Information recommendation method, training method of neural network model for information recommendation

Abstract

The application provides an information recommendation method and a training method of a neural network model for information recommendation. The information recommendation method comprises the following steps: acquiring at least one piece of multi-mode information to be sent to an object, wherein each piece of multi-mode information comprises a plurality of display strategies; acquiring display strategy characteristic information corresponding to the plurality of display strategies respectively and object characteristic information of the object; performing prediction processing on each display strategy characteristic information and the object characteristic information based on a neural network model to obtain a ranking parameter corresponding to each display strategy, wherein the neural network model is obtained by training based on ranking results of a plurality of display strategy samples that have been applied; determining a target display strategy corresponding to each piece of multi-mode information; and performing a recommendation operation on the object based on at least one piece of multi-mode information to which the target display strategy is applied. Through the application, the recommendation effect can be optimized, and the utilization rate of recommended resources can be improved.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and big data technology, and in particular to an information recommendation method and a training method for a neural network model used for information recommendation. Background Technology

[0002] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, 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 attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence.

[0003] Information recommendation is an important branch of artificial intelligence, also involving big data processing using cloud technology. It primarily studies how to recommend appropriate information to a specific target audience. For multi-modal information with multiple display strategies, such as multi-creative ads with multiple creatives, the solutions provided by related technologies typically apply a random strategy, meaning all creatives in a multi-creative ad are recommended with equal probability. However, this approach can easily lead to poor recommendation performance and low utilization of recommendation resources. Summary of the Invention

[0004] This application provides an information recommendation method, a training method for a neural network model for information recommendation, an apparatus, an electronic device, a computer-readable storage medium, and a computer program product, which can optimize recommendation performance and improve the utilization rate of recommendation resources.

[0005] The technical solution of this application embodiment is implemented as follows: This application provides an information recommendation method, including: Obtain at least one multi-mode message to be sent to the object, wherein each multi-mode message includes multiple display strategies; Obtain the display strategy feature information corresponding to the multiple display strategies and the object feature information of the object; Based on the feature information of each display strategy and the feature information of the object, a neural network model is invoked to perform prediction processing to obtain the ranking parameters corresponding to each display strategy. The neural network model is trained based on the ranking results of multiple display strategy samples that have been applied. Determine the target display strategy corresponding to each piece of multi-mode information, wherein the target display strategy is the display strategy with the largest sorting parameter among the multi-mode information; Based on at least one piece of multimodal information that has the corresponding target display strategy applied, a recommendation operation is performed for the object.

[0006] This application provides an information recommendation device, including: The acquisition module is used to acquire at least one piece of multi-mode information to be sent to the object, wherein each piece of multi-mode information includes multiple display strategies; The acquisition module is further configured to acquire display strategy feature information corresponding to the plurality of display strategies and object feature information of the object; The prediction module is used to call a neural network model to perform prediction processing based on the feature information of each of the display strategies and the feature information of the object, so as to obtain the ranking parameters corresponding to each of the display strategies. The neural network model is trained based on the ranking results of multiple display strategy samples that have been applied. The determining module is used to determine the target display strategy corresponding to each piece of multi-mode information, wherein the target display strategy is the display strategy with the largest sorting parameter among the multi-mode information; The recommendation module is used to perform a recommendation operation for the object based on at least one piece of multi-modal information that has the corresponding target display strategy applied to it.

[0007] This application provides a method for training a neural network model for information recommendation, comprising: Obtain the display strategy ranking results corresponding to multiple multi-mode information samples, wherein the display strategy ranking results are obtained by sorting the display strategy samples in the multi-mode information samples based on the recommendation parameters; Multiple display strategy samples were obtained based on the ranking results of the aforementioned display strategies. The neural network model initialized based on the multiple display strategy samples is used to perform a training task to update the parameters of the neural network model; The trained neural network model is generated based on the updated parameters.

[0008] This application provides a training apparatus for a neural network model for information recommendation, comprising: The acquisition module is used to acquire the display strategy ranking results corresponding to multiple multi-mode information samples, wherein the display strategy ranking results are obtained by sorting the display strategy samples in the multi-mode information samples based on the recommendation parameters; The sampling module is used to obtain multiple display strategy samples based on the sorting results of the display strategy; An execution module is used to call the initialized neural network model based on the multiple display strategy samples to perform a training task, so as to update the parameters of the neural network model; A generation module is used to generate the trained neural network model based on the updated parameters.

[0009] This application provides an electronic device, including: Memory, used to store executable instructions; When a processor executes executable instructions stored in the memory, it implements the information recommendation method provided in the embodiments of this application, or the training method for a neural network model for information recommendation.

[0010] This application provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the information recommendation method provided in this application, or the training method for a neural network model for information recommendation.

[0011] This application provides a computer program product, including a computer program or computer-executable instructions, which, when executed by a processor, implements the information recommendation method provided in this application, or a training method for a neural network model for information recommendation.

[0012] The embodiments of this application have the following beneficial effects: The neural network model is trained based on the ranking results of multiple display strategy samples that have been applied. This enables the neural network model to learn the ability to accurately predict the recommendation parameters of multiple display strategies from the ranking results. In this way, for at least one multi-modal message to be sent, the neural network model can select the best display strategy for each multi-modal message, which enhances the targeting and accuracy of the recommendation. This makes the final selected display strategy more in line with the needs of the target audience and improves the utilization rate of recommendation resources. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the architecture of the advertising management system provided in the embodiments of this application; Figure 2 This is a schematic diagram illustrating the principle of the MAB-based advertising creative selection strategy provided in the embodiments of this application; Figure 3 This is a schematic diagram of the architecture of the information recommendation system 100 provided in an embodiment of this application; Figure 4A This is a schematic diagram of the structure of the server 200 provided in the embodiments of this application; Figure 4B This is a schematic diagram of the structure of the server 200 provided in the embodiments of this application; Figure 5This is a flowchart illustrating the training method for a neural network model for information recommendation provided in an embodiment of this application. Figure 6 This is a flowchart illustrating the information recommendation method provided in an embodiment of this application; Figure 7 This is a flowchart illustrating the information recommendation method provided in an embodiment of this application; Figure 8A This is a schematic diagram illustrating the principles of the training and prediction phases of the neural network model provided in this application embodiment; Figure 8B This is a schematic diagram illustrating the principles of the training and prediction phases of the neural network model provided in this application embodiment; Figure 9 This is a flowchart illustrating the information recommendation method provided in an embodiment of this application; Figure 10 This is a schematic diagram of the architecture of the advertising management system provided in the embodiments of this application; Figure 11A This is a schematic diagram of the structure of the tLTR model provided in the embodiments of this application; Figure 11B This is a schematic diagram of the structure of the tLTR model provided in the embodiments of this application; Figure 12 This is a schematic diagram illustrating the principle of the word embedding process provided in the embodiments of this application. Detailed Implementation

[0014] To make the objectives, 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 limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0015] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is 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.

[0016] It is understood that in the embodiments of this application, data such as user information are involved. When the embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0017] In the following description, the terms “first, second, ...” are used merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that “first, second, ...” may be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0019] Before providing a further detailed description of the embodiments of this application, the nouns and terms involved in the embodiments of this application will be explained, and the nouns and terms involved in the embodiments of this application shall be interpreted as follows.

[0020] 1) Multi-modal information: This includes information on multiple display strategies, such as multiple creative advertisements. The creative refers to the display strategy of the advertising elements, including the elements used (such as titles, descriptions, images, videos, etc.). It can also include display parameters, such as the display position and size of each element, and the number of times dynamic elements (such as GIFs) are displayed. Furthermore, it can include human-computer interaction modes, such as a link to a landing page when a text element is clicked, or a zoomed-in image when an image element is clicked.

[0021] Furthermore, ads with multiple creative elements can be further categorized into MultiCreative (MC) ads and Dynamic Creative (DC) ads. MC ads consist entirely of creative elements manually uploaded by the user (e.g., the advertiser). DC ads, on the other hand, have multiple creative elements generated by the Advertising Management System (AMS) after the advertiser uploads discrete ad elements, which are then combined according to display requirements. In each ad display, the best creative is selected from the combined elements for display. For example, if a user uploads elements in four dimensions (title, description, image, and video), the AMS can automatically perform a cross-product and store the resulting creatives in the database. For instance, if a user uploads 3 images and 3 titles (6 elements in total), the AMS can perform a cross-product to generate 3... 3 = 9 combinations, meaning 9 creatives will be added to the database. It can be seen that the difference between MC ads and DC ads is that the former's creatives are uploaded by the user after customizing and combining elements, while the latter's creatives are obtained by AMS automatically performing a cross product of elements uploaded by the user across multiple dimensions.

[0022] 2) Neural Networks (NN) Model: A neural network is a complex network system formed by the extensive interconnection of a large number of simple processing units (also known as neurons). It reflects many basic characteristics of human brain function and is a highly complex nonlinear dynamic learning system. Neural network models have the capabilities of large-scale parallelism, distributed storage and processing, self-organization, self-adaptation, and self-learning. They are particularly suitable for handling imprecise and fuzzy information processing problems that require simultaneous consideration of many factors and conditions.

[0023] 3) Learning to Rank (LTR): This method uses machine learning to train a model when dealing with ranking problems. It can be applied to information retrieval, natural language processing, and data mining. For each given query and candidate result, features are extracted, and real data labels are obtained through log mining or manual annotation. Then, the ranking model is used to obtain the most relevant results. For example, in search engines, the model learns the order of web pages in search results. In e-commerce recommendations, the model needs to learn the order of product recommendation lists. In scenarios of coarse-grained ad ranking, the model can learn the order of ads in fine-grained ranking. Common learning to rank methods can be divided into three types: single-document methods (PointWise), document-pair methods (PairWise), and document-list methods (ListWise).

[0024] 4) Response: used to indicate the conditions or states on which the operation is performed depends. When the conditions or states on which it depends are met, one or more operations can be performed in real time or with a set delay. Unless otherwise specified, there is no restriction on the order in which the multiple operations are performed.

[0025] 5) Object: Used to receive multi-modal information. It can be a real user or a virtual account, such as a user account registered on an e-commerce platform.

[0026] 6) One-Hot Encoding: Also known as one-bit valid encoding, this method uses an N-bit state register to encode N states. Each state has its own independent register bit, and at any given time, only one bit is valid. For example, for encoding six states, the corresponding one-hot encoded vectors are: 000001, 000010, 000100, 001000, 010000, and 10000.

[0027] 7) Word Embedding: This is a general term for language modeling and representation learning techniques in natural language processing. Conceptually, it refers to embedding a high-dimensional space containing the number of words into a much lower-dimensional continuous vector space, where each word or phrase is mapped to a vector in the real number field. Word embedding methods include artificial neural networks (such as Word2Vec), dimensionality reduction of word co-occurrence matrices, probabilistic models, and explicit representations of the context of words. For example, in a text containing words like "cat" and "dog," these words are mapped to vector spaces such as (0.1 0.2 0.3) for "cat" and (0.2 0.2 0.4) for "dog." This process of mapping text to a multi-dimensional vector space is called word embedding.

[0028] 8) Match: This refers to triggering as many correct results as possible from the entire information set and returning the results to the user. There are various ways to match, including collaborative filtering, topic modeling, content matching, and hot topic matching. For example, in a search system, information matching the user's input keywords can be retrieved from the entire information set. In a recommendation system, since the user does not have a specific search term, content that the user may be interested in can be recommended based on the user's interest tags, browsing history, etc.

[0029] 9) Coarse Ranking: Coarse ranking is the link following recall in the advertising management system. Its main goal is to complete the coarse selection of information from tens of thousands to hundreds through multiple strategies and models. The strategies can include a large number of diversity, relevance, cold start and exploration strategies, etc. The models can include LTR and LiteCXR models, etc. Among them, the LiteCXR model is one of the most important models in the coarse ranking stage. It can filter information through eCPM.

[0030] 10) Fine ranking: Fine ranking is the link following coarse ranking in the advertising management system. Its main goal is to accurately personalize the information and complete the selection from hundreds to single digits. For example, multiple pieces of information obtained in the coarse ranking stage can be sorted by the click-through rate (CTR) prediction model.

[0031] With the evolution of internet technology, the changing forms of products, and the actual needs of advertisers, advertising has gradually evolved from single-creative ad formats to those incorporating multiple creative elements, such as dynamic creative ads and multi-creative ads. Dynamic creative ads only require advertisers to upload original materials, and the ad management system can automatically combine them into different creative formats. This significantly lowers the barrier to entry for advertisers in creating and placing ads, giving them more opportunities for trial and error. Multi-creative ads help advertisers overcome the predicament of new ads failing to achieve significant scale by reusing data from already stable ads to increase the exposure of new creatives. At the same time, these two new ad formats also bring new technical challenges to ad management systems.

[0032] Limited by computing resources and response latency (e.g., hundreds of milliseconds), advertising management systems typically employ a phased, funnel-shaped system architecture to balance computing performance and advertising effectiveness. For example... Figure 1 As shown, when a user request arrives at the advertising management system, the system first filters tens of thousands of ads from a database of millions of ads based on certain rules, such as audience targeting, behavioral interest targeting, contextual targeting, and application (APP) installation targeting selected by the advertiser. This stage is called the recall stage. Ads recalled in the recall stage enter the coarse ranking stage. The purpose of the coarse ranking stage is to efficiently and lightweightly sort and truncate the recall queue based on the high-quality ads with exploration potential, thereby quickly selecting a small number of relatively high-quality ads for the next stage, the fine ranking stage. The fine ranking stage determines the ads that will ultimately be exposed to the user. Its main task is to rank the small number of high-quality candidate ads according to the value of the ad input (eCPM, effective cost per million) obtained per thousand impressions, and select the ad with the highest eCPM to display to the user.

[0033] When advertising in an ad management system shifts from purely single-creative ad formats to a coexistence of single-creative ads (i.e., ads containing only one creative) and ads containing multiple creatives, such as DC ads and MC ads, the current ad management system, considering the requirements of computing resources and response latency, chooses to differentiate between single-creative ads and ads containing multiple creatives during the coarse-ranking stage. Specifically, for ads containing multiple creatives, the coarse-ranking stage selects the creatives to be exposed. The relevant technologies mainly include two solutions: a creative carousel mode and an ad creative priority strategy based on Multi-Armed Bandit (MAB). These two solutions are explained below.

[0034] First, let's explain the creative carousel mode. In carousel mode, for ads containing multiple creatives, such as MC ads, the ad management system doesn't consider multiple different creatives within the same ad to have significant differences in metrics like Click-Through-Rate (CTR) and Conversion Rate (CVR). When a user request reaches the ad management system, for a particular recalled DC or MC ad, the system will randomly select one creative from the multiple creatives belonging to that ad, or select one to be displayed according to other rules, such as playing different creatives at different times, for the subsequent fine-tuning stage. The carousel mode has simple rules, requires minimal modification to the existing ad management system, and adds almost no latency; however, its effectiveness is obviously the least guaranteed.

[0035] The following section continues to explain the MAB-based creative-first strategy for advertising. For example... Figure 2 As shown, it is based on the strategy balancing exploration and utilization of multi-armed slot machines. The information used internally includes exposure, clicks, conversions, cost achievement, and random exploration factors. For each online user request, the ad management system can sample competing creatives (tid) for each ad containing multiple creatives based on the probability distribution D of the MAB strategy. ) and play creative (tid) ’ During the fine-tuning phase, competing creatives are used to rank the ads. When an ad wins, it is played using the playable creative. The specific process is as follows: 1. Extract creative performance data; 2. Calculate the creative probability using historical data of each creative using the Upper Confidence Bound algorithm; 3. Sample the creatives using the play probability. However, because different creatives compete in different environments, the actual number of plays will not match the play probability. Therefore, it is necessary to sample the playable creative and competing creatives independently, using the competing creative to compete, but exposing the playable creative.

[0036] It can be seen that the creative carousel mode cannot effectively distinguish between high-performing and low-performing creatives, which to some extent leads to a waste of advertisers' budgets. On the other hand, the ad creative priority strategy based on MAB has a complex process due to the inconsistency between competing creatives and playing creatives. It also does not take user information into account, resulting in poor creative selection and difficulty in improving and enhancing the algorithm, and poor scalability.

[0037] In view of this, embodiments of this application provide an information recommendation method, a training method for a neural network model for information recommendation, an apparatus, an electronic device, a computer-readable storage medium, and a computer program product, which can optimize recommendation performance and improve the utilization rate of recommendation resources. The exemplary application of the electronic device provided in the embodiments of this application is described below. The electronic device provided in the embodiments of this application can be implemented as a terminal device, or as a server, or implemented collaboratively by a terminal device and a server.

[0038] The following description uses the information recommendation method provided in the embodiments of this application, implemented solely by a server, as an example.

[0039] See Figure 3 , Figure 3 This is a schematic diagram of the architecture of the information recommendation system 100 provided in this application embodiment. It aims to support an application that can optimize recommendation performance and improve the utilization rate of recommendation resources, such as... Figure 3 As shown, the information recommendation system 100 includes: a server 200, a network 300, and a terminal device 400. A client 410 runs on the terminal device 400. The client 410 can be various types of clients, such as instant messaging clients, news and information reading clients, e-commerce shopping clients, browsers, etc. The network 300 can be a wide area network or a local area network, or a combination of both.

[0040] In some embodiments, server 200 may first retrieve data from a database ( Figure 3 In the process of obtaining at least one piece of multi-mode information to be sent to an object (e.g., user 1), the server 200 can then obtain, for each piece of multi-mode information, the display strategy feature information (e.g., element type corresponding to the display strategy, label of the display strategy, etc.) and the object feature information (e.g., user 1's interest tags, behavior sequence, etc.) corresponding to the multiple display strategies included in the multi-mode information, as well as the object feature information (e.g., user 1's interest tags, behavior sequence, etc.). Subsequently, the server 200 can call a neural network model to perform prediction processing based on each display strategy feature information and the object feature information to obtain the ranking parameters corresponding to each display strategy, and determine the target display strategy corresponding to each piece of multi-mode information based on the ranking parameters. Finally, the server 200 can send at least one piece of multi-mode information with the corresponding target display strategy to the terminal device 400 through the network 300.

[0041] In other embodiments, the embodiments of this application can also be implemented with the aid of cloud technology, which refers to a hosting technology that unifies a series of resources such as hardware, software, and networks within a wide area network or local area network to realize the computation, storage, processing, and sharing of data.

[0042] Cloud technology is a general term encompassing network technology, information technology, integration technology, management platform technology, and application technology based on the cloud computing business model. It can form resource pools, allowing for on-demand use with flexibility and convenience. Cloud computing technology will become a crucial support. The backend services of cloud computing systems require substantial computing and storage resources.

[0043] Example, Figure 3 The server 200 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. The terminal device 400 can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, in-vehicle terminal, etc., but is not limited to these. The terminal device 400 and the server 200 can be directly or indirectly connected via wired or wireless communication, which is not limited in this embodiment.

[0044] The following continues... Figure 3 The structure of server 200 shown in the figure will be described below. See [link to documentation]. Figure 4A , Figure 4A This is a schematic diagram of the structure of the server 200 provided in the embodiments of this application. Figure 4A The server 200 shown includes at least one processor 210, memory 240, and at least one network interface 220. The various components of server 200 are coupled together via a bus system 230. It is understood that the bus system 230 is used to implement communication between these components. In addition to a data bus, the bus system 230 also includes a power bus, a control bus, and a status signal bus. However, for clarity, ... Figure 4A The general labeled all buses as Bus System 230.

[0045] 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. Among them, the general-purpose processor can be a microprocessor or any conventional processor, etc.

[0046] The memory 240 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state storage, hard disk drives, optical disk drives, etc. The memory 240 may optionally include one or more storage devices physically located away from the processor 210.

[0047] The memory 240 may include volatile memory or non-volatile memory, or both. The non-volatile memory may be read-only memory (ROM), and the volatile memory may be random access memory (RAM). The memory 240 described in this application embodiment is intended to include any suitable type of memory.

[0048] In some embodiments, memory 240 is capable of storing data to support various operations, examples of which include programs, modules, and data structures or subsets or supersets thereof, as illustrated below.

[0049] Operating system 241 includes system programs for handling various basic system services and performing hardware-related tasks, such as the framework layer, core library layer, driver layer, etc., for implementing various basic business functions and handling hardware-based tasks; The network communication module 242 is used to reach other computing devices via one or more (wired or wireless) network interfaces 220, such as Bluetooth, WiFi, and Universal Serial Bus (USB). In some embodiments, the apparatus provided in this application can be implemented in software. Figure 4A An information recommendation device 243 stored in memory 240 is shown. This device can be software in the form of programs and plug-ins, and includes the following software modules: an acquisition module 2431, a prediction module 2432, a determination module 2433, a recommendation module 2434, a transfer module 2435, a training module 2436, a generation module 2437, a setting module 2438, and a sorting module 2439. These modules are logically connected and can therefore be arbitrarily combined or further split according to the functions they implement. It should be noted that... Figure 4A For ease of explanation, all the above modules are shown at once, but this should not be interpreted as excluding the implementation of the information recommendation device 243, which may only include the acquisition module 2431, prediction module 2432, determination module 2433, and recommendation module 2434. The functions of each module will be explained below.

[0050] In other embodiments, such as Figure 4B As shown, the memory 240 may also store a training device 244 for a neural network model for information recommendation, which may be software in the form of programs and plug-ins, including the following software modules: acquisition module 2441, sampling module 2442, execution module 2443 and generation module 2444, the functions of each module will be described below.

[0051] The following will describe in detail the information recommendation method and the training method of the neural network model for information recommendation provided in the embodiments of this application, with reference to the exemplary application and implementation of the server provided in the embodiments of this application.

[0052] Before describing the information recommendation method provided in the embodiments of this application, the structure of the neural network model provided in the embodiments of this application will first be described. The structure of the neural network model provided in the embodiments of this application may include... Figure 11A The twin-tower structure shown, or Figure 11B The single-tower structure shown allows the server to select based on the data volume of at least one multi-mode message to be sent. Figure 11A The twin-tower structure shown or Figure 11B The single-tower structure shown performs predictive processing; for example, when the data volume exceeds a data volume threshold, the server can call... Figure 11A The illustrated dual-tower structure is used for prediction processing to accelerate computation; however, when the data volume is less than a data volume threshold, the server can call... Figure 11B The single-tower structure shown is subjected to predictive processing.

[0053] To make the information recommendation method provided in the embodiments of this application easier to understand, before describing the information recommendation method provided in the embodiments of this application, the training method of the neural network model for information recommendation provided in the embodiments of this application will be described first.

[0054] See Figure 5 , Figure 5 This is a flowchart illustrating the training method for a neural network model for information recommendation provided in this application embodiment, which will be combined with... Figure 5 The steps shown are explained.

[0055] In step 101, the ranking results of the display strategies corresponding to multiple multi-mode information samples are obtained.

[0056] Here, the ranking result of the display strategy can be obtained by ranking the display strategy samples based on the recommendation parameters (i.e., the parameters used to quantify the recommendation effect of the display strategy samples) in the multimodal information sample.

[0057] In some embodiments, the server can obtain historical recommendation records for the sample object and obtain multiple multimodal information samples from the historical recommendation records (e.g., obtaining multiple multimodal information samples recommended to the sample object in the past month). Then, for each multimodal information sample, the server can further obtain the display strategy ranking result corresponding to each multimodal information sample. Here, the display strategy ranking result can be the ranking result of multiple display strategy samples in the fine ranking stage (the fine ranking stage is the most important stage to determine whether the display strategy sample can be exposed, and can accurately reflect the recommendation parameters of the display strategy sample), that is, the ranking result of multiple display strategy samples that have been applied.

[0058] In step 102, multiple display strategy samples are obtained based on the display strategy ranking results.

[0059] In some embodiments, taking multi-mode information sample A as an example, assuming that multi-mode information sample A includes 10 display strategy samples, namely display strategy sample 1, display strategy sample 2, display strategy sample 3, ..., display strategy sample 10, and the display strategy ranking result of these 10 display strategy samples is: display strategy sample 2, display strategy sample 1, display strategy sample 4, display strategy sample 3, display strategy sample 5, display strategy sample 7, display strategy sample 6, display strategy sample 10, display strategy sample 8, and display strategy sample 9, then the server can sample the display strategy ranking result of these 10 display strategy samples (for example, the server can read the identifiers of display strategy samples that meet the proportion from the display strategy ranking result and use the display strategy sample corresponding to the identifier as the sampling result). For example, assuming that the server reads the identifiers of display strategy sample 2, display strategy sample 4, and display strategy sample 5 from the display strategy ranking result of these 10 display strategy samples, then display strategy sample 2, display strategy sample 4, and display strategy sample 5 can be used as the sampling result, that is, the server can perform subsequent processing on the sampled display strategy sample 2, display strategy sample 4, and display strategy sample 5.

[0060] In step 103, the initialized neural network model is invoked based on multiple display strategy samples to perform a training task in order to update the parameters of the neural network model.

[0061] In some embodiments, the server can implement the above-mentioned training task of calling the initialized neural network model based on multiple display strategy samples to update the parameters of the neural network model in the following manner: performing pairwise combination processing on multiple display strategy samples to obtain multiple display strategy sample pairs; comparing the recommendation parameters of the two display strategy samples included in each display strategy sample pair, and generating label data for each display strategy sample pair according to the comparison results; and calling the initialized neural network model to perform a first training task based on the multiple display strategy sample pairs and the corresponding label data; wherein, the first training task includes: performing prediction processing on the initialized neural network model based on the display strategy samples, and updating the parameters of the neural network model based on the difference between the comparison results obtained from the prediction and the label data.

[0062] It should be noted that, in addition to using display strategy sample pairs to train the neural network model, a single display strategy sample can also be used to train the neural network model. For example, after sampling multiple display strategy samples, the label data corresponding to each display strategy sample can be determined. Then, based on the multiple display strategy samples and the corresponding label data, the initialized neural network model is called to perform the second training task. The second training task includes: calling the initialized neural network model to perform prediction processing based on the display strategy samples, and updating the parameters of the neural network model based on the difference between the prediction results and the label data. This application embodiment does not limit the training method of the neural network model.

[0063] In step 104, a trained neural network model is generated based on the updated parameters.

[0064] In some embodiments, after the server calls the initialized neural network model to perform a training task based on multiple display strategy samples to update the parameters of the neural network model, it can generate a trained neural network model based on the updated parameters. Subsequently, the server can call the trained neural network model to score and predict the multiple display strategies included in at least one multi-mode message to be sent, thereby selecting the best display strategy for application.

[0065] The following will continue to combine Figure 6 The information recommendation method provided in the embodiments of this application will be described.

[0066] See Figure 6 , Figure 6 This is a flowchart illustrating the information recommendation method provided in the embodiments of this application, which will be combined with... Figure 6 The steps shown are explained.

[0067] In step 201, at least one multi-mode message to be sent to the object is obtained.

[0068] Here, the target audience can refer to users of various online applications, such as users of e-commerce shopping applications or users of news and information reading applications. Each multimodal message includes multiple display strategies, and the elements used by different display strategies are not exactly the same. For example, taking a multimodal message as a multi-creative advertisement with multiple creatives as an example, suppose multi-creative advertisement 1 includes 4 creatives, namely creative 1, creative 2, creative 3 and creative 4. Creative 1 includes title 1 and image 1, creative 2 includes title 2 and image 1, creative 3 includes title 1 and image 2, and creative 4 includes title 2 and image 2.

[0069] In some embodiments, the server may obtain at least one piece of multimodal information to be sent to the object by: obtaining multimodal information that meets at least one of the following conditions: matching the object feature information of the object (e.g., user interest tags, behavior sequences, browsing history, etc.); having a similarity (e.g., cosine similarity) between the multimodal information and the object feature vector of the object that is less than a similarity threshold; sorting multiple pieces of multimodal information based on the predicted click-through rate, and filtering out at least one piece of multimodal information that ranks first from the sorting results.

[0070] For example, taking multi-modal information as a multi-creative ad, the server can first recall multi-creative ads from the ad library that match the user's interest tags or keywords, such as multi-creative ads matching the region, gender, age, etc. in the ad targeting conditions; of course, the server can also perform word embedding processing by selecting effective original features, for example, through similar Figure 11A The illustrated dual-tower structure learns vector representations of users and advertisements respectively. Then, it calculates vector similarity using methods such as dot product, cosine similarity, or Euclidean distance to recall multiple creative advertisements from the advertisement library whose feature vectors have a similarity of less than a similarity threshold with the user. After recalling multiple creative advertisements from the advertisement library, the server can also perform coarse ranking on the recalled multiple creative advertisements. For example, it can rank the multiple creative advertisements based on the predicted click-through rate and select at least one of the top-ranked creative advertisements from the ranking results.

[0071] The following example uses multi-modal information as a multi-creative advertisement to illustrate the generation process of multiple creatives included in a multi-creative advertisement. When multi-modal information is used as a multi-creative advertisement, the target audience can be the advertiser's intended users. In some embodiments, before executing step 201, the server may also perform any of the following processes: obtaining a material corresponding to each of the multiple parts to be filled in the multi-modal information, and filling each material into the corresponding part to be filled to obtain a display strategy for the multi-modal information; obtaining multiple materials corresponding to each of the multiple parts to be filled in the multi-modal information, and randomly selecting one material from the multiple materials corresponding to each part to be filled, and filling the selected material into the corresponding part to be filled to obtain a display strategy for the multi-modal information.

[0072] This application provides two methods for obtaining multi-modal information. The first method involves obtaining a material corresponding to each of the multiple parts to be filled in the multi-modal information, and filling each material into its corresponding part to obtain a display strategy for the multi-modal information. Taking a multi-creative advertisement as an example, assuming the parts to be filled in the multi-creative advertisement include an image part and a text part, for the image part, assuming image material 1 uploaded by the advertiser is obtained; for the text part, assuming text material 1 uploaded by the advertiser is obtained, the server can fill the image part with image material 1 and the text part with text material 1, obtaining one creative for the multi-creative advertisement. The advertiser can continue to create the next creative for the multi-creative advertisement by uploading materials.

[0073] The second approach involves acquiring multiple materials corresponding to the various parts of the multimodal information to be filled, and randomly selecting one material from these materials for each part. This selected material is then used to fill the corresponding part, resulting in a display strategy for the multimodal information. For example, for the image part of an advertisement, assuming we acquire image material 1 and image material 2 uploaded by the advertiser; and for the text part, assuming we acquire text material 1 and text material 2 uploaded by the advertiser, the server can fill the image part with a randomly selected image material (e.g., image material 1) and the text part with a randomly selected text material (e.g., text material 1), thus generating a creative for the advertisement. It's worth noting that this approach can exhaustively explore all possible combinations of materials corresponding to different parts to be filled, until all possible creatives are obtained. For example, it can generate four different creatives: (image material 1, text material 1), (image material 1, text material 2), (image material 2, text material 1), and (image material 2, text material 2). This increases the number of creatives obtained, thereby enhancing the flexibility of generating multimodal information.

[0074] It should be noted that multi-creative advertising is just one specific form of multi-modal information. Multi-modal information can also include news with multiple display strategies, or animated images (GIFs), etc.

[0075] In step 202, the display strategy feature information and object feature information of the objects corresponding to the multiple display strategies are obtained.

[0076] In some embodiments, for each multi-mode information obtained in step 201, the server may obtain display strategy feature information corresponding to the multiple display strategies included in the multi-mode information, such as the basic attributes of the display strategy (e.g., the element type corresponding to the display strategy), the tags of the display strategy (used to characterize the domain to which the multi-mode information belongs), and the statistical features of the display strategy (e.g., the predicted click-through rate of the display strategy). In addition, the server may also obtain object feature information of the object, such as the object's interest tags, behavior sequences, browsing history, etc.

[0077] For example, taking a multi-creative ad with multiple creatives as an example, suppose multi-creative ad 1 includes 4 creatives, namely creative 1, creative 2, creative 3 and creative 4. The server can obtain the creative feature information corresponding to each creative, such as the basic attributes of the creative (e.g. the number and type of elements used in the creative), creative tags (used to represent the type of ad to which the creative belongs, i.e., which field the ad belongs to), statistical features (e.g. the predicted click-through rate of the creative), etc.

[0078] In step 203, based on the feature information of each display strategy and the feature information of the object, a neural network model is invoked to perform prediction processing to obtain the ranking parameters corresponding to each display strategy.

[0079] Here, the neural network model is trained based on the ranking results of multiple display strategy samples that have been applied. For example, it could be trained using... Figure 5 The neural network model obtained by training in steps 101 to 104 is shown.

[0080] In some embodiments, a neural network model can be trained using demonstration strategy sample pairs, then the server executes... Figure 6 Before step 203 shown, the following steps can also be performed: Figure 7 Steps 206 to 210 shown will combine Figure 7 The steps shown are explained.

[0081] In step 206, the ranking results of the display strategies corresponding to the multiple multi-mode information samples are obtained.

[0082] Here, the display strategy ranking result is obtained by sorting the display strategy samples in the multi-modal information sample based on the recommendation parameters. For example, multiple display strategy samples in the fine-tuning stage can be sorted to obtain the display strategy ranking result.

[0083] In step 207, multiple display strategy samples are obtained based on the sorting results of the display strategy samples.

[0084] In some embodiments, the server may implement step 207 above in the following manner: determine the sampling ratio according to the training speed of the neural network model, wherein the ratio is negatively correlated with the training speed; sample the display strategy ranking result corresponding to each multi-mode information sample according to the ratio to obtain multiple display strategy samples. For example, the server may read the identifier (e.g., ID) of the display strategy sample that meets the ratio from the display strategy ranking result as the sampling result (e.g., assuming the server reads the identifier of display strategy sample 1 from the display strategy ranking result, then display strategy sample 1 can be used as the sampling result). The reading method may be sequential reading or random reading.

[0085] For example, in order to improve the training speed of the neural network model, the sampling ratio can be appropriately reduced. For instance, only 20% of the display strategy samples can be sampled from the ranking results for subsequent model training. Of course, in order to improve the prediction accuracy of the neural network model (i.e., the training speed of the neural network model can be slower), the sampling ratio can be appropriately increased. For instance, 50% of the display strategy samples can be sampled from the ranking results for subsequent model training. Those skilled in the art can determine the sampling ratio according to actual needs, and the embodiments of this application do not make specific limitations on this.

[0086] In step 208, multiple display strategy samples are combined in pairs to obtain multiple display strategy sample pairs.

[0087] In some embodiments, after sampling the sorting results of multiple display strategy samples in each multi-mode information sample to obtain multiple display strategy samples, the server can also perform pairwise combination processing on the sampled multiple display strategy samples to obtain multiple display strategy sample pairs.

[0088] For example, taking multi-modal information sample A as an example, suppose the server samples the sorting results of multiple display strategy samples included in multi-modal information sample A, and obtains 4 display strategy samples, namely display strategy sample 1, display strategy sample 2, display strategy sample 3 and display strategy sample 4. Then the server can perform pairwise combination processing on these 4 display strategy samples to obtain 6 display strategy sample pairs, namely: (display strategy sample 1, display strategy sample 2), (display strategy sample 1, display strategy sample 3), (display strategy sample 1, display strategy sample 4), (display strategy sample 2, display strategy sample 3), (display strategy sample 2, display strategy sample 4) and (display strategy sample 3, display strategy sample 4).

[0089] In step 209, the recommended parameters of the two display strategy samples included in each display strategy sample pair are compared, and label data for each display strategy sample pair is generated based on the comparison results.

[0090] Here, recommendation parameters refer to parameters used to quantify the recommendation effect.

[0091] In some embodiments, after comparing the recommendation parameters of the two display strategy samples included in each display strategy sample pair, the server can generate the label data for each display strategy sample pair based on the comparison result as described above in the following manner: Perform the following processing for each display strategy sample pair: when the recommendation parameter of the first display strategy sample in the display strategy sample pair is greater than the recommendation parameter of the second display strategy sample, determine the label data of the display strategy sample pair as 1; when the recommendation parameter of the first display strategy sample in the display strategy sample pair is less than the recommendation parameter of the second display strategy sample, determine the label data of the display strategy sample pair as 0.

[0092] For example, taking display strategy sample pair A as an example, assuming that display strategy sample pair A includes display strategy sample 1 and display strategy sample 2, when the recommendation parameter of display strategy sample 1 is greater than the recommendation parameter of display strategy sample 2 (for example, the ranking of display strategy sample 1 in the ranking result is greater than that of display strategy sample 2, that is, the recommendation effect of display strategy sample 1 is better than that of display strategy sample 2), the server can determine the label data of display strategy sample pair A as 1; when the recommendation parameter of display strategy sample 1 is less than the recommendation parameter of display strategy sample 2 (for example, the ranking of display strategy sample 1 in the ranking result is less than that of display strategy sample 2, that is, the recommendation effect of display strategy sample 2 is better than that of display strategy sample 1), the server can determine the label data of display strategy sample pair A as 0.

[0093] In step 210, a neural network model is trained based on multiple display strategy sample pairs and their corresponding label data.

[0094] In some embodiments, when the neural network model includes multiple hidden layers, and these hidden layers are cascaded sequentially to form a pyramid network, the server may perform the following processing before training the neural network model based on multiple display strategy sample pairs and their corresponding label data: setting a mirror network of the pyramid network in the neural network model, wherein the pyramid network is used to receive the display strategy feature information of one display strategy sample in the display strategy sample pair, and the mirror network is used to receive the display strategy feature information of the other display strategy sample in the display strategy sample pair. It should be noted that the structure and parameters of the mirror network are completely identical to those of the pyramid network; that is, the mirror network and the pyramid network are two completely identical networks.

[0095] Taking display strategy sample pair A as an example, assuming display strategy sample pair A includes display strategy sample 1 and display strategy sample 2, during the training phase of the neural network model, the server can first set up a mirror network of the pyramid network in the neural network model. Then, the first concatenation vector (i.e., the vector obtained by concatenating the display strategy feature vector and the object feature vector of display strategy sample 1) can be input into the pyramid network to obtain the ranking parameters of display strategy sample 1. The second concatenation vector (i.e., the vector obtained by concatenating the display strategy feature vector and the object feature vector of display strategy sample 2) can be input into the mirror network to obtain the ranking parameters of display strategy sample 2. Next, the server can determine the difference between the ranking parameters of display strategy sample 1 and the ranking parameters of display strategy sample 2, and substitute the difference and the label data of display strategy sample pair A into the loss function to determine the corresponding error. Finally, the gradient of the neural network model is determined based on the error, and the parameters of the neural network model are updated based on the gradient. After training the neural network model based on multiple display strategy sample pairs and their corresponding label data, the server can also perform the following processing: remove the mirror network from the neural network model. In other words, during the prediction phase, the server can simply encode the concatenated vector (i.e., the vector obtained by concatenating the display strategy feature vector and the object feature vector) through a tower network to obtain the ranking parameters corresponding to each display strategy.

[0096] For example, see Figure 8A , Figure 8A This is a schematic diagram illustrating the principles of the training and prediction phases of the neural network model provided in this application embodiment, as shown below. Figure 8AAs shown, during the training phase of the neural network model, the server can set up a mirror network 802 of the tower network 801 in the neural network model. The tower network 801 and the mirror network 802 encode the display strategy feature vectors of the two display strategy samples in the display strategy sample pair, respectively. After training the neural network model based on multiple display strategy sample pairs and their corresponding label data, the server can remove the mirror network 802 from the neural network model. It can be seen that there is a significant difference in the structure of the neural network model between the training and prediction phases. During the training phase, input can be based on display strategy sample pairs, while during the prediction phase, input is not required based on display strategy pairs; only the display strategy feature vector of a single display strategy needs to be input. In other words, the prediction phase uses half the size of the model from the training phase. Furthermore, Figure 8A The loss function shown can be of various types, such as cross-entropy loss function, hinge loss function, etc.

[0097] In other embodiments, when the hidden layers of the neural network model are divided into two groups, with the hidden layers of the first group cascaded sequentially to form a first tower network and the hidden layers of the second group cascaded sequentially to form a second tower network, the server may perform the following processing before training the neural network model based on multiple display strategy sample pairs and corresponding label data: setting a mirror network of the second tower network in the neural network model, wherein the second tower network is used to receive the display strategy feature information of one display strategy sample in the display strategy sample pair, and the mirror network is used to receive the display strategy feature information of the other display strategy sample in the display strategy sample pair. It should be noted that the structure and parameters of the mirror network are completely identical to those of the second tower network; that is, the mirror network and the second tower network are two completely identical networks.

[0098] Taking display strategy sample pair A as an example, assuming display strategy sample pair A includes display strategy sample 1 and display strategy sample 2, during the training phase of the neural network model, the server can first set up a mirror network of the second tower network in the neural network model. Then, the display strategy feature information of display strategy sample 1 can be input into the second tower network to obtain the ranking parameters of display strategy sample 1, and the display strategy feature information of display strategy sample 2 can be input into the mirror network to obtain the ranking parameters of display strategy sample 2. Next, the server can determine the difference between the ranking parameters of display strategy sample 1 and the ranking parameters of display strategy sample 2, and substitute the difference and the label data of display strategy sample pair A into the loss function to determine the corresponding error. Finally, the gradient of the neural network model is determined based on the error, and the parameters of the neural network model are updated based on the gradient. After training the neural network model based on multiple display strategy sample pairs and their corresponding label data, the server can also perform the following processing: remove the mirror network from the neural network model. That is, during the prediction phase, the server can encode the display strategy feature vector of each display strategy only through the second tower network to obtain the ranking parameters corresponding to each display strategy.

[0099] For example, see Figure 8B , Figure 8B This is a schematic diagram illustrating the principles of the training and prediction phases of the neural network model provided in this application embodiment, as shown below. Figure 8B As shown, during the training phase of the neural network model, the server can set up a mirror network 804 of the second tower network 803 in the neural network model, and encode the display strategy feature vectors of the two display strategy samples included in the display strategy sample pair through the second tower network 803 and the mirror network 804 respectively. After training the neural network model based on multiple display strategy sample pairs and corresponding label data, the server can remove the mirror network 804 from the neural network model. That is, in the prediction phase, the model in the training phase is reduced by half for use.

[0100] It should be noted that, in addition to using display strategy samples to train the neural network model, a single display strategy sample can also be used to train the neural network model. For example, a neural network model can be trained based on multiple display strategy samples and the corresponding label data for each display strategy sample. The label data of each display strategy sample is used to represent the ranking of the display strategy sample in the ranking result. For example, for display strategy samples with higher rankings, the corresponding label data value is also larger. For example, if display strategy sample 2 ranks second in the ranking result, its corresponding label data value can be 0.9; if display strategy sample 3 ranks fifth in the ranking result, its corresponding label data value can be 0.7. That is to say, when using a single display strategy sample for training, there is no need to set up an additional mirror network in the neural network model during the training phase. That is, the structure of the neural network model is the same in the training phase and the prediction phase. This application embodiment does not specifically limit the training method of the neural network model.

[0101] In other embodiments, the neural network model may include an input layer, at least one hidden layer, and an output layer. Figure 6 Step 203 shown can be achieved through Figure 9 Steps 2031 to 2033 shown are implemented in conjunction with Figure 9 The steps shown are explained.

[0102] For each display strategy feature, the neural network model is invoked to perform the following processing: In step 2031, the display strategy feature information and the object feature information are vectorized respectively to obtain the display strategy feature vector of the display strategy and the object feature vector of the object.

[0103] In some embodiments, for each display strategy feature information, the server can call the input layer of the neural network model to perform vectorization processing based on the display strategy feature information and the object feature information to obtain the display strategy feature vector of the display strategy and the object feature vector of the object.

[0104] For example, the server can vectorize the display strategy feature information in the following way to obtain the display strategy feature vector: call the input layer of the neural network model to perform one-hot encoding on the display strategy feature information to obtain the first one-hot encoded vector; after obtaining the first one-hot encoded vector, the server can then call the input layer of the neural network model to perform word embedding on the first one-hot encoded vector, and determine the obtained first word embedding vector as the display strategy feature vector of the display strategy.

[0105] For example, the server can vectorize object feature information to obtain the object feature vector by calling the input layer of the neural network model to perform one-hot encoding on the object feature information to obtain a second one-hot encoded vector; after obtaining the second one-hot encoded vector, the server can continue to call the input layer of the neural network model to perform word embedding on the second one-hot encoded vector, and determine the obtained second word embedding vector as the object feature vector.

[0106] In other embodiments, the input layer of the neural network model may include sub-input layers corresponding to multiple domains (e.g., a first sub-input layer corresponding to the education industry, a second sub-input layer corresponding to the e-commerce industry, etc.). The server can then perform the one-hot encoding of the object feature information as described above to obtain the second one-hot encoded vector in the following way: determine the domain to which the object belongs; call the sub-input layer corresponding to the domain to perform one-hot encoding of the object feature information to obtain the second one-hot encoded vector. For example, assuming that the user's domain is determined to be the education industry, the server can call the first sub-input layer corresponding to the education industry to perform one-hot encoding of the user's feature information to obtain the second one-hot encoded vector.

[0107] It should be noted that, according to the characteristics of the traffic scenario, the embodiments of this application can also add a new sub-input layer corresponding to the domain to the input layer of the neural network model. For example, when it is necessary to recommend multi-mode information to users in the manufacturing industry, a third sub-input layer corresponding to the manufacturing industry can be added to the input layer of the neural network model, which has strong scalability.

[0108] In step 2032, the display strategy feature vector and the object feature vector are encoded to obtain the hidden layer feature vector.

[0109] In some embodiments, the encoding process can be implemented by calling at least one hidden layer of the neural network model. When there are multiple hidden layers, they can be cascaded to form a tower network. The server can then implement step 1032 as follows: The server first concatenates the display strategy feature vector and the object feature vector to obtain a concatenated vector. For example, assuming the display strategy feature vector is... E = (e1, e2, e3), the object feature vector is D = (d1, d2), then the feature vector of the display strategy E and object feature vector D The concatenation process is performed, resulting in a concatenated vector of (e1, e2, e3, d1, d2). After obtaining the concatenated vector, the server can call the tower network to encode the concatenated vector to obtain the hidden layer feature vector.

[0110] For example, the server can implement the above-mentioned call to the pyramid network to encode the concatenated vector and obtain the hidden feature vector in the following way: encode the concatenated vector through the first hidden layer of the pyramid network; input the encoding result of the first hidden layer into the subsequent cascaded hidden layers, and continue to encode through the subsequent cascaded hidden layers until the last hidden layer of the pyramid network; determine the encoding result of the last hidden layer as the hidden feature vector. Taking a network with four hidden layers as an example, namely Hidden Layer 1, Hidden Layer 2, Hidden Layer 3, and Hidden Layer 4, these four hidden layers can be cascaded to form a tower network. The server can first encode the concatenated vector through Hidden Layer 1 of the tower network, then input the encoded result of Hidden Layer 1 into Hidden Layer 2, so that Hidden Layer 2 encodes the encoded result of Hidden Layer 1. Subsequently, the encoded result of Hidden Layer 2 can be input into Hidden Layer 3, so that Hidden Layer 3 encodes the encoded result of Hidden Layer 2. Finally, the encoded result of Hidden Layer 3 can be input into Hidden Layer 4, so that Hidden Layer 4 encodes the encoded result of Hidden Layer 3, and the encoded result of Hidden Layer 4 is determined as the hidden layer feature vector.

[0111] In other embodiments, before encoding the display strategy feature vector and the object feature vector to obtain the hidden layer feature vector, the server may also perform the following processing: obtain the data volume of at least one piece of multi-mode information; in response to the data volume being less than the data volume threshold, proceed to the step of calling the tower network for encoding processing based on the display strategy feature vector and the object feature vector.

[0112] For example, the server can determine whether to use a single-tower network or a dual-tower network for encoding processing based on the characteristics of different traffic scenarios (such as data volume, time consumption, etc.). For instance, when the server determines that the data volume of at least one multi-mode message to be sent is less than the data volume threshold, it can concatenate the display strategy feature vector and the object feature vector, and then call the aforementioned tower network (i.e., single-tower network) to encode the concatenated vector. In this way, while ensuring the time consumption, the interaction between the display strategy feature vector and the object feature vector can be fully realized, thereby obtaining more accurate sorting parameters.

[0113] In other embodiments, the encoding process can be implemented by calling at least one hidden layer of the neural network model. When there are multiple hidden layers, they can be divided into two groups. The hidden layers in the first group are cascaded to form a first pyramid network, and the hidden layers in the second group are cascaded to form a second pyramid network. The server can then implement step 1032 as follows: For each display strategy feature vector, perform the following processing: call the first pyramid network to encode the object feature vector to obtain the first sub-hidden layer feature vector; call the second pyramid network to encode the display strategy feature vector to obtain the second sub-hidden layer feature vector; perform a dot product on the first and second sub-hidden layer feature vectors to obtain the hidden layer feature vector. The dot product, also known as the dot product or scalar product, is a binary operation that accepts two vectors on a real number R and returns a real scalar value. It is the standard inner product in Euclidean space. For example, for two vectors a = [a1, a2, ..., a2], ... n ] and b=[b1, b2, ..., b n The dot product of a is defined as: a·b = a1b1 + a2b2 + ... + a n b n .

[0114] For example, taking a network with at least six hidden layers (hidden layer 1, hidden layer 2, hidden layer 3, hidden layer 4, hidden layer 5, and hidden layer 6), these six hidden layers can be divided into two groups. For instance, hidden layers 1 to 3 can be grouped into one group, and hidden layers 4 to 6 into another. The hidden layers in the first group (hidden layers 1 to 3) can be cascaded to form a first pyramid network, and the hidden layers in the second group (hidden layers 4 to 6) can be cascaded to form a second pyramid network. The server can then use the first pyramid network to encode the object's feature vector. For example, the server can first encode the object's feature vector using hidden layer 1 of the first pyramid network, then input the encoded result from hidden layer 1 into hidden layer 2, so that hidden layer 2 can continue to encode the result from hidden layer 1. Subsequently, the encoded result from hidden layer 2 is input into hidden layer 3, so that hidden layer 3 can encode the result from hidden layer 2, and then... The encoding result output by hidden layer 3 is determined as the first sub-hidden layer feature vector. For each display strategy feature vector, the server can call the second tower network to encode the display strategy feature vector. For example, the server can first encode the display strategy feature vector through hidden layer 4 of the second tower network, and input the encoding result output by hidden layer 4 into hidden layer 5 so that hidden layer 5 encodes the encoding result output by hidden layer 4. Then, the encoding result output by hidden layer 5 is input into hidden layer 6 so that hidden layer 6 encodes the encoding result output by hidden layer 5, and the encoding result output by hidden layer 6 is determined as the second sub-hidden layer feature vector. After obtaining the first sub-hidden layer feature vector (i.e., the hidden layer feature vector obtained by encoding the object feature vector) and the second sub-hidden layer feature vector (i.e., the hidden layer feature vector obtained by encoding the display strategy feature vector), the server can perform a dot product operation on the first sub-hidden layer feature vector and the second sub-hidden layer feature vector, and determine the dot product result as the hidden layer feature vector.

[0115] In other embodiments, before the server encodes the display strategy feature vector and the object feature vector to obtain the hidden layer feature vector, it may also perform the following processing: obtain the data volume of at least one piece of multi-mode information; in response to the data volume being greater than or equal to the data volume threshold, proceed to the step of encoding based on the object feature vector and the display strategy feature vector by calling the first tower network and the second tower network respectively.

[0116] For example, the server can determine whether to call a single-tower network or a dual-tower network for encoding processing based on the characteristics of different traffic scenarios (such as data volume, time consumption, etc.). For instance, when the server determines that the data volume of at least one multi-mode message to be sent is greater than the data volume threshold, it can call a dual-tower network (i.e., the first tower network and the second tower network mentioned above) to encode the object feature information and the display strategy feature information respectively, so as to improve the computational efficiency and save the time consumed in the scoring of the display strategy.

[0117] In step 2033, the hidden layer feature vectors are activated to obtain the ranking parameters of the display strategy.

[0118] In some embodiments, after obtaining the hidden layer feature vector, the server can, based on the hidden layer feature vector, call the output layer of the neural network model for activation processing to obtain the ranking parameters of the display strategy (parameters used to quantify the recommendation effect of the display strategy). For example, the output layer of the neural network model can include various types of non-linear activation functions (e.g., ReLU, sigmoid, tanh function, etc.), that is, after obtaining the hidden layer feature vector, the server can call the activation function included in the output layer of the neural network model to activate the hidden layer feature vector to obtain the ranking parameters of the display strategy (e.g., score).

[0119] See also Figure 6 In step 204, the target display strategy corresponding to each multi-mode information is determined.

[0120] Here, the target display strategy is the display strategy with the largest ranking parameter among the multi-modal information.

[0121] In some embodiments, taking a total of three multi-mode information items as an example, let's assume they are multi-mode information 1, multi-mode information 2, and multi-mode information 3. Multi-mode information 1 includes three display strategies: display strategy 1, display strategy 2, and display strategy 3; multi-mode information 2 includes two display strategies: display strategy 4 and display strategy 5; and multi-mode information 3 includes four display strategies: display strategy 6, display strategy 7, display strategy 8, and display strategy 9. Assuming that display strategy 1 has the largest sorting parameter in multi-mode information 1, display strategy 4 has the largest sorting parameter in multi-mode information 2, and display strategy 9 has the largest sorting parameter in multi-mode information 3, then display strategy 1 can be determined as the target display strategy for multi-mode information 1 (i.e., multi-mode information 1 will be displayed using display strategy 1), display strategy 4 can be determined as the target display strategy for multi-mode information 2 (i.e., multi-mode information 2 will be displayed using display strategy 4), and display strategy 9 can be determined as the target display strategy for multi-mode information 3 (i.e., multi-mode information 3 will be displayed using display strategy 9).

[0122] In step 205, a recommendation operation is performed on the target based on at least one piece of multimodal information that has a corresponding target display strategy applied.

[0123] In some embodiments, when there are multiple pieces of at least one piece of multi-mode information, before the server performs the recommendation operation for the target based on at least one piece of multi-mode information with corresponding target display strategies, the server may also perform the following processing: sorting the multiple pieces of multi-mode information with corresponding target display strategies according to any of the following sorting strategies: sorting the multiple pieces of multi-mode information with corresponding target display strategies according to information type; randomly sorting the multiple pieces of multi-mode information with corresponding target display strategies; sorting the multiple pieces of multi-mode information with corresponding target display strategies according to operating rules.

[0124] For example, taking multi-modal information as multi-creative ads, after determining the best creative (i.e., the target display strategy) in each multi-creative ad, the ads can be sorted. For instance, when the multi-creative ads include different types of ads such as e-commerce and education, for visual appeal, they can be categorized and grouped together according to information type, making it easier for users to find them and improving the user's visual experience. Of course, to explore new user interests, the multi-creative ads can also be randomly sorted and pushed to users. In addition, the multi-creative ads can be sorted according to operational rules. For example, to promote product A, multi-creative ads targeting product A can be placed at the top, thereby greatly increasing the exposure of product A's ads.

[0125] The information recommendation method provided in this application training method trains a neural network model based on the ranking results of multiple display strategy samples that have been applied. This enables the trained neural network model to score multiple display strategies included in the multi-modal information to be sent, and then selects the best display strategy to recommend to the user. This enhances the targeting and accuracy of the recommendation, making the final selected display strategy more in line with the needs of the user and improving the utilization rate of recommendation resources.

[0126] The following example illustrates an exemplary application of this application in a real-world application scenario, using multi-modal information as an example of an advertisement that includes multiple creative ideas (such as DC ads and MC ads).

[0127] This application provides an information recommendation method that employs a ranking learning approach. A neural network model is constructed to learn the eCPM ranking results of multiple creatives during the fine-tuning stage. Simultaneously, by utilizing a group approach, in the creative priority task, each advertisement constitutes a group. The tLTR model (corresponding to the aforementioned neural network model) only needs to rank different creatives within a single advertisement, without needing to consider the ranking between different advertisements. Thus, compared to the MAB-based ad creative priority strategy provided by related technologies, the solution provided in this application considers the information requested by the user and is an online, real-time personalized creative priority model, resulting in a significantly higher optimization effect than the MAB-based ad creative priority strategy. Furthermore, the method provided in this application can be improved and enhanced according to different scenarios. For example, due to significant differences in traffic scenarios and audience distribution, the model can be split by site (i.e., the model's input layer can be split into multiple sub-input layers corresponding to different domains). Additionally, different features can be added to the model based on traffic characteristics (i.e., new sub-input layers corresponding to new domains can be added to the model's input layer), demonstrating strong scalability.

[0128] The information recommendation method provided in the embodiments of this application will be described in detail below.

[0129] DC (Distributed Advertising) and MC (Modified Content) ads are the most important ad formats in current advertising management systems, accounting for a significant share of overall ad spending. Therefore, improving the effectiveness of DC and MC ads can significantly enhance the ad experience for advertisers and the exposure experience for ad audiences. The information recommendation method provided in this application can be applied to DC and MC ads, enabling personalized and automatic determination of the best creative for DC and MC ads, and automatically implementing traffic allocation.

[0130] In some embodiments, see Figure 10 , Figure 10 This is a schematic diagram of the architecture of the advertising management system provided in the embodiments of this application, such as... Figure 10 As shown, the advertising management system provided in this application upgrades the ad creative ranking strategy from the original MAB-based creative priority strategy to an online tLTR (tidwise Learning To Rank) model scoring + creative priority. The tLTR model is located after the coarse ranking stage and before the fine ranking stage. This model, for each ad containing multiple creatives, such as an MC ad, individually selects the optimal creative from different requests and sends it to the subsequent fine ranking stage. For example, assuming that in a user request, the ad queue obtained in the coarse ranking stage contains 200 ads, of which 50 are multi-creative ads, then the tLTR model can select the corresponding best creative for each of these 50 multi-creative ads. For example, the tLTR model can employ a coarse-ranking algorithm based on the fine-ranking results, fitting the ranking results of multiple creatives in the fine-ranking stage. This allows for the scoring, prediction, and ranking of multiple creatives within an ad. The reasons for learning the ranking results of multiple creatives in the fine-ranking stage are: 1. The fine-ranking stage is the most accurate recommendation stage in the ad management system; 2. The fine-ranking stage is the final stage that determines whether an ad will be exposed. Therefore, the coarse-ranking stage only needs to fit the ranking results of the fine-ranking stage to increase the overall consistency and overall performance of the ranking. The training samples, model structure, and loss function of the model will be explained below.

[0131] In some embodiments, the tLTR model can be trained in a pairwise manner to learn the eCPM order relationship between pairs of creatives. Therefore, the training samples for the model can be obtained by sampling creatives from the finely ranked queue on a request-by-request basis. Meanwhile, considering the large number of creatives in the creative set and the limitations of the model's training speed, a certain amount of sampling can be performed. The sampling ratio can be determined according to the model's training speed. For example, the top 1 creative can be sampled from the ranking results, 5 creatives can be randomly sampled from the top 2 to the top 20, and 9 creatives can be sampled from the top 20 onwards, resulting in a total of 15 creatives. These 15 sampled creatives can then be combined in pairs to obtain 105 partial order pairs (i.e., training sample pairs, corresponding to the display strategy sample pairs mentioned above). For example, taking partial order pair 1 as an example, assuming that partial order pair 1 includes creative 1 and creative 2, if the eCPM1 of creative 1 is greater than the eCPM2 of creative 2, then the label of partial order pair 1 is set to 1; otherwise, the label of partial order pair 1 is set to 0.

[0132] In other embodiments, the structure of the tLTR model can be further divided according to the characteristics of different traffic scenarios (e.g., data volume, time consumption, etc.). Figure 11A The twin-tower structure shown and Figure 11B The diagram illustrates a non-dual-tower structure (i.e., a single-tower structure). A dual-tower structure divides the user's feature vector and the idea's feature vector into two separate network towers for computation. Finally, the encoded results of the user's feature vector and the idea's feature vector are multiplied by a dot product to obtain the final predicted score for the idea (corresponding to the ranking parameters mentioned above). A non-dual-tower structure, i.e., a single-tower network, concatenates the user's feature vector and the idea's feature vector, and inputs the concatenated vector into a single tower for computation to obtain the final predicted score for the idea. The overall structure of the tLTR model can be divided into: a feature input layer, a hidden layer, and an output layer. The input layer, hidden layer, and loss function of the tLTR model are explained below.

[0133] 1) Feature Input Layer In some embodiments, the feature information input to the tLTR model includes user feature information (e.g., user interest tags, behavioral sequences, browsing history, etc.) and creative feature information (e.g., creative basic attributes, creative tags, statistical features, etc.). The user and creative feature information contains a large number of categorical and character-based features, therefore, the feature information needs to be vectorized first to facilitate subsequent model computation. For example, a query word embedding table (where the embedding table can be a randomly initialized matrix, the shape of which depends on the number of features and the dimensions of the embedded features) can be used to construct the word embedding vector for each feature information.

[0134] For example, see Figure 12 , Figure 12 This is a schematic diagram illustrating the principle of the word embedding process provided in the embodiments of this application. Figure 12 An example of an embedding table is given; for example, the embedding table could be a 5-column table. Given a matrix W of size 3, assuming that for a certain feature value, its corresponding one-hot encoding vector is x = (0, 0, 0, 1, 0), then performing word embedding processing on this one-hot encoding vector yields a word embedding vector X = xW = (10, 12, 19).

[0135] 2) Hidden layer After vectorizing all feature information through the input layer of the tLTR model to obtain feature vector X, feature vector X can be fed into the hidden layer of the tLTR model so that the hidden layer encodes feature vector X to obtain the deep representation H of the features. （1） (i.e., hidden layer feature vectors), as shown in the following formula: Formula 1 Where X represents the feature vector obtained after passing through the input layer of the tLTR model, and W （1） Let b represent the weight matrix of the first hidden layer. （1） This represents the bias matrix of the first hidden layer. This represents the non-linear activation function corresponding to the first hidden layer, such as ReLU, sigmoid, tanh, etc. This represents the encoded result of the first hidden layer output.

[0136] In other embodiments, the tLTR model can learn deeper representations of features by stacking more hidden layers, as shown in the following equation: Formula 2 in, This represents the weight matrix of the second hidden layer. This represents the bias matrix of the second hidden layer. This represents the non-linear activation function corresponding to the second hidden layer. This represents the encoded result of the second hidden layer output.

[0137] In some embodiments, the encoded result of the last hidden layer output, i.e., the vector, can be... The elements in ( , , ..., The summation is performed to obtain a scalar z. Then, the sigmoid function is used to activate the scalar z to obtain the final prediction score s of the creative idea i.i The specific formula is as follows: Formula 3 =sigmoid(z) Formula 4 3) Loss Function In some embodiments, adaptive training of the tLTR model can be achieved by designing a loss function and combining it with backpropagation. The loss function can be a pairwise loss function, which, simply put, is the difference (s) between the predicted scores of training sample pairs (creation i, creation j). i -s j The labels of the sample pairs are substituted into the loss function (e.g., the cross-entropy loss function) for calculation. The specific formula of the loss function is as follows:

[0138] =

[0139] = Formula 5 Among them, when When =1, we have: Formula 6 when When =-1, we have: Formula 7 Among them, s i s j Let i and j represent the prediction scores obtained by applying the sigmoid algorithm to each of the two ideas. Represents the sigmoid function, and when s i >s j hour, =1, when s i <s j hour, =0.

[0140] In some embodiments, considering the potential surge in the number of ideas in the future, which could lead to a decline in model performance, binarization embedding can be used to improve the model's inference speed. Binarization may result in a small loss in model performance but can significantly reduce machine costs. Furthermore, in addition to pairwise training, listwise training can also be considered. Figure 11A The dual-tower structure in the model does not fully consider the interaction between features. The model structure can be further explored. In addition, an attention mechanism can be introduced to learn the importance of features.

[0141] The information recommendation method provided in this application, based on the coarse ranking results of advertisements, selects the best-performing creatives from multiple creative advertisements for exposure and display. By adopting the LTR+Group concept, a tLTR model is proposed. This model creates a personalized, scalable, and automated advertising creative selection strategy by fitting the order of eCPM or pCTCVR of each creative in the fine ranking stage. It enables scoring, prediction, and ranking of multiple creatives within an advertisement, optimizes the recommendation effect of advertisements, and thus improves the utilization rate of advertising resources.

[0142] The following description continues to illustrate the exemplary structure of the information recommendation device 243 provided in the embodiments of this application as a software module. In some embodiments, such as Figure 4A As shown, the software modules stored in the information recommendation device 243 of the memory 240 may include: an acquisition module 2431, a prediction module 2432, a determination module 2433, and a recommendation module 2434.

[0143] The acquisition module 2431 is used to acquire at least one piece of multi-mode information to be sent to the object, wherein each piece of multi-mode information includes multiple display strategies; the acquisition module 2431 is also used to acquire display strategy feature information corresponding to the multiple display strategies and object feature information of the object; the prediction module 2432 is used to call a neural network model to perform prediction processing based on each display strategy feature information and object feature information to obtain the ranking parameters corresponding to each display strategy, wherein the neural network model is trained based on the ranking results of multiple display strategy samples that have been applied; the determination module 2433 is used to determine the target display strategy corresponding to each piece of multi-mode information, wherein the target display strategy is the display strategy with the largest ranking parameter among the multi-mode information; the recommendation module 2434 is used to perform a recommendation operation for the object based on at least one piece of multi-mode information that has a corresponding target display strategy applied.

[0144] In some embodiments, the prediction module 2432 is further configured to, for each display strategy feature information, call the neural network model to perform the following processing: vectorize the display strategy feature information and the object feature information respectively to obtain the display strategy feature vector of the display strategy and the object feature vector of the object; encode the display strategy feature vector and the object feature vector to obtain the hidden layer feature vector; and activate the hidden layer feature vector to obtain the sorting parameters of the display strategy.

[0145] In some embodiments, the encoding process is implemented by calling at least one hidden layer of the neural network model. When there are multiple hidden layers, the multiple hidden layers are cascaded in sequence to form a tower network. The prediction module 2432 is also used to concatenate the display strategy feature vector and the object feature vector to obtain a concatenated vector. It is also used to call the tower network to encode the concatenated vector to obtain the hidden layer feature vector.

[0146] In some embodiments, the acquisition module 2431 is further configured to acquire the data volume of at least one multi-mode information; the information recommendation device 243 further includes a transfer module 2435, configured to, in response to the data volume being less than the data volume threshold, transfer to the step of performing encoding processing based on the display strategy feature vector and the object feature vector by calling the tower network.

[0147] In some embodiments, the prediction module 2432 is further configured to encode the concatenated vector through the first hidden layer of the tower network; input the encoding result output by the first hidden layer into the subsequent cascaded hidden layers, and continue the encoding process through the subsequent cascaded hidden layers until the last hidden layer; and determine the encoding result output by the last hidden layer as the hidden layer feature vector.

[0148] In some embodiments, the encoding process is implemented by calling at least one hidden layer of the neural network model. When there are multiple hidden layers, the multiple hidden layers are divided into two groups. The multiple hidden layers in the first group are cascaded to form a first pyramid network, and the multiple hidden layers in the second group are cascaded to form a second pyramid network. The prediction module 2432 is also used to perform the following processing for each display strategy feature vector: calling the first pyramid network to encode the object feature vector to obtain a first sub-hidden layer feature vector; calling the second pyramid network to encode the display strategy feature vector to obtain a second sub-hidden layer feature vector; and performing a dot product of the first sub-hidden layer feature vector and the second sub-hidden layer feature vector to obtain a hidden layer feature vector.

[0149] In some embodiments, the acquisition module 2431 is further configured to acquire the data volume of at least one piece of multi-mode information; the transfer module 2435 is further configured to, in response to the data volume being greater than or equal to a data volume threshold, transfer to the step of executing the encoding process based on the object feature vector and the display strategy feature vector, respectively calling the first tower network and the second tower network for encoding processing.

[0150] In some embodiments, the prediction module 2432 is further configured to perform one-hot encoding on the display strategy feature information to obtain a first one-hot encoded vector; and to perform word embedding on the first one-hot encoded vector to obtain a display strategy feature vector of the display strategy; the prediction module 2432 is further configured to perform one-hot encoding on the object feature information to obtain a second one-hot encoded vector; and to perform word embedding on the second one-hot encoded vector to obtain an object feature vector of the object.

[0151] In some embodiments, the input layer of the neural network model includes multiple sub-input layers corresponding to multiple domains respectively; the determination module 2433 is further configured to determine the domain to which the object belongs; the prediction module 2432 is further configured to call the sub-input layer corresponding to the domain to perform one-hot encoding processing on the object feature information to obtain a second one-hot encoded vector.

[0152] In some embodiments, the acquisition module 2431 is further configured to acquire the display strategy ranking results corresponding to multiple multi-mode information samples, wherein the display strategy ranking results are obtained by ranking the recommendation parameters of the display strategy samples in the multi-mode information samples; the information recommendation device 243 further includes a training module 2436, configured to sample multiple display strategy samples based on the display strategy ranking results; perform pairwise combination processing on the multiple display strategy samples to obtain multiple display strategy sample pairs; compare the recommendation parameters of the two display strategy samples included in each display strategy sample pair; generate label data for each display strategy sample pair based on the comparison results; and train a neural network model based on the multiple display strategy sample pairs and the corresponding label data.

[0153] In some embodiments, the information recommendation device 243 further includes a generation module 2437, which performs the following processing for each display strategy sample pair: when the recommendation parameter of the first display strategy sample in the display strategy sample pair is greater than the recommendation parameter of the second display strategy sample, the label data of the display strategy sample pair is determined to be 1; when the recommendation parameter of the first display strategy sample in the display strategy sample pair is less than the recommendation parameter of the second display strategy sample, the label data of the display strategy sample pair is determined to be 0.

[0154] In some embodiments, the determining module 2433 is further configured to determine the sampling ratio based on the training speed of the neural network model, wherein the ratio is negatively correlated with the training speed; the training module 2436 is further configured to sample the display strategy ranking result corresponding to each multi-mode information sample according to the ratio to obtain multiple display strategy samples.

[0155] In some embodiments, the information recommendation device 243 further includes a setting module 2438, which is used to set a mirror network of a pyramid network in the neural network model before training the neural network model based on multiple display strategy sample pairs and corresponding label data. The pyramid network is used to receive display strategy feature information of one display strategy sample in the display strategy sample pair, and the mirror network is used to receive display strategy feature information of the other display strategy sample in the display strategy sample pair. The setting module 2438 is also used to remove the mirror network from the neural network model after training the neural network model based on multiple display strategy sample pairs and corresponding label data.

[0156] In some embodiments, the setting module 2438 is further configured to set a mirror network of a second tower network in the neural network model before training the neural network model based on multiple display strategy sample pairs and corresponding label data, wherein the second tower network is used to receive display strategy feature information of one display strategy sample in the display strategy sample pair, and the mirror network is used to receive display strategy feature information of the other display strategy sample in the display strategy sample pair; and to remove the mirror network from the neural network model after training the neural network model based on multiple display strategy sample pairs and corresponding label data.

[0157] In some embodiments, the acquisition module 2431 is further configured to acquire multi-mode information that satisfies at least one of the following conditions: matching the object feature information of the object; the similarity between the multi-mode information and the object feature vector of the object is less than a similarity threshold; sorting multiple multi-mode information based on the predicted click-through rate; and filtering out at least one multi-mode information that ranks first from the sorting results.

[0158] In some embodiments, when the number of at least one multi-mode information is multiple, the information recommendation device 243 further includes a sorting module 2439, which is used to sort the multiple multi-mode information with corresponding target display strategies according to any of the following sorting strategies: sorting the multiple multi-mode information with corresponding target display strategies according to information type; randomly sorting the multiple multi-mode information with corresponding target display strategies; and sorting the multiple multi-mode information with corresponding target display strategies according to operating rules.

[0159] It should be noted that the description of the apparatus in this application embodiment is similar to the description of the method embodiment above, and has similar beneficial effects as the method embodiment; therefore, it will not be repeated. For technical details not covered in the information recommendation apparatus provided in this application embodiment, please refer to... Figure 6 ,or Figure 7 ,or Figure 9 The meaning is understood in accordance with the description of any of the accompanying drawings.

[0160] The following description continues to illustrate the exemplary structure of the training device 244 for the neural network model for information recommendation provided in the embodiments of this application as a software module. In some embodiments, such as Figure 4B As shown, the software modules stored in the training device 244 of the neural network model for information recommendation stored in the memory 240 may include: an acquisition module 2441, a sampling module 2442, an execution module 2443, and a generation module 2444.

[0161] The acquisition module 2441 is used to acquire the display strategy ranking results corresponding to multiple multi-mode information samples, wherein the display strategy ranking results are obtained by ranking the recommendation parameters of the display strategy samples in the multi-mode information samples; the sampling module 2442 is used to sample multiple display strategy samples based on the display strategy ranking results; the execution module 2443 is used to call the initialized neural network model to perform training tasks based on the multiple display strategy samples to update the parameters of the neural network model; and the generation module 2444 is used to generate the trained neural network model based on the updated parameters.

[0162] In some embodiments, the execution module 2443 is further configured to perform pairwise combination processing on multiple display strategy samples to obtain multiple display strategy sample pairs; compare the recommendation parameters of the two display strategy samples included in each display strategy sample pair, and generate label data for each display strategy sample pair based on the comparison results; and, based on the multiple display strategy sample pairs and the corresponding label data, call an initialized neural network model to perform a first training task; wherein the first training task includes: performing prediction processing on the initialized neural network model based on the display strategy sample pairs, and updating the parameters of the neural network model based on the difference between the comparison results obtained from the prediction and the label data.

[0163] It should be noted that the description of the apparatus in this application embodiment is similar to the description of the method embodiment above, and has similar beneficial effects as the method embodiment, therefore it will not be repeated. For any technical details not covered in the training apparatus for the neural network model for information recommendation provided in this application embodiment, please refer to... Figure 5 The explanation is for understanding.

[0164] This application provides a computer program product comprising a computer program or computer-executable instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer-executable instructions from the computer-readable storage medium and executes the computer-executable instructions, causing the computer device to perform the information recommendation method described in this application embodiment, or the training method for a neural network model used for information recommendation.

[0165] This application provides a computer-readable storage medium storing computer-executable instructions. When these computer-executable instructions are executed by a processor, they cause the processor to execute the information recommendation method provided in this application, for example... Figure 6 , Figure 7 ,or Figure 9 The information recommendation method shown, or the training method of the neural network model for information recommendation provided in the embodiments of this application, for example, such as... Figure 5 The training method for a neural network model used for information recommendation is shown.

[0166] In some embodiments, the computer-readable storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface memory, optical disk, or CD-ROM; or it may be a variety of devices including one or any combination of the above-mentioned memories.

[0167] In some embodiments, executable instructions may take 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 standalone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.

[0168] As an example, executable instructions can be deployed to execute on a single electronic device, or on multiple electronic devices located in one location, or on multiple electronic devices distributed across multiple locations and interconnected via a communication network.

[0169] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, and improvements made within the spirit and scope of this application are included within the scope of protection of this application.

Claims

1. An information recommendation method, characterized in that, The method includes: Obtain at least one multi-mode message to be sent to the object, wherein each multi-mode message includes multiple display strategies; Obtain the display strategy feature information corresponding to the multiple display strategies and the object feature information of the object; For each of the aforementioned display strategy feature information, the neural network model is invoked to perform the following processing: The display strategy feature information and the object feature information are vectorized respectively to obtain the display strategy feature vector of the display strategy and the object feature vector of the object. Obtain the data volume of the at least one multi-mode information; in response to the data volume being less than the data volume threshold, call the tower network for encoding processing based on the display strategy feature vector and the object feature vector to obtain the hidden layer feature vector, wherein the tower network is formed by cascading multiple hidden layers of the neural network model in sequence; In response to the data volume being greater than or equal to the data volume threshold, based on the object feature vector and the display strategy feature vector, the first tower network and the second tower network are respectively invoked for encoding processing to obtain the hidden layer feature vector; wherein, the multiple hidden layers are divided into two groups, the multiple hidden layers of the first group are sequentially cascaded to form the first tower network, and the multiple hidden layers of the second group are sequentially cascaded to form the second tower network. The hidden layer feature vectors are activated to obtain the ranking parameters of the display strategy; wherein, the neural network model is trained using the following method: A mirror network of the tower network is configured, wherein the structure and parameters of the mirror network of the tower network are consistent with those of the tower network. The tower network is used to receive the display strategy feature information of one display strategy sample in the display strategy sample pair, and the mirror network of the tower network is used to receive the display strategy feature information of the other display strategy sample in the display strategy sample pair. A mirror network of the second tower network is configured, wherein the structure and parameters of the mirror network of the second tower network are consistent with those of the second tower network. The second tower network is used to receive the display strategy feature information of one display strategy sample in the display strategy sample pair, and the mirror network of the second tower network is used to receive the display strategy feature information of the other display strategy sample in the display strategy sample pair. The neural network model is trained based on multiple display strategy sample pairs and corresponding label data. The label data is obtained by comparing the recommendation parameters of the two display strategy samples included in the display strategy sample pair, and the mirror network of the tower network is removed from the neural network model. Determine the target display strategy corresponding to each piece of multi-mode information, wherein the target display strategy is the display strategy with the largest sorting parameter among the multi-mode information; Based on at least one piece of multimodal information that has the corresponding target display strategy applied, a recommendation operation is performed for the object.

2. The method according to claim 1, characterized in that, The encoding process is implemented by calling at least one hidden layer of the neural network model; In response to the data volume being less than a data volume threshold, a tower network is invoked for encoding processing based on the display strategy feature vector and the object feature vector to obtain a hidden layer feature vector, including: The display strategy feature vector and the object feature vector are concatenated to obtain a concatenated vector. The concatenated vector is encoded using the tower network to obtain the hidden layer feature vector.

3. The method according to claim 2, characterized in that, The process of encoding the concatenated vector using the tower network to obtain the hidden layer feature vector includes: The spliced ​​vector is encoded through the first hidden layer of the tower network; The encoded result output from the first hidden layer is input into the subsequent cascaded hidden layers, and the encoding process continues through the subsequent cascaded hidden layers until the last hidden layer. The encoding result output by the last hidden layer is determined as the hidden layer feature vector.

4. The method according to claim 1, characterized in that, The encoding process is implemented by calling at least one hidden layer of the neural network model; the encoding process, based on the object feature vector and the display strategy feature vector, involves calling the first tower network and the second tower network respectively to obtain the hidden layer feature vector, including: For each of the aforementioned display strategy feature vectors, the following processing is performed: The first tower network is invoked to encode the feature vector of the object to obtain the first sub-hidden layer feature vector; The second tower network is invoked to encode the feature vector of the display strategy to obtain the feature vector of the second sub-hidden layer. The first sub-hidden layer feature vector and the second sub-hidden layer feature vector are multiplied by a dot product to obtain the hidden layer feature vector.

5. The method according to claim 4, characterized in that, The step of vectorizing the display strategy feature information and the object feature information respectively to obtain the display strategy feature vector of the display strategy and the object feature vector of the object includes: The display strategy feature information is processed by one-hot encoding to obtain a first one-hot encoded vector; the first one-hot encoded vector is processed by word embedding to obtain the display strategy feature vector of the display strategy. The object feature information is subjected to one-hot encoding to obtain a second one-hot encoded vector; the second one-hot encoded vector is subjected to word embedding to obtain the object feature vector of the object.

6. The method according to claim 5, characterized in that, The input layer of the neural network model includes multiple sub-input layers corresponding to multiple domains; The step of performing one-hot encoding on the object feature information to obtain a second one-hot encoded vector includes: Determine the domain to which the object belongs; The sub-input layer corresponding to the domain is invoked to perform one-hot encoding on the object feature information to obtain a second one-hot encoded vector.

7. The method according to any one of claims 1-6, characterized in that, Before invoking the neural network model for prediction processing, the method further includes: Obtain the display strategy ranking results corresponding to multiple multi-mode information samples, wherein the display strategy ranking results are obtained by sorting based on the recommendation parameters of the display strategy samples in the multi-mode information samples; Multiple display strategy samples are obtained based on the ranking results of the display strategy; The multiple display strategy samples are combined in pairs to obtain multiple display strategy sample pairs; Compare the recommendation parameters of the two display strategy samples included in each display strategy sample pair, and generate label data for each display strategy sample pair based on the comparison results; The neural network model is trained based on the multiple display strategy sample pairs and the corresponding label data.

8. The method according to claim 7, characterized in that, The step of generating label data for each of the display strategy sample pairs based on the comparison results includes: For each of the aforementioned display strategy sample pairs, the following processing is performed: When the recommendation parameter of the first display strategy sample in the display strategy sample pair is greater than the recommendation parameter of the second display strategy sample, the label data of the display strategy sample pair is determined to be 1; When the recommendation parameter of the first display strategy sample in the display strategy sample pair is less than the recommendation parameter of the second display strategy sample, the label data of the display strategy sample pair is set to 0.

9. The method according to claim 7, characterized in that, The sampling based on the ranking result of the display strategy yields multiple display strategy samples, including: The sampling ratio is determined based on the training speed of the neural network model, wherein the ratio is negatively correlated with the training speed; The display strategy ranking results corresponding to each of the multi-mode information samples are sampled according to the stated ratio to obtain multiple display strategy samples.

10. The method according to claim 7, characterized in that, When the hidden layers of the neural network model are divided into two groups, with the hidden layers of the first group cascaded sequentially to form a first tower network and the hidden layers of the second group cascaded sequentially to form a second tower network, after training the neural network model based on the multiple display strategy sample pairs and the corresponding label data, the method further includes: Remove the mirror network of the second tower network from the neural network model.

11. The method according to any one of claims 1-6, characterized in that, The step of obtaining at least one piece of multi-mode information to be sent to the object includes: Obtain multi-modal information that satisfies at least one of the following conditions: matches the object feature information of the object; or has a similarity to the object feature vector of the object that is less than a similarity threshold. Based on the predicted click-through rate, multiple pieces of multi-modal information are sorted, and at least one piece of multi-modal information with the highest ranking is selected from the sorting results.

12. The method according to any one of claims 1-6, characterized in that, When the number of the at least one piece of multimodal information is multiple, before performing a recommendation operation for the object based on the at least one piece of multimodal information for which the corresponding target display strategy is applied, the method further includes: Sort the multiple pieces of multi-mode information, each with its corresponding target display strategy, according to any of the following sorting strategies: Sort the multiple pieces of multi-mode information that have the corresponding target display strategy applied to them according to the information type; Randomly sort the multiple pieces of multi-mode information that each applies the corresponding target display strategy; According to the operating rules, the multiple pieces of multi-mode information that have applied the corresponding target display strategies are sorted.

13. A training method for a neural network model for information recommendation, characterized in that, The method includes: Obtain the display strategy ranking results corresponding to multiple multi-mode information samples, wherein the display strategy ranking results are obtained by sorting based on the recommendation parameters of the display strategy samples in the multi-mode information samples; Multiple display strategy samples were obtained based on the ranking results of the aforementioned display strategies. A mirror network of a tower network is configured, wherein the structure and parameters of the mirror network are consistent with those of the tower network. The tower network is used to receive the display strategy feature information of one display strategy sample in the display strategy sample pair, and the mirror network of the tower network is used to receive the display strategy feature information of the other display strategy sample in the display strategy sample pair. A mirror network of the second tower network is configured, wherein the structure and parameters of the mirror network of the second tower network are consistent with those of the second tower network. The second tower network is used to receive the display strategy feature information of one display strategy sample in the display strategy sample pair, and the mirror network of the second tower network is used to receive the display strategy feature information of the other display strategy sample in the display strategy sample pair. Based on multiple display strategy sample pairs and corresponding label data, the neural network model is trained to update the parameters of the neural network model, and the mirror network of the tower network is removed from the neural network model. The label data is obtained by comparing the recommendation parameters of the two display strategy samples included in the display strategy sample pair. The trained neural network model is generated based on the updated parameters.

14. The method according to claim 13, characterized in that, The display strategy sample pairs are obtained by combining multiple display strategy samples pairwise; the step of training the neural network model based on multiple display strategy sample pairs and corresponding label data to update the parameters of the neural network model includes: Based on multiple display strategy sample pairs and the corresponding label data, the initialized neural network model is invoked to perform a first training task; wherein, the first training task includes: performing prediction processing on the initialized neural network model based on the display strategy samples, and updating the parameters of the neural network model based on the difference between the prediction comparison result and the label data.

15. An information recommendation device, characterized in that, The device includes: The acquisition module is used to acquire at least one piece of multi-mode information to be sent to the object, wherein each piece of multi-mode information includes multiple display strategies; The acquisition module is further configured to acquire display strategy feature information corresponding to the plurality of display strategies and object feature information of the object; The prediction module is used to perform the following processing by calling a neural network model for each of the display strategy feature information: vectorizing the display strategy feature information and the object feature information respectively to obtain the display strategy feature vector of the display strategy and the object feature vector of the object; acquiring the data volume of the at least one piece of multi-mode information; in response to the data volume being less than a data volume threshold, calling a pyramid network for encoding processing based on the display strategy feature vector and the object feature vector to obtain a hidden layer feature vector, wherein the pyramid network is formed by cascading multiple hidden layers of the neural network model; in response to the data volume being greater than or equal to the data volume threshold, calling a first pyramid network and a second pyramid network for encoding processing based on the object feature vector and the display strategy feature vector respectively to obtain the hidden layer feature vector; wherein the multiple hidden layers are divided into two groups, the multiple hidden layers of the first group are cascaded to form the first pyramid network, and the multiple hidden layers of the second group are cascaded to form the second pyramid network; and activating the hidden layer feature vector to obtain the sorting parameters of the display strategy. A configuration module is used to configure a mirror network of the tower network, wherein the structure and parameters of the mirror network are consistent with those of the tower network. The tower network is used to receive display strategy feature information of one display strategy sample in a display strategy sample pair, and the mirror network is used to receive display strategy feature information of the other display strategy sample in the same pair. A mirror network of the second tower network is also configured, wherein the structure and parameters of the mirror network are consistent with those of the second tower network. The second tower network is used to receive display strategy feature information of one display strategy sample in a display strategy sample pair, and the mirror network is used to receive display strategy feature information of the other display strategy sample in the same pair. Based on multiple display strategy sample pairs and corresponding label data, a neural network model is trained. The label data is obtained by comparing the recommendation parameters of the two display strategy samples included in each display strategy sample pair, and the mirror network of the tower network is removed from the neural network model. The determining module is used to determine the target display strategy corresponding to each piece of multi-mode information, wherein the target display strategy is the display strategy with the largest sorting parameter among the multi-mode information; The recommendation module is used to perform a recommendation operation for the object based on at least one piece of multi-modal information that has the corresponding target display strategy applied to it.

16. The apparatus according to claim 15, characterized in that, The encoding process is implemented by calling at least one hidden layer of the neural network model; The prediction module is further configured to concatenate the display strategy feature vector and the object feature vector to obtain a concatenated vector; and call the tower network to encode the concatenated vector to obtain the hidden layer feature vector.

17. The apparatus according to claim 16, characterized in that, The prediction module is also used to encode the spliced ​​vector through the first hidden layer of the tower network; The encoded result output from the first hidden layer is input into the subsequent cascaded hidden layers, and the encoding process continues through the subsequent cascaded hidden layers until the last hidden layer. The encoding result output by the last hidden layer is determined as the hidden layer feature vector.

18. A training apparatus for a neural network model for information recommendation, characterized in that, The device includes: The acquisition module is used to acquire the display strategy ranking results corresponding to multiple multi-mode information samples, wherein the display strategy ranking results are obtained by sorting based on the recommendation parameters of the display strategy samples in the multi-mode information samples; The sampling module is used to obtain multiple display strategy samples based on the sorting results of the display strategy; A mirror network of a tower network is configured, wherein the structure and parameters of the mirror network are consistent with those of the tower network. The tower network is used to receive the display strategy feature information of one display strategy sample in the display strategy sample pair, and the mirror network of the tower network is used to receive the display strategy feature information of the other display strategy sample in the display strategy sample pair. A mirror network of the second tower network is configured, wherein the structure and parameters of the mirror network of the second tower network are consistent with those of the second tower network. The second tower network is used to receive the display strategy feature information of one display strategy sample in the display strategy sample pair, and the mirror network of the second tower network is used to receive the display strategy feature information of the other display strategy sample in the display strategy sample pair. An execution module is used to train the neural network model based on multiple display strategy sample pairs and corresponding label data, so as to update the parameters of the neural network model and remove the mirror network of the tower network from the neural network model, wherein the label data is obtained by comparing the recommendation parameters of the two display strategy samples included in the display strategy sample pair; A generation module is used to generate the trained neural network model based on the updated parameters.

19. An electronic device, characterized in that, The electronic device includes: Memory, used to store executable instructions; The processor, when executing executable instructions stored in the memory, implements the information recommendation method according to any one of claims 1 to 12, or the training method for a neural network model for information recommendation according to any one of claims 13 to 14.

20. A computer-readable storage medium storing computer-executable instructions, characterized in that, When the computer-executable instructions are executed by the processor, they implement the information recommendation method according to any one of claims 1 to 12, or the training method for the neural network model for information recommendation according to any one of claims 13 to 14.

21. A computer program product, comprising a computer program or computer-executable instructions, characterized in that, When the computer program or computer-executable instructions are executed by a processor, they implement the information recommendation method according to any one of claims 1 to 12, or the training method for a neural network model for information recommendation according to any one of claims 13 to 14.

Images