Information recommendation method, system and device

Abstract

The present invention discloses an information recommendation method, system and device, the method comprising: obtaining relevant data to form a data set, preprocessing the data set to obtain a processed data set; forming a cross feature set based on a user feature set and information text labels; constructing an information recommendation pre-training model, training the information recommendation pre-training model based on the user feature set, the cross feature set and the sequence data set to obtain an information recommendation model; the information recommendation model infers and predicts the target related data to obtain an information recommendation result. The present invention forms a cross feature set based on user related data and information related data, and then trains and generates an information recommendation model, thereby improving the accuracy and diversity of the information recommendation model from different dimensions and directions. In addition, due to the factual update of the data and the breadth of the data, the accuracy in the reasoning process is improved.

Description

Technical Field

[0001] The present invention relates to the field of algorithm technology, and in particular to an information recommendation method, system and device. Background Art

[0002] Up to now, the securities information recommendation method is mainly based on the user-information correlation model constructed based on user attributes, information attributes and user behavior attributes to recommend information IDs. The user behavior attributes mentioned in this model are either the user's behavior attributes in the information dimension or the user's behavior attributes in the individual stock dimension. These attributes only consider the user's single-dimensional behavior attributes or simply calculate user preferences based on different behavior attributes, and cannot better realize the personalized distribution of securities information based on the multi-dimensional attributes of the securities industry. Furthermore, the existing securities information recommendation method does not take into account the particularity of the new and easily missing features of securities information effectiveness. Therefore, it is not comprehensive enough in fitting the reasoning and scoring of new information when ranking in the model, which will lead to inaccurate reasoning of new information due to missing features.

[0003] In other words, simply relying on one-dimensional behavioral attributes or simply superimposing different behavioral attributes to calculate user preferences is difficult to accurately capture users' multi-dimensional and complex preferences in the field of securities information. Ignoring the depth and breadth of user preferences will lead to a large deviation between the recommendation results and the actual needs of users. Due to the timeliness of securities information, features are easily missing, which in turn affects the accuracy of reasoning. Summary of the invention

[0004] The present invention aims at the shortcomings of the prior art and provides an information recommendation method, system and device.

[0005] In order to solve the above technical problems, the present invention is solved by the following technical solutions:

[0006] An information recommendation method comprises the following steps:

[0007] Acquire relevant data to form a data set, and preprocess the data set to obtain a processed data set, wherein the relevant data includes user-related data and information-related data;

[0008] A user feature set is obtained based on relevant data, an information feature set and information text labels are obtained based on information-related data, and a cross feature set is formed based on the user feature set and the information text labels;

[0009] Constructing an information recommendation pre-training model, which includes an input layer, a shared feature structure layer, a main structure layer and an output layer. The input layer receives a user feature set, a cross feature set and an information feature set. The shared feature structure layer processes the user feature set, the cross feature set and the information feature set to obtain corresponding vector representations. The main structure layer processes the vector representation corresponding to the information feature set to obtain a user information vector representation. The output layer fuses the user vector representation, the cross vector representation and the user information vector representation to obtain a recommendation score.

[0010] The information recommendation pre-training model is trained based on the cross feature set and the sequence data set to obtain the information recommendation model;

[0011] The information recommendation model performs inference and prediction on target-related data to obtain information recommendation results.

[0012] As an implementable method, obtaining the information text label based on the information-related data includes the following steps:

[0013] The information-related data is information text information, and the information text information includes a title, an abstract, and text information, and an information text information set is formed based on the information text information;

[0014] Performing word segmentation on the information text information in the information text information set to obtain word units and then form a word unit set;

[0015] Build a graph model, use word units as word units in the graph model, and assign corresponding weights to the edges in the graph model according to word unit similarity;

[0016] Based on the corresponding weights, the importance score of each word is calculated, and the words corresponding to the first few words of the importance score are used as keywords;

[0017] Identify information text information through keywords and in combination with the tag system to obtain information text tags;

[0018] The importance score of each word is obtained by:

[0019]

[0020] in, Representation word The importance score of represents the set of all words, Representation word To word The shortest path of Represents the weight of the edge.

[0021] As an implementable method, forming a cross feature set based on a user feature set and information text tags includes the following steps:

[0022] The user characteristics include at least one or more of user attribute characteristics, user device characteristics, user activity characteristics, user information preference characteristics, and user stock preference characteristics;

[0023] Based on user characteristics, the user's preference for information text labels is obtained, and then the preference score is obtained;

[0024] Determine the cross-features of user features and information text labels based on the preference scores;

[0025] The preference score is obtained in the following way:

[0026]

[0027] in, Indicates user Text label for information The preference score of Indicates the adjustment item. Indicates Weight coefficient, A function mapping representing user features, Function mapping representing information text labels, Indicates the number of user features.

[0028] As an implementable method, the information-related data further includes sequence data, and the sequence data is processed to obtain sequence features, specifically:

[0029] The sequence data includes information sequence data and individual stock sequence data. The information sequence data is information click records and / or information browsing records within a certain time period, and the individual stock sequence data is holding individual stock records and / or self-selected individual stock records within a certain time period;

[0030] Feature extraction is performed on the information sequence data to obtain information sequence features; feature extraction is performed on the individual stock sequence data to obtain individual stock sequence features, and the information sequence features and individual stock sequence features are processed by the shared feature structure layer to obtain information sequence features and individual stock sequence features.

[0031] As an implementable method, the following steps are also included:

[0032] Determine whether there is an information feature corresponding to the target information-related data in the target-related data in the information feature set;

[0033] If it does not exist, search for information-related data similar to the target information-related data in the information-related data;

[0034] The information features corresponding to similar information-related data are averaged and pooled to obtain the information features of the target information-related data.

[0035] As an implementable method, searching the information-related data for information-related data that is similar to the target information-related data includes the following steps:

[0036] Representing the information-related data and the target information-related data as a first information TF-IDF vector and a second information TF-IDF vector respectively;

[0037] Calculate the similarity between the TF-IDF vector of the first information and the TF-IDF vector of the second information based on the cosine similarity model to obtain a similarity score between the information, and determine similar information to the target information based on the similarity score;

[0038] The information TF-IDF vector is expressed as follows:

[0039]

[0040] The cosine similarity model is expressed as follows:

[0041]

[0042] in, represents the information TF-IDF vector, represents the similarity score, Represents a word, represents the corresponding score, Indicates Information TF-IDF vectors, Indicates Information TF-IDF vector.

[0043] As an implementable method, the main structure layer processes the vector representation corresponding to the information feature set to obtain the user information vector representation, including the following steps:

[0044] The information sequence features, information vector representation and individual stock sequence features are processed through the attention mechanism to obtain the user information vector representation, and then the user portrait features are obtained;

[0045] The attention mechanism is expressed as follows:

[0046]

[0047]

[0048]

[0049] in, represents the user information vector representation, is the target information vector representation, Represents the i-th information vector representation.

[0050] As an implementation method, the output layer fuses the user vector representation, the cross vector representation and the user information vector representation to obtain a recommendation score, including the following steps:

[0051] At least three fully connected layers are set, and nonlinear functions are used as activation functions between each fully connected layer to fuse the user vector representation, cross vector representation and user information vector representation to obtain the output result;

[0052] Convert the output of the fully connected layer into a score value based on the sigmoid function;

[0053] The actual value and the predicted value are evaluated based on the loss function until the loss result is minimized, and then the predicted result is obtained;

[0054] The sigmoid function is expressed as follows:

[0055]

[0056] The loss function is expressed as follows:

[0057]

[0058] in, Indicates the actual label of the user for the information data. Indicates the user's predicted label for news data.

[0059] An information recommendation system, comprising:

[0060] A data acquisition module, used to acquire relevant data to form a data set, and pre-process the data set to obtain a processed data set, wherein the relevant data includes user-related data and information-related data;

[0061] A feature extraction module obtains a user feature set based on relevant data, obtains an information feature set and information text labels based on information-related data, and forms a cross feature set based on the user feature set and the information text labels;

[0062] A model building module is used to build an information recommendation pre-training model. The information recommendation pre-training model includes an input layer, a shared feature structure layer, a main structure layer and an output layer. The input layer receives a user feature set, a cross feature set and an information feature set. The shared feature structure layer processes the user feature set, the cross feature set and the information feature set to obtain corresponding vector representations. The main structure layer processes the vector representation corresponding to the information feature set to obtain a user information vector representation. The output layer fuses the user vector representation, the cross vector representation and the user information vector representation to obtain a recommendation score.

[0063] Model training module, which trains the information recommendation pre-training model based on the cross feature set and sequence data set to obtain the information recommendation model;

[0064] The result recommendation module is used for the information recommendation model to perform inference and prediction on target-related data to obtain information recommendation results.

[0065] As an implementable method, the feature extraction module is configured as follows:

[0066] The information-related data is information text information, and the information text information includes a title, an abstract, and text information, and an information text information set is formed based on the information text information;

[0067] Performing word segmentation on the information text information in the information text information set to obtain word units and then form a word unit set;

[0068] Build a graph model, use word units as word units in the graph model, and assign corresponding weights to the edges in the graph model according to word unit similarity;

[0069] Based on the corresponding weights, the importance score of each word is calculated, and the words corresponding to the first few words of the importance score are used as keywords;

[0070] Identify information text information through keywords and in combination with the tag system to obtain information text tags;

[0071] The importance score of each word is obtained by:

[0072]

[0073] in, Representation word The importance score of represents the set of all words, Representation word To word The shortest path of Represents the weight of the edge.

[0074] As an implementable method, the feature extraction module is configured as follows:

[0075] The user characteristics include at least one or more of user attribute characteristics, user device characteristics, user activity characteristics, user information preference characteristics, and user stock preference characteristics;

[0076] Based on user characteristics, the user's preference for information text labels is obtained, and then the preference score is obtained;

[0077] Determine the cross-features of user features and information text labels based on the preference scores;

[0078] The preference score is obtained in the following way:

[0079]

[0080] in, Indicates user Text label for information The preference score of Represents an adjustment item. Indicates Weight coefficient, Function mapping representing user features, Function mapping representing information text labels, Indicates the number of user features.

[0081] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the following method is implemented:

[0082] Acquire relevant data to form a data set, and preprocess the data set to obtain a processed data set, wherein the relevant data includes user-related data and information-related data;

[0083] A user feature set is obtained based on relevant data, an information feature set and information text labels are obtained based on information-related data, and a cross feature set is formed based on the user feature set and the information text labels;

[0084] Constructing an information recommendation pre-training model, which includes an input layer, a shared feature structure layer, a main structure layer and an output layer. The input layer receives a user feature set, a cross feature set and an information feature set. The shared feature structure layer processes the user feature set, the cross feature set and the information feature set to obtain corresponding vector representations. The main structure layer processes the vector representation corresponding to the information feature set to obtain a user information vector representation. The output layer fuses the user vector representation, the cross vector representation and the user information vector representation to obtain a recommendation score.

[0085] The information recommendation pre-training model is trained based on the cross feature set and the sequence data set to obtain the information recommendation model;

[0086] The information recommendation model performs inference and prediction on target-related data to obtain information recommendation results.

[0087] An information recommendation device includes a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor implements the following method when executing the computer program:

[0088] Acquire relevant data to form a data set, and preprocess the data set to obtain a processed data set, wherein the relevant data includes user-related data and information-related data;

[0089] A user feature set is obtained based on relevant data, an information feature set and information text labels are obtained based on information-related data, and a cross feature set is formed based on the user feature set and the information text labels;

[0090] Constructing an information recommendation pre-training model, which includes an input layer, a shared feature structure layer, a main structure layer and an output layer. The input layer receives a user feature set, a cross feature set and an information feature set. The shared feature structure layer processes the user feature set, the cross feature set and the information feature set to obtain corresponding vector representations. The main structure layer processes the vector representation corresponding to the information feature set to obtain a user information vector representation. The output layer fuses the user vector representation, the cross vector representation and the user information vector representation to obtain a recommendation score.

[0091] The information recommendation pre-training model is trained based on the cross feature set and the sequence data set to obtain the information recommendation model;

[0092] The information recommendation model performs inference and prediction on target-related data to obtain information recommendation results.

[0093] The present invention has significant technical effects due to the adoption of the above technical solution:

[0094] The present invention forms a cross-feature set based on user-related data and information-related data, and then trains and generates an information recommendation model, thereby improving the accuracy and diversity of the information recommendation model from different dimensions and directions. In addition, due to the factual update of the data and the breadth of the data, the accuracy of the reasoning process is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0095] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0096] Figure 1 It is a schematic flow diagram of the method of the present invention;

[0097] Figure 2 It is a structural schematic diagram of the system of the present invention;

[0098] Figure 3 It is a structural diagram of the information recommendation model of the present invention;

[0099] Figure 4 It is a schematic diagram of the process of obtaining information text labels of the present invention;

[0100] Figure 5 It is a flow chart of another embodiment of the present invention. DETAILED DESCRIPTION

[0101] The present invention is further described in detail below in conjunction with embodiments. The following embodiments are for explanation of the present invention but the present invention is not limited to the following embodiments.

[0102] Embodiment 1:

[0103] An information recommendation method, such as Figure 1 As shown, the following steps are included:

[0104] S100, acquiring relevant data to form a data set, and preprocessing the data set to obtain a processed data set, wherein the relevant data includes user-related data and information-related data;

[0105] S200, obtaining a user feature set based on relevant data, obtaining an information feature set and information text labels based on information-related data, and forming a cross feature set based on the user feature set and the information text labels;

[0106] S300, constructing an information recommendation pre-training model, which includes an input layer, a shared feature structure layer, a main structure layer and an output layer. The input layer receives a user feature set, a cross feature set and an information feature set. The shared feature structure layer processes the user feature set, the cross feature set and the information feature set to obtain corresponding vector representations. The main structure layer processes the vector representation corresponding to the information feature set to obtain a user information vector representation. The output layer fuses the user vector representation, the cross vector representation and the user information vector representation to obtain a recommendation score.

[0107] S400, training the information recommendation pre-training model based on the cross feature set and the sequence data set to obtain an information recommendation model;

[0108] S500: The information recommendation model performs inference and prediction on target-related data to obtain information recommendation results.

[0109] The present invention forms a cross-feature set based on user-related data and information-related data, and then trains and generates an information recommendation model, thereby improving the accuracy and diversity of the information recommendation model from different dimensions and directions. In addition, due to the factual update of the data and the breadth of the data, the accuracy of the reasoning process is improved. After the output layer outputs the recommendation score, the recommendation score will be sorted, and the information-related data corresponding to the top ranking will be used as the recommendation result.

[0110] The relevant data of the present invention has a relatively wide range. It is precisely because of the wide range of relevant data that the accuracy and diversity of the information recommendation model can be improved. The relevant data covers different data sources, and key features are extracted from different data sources, so that it can cover multiple dimensions such as users, information, individual stock behavior sequences and information behavior sequences, and then cross-features are formed. The information recommendation pre-training model is trained based on the cross-feature set and the sequence data set to obtain the information recommendation model. During the training process, it is also ensured that the information recommendation model can accurately capture the inherent laws of the data. Then, through the information recommendation model, reasoning is performed to obtain the information recommendation results, which realizes efficient and accurate information distribution, provides users with a more personalized and intelligent information service experience, and significantly improves the efficiency and accuracy of information distribution.

[0111] The relevant data used in the present invention include user-related data and information-related data. The user-related data include user data and behavior data. The behavior data include user individual stock behavior data and user information behavior data. The information-related data include information data and sequence data. The sequence data includes information sequence data and individual stock sequence data. The information sequence data is the information click record and / or information browsing record within a certain time period, and the individual stock sequence data is the holding individual stock record and / or self-selected individual stock record within a certain time period. A user feature set is obtained based on the relevant data, an information feature set and information text labels are obtained based on the information-related data, and a cross feature set is formed based on the user feature set and the information text labels.

[0112] The cross-feature set also includes sequence features, which are obtained by extracting features from information sequence data; extracting features from individual stock sequence data to obtain individual stock sequence features, and processing the information sequence features and individual stock sequence features at the shared feature structure layer to obtain information sequence features and individual stock sequence features.

[0113] In one embodiment, user characteristics include at least one or more of user attribute characteristics, user device characteristics, user activity characteristics, user information preference characteristics, and user stock preference characteristics. User characteristics mainly include user basic attributes and user tag preferences. User attributes include basic information such as user age, region, gender, and user self-selected stock data, user position data, etc. User device characteristics include device operating system and device operator, etc. User activity characteristics include platform activity, scene activity, and user account level, etc. User information preference characteristics are calculated based on user clicks, browsing information, and other behaviors. User stock preference characteristics are calculated based on user positions and self-selected stocks. Information characteristics include information attribute characteristics, such as tags, release time, source, etc.

[0114] Furthermore, in step S200, a cross feature set is formed based on the user feature set and the information text label, including the following steps:

[0115] Based on user characteristics, the user's preference for information text labels is obtained, and then the preference score is obtained;

[0116] Determine the cross-features of user features and information text labels based on the preference scores;

[0117] The preference score is obtained in the following way:

[0118]

[0119] in, Indicates user Text label for information The preference score of Represents an adjustment item. Indicates Weight coefficient, Function mapping representing user features, Function mapping representing information text labels, Indicates the number of user features.

[0120] That is to say, the cross-feature is determined by the user's preference for information text tags. In other words, the user has a certain preference for information text tags and has different preference scores for different information text tags. The user's preference score for information text tags in the information text tag set is considered to be the cross-feature of the user and the information text tag.

[0121] Specifically, cross-features can be obtained in the following ways: Data collection: Collect historical user data, such as clicks, browsing, collections, sharing, etc. Feature extraction: Extract features from user data, including basic user attributes (such as age, gender, occupation) and behavioral characteristics (such as active time, preferred topics, etc.). Association analysis between features and tags: Analyze the relationship between user features and information text tags, and determine which user features are related to information text tag preferences. Preference degree model construction: Use machine learning algorithms (such as logistic regression, random forest, gradient boosting tree, etc.) to build a model to learn the relationship between user features and tag preferences. Preference degree score calculation: Use the constructed model to calculate the user's preference score for each tag. Cross-feature determination: According to the preference score, determine the cross-features of user features and information text tags, that is, those features that have the greatest impact on the preference score. The above modeling process and training process can be implemented with the help of existing technologies and will not be repeated here.

[0122] The information text label is obtained based on the information-related data, including the following steps:

[0123] The information-related data is information text information, and the information text information includes a title, an abstract, and text information, and an information text information set is formed based on the information text information;

[0124] Performing word segmentation on the information text information in the information text information set to obtain word units and then form a word unit set;

[0125] Build a graph model, use word units as word units in the graph model, and assign corresponding weights to the edges in the graph model according to word unit similarity;

[0126] Based on the corresponding weights, the importance score of each word is calculated, and the words corresponding to the first few words of the importance score are used as keywords;

[0127] Identify information text information through keywords and in combination with the tag system to obtain information text tags;

[0128] The importance score of each word is obtained by:

[0129]

[0130] in, Representation word The importance score of represents the set of all words, Representation word To word The shortest path of Represents the weight of the edge.

[0131] The generation process of the information text label in this embodiment can be found in the attached Figure 4 As shown, the information text information is obtained, and the full text is segmented based on the securities industry vocabulary to obtain the relevant word units corresponding to the securities industry. After obtaining the word units, filtering and cleaning operations will be performed first to remove stop words / meaningless words, and then combined with the graph model to obtain keywords, and combined with the securities industry label system to perform label feature recognition, and of course there is also a label weight normalization operation to obtain the information text label.

[0132] A graph model is a mathematical structure used to represent the relationship between entities. In the field of machine learning and artificial intelligence, graph models are particularly important because they can capture complex relationships and interactions. A node in a graph model represents an entity. In this application, each word is used as a node in the graph model. The line connecting two nodes is called an edge in the graph model, which represents the relationship between entities. By combining the graph model, we can get the relationship between word units and then determine the keywords.

[0133] Here, we will not emphasize the calculation method of the weight of each edge. The weight is assigned according to the similarity, that is, it can be directly assigned, and the similarity can be calculated by cosine similarity or other similarities. The shortest path can be determined by the following methods:

[0134] Step 1: Set the shortest path estimate of the starting token to 0 and the estimates of all other tokens to infinity;

[0135] Step 2: Create a set of unvisited tokens and a priority queue (sorted by path estimate);

[0136] Step 3: Starting from the starting point, add the starting point to the priority queue;

[0137] Step 4: When the priority queue is not empty, perform the following operations: 1. Take the token with the smallest path estimate from the queue; 2. Traverse all neighbors of the token; 3. For each neighbor, calculate the new path length from the current token to the neighbor; 4. If the new path length is less than the neighbor's current path estimate, update the neighbor's path estimate and add it to the priority queue;

[0138] Repeat step 4 until the target word is reached or the priority queue is empty, and the shortest path is obtained.

[0139] In this embodiment, the tag system mainly includes 9 tag system dimensions, which can be preset, including the theme tag system, industry tag system, news tag system, individual stock tag system, geographical area tag system, person tag system, organization tag system, fund tag system, and convertible bond attribute tag system. There are corresponding tags under each tag system. For example, the individual stock tag system contains corresponding individual stock tags (the individual stock tags here can be represented by the company's name or abbreviation), such as "#xx Securities#", "#xxx An#", etc.

[0140] In addition, due to the time-sensitive particularity of securities information, when the data related to the target information arrives, the corresponding information features may not be found. Therefore, the following steps are used for judgment and search. For example:

[0141] Judge whether there is an information feature corresponding to the target information related data in the information feature set;

[0142] If not, search for information related data similar to the target information related data in the information related data;

[0143] Perform average pooling operation on the information features corresponding to the similar information related data to obtain the information features of the target information related data.

[0144] And the step of searching for information related data similar to the target information related data in the information related data includes the following steps:

[0145] Characterize the information related data and the target information related data as the first information TF-IDF vector and the second information TF-IDF vector respectively;

[0146] Calculate the similarity between the first information TF-IDF vector and the second information TF-IDF vector based on the cosine similarity model to obtain the information similarity score, and judge the similar information of the target information based on the similarity score;

[0147] The information TF-IDF vector is expressed as follows:

[0148]

[0149] The cosine similarity model is expressed as follows:

[0150]

[0151] Among them, represents the information TF-IDF vector, represents the similarity score, represents the token, represents the corresponding score, represents the Information TF-IDF vectors, Indicates Information TF-IDF vector.

[0152] TF-IDF is a commonly used text processing technology to evaluate the importance of a word to an article or an article in a corpus. TF stands for Term Frequency and IDF stands for Inverse Document Frequency. Here, the TF-IDF vector of the information is obtained with the help of TF-IDF technology, and then a series of subsequent operations are performed.

[0153] The whole process is like Figure 5 As shown, if there is no information feature corresponding to the target information-related data in the target-related data, information-related data similar to the target information-related data is searched in the information-related data, and the information features corresponding to the similar information-related data are averaged pooled to obtain the information features of the target information-related data.

[0154] In addition, since the content and data involved in the user's understanding of information are huge, it is necessary to process the attention mechanism to enrich the vector representation. The main structure layer processes the vector representation corresponding to the information feature set to obtain the user information vector representation, including the following steps:

[0155] The information sequence features, information vector representation and individual stock sequence features are processed through the attention mechanism to obtain the user information vector representation, and then the user portrait features are obtained;

[0156] The attention mechanism is expressed as follows:

[0157]

[0158]

[0159]

[0160] in, represents the user information vector representation, is the target information vector representation, Represents the i-th information vector representation.

[0161] In step S300, the output layer fuses the user vector representation, the cross vector representation and the user information vector representation to obtain a recommendation score, including the following steps:

[0162] At least three fully connected layers are set, and nonlinear functions are used as activation functions between each fully connected layer to fuse the user vector representation, the cross vector representation and the user information vector representation to obtain the output result. The fusion here can refer to the relevant fusion of the technology in this field, which will not be repeated here;

[0163] The output of the fully connected layer is converted into a score value based on the sigmoid function; this score value is the recommendation score, and the recommendation scores are sorted to obtain the top rankings of the recommendation scores, and the information-related data corresponding to the top rankings are used as the recommendation results;

[0164] The actual value and the predicted value are evaluated based on the loss function until the loss result is minimized, and then the predicted result is obtained;

[0165] The sigmoid function is expressed as follows:

[0166]

[0167] The loss function is expressed as follows:

[0168]

[0169] in, Indicates the actual label of the user for the information data. Indicates the user's predicted label for news data.

[0170] In one embodiment, the architecture of the information recommendation model is as follows: Figure 3As shown, the information recommendation model or the constructed information recommendation pre-trained model includes an input layer, a shared feature structure layer, a main structure layer and an output layer. The input layer receives a user feature set, a cross feature set and an information feature set. The shared feature structure layer processes the user feature set, the cross feature set and the information feature set respectively to obtain corresponding vector representations. The main structure layer processes the vector representation corresponding to the information feature set to obtain a user information vector representation. The output layer fuses the user vector representation, the cross vector representation and the user information vector representation to obtain a recommendation score. In combination with information, stocks and other related data, the following detailed description can be made: when training the model, the input layer actually receives various related data. This application uses the pre-processing of the data as a pre-processing for model training, and obtains a cross feature set through pre-processing. The cross feature set actually also contains various features, and the cross feature set is used as a training set for the information recommendation pre-trained model. The shared feature structure layer may actually include an embedding layer, a convolutional layer, a recurrent layer or other feature extraction methods, which are used to convert the original features into vector form, and process the user feature set, the cross feature set and the information feature set respectively to obtain corresponding vector representations. The vector representation is actually equivalent to obtaining user portrait features. The goal of the main structure layer is to learn the preference representation of users and information. The obtained preference vector representations in multiple dimensions reflect the user's potential interest in different information text tags. The output layer will fuse the preference vector representations and finally output the recommendation results.

[0171] Of course, it will also include an evaluation and iteration process, using appropriate evaluation indicators such as accuracy, recall, or AUC to evaluate the performance of the model, and performing iterative optimization based on the evaluation results.

[0172] Embodiment 2:

[0173] An information recommendation system, such as Figure 2 As shown, including:

[0174] The data acquisition module 100 is used to acquire relevant data to form a data set, and pre-process the data set to obtain a processed data set, wherein the relevant data includes user-related data and information-related data;

[0175] The feature extraction module 200 obtains a user feature set based on the relevant data, obtains an information feature set and information text labels based on the information-related data, and forms a cross feature set based on the user feature set and the information text labels;

[0176] The model building module 300 is used to build an information recommendation pre-training model. The information recommendation pre-training model includes an input layer, a shared feature structure layer, a main structure layer and an output layer. The input layer receives a user feature set, a cross feature set and an information feature set. The shared feature structure layer processes the user feature set, the cross feature set and the information feature set to obtain corresponding vector representations. The main structure layer processes the vector representation corresponding to the information feature set to obtain a user information vector representation. The output layer fuses the user vector representation, the cross vector representation and the user information vector representation to obtain a recommendation score.

[0177] Model training module 400, training the information recommendation pre-training model based on the cross feature set and the sequence data set to obtain the information recommendation model;

[0178] The result recommendation module 500 is used for the information recommendation model to perform inference and prediction on target related data to obtain information recommendation results.

[0179] Specifically, the feature extraction module 200 is configured as follows:

[0180] The information-related data is information text information, and the information text information includes a title, an abstract, and text information, and an information text information set is formed based on the information text information;

[0181] Performing word segmentation on the information text information in the information text information set to obtain word units and then form a word unit set;

[0182] Build a graph model, use word units as word units in the graph model, and assign corresponding weights to the edges in the graph model according to word unit similarity;

[0183] Based on the corresponding weights, the importance score of each word is calculated, and the words corresponding to the first few words of the importance score are used as keywords;

[0184] Identify information text information through keywords and in combination with the tag system to obtain information text tags;

[0185] The importance score of each word is obtained by:

[0186]

[0187] in, Representation word The importance score of represents the set of all words, Representation word To word The shortest path of Represents the weight of the edge.

[0188] In addition, the feature extraction module 200 is configured as follows:

[0189] The user characteristics include at least one or more of user attribute characteristics, user device characteristics, user activity characteristics, user information preference characteristics, and user stock preference characteristics;

[0190] Based on user characteristics, the user's preference for information text labels is obtained, and then the preference score is obtained;

[0191] Determine the cross-features of user features and information text labels based on the preference scores;

[0192] The preference score is obtained in the following way:

[0193]

[0194] in, Indicates user Text label for information The preference score of Indicates the adjustment item. Indicates Weight coefficient, A function mapping representing user features, Function mapping representing information text labels, Indicates the number of user features.

[0195] Various changes and modifications can be made without departing from the spirit and scope of the present invention, and all equivalent technical solutions also belong to the scope of the present invention.

[0196] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0197] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, devices or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0198] The present invention is described with reference to the flowcharts and / or block diagrams of the method, terminal device (system), and computer program product according to the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of the processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing terminal device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing terminal device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0199] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing terminal device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0200] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal device so that a series of operating steps are executed on the computer or other programmable terminal device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable terminal device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0201] It should be noted that:

[0202] The "one embodiment" or "embodiment" mentioned in the specification means that the specific features, structures or characteristics described in conjunction with the embodiment are included in at least one embodiment of the present invention. Therefore, the phrases "one embodiment" or "embodiment" appearing in various places throughout the specification do not necessarily refer to the same embodiment.

[0203] In addition, it should be noted that the shapes and names of the parts and components of the specific embodiments described in this specification may be different. Any equivalent or simple changes made based on the structure, features and principles described in the patent concept of the present invention are included in the protection scope of the patent of the present invention. The technicians in the technical field of the present invention can make various modifications or supplements to the specific embodiments described or replace them in a similar manner, as long as they do not deviate from the structure of the present invention or exceed the scope defined by the claims, they should all fall within the protection scope of the present invention.

Claims

1. An information recommendation method, characterized in that: The following steps are involved: Acquire relevant data to form a data set, and preprocess the data set to obtain a processed data set, wherein the relevant data includes user-related data and information-related data; A user feature set is obtained based on relevant data, an information feature set and information text labels are obtained based on information-related data, and a cross feature set is formed based on the user feature set and the information text labels; Constructing an information recommendation pre-training model, which includes an input layer, a shared feature structure layer, a main structure layer and an output layer. The input layer receives a user feature set, a cross feature set and an information feature set. The shared feature structure layer processes the user feature set, the cross feature set and the information feature set to obtain corresponding vector representations. The main structure layer processes the vector representation corresponding to the information feature set to obtain a user information vector representation. The output layer fuses the user vector representation, the cross vector representation and the user information vector representation to obtain a recommendation score. The information recommendation pre-training model is trained based on the cross feature set and the sequence data set to obtain the information recommendation model; The information recommendation model performs inference and prediction on target-related data to obtain information recommendation results; The information-related data also includes sequence data, and the sequence data is processed to obtain sequence features, specifically: The sequence data includes information sequence data and individual stock sequence data. The information sequence data is information click records and / or information browsing records within a certain time period, and the individual stock sequence data is holding individual stock records and / or self-selected individual stock records within a certain time period; Extract features from information sequence data to obtain information sequence features; extract features from individual stock sequence data to obtain individual stock sequence features, and process the information sequence features and individual stock sequence features at the shared feature structure layer to obtain information sequence features and individual stock sequence features; The main structure layer processes the vector representation corresponding to the information feature set to obtain the user information vector representation, including the following steps: The information sequence features, information vector representation and individual stock sequence features are processed through the attention mechanism to obtain the user information vector representation, and then the user portrait features are obtained; The attention mechanism is expressed as follows: in, represents the user information vector representation, is the target information vector representation, Represents the i-th information vector representation.

2. The information recommendation method according to claim 1, characterized in that: The step of obtaining the information text label based on the information-related data includes the following steps: The information-related data is information text information, and the information text information includes a title, an abstract, and text information, and an information text information set is formed based on the information text information; Performing word segmentation on the information text information in the information text information set to obtain word units and then form a word unit set; Build a graph model, use word units as word units in the graph model, and assign corresponding weights to the edges in the graph model according to word unit similarity; Based on the corresponding weights, the importance score of each word is calculated, and the words corresponding to the first few words of the importance score are used as keywords; Identify information text information through keywords and in combination with the tag system to obtain information text tags; The importance score of each word is obtained by: in, Representation word The importance score of represents the set of all words, Representation word To word The shortest path of Represents the weight of the edge.

3. The information recommendation method according to claim 1, characterized in that: The forming of a cross feature set based on the user feature set and the information text label comprises the following steps: The user characteristics include at least one or more of user attribute characteristics, user device characteristics, user activity characteristics, user information preference characteristics, and user stock preference characteristics; Based on user characteristics, the user's preference for information text labels is obtained, and then the preference score is obtained; Determine the cross-features of user features and information text labels based on the preference scores; The preference score is obtained in the following way: in, Indicates user Text label for information The preference score of Indicates the adjustment item. Indicates Weight coefficient, A function mapping representing user features, Function mapping representing information text labels, Indicates the number of user features.

4. The information recommendation method according to claim 1, characterized in that: The following steps are also included: Determine whether there is an information feature corresponding to the target information-related data in the target-related data in the information feature set; If it does not exist, search for information-related data similar to the target information-related data in the information-related data; The information features corresponding to similar information-related data are averaged and pooled to obtain the information features of the target information-related data.

5. The information recommendation method according to claim 4, characterized in that: The step of searching the information-related data for information-related data that is similar to the target information-related data comprises the following steps: Representing the information-related data and the target information-related data as a first information TF-IDF vector and a second information TF-IDF vector respectively; Calculate the similarity between the TF-IDF vector of the first information and the TF-IDF vector of the second information based on the cosine similarity model to obtain a similarity score between the information, and determine similar information to the target information based on the similarity score; The information TF-IDF vector is expressed as follows: The cosine similarity model is expressed as follows: in, represents the information TF-IDF vector, represents the similarity score, Represents a word, represents the corresponding score, Indicates Information TF-IDF vectors, Indicates Information TF-IDF vector.

6. The information recommendation method according to claim 1, characterized in that: The output layer fuses the user vector representation, the cross vector representation and the user information vector representation to obtain a recommendation score, including the following steps: At least three fully connected layers are set, and nonlinear functions are used as activation functions between each fully connected layer to fuse the user vector representation, cross vector representation and user information vector representation to obtain the output result; Convert the output of the fully connected layer into a score value based on the sigmoid function; The actual value and the predicted value are evaluated based on the loss function until the loss result is minimized, and then the predicted result is obtained; The sigmoid function is expressed as follows: The loss function is expressed as follows: in, Indicates the actual label of the user for the information data. Indicates the user's predicted label for news data.

7. An information recommendation system, characterized in that: include: A data acquisition module, used to acquire relevant data to form a data set, and pre-process the data set to obtain a processed data set, wherein the relevant data includes user-related data and information-related data; A feature extraction module obtains a user feature set based on relevant data, obtains an information feature set and information text labels based on information-related data, and forms a cross feature set based on the user feature set and the information text labels; A model building module is used to build an information recommendation pre-training model. The information recommendation pre-training model includes an input layer, a shared feature structure layer, a main structure layer and an output layer. The input layer receives a user feature set, a cross feature set and an information feature set. The shared feature structure layer processes the user feature set, the cross feature set and the information feature set to obtain corresponding vector representations. The main structure layer processes the vector representation corresponding to the information feature set to obtain a user information vector representation. The output layer fuses the user vector representation, the cross vector representation and the user information vector representation to obtain a recommendation score. Model training module, which trains the information recommendation pre-training model based on the cross feature set and sequence data set to obtain the information recommendation model; A result recommendation module, which is used for the information recommendation model to perform inference and prediction on target-related data to obtain information recommendation results; The information-related data also includes sequence data, and the sequence data is processed to obtain sequence features, specifically: The sequence data includes information sequence data and individual stock sequence data. The information sequence data is information click records and / or information browsing records within a certain time period, and the individual stock sequence data is holding individual stock records and / or self-selected individual stock records within a certain time period; Extract features from information sequence data to obtain information sequence features; extract features from individual stock sequence data to obtain individual stock sequence features, and process the information sequence features and individual stock sequence features at the shared feature structure layer to obtain information sequence features and individual stock sequence features; The main structure layer processes the vector representation corresponding to the information feature set to obtain the user information vector representation, including the following steps: The information sequence features, information vector representation and individual stock sequence features are processed through the attention mechanism to obtain the user information vector representation, and then the user portrait features are obtained; The attention mechanism is expressed as follows: in, represents the user information vector representation, is the target information vector representation, Represents the i-th information vector representation.

8. The information recommendation system according to claim 7, characterized in that: The feature extraction module is configured as follows: The information-related data is information text information, and the information text information includes a title, an abstract, and text information, and an information text information set is formed based on the information text information; Performing word segmentation on the information text information in the information text information set to obtain word units and then form a word unit set; Build a graph model, use word units as word units in the graph model, and assign corresponding weights to the edges in the graph model according to word unit similarity; Based on the corresponding weights, the importance score of each word is calculated, and the words corresponding to the first few words of the importance score are used as keywords; Identify information text information through keywords and in combination with the tag system to obtain information text tags; The importance score of each word is obtained by: in, Representation word The importance score of represents the set of all words, Representation word To word The shortest path of Represents the weight of the edge.

9. The information recommendation system according to claim 7, characterized in that: The feature extraction module is configured as follows: The user characteristics include at least one or more of user attribute characteristics, user device characteristics, user activity characteristics, user information preference characteristics, and user stock preference characteristics; Based on user characteristics, the user's preference for information text labels is obtained, and then the preference score is obtained; Determine the cross-features of user features and information text labels based on the preference scores; The preference score is obtained in the following way: in, Indicates user Text label for information The preference score of Indicates the adjustment item. Indicates Weight coefficient, A function mapping representing user features, Function mapping representing information text labels, Indicates the number of user features.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

11. An information recommendation device, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, the method according to any one of claims 1 to 6 is implemented.

Images