56results about How to "Solve data sparsity" patented technology

The invention discloses a Chinese domain term recognition method based on mutual information and a conditional random field model. The Chinese domain term recognition method includes the following steps: (1) gathering domain text corpus and marking all the punctuations, spaces, numbers, ASSCII (American Standard Code for Information Interchange) characters and characters except Chinese characters in the corpus; (2) setting character strings and computing the mutual information values of the character strings, (3) computing the left comentropy and the right comentropy of every character string, (4) defining character string evaluation function, setting evaluation function threshold, computing the evaluation function values of every character string, determining that every character string is a word, comparing in sequence the evaluation function value of the former character with the evaluation function value of the latter character in the character string and segmenting character meaning character strings one by one, (5) utilizing conditional random fields to train a conditional random field model and recognizing domain terms with the conditional random field model. When the Chinese domain term recognition method is used to recognize terms, the data sparsity of legitimate terms is overcome, the amount of calculation of conditional random fields is reduced, and the accuracy of the Chinese domain term recognition is improved.

The application relates to a personalized data searching method and device. The device comprises the following steps: searching a data object according to a query word in a search request of a current user; determining a first behavior characteristic of a historic user in the search access process utilizing the query word according to a historic behavior log, and generating an intention vector of the query word; counting a second behavior characteristic of each user group to the data object according to a user attribute, and generating a preference vector of the user group; computing the similarity of the intention vector of the query word and the preference vector of each user group; using the user group corresponding to the condition that the similarity is greater than a set threshold value as a reference group for determining the intention preference of the current user; adjusting the sorting of the data objects searched by the current user through the query word through adoption of the historic behavior characteristics of the reference group. Thus the performance of a search platform is improved and promoted, the accuracy of a search result output to the user is improved, and the result, which is the most reasonable and the best for the search intention, is output for the user.

The invention discloses a shop recommendation method based on a position of a mobile user. The method comprises sending position information of the user to a server by a mobile phone used by the user, finding all shops related to the position by the server according to the position information sent by the user, finding all users related to the shops, and finally generating a user-shop scoring matrix; and the method further comprises initially filling the obtained user-shop scoring matrix, performing similarity calculation on a target user vector and other user vectors in the obtained matrix, generating neighbor users, performing scoring prediction on an unscored shop of the target use according to the neighbor users, and performing recommendation for the user according to a prediction result. After prediction scores of the unscored shops of the target user are obtained, the recommendation is performed for the user; and all the prediction scores are ranked, and then the shops, which have the high prediction scores and are the former N items, are recommended to the user.

The invention discloses a Mongolian large vocabulary continuous speech recognition method. The method is composed of a preprocessing phase, a preparation phase, a training phase, a decoding phase and a synthesis conversion phase. In the preprocessing phase, text training data is segmented and a pronunciation dictionary is established; in the preparation phase, acoustic features are extracted from input voice signals; in the training phase, an acoustic model is trained by use of a whole-word pronunciation dictionary, and a language model is trained by use of a training text after segmentation; in the decoding phase, by use of the acoustic model, the language model and the pronunciation dictionary, the input acoustic features are recognized into text information; and in the synthesis conversion phase, case suffix errors in a decoding process are corrected by use of rules, stems are merged with case suffixes, and finally, sentences composed of Mongolian words are output. According to the invention, the problems of too long voice recognition time and language model data sparsity in a voice recognition system because the voice recognition system cannot include large-scale Mongolian words and the vocabulary is too large in the prior art are solved.

The invention discloses a self-adaptive identification method and system wherein the system comprises: according to a user's historical corpus, establishing a customized dictionary; clustering the customized words in the user's customized dictionary to obtain the type number of each customized word; constructing language models according to the type numbers of the customized words; during the identification of the inputted information of the user, if the words contained in the information exist in the user's customized dictionary, expanding the de-coding paths according to the type numbers of the customized words corresponding to the words; decoding the information according to the expanded decoding paths to obtain a plurality of selective decoding results; according to the language models, calculating the scores of the language models of the selective decoding results; and selecting the selective decoding result whose language model obtains the highest score as the information identification result. With the method and system of the invention, it is possible to increase the identification accuracy for a user's customized words and the consumption of the system can be reduced.

The invention relates to the technical field of topic detection, and discloses a microblog topic detecting method and system. The method comprises the following steps: S1, segmenting a microblog text into vocabularies; S2, constructing a microblog text clue and a microblog text forest; S3, analyzing a microblog topic aiming at a specific microblog text clue so as to find out the main topic and noise topic in the microblog text clue; S4, combining the microblog text in the main topic aiming at each microblog text clue, thereby generating a microblog clue text for each microblog text clue; and S5, analyzing overall microblog topics to detect an overall microblog topic, thus forming a microblog topic base. The microblog topic detecting method and system can be used for rapidly and accurately detecting the microblog topic, thereby improving the hit rate of the microblog search, shortening the microblog search time of a user and improving the user experience.

The invention discloses a gateway and floating vehicle data fusion method for road real-time speed calculation. The method carries out weight fusion on the real-time road speed calculated based on floating vehicle data and real-time gateway pair speed calculated based on gateway vehicle passing data. A gateway in the invention is a road monitoring point provided with a traffic monitoring device, and is used for collecting the vehicle passing data, wherein the vehicle passing data is information of vehicles passing through the gateway, the information comprising license number of each vehicle, gateway serial number and vehicle passing time; and a gateway pair is the gateways, which are physically adjacent, and can be passed through by vehicles in sequence. The method has the advantages of high accuracy, fast speed and constant real-time performance and the like, and effectively solves the problems of single data source, sparse data and not high accuracy and the like in road real-time speed calculation.

The invention discloses a short-text content classification method. The method includes: obtaining short-text contents of a social network platform; obtaining context emotion feature values and a priori emotion feature values of the short-text contents; using model training to generate a word vector of the short-text contents; utilizing multi-window convolution operations to obtain semantic relationships of the short-text contents of different granularities, and combining pooling operations to abstract semantic representations of the short-text content from different levels; using a bi-directional long-term memory network to obtain a semantic representation of the short-text contents; and combining different-level emotion feature vectors to obtain an output vector, using a function to carry out calculation on the output vector to obtain a value of probability that the short-text contents belong to one or more content categories, and using a content category, of which a value of probability is highest, as the category of the short-text contents. The invention also discloses a short-text content classification system, which can realize the aforementioned short-text content classification method.

The invention is suitable for the technical field of information and provides a method and a device for recommending items. The method comprises the following steps: acquiring item information within a preset distance scope on a present position of a to-be-recommended user; acquiring a candidate user group according to the item information; calculating an interactive behavior similarity and an item scoring similarity between the to-be-recommended user and each candidate user in the candidate user group; calculating a compound similarity between the to-be-recommended user and each candidate user according to the interactive behavior similarity and the item scoring similarity; selecting N candidate users with the maximal compound similarity and forming a neighbor set of the to-be-recommended user; acquiring the item concerned by each candidate user in the neighbor set; and forecasting a score of the item scored by the to-be-recommended user and recommending M items in the highest scores to the to-be-recommended user. According to the invention, the problems of sparse data and cold starting of the present individual recommendation technique can be solved and the item recommendation accuracy is promoted.

The invention relates to the technical field of recommendation systems, in particular to a personalized intelligent clothes matching recommendation method in combination with a knowledge graph, which comprises the following steps: acquiring clothes commodity information and user clothing interaction information of an e-commerce platform, and constructing the knowledge graph; modeling the knowledge graph, and learning each entity and relationship in the knowledge graph to obtain an entity vector and a relationship vector; combining the entity vector and the relationship vector with a hybrid recommendation system, and performing score prediction on the single clothing to obtain a clothes single recommendation result; and calculating a matching index of every two clothes single items, automatically matching the two clothes single items with high matching indexes, performing score prediction, and performing a final TOP-N suit recommendation result according to scores. Online clothes suit recommendation is more intelligent.

The invention discloses a recommendation method based on a user personality label. Development of a user behavior index list, modeling of a user personality-recommendation mode matching rule, development of a recommendation mode list, development of the user personality label, modeling of user online behavior indexes and the like are achieved. According to the invention, a psychological charactertheory is used as a basis; data such as use behaviors and comments of the user on the product are comprehensively analyzed by utilizing a big data technology; according to the method, the personalityof the user is mined, the user behavior model is constructed, a user personality label and a corresponding recommendation mode are developed, related recommendation can be carried out on the user according to a specific application scene, and the method is suitable for recommendation systems of platforms such as e-commerce and social networks. Due to the fact that the internal stability, interpretability and mobility of user characters are high, the cold start problems that system data are sparse and cannot be migrated, and the new user recommendation effect is poor can be effectively solved by utilizing character labels to recommend.

The invention provides a cross-domain recommendation method based on multi-view knowledge representation. The method comprises the steps of integrating different projects in a heterogeneous graph formaccording to similar attributes of the projects in different fields to form a plurality of views, taking the views as inputs of a graph attention network respectively, and obtaining initial knowledgerepresentation of the projects under the views through the graph attention network; taking the initial knowledge representation of the project under each view as the input of a multi-head attention network, obtaining and integrating project representation vectors with user preferences under different views through the multi-head attention network, and obtaining the final representation of the project with the user preferences; and recommending a corresponding project in the target domain to the user according to the final representation of the project with the user preference and the information of the target domain. The multi-view multi-head attention network learning method is set among multiple fields, project knowledge representation is fully learned, cross-field recommendation is carried out, and therefore the recommendation effect of the target field is improved.

Provided are a personalized query word completion recommendation method and device based on a like user model. The method comprises the steps that query word prefixes input by a user are obtained; a set of query words to be completed is obtained according to query logs and the query word prefixes; the frequency scores of the query words to be completed are calculated; the similarity of the user and members of a like user group which the user belongs to and the similarity of the query words to be completed and query words submitted by like user group members are obtained according to the like user model, and the similarity of the query words to be completed and query words submitted by like users is calculated; according to the frequency scores and the similarity of the query words to be completed and query words submitted by like users, the sequence of the query words to be completed is obtained; the query words to be completed are sequenced, and the sequenced query words to be completed are recommended to the users. Accordingly, the problem of data sparsity of single users is solved, the stability of query word recommendation is improved, and user experience is improved.

The invention discloses a Spark-based tour interest recommendation system. The system includes: a data warehouse module, which is used for storing data; a data collection module, which stores the collected data into the data warehouse module; a recommendation engine group module, which extracts the data from the data warehouse module; a result processing module, which unifies results, which are output by the recommendation engine group module, according to weights; an evaluation module, which evaluates each engine of the recommendation engine group module for accuracy and diversity; an engine management module, which dynamically adds or deletes the recommendation engines according to results of the evaluation module, and determining the weight of each engine; and a user feedback processing module, which collects user feedback data of a user interaction interface. The invention also discloses a recommendation method of the above-mentioned recommendation system. Problems that existing recommendation algorithms are low in analysis and calculation efficiency and insufficient in storage space expansibility are solved.

The invention provides a day-ahead transaction strategy method and system based on market supply and demand and regional meteorological prediction, and the method comprises the steps: obtaining regional prediction meteorological data and station prediction meteorological data, and obtaining market supply and demand prediction data and historical power transaction data; carrying out data preprocessing; modeling the preprocessed data by adopting a multi-task learning method in deep learning to obtain a prediction model for electric power transaction market pre-judgment; inputting regional prediction meteorological data, station prediction meteorological data and market supply and demand prediction data of the D day, and predicting market pre-judgment information of power transaction of the Dday through the prediction model; and substituting the historical medium and long term average price, the electric field installed capacity, the short term prediction of the station and the market pre-judgment information into the optimization model, and solving the day-ahead 96-point power declaration and the expected strategy income. The method and system can adapt to scenes of multiple tasks,strong correlation among the tasks can be fully utilized, the prediction accuracy is improved, and the generalization ability of the model is greatly enhanced.

The invention discloses a method of identifying a target terminal. The method includes the following steps: extracting data from a data source, pre-processing the data according to preset strategies to remove abnormal first data, and keeping normal second data used for obtaining an analysis data set; performing data verification and/or data conversion on the second data and then obtaining an analysis data set; obtaining the analysis data set, and extracting characteristic vectors of users from the analysis data set on the basis of communication characteristics of the users, wherein the characteristic vectors of the users is used for representing the communication characteristics of the users; dividing all the users into first users and second users on the basis of the characteristic vectors of the users, obtaining cluster results corresponding to all the first users on the basis of data of the first users, and taking the cluster results as first cluster results; performing clustering on the basis of the characteristic vectors of the second users and the first cluster results, obtaining cluster results corresponding to all the users, and taking the cluster results as second clusterresults; and identifying a target terminal on the basis of the second cluster results. The invention also discloses a device of identifying a target terminal.

The invention discloses a text detection and correction method based on Pinyin similarity and a language model. The method comprises the steps: collecting a large number of correct instruction text statements to serve as training statements; selecting words of a professional field from the training statements, and constructing a custom dictionary; carrying out word segmentation on the training statements by utilizing a HanLP language processing toolkit and a custom dictionary; counting the occurrence frequency of each word and each word combination in the word segmentation result in all the training statements, and constructing a Bi-Gram language model; converting the to-be-corrected statement into corresponding to-be-corrected pinyin, and converting words of the custom dictionary into corresponding dictionary pinyin; and correcting the to-be-corrected statement according to the pinyin similarity between the to-be-corrected pinyin and the dictionary pinyin in combination with the sentence rationality of the to-be-corrected statement to obtain a corrected statement. Through word pinyin similarity calculation and sentence rationality analysis, semantic information and contexts of sentences are considered, wrong words in the sentences can be detected, and the correction accuracy is improved.

The invention discloses a POI recommendation method based on heterogeneous network, as to the technical field recommend by the POI, a new probability generation model of comment sentences is established, this model can deal with heterogeneous data across websites, and consider the influence of geographic region division, user community division, review content and user interaction behavior on POIprediction results, and solve the problem of data sparseness and low user trust in the recommendation process, which can significantly improve the accuracy of POI recommendation.

According to the Chinese medical question classification system based on deep encyclopedia learning, by using a semantic structure of Chinese search encyclopedia in combination with a deep learning method, a method for constructing a feature vector more efficiently and accurately is provided, which comprises: using a semantic association degree efficient convergence method based on the semantic structure of the Chinese search encyclopedia for constructing a network inquiry question feature vector; based on the features of the medical questions, improving a semantic association degree algorithm, solving the defect that the speed is low when feature vectors are constructed, and expanding feature words by extracting Chinese search encyclopedia word links; on the basis of a distributed Chinese word vector space of a CB-CBS language model, achieving efficient dimensionality reduction of network inquiry question feature vectors, avoiding the problem of data sparseness, greatly improving the inquiry classification efficiency; and using the CB-CBS model in combination with Chinese search encyclopedia and deep learning to construct distributed medical question word vectors, constructing a professional medical question corpus, and improving the accuracy of the word association degree and the medical question classification efficiency remarkably.

The embodiment of the invention relates to the field of machine learning, in particular to a federated learning multi-party security computing method and a device, and aims to improve the model training efficiency and accuracy on the basis of protecting the data security in a multi-party computing process. The method comprises the steps that adopting a data node to divide all individual objects inthe data node into a plurality of object sets according to classification standards; for each object set, adopting the data node to determine feature data of the object set according to the feature data of all the individual objects in the object set; adopting the data node to send the feature data of the object set to a model node, so that the model node performs sample alignment on all the feature data of the same object set according to the feature data of the object set sent by the plurality of data nodes to obtain sample data of the object set, training the federated learning model according to the sample data of all the object sets; wherein the classification standards for dividing the individual objects in the plurality of data nodes are the same.

The invention provides a road condition determination method and device, a medium and electronic equipment. The method comprises the following steps: determining a shunting path of a traffic marking road section and at least two associated road sections, and obtaining vehicle driving information; determining at least two pieces of marking road condition information according to the vehicle drivinginformation, and dividing the vehicle driving information according to the at least two pieces of marking road condition information to obtain at least two road condition sets; calculating the vehicle driving information according to the at least two road condition sets to obtain at least two sets of driving information; determining at least two pieces of associated road condition information andat least two pieces of associated driving information of the at least two associated road sections; and according to the at least two pieces of associated road condition information, the at least twopieces of associated driving information, the at least two pieces of marking road condition information and the at least two pieces of set driving information, determining shunting road condition information of the shunting path in the at least two pieces of marking road condition information. The calculation accuracy of the road condition information is improved, and the service experience of using the electronic map by a user is improved.

The invention provides an English lexical analysis method for neural network machine translation. The English lexical analysis method comprises the following steps: performing English word segmentation on a to-be-processed English text; carrying out query screening on each word obtained after English word segmentation by utilizing a special vocabulary; reserving result information of the queried words in a lexical analysis result; performing rule processing on the words which are not queried; carrying out word architecture restoration on the words meeting the rule processing conditions, and directly storing the words which do not meet the rule processing conditions in a lexical analysis result; and outputting a lexical analysis result, and adding the lexical analysis result into machine deep learning training. The invention further provides an English lexical analysis system for neural network machine translation. According to the English lexical analysis method for neural network machine translation, the problems that the machine learning efficiency is reduced and the translation quality is poor due to the fact that training corpus data of neural network machine translation is sparse can be solved.

The invention relates to a method for generating a personalized page. The method comprises the following steps of: recording user behavior data of a user in a first Internet application program and a second Internet application program; analyzing interested user distribution features of a target object in the first Internet application program and the second Internet application program according to the user behavior data; calculating a correlation coefficient between an object in the first Internet application program and an object in the second Internet application program according to the interested user distribution features; obtaining a first interested object of the first user in the first Internet application program when the condition that the first user logs in the second Internet application program is detected; objecting a second interested object from the second Internet application program, wherein the correlation coefficient between the first interested object and the second interested object is greater than a preset value; and generating the personalized page of the first user according to the second interested object. The method can solve the new user data sparsity in the personalized page generating process.