60results about How to "Alleviate cold start problems" patented technology

The invention discloses an association-rule recommending method based on self-adaptive multiple minimum supports. The method comprises the following steps of firstly, establishing a commodity-classifying hierarchical tree according to commodity classification, and classifying concrete commodities according to the classifying hierarchical tree; next respectively carrying out minimum-support threshold-value setting on each concrete commodity and the upper-layer class of a concrete-commodity layer, and then mining frequent item sets and generating rules by utilizing a multiple-minimum-support association-rule expanding algorithm on the basis of the support threshold-value setting, wherein the threshold-value setting relates to the influences of time factors, concrete-commodity price factors and concrete-commodity brand factors; finally generating recommendation for each user by adopting a TOP-N recommending method. When personalized recommendation is made for the user by the association-rule recommending method, the characteristics of different objects can be better embodied by considering the influences of many factors on the multiple-minimum-support threshold-value setting for the concrete commodities and the classes; meanwhile, a data-sparsity problem and a cold-starting problem in a recommending system are relieved, so that the personalized recommendation can be more accurately made for the user.

The invention discloses a TV shopping commodity recommendation method based on a classification algorithm. The TV shopping commodity recommendation method comprises the steps of: converting a prediction problem into a classification problem by utilizing logistic regression and a random forest, namely, predicting that purchasing behaviors of a user about a commodity can be divided into two categories: purchasing and not purchasing; extracting features from commodity information, user information and a user behavior record as input, and taking a prediction score of the user as output, thereby forming a function; and training a model by adopting a linear regression method, and converting the problem into a training classifier problem. The TV shopping commodity recommendation method does not carry out prediction and calculation on the basis of a heuristic rule, but carries out prediction on the basis of data analysis and statistics as well as machine learning for training model; and new users and new commodities can be calculated and predicted quickly as long as the model is trained.

The invention discloses a project recommendation method based on attribute coupled matrix decomposition. The method comprises steps as follows: firstly, giving attribute information of projects and calculating the similarity between the projects with a coupled object similarity measuring index; then learning conceal eigenvectors of users and the projects with a matrix decomposition algorithm, during learning of the conceal eigenvectors of the projects, constructing a regularized term by means of the attribute information of the projects, and constraining the execution process of matrix decomposition, so that the projects with similar attribute information have similar conceal eigenvectors; finally, according to the learned conceal eigenvectors of the users and the projects, projecting scores of projects which are not scored by the users by use of inner products of the conceal eigenvectors of the users and the predicts, and providing personalized project recommendation for the users according to the predicted scores. The problems of similarity calculation of the projects, cold start of project terminals and recommendation accuracy in a recommendation system are solved with the method.

The invention discloses a matrix decomposition recommending method based on an expanded label, comprising: constructing an item-label matrix based on the item label data and calculating the label similarity; constructing a first item pair-label vector based on the item-label matrix; and according to the label similarity, extending the first item pair label-vector as a second item pair-label vector; establishing an item similarity matrix based on the second item pair-label vector; solving a user implicit characteristics matrix and an item implicit characteristics matrix based on the item similarity matrix and the pre-constructed item scoring matrixes; predicting the score by the user to the unrated items; and building a list of recommended items for the user. At the same time, the invention also discloses a matrix decomposition recommending system based on the expanded label. The invention can provide the user with more accurate personalized recommendations when the label is sparse, and relieves the problem of cold starting of the item to a certain extent.

The invention relates to a recommendation prediction method based on attribute information reference self-leaning. Attribution information is utilized for slowing down cold start, the advantages of user and commodity collaborative filtering and model decomposition based on a matrix are combined, the training speed is high, and the interpretability is achieved; meanwhile, on the condition that rating data is sparse, the prediction precision is superior to that of user and commodity collaborative filtering and the matrix-based decomposition method.

The invention relates to a recommendation system cold start method based on multi-arm bandit confidence upper limit. The method comprises steps of: collecting data to build and preprocess a commoditydata set and obtaining a commodity dominant characteristic in a standard format; according to the commodity dominant characteristic, constructing a commodity recessive feature based on a latent Dirichlet algorithm, setting an output commodity recessive feature dimension, and remarking a commodity; creating a candidate commodity set based on the commodity data set: clustering the commodity data setaccording to the commodity recessive feature, clustering the commodities, wherein the commodities in the same cluster have similar properties and the commodities in different clusters have differentproperties, and extracting a commodity from each cluster at random to create a candidate commodity set; view the selection of an optimal commodity from the candidate commodity set as a multi-arm bandit problem, calculating the commodity with the highest score as a recommended product based on a confidence interval upper limit algorithm; after the commodity with the highest score in the candidate commodity set is recommended to a user, updating a user feature and a weight parameter according to the feedback.

The invention discloses a tourist destination recommendation method based on user emotion and time dynamic. The method includes the following steps: firstly, obtaining a quantitative value of emotional inclination of a user by using opinion mining techniques, integrating emotion degree into a matrix decomposition model, and expressing user preferences and tourist destination popularity with the change of time by using a time dynamic mechanism; and then constructing a synthetic recommendation model by combining two elements of user emotion and time influence with an SVD++ method. The user preferences are automatically learned by analyzing explicit or implicit feedback, and so features of tourism leisure resources can match user needs. Compared with a plurality of famous recommendation method, the method of the invention is improved in recommendation accuracy and quality, and a series of experimental evaluations are carried out on publicly available datasets to show that a provided recommendation system is better than a conventional recommendation system.

The invention discloses a method of improving accuracy of a recommendation system. The method of improving the accuracy of the recommendation system comprises the steps of firstly establishing three data subsets, then marking label-free data in each data subset by respectively applying a scoring model based on Gaussian mixture distribution established by the invention onto the established data subsets to obtain each training set, adding the original label-free data marked in the obtained training sets into labeled data in other subsets, iteratively updating labels of the original label-free data in the other subsets, and finally outputting a final recommendation result. The method enables the recommendation system to be capable of effectively resisting scoring noises, and good robustness is obtained and the recommendation accuracy is improved. Besides, the problem of cold start can be effectively relieved.

The invention relates to a recommendation method based on a Mask mechanism and a hierarchical attention mechanism. The recommendation method comprises the steps of obtaining a node data set U and a project data set I; inputting U and I into a context description layer to obtain sequences Su and Si, processing to obtain a node vector and a project vector, fusing the node vector and the project vector to obtain an embedded vector of the ith node u in I, and forming a vector set Z by the embedded vectors of all nodes; calculating a k-head attention coefficient of an L-order neighbor node v closely related to the ith node u, calculating k-head attention of the node u according to an embedded vector of the v, aggregating to obtain an aggregated attention vector of the ith node u, and splicing and linearizing the aggregated attention vector; obtaining a recommendation vector of the ith node u, and forming a vector set Z' by the recommendation vectors of all the nodes; and obtaining a projectrecommendation list of the target node according to Z and / or Z '. According to the method, the recommendation accuracy is improved, the flexibility of the network is improved, the cold start problemis solved, the model is simpler, and the time consumed for recommendation is shorter.

The invention discloses a social recommendation method based on a heterogeneous graph neural network. The method comprises the following steps that: a local heterogeneous graph is constructed according to local data on a client; the client requests model parameters from the server, and performs embedded learning on the local heterogeneous graph by using the graph attention network model so as to process the isomerism of the local graph and the personalized information of the client; a user is associated with a client, after a pseudo item label is added to the client, the gradient of the client is calculated by using a loss function, and then the gradient is uploaded to a server after passing through a local differential privacy model; the server collects gradients of a plurality of clients, and further updates model parameters to train a social recommendation model; and social recommendation is performed through local client embedding output by the social recommendation model. Data storage is dispersed, local user privacy data of the client is comprehensively fused, and social recommendation is cooperatively trained by using the server, so that social recommendation can be effectively realized, and the privacy of the data is protected.

The invention discloses a deep alignment matrix decomposition model-based academic paper recommendation method. The method comprises the following steps of: respectively mapping a user and a paper andthe paper and a word into low-dimensional feature spaces with a same dimensionality through two nonlinear multilayer sensors; building a bridge between the two nonlinear multilayer sensors through maximizing a similarity between two low-dimensional expression vectors of the same paper; and finally transferring information through the paper information bridge and training the two sensors in turn.Through the method, when the sensor of the user-paper part is trained, information of the paper-word part can be utilized, and when the sensor of the paper-word part is trained, information of the user-paper part can be utilized, so that user-paper collection records and paper content text information can be used at the same time so as to jointly make contribution to user paper recommendation.

The invention discloses a hyperspherical collaborative metric recommendation device and method based on a pre-training semantic model, the device comprises a pre-training latent semantic module and acollaborative metric recommendation module, the pre-training latent semantic module comprises an article text information encoder and a decoder, and the collaborative metric recommendation module comprises a hyperspherical mapping module and a fusion loss function module; text information vectorization representation of the article is obtained by an encoder and a decoder; the hyperspherical surface mapping module maps the initialized positive and negative users and the article implicit vectors into the same high-dimensional hyperspherical surface manifold in an angle measurement mode, and thefusion loss function module trains the user and article implicit vectors after hyperspherical surface mapping and optimizes the intra-class and inter-class distances of the positive and negative userarticle sample pairs; and during prediction, a text vector of the article and a corresponding user article implicit vector are obtained while prediction is performed, and an article recommendation result of the user is obtained by calculating a cosine distance between the text vector and the corresponding user article implicit vector.

The invention discloses a family diet recommendation method and device based on demand similarity of multiple targets. The method mainly aims at meeting the health demands of family members and balancing different diet demands of the family members. The method comprises the steps of collecting health information, role information and demand information of all family members; establishing a diet preference model and a family diet demand model according to the information of the family members; and constructing a food recommendation interval time function model, increasing the types of recommended food to enrich the diet structure, and improving the diversity of a recommendation result. In the aspect of recipe recommendation, nutrition information of disease patients and family members is considered, the requirements of the family members in other aspects are also considered, and multi-target recommendation is realized; and in the aspect of the diet structure of the recommended recipe, change of user preferences along with time and the diversity of the recommended recipe are considered, and through a user preference drifting function and a food recommendation interval time function,a certain food loved by the user is prevented from being continuously recommended, and the diet structure is enriched.

The invention relates to a matrix decomposition method and device integrating a convolutional neural network and explicit feedback and electronic equipment. The method comprises the following steps: expressing a user description document of an article as a word vector matrix; mapping the score set of the user to the article into a binary vector; obtaining hidden factors of the article and the userthrough the convolutional attention neural network and the stacked noise reduction automatic encoder respectively; Solving a vector inner product of the hidden factors of the article and the hidden factors of the user to obtain a prediction scoring matrix; and optimizing parameters of the convolutional attention neural network and the stacked noise reduction automatic encoder according to a lossfunction between the prediction scoring matrix and the real scoring matrix. The matrix decomposition method fusing the convolutional neural network and the explicit feedback is good in interpretability, can alleviate the cold start problem of an article, can improve the data sparsity problem, and improves the accuracy of score prediction.

The invention discloses a recommendation method based on a knowledge graph and long and short term interests of a user, which comprises the following steps: acquiring an item set and mapping the item set to a general knowledge graph, and acquiring user-item interaction information; aggregating neighborhood entities by adopting an entity neighborhood aggregation mode based on the knowledge graph convolutional network to obtain project feature vector representation of the to-be-recommended project; learning user long-term preference vector representation through a preference propagation method; inputting interested items in the historical preference set of the user into a gating loop unit for training according to a time sequence to obtain short-term preference vector representation of the user; and overlapping and fusing the user long-term and short-term preference vector representations according to columns, performing full connection layer processing to obtain a final user preference vector representation, performing inner product calculation on the similarity of corresponding feature dimensions with the item feature vector representation of the to-be-recommended item, inputting the similarity into a multi-layer perceptron, and predicting the probability that the user is interested in the to-be-recommended item. The method improves the recommendation effect, and has the characteristics of high interpretability, strong adaptability and high precision.

The invention relates to data analysis, and particularly discloses a product recommendation method, a device, equipment and a storage medium. The method comprises the steps of when an access request of a first user is received, acquiring the historical product information corresponding to the first user; based on the historical product information, determining feature similarities and historical product similarities of a plurality of different first users, the historical product information being stored in a block chain; determining a candidate product information set corresponding to the candidate target user based on the feature similarity and the historical product similarity; and when the product information corresponding to the access record corresponding to the candidate target userexists in the candidate product information set, calculating the interestingness of the candidate target user for each candidate product, and pushing according to the interestingness. When a product is recommended to a user, the cold start problem of common collaborative filtering is effectively relieved.

The invention discloses a MOOC forum post recommendation method integrating forum interaction behaviors and user reading preferences. The MOOC forum post recommendation method comprises the followingsteps: 1) constructing a target scoring matrix, a user interaction scoring matrix and a user interaction frequency matrix of a user to posts; 2) decomposing the user interaction score matrix to obtaina user objective function and a user behavior feature matrix; 3) calculating the number of times of user interaction behaviors, extracting embedded features, and obtaining a user matrix based on theuser behavior feature matrix and the user embedded matrix; 4) extracting post topics through a noise reduction auto-encoder to obtain an object target function, and decomposing the target score matrixto obtain a score matrix target function; and 5) optimizing the scoring matrix objective function, the user objective function and the article objective function, and providing a recommendation listfor the user. Through deep learning and probability-based matrix decomposition, subtle interaction between the users and between the posts and between the users is fused into the model, the cold startproblem is relieved, and accurate recommendation is realized in combination with user reading preferences.

The invention discloses a personalized recommendation method based on probability matrix decomposition. The personalized recommendation method comprises the steps of 1, calculating user attribute similarity; 2, calculating the attention similarity of the user to the label and the score similarity of the user to the label, and integrating the two similarities to generate the preference similarity of the user; 3, proportionally adding the user attribute similarity and the user preference similarity into 1 to generate a final user comprehensive similarity matrix; 4, initializing a user implicit feature matrix and a project implicit feature matrix into normal distribution, and fusing a user comprehensive similarity matrix into the probability matrix decomposition model; and 5, iteratively calculating a user feature matrix and a project feature matrix, generating a prediction scoring matrix of the user for the projects, and recommending the first K projects with the highest scores to the user as final recommendation results by adopting a Top-K sorting method. The method has the advantages that the user cold start problem in the recommendation system is relieved, the accuracy of the recommendation algorithm is improved, and the expandability of the algorithm is enhanced.

The invention discloses a geographic feature-based real estate recommendation method. The method comprises the following steps of: firstly calculating similarities between real estate projects according to real estate project attribute information and building geographic coordinate vectors; constructing geographic feature-included regularization items according to the similarities between the realestate projects; learning implicit feature vectors of a user and the real estate projects by combining the geographic feature-included regularization items and adopting a random gradient descent algorithm on the basis of a weight matrix decomposition model; and finally predicting clicking frequencies, for non-clicked real estate projects, of the users by using an inner product of the implicit feature vectors of the user and the real estate projects, and providing a possible interested real estate information list for the user according to the prediction value. According to the method, the geographic feature information of real estate projects are fused on the basis of a weight matrix decomposition technology, and the execution process of weight matrix decomposition is restrained so that implicit feature vectors of the real estate projects can be learnt more correctly, and the cold starting problem in real estate recommendation is lightened.

The invention discloses an information pushing method and system fusing a knowledge graph structure and path semantics. The method comprises the following steps: constructing a fusion graph based on a user article behavior graph and an article association knowledge graph; mapping the heterogeneous node representation to a unified semantic space; alternately updating the behavior sub-graphs and the article sub-graphs in the fusion graph layer by layer by using a graph neural network; sampling migration between the user and the article node to obtain a path set, and carrying out weighted fusion to generate a path representation; calculating a matching score by combining user and article node representations and path representations between the user and the article node representations to train a recommendation system model; and pushing interested article information to the user based on the recommendation system model. According to the method, the structure information and the path semantics are comprehensively considered to calculate the similarity between the user and the article to push the article information, a more reasonable information fusion means is provided, and the technology can be applied to article information pushing, especially user article representation learning in a sparse scene.

The invention discloses a commodity personalized recommendation method and system based on community discovery and a graph neural network. The method comprises the steps of obtaining the interaction information of a user and a commodity, and constructing a user interaction network with the mapping of the user and the commodity as nodes and the mapping of the interaction information as edges, conducting community discovery in the constructed user interaction network, and endowing attributes of user nodes again according to communities to which users belong, reconstructing a bipartite graph for user nodes with attributes and commodity interaction information, and mapping the user nodes and commodity nodes into vector representation, inputting the user vector representation, the commodity vector representation and the constructed bigraph into an embedded propagation layer for graph representation learning, and optimizing the user vector representation and the commodity vector representation, performing inner product operation on the optimized user vector and the commodity vector, and predicting the probability that the user clicks the commodity according to an inner product result, and obtaining a personalized commodity recommendation sequence for the user. According to the method, the recommended cold start problem can be relieved, and a good recommendation effect is achieved.

The invention discloses a multisource heterogeneous information-fused personalized recommendation method and system. The personalized recommendation method comprises the following steps of: acquiringuser behavior record information and storing the user behavior record information into a database, wherein the user behavior record information comprises but is not limited to various behaviors, for projects, of users, and occurrence times and places of the behaviors; carrying out linear weighting conversion on the behaviors, for the projects, of the users according to different weights, so as toobtain implicit feedback, for the projects, of the users; carrying out interaction matrix construction according to the implicit feedback, for the projects, of the users so as to obtain a user-projectinteraction matrix; and constructing a personalized recommendation model according to the user-project interaction matrix and obtaining corresponding projects recommended to the users. According to the method and system, the data dilution problem and cold start problem in collaborative filtering of data can be effectively relieved, and proper projects can be recommended to users according to specific conditions of the users, so that the friendly using experience of the users can be effectively improved.

The invention provides a graph model-based QoS prediction method applied to cloud service recommendation. The method comprises the steps that a full-graph model containing multi-source information isconstructed, and the full-graph model comprises nodes representing users and services and edges using the similarity between the users, the similarity between the services and the similarity between the users and the services as weights; the full-graph model is segmented into a plurality of sub-graph models; a probability matrix decomposition algorithm for respectively optimizing the full-graph model and the sub-graph model is used for obtaining QoS of global and local prediction; and adaptive fusion processing is carried out on the QoS of global and local prediction to obtain the QoS of finalprediction. According to the method provided by the invention, the influence of multi-source information on QoS is fully considered, and local and global features are adaptively fused, so that the QoS prediction precision is improved. According to the method, the missing QoS value can be accurately predicted, then the sparse matrix is filled, the density of the matrix is improved, and the problemof QoS sparsity in the field of cloud service recommendation is solved to a certain extent.

The invention discloses an academic cooperation relationship prediction method based on a heterogeneous graph neural network, and the method comprises the steps: collecting the work information of an author to construct a work data set and an academic heterogeneous graph, and the nodes in the graph comprise author nodes, work nodes, affiliation nodes and content nodes; using a DeepWalk algorithm and a Text-CNN (Convolutional Neural Network) model to extract information in the learning heterogeneous graph to obtain an embedding vector of an author node; aggregating meta-paths in the learning heterogeneous graph based on author nodes by using a multi-head attention mechanism, and obtaining a long-term interest embedding representation of each author node; obtaining a short-term interest embedding representation of the author based on LSTM and an attention mechanism; and constructing a binary classification model by using a binary classification prediction method, and inputting the work data set into the binary classification model for training to obtain an academic cooperation prediction model. According to the method, the potential cooperation interest of the author in a certain time can be captured, a better recommendation effect is obtained, and the prediction accuracy is improved.

The invention belongs to the technical field of information retrieval and database structures, and discloses a book heat degree correlation detection method based on publishing time and a forgetting curve. The book heat degree correlation detection method based on publishing time and a forgetting curve comprises the following steps: calculating objective heat degree correlation; the forgetting curve calculating the subjective heat; combining subjective heat and objective heat to calculate the correlation of book heat. The method can calculate the heat degree correlation of a book based on thepublishing time, and can alleviate the cold start problem of the book recommendation and improve the performance of the recommendation by combining the collaborative filtering algorithm.

The invention provides a multi-modal news recommendation method and device based on a multi-head self-attention neural mechanism, and the method comprises the steps: collecting data information, including news data, feature data and trace data; fusing the data information into uniform news features based on a multi-component feature cross model of a view-level attention mechanism, a hot news real-time prediction technology of streaming data and a multi-modal information fusion technology of intelligent frame extraction; and taking the unified news features as model input, and completing a personalized accurate recommendation function through a user interest representation model and in combination with a highest future influence strategy. According to the scheme, the user cold start problem is effectively relieved, feature collection and fusion are conducted on multi-modal information in news through multi-modal information fusion, high-order cross feature mining and user interest characterization learning are conducted through a multi-head self-attention mechanism, time sequence weights are given to the news according to the highest future influence strategy and real-time news hot spot mining, and the user interest characterization learning efficiency is improved. And participating in final user recommendation.

The invention discloses an academic paper recommendation method based on network representation and auxiliary information embedding, and the method comprises the following steps: 1, constructing a citation network, carrying out the dimensionality reduction of variables with significant influence in each paper through principal component analysis, obtaining a paper edge weight composed of multiple factors, adding paper edge weights on the basis of the citation network, and constructing a paper influence network; 2, generating a paper sequence in the paper influence network, learning an embedding vector of the paper sequence by an algorithm to obtain a primary graph embedding function of the paper, and adding auxiliary information of the paper into the primary graph embedding function to obtain an ultimate graph embedding function of the paper; 3, taking the ultimate diagram embedding function of the papers as an embedding model, calculating the similarity between each paper and the papers in which the user is interested, and generating a recommendation list. The method has the advantages that the accuracy of the paper recommendation result is high, and the cold start problem of the paper is relieved to a certain extent.

The invention discloses a network service recommending method based on mobile subscriber positions; mobile subscriber positions are applied to a network service recommending process; similarity based on positions can be introduced into a network service recommending selection process, and credibility is combined, thus forming the network service recommending method based on the mobile subscriber positions; the method can effectively improve network service recommending accuracy and reliability, and simultaneously alleviating data sparsity and cold start-up problems possibly existing in the recommending process.

The invention relates to a user portrait construction method for the microblog heterogeneous information and belongs to the field of machine learning and deep learning. The method comprises the following steps of 1) collecting appropriate text and image data, and respectively training an image classification network and a text classification network; 2) using a crawler technology to crawl text and image information of a microblog, using the constructed classification network to obtain a label, namely a user-topic matrix, and using the user-topic matrix as a user portrait; 3) performing fuzzy clustering on user-topic, and discovering a user community; 4) performing keyword discovery on the microblogs classified by the user by using TF-IDF to obtain a topic-item matrix, and calculating associated words by using a knowledge graph thought; and 5) performing similarity user calculation on the community to which the users belong, selecting the users with relatively high similarity, repeating the method in the step 4), and selecting keywords as content supplementary recommendation according to a word frequency threshold. According to the invention, users with high similarity can be recommended to the users.

The invention provides a collaborative filtering recommendation method based on explicit trust and implicit trust, and the method comprises the steps: calculating the prediction scoring accuracy of auser u for a user v according to a user article scoring matrix R; calculating the degree of dependence of the user u on the user v; calculating the implicit trust degree of the user u to the user v according to the accuracy and the dependence degree; filling the user article scoring matrix R according to the implicit trust degree; calculating the scoring confidence of the user u to the article; calculating the similarity between the user u and the user v according to the scoring confidence coefficient and the filled user article scoring matrix; calculating the global trust degree of the user uand the local trust degree of the user u to the user v; calculating the final trust degree of the user u to the user v according to the global trust degree and the local trust degree; correcting a potential characteristic user matrix of the user u; and according to the potential user feature matrix, predicting the score of the user u on the article i. The problems of data sparseness and user coldstart caused by information overload can be solved, and the recommendation accuracy is improved.