature vector; and finally finishing recommendation through trusted users of new users. The invention aims to provide the probability matrix decomposition recommendation method based on user context coupling similarity. The user context information coupling similarity serves as a special trust relationship, a user trust relationship matrix is constructed through the coupling similarity between the user context information and the user trust degree, and then potential feature vectors are learned through probability decomposition in combination with a user scoring matrix, so that recommendation is achieved. The problems of cold start and sparsity which widely occur in a recommendation system are solved.

The invention discloses a text information recommendation method and device and a readable medium, and the method comprises the steps: obtaining user data and a corresponding text corpus, carrying out the preprocessing of the text corpus, and obtaining text data; processing the user data to obtain user attribute features, generating text features from the text data through a word attention network, generating user knowledge graph features from the user data by using a knowledge graph representation learning model, and splicing the user attribute features, the text features and the user knowledge graph features to obtain comprehensive features; and inputting the comprehensive features into a multi-layer perceptron to obtain a user preference category and a corresponding pushing quantity. According to the method, the recognition degree and expansibility of user interests are enhanced, so that the accuracy and diversity of recommendation results are improved, multi-dimensional abstract feature representation can be obtained through comprehensive features, and the problems of data sparsity, cold start and the like in traditional personalized information recommendation are solved. And the attention mechanism is applied to text recommendation, so that the accuracy is further improved.

The invention discloses a network service recommending method based on the positions of mobile users. Preference models are applied to the network service recommending process based on the positions of the mobile users, the similarity of the positions is introduced into the network service recommending selection process and is combined with credibility, and then the network service recommending method based on the positions of the mobile users is formed. By adopting the network service recommending method, the accuracy and the reliability of recommendation of network service are effectively improved, and meanwhile the problems of data sparsity and cold start in the recommending process are simultaneously relieved.

The invention relates to a science and technology resource recommendation information processing method and device. The method comprises the steps of obtaining science and technology information resources and establishing a science and technology resource library; constructing a technical map based on the science and technology information resources in the science and technology resource library and dynamically updating; establishing an interest model of the user based on the interest type input by the user, and updating the interest model of the user according to the browsing history of the user; and pushing recommendation information including science and technology information resources to the user based on the interest model. Compared with the prior art, the recommendation informationis generated based on the technical map and the interest model of the user, an independent database does not need to be established, the technical map system is used for sharing the resource database,and the response speed of the system and the recommendation accuracy are improved.

The invention discloses a time-aware adaptive interest point recommendation method based on K-means clustering. The method comprises the following steps: 1, converting a sign-in data set into a three-dimensional scoring matrix; 2, counting the number of sign-in users, the number of accessed interest points and sign-in times in each time slot, and constructing a three-dimensional sign-in feature vector of each time slot; 3, performing K-means clustering on the time slots, and calculating the time similarity between the time slots in the same cluster; 4, calculating the user similarity at the current time by using the score information in other time slots in the same time cluster; step 5, improving a traditional user-based collaborative filtering method by using a time clustering result and time similarity in a cluster, so that the traditional user-based collaborative filtering method can adaptively generate interest point prediction scores according to current recommendation time; and 6, comparing the recommendation precision of the recommendation system provided by the invention with the recommendation precision of other classic recommendation systems, and evaluating the accuracy and effectiveness of the provided technology.

The invention provides a multivariate information-driven approximate fusion network recommendation propagation method. A propagation algorithm is recommended based on an approximate fusion network, the problems of single information source and data initial stage recommendation quality of a traditional recommendation algorithm can be effectively improved, participants in a network recommendation system are divided into four entity classes and six relationships from three key steps of an approximate fusion network recommendation propagation algorithm, and a probability transfer matrix is determined according to different types of relationships among entities. According to the invention, various types of information such as recommended objects, projects, labels and attributes and relations thereof are effectively fused; the problem of data sparseness and the problem of data initial stage recommendation quality caused by a single information source are relieved, so that recommendation results are more diversified, recommendation accuracy is obviously improved, robustness and robustness are good, calculation complexity is moderate, overall implementation is easy, and the method can be rapidly popularized to network recommendation system application and is high in market practical value.

The invention relates to a chest X-ray film diagnostic report anomaly detection method based on a subject model, and belongs to the technical field of computer natural language processing. The methodcomprises the following steps: firstly, proposing entity extraction of a diagnosis report based on an LSTM-CRF model according to the characteristics of the diagnosis report; performing effective feature extension on the diagnosis report according to domain knowledge and a template to relieve the problem of data sparseness; obtaining instance theme distribution of two diagnosis reports, namely image description and diagnosis conclusion, by utilizing the improved LDA model; and detecting the abnormal diagnosis report by calculating and comparing whether instance theme distribution obtained by the image description entity and the diagnosis conclusion entity is matched or not. The method is high in detection accuracy.