Information delivery method, device, equipment, storage medium, and program product

Abstract

The present application provides an information delivery method, device, equipment, storage medium, and program product; the embodiments of the present application can be applied to various scenarios such as cloud technology, artificial intelligence, smart transportation, and vehicle-mounted, and involve artificial intelligence technology. The method includes: performing correlation mining on the delivery information features of the information to be delivered and the historical information features of the historical interaction information, obtaining the preference information of the candidate object for the delivery information features, and then determining the interest features of the candidate object for the information to be delivered at the information granularity and feature granularity based on the mined preference information and the delivery information features. Finally, based on the interest features, the target object that is interested in the information to be delivered is screened out from the candidate object to deliver the information. Through the present application, the accuracy of information delivery can be improved.

Description

Technical Field

[0001] The present application relates to artificial intelligence technology, and in particular to an information delivery method, device, equipment, storage medium, and program product. Background Art

[0002] Information delivery refers to finding matching recipients for the information being delivered and then sending it to them to increase its visibility. Related technologies often leverage artificial intelligence (AI) to analyze the historical behavior of all users to identify matching recipients. However, these technologies often struggle with modeling the diverse and personalized interests of users, resulting in low accuracy in information delivery. Summary of the Invention

[0003] The embodiments of the present application provide an information delivery method, apparatus, device, computer-readable storage medium, and program product, which can improve the accuracy of information delivery.

[0004] The technical solution of the embodiment of the present application is implemented as follows:

[0005] The present invention provides an information delivery method, including:

[0006] Obtaining delivery information features of the information to be delivered and historical information features corresponding to the historical interaction information of the candidate object;

[0007] Determining the preference information of the candidate object with respect to the delivery information feature by performing correlation mining on the delivery information feature and the historical information feature;

[0008] constructing an interest feature of the candidate object based on the preference information and the delivery information feature; the interest feature describes the candidate object's interest in the information to be delivered and the delivery information feature;

[0009] According to the interest characteristics, a target object is screened out from the candidate objects, and the information to be delivered is sent to the target object.

[0010] An embodiment of the present application provides an information delivery device, including:

[0011] An information acquisition module is used to obtain the delivery information features of the information to be delivered and the historical information features corresponding to the historical interaction information of the candidate object;

[0012] An information mining module, configured to determine the preference information of the candidate object with respect to the delivery information feature by performing correlation mining on the delivery information feature and the historical information feature;

[0013] A feature construction module is used to construct an interest feature of the candidate object based on the preference information and the delivery information feature; the interest feature describes the candidate object's interest in the information to be delivered and the delivery information feature;

[0014] An object screening module, configured to screen a target object from the candidate objects based on the interest characteristics;

[0015] The information sending module is used to send the information to be delivered to the target object.

[0016] In some embodiments of the present application, the information mining module is further used to perform correlation mining on the delivery information features and the historical information features to obtain feature correlation; the feature correlation characterizes the impact of the historical interaction information on the preference information; based on the feature correlation, the historical information features are fused to obtain the preference information of the candidate object for the delivery information features.

[0017] In some embodiments of the present application, the information mining module is further used to perform inner product processing on the delivery information feature and the historical information feature to obtain a feature inner product result; and normalize the feature inner product result to obtain the feature correlation.

[0018] In some embodiments of the present application, the feature construction module is further used to perform correlation mining on the preference information and the delivery information features to obtain the influence weight of the preference information on the delivery information features; and based on the influence weight and the delivery information features, determine the interest features of the candidate object.

[0019] In some embodiments of the present application, the feature construction module is further used to use the influence weight to perform weighted fusion on the delivery information features to obtain the fusion features corresponding to the preference information; and to perform averaging processing on the fusion features corresponding to the preference information to obtain the interest features of the candidate object.

[0020] In some embodiments of the present application, the object screening module is further used to predict the preference value of the candidate object for the information to be delivered based on the interest characteristics; and to screen the target object from the candidate objects using the preference value.

[0021] In some embodiments of the present application, the information acquisition module is further used to find the neighbor nodes of the information to be delivered from a preset knowledge graph; wherein the neighbor nodes are nodes corresponding to the auxiliary information of the information to be delivered; from a preset feature table, the features corresponding to the information to be delivered, the features corresponding to the neighbor nodes, and the historical information features corresponding to the historical interaction information of the candidate object are screened out; the features corresponding to the information to be delivered and the features corresponding to the neighbor nodes are determined as the delivery information features of the information to be delivered.

[0022] In some embodiments of the present application, the information delivery device also includes: a feature integration module; the feature integration module is used to construct a preset knowledge graph using target information screened from the information library; the target information at least includes information in the information library whose conversion times are greater than a threshold; for each entity node in the preset knowledge graph, a corresponding node sequence is sampled from the preset knowledge graph; feature encoding is performed on the node sequence to obtain the encoding features corresponding to each of the entity nodes; and the encoding features corresponding to each of the entity nodes are used to integrate into the preset feature table.

[0023] In some embodiments of the present application, the feature integration module is also used to crawl information for the target information to obtain descriptive information of the target information; perform text processing on the descriptive information to obtain auxiliary information; wherein the text processing includes at least filtering, word segmentation and deduplication; extract the association between the auxiliary information and the target information to obtain association information; use the target information and the auxiliary information as entity nodes, and use the association information to connect the entity nodes to obtain the preset knowledge graph.

[0024] In some embodiments of the present application, the feature integration module is also used to use the target information as the head entity node, the auxiliary information as other entity nodes other than the head entity node, and use the association information to connect the head entity node and the other entity nodes to obtain the preset knowledge graph.

[0025] In some embodiments of the present application, the feature integration module is further used to perform feature mapping on the node sequence to obtain initial mapping features; and perform feature extraction on the initial mapping features to obtain the coding features of each of the entity nodes.

[0026] In some embodiments of the present application, the feature integration module is further used to perform association extraction on the auxiliary information and the target information to obtain a preliminary extraction result; filter the target engine in which the preliminary extraction result appears from multiple search engines; when the proportion of the target engine in the number of the multiple search engines reaches a proportion threshold, determine the preliminary extraction result as the association information.

[0027] The present invention provides an information delivery device, including:

[0028] a memory for storing executable instructions;

[0029] The processor is used to implement the information delivery method provided in the embodiment of the present application when executing the executable instructions stored in the memory.

[0030] An embodiment of the present application provides a computer-readable storage medium storing executable instructions for causing a processor to execute and implement the information delivery method provided in the embodiment of the present application.

[0031] An embodiment of the present application provides a computer program product that stores executable instructions. When the computer program or instructions are executed by a processor, the information delivery method provided in the embodiment of the present application is implemented.

[0032] The embodiments of the present application have the following beneficial effects: the information delivery device can first determine the candidate object's preference for the delivery information characteristics, that is, the characteristic dimension of the information to be delivered, based on correlation mining of the delivery information characteristics and historical information characteristics, so as to distinguish the preferences of the candidate user objects at a finer granularity, and based on the determined preferences and delivery information characteristics, realize fine-grained modeling of the personalized interests of the user object to obtain more accurate interest characteristics, and finally, based on the interest characteristics, accurately screen out the delivery objects that are interested in the delivery information from the user object, thereby ultimately improving the accuracy of information delivery. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a schematic diagram of the architecture of the information delivery system provided in an embodiment of the present application;

[0034] Figure 2 This embodiment of the present application provides Figure 1 A schematic diagram of the server structure in FIG;

[0035] Figure 3 This is a flow diagram of the information delivery method provided in the embodiment of the present application. Figure 1 ;

[0036] Figure 4 This is a flow diagram of the information delivery method provided in the embodiment of the present application. Figure 2 ;

[0037] Figure 5 This is a flow diagram of the information delivery method provided in the embodiment of the present application. Figure 3 ;

[0038] Figure 6 This is a type of information to be delivered provided by the embodiment of the present application;

[0039] Figure 7 This is another type of information to be delivered provided by the embodiment of the present application;

[0040] Figure 8 This is a flow diagram of the information delivery method provided in the embodiment of the present application. Figure 4 ;

[0041] Figure 9 is a schematic diagram of a meta-path provided in an embodiment of the present application;

[0042] Figure 10 is a schematic diagram of a preset knowledge graph provided in an embodiment of the present application;

[0043] Figure 11 This is a framework diagram of the advertising delivery process provided by the embodiment of the present application;

[0044] Figure 12 This is a schematic diagram of the changes in features when calculating the interaction probability provided by an embodiment of the present application;

[0045] Figure 13 This is a schematic diagram of the construction process of the knowledge graph provided in the embodiment of the present application. DETAILED DESCRIPTION

[0046] In order to make the purpose, technical solutions and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limiting this application. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0047] In the following description, reference is made to “some embodiments”, which describes a subset of all possible embodiments, but it will be understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0048] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this application pertains. The terms used herein are for the purpose of describing the embodiments of this application only and are not intended to limit this application.

[0049] Before further describing the embodiments of the present application in detail, the nouns and terms involved in the embodiments of the present application are explained. The nouns and terms involved in the embodiments of the present application are subject to the following interpretations.

[0050] 1) Artificial Intelligence (AI) refers to the theories, methods, techniques, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, to perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that seeks to understand the essence of intelligence and produce new intelligent machines that can respond in the same way as human intelligence. AI also involves studying the design principles and implementation methods of various intelligent machines, enabling them to possess the capabilities of perception, reasoning, and decision-making.

[0051] Artificial intelligence (AI) technology is a comprehensive discipline encompassing a wide range of fields, encompassing both hardware and software technologies. Foundational AI technologies generally include sensors, specialized AI chips, cloud computing, distributed storage, big data processing, operating / interaction systems, and mechatronics. AI software technologies primarily encompass computer vision, speech processing, natural language processing, as well as machine learning / deep learning, autonomous driving, and smart transportation.

[0052] 2) Machine Learning (ML) is a multidisciplinary field that encompasses probability theory, statistics, approximation theory, convex analysis, and algorithmic complexity theory. It specifically studies how computers can simulate or implement human learning behaviors to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is at the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications span all areas of AI. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and self-learning.

[0053] 3) Knowledge Graph (KG): A semantic network that reveals the relationships between natural entities. It is typically composed of triples such as (entity 1, relationship, entity 2). For example, the triple ("XX Theory", author, Zhang San) represents the knowledge that the author of the book "XX Theory" is Zhang San.

[0054] 4) Meta-path: In a heterogeneous graph (i.e., a graph containing different types of nodes), any node may have multiple types of first-order neighboring nodes. A meta-path is a sequence of node types that specifies the rules for navigating the node. Using the meta-path, a node sequence consisting of a series of nodes can be extracted from the knowledge graph. This node sequence can express the structural and semantic relationships between nodes of different types.

[0055] 5) Knowledge Graph Embedding (KGE) refers to embedding entities and relationships in a knowledge graph into a continuous vector space, and encoding the structure of the knowledge graph into a latent vector through machine learning methods, so that each adjacent entity in the latent space has semantic information.

[0056] 6) Collaborative Filtering (CF) is a classic algorithm in the recommendation field. It uses similarity calculations (such as cosine similarity) to find users with similar interests to the target user, and then recommends items liked by the user to the target user. Alternatively, it uses similarity calculations to find items similar to items liked by the target user and recommends these items to the target user.

[0057] 7) Attention Neural Networks (ANNs) are a resource allocation strategy in deep learning. They typically consist of a query, a key, and a value. By calculating the matching degree between the query and the key, they assign greater weight to important features, allowing them to occupy a larger proportion in the final feature combination process.

[0058] 8) The Dense Vector Search and Matching Framework (FAISS) uses metrics such as Euclidean distance and vector inner product to find the vector most similar to the target vector from a vector database. In addition to the brute force search method mentioned above, it also provides an approximate search method for querying cluster centers.

[0059] 9) Side Information: Supplementary features of the user or item are used to enrich the expression of the user and item, such as the user's age and gender, and the item's category and label.

[0060] 10) Cross Entropy Loss is a classification loss function commonly used in machine learning. When the target task is binary classification, the calculation formula of the cross entropy loss function is shown in formula (1):

[0061]

[0062] Where N is the total number of samples, y iIndicates the label of sample i, the positive class is 1, the negative class is 0, p i It represents the probability that sample i is predicted to be a positive class.

[0063] 11) Skip-Gram: By setting a central word and observing its appearance in the context (in a sliding window of a specific size), it analyzes the conditional probability distribution of the relationship between the central word and the context to learn the embedding representation of each word.

[0064] 12) Click-Through-Rate (CTR) refers to the ratio of the number of times a piece of content is clicked to the number of times it is displayed, reflecting the level of attention the content receives. The formula for CTR is actual clicks / number of displays.

[0065] 13) Seed objects, that is, users who have positive behaviors (such as clicks, downloads, and payments) on a certain item or information, are generally used as positive samples for model training.

[0066] 14) Cost Per Action, which refers to the cost paid by the information delivery party for each action.

[0067] 15) Area Under the Curve (AUC) is a metric used in machine learning to evaluate classification tasks. AUC values ​​closer to 1 indicate better classification results, while values ​​closer to 0.5 indicate poorer results.

[0068] With the advancement of AI research and technology, AI is being studied and applied in a wide range of fields, including smart homes, smart wearable devices, virtual assistants, smart speakers, smart marketing, autonomous driving, drones, robotics, smart healthcare, smart customer service, connected vehicles, and smart transportation. We believe that as technology develops, AI will be applied in even more areas and play an increasingly important role.

[0069] Information placement refers to finding matching recipients for the information being delivered and then sending it to them to increase its visibility. Related technologies often leverage artificial intelligence to analyze historical user behavior and identify matching recipients.

[0070] For example, the word vector generation (Word2Vec) model is used to construct the item representation, and then the recommendation algorithm (such as CF) and machine learning algorithm (such as extreme gradient boosting (eXtreme Gradient Boosting, XGBoost)) are used to determine the matching delivery object. The first step is to process the historical interaction behavior of the object u into a sequence form Tu =[v1,v2,v3,…,v k ], each item v i Can be seen as T u A "word" in this "sentence". Then each item is mapped into a latent vector, and under the assumption of the bag-of-words model, the next "word" is predicted, thereby learning the relationship between "words" and "words". Finally, the latent vector obtained is Represent the relationship between items. The representation of the object can be expressed by the average value of the embedding expression of historical interaction items, that is, After obtaining the representation of the object used and the item representation, the probability of interaction between the representation of the object used and the item representation is predicted by the inner product, or a neural network or XGBoost is used to identify whether they interact as a classification task (0 represents no interaction, 1 represents interaction).

[0071] For example, using the dual-tower model (a general semantic matching model used in the field of information retrieval), which includes two "towers" of query text (Query) and content text (Doc), the respective features of the Query and Doc sides are spliced ​​into a high-dimensional vector, and then deep neural networks (DNN) are used to reduce the dimension of the Query and Doc and condense them into a low-dimensional vector. The similarity between the Query and Doc is then calculated using the vector dot product (or cosine similarity) to find matching delivery objects based on the similarity between the Query and Doc.

[0072] For another example, through the Deep Interest Network (DIN), the expression of the usage object is replaced by the original simple mean of the embedded expressions of historical interaction items to a vector of the weighted sum of attention, thereby obtaining a personalized expression of the usage object, and finding matching delivery objects based on the personalized expression of the usage object.

[0073] For example, the KGE technology is used to first train the embedded representation of the item, and then it is integrated into the CF framework for training as auxiliary information of the item, thereby enhancing the embedded expression of the item, so that the matching delivery object can be found using the enhanced embedded expression.

[0074] However, when constructing item representations based on the Word2Vec model, the representation of the user is simply represented as the mean of the embedding representations of historically interacted items, making the representation of the user too coarse, losing a lot of information, and failing to consider the user's personalized interests. When calculating similarity based on the Twin Towers model, the query and document are already highly condensed, losing a lot of information, making it impossible for the query and document to exchange information with each other. Moreover, the Twin Towers model unbiasedly uses the user's historical click behavior, failing to consider the diversity of the reasons behind the user's click behavior, thereby ignoring the diversity of the user's interests. When determining the representation of the user based on the DIN model, only the correlation between the candidate item and the historically interacted items is considered, failing to model the diverse interests of the user. Moreover, this solution only processes the auxiliary information of the item into randomly initialized latent vectors, underutilizing this auxiliary information. After enhancing the item embedding representation using KGE technology, the subsequent recommendation model does not fully utilize the enhanced embedding representation and is unable to model the user's interests. Instead, it is used as an extension of the item representation. Moreover, the CF framework and the enhanced embedding representation are loosely coupled, failing to fully integrate these features.

[0075] To sum up, in the related technologies, when information is delivered, there is a problem that the modeling of interaction data is shallow based on the single interest of the user, resulting in poor modeling effect on the diverse personalized interests of the user, which makes the accuracy of information delivery low.

[0076] In addition, when constructing item expressions based on the word vector generation (Word2Vec) model, it is necessary to periodically train the model on the entire historical interaction data. Each time the historical interaction data is modeled, a large amount of computing resources and time costs are consumed, which makes the information delivery require more resources.

[0077] The embodiments of the present application provide an information delivery method, apparatus, device, computer-readable storage medium, and program product, which can improve the accuracy of information delivery. The following describes an exemplary application of the information delivery device provided by the embodiments of the present application. The information delivery device provided by the embodiments of the present application can be implemented as various types of terminals such as laptop computers, tablet computers, desktop computers, set-top boxes, mobile devices (for example, mobile phones, portable music players, personal digital assistants, dedicated messaging devices, portable gaming devices), etc., and can also be implemented as a server, and can also be implemented as a device cluster consisting of a terminal and a server. Below, an exemplary application of the information delivery device when it is implemented as a server will be described.

[0078] See also Figure 1 , Figure 11 is a schematic diagram of the architecture of the information delivery system provided in an embodiment of the present application. To support an information delivery application, in information delivery system 100, terminals 400 (terminals 400-1 and 400-2 are shown as examples) are connected to server 200 via network 300. Network 300 can be a wide area network (WAN), a local area network (LAN), or a combination of the two. Information delivery system 100 also includes server 500, which can be independent of server 200 or configured within server 200. Figure 1 The illustrated case is that the server 500 is independent of the server 200 .

[0079] The server 200 is used to obtain the delivery information features of the information to be delivered, and the historical information features corresponding to the historical interaction information of the candidate object; by performing correlation mining on the delivery information features and the historical information features, the preference information of the candidate object for the delivery information features is determined; based on the preference information and the delivery information features, the interest features of the candidate object are constructed; the interest features describe the interest of the candidate object in the information to be delivered and the delivery information features; based on the interest features, the target object is screened out from the candidate objects, and the information to be delivered is sent to the target object via the network 300.

[0080] The terminal 400 is a terminal used by the target object and is used to receive the information to be delivered sent by the server 200 and display the information to be delivered in a graphical interface 410 (graphic interface 410 - 1 and graphical interface 410 - 2 are shown as examples).

[0081] In some embodiments, the server 200 can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers. It can also be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The terminal 400 can be a smart phone, tablet computer, laptop computer, desktop computer, smart speaker, smart watch, smart home appliance, car-mounted device, etc., but is not limited to these. The terminal and the server can be directly or indirectly connected via wired or wireless communication, which is not limited in the embodiments of the present invention.

[0082] See also Figure 2 , Figure 2 This embodiment of the present application provides Figure 1 The structural diagram of the server in Figure 2The server 200 shown includes: at least one processor 210, a memory 250, at least one network interface 220, and a user interface 230. The various components in the server 200 are coupled together via a bus system 240. It is understood that the bus system 240 is used to achieve connection and communication between these components. In addition to including a data bus, the bus system 240 also includes a power bus, a control bus, and a status signal bus. However, for the sake of clarity, the bus system 240 is not described in detail. Figure 2 Various buses are labeled as bus system 240 .

[0083] The processor 210 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc., where the general-purpose processor can be a microprocessor or any conventional processor, etc.

[0084] The user interface 230 includes one or more output devices 231 that enable presentation of media content, including one or more speakers and / or one or more visual display screens. The user interface 230 also includes one or more input devices 232, including user interface components that facilitate user input, such as a keyboard, mouse, microphone, touch screen display, camera, other input buttons and controls.

[0085] The memory 250 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state memory, hard drives, optical drives, etc. The memory 250 may optionally include one or more storage devices that are physically remote from the processor 210.

[0086] The memory 250 includes volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory may be a read-only memory (ROM), and the volatile memory may be a random access memory (RAM). The memory 250 described in the embodiments of the present application is intended to include any suitable type of memory.

[0087] In some embodiments, the memory 250 can store data to support various operations, examples of which include programs, modules, and data structures, or a subset or superset thereof, as exemplified below.

[0088] Operating system 251, including system programs for processing various basic system services and performing hardware-related tasks, such as the framework layer, core library layer, and driver layer, which are used to implement various basic services and process hardware-based tasks;

[0089] A network communication module 252 for reaching other computing devices via one or more (wired or wireless) network interfaces 220 , exemplary network interfaces 220 including Bluetooth, Wi-Fi, and Universal Serial Bus (USB);

[0090] a presentation module 253 for enabling presentation of information via one or more output devices 231 (e.g., a display screen, a speaker, etc.) associated with the user interface 230 (e.g., a user interface for operating peripheral devices and displaying content and information);

[0091] The input processing module 254 is configured to detect one or more user inputs or interactions from one of the one or more input devices 232 and to translate the detected inputs or interactions.

[0092] In some embodiments, the apparatus provided in the embodiments of the present application may be implemented in software. Figure 2 The information delivery device 255 stored in the memory 250 is shown. This device can be software in the form of a program or plug-in, and includes the following software modules: an information acquisition module 2551, an information mining module 2552, a feature construction module 2553, an object screening module 2554, an information sending module 2555, and a feature integration module 2556. These modules are logical and can be arbitrarily combined or further separated according to the functions they implement. The functions of each module will be described below.

[0093] In other embodiments, the information delivery device provided in the embodiments of the present application can be implemented in hardware. As an example, the information delivery device provided in the embodiments of the present application can be a processor in the form of a hardware decoding processor, which is programmed to execute the information delivery method provided in the embodiments of the present application. For example, the processor in the form of a hardware decoding processor can adopt one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components.

[0094] In some embodiments, the information delivery device can implement the information delivery method provided in the embodiments of the present application by running a computer program. For example, the computer program can be a native program or software module in the operating system; it can be a native application (APP, Application), that is, a program that needs to be installed in the operating system to run, such as a text message APP, a social APP, etc.; it can also be a small program, that is, a program that can be run only by downloading it into a browser environment; it can also be a small program that can be embedded in any APP. In short, the above-mentioned computer program can be any form of application, module or plug-in.

[0095] The embodiments of the present application can be applied to various scenarios such as cloud technology, artificial intelligence, smart transportation, and vehicle-mounted. Below, the information delivery method provided by the embodiments of the present application will be described in conjunction with the exemplary application and implementation of the information delivery device provided by the embodiments of the present application.

[0096] See also Figure 3 , Figure 3 This is a flow diagram of the information delivery method provided in the embodiment of the present application. Figure 1 , will combine Figure 3 The steps shown are explained.

[0097] S101: Acquire delivery information features of information to be delivered and historical information features corresponding to historical interaction information of candidate objects.

[0098] The embodiment of the present application is implemented in the scenario of information delivery to users, that is, screening suitable users for the information to be delivered, and then sending the information to the screened users. First, the information delivery device searches for corresponding delivery information features in a database or its own storage space for the information to be delivered, and at the same time determines the candidate objects. It searches the database or its own storage space for information that the candidate objects have interacted with in the past period, that is, features of the historical interaction information, and determines the features as historical information features.

[0099] It is understood that the information to be released can be any type of information waiting to be released. The information to be released can be provided by the information release party, such as a product promotion video or promotional link. The information to be released can also be automatically generated by the information release device, such as a notification generated by the information release device for a certain event, a copy generated for a festival, etc. Furthermore, the information to be released can be cold start information or information that has already gained a certain degree of popularity, and this application does not limit this.

[0100] The delivery information feature is a descriptive feature of the delivery information. It can be a descriptive feature of the appearance of the delivery information itself, or a descriptive feature of auxiliary information such as the supplier or category of the delivery information. Furthermore, in this application, the delivery information of the delivery information can have only one or more delivery information features.

[0101] Candidate objects are a set of objects used to filter the target objects for information delivery. Candidate objects can be all online users, all registered users, or users selected from all registered users with high activity or a certain correlation with the tag of the information to be delivered. This embodiment of the application is not limited here.

[0102] It should be noted that historical interaction information can be information about operations performed by the candidate object during a specific historical time period, or it can be all information performed by the candidate object before the current time point. The above operations can be clicks, double clicks, forwarding, comments, orders, favorites, etc., which are not limited in this application. Furthermore, a candidate object can have only one historical interaction information or multiple historical interaction information.

[0103] The historical information feature can be a descriptive feature of the appearance of the historical interaction information itself, a descriptive feature of the semantics of the historical interaction information, or a descriptive feature of auxiliary information such as the category or supplier of the historical interaction information. Each historical interaction information has its corresponding historical information feature. Therefore, in the embodiment of the present application, the information delivery device can obtain one or more historical information features.

[0104] S102: Determine the preference information of the candidate object with respect to the delivery information feature by performing correlation mining on the delivery information feature and the historical information feature.

[0105] It should be noted that the interest preferences of candidate objects can be mined from historical interaction information. Therefore, the information delivery device mines the correlation between the delivery information characteristics and historical information characteristics, and can clarify the degree of interest of the candidate objects in the delivery information characteristics, that is, clarify the candidate objects' preferences for the delivery information characteristics and obtain preference information.

[0106] That is to say, in the embodiment of the present application, based on the characteristics of the delivery information and the characteristics of the historical information, the preferences of the candidate objects for the characteristic dimensions of the information to be delivered are mined, so that the granularity of the mined preference information is finer, reaching the characteristic dimension, and the granularity of the interest analysis of the candidate objects is finer, so as to facilitate more accurate interest feature modeling at a finer granularity in the future.

[0107] In some embodiments, the information delivery device can realize correlation mining by calculating the similarity between the delivery information features and the historical information features, obtain the correlation between the delivery information features and the historical information features, and then fuse the obtained correlation with the historical information features, that is, use the correlation to adjust the historical information features, and then use the adjusted historical information features to describe the candidate object's interest preferences for each delivery information feature of the information to be delivered.

[0108] In other embodiments, the information delivery device can also match the delivery information features and historical information features element by element to achieve correlation mining, generate preference weights representing positive impacts for the delivery information features that match the historical information features, and generate preference weights representing negative impacts for the delivery information features that do not match the historical information features, thereby obtaining preference information for each delivery information feature.

[0109] It can be understood that the number of delivery information features and the number of preference information are the same, that is, when the information delivery device obtains multiple delivery information features, multiple preference information will be generated accordingly; when the information delivery device only obtains one delivery information feature, only one preference information will be generated.

[0110] S103: Construct interest features of the candidate objects based on the preference information and delivery information features.

[0111] The information delivery device jointly analyzes the preference information and delivery information features to clarify the weight distribution of the preference information on the delivery information features, that is, to clarify the attention of the candidate objects to different delivery information features, and finally combines this attention to construct the interest features of the candidate objects.

[0112] It should be noted that, based on preference information, the information delivery device can determine the candidate's interest in the delivery information features. This interest in the delivery information features can actually represent whether the candidate is interested in the delivery information. Therefore, the interest profile derived based on the preference information and delivery information features can simultaneously indicate whether the candidate is interested in the delivery information, as well as specifically whether they are interested in each delivery information feature. In other words, the interest profile describes the candidate's interest in both the delivery information and the delivery information features.

[0113] It is understandable that since the granularity of preference information reaches the feature dimension, the granularity of interest features constructed based on preference information and delivery information features is also finer, so that target objects can be screened more accurately based on finer-grained interest features.

[0114] In some embodiments, the information delivery device can calculate the similarity between the preference information and the delivery information features to clarify the attention of the preference information on the delivery information features, and then fuse the delivery information features based on the calculated attention to obtain the interest features.

[0115] In other embodiments, the information delivery device can also match the preference information with the delivery information characteristics element by element (i.e., compare whether the various components of the preference information and the various components of the delivery information characteristics are the same), allocate different attention to the delivery information characteristics according to the proportion of the matched elements, and finally use the attention to update the delivery information characteristics to obtain interest characteristics.

[0116] It is understandable that the information delivery device will only obtain one interest feature for one candidate object, and the interest features of different candidate objects are different.

[0117] S104: Filter out target objects from candidate objects based on interest characteristics, and send the information to be delivered to the target objects.

[0118] After obtaining the interest characteristics of the candidate subjects, the information delivery device analyzes these characteristics to determine which of the candidate subjects are interested in receiving the information delivery and selects these interested subjects as target subjects. The information delivery device then pushes the information to the selected target subjects on a scheduled or real-time basis to complete the information delivery process.

[0119] In some embodiments, the information delivery device may analyze the preference values ​​of the candidate objects for the information to be delivered from the interest characteristics, and then filter out the target objects from the candidate objects based on the preference values.

[0120] In other embodiments, the information delivery device may also cluster the interest features of the candidate objects to obtain different clusters, and then use the clusters hit by the labels of the information to be delivered to determine the target cluster, and determine the objects in the target cluster as target objects.

[0121] It can be understood that compared with the related technologies of constructing item expressions based on the word vector generation (Word2Vec) model, calculating similarity based on the twin-tower model, determining the expression of the user object based on the DIN model, and enhancing the embedded expression of the item through KGE technology, in the embodiment of the present application, the information delivery device can first determine the candidate object's preference for the delivery information characteristics, that is, the characteristic dimension of the information to be delivered, based on the correlation mining of the delivery information characteristics and the historical information characteristics, so as to distinguish the preferences of the candidate user objects at a finer granularity, and based on the determined preferences and delivery information characteristics, realize the modeling of the personalized interests of the user object at a finer granularity, obtain more accurate interest characteristics, and finally, based on the interest characteristics, accurately screen out the delivery objects that are interested in the delivery information from the user object, thereby ultimately improving the accuracy of information delivery.

[0122] based on Figure 3 , see Figure 4 , Figure 4 This is a flow diagram of the information delivery method provided in the embodiment of the present application. Figure 2 By mining the correlation between the delivery information features and the historical information features, the preference information of the candidate object with respect to the delivery information features is determined, that is, the specific implementation process of S102 may include: S1021-S1022, as follows:

[0123] S1021. Perform correlation mining on the delivery information features and the historical information features to obtain feature correlation.

[0124] The information delivery device can obtain feature correlation by calculating the distance between the delivery information feature and the historical information feature in the vector space, or obtain feature similarity by calculating the projection of the delivery information feature on the historical information feature. This application does not limit this.

[0125] It's important to note that historical information features correspond to historical interaction information, which is generated based on the candidate's own preferences and interests. Calculating the correlation between historical information features and interaction information features clearly identifies which of the interaction information features the candidate prefers, and the degree of preference. In other words, feature correlation characterizes the influence of historical interaction information on preference information.

[0126] S1022: Fusing historical information features based on feature relevance to obtain preference information of candidate objects with respect to delivery information features.

[0127] The information delivery device adjusts the historical information features using feature relevance to obtain weight-adjusted historical information features, and then fuses the adjusted historical information features into one feature, which represents the candidate's preference information for the delivery information features.

[0128] In some embodiments, the information delivery device may complete the fusion of the adjusted historical information features through weighted summation. In other embodiments, the information delivery device may complete the fusion of the adjusted historical information features through splicing.

[0129] For example, when the historical information features obtained by the information delivery device are The information delivery characteristics obtained are When k historical information features and L information delivery features are obtained, the embodiment of the present application provides a formula for fusing the adjusted historical information features, see formula (2):

[0130]

[0131] Among them, α ij is the similarity between the i-th delivery information feature and the j-th historical information feature, e vj is the jth historical information feature, k is the total number of historical information features, b i It is the candidate’s preference information for the i-th delivery information feature.

[0132] In an embodiment of the present application, the information delivery device can mine the correlation between historical information features and delivery information features to clarify the impact of historical interaction information on preference information, and then fuse the historical information features based on the impact to achieve modeling of candidate preferences for candidate objects at the feature granularity, so as to facilitate more accurate screening of target objects in the future.

[0133] based on Figure 4 , see Figure 5 , Figure 5 This is a flow diagram of the information delivery method provided in the embodiment of the present application. Figure 3 In some embodiments of the present application, correlation mining is performed on the delivery information features and the historical information features to obtain the feature correlation, that is, the specific implementation process of S1021, which may include: S1021a-S1021b, as follows:

[0134] S1021a. Perform inner product processing on the delivery information feature and the historical information feature to obtain a feature inner product result.

[0135] S1021b. Normalize the feature inner product result to obtain feature correlation.

[0136] When the number of delivery information features and historical information features is 1, the information delivery device will directly calculate the inner product of the delivery information features and the historical information features, and normalize the obtained feature inner product result using a preset normalization parameter to obtain feature relevance.

[0137] If the number of either the delivery information feature or the historical information feature is not 1, the delivery device pairs each delivery information feature with each historical information feature to obtain multiple information feature pairs. It then performs inner product calculations on each of these pairs to obtain multiple feature inner product results. Finally, the delivery device accumulates these multiple feature inner product results and uses the accumulated result to normalize each feature inner product result. The resulting normalized result represents the feature correlation between each delivery information feature and each historical information feature.

[0138] For example, the present embodiment provides a calculation formula for feature correlation, see formula (3):

[0139]

[0140] in, is the jth historical information feature, is the i-th delivery information feature, e is the natural base number, c i,j is the feature correlation between the i-th information delivery feature and the j-th historical information feature, represents the inner product.

[0141] In an embodiment of the present application, the information delivery device can perform inner product calculation on the delivery information features and the historical information features, and then normalize the result obtained by the inner product calculation to achieve correlation mining of the delivery information features and the historical information features, so as to facilitate subsequent modeling of preference information based on feature correlation.

[0142] Continue to see Figure 5 In some embodiments of the present application, based on the preference information and delivery information features, the interest features of the candidate objects are constructed, that is, the specific implementation process of S103 may include: S1031-S1032, as follows:

[0143] S1031. Perform correlation mining on the preference information and delivery information features to obtain the influence weight of the preference information on the delivery information features.

[0144] In some embodiments, the information delivery device may perform inner product processing on the preference information and the delivery information characteristics to project the preference information onto the delivery information characteristics, and then normalize the obtained inner product result (normalization may be performed using the accumulated result of all inner product results, or the largest inner product result) to determine the impact of the preference information on the delivery information characteristics, thereby obtaining the impact weight.

[0145] For example, formula (4) is a calculation formula for the influence weight provided in the embodiment of the present application, as follows:

[0146]

[0147] Among them, b i is the preference information of the i-th information delivery feature, is the jth information delivery feature, is the inner product processing, c i,j is the influence weight of the i-th preference information on the j-th delivery information feature.

[0148] In other embodiments, the information delivery device can also transform the delivery information features, such as performing transposition transformation, scaling transformation, etc., to obtain the transformed information features, and then calculate the feature distance between the transformed information features and the preference information, normalize the feature distance, and obtain the influence weight of the preference information on the delivery information features.

[0149] S1032: Determine the interest characteristics of the candidate object based on the influence weight and the delivery information characteristics.

[0150] In some embodiments, the information delivery device can adjust the delivery information characteristics by influencing the weights, so that the adjusted delivery information characteristics will reflect the interests and preferences of the candidate. The information delivery device can then directly determine the adjusted delivery information characteristics as the interest characteristics, or combine different adjusted delivery information characteristics to obtain the interest characteristics.

[0151] In other embodiments, the information delivery device may also compare the impact weight with the weight threshold, filter out delivery information features whose impact weight is greater than the weight threshold, and then average or fuse the filtered delivery information features to obtain the interest features of the candidate object.

[0152] In an embodiment of the present application, the information delivery device will first analyze the influence weight of the preference information on the delivery information characteristics, so as to clarify the interest distribution of the candidate objects for the information to be delivered through the influence weight, and then use the influence weight and the delivery information characteristics to model the interest characteristics of the candidate objects. In this way, it is possible to model the interests of the candidate users at the feature granularity to clarify the personalized interests of different users.

[0153] In some embodiments of the present application, based on the influence weight and the delivery information characteristics, determining the interest characteristics of the candidate object, that is, the specific implementation process of S1032, may include: S1032a-S1032b, as follows:

[0154] S1032a: Using the influence weights, perform weighted fusion on the delivery information features to obtain fusion features corresponding to the preference information.

[0155] It should be noted that the same preference information can have different impacts on different delivery information features, resulting in different impact weights for the same preference information on different delivery information features. In this case, the information delivery device uses the different impact weights for the same preference information as the weighted weights for the corresponding delivery information features, and then performs a weighted summation of the different delivery information features to obtain the fused features corresponding to the preference information.

[0156] S1032b: Perform average processing on the fusion features corresponding to the preference information to obtain the interest features of the candidate object.

[0157] Next, the information delivery device averages the fusion features corresponding to different preference information, so that the interest features can express different preference information, thereby achieving complete modeling of the interest preferences of the candidate objects.

[0158] For example, the embodiment of the present application provides a formula for determining the interest characteristics of a candidate object, see formula (5):

[0159]

[0160] Among them, e tj is the jth delivery information feature, c ij is the influence weight of the i-th preference information on the j-th delivery information feature, ∑ j c ij e tj is the fusion feature corresponding to the i-th preference information, L is the number of preference information (the same as the number of delivery information features), e u is the feature of interest of the candidate.

[0161] It should be noted that, in other embodiments, the information delivery device may also average the fused features and then combine the processed features with the temporal weights to obtain the interest features of the candidate objects.

[0162] In an embodiment of the present application, the information delivery device will fuse the delivery information features together through the influence weight, and then determine the average of the fused features corresponding to different preference information as the interest feature, so that the interest feature can express the interest of the candidate object in different feature dimensions, so that the obtained interest feature is more accurate.

[0163] In some embodiments of the present application, the specific implementation process of selecting a target object from candidate objects based on interest characteristics, that is, S104, may include: S1041-S1042, as follows:

[0164] S1041. Predict the candidate's preference value for the information to be delivered based on the interest characteristics.

[0165] The information delivery device analyzes and predicts the preference of the interest characteristics to determine whether the candidate is interested in the information to be delivered, the degree of interest, etc. from the interest characteristics. In this way, the information delivery device obtains the preference value.

[0166] S1042. Filter the target object from the candidate objects using the preference value.

[0167] The information delivery device can sort the preference values ​​of the candidate objects, filter out the k candidate objects with the largest preference values, and determine them as target objects, or compare the preference values ​​with a set threshold, filter out the candidate objects with preference values ​​greater than the threshold, and determine them as target objects.

[0168] It should be noted that the target objects selected based on the preference values ​​for the information to be delivered can be guaranteed to be users who are interested in one or more features of the information to be delivered. Figure 6 ( Figure 6 The target objects selected by the information delivery device are any one or more interested users of the features: fireworks, shooting merchants, etc. Figure 7 ( Figure 7 In the information 7-1 to be delivered provided in another embodiment of the present application, the target objects selected by the information delivery device are users who are interested in one or more of the following features: web games, limited-edition IDs. In this way, it is possible to ensure that the information to be delivered has a better delivery effect.

[0169] In the embodiment of the present application, the information delivery device can use interest characteristics to filter out candidates who are interested in the information to be delivered and use them as target objects. In this way, the target objects can be all interested in the information to be delivered, thereby improving the accuracy of information delivery and thus improving the benefits brought by information delivery.

[0170] based on Figure 5 , see Figure 8 , Figure 8 This is a flow diagram of the information delivery method provided in the embodiment of the present application. Figure 4 In some embodiments of the present application, the specific implementation process of obtaining the delivery information features of the information to be delivered and the historical information features corresponding to the historical interaction information of the candidate object, that is, S101, may include: S1011-S1013, as follows:

[0171] S1011. Find neighbor nodes of the information to be delivered from the preset knowledge graph.

[0172] The information delivery device first determines the node corresponding to the information to be delivered from the pre-built knowledge graph, and then finds the node's neighbor nodes. It is understandable that the neighbor nodes can be first-order neighbor nodes or second-order neighbor nodes of the information to be delivered.

[0173] It should be noted that the neighbor node is the node corresponding to the auxiliary information of the information to be delivered. Therefore, the information delivery device searches for the neighbor node in order to clarify the auxiliary information of the information to be delivered. The auxiliary information can be the supplier information of the information to be delivered or the category information of the information to be delivered, which is not limited in this application.

[0174] S1012: Filter out features corresponding to the information to be delivered, features corresponding to neighboring nodes, and historical information features corresponding to historical interaction information of candidate objects from a preset feature table.

[0175] The information delivery device searches the pre-built feature table for the features corresponding to the information to be delivered, the features corresponding to neighboring nodes, and the features of historical delivery information. It should be noted that the features corresponding to historical interaction information are simply historical information features. That is, historical information features may only include the features of the historical delivery information itself and may not include features corresponding to auxiliary information of the historical interaction information.

[0176] S1013: Determine the features corresponding to the information to be delivered and the features corresponding to the neighboring nodes as delivery information features of the information to be delivered.

[0177] The information delivery device determines the features corresponding to the information to be delivered and the features corresponding to the neighboring nodes as delivery information features. In this way, the obtained delivery information features can describe the information to be delivered from various angles.

[0178] In the embodiment of the present application, the preset knowledge graph and the preset feature table are both determined offline. When making online predictions, the information delivery device can directly find neighbor nodes from the preset knowledge graph, and find historical delivery features of historical interaction information and delivery information features of the information to be delivered from the preset feature table, thereby speeding up the calculation speed during information delivery.

[0179] In some embodiments of the present application, before searching for neighbor nodes to which information is to be delivered from a preset knowledge graph, that is, before S1011, the method may further include: S201-S204, as follows:

[0180] S201. Utilize the target information selected from the information database to construct a preset knowledge graph.

[0181] The information delivery device filters the information in the information database to obtain target information for constructing the knowledge graph, and then uses the target information to construct the preset knowledge graph. It should be noted that the target information at least includes information in the information database whose conversion count is greater than a threshold number of times, where the threshold number of times can be set according to actual circumstances and is not limited in this application.

[0182] It can be understood that the information to be delivered can be any one of the target information.

[0183] S202. For each entity node in the preset knowledge graph, sample a corresponding node sequence from the preset knowledge graph.

[0184] The information delivery device can take any entity node as the starting node of the walk and walk in the preset knowledge graph according to the meta-path until the walk ends. All the nodes passed during the walk are connected to obtain the node sequence corresponding to each entity node.

[0185] It is understandable that the meta-path is a sequence of node types that specifies the walking rules. The meta-path can be set according to actual needs and is not limited in this application.

[0186] For example, Figure 9Schematic diagram of the meta-path provided in the embodiment of the present application. When information is represented by V, secondary classification is represented by C, primary classification is represented by F, label is represented by T, and supplier is represented by S, the meta-path can be any of path 9-1 (i.e., VCFV), path 9-2 (i.e., VCFCV), path 9-3 (i.e., VCV), path 9-4 (i.e., TVT), and path 9-5 (i.e., SVCVS).

[0187] Furthermore, in some embodiments, the information delivery device can also classify the neighbor nodes of each entity node, that is, when roaming, only select the next node from a certain specific type of neighbor nodes to form a node sequence, so as to ensure that the different semantic information contained in different types of neighbor nodes can be fully utilized.

[0188] S203: Perform feature encoding on the node sequence to obtain the encoding feature corresponding to each entity node.

[0189] After obtaining the node sequence, the information delivery device will perform encoding learning on the node sequence to generate corresponding encoding features for each entity node, so as to facilitate the subsequent generation of a preset feature table based on the encoding features.

[0190] It is understandable that the information delivery device can use one-hot encoding to encode the node sequence to directly obtain the encoding features, or it can further use the Skip-Gram model to learn on the features obtained by one-hot encoding to obtain the encoding features.

[0191] S204: Utilize the coding features corresponding to each entity node to integrate into a preset feature table.

[0192] The information delivery device integrates the coding features of each entity node into a feature table, thus obtaining a preset feature table for subsequent feature search.

[0193] In an embodiment of the present application, the information delivery device can construct a knowledge graph and feature table at least for information with a large number of conversions in the information database, so that the constructed preset knowledge graph contains rich knowledge, which facilitates more effective interest modeling of candidate objects in the subsequent process.

[0194] Of course, in other embodiments, the information delivery device can also learn each node on the preset knowledge graph based on the graph attention network to obtain the encoding features of each node.

[0195] In some embodiments of the present application, a preset knowledge graph is constructed using target information screened from an information database, i.e., a specific implementation process of S201 may include: S2011-S2014, as follows:

[0196] S2011. Crawl target information to obtain description information of the target information.

[0197] The information delivery device uses the target information as a clue to crawl information from a database or search engine, and uses the crawled information as the description information of the target information. It is understood that the description information can be a text paragraph that explains and categorizes the target information, or a short phrase such as the category of the target information or the name of the supplier, and this application does not limit this.

[0198] S2012: Perform text processing on the description information to obtain auxiliary information.

[0199] It should be noted that text processing includes at least filtering, word segmentation and deduplication, that is, the information delivery device simplifies and extracts key points of the description text to obtain auxiliary information. Therefore, the auxiliary information can include a concise description of the category, supplier, etc. of the target information.

[0200] S2013: Extract the correlation between the auxiliary information and the target information to obtain correlation information.

[0201] Next, the information delivery device extracts associations between the auxiliary information and the target information, identifying the associations between the auxiliary information and the target information, and identifies the resulting information as associated information. Of course, the information delivery device can also extract associations for different target information and different auxiliary information, and the resulting information can also be used as associated information.

[0202] S2014. The target information and the auxiliary information are used as entity nodes, and the entity nodes are connected using the associated information to obtain a preset knowledge graph.

[0203] Finally, the information delivery device uses the target information and auxiliary information as entity nodes in the knowledge graph, and generates corresponding connection edges for the related information to connect different entity nodes. In this way, the construction of the knowledge graph is completed and the preset knowledge graph is obtained.

[0204] For example, Figure 10 This is a schematic diagram of the preset knowledge graph provided in the embodiment of this application. Figure 10In the preset knowledge graph 10-1, the information delivery device determines tag 10-3, secondary category 10-4, and supplier 10-5 for item 10-1 (information to be delivered). It also determines tag 10-6, tag 10-3, supplier 10-5, and secondary category 10-7 for item 10-2 (information to be delivered). Furthermore, it determines primary category 10-8 for secondary categories 10-4 and 10-7. The information delivery device then connects the different entity nodes using associations to create the preset knowledge graph.

[0205] In the embodiment of the present application, the information delivery device obtains a preset knowledge graph based on processing the descriptive information of the target information, and the descriptive information, especially the category, supplier and other information, is difficult to change after the target information is generated. Therefore, the obtained descriptive information will not be as changeable as the interaction between the object and the information. Therefore, the obtained preset knowledge graph is also relatively stable and usually needs to be updated after a long time. Therefore, the information delivery device does not have to frequently construct the knowledge graph, which reduces the workload required for offline processing of information delivery and saves computing resources.

[0206] In some embodiments of the present application, the target information and the auxiliary information are used as entity nodes, and the entity nodes are connected using the association information to obtain a preset knowledge graph. The specific implementation process of S2014 may include: S2014a, as follows:

[0207] S2014a. Use the target information as the head entity node, and the auxiliary information as other entity nodes other than the head entity node, and use the associated information to connect the head entity node and other entity nodes to obtain a preset knowledge graph.

[0208] That is to say, in the embodiment of the present application, the preset knowledge graph constructed by the information delivery device uses the target information as the head entity node. In this way, when sampling the node sequence for the target information, a longer node sequence can be obtained, so as to learn more knowledge for the target information.

[0209] In some embodiments of the present application, feature encoding is performed on the node sequence to obtain the encoding feature corresponding to each entity node, that is, the specific implementation process of S203 may include: S2031-S2032, as follows:

[0210] S2031. Perform feature mapping on the node sequence to obtain initial mapping features.

[0211] S2032: Extract the initial mapping features to obtain the encoding features of each entity node.

[0212] In the embodiment of the present application, the information delivery device first maps the node sequence into an initial mapping feature, and then further extracts features from the initial mapping feature, and the resulting feature is the coding feature.

[0213] It is understandable that the information delivery device can directly input the initial mapping features into the feature coding model to extract the coding features.

[0214] Furthermore, the feature encoding model can be a Skip-Gram model, in which the output of the intermediate layer is the encoding feature. During training, the information delivery device can further perform probability prediction on the encoding feature output by the intermediate layer, calculate the loss value based on the predicted probability, and use the loss value to adjust the parameters of the model.

[0215] For example, the information delivery device may input the training data v into the Skip-Gram model to extract the training coding features using the Skip-Gram model, and use the training coding features to predict the occurrence of v in its context c. t The probability of is calculated as follows:

[0216]

[0217] in, is context c t Features, X v is the feature of v, θ is the parameter of Skip-Gram model, p(c t |v;θ) is the calculated probability.

[0218] Then, the loss value is calculated using formula (7) to perform back propagation and adjust the parameters of the model.

[0219]

[0220] Among them, V is the set of nodes in the knowledge graph, T V is a collection of node types in the knowledge graph. Is a collection of contexts.

[0221] In the embodiment of the present application, the information delivery device performs feature encoding on the node sequence to extract target information encoding features from the node sequence, so as to facilitate the subsequent generation of a preset feature table.

[0222] In some embodiments of the present application, the extraction of association between the auxiliary information and the target information to obtain the association information, i.e., the specific implementation process of S2013, may include: S2013a-S2013c, as follows:

[0223] S2013a, perform correlation extraction on auxiliary information and target information to obtain preliminary extraction results,

[0224] S2013b. Filter the target engine that has the preliminary extraction results from multiple search engines.

[0225] The information delivery device searches for preliminary extraction results of auxiliary information and target information in multiple search engines respectively, and then counts the engines where the preliminary extraction results appear and uses them as target engines.

[0226] It is understandable that the multiple search engines may be search engines that crawl the description information, or may be any search engine, and this application does not limit this.

[0227] S2013c: When the target engine accounts for a certain proportion of the search engines, the initially extracted information is determined as related information.

[0228] The information delivery device counts the number of target engines and the total number of search engines. It then divides the number of target engines by the total number to calculate the target engine's share of the total number of search engines. This share is then compared with a threshold. If the share reaches the threshold, the preliminary extraction result is considered reliable and is then directly identified as relevant information.

[0229] In an embodiment of the present application, the information delivery device will verify the associations extracted from the auxiliary information and the target information, that is, the preliminary extraction results. Only after the verification is passed will the preliminary extraction results be used as the association information, so that the association information is credible and the accuracy of the preset knowledge graph constructed based on the association information is guaranteed.

[0230] The following describes an exemplary application of the embodiments of the present application in a practical application scenario.

[0231] The embodiment of the present application is implemented in the scenario of advertising via SMS, that is, targeting the group of people interested in the advertisement and sending the advertisement to this group via SMS. The embodiment of the present application mainly builds a lightweight recommendation model based on the mobile phone APP installation list installed by the user (candidate object) to model the user's diverse and personalized interests, and deliver specific advertisements to the user regularly and accurately.

[0232] Figure 11 This is a framework diagram of the advertising delivery process provided by the embodiment of this application. Figure 11This process mainly consists of two parts: constructing a knowledge graph embedding (KGE) 11-1 and estimating click-through rate (CTR) 11-2. The knowledge graph embedding construction 11-1 is divided into building a knowledge graph 11-12 (preset knowledge graph) based on the app name 11-11, screening meta-path templates (meta-paths) 11-13, performing random walks based on the meta-path templates 11-14, initializing node embeddings 11-15 (initial mapping features), and modeling the meta-path (node ​​sequence) using the Skip-Gram model 11-16. CTR estimation 11-2 involves searching the embedding table (preset feature table) 11-23 for historical click sequences 11-21 (historical interaction information) and candidate information 11-22 (information to be delivered), calculating item-level attention 11-24, calculating feature-level attention 11-25, calculating the fully connected layer 11-26, and predicting the score (preference value) 11-27.

[0233] based on Figure 11 It can be seen that when the server constructs the knowledge graph embedding, it is divided into the following steps:

[0234] Step 1: Build a knowledge graph where the head entity (head entity node) or tail entity is the APP, and filter the meta-path template.

[0235] Step 2: Random walk based on the meta-path template to achieve sampling. For the knowledge graph G = (V, E, T), V represents the vertex set, E represents the edge set, and T represents the vertex and edge type set. Assume that the meta-path template is Then the calculation formula for the probability transfer of step i is There are three situations:

[0236] 1) When v i+1 and There is an edge between two nodes, and v i+1 The node's type belongs to the next type V defined by the meta-path t+1 , then the calculation formula for the probability transfer of step i is shown in formula (8):

[0237]

[0238] in, express Among the neighbor nodes of V t+1 A collection of nodes of type For the meta path template.

[0239] 2) When v i+1 and There is an edge between two nodes, and node v i+1 The type does not belong to the next type V defined by the meta-path template t+1When , the calculation formula of the probability transfer of step i is shown in formula (9):

[0240]

[0241] 3) When v i+1 and When there is no edge between two nodes, the calculation formula for the probability transfer of the i-th step is also shown in formula (9).

[0242] The server performs random walk based on this formula and can sample a meta-path (node ​​sequence) of length l triggered by node v from the knowledge graph.

[0243] Step 3: Initialize node embedding and get an initialized embedding table X∈R |V|×D , map all nodes on the knowledge graph into vectors in the vector space (initial mapping features), for example, X v ∈R D , where D is the dimension of the vector.

[0244] Step 4: Skip-Gram model models the meta-path. During training, the normalization (such as Softmax) model needs to be trained on the context c of different types of neighbor nodes. t Normalize, that is, predict that node v appears in c t The probability of , the calculation formula can be shown in formula (6), and then the calculated probability is input into formula (7) to calculate the loss value, and then the loss value is used to back propagate to adjust the parameters of the Softmax model and Skip-Gram model.

[0245] Furthermore, considering that different types of neighbor nodes have different semantics in a heterogeneous graph, the server will sample different types of nodes separately. Therefore, during training, it is not necessary to sample all nodes in the knowledge graph. Only M negative samples need to be sampled from the central node v. Thus, for node v, the final loss function is shown in Equation (10):

[0246]

[0247] Among them, P t (u t ) represents the predefined distribution of negative samples of node type t.

[0248] Step 5: Save the embeddings of all modeled nodes as a table (preset feature table) for subsequent delivery tasks.

[0249] Click-through rate prediction is mainly based on the historical access sequence T of user u u =[v1,v2,v3,…,vk ] and candidate item t, predict the interaction probability of user u to candidate item t Figure 12 This is a schematic diagram of the changes in features when calculating the interaction probability provided by the embodiment of the present application. Figure 12 It can be seen that the process mainly includes the following steps:

[0250] Step 1: Find the embedding table, that is, for the historical access sequence T u =[v1,v2,v3,…,v k ] (including multiple historical interaction information) to query the embedding respectively, and query the corresponding embedding for the candidate item t (to be released information), so as to obtain (including multiple historical information features) and (Contains multiple delivery information features). Among them, k is the number of items in the historical access sequence, L is the number of features of the candidate item, and d is the dimension of the embedding. Here, in order to make full use of the information of the candidate item, all its first-order neighbor nodes (neighbor nodes of the information to be delivered) in the knowledge graph are used as auxiliary information of the candidate item, that is, the embedding of the first-order neighbor nodes and the embedding of the candidate item itself are combined to obtain

[0251] Step 2: Item-level attention calculation: This involves calculating the correlation between historically visited items and candidate items to help assign weights when modeling interests. For example, if a user's historical visit sequence is [Swordsman Games, Xianxia TV Series, XX Browser, and XX Securities], then ancient-style mobile games are more in line with the user's interests, and thus Swordsman Games and Xianxia TV Series should be given greater weight when modeling the user's interests.

[0252] The formula for calculating item-level attention can be shown as formula (11):

[0253]

[0254] in, yes The transpose of Is an attention score table, each row vector They all represent the weight distribution of the user's historical visited items on the candidate items. The calculation method of a single weight (feature relevance) is shown in formula (3).

[0255] Afterwards, the server performs a weighted summation on each weight distribution (preference information) and the corresponding historical access sequence embedding to obtain the user's preference b for each feature of the candidate item.i ∈R d . b i The calculation formula can be shown as formula (2).

[0256] Step 3: Calculate feature-level attention to model the user's interests. When recommending candidate items to users, different features will occupy different importance. For example, when a user is interested in 2D games, their historical installation sequence is mostly 2D games. At this time, recommending cartoon-themed games to them is more likely to attract users to click than recommending fighting games. Therefore, it is necessary to use weight distribution to further calculate the user's attention to the items. The calculation process is shown in formula (12):

[0257]

[0258] in, Calculated by item-level attention, It is an attention score table, with the vector of each row object represents the weight distribution of the user's preference on the i-th feature dimension of the candidate item. The single weight (influence weight) can be calculated by formula (4).

[0259] Finally, the weighted sum is performed to obtain the embedding (interest feature) of the user’s interest. This process can be implemented by formula (5).

[0260] Step 4: Through fully connected layer calculation and score prediction, the intended group (target object) is screened. After obtaining a value that can represent the user's diverse and personalized interests, the server will use the classifier to calculate the user's preference value for the candidate item. The preference value range is [0, 1]. The calculation formula is shown in formula (13):

[0261]

[0262] Among them, W and b are the parameters of sigmoid, is the preference value.

[0263] Next, the construction process of the knowledge graph is explained. Figure 13 This is a schematic diagram of the construction process of the knowledge graph provided in the embodiment of this application, see Figure 13 , the process includes:

[0264] S301. Filter apps (target information) with fewer than 5 installation times, that is, filter out apps with too few installation times.

[0265] S302: Crawl description information using the app name as a keyword. The server can crawl description information from multiple different search engines, such as Chinese descriptions searched from search engine 1, Chinese descriptions searched from search engine 2, and relationships searched from a knowledge graph engine.

[0266] S303, Chinese processing (filtering, word segmentation, deduplication): Chinese processing is performed to enable smooth relationship extraction in the subsequent process.

[0267] S304: Relationship Extraction. At this point, the server will determine that relationships with a credibility score greater than a threshold, as well as relationships that appear in two or more search engines (the target engine's number of appearances in multiple search engines reaches a threshold), are reliable. These relationships mainly include: categories, tags, suppliers, etc.

[0268] S305. Construct a knowledge graph represented by triples.

[0269] After obtaining the knowledge graph, we can determine the meta-path template, perform random walks based on the meta-path template, sample meta-paths with specific relationships, model the meta-paths using the sampling Skip-Gram algorithm, and obtain the node embeddings. The embeddings of all nodes are persistently stored in a table for downstream tasks to query and obtain the embedding vectors.

[0270] In an embodiment of the present application, click-through rate estimation can be achieved by an estimation model. When training the estimation model, the server will sample negative samples from all users at a ratio of 1:3 for a given seed sample, thereby constructing a complete sample set, that is, inputting it into the model to be trained. At the same time, the model parameters are randomly initialized using the overall distribution, and the network parameters are trained using a supervised learning method with a cross-entropy loss function. The parameters are updated using the Adam algorithm, and iterated multiple times until the model converges. The final model is saved for use in the offline application stage.

[0271] In the application stage, the historical click sequences of all users (candidate objects) and candidate items are first input into the trained prediction model to calculate the score of each user and candidate item through the prediction model, and sort them from high to low.

[0272] Next, the server can select a fixed score based on experience and online testing, filter out users (target objects) with a score greater than or equal to the score, and obtain a population package for delivery; or sort the users (target objects) with the highest scores, such as the top 5 million, and filter out a population package for delivery.

[0273] Next, the effect of the advertisement delivery provided by the embodiment of the present application is described.

[0274] For example, Table 1 shows the AUC improvement comparison between the advertising delivery method provided in the embodiment of the present application and the advertising delivery method in the related art, wherein the word vector generation model is the baseline for comparison.

[0275] Table 1

[0276] Project Name AUC improvement Word vector generation model (Word2Vec) - Embedding expression enhancement via KGE 0.13％ Deep Interest Network (DIN) 0.17％ Embodiments of the present application 0.25％

[0277] As can be seen, the AUC improvement achieved by the advertising delivery method in the present embodiment is higher than that achieved by the related art methods of word vector generation models, embedding enhancement through KGE, and deep interest network advertising delivery. Therefore, the present embodiment achieves better advertising delivery results than the related art methods.

[0278] Table 2 shows a comparison of the CPA reduction of the advertising delivery method provided by the embodiment of the present application and the advertising delivery method in the related art. The word vector generation model is used as the baseline for comparison.

[0279] Table 2

[0280] Project Name Lower CPA Word vector generation model (Word2Vec) - Embedding expression enhancement via KGE 2.51％ Deep Interest Network (DIN) 2.18％ Embodiments of the present application 3.89％

[0281] As can be seen, the CPA reduction achieved by the advertising delivery method of the present application embodiment is greater than the CPA reduction achieved by the related art methods of word vector generation models, embedding enhancement through KGE, and deep interest network-based advertising delivery. Therefore, compared with the related art methods, the present application embodiment can achieve greater savings in delivery costs.

[0282] It is understandable that in the embodiments of the present application, when the embodiments of the present application are applied to specific products or technologies, the user's historical click sequence, historical interaction information and other related data need to be obtained, that is, the user's permission or consent, and the collection, use and processing of relevant data need to comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0283] The following continues to describe the exemplary structure of the information delivery device 255 provided in the embodiment of the present application as a software module. In some embodiments, such as Figure 2 As shown, the software modules stored in the information delivery device 255 of the memory 250 may include:

[0284] The information acquisition module 2551 is used to obtain the delivery information characteristics of the information to be delivered and the historical information characteristics corresponding to the historical interaction information of the candidate object;

[0285] An information mining module 2552 is configured to determine the candidate's preference information for the delivery information feature by performing correlation mining on the delivery information feature and the historical information feature;

[0286] Feature construction module 2553, configured to construct interest features of the candidate object based on the preference information and the delivery information features; the interest features describe the candidate object's interest in the information to be delivered and the delivery information features;

[0287] An object screening module 2554 is configured to screen a target object from the candidate objects based on the interest characteristics;

[0288] The information sending module 2555 is used to send the information to be delivered to the target object.

[0289] In some embodiments of the present application, the information mining module 2552 is further used to perform correlation mining on the delivery information features and the historical information features to obtain feature correlation; the feature correlation represents the influence of the historical interaction information on the preference information; based on the feature correlation, the historical information features are fused to obtain the preference information of the candidate object for the delivery information features.

[0290] In some embodiments of the present application, the information mining module 2552 is further used to perform inner product processing on the delivery information feature and the historical information feature to obtain a feature inner product result; and normalize the feature inner product result to obtain the feature correlation.

[0291] In some embodiments of the present application, the feature construction module 2553 is also used to perform correlation mining on the preference information and the delivery information features to obtain the influence weight of the preference information on the delivery information features; and determine the interest features of the candidate object based on the influence weight and the delivery information features.

[0292] In some embodiments of the present application, the feature construction module 2553 is further used to use the influence weight to perform weighted fusion on the delivery information features to obtain the fusion features corresponding to the preference information; and to perform averaging processing on the fusion features corresponding to the preference information to obtain the interest features of the candidate object.

[0293] In some embodiments of the present application, the object screening module 2554 is further used to predict the preference value of the candidate object for the information to be delivered based on the interest characteristics; and to screen the target object from the candidate objects using the preference value.

[0294] In some embodiments of the present application, the information acquisition module 2551 is further used to find the neighbor nodes of the information to be delivered from a preset knowledge graph; wherein the neighbor nodes are nodes corresponding to the auxiliary information of the information to be delivered; from a preset feature table, the features corresponding to the information to be delivered, the features corresponding to the neighbor nodes, and the historical information features corresponding to the historical interaction information of the candidate object are screened out; the features corresponding to the information to be delivered and the features corresponding to the neighbor nodes are determined as the delivery information features of the information to be delivered.

[0295] In some embodiments of the present application, the information delivery device 255 also includes: a feature integration module 2556; the feature integration module 2556 is used to construct a preset knowledge graph using target information filtered out from the information library; the target information at least includes information in the information library whose conversion times are greater than a threshold; for each entity node in the preset knowledge graph, a corresponding node sequence is sampled from the preset knowledge graph; feature encoding is performed on the node sequence to obtain the encoding features corresponding to each of the entity nodes; and the encoding features corresponding to each of the entity nodes are used to integrate into the preset feature table.

[0296] In some embodiments of the present application, the feature integration module 2556 is also used to crawl information for the target information to obtain descriptive information of the target information; perform text processing on the descriptive information to obtain auxiliary information; wherein the text processing includes at least filtering, word segmentation and deduplication; extract the association between the auxiliary information and the target information to obtain association information; use the target information and the auxiliary information as entity nodes, and use the association information to connect the entity nodes to obtain the preset knowledge graph.

[0297] In some embodiments of the present application, the feature integration module 2556 is also used to use the target information as the head entity node, the auxiliary information as other entity nodes other than the head entity node, and use the association information to connect the head entity node and the other entity nodes to obtain the preset knowledge graph.

[0298] In some embodiments of the present application, the feature integration module 2556 is further used to perform feature mapping on the node sequence to obtain initial mapping features; and perform feature extraction on the initial mapping features to obtain the coding features of each of the entity nodes.

[0299] In some embodiments of the present application, the feature integration module 2556 is further used to perform association extraction on the auxiliary information and the target information to obtain a preliminary extraction result; filter the target engine in which the preliminary extraction result appears from multiple search engines; when the proportion of the target engine in the number of the multiple search engines reaches a proportion threshold, determine the preliminary extraction result as the association information.

[0300] The present invention provides a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the information delivery method described above in the present invention.

[0301] The embodiment of the present application provides a computer-readable storage medium storing executable instructions, wherein the executable instructions are stored. When the executable instructions are executed by a processor, the processor will execute the information delivery method provided by the embodiment of the present application, for example, Figure 3 The information delivery method shown.

[0302] In some embodiments, the computer-readable storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface storage, optical disk, or CD-ROM; or various devices including one or any combination of the above memories.

[0303] In some embodiments, executable instructions may be in the form of a program, software, software module, script, or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as a stand-alone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.

[0304] As an example, executable instructions may, but need not, correspond to a file in a file system, may be stored as part of a file that stores other programs or data, such as in one or more scripts in a HyperText Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple coordinating files (e.g., files storing one or more modules, subroutines, or code portions).

[0305] As an example, executable instructions may be deployed to be executed on one computing device (information delivery device), or on multiple computing devices located at one location, or on multiple computing devices distributed at multiple locations and interconnected by a communication network.

[0306] To sum up, through the embodiments of the present application, the information delivery device can first determine the candidate object's preference for the delivery information characteristics, that is, the characteristic dimension of the information to be delivered, based on the correlation mining of the delivery information characteristics and the historical information characteristics, so as to distinguish the preferences of the candidate user objects at a finer granularity, and based on the determined preferences and delivery information characteristics, realize the modeling of the personalized interests of the user object at a finer granularity, and obtain more accurate interest characteristics of the user object. Finally, based on the interest characteristics, it can accurately screen out the delivery objects that are interested in the delivery information, and ultimately improve the accuracy of information delivery; the obtained preset knowledge graph is relatively stable and usually needs to be updated after a long time. Therefore, the information delivery device does not have to frequently construct the knowledge graph, which reduces the workload required for offline processing of information delivery and competes for computing resources.

[0307] The above description is merely an embodiment of the present application and is not intended to limit the scope of protection of the present application. Any modifications, equivalent replacements, and improvements made within the spirit and scope of the present application are included in the scope of protection of the present application.

Claims

1. An information delivery method, characterized in that: The information delivery method includes: Obtaining delivery information features of the information to be delivered and historical information features corresponding to the historical interaction information of the candidate object; Determining the preference information of the candidate object with respect to the delivery information feature by performing correlation mining on the delivery information feature and the historical information feature; performing correlation mining on the preference information and the delivery information features to obtain an influence weight of the preference information on the delivery information features, performing weighted fusion on the delivery information features using the influence weight to obtain a fused feature corresponding to the preference information, and performing averaging processing on the fused features corresponding to the preference information to obtain an interest feature of the candidate object; wherein the interest feature describes the candidate object's interest in the information to be delivered and the delivery information features; According to the interest characteristics, a target object is screened out from the candidate objects, and the information to be delivered is sent to the target object.

2. The method according to claim 1, characterized in that The determining of the candidate object's preference information for the delivery information feature by performing correlation mining on the delivery information feature and the historical information feature includes: Performing correlation mining on the delivery information feature and the historical information feature to obtain a feature correlation; the feature correlation represents the influence of the historical interaction information on the preference information; The historical information features are fused based on the feature relevance to obtain the preference information of the candidate object for the delivery information feature.

3. The method according to claim 2, characterized in that The performing correlation mining on the delivery information feature and the historical information feature to obtain feature correlation includes: Performing inner product processing on the delivery information feature and the historical information feature to obtain a feature inner product result; The feature inner product result is normalized to obtain the feature correlation.

4. The method according to any one of claims 1 to 3, characterized in that The step of selecting a target object from the candidate objects based on the interest feature includes: Predicting the candidate object's preference value for the information to be delivered based on the interest characteristics; The target object is selected from the candidate objects by using the preference value.

5. The method according to any one of claims 1 to 3, characterized in that The acquiring of the delivery information features of the information to be delivered and the historical information features corresponding to the historical interaction information of the candidate object includes: Finding neighbor nodes of the information to be delivered from a preset knowledge graph; wherein the neighbor nodes are nodes corresponding to the auxiliary information of the information to be delivered; Filtering out features corresponding to the information to be delivered, features corresponding to the neighboring nodes, and historical information features corresponding to the historical interaction information of the candidate object from a preset feature table; The features corresponding to the information to be delivered and the features corresponding to the neighboring nodes are determined as the delivery information features of the information to be delivered.

6. The method according to claim 5, characterized in that Before searching for the neighboring nodes of the information to be delivered from the preset knowledge graph, the method further includes: Using the target information screened from the information database, a preset knowledge graph is constructed; the target information at least includes information in the information database whose conversion count is greater than a threshold; For each entity node in the preset knowledge graph, sampling a corresponding node sequence from the preset knowledge graph; Performing feature encoding on the node sequence to obtain a coding feature corresponding to each entity node; The coding features corresponding to each of the entity nodes are integrated into the preset feature table.

7. The method according to claim 6, characterized in that The method of constructing a preset knowledge graph by using the target information filtered from the information database includes: Crawling the target information to obtain description information of the target information; Performing text processing on the description information to obtain auxiliary information; wherein the text processing includes at least filtering, word segmentation, and deduplication; Extracting correlation information between the auxiliary information and the target information to obtain correlation information; The target information and the auxiliary information are used as entity nodes, and the entity nodes are connected using the association information to obtain the preset knowledge graph.

8. The method according to claim 7, characterized in that The step of taking the target information and the auxiliary information as entity nodes and connecting the entity nodes using the association information to obtain the preset knowledge graph includes: The target information is used as a head entity node, the auxiliary information is used as other entity nodes other than the head entity node, and the head entity node and the other entity nodes are connected using the association information to obtain the preset knowledge graph.

9. The method according to claim 6, characterized in that The feature encoding of the node sequence to obtain the encoding feature corresponding to each entity node includes: Perform feature mapping on the node sequence to obtain initial mapping features; Feature extraction is performed on the initial mapping features to obtain the encoding features of each of the entity nodes.

10. The method according to claim 7, characterized in that The extracting of the association between the auxiliary information and the target information to obtain the association information includes: Performing correlation extraction on the auxiliary information and the target information to obtain a preliminary extraction result; Filtering a target engine that displays the preliminary extraction result from a plurality of search engines; When the target engine accounts for a proportion of the plurality of search engines reaching a proportion threshold, the preliminary extraction result is determined as the associated information.

11. An information delivery device, characterized in that: The information delivery device includes: An information acquisition module is used to obtain the delivery information features of the information to be delivered and the historical information features corresponding to the historical interaction information of the candidate object; An information mining module, configured to determine the preference information of the candidate object with respect to the delivery information feature by performing correlation mining on the delivery information feature and the historical information feature; a feature construction module configured to perform correlation mining on the preference information and the delivery information features to obtain an influence weight of the preference information on the delivery information features, perform weighted fusion on the delivery information features using the influence weight to obtain a fused feature corresponding to the preference information, and perform averaging processing on the fused features corresponding to the preference information to obtain an interest feature of the candidate object; wherein the interest feature describes the candidate object's interest in the information to be delivered and the delivery information features; An object screening module, configured to screen a target object from the candidate objects based on the interest characteristics; The information sending module is used to send the information to be delivered to the target object.

12. The device according to claim 11, characterized in that The information mining module is further configured to perform correlation mining on the delivery information features and the historical information features to obtain feature correlation; The feature relevance characterizes the influence of the historical interaction information on the preference information; the historical information features are fused based on the feature relevance to obtain the preference information of the candidate object for the delivery information feature.

13. The device according to claim 11, characterized in that The information acquisition module is further used to find the neighbor nodes of the information to be delivered from a preset knowledge graph; wherein the neighbor nodes are nodes corresponding to the auxiliary information of the information to be delivered; from a preset feature table, filter out the features corresponding to the information to be delivered, the features corresponding to the neighbor nodes, and the historical information features corresponding to the historical interaction information of the candidate object; and determine the features corresponding to the information to be delivered and the features corresponding to the neighbor nodes as the delivery information features of the information to be delivered.

14. An information delivery device, characterized in that: The information delivery device includes: a memory for storing executable instructions; The processor is configured to implement the information delivery method according to any one of claims 1 to 10 when executing the executable instructions stored in the memory.

15. A computer-readable storage medium storing executable instructions, characterized in that: When the executable instructions are executed by the processor, the information delivery method according to any one of claims 1 to 10 is implemented.

16. A computer program product comprising a computer program or instructions, characterized in that When the computer program or instruction is executed by a processor, the information delivery method according to any one of claims 1 to 10 is implemented.

Images