Content recommendation model processing method and apparatus, electronic device, and medium

Abstract

The application discloses a content recommendation model processing method and device, electronic equipment and medium. The method comprises the following steps: inputting the account attribute information of a sample account and the content description information of sample media content corresponding to the sample account into a to-be-trained double-tower model, performing recommendation prediction on the sample account for the sample media content on the basis of feature extraction of the to-be-trained double-tower model from the content income index dimension and the quality correction factor dimension, obtaining content income index data and content recommendation index data; the content income index data is an index factor of the content recommendation index data; based on the content income index label, the content recommendation index label, the content income index data and the content recommendation index data corresponding to the sample media content, the income index loss information and the recommendation index loss information are determined to perform content recommendation training on the to-be-trained double-tower model, and a content recommendation model is obtained, so that the content income and the user browsing experience of the model recommended content are effectively balanced.

Description

Technical Field

[0001] This application relates to the field of multimedia content delivery technology, and in particular to a content recommendation model processing method, apparatus, electronic device and medium. Background Technology

[0002] With the development of the internet industry, multimedia content placement and promotion on various internet platforms (such as mobile internet) has become a major form of information dissemination. For example, internet advertising targeting applications or physical products is a typical example of information placement, serving as an important means of exposure and promotion to target audiences (such as user accounts). The selection of which media resources to target, and which audience to allocate, is typically achieved through a preliminary selection phase and a final selection phase.

[0003] However, in the coarse ranking stage of existing technologies, the optimization goal of content recommendation models is usually to maximize the revenue of content recommendation platforms and content providers, ignoring the user's browsing experience of recommended content. This results in low accuracy of recommended content recall and ranking in the coarse ranking stage, affecting the consistency of ranking between the coarse and fine ranking stages, and thus affecting the overall recommendation effect of the content recommendation system. Summary of the Invention

[0004] This application provides a content recommendation model processing method, apparatus, device, storage medium, and computer program product. By explicitly separating the feature extraction dimension of the dual-tower model into a content benefit index dimension and a quality correction factor dimension, the content recommendation index data output by the dual-tower model simultaneously measures content benefit and content quality. During model training, multi-objective collaborative optimization is performed on the content benefit index and content recommendation index. While improving the stability of the content recommendation training of the dual-tower model, it can effectively balance the potential content benefit of the recommended content and the user browsing experience, improve the ranking consistency between the coarse ranking stage and the fine ranking stage, and thus improve the overall recommendation effect of the content recommendation system.

[0005] According to one aspect of the embodiments of this application, a content recommendation model processing method is provided, the method comprising: Obtain the account attribute information of the sample account, the content description information of the sample media content corresponding to the sample account, the content revenue indicator tags of the sample media content, and the content recommendation indicator tags of the sample media content; The account attribute information and the content description information are input into the dual-tower model to be trained. Based on the feature extraction of the dual-tower model from the dimensions of content revenue indicators and quality correction factors, the model performs recommendation prediction on the sample accounts for the sample media content, thereby obtaining the content revenue indicator data and the content recommendation indicator data corresponding to the sample media content; the content revenue indicator data is the indicator factor of the content recommendation indicator data. Based on the content revenue indicator tags, the content recommendation indicator tags, the content revenue indicator data, and the content recommendation indicator data, determine revenue indicator loss information and recommendation indicator loss information; Based on the revenue loss information and the recommendation loss information, the dual-tower model to be trained is trained for content recommendation to obtain a content recommendation model.

[0006] According to one aspect of the embodiments of this application, a content recommendation model processing apparatus is provided, the apparatus comprising: The training data acquisition module is used to acquire account attribute information of sample accounts, content description information of sample media content corresponding to the sample accounts, content revenue indicator tags and content recommendation indicator tags corresponding to the sample media content; The recommendation prediction module is used to input the account attribute information and the content description information into the dual-tower model to be trained. Based on the feature extraction of the dual-tower model from the dimensions of content revenue indicators and quality correction factors, the module performs recommendation prediction on the sample account for the sample media content, and obtains the content revenue indicator data and the content recommendation indicator data corresponding to the sample media content; the content revenue indicator data is the indicator factor of the content recommendation indicator data. The loss determination module is used to determine revenue indicator loss information and recommendation indicator loss information based on the content revenue indicator label, the content recommendation indicator label, the content revenue indicator data and the content recommendation indicator data. The model training module is used to train the dual-tower model to be trained on content recommendation based on the revenue indicator loss information and the recommendation indicator loss information, so as to obtain a content recommendation model.

[0007] According to one aspect of the embodiments of this application, an electronic device is provided, including: a processor; Memory used to store the processor's executable instructions; The processor is configured to execute the instructions to implement the content recommendation model processing method described above.

[0008] According to one aspect of the embodiments of this application, a computer-readable storage medium is provided, which, when the instructions in the storage medium are executed by a processor of an electronic device, enables the electronic device to perform any of the above-described content recommendation model processing methods.

[0009] According to one aspect of the embodiments of this application, a computer program product is provided, the computer program product including a computer program stored in a computer-readable storage medium, wherein a processor reads from the computer-readable storage medium and executes the computer program to implement the content recommendation model processing method provided in the various optional implementations described above.

[0010] The recommendation model processing method, apparatus, equipment, storage medium, and computer program product provided in this application have the following technical effects: In the coarse ranking stage of content recommendation, this application obtains the account attribute information of sample accounts, the content description information of the sample media content corresponding to the sample accounts, the content revenue indicator tags of the sample media content, and the content recommendation indicator tags of the sample media content. The account attribute information and content description information are input into the dual-tower model to be trained. Based on the feature extraction of the dual-tower model from the content revenue indicator dimension and the quality correction factor dimension, recommendation prediction is performed on the sample accounts for the sample media content, resulting in content revenue indicator data and content recommendation indicator data corresponding to the sample media content. The content revenue indicator data is the indicator factor of the content recommendation indicator data, achieved by explicitly separating the feature extraction dimensions of the dual-tower model into the content revenue indicator dimension and the quality correction factor dimension. This approach allows the dual-tower model to simultaneously measure content revenue and content quality in its output content recommendation metrics. Based on content revenue metric labels, content recommendation metric labels, content revenue metric data, and content recommendation metric data, revenue metric loss information and recommendation metric loss information are determined. Then, based on these loss information, the dual-tower model is trained for content recommendation, resulting in a content recommendation model. During model training, multi-objective collaborative optimization of content revenue and recommendation metrics can be performed. This improves the stability of the dual-tower model's content recommendation training, effectively balancing the potential content revenue of recommended content with the user browsing experience, enhancing the consistency of ranking between the coarse and fine ranking stages, and ultimately improving the overall recommendation performance of the content recommendation system. Attached Figure Description

[0011] To more clearly illustrate the technical solutions and advantages in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a schematic diagram of the application environment of a content recommendation model processing method provided in an embodiment of this application; Figure 2 This is a flowchart illustrating a content recommendation model processing method provided in an embodiment of this application; Figure 3 This is a flowchart illustrating another content recommendation model processing method provided in an embodiment of this application; Figure 4 This is a flowchart illustrating another content recommendation model processing method provided in an embodiment of this application; Figure 5a This is a flowchart illustrating another content recommendation model processing method provided in an embodiment of this application; Figure 5b This is a schematic diagram of the structure of a content recommendation model provided in an embodiment of this application; Figure 6 This is a schematic diagram of the structure of a content recommendation model processing device provided in an embodiment of this application; Figure 7 This is a block diagram of an electronic device for content recommendation model processing provided in an embodiment of this application; Figure 8 This is a block diagram of another electronic device for content recommendation model processing provided in the embodiments of this application. Detailed Implementation

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0014] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or server that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0015] In the embodiments of this application, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0016] 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.

[0017] Retrieval: Quickly retrieve a small subset of candidate ads relevant to the current request (user, context, query terms, etc.) from the full ad library (which may contain millions or even hundreds of millions of ads).

[0018] Pre-ranking: The thousands of ads returned during the retrieval phase are initially scored and screened, and the few hundred most promising ads are retained for fine-ranking.

[0019] Fine-grained ranking: This involves giving high-precision, personalized scores and final rankings to the small number of ads selected from the initial screening, determining their display order.

[0020] bid: The highest price an advertiser sets for a unit of advertising performance event. It is the highest price or expected price that an advertiser is willing to pay to achieve a specific advertising goal (such as one click, one conversion, or thousands of impressions).

[0021] eCPM (Effective Cost Per Mille): Expected revenue per thousand impressions, used to measure the expected revenue an ad generates for the platform (or advertiser) per thousand impressions. The formula is described below. .

[0022] Ad Quality is a comprehensive metric that measures the value of an ad to user experience, the platform ecosystem, and advertiser performance. It is typically not a single metric but a composite signal comprised of multiple sub-dimensions, introduced as a regularization term, multiplier factor, or independent scoring item in ad ranking (especially eCPM calculation). Some common component dimensions include relevance, user experience (UX), engagement quality, historical performance, and policy compliance, among others.

[0023] Final Value: The final score of an ad, typically calculated by weighting the ad's eCPM and Ad Quality, is a comprehensive metric that measures the ad's value based on both its commercial and user experience aspects. .

[0024] The Dual Tower Model separates the processing of user features and ad features into two independent neural networks ("towers") to achieve efficient large-scale ad retrieval or coarse ranking.

[0025] LTR (Learning to Rank): A type of machine learning method used to rank candidate results, widely applied in search engines, recommendation systems, question answering systems, and other scenarios. Given a query and a set of documents (or items), the model learns how to rank these documents from highest to lowest relevance.

[0026] Multi-Task Learning (MTL): Under a unified model framework, multiple objective tasks (such as CTR, CVR, dwell time, etc.) are optimized simultaneously. These tasks are usually related to each other, and the learning effect of each task is improved through parameter sharing or knowledge transfer.

[0027] NDCG: A listwise ranking evaluation metric widely used in information retrieval and recommendation systems, used to measure how close the ranking result is to the ideal ranking.

[0028] Pairwise Loss: This transforms the sorting problem into a loss function that determines whether the relative order of any two samples is correct.

[0029] Please see Figure 1 , Figure 1This is a schematic diagram of an application environment for a content recommendation model processing method provided in an embodiment of this application. The application environment may include at least a terminal device 101 and a server 102.

[0030] Terminal device 101 and server 102 are connected via a wireless or wired network. Terminal device 101 is used to obtain user-triggered model training requests for the content recommendation model and send these requests to server 102. Server 102 is used to obtain account attribute information of sample accounts, content description information of sample media content corresponding to the sample accounts, content revenue indicator labels of sample media content, and content recommendation indicator labels of sample media content; input the account attribute information and content description information into the dual-tower model to be trained; based on feature extraction from the content revenue indicator dimension and quality correction factor dimension of the dual-tower model to be trained, perform recommendation prediction on the sample accounts for the sample media content, and obtain content revenue indicator data and content recommendation indicator data corresponding to the sample media content; the content revenue indicator data is the indicator factor of the content recommendation indicator data; based on the content revenue indicator labels, content recommendation indicator labels, content revenue indicator data, and content recommendation indicator data, determine revenue indicator loss information and recommendation indicator loss information; based on the revenue indicator loss information and recommendation indicator loss information, train the dual-tower model to be trained for content recommendation to obtain the content recommendation model. Optionally, the terminal device 101 may be a smartphone, tablet computer, laptop computer, desktop computer, smart speaker, smartwatch, smart voice interaction device, smart home appliance, or in-vehicle terminal, but is not limited to these.

[0031] Optionally, the server 102 may be an independent physical server, or a server cluster or distributed system consisting of multiple physical servers. Optionally, it may be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms.

[0032] Those skilled in the art should understand that the terminal device 101 and server 102 described above are merely illustrative examples. Other existing or future terminal devices or servers that are applicable to this application should also be included within the scope of protection of this application, and are hereby incorporated by reference.

[0033] During their research, the inventors discovered the following technical problems in the content recommendation model processing technology: Question 1: In the coarse-grained stage of existing technologies, the optimization goal of content recommendation models is usually to maximize the revenue of content recommendation platforms and content providers, while ignoring the user's browsing experience of the recommended content.

[0034] Question 2: Although a multi-objective collaborative optimization scheme can be achieved in the fine-ranking stage, it requires maintaining multiple independent task models. The multi-objective ranking is optimized by iterating Pareto strategies in the loss space. Furthermore, due to the semantic gap between the loss function and the business objectives, relying solely on the Pareto balance of the loss space can easily lead to the trap of "optimal local indicators but suboptimal global business indicators".

[0035] Based on this, the content recommendation model processing method provided in this application embodiment can at least solve the following problems: 1. By explicitly separating the feature extraction dimensions of the dual-tower model into the content revenue indicator dimension and the quality correction factor dimension, the content recommendation indicator data output by the dual-tower model can simultaneously measure content revenue and content quality.

[0036] 2. Based on factorization modeling of the feature space of the dual-tower model, multi-objective collaborative optimization is performed on content revenue indicators and content recommendation indicators during model training. This not only improves the stability of content recommendation training of the dual-tower model, but also effectively balances the potential content revenue of recommended content with user browsing experience, improves the consistency of ranking in the coarse ranking stage and the fine ranking stage, and thus improves the overall recommendation effect of the content recommendation system.

[0037] The following describes the application scenarios of the content recommendation model processing method provided in the embodiments of this application. The embodiments of this application provide a content recommendation model processing method, apparatus, electronic device, computer-readable storage medium, and computer program product, which can be applied to various scenarios. Examples are given below.

[0038] In some embodiments, the above-described content recommendation model processing method can be applied to advertising recommendation scenarios. This involves obtaining account attribute information of sample accounts, advertising description information of sample ads corresponding to the sample accounts, advertising revenue indicator tags corresponding to the sample ads, and advertising recommendation indicator tags corresponding to the sample ads. The account attribute information and advertising description information are input into the dual-tower model to be trained. Based on feature extraction from the advertising revenue indicator dimension and quality correction factor dimension by the dual-tower model to be trained, recommendation prediction is performed on the sample accounts for the sample ads, obtaining advertising revenue indicator data and advertising recommendation indicator data corresponding to the sample ads. The advertising revenue indicator data serves as the indicator factor for the advertising recommendation indicator data. Based on the advertising revenue indicator tags, advertising recommendation indicator tags, advertising revenue indicator data, and advertising recommendation indicator data, revenue indicator loss information and recommendation indicator loss information are determined. Based on the revenue indicator loss information and recommendation indicator loss information, the dual-tower model to be trained is used for advertising recommendation training to obtain an advertising recommendation model. Based on the trained advertising recommendation model, targeted ads with advertising potential are selected, effectively balancing the potential content ads and user browsing experience of the targeted ads.

[0039] It should be noted that, in addition to the above-mentioned application scenarios, the content recommendation model processing method provided in this application can also be widely applied to other scenarios to meet the diverse needs of different user groups, and this application does not impose any restrictions.

[0040] Figure 2 This is a flowchart illustrating a content recommendation model processing method provided in an embodiment of this application. The method is executed by a computer device, which can be a terminal device, a server, or an interaction between the terminal device and the server. Figure 2 As shown, the method may include: S201, obtain the account attribute information of the sample account, the content description information of the sample media content corresponding to the sample account, the content revenue indicator tags of the sample media content, and the content recommendation indicator tags of the sample media content.

[0041] Specifically, the sample account can be the object account corresponding to the sample object. For example, the sample object can be an online user on the content recommendation platform, and the object account can be the user account used by the online user on the content recommendation platform. In a specific embodiment, the account attribute information of the sample account can characterize the basic attribute features, account usage features, and content preference features of the sample account. Optionally, the account attribute information may include, but is not limited to: gender, age, hobbies, registration time, etc.

[0042] Specifically, the sample media content corresponding to the sample account can be historical media content recalled for the sample account. The content modality of the sample media content can include at least one of the following: image, text, audio, or video. Illustratively, the content function of the sample media content can include, but is not limited to, news information, knowledge dissemination, and advertising. In a specific embodiment, the content description information of the sample media content can be used to describe the information conveyed by the sample media content. Optionally, the presentation format of the content description information can be determined in conjunction with the content modality of the sample media content. Illustratively, when the content modality is text, the content description information can be a text summary; when the content modality is video, the content description information can be video keyframes, video description text, etc.

[0043] Specifically, the content revenue indicator (CPI) tag can be a preset tag for the sample media content regarding the content revenue indicator. The content revenue indicator can be used to measure the expected revenue of the sample media content in the content recommendation platform. The content revenue indicator can be a revenue indicator pre-set by the content recommendation platform in combination with the revenue measurement needs in actual applications. For example, the content revenue indicator can adopt eCPM.

[0044] In a specific embodiment, the content revenue metric can be obtained by fusing multiple revenue metric factors. Optionally, these multiple revenue metric factors can be multiple linear factors of the content revenue metric or multiple multiplicative factors of the content revenue metric. For example, taking eCPM as the content revenue metric, the multiple revenue metric factors can be as follows: , (Estimated click-through rate) (Estimated conversion rate), accordingly, , where u represents media content and a represents account.

[0045] Specifically, content recommendation metric tags can be preset tags for sample media content targeting specific content recommendation metrics. Content recommendation metrics can be used to measure the degree of recommendation (matching degree) of sample media content to sample accounts. Illustratively, content recommendation metrics can be represented as content recommendation scores. Specifically, content recommendation metrics can be obtained by fusing content revenue metrics and quality correction factors. The quality correction factor can be a learnable, general factor used to compensate for higher-order metric factors related to content quality metrics (such as contextual matching, novelty, fairness, delayed conversion compensation, etc.). Here, content quality metrics are used to quantify the browsing experience of sample accounts with sample media content (e.g., content satisfaction).

[0046] Optionally, the content recommendation metric can be obtained by linearly fusing the content revenue metric and the quality correction factor, or by multiplicatively fusing the content revenue metric and the quality correction factor. For example, using eCPM as the content revenue metric, the multiplicative fusing method can be used to obtain the content recommendation metric: .

[0047] In one specific embodiment, the sample media content corresponding to the above-mentioned sample accounts was collected in the following manner: S1, obtain the historical ranking candidate content set corresponding to the sample account.

[0048] Specifically, the historical fine-ranking candidate content set can be the set of candidate content that has passed the coarse-ranking recall stage and entered the fine-ranking scoring stage for the historical content recommendation requests of the sample account.

[0049] S2, perform hierarchical negative sampling on the historical fine-ranking candidate content set to obtain the sampled content set.

[0050] Specifically, to avoid the large amount of "obviously irrelevant" noise introduced by random sampling of the entire media content library, the historical fine-ranking candidate content set that has passed coarse-ranking recall and preliminary filtering can be sampled. Since the indicator feature distribution of the historical fine-ranking candidate content set is closer to the real online delivery scenario, it can significantly improve the training efficiency and generalization ability of the content recommendation model in the coarse-ranking stage (hereinafter referred to as the coarse-ranking model). Furthermore, to enhance the coarse-ranking model's ability to jointly model the sensitivity of the ranking head and the diversity of the long tail, a segmented and hierarchical negative sampling strategy can be adopted for the historical fine-ranking candidate content set. For example, one candidate content is sampled from the top 5 candidates in the fine-ranking (as a high-confidence positive example, a strong eCPM signal); four candidate content is sampled from the candidates ranked 6-50 in the fine-ranking (as medium-to-high-value candidates, balancing content benefits and user experience); ten candidate content is sampled from the candidates ranked 51-100 in the fine-ranking (medium-value, improving the robustness of head coverage); and fifteen candidate content is sampled from the remaining candidate content ranked 101 and below in the fine-ranking (simulating simple negative examples, enhancing global discrimination ability). S3 uses the sampled content set as the sample media content.

[0051] As can be seen from the above embodiments, the segmented and hierarchical negative sampling strategy enhances the content recommendation model's ability to jointly model the sensitivity of the ranking head and the diversity of the long tail in the coarse ranking stage, thereby improving training efficiency and generalization ability.

[0052] In one specific embodiment, the content revenue metric tag corresponding to the sample media content is obtained in the following way: The sample accounts and sample media content are input into the pre-trained fine ranking model to label the content revenue indicators, thus obtaining the content revenue indicator labels corresponding to the sample media content.

[0053] Specifically, the content revenue indicator labels corresponding to the sample media content can be the prediction data obtained by the pre-trained fine-ranking model after predicting the content revenue of sample accounts and sample media content.

[0054] For example, taking content revenue metric eCPM as an example, the eCPM tag can be... The calculation yielded that, This refers to the click-through rate (CTR) tag. The conversion rate label is illustrative. It can be a constant set by the content provider or content recommendation platform. It can come from the output of the fine-ranking model or from the log statistics after calibration.

[0055] As can be seen from the above embodiments, by pre-training the fine ranking model to label the content revenue indicators for sample accounts and sample media content, the corresponding content revenue indicator tags for the sample media content are obtained. Based on the corresponding content revenue indicator tags for the sample media content, the content recommendation model in the coarse ranking stage is trained, which can improve the consistency of content recommendation ranking between the coarse ranking stage and the fine ranking stage.

[0056] In a specific embodiment, the content recommendation metric tags corresponding to the sample media content are obtained in the following way: Collect multiple interactive feedback metrics from sample accounts for each sampled content; Multiple interactive feedback metrics are fused to obtain content recommendation metric tags for each sampled content.

[0057] Specifically, various interactive feedback metrics can be pre-set based on the content recommendation scoring requirements of the content recommendation platform in actual applications. For example, various interactive feedback metrics may include, but are not limited to, click count, conversion count, and content dwell time.

[0058] In a specific embodiment, the above-mentioned index fusion processing of multiple interactive feedback indicators to obtain the content recommendation index label for each sampled content may include: performing weighted fusion processing on multiple interactive feedback indicators for each sampled content to obtain the content recommendation index label for each sampled content.

[0059] Alternatively, a pre-trained ranking model can be used to score the matching degree between sample accounts and sample media content outputs as content recommendation indicator labels. Alternatively, content recommendation indicator labels can be generated through the reward function of a reinforcement learning model that has been pre-trained for content recommendation.

[0060] As can be seen from the above embodiments, by collecting multiple interactive feedback indicators from sample accounts for each sampled content, and performing indicator fusion processing on these multiple interactive feedback indicators to obtain content recommendation indicator tags for each sampled content, the accuracy of content recommendation indicator labeling can be effectively improved.

[0061] S202, input the account attribute information and content description information into the dual-tower model to be trained. Based on the feature extraction of the dual-tower model from the dimensions of content revenue indicators and quality correction factors, make recommendation predictions for the sample accounts for the sample media content, and obtain the content revenue indicator data and the content recommendation indicator data corresponding to the sample media content; the content revenue indicator data is the indicator factor of the content recommendation indicator data.

[0062] In one specific embodiment, the dual-tower model to be trained can be applied to the coarse ranking stage of the content recommendation process to perform preliminary recall of media content.

[0063] In an optional embodiment, the content revenue indicator data can be the predicted indicator data obtained by the dual-tower model to be trained based on account attribute information and content description information to predict the content revenue indicator, and the content recommendation indicator data can be the predicted indicator data obtained by the dual-tower model to be trained based on account attribute information and content description information to predict the content recommendation indicator.

[0064] In one specific embodiment, the dual-tower model to be trained includes: an account tower and a content tower, such as... Figure 3 As shown above, the account attribute information and content description information are input into the dual-tower model to be trained. Based on the feature extraction of the dual-tower model from the dimensions of content revenue indicators and quality correction factors, the model performs recommendation prediction on the sample accounts for the sample media content. The resulting content revenue indicator data and content recommendation indicator data corresponding to the sample media content may include: S301. Input the account attribute information into the account tower, extract account features from the content revenue indicator dimension and the quality correction factor dimension, and obtain the account representation vector.

[0065] S302, input the content description information into the content pyramid, extract content features from the content description information from the dimensions of content revenue indicators and quality correction factors, and obtain the content representation vector.

[0066] Specifically, Account Tower is used to extract account features from the content revenue index dimension and the quality correction factor dimension of account attribute information to obtain the account representation vector; correspondingly, the account representation vector can represent the account features of the sample account from the content revenue index dimension and the quality correction factor dimension.

[0067] Specifically, the content pyramid is used to extract content features from the content revenue index dimension and the quality correction factor dimension of content description information to obtain a content representation vector; correspondingly, the content representation vector can represent the content features of the sample media content from the content revenue index dimension and the quality correction factor dimension.

[0068] Specifically, the content revenue metric dimension can be the feature fitting dimension corresponding to the content revenue metric, and the quality correction factor dimension can be the feature fitting dimension corresponding to the quality correction factor.

[0069] It is understandable that, since the dual-tower model needs to calculate the inner product of the account representation vector and the content representation vector when performing content recommendation (content recall), the inner product can be used as content recommendation indicator data. The content recommendation indicator can be obtained by fusing the content revenue indicator and the quality correction factor. Therefore, the account representation vector and the content representation vector can be divided into an orthogonal subspace of the content revenue indicator dimension (i.e., the vector elements corresponding to the content revenue indicator dimension) and an orthogonal subspace of the quality correction factor dimension (i.e., the vector elements corresponding to the quality correction factor dimension) according to semantic function.

[0070] Optionally, if the content recommendation metric is obtained by linearly fusing the content revenue metric and the quality correction factor, both the account representation vector and the content representation vector can include vector elements corresponding to the content revenue metric dimension and vector elements corresponding to the quality correction factor dimension. Optionally, if the content recommendation metric is obtained by multiplicatively fusing the content revenue metric and the quality correction factor, the logarithm of the content recommendation metric can be obtained by linearly fusing the logarithm of the content revenue metric and the logarithm of the quality correction factor. Accordingly, both the account representation vector and the content representation vector can include vector elements corresponding to the logarithm dimension of the content revenue metric and vector elements corresponding to the logarithm dimension of the quality correction factor.

[0071] For example, the content revenue metrics corresponding to the content revenue metrics dimension are... The quality correction factor corresponding to the quality correction factor dimension is: For example, content recommendation metrics Content revenue metrics and quality correction factor The corresponding content recommendation metric is obtained by multiplicative fusion. Logarithm of content revenue metrics Logarithm of quality correction factor After linear fusion, the account representation vector can be expressed as: The content representation vector can be represented as ,in, , , .

[0072] S303 performs vector fusion processing on the account representation vector and the content representation vector to determine the content revenue indicator data corresponding to the content revenue indicator dimension and the quality correction factor data corresponding to the quality correction factor dimension.

[0073] In a specific embodiment, the vector fusion processing here can take the form of vector dot product, calculating the product of vector elements in the corresponding vector dimensions. Optionally, if the content recommendation metric is obtained by linearly fusing the content revenue metric and the quality correction factor, the content revenue metric data can be the predicted data of the content revenue metric, and the quality correction factor data can be the predicted data of the quality correction factor; if the content recommendation metric is obtained by multiplicatively fusing the content revenue metric and the quality correction factor, the content revenue metric data can be the predicted data of the logarithm of the content revenue metric, and the quality correction factor data can be the predicted data of the logarithm of the quality correction factor.

[0074] S304, the content revenue indicator data and quality correction factor data are fused to obtain content recommendation indicator data.

[0075] Specifically, content recommendation metrics data can be a fusion of content revenue metrics data and quality correction factor data.

[0076] In one specific embodiment, the content recommendation index data can be obtained by adding the content revenue index data and the quality correction factor data. Optionally, if the content recommendation index is obtained by linearly fusing the content revenue index and the quality correction factor, the content revenue index data can be the predicted data of the content revenue index, the quality correction factor data can be the predicted data of the quality correction factor, and correspondingly, the content recommendation index data can be the predicted data of the content recommendation index; if the content recommendation index is obtained by multiplicatively fusing the content revenue index and the quality correction factor, the content revenue index data can be the predicted data of the logarithm of the content revenue index, the quality correction factor data can be the predicted data of the logarithm of the quality correction factor, and correspondingly, the content recommendation index data can be the predicted data of the logarithm of the content recommendation index.

[0077] As can be seen from the above embodiments, by explicitly separating the feature extraction dimension of the dual-tower model into the content revenue indicator dimension and the quality correction factor dimension, the content recommendation indicator data output by the dual-tower model can simultaneously measure content revenue and content quality.

[0078] S203, based on content revenue indicator tags, content recommendation indicator tags, content revenue indicator data, and content recommendation indicator data, determine revenue indicator loss information and recommendation indicator loss information.

[0079] In a specific embodiment, the sample media content includes: a sampled content set. The determination of revenue indicator loss information and recommendation indicator loss information based on content revenue indicator tags, content recommendation indicator tags, content revenue indicator data, and content recommendation indicator data may include: S2031, Based on the first preset ranking loss function, perform content revenue difference analysis on the content revenue index labels of each sampled content set and the content revenue index data of each sampled content set to obtain revenue index loss information; S2032, based on the second preset ranking loss function, perform content recommendation difference analysis on the content recommendation index labels and content recommendation index data of each sampled content set to obtain recommendation index loss information.

[0080] Specifically, the revenue indicator loss information is used to characterize the ranking difference between the content revenue indicator labels and the content revenue indicator data of each sampled content set; the recommendation indicator loss information is used to characterize the ranking difference between the content recommendation indicator labels and the content recommendation indicator data of each sampled content set.

[0081] In an optional embodiment, the first preset sorting loss function and the second preset sorting loss function can be the same or different preset sorting loss functions. Optionally, the preset sorting loss function can be a pairwise sorting loss function or a list sorting loss function.

[0082] For example, taking the pairwise ranking loss function as the preset ranking loss function, a partial order pair of revenue metrics can be constructed for every two sampled contents corresponding to the sample account, resulting in a set of partial order pairs of revenue metrics. Taking eCPM as the content revenue metric as an example, if... Then construct positive sample pairs Based on the pairwise ranking loss function, the loss information of the revenue indicator is determined. The optimization objective of the model is to minimize the loss information of the profit metric, where... = , Indicates the temperature coefficient. This represents the set of partially ordered pairs of revenue metrics; and, for each pair of sampled content corresponding to a sample account, a partially ordered pair of recommendation metrics can be constructed, resulting in the set of partially ordered pairs of recommendation metrics. Taking the recommendation revenue metric FinalValue (FV) as an example, if... Then construct positive sample pairs Based on the pairwise ranking loss function, the loss information of the recommendation metric is determined. The optimization objective of the model is to minimize the loss information of the recommendation metric, where... = , Indicates the temperature coefficient. This represents the set of partially ordered pairs of recommendation metrics.

[0083] For example, taking the list ranking loss function as the preset ranking loss function, the content revenue metric eCPM and the recommendation revenue metric FV are regarded as two "pseudo-relevance labels". Differentiable approximations (such as Soft-NDCG, ApproxNDCG) are used to directly optimize the ranking metrics, resulting in the revenue metric loss information shown below. and recommendation metric loss information : ,in, = ; Indicates the sampled content Corresponding content revenue metrics tags; ,in, = , Indicates the sampled content The corresponding content recommendation metrics tags.

[0084] As can be seen from the above embodiments, based on the preset ranking loss function, the loss information of the revenue indicator and the loss information of the recommendation indicator are determined, thereby improving the accuracy of the loss information representation.

[0085] S204. Based on the loss information of the revenue indicator and the loss information of the recommendation indicator, the dual-tower model to be trained is trained on content recommendation to obtain the content recommendation model.

[0086] In a specific embodiment, such as Figure 4 As shown, based on the loss information of the revenue metric and the loss information of the recommendation metric, the dual-tower model to be trained is trained on content recommendation, resulting in a content recommendation model including: S401, determine the first loss weight corresponding to the loss information of the profit indicator and the second loss weight corresponding to the loss information of the recommendation indicator.

[0087] In an optional embodiment, the determination of the first loss weight corresponding to the profit indicator loss information and the second loss weight corresponding to the recommendation indicator loss information includes: Obtain incremental online metrics data from the content recommendation platform where the content recommendation model resides; Based on incremental online metrics data, the loss weight adjustment model is used to adjust the loss weights and determine the first loss weight corresponding to the loss information of the revenue metrics and the second loss weight corresponding to the loss information of the recommendation metrics.

[0088] Specifically, the incremental data of online metrics is used to measure the growth of online metrics of the content recommendation platform within a preset statistical period. In a specific embodiment, online metrics may include, but are not limited to, RPM (Revenue Per Mille). The preset statistical period can be set according to the weight update cycle requirements in actual applications. For example, the preset statistical period can be weekly or monthly.

[0089] In a specific embodiment, the loss weight adjustment model is a reinforcement learning model. The above-mentioned adjustment of loss weights based on incremental online indicator data, input into the loss weight adjustment model, and determination of the first loss weight corresponding to the revenue indicator loss information and the second loss weight corresponding to the recommendation indicator loss information include: The incremental online indicator data is input into the reinforcement learning model. Based on the adjustment reward representation information corresponding to the reinforcement learning model, the target weight adjustment data with the largest long-term reward data is determined from a variety of preset weight adjustment data under the incremental online indicator data. The adjustment reward representation information is used to represent the long-term reward data obtained by making decisions on different preset weight adjustment data under different incremental online indicator data. The long-term reward data is used to represent the feedback effect of weight adjustment accumulated since the weight adjustment training of the reinforcement learning model. Based on the target weight adjustment data, the current loss weights of the dual-tower model to be trained are adjusted to obtain the first loss weight and the second loss weight.

[0090] Specifically, the weight adjustment strategy is modeled as an MDP (Markov Decision Process) model to accumulate long-term rewards (e.g., discounted RPM + ...). • Retention rate, where discounted RPM Instant earnings+ The MDP model is trained with the expected future returns as the optimization objective.

[0091] In an optional embodiment, the determination of the first loss weight corresponding to the profit indicator loss information and the second loss weight corresponding to the recommendation indicator loss information may include: Determine the group characteristics of the target account group and the scenario characteristics of the target recommendation scenario; The loss weights are adjusted based on group characteristic information and scene characteristic information to determine the first loss weight and the second loss weight.

[0092] Specifically, the group characteristic information of the target account group can be used to characterize the target account group's usage characteristics of the content browsing platform and the target account group's own attribute characteristics. The group characteristic information may include: platform usage characteristic information and group attribute characteristic information. Optionally, the platform usage characteristic information may include, but is not limited to: usage frequency information, registration time information, etc., and the group attribute characteristic information may include, but is not limited to: age, gender, education level, etc.

[0093] Specifically, the scene feature information of the target recommendation scenario can be used to characterize the content recommendation demand features of the target recommendation scenario. The scene feature information may include, but is not limited to, recommendation time, recommendation channel, etc.

[0094] For example, for high-value user groups (such as high activity and high retention rates), the weight of the second loss corresponding to the loss information of recommendation metrics can be increased. This will enable the orthogonal subspace of the quality correction factor dimension to fit the quality correction factor signal, which is oriented towards long-term experience, more closely during model training. For example, the quality correction factor signal may include, but is not limited to: deep interaction duration (>30 seconds), negative feedback rate (such as "not interested" clicks), conversion quality measurement metrics (such as high average order value, repeat purchase behavior), etc. For price-sensitive scenarios or new users (such as cold start traffic), the weight of the first loss corresponding to the loss information of revenue metrics can be increased. This will enable the model to focus more on the fitting accuracy of the orthogonal subspace of the content revenue metric dimension during model training, prioritizing short-term revenue efficiency.

[0095] In an optional embodiment, the determination of the first loss weight corresponding to the profit indicator loss information and the second loss weight corresponding to the recommendation indicator loss information includes: Multiple candidate loss weight reorganizations are obtained. Each candidate loss weight reorganization includes: candidate loss weights corresponding to the loss information of the return indicator and candidate loss weights corresponding to the loss information of the recommendation indicator. A grouped comparative test was conducted on multiple candidate loss weight reorganizations to obtain the test results; Based on the test results, target loss right restructuring is selected from multiple candidate loss right restructurings; Based on the restructuring of the target loss weight, the first loss weight and the second loss weight are determined.

[0096] Specifically, an account interaction prediction model can be trained based on historical logs. This allows the trained model to learn the interaction probabilities of an account with different sample media content (corresponding to combinations of different content revenue metrics and different content recommendation metrics). These interaction probabilities can include, but are not limited to, click probability, conversion probability, and negative feedback probability. Then, through a metric mapping function, the interaction probabilities output by the account interaction prediction model are transformed into predictable online business metrics (e.g., RPM, click-through rate, dwell time). Based on grouped comparative testing, multiple candidate loss weight reorganizations are executed in parallel in a simulation environment to evaluate their long-term returns and stability, obtaining test results. Based on these test results, the best-performing target loss weight reorganization is selected.

[0097] As can be seen from the above embodiments, the dynamic loss weight scheduling mechanism driven by online business feedback can effectively improve the pertinence and adaptability of loss weight updates.

[0098] S402, based on the first loss weight and the second loss weight, performs loss fusion on the revenue indicator loss information and the recommendation indicator loss information to obtain content recommendation loss information.

[0099] In a specific embodiment, the revenue indicator loss information and the recommendation indicator loss information can be weighted and fused based on the first loss weight and the second loss weight to obtain the content recommendation loss information.

[0100] For illustrative purposes, the content recommendation loss information can be expressed as the following formula:

[0101] in, It can be fixed or determined through an adaptive scheduling mechanism. This is for embedding regular expressions.

[0102] S403, based on the content recommendation loss information, updates the model parameters of the dual-tower model to be trained, and obtains the content recommendation model.

[0103] In one specific embodiment, the content recommendation model can be applied to the coarse ranking stage of the content recommendation system. The content recommendation model can be obtained by updating the model parameters of the dual-tower model to be trained based on the content recommendation loss information.

[0104] In a specific embodiment, updating the model parameters of the dual-tower model to be trained based on the content recommendation loss information to obtain the content recommendation model may include: S4031 (not shown in the figure) updates the model parameters of the dual-tower model to be trained based on content recommendation loss information.

[0105] Specifically, updating the model parameters of the dual-tower model to be trained can include updating the model parameters of the account tower and updating the model parameters of the content tower.

[0106] S4032 (not shown in the figure) is based on the updated dual-tower model. The content recommendation training iteration operation, which includes steps S202, S203, S401, S402 and S4031, is repeatedly executed until the content recommendation convergence condition is met.

[0107] S4033 (not shown in the figure) is the dual-tower model obtained when the content recommendation convergence condition is met, and is used as the content recommendation model.

[0108] In an optional embodiment, the aforementioned convergence condition for content recommendation can be that the number of training iterations reaches a preset number of training iterations. Optionally, the convergence condition for content recommendation can also be that the content recommendation loss information is less than a specified threshold. In the embodiments of this specification, the preset number of training iterations and the specified threshold can be preset in conjunction with the training speed and accuracy of the network in practical applications.

[0109] As can be seen from the above embodiments, in the coarse ranking stage of content recommendation, the account attribute information of the sample account, the content description information of the sample media content corresponding to the sample account, the content revenue indicator tags of the sample media content, and the content recommendation indicator tags of the sample media content are obtained. The account attribute information and content description information are input into the dual-tower model to be trained. Based on the feature extraction of the dual-tower model from the dimensions of content revenue indicator and quality correction factor, the recommendation prediction is performed on the sample account for the sample media content, resulting in the content revenue indicator data and the content recommendation indicator data corresponding to the sample media content. The content revenue indicator data is the indicator factor of the content recommendation indicator data. This is achieved by explicitly separating the feature extraction dimensions of the dual-tower model into the content revenue indicator dimension and the quality correction factor. The factor dimension allows the content recommendation metrics output by the dual-tower model to simultaneously measure content revenue and content quality. Based on content revenue metric labels, content recommendation metric labels, content revenue metric data, and content recommendation metric data, revenue metric loss information and recommendation metric loss information are determined. Then, based on these loss information, the dual-tower model to be trained is used for content recommendation training, resulting in a content recommendation model. During model training, multi-objective collaborative optimization can be performed on content revenue and content recommendation metrics. This improves the stability of the dual-tower model's content recommendation training, effectively balancing the potential content revenue of recommended content with the user browsing experience, enhancing the consistency of ranking between the coarse and fine ranking stages, and thus improving the overall recommendation performance of the content recommendation system.

[0110] In a specific embodiment, the aforementioned content revenue metric dimension includes: at least one revenue metric factor dimension. Specifically, the at least one revenue metric factor dimension can be a feature fitting dimension corresponding to each of the at least one revenue metric factor. This at least one revenue metric factor can be a driving variable constituting the calculation logic of the content revenue metric, that is, the content revenue metric can be obtained by fusing the at least one revenue metric factor. For example, taking eCPM as the content revenue metric, the at least one revenue metric factor may include: , (Estimated click-through rate) (Estimated conversion rate), accordingly, , where u represents media content and a represents account.

[0111] In a specific embodiment, such as Figure 5a As shown, the above method may further include: S501, Obtain the revenue indicator factor labels for each of the sample media content for at least one revenue indicator factor dimension.

[0112] Specifically, the revenue indicator factor labels can be obtained by a pre-trained fine-ranking model predicting the revenue indicator factors for the sample media content.

[0113] The above vector fusion processing of account representation vector and content representation vector determines the content revenue indicator data corresponding to the content revenue indicator dimension and the quality correction factor data corresponding to the quality correction factor dimension, including: S502, perform vector fusion processing on the account representation vector and the content representation vector to determine the revenue indicator factor data for each of the at least one revenue indicator factor dimension and the quality correction factor data corresponding to the quality correction factor dimension.

[0114] Specifically, the account representation vector and content representation vector can be semantically divided into orthogonal subspaces of the content revenue indicator dimension (i.e., vector elements corresponding to the content revenue indicator dimension) and orthogonal subspaces of the quality correction factor dimension (i.e., vector elements corresponding to the quality correction factor dimension). The content revenue indicator dimension can include at least one revenue indicator factor dimension. Correspondingly, the account representation vector and content representation vector can be semantically divided into orthogonal subspaces of at least one revenue indicator factor dimension (i.e., vector elements corresponding to each of the at least one revenue indicator factor dimension) and orthogonal subspaces of the quality correction factor dimension (i.e., vector elements corresponding to the quality correction factor dimension).

[0115] Optionally, if the content revenue metric is obtained by linearly fusing at least one revenue metric factor, both the account representation vector and the content representation vector may include: vector elements corresponding to each dimension of at least one revenue metric factor and vector elements corresponding to the dimension of the quality correction factor; alternatively, if the content revenue metric is obtained by multiplicatively fusing at least one revenue metric factor, the logarithm of the content revenue metric can be obtained by linearly fusing the logarithms of at least one revenue metric factor, and correspondingly, both the account representation vector and the content representation vector may include: vector elements corresponding to each dimension of the logarithm of at least one revenue metric factor and vector elements corresponding to the dimension of the logarithm of the quality correction factor.

[0116] In a specific embodiment, the revenue indicator factor data corresponding to each revenue indicator factor dimension can be the prediction indicator data corresponding to each revenue indicator factor dimension. The revenue indicator factor data corresponding to each revenue indicator factor dimension can be obtained by fusing the vector elements of the corresponding revenue indicator factor dimension in the account representation vector and the vector elements of the corresponding revenue indicator factor dimension in the content representation vector. For illustration, the fusing process here can be done by multiplying the vector elements.

[0117] Optionally, if the content revenue indicator is obtained by linearly fusing at least one revenue indicator factor, the revenue indicator factor data corresponding to each revenue indicator factor dimension can be the prediction indicator data corresponding to each revenue indicator factor; if the content revenue indicator is obtained by multiplicatively fusing at least one revenue indicator factor, then the logarithm of the content revenue indicator is obtained by linearly fusing the logarithms of at least one revenue indicator factor, and the revenue indicator factor data corresponding to each revenue indicator factor dimension can be the prediction indicator data corresponding to the logarithm of each revenue indicator factor.

[0118] For example, the content revenue metrics corresponding to the content revenue metrics dimension are... The quality correction factor corresponding to the quality correction factor dimension is: For example, Each of the at least one return indicator factor dimension corresponds to a return indicator factor. , (Estimated click-through rate) The estimated conversion rate is obtained through multiplicative fusion, and correspondingly, the logarithm of the content revenue metric is... Depend on , and After linear fusion, the account representation vector can be expressed as: The content representation vector can be represented as ,in, , , .

[0119] S503, perform index fusion on the revenue index factor data of each of at least one revenue index factor dimension to obtain content revenue index data.

[0120] Specifically, the content revenue indicator data can be obtained by adding the revenue indicator factor data of each of the at least one revenue indicator factor dimensions.

[0121] Optionally, if the content revenue indicator is obtained by linearly fusing at least one revenue indicator factor, the revenue indicator factor data for each dimension of the at least one revenue indicator factor can be the predictive indicator data for the at least one revenue indicator factor, and correspondingly, the content revenue indicator data can be the predictive indicator data for the content revenue indicator; if the content revenue indicator is obtained by multiplicatively fusing at least one revenue indicator factor, the revenue indicator factor data for each dimension of the at least one revenue indicator factor can be the predictive indicator data for the logarithm of the at least one revenue indicator factor, and correspondingly, the content revenue indicator data can be the predictive indicator data for the logarithm of the content revenue indicator.

[0122] Based on the loss information of the revenue metric and the loss information of the recommendation metric, the dual-tower model to be trained is trained for content recommendation, resulting in a content recommendation model including: S504, based on the profit indicator factor labels and profit indicator factor data of each of the at least one profit indicator factor dimension, determine the loss information of each of the at least one profit indicator factor dimension.

[0123] In a specific embodiment, determining the loss information of each of the at least one return indicator factor dimensions based on the return indicator factor labels and return indicator factor data of each of the at least one return indicator factor dimensions may include: Based on the preset regression loss function, a difference analysis is performed on the return indicator factor labels corresponding to each return indicator factor dimension and the return indicator factor data corresponding to each return indicator factor dimension to determine the indicator factor loss information corresponding to each return indicator factor dimension.

[0124] Specifically, the indicator factor loss information corresponding to each revenue indicator factor dimension is used to characterize the indicator difference between the revenue indicator factor label and the revenue indicator factor data of the corresponding revenue indicator factor dimension.

[0125] Specifically, since the fitting target here is the return index factor (e.g., Since X is V or T, the traditional cross-entropy loss cannot be applied. Furthermore, considering that the training samples in the coarse-ranking stage come from massive online logs and have significant label noise and long-tail distribution, this application abandons the traditional mean squared error loss and instead adopts Huber Loss or smooth and stable Log-Cosh Loss, which are robust to outliers. This effectively alleviates gradient oscillations caused by extreme clicks / conversion labels and improves the model's convergence stability and generalization ability.

[0126] Optionally, the preset regression loss function can be a smoothing loss function, such as the Huber loss function (smoothed mean absolute error loss) or the log-cosh loss function (log-hyperbolic cosine loss function).

[0127] In a specific embodiment, when the preset regression loss function adopts the Huber loss function, the indicator factor loss information of the return indicator factor dimension can be expressed as the following formula: , in, For numerical stability constants (e.g.) ), This is a threshold hyperparameter (e.g., 1.0).

[0128] In a specific embodiment, when the preset regression loss function adopts the Huber loss function, the indicator factor loss information of the return indicator factor dimension can be expressed as the following formula:

[0129] in, For numerical stability constants (e.g.) ).

[0130] As can be seen from the above embodiments, by using a smoothing loss function, the indicator factor loss information between the indicator factor labels of each of the at least one revenue indicator factor dimension and the revenue indicator factor data of each of the at least one revenue indicator factor dimension is determined. Based on the indicator factor loss information, the dual-tower model to be trained is trained for content recommendation, which effectively alleviates the gradient oscillation caused by extreme click / conversion labels and improves the convergence stability and generalization ability of the model.

[0131] S505, based on the indicator factor loss information, the revenue indicator loss information, and the recommendation indicator loss information, performs content recommendation training on the dual-tower model to be trained, and obtains the content recommendation model.

[0132] In a specific embodiment, the content recommendation model obtained by training the dual-tower model to be trained based on the indicator factor loss information, the revenue indicator loss information, and the recommendation indicator loss information can include: S5051 (not shown in the figure) determines the first loss weight corresponding to the loss information of the return indicator, the second loss weight corresponding to the loss information of the recommendation indicator, and the third loss weight corresponding to the loss information of the indicator factor. S5052 (not shown in the figure) uses the first loss weight, the second loss weight and the third loss weight to perform loss fusion on the loss information of the revenue indicator, the loss information of the recommendation indicator and the loss information of the indicator factor to obtain the content recommendation loss information. S5053 (not shown in the figure) updates the model parameters of the dual-tower model to be trained based on the content recommendation loss information, thus obtaining the content recommendation model.

[0133] Specifically, the detailed content of "training the dual-tower model to be trained on content recommendation based on the indicator factor loss information, revenue indicator loss information and recommendation indicator loss information to obtain the content recommendation model" in step S505 is similar to the detailed content of "training the dual-tower model to be trained on content recommendation based on the revenue indicator loss information and recommendation indicator loss information to obtain the content recommendation model" in step S204, and will not be repeated here.

[0134] As can be seen from the above embodiments, based on the loss information of the revenue indicator and the loss information of the recommendation indicator, further combining the content recommendation training of the dual-tower model to be trained with the loss information of the indicator factor can effectively improve the content recommendation performance optimization effect of the dual-tower model.

[0135] See Figure 5b , Figure 5b This is a schematic diagram of the structure of a content recommendation model provided in an embodiment of this application. Specifically, the content recommendation model is based on a dual-tower model, which is explicitly decoupled into a semantically orthogonal subspace embedding structure. This allows the account tower and content tower to extract features from N revenue indicator factor dimensions (e.g., bid bias dimension, pCTR click-through rate dimension, and pCVR conversion rate dimension) and quality correction factor dimensions (e.g., factor general factor dimension), respectively, to obtain account representation vectors and content representation vectors. The account representation vectors and content representation vectors are then multiplied to obtain content recommendation indicators that serve as the basis for coarse ranking and recall, thereby enabling preliminary content recall based on the content recommendation indicators.

[0136] As can be seen from the above embodiments, by explicitly decoupling the dual-tower model in the coarse-ranking stage into a semantically orthogonal subspace embedding structure, and introducing a learnable general factor, a more efficient and efficient model can be achieved. This allows the model to maintain a single-tower dot product structure and support efficient ANN retrieval while explicitly learning the ranking preferences of two business objectives: eCPM and user experience (Final Value). This avoids the error accumulation problem of the traditional method of "fitting pCTR / pCVR first and then multiplying by bid," and naturally supports personalized balancing (because...). The space can learn user-specific offsets through the Final Value signal; in the design of loss weights, either a rule-based system or a relatively complex feedback signal design can be used. Through differentiated sub-task loss function design and adaptive loss weight scheduling mechanism, the general factor terms can be optimized. It can not only effectively approximate the final sorting target It can also achieve personalized offset compensation at the user level, thereby achieving a dynamic balance between business value (eCPM) and user experience (Final Value).

[0137] This application also provides a content recommendation model processing device. Figure 6 This is a schematic diagram of the structure of a content recommendation model processing device provided in an embodiment of this application, as shown below. Figure 6 As shown, the above-mentioned device includes: The training data acquisition module 610 is used to acquire the account attribute information of the sample account, the content description information of the sample media content corresponding to the sample account, the content revenue indicator tags of the sample media content, and the content recommendation indicator tags of the sample media content. The recommendation prediction module 620 is used to input account attribute information and content description information into the dual-tower model to be trained. Based on the feature extraction of the dual-tower model from the dimensions of content revenue indicators and quality correction factors, the module performs recommendation prediction on the sample accounts for the sample media content, and obtains the content revenue indicator data and the content recommendation indicator data corresponding to the sample media content; the content revenue indicator data is the indicator factor of the content recommendation indicator data. The loss determination module 630 is used to determine revenue indicator loss information and recommendation indicator loss information based on content revenue indicator tags, content recommendation indicator tags, content revenue indicator data and content recommendation indicator data. The model training module 640 is used to train the dual-tower model to be trained on content recommendation based on the loss information of the revenue indicator and the loss information of the recommendation indicator, so as to obtain the content recommendation model.

[0138] In one specific embodiment, the dual-tower model to be trained includes: an account tower and a content tower, and the aforementioned recommendation prediction module 620 includes: The account feature extraction unit is used to input account attribute information into the account tower, extract account features from the content revenue index dimension and the quality correction factor dimension, and obtain the account representation vector. The content feature extraction unit is used to input content description information into the content tower, extract content features from the content description information from the dimensions of content revenue indicators and quality correction factors, and obtain the content representation vector. The vector fusion processing unit is used to perform vector fusion processing on the account representation vector and the content representation vector to determine the content revenue indicator data corresponding to the content revenue indicator dimension and the quality correction factor data corresponding to the quality correction factor dimension. The indicator fusion unit is used to fuse content revenue indicator data and quality correction factor data to obtain content recommendation indicator data.

[0139] In one specific embodiment, the content revenue indicator dimension includes at least one revenue indicator factor dimension. The training data acquisition module 610 is also used to acquire the revenue indicator factor labels of the sample media content for each of the at least one revenue indicator factor dimension. The aforementioned vector fusion processing unit is also used to: perform vector fusion processing on the account representation vector and the content representation vector to determine the revenue indicator factor data and the quality correction factor data corresponding to the quality correction factor dimension of at least one revenue indicator factor dimension; and perform indicator fusion on the revenue indicator factor data of at least one revenue indicator factor dimension to obtain content revenue indicator data. The aforementioned model training module 640 is further configured to: determine the indicator factor loss information for each of the at least one revenue indicator factor dimensions based on the revenue indicator factor labels and revenue indicator factor data for each of the at least one revenue indicator factor dimensions; and perform content recommendation training on the dual-tower model to be trained based on the indicator factor loss information, revenue indicator loss information, and recommendation indicator loss information to obtain a content recommendation model.

[0140] In one specific embodiment, the model training module 640 includes: The loss weight determination unit is used to determine the first loss weight corresponding to the loss information of the revenue indicator and the second loss weight corresponding to the loss information of the recommendation indicator. The loss fusion unit is used to perform loss fusion on the revenue indicator loss information and the recommendation indicator loss information based on the first loss weight and the second loss weight to obtain content recommendation loss information. The model update unit is used to update the model parameters of the dual-tower model to be trained based on the content recommendation loss information, so as to obtain the content recommendation model.

[0141] In one specific embodiment, the loss weight determination unit includes: The online data acquisition unit is used to acquire incremental online metrics data from the content recommendation platform where the content recommendation model resides. The first loss weight adjustment unit is used to adjust the loss weight based on the incremental data of online indicators, inputting it into the loss weight adjustment model to determine the first loss weight corresponding to the loss information of the revenue indicator and the second loss weight corresponding to the loss information of the recommendation indicator.

[0142] In a specific embodiment, the loss weight adjustment model is a reinforcement learning model. The first loss weight adjustment unit is further configured to: input online indicator incremental data into the reinforcement learning model; based on the adjustment reward representation information corresponding to the reinforcement learning model, determine the target weight adjustment data with the largest long-term reward data from a variety of preset weight adjustment data under the online indicator incremental data; the adjustment reward representation information is used to represent the long-term reward data obtained by deciding different preset weight adjustment data under different online indicator incremental data; the long-term reward data is used to represent the feedback effect of weight adjustment accumulated since the weight adjustment training of the reinforcement learning model began; and adjust the current loss weight of the dual-tower model to be trained based on the target weight adjustment data to obtain the first loss weight and the second loss weight.

[0143] In one specific embodiment, the loss weight determination unit includes: The group characteristic information determination unit is used to determine the group characteristic information of the target account group and the scene characteristic information of the target recommendation scenario; The second loss weight adjustment unit is used to adjust the loss weight based on group feature information and scene feature information, and to determine the first loss weight and the second loss weight.

[0144] In one specific embodiment, the loss weight determination unit includes: The candidate loss acquisition unit is used to acquire multiple candidate loss weight recombinations. Each candidate loss weight recombination includes: candidate loss weights corresponding to the loss information of the revenue indicator and candidate loss weights corresponding to the loss information of the recommendation indicator. The grouped control test unit is used to conduct grouped control tests on multiple candidate loss weight reorganizations and obtain test results. The loss screening unit is used to screen the target loss weight reorganization from multiple candidate loss weight reorganizations based on the test results; The third loss weight adjustment unit is used to determine the first loss weight and the second loss weight based on the target loss weight reorganization.

[0145] In one specific embodiment, the sample media content includes a sampled content set, and the loss determination module 630 includes: The revenue indicator loss determination unit is used to perform content revenue difference analysis on the content revenue indicator labels and content revenue indicator data of each sampled content set based on the first preset ranking loss function, and to obtain revenue indicator loss information. The recommendation index loss determination unit is used to perform content recommendation difference analysis on the content recommendation index labels and content recommendation index data of each sampled content set based on the second preset ranking loss function, and obtain recommendation index loss information.

[0146] In one specific embodiment, the model training module 640 includes: The indicator factor loss determination unit is used to perform difference analysis on the indicator factor label and the indicator factor data corresponding to each indicator factor dimension based on a preset regression loss function, and to determine the indicator factor loss information corresponding to each indicator factor dimension.

[0147] In one specific embodiment, the sample media content corresponding to the sample account is collected by the following device: The historical candidate content acquisition module is used to acquire the historical fine-ranking candidate content set corresponding to the sample account; The hierarchical negative sampling module is used to perform hierarchical negative sampling on the historical fine-ranking candidate content set to obtain the sampled content set; the sampled content set is used as the sample media content.

[0148] In one specific embodiment, the content revenue metric tag corresponding to the sample media content is obtained in the following way: The revenue indicator labeling module is used to input sample media content and sample media content into the pre-trained fine ranking model to label the content revenue indicators and obtain the content revenue indicator labels corresponding to the sample media content.

[0149] In a specific embodiment, the content recommendation metric tags corresponding to the sample media content are obtained in the following way: The interactive feedback metric collection module is used to collect various interactive feedback metrics from sample accounts for each sampled content; The recommendation indicator labeling module is used to perform indicator fusion processing on multiple interactive feedback indicators to obtain content recommendation indicator labels for each sampled content.

[0150] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0151] Figure 7 This is a block diagram of an electronic device for content recommendation model processing provided in an embodiment of this application. The electronic device can be a terminal, and its internal structure diagram can be as follows. Figure 7As shown, the electronic device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a content recommendation model processing method. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the device's casing, or an external keyboard, touchpad, or mouse. Figure 8 This is a block diagram of another electronic device for content recommendation model processing provided in the embodiments of this application. The electronic device can be a server, and its internal structure diagram can be as follows: Figure 8 As shown, the electronic device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a content recommendation model processing method. Those skilled in the art will understand that Figure 7 or Figure 8 The structure shown is merely a block diagram of a portion of the structure related to the present disclosure and does not constitute a limitation on the electronic device to which the present disclosure is applied. A specific electronic device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements. In an exemplary embodiment, an electronic device is also provided, including: a processor; and a memory for storing processor-executable instructions; wherein the processor is configured to execute the instructions to implement the content recommendation model processing method as described in the embodiments of this disclosure.

[0152] In an exemplary embodiment, a computer-readable storage medium is also provided, wherein when the instructions in the storage medium are executed by a processor of an electronic device, the electronic device is enabled to perform the content recommendation model processing method of the embodiments of this disclosure. In an exemplary embodiment, a computer program product or computer program is also provided, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the content recommendation model processing method provided in the various optional implementations described above.

[0153] As can be seen from the above embodiments of this application, in the coarse ranking stage of content recommendation, the account attribute information of the sample account, the content description information of the sample media content corresponding to the sample account, the content revenue indicator tags of the sample media content, and the content recommendation indicator tags of the sample media content are obtained. The account attribute information and content description information are input into the dual-tower model to be trained. Based on the feature extraction of the dual-tower model from the content revenue indicator dimension and the quality correction factor dimension, the recommendation prediction of the sample account for the sample media content is performed, and the content revenue indicator data and the content recommendation indicator data corresponding to the sample media content are obtained. The content revenue indicator data is the indicator factor of the content recommendation indicator data. This is achieved by explicitly separating the feature extraction dimensions of the dual-tower model into the content revenue indicator dimension and the quality correction factor dimension. The positive factor dimension allows the content recommendation metrics output by the dual-tower model to simultaneously measure content revenue and content quality. Based on content revenue metric labels, content recommendation metric labels, content revenue metric data, and content recommendation metric data, revenue metric loss information and recommendation metric loss information are determined. Then, based on these loss information, the dual-tower model to be trained is used for content recommendation training, resulting in a content recommendation model. During model training, multi-objective collaborative optimization can be performed on content revenue and content recommendation metrics. This improves the stability of the dual-tower model's content recommendation training, effectively balancing the potential content revenue of recommended content with the user browsing experience, enhancing the consistency of ranking between the coarse and fine ranking stages, and thus improving the overall recommendation performance of the content recommendation system.

[0154] It is understood that in the specific implementation of this application, user-related data is involved. When the above 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.

[0155] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. This computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media in the processing steps of the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0156] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the claims.

[0157] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims.

Claims

1. A content recommendation model processing method, characterized in that, The method includes: Obtain the account attribute information of the sample account, the content description information of the sample media content corresponding to the sample account, the content revenue indicator tags of the sample media content, and the content recommendation indicator tags of the sample media content; The account attribute information and the content description information are input into the dual-tower model to be trained. Based on the feature extraction of the dual-tower model from the dimensions of content revenue indicators and quality correction factors, the model performs recommendation prediction on the sample accounts for the sample media content, thereby obtaining the content revenue indicator data and the content recommendation indicator data corresponding to the sample media content; the content revenue indicator data is the indicator factor of the content recommendation indicator data. Based on the content revenue indicator tags, the content recommendation indicator tags, the content revenue indicator data, and the content recommendation indicator data, determine revenue indicator loss information and recommendation indicator loss information; Based on the revenue loss information and the recommendation loss information, the dual-tower model to be trained is trained for content recommendation to obtain a content recommendation model.

2. The method according to claim 1, characterized in that, The dual-tower model to be trained includes an account tower and a content tower. The account attribute information and content description information are input into the dual-tower model to be trained. Based on feature extraction from the content revenue indicator dimension and quality correction factor dimension, the model performs recommendation prediction on the sample accounts for the sample media content, obtaining content revenue indicator data and content recommendation indicator data corresponding to the sample media content, including: The account attribute information is input into the account tower, and account features are extracted from the account attribute information from the dimensions of content revenue indicators and quality correction factors to obtain the account representation vector. The content description information is input into the content tower, and content features are extracted from the content description information from the dimensions of content revenue indicators and quality correction factors to obtain a content representation vector. The account representation vector and the content representation vector are fused to determine the content revenue indicator data corresponding to the content revenue indicator dimension and the quality correction factor data corresponding to the quality correction factor dimension. The content revenue indicator data and the quality correction factor data are fused to obtain the content recommendation indicator data.

3. The method according to claim 2, characterized in that, The content revenue metrics dimension includes at least one revenue metric factor dimension, and the method further includes: Obtain the revenue indicator factor labels for each of the at least one revenue indicator factor dimension of the sample media content. The step of performing vector fusion processing on the account representation vector and the content representation vector to determine the content revenue indicator data corresponding to the content revenue indicator dimension and the quality correction factor data corresponding to the quality correction factor dimension includes: The account representation vector and the content representation vector are fused to determine the revenue indicator factor data for each of the at least one revenue indicator factor dimension and the quality correction factor data corresponding to the quality correction factor dimension. The content revenue indicator data is obtained by fusing the revenue indicator factor data of each of the at least one revenue indicator factor dimensions. The step of training the dual-tower model to be trained on content recommendation based on the revenue indicator loss information and the recommendation indicator loss information to obtain the content recommendation model includes: Based on the respective return indicator factor labels and return indicator factor data of each of the at least one return indicator factor dimension, determine the indicator factor loss information of each of the at least one return indicator factor dimension. Based on the loss information of the indicator factors, the loss information of the revenue indicator, and the loss information of the recommendation indicator, the dual-tower model to be trained is trained for content recommendation to obtain the content recommendation model.

4. The method according to claim 1, characterized in that, The step of training the dual-tower model to be trained on content recommendation based on the revenue indicator loss information and the recommendation indicator loss information to obtain the content recommendation model includes: Determine the first loss weight corresponding to the loss information of the revenue indicator and the second loss weight corresponding to the loss information of the recommendation indicator; Based on the first loss weight and the second loss weight, loss fusion is performed on the revenue indicator loss information and the recommendation indicator loss information to obtain content recommendation loss information; Based on the content recommendation loss information, the model parameters of the dual-tower model to be trained are updated to obtain the content recommendation model.

5. The method according to claim 4, characterized in that, Determining the first loss weight corresponding to the loss information of the revenue indicator and the second loss weight corresponding to the loss information of the recommendation indicator includes: Obtain incremental online metrics data from the content recommendation platform where the content recommendation model resides; Based on the incremental data of the online indicators, the loss weight is adjusted by inputting the loss weight adjustment model to determine the first loss weight corresponding to the loss information of the revenue indicator and the second loss weight corresponding to the loss information of the recommendation indicator.

6. The method according to claim 5, characterized in that, The loss weight adjustment model is a reinforcement learning model. The step of adjusting the loss weights based on the incremental online indicator data, and determining the first loss weight corresponding to the loss information of the revenue indicator and the second loss weight corresponding to the loss information of the recommendation indicator, includes: The online indicator incremental data is input into the reinforcement learning model. Based on the adjustment reward representation information corresponding to the reinforcement learning model, the target weight adjustment data with the largest long-term reward data is determined from a variety of preset weight adjustment data under the online indicator incremental data. The adjustment reward representation information is used to represent the long-term reward data obtained by deciding different preset weight adjustment data under different online indicator incremental data. The long-term reward data is used to represent the feedback effect of weight adjustment accumulated since the weight adjustment training of the reinforcement learning model. Based on the target weight adjustment data, the current loss weights of the dual-tower model to be trained are adjusted to obtain the first loss weight and the second loss weight.

7. The method according to claim 4, characterized in that, Determining the first loss weight corresponding to the loss information of the revenue indicator and the second loss weight corresponding to the loss information of the recommendation indicator includes: Determine the group characteristics of the target account group and the scenario characteristics of the target recommendation scenario; Based on the group feature information and the scene feature information, the loss weight is adjusted to determine the first loss weight and the second loss weight.

8. The method according to claim 4, characterized in that, Determining the first loss weight corresponding to the loss information of the revenue indicator and the second loss weight corresponding to the loss information of the recommendation indicator includes: Multiple candidate loss weight reorganizations are obtained, and each candidate loss weight reorganization includes: the candidate loss weight corresponding to the loss information of the revenue indicator and the candidate loss weight corresponding to the loss information of the recommendation indicator. The multiple candidate loss weight reorganizations were subjected to a grouped comparative test to obtain the test results; Based on the test results, a target loss right restructuring is selected from the plurality of candidate loss right restructurings; Based on the target loss weight reorganization, the first loss weight and the second loss weight are determined.

9. The method according to claim 1, characterized in that, The sample media content includes: a sampled content set. The determination of revenue indicator loss information and recommendation indicator loss information based on the content revenue indicator label, the content recommendation indicator label, the content revenue indicator data, and the content recommendation indicator data includes: Based on the first preset ranking loss function, a content revenue difference analysis is performed on the content revenue index labels of each of the sampled content sets and the content revenue index data of each of the sampled content sets to obtain the revenue index loss information. Based on the second preset ranking loss function, a content recommendation difference analysis is performed on the content recommendation index tags and content recommendation index data of each of the sampled content sets to obtain the recommendation index loss information.

10. The method according to claim 3, characterized in that, The step of determining the indicator factor loss information for each of the at least one revenue indicator factor dimensions based on the respective revenue indicator factor labels and the respective revenue indicator factor data for each of the at least one revenue indicator factor dimensions includes: Based on a preset regression loss function, a difference analysis is performed on the return indicator factor label corresponding to each return indicator factor dimension and the return indicator factor data corresponding to each return indicator factor dimension to determine the indicator factor loss information corresponding to each return indicator factor dimension.

11. The method according to claim 1, characterized in that, The sample media content corresponding to the sample accounts was collected in the following manner: Obtain the historical ranking candidate content set corresponding to the sample account; Hierarchical negative sampling is performed on the historical fine-rank candidate content set to obtain the sampled content set; The sampled content set is used as the sample media content.

12. The method according to claim 11, characterized in that, The content revenue metric tags corresponding to the sample media content were obtained in the following way: The sample media content and the sample media content are input into a pre-trained fine ranking model to label the content revenue indicators, thereby obtaining the content revenue indicator labels corresponding to the sample media content.

13. The method according to claim 11, characterized in that, The content recommendation metric tags corresponding to the sample media content were obtained in the following way: Collect multiple interactive feedback metrics from the sample accounts for each sampled content; The various interactive feedback metrics are fused to obtain the content recommendation metric tag for each sampled content.

14. A content recommendation model processing device, characterized in that, The device includes: The training data acquisition module is used to acquire account attribute information of sample accounts, content description information of sample media content corresponding to the sample accounts, content revenue indicator tags and content recommendation indicator tags corresponding to the sample media content; The recommendation prediction module is used to input the account attribute information and the content description information into the dual-tower model to be trained. Based on the feature extraction of the dual-tower model from the dimensions of content revenue indicators and quality correction factors, the module performs recommendation prediction on the sample account for the sample media content, and obtains the content revenue indicator data and the content recommendation indicator data corresponding to the sample media content; the content revenue indicator data is the indicator factor of the content recommendation indicator data. The loss determination module is used to determine revenue indicator loss information and recommendation indicator loss information based on the content revenue indicator label, the content recommendation indicator label, the content revenue indicator data and the content recommendation indicator data. The model training module is used to train the dual-tower model to be trained on content recommendation based on the revenue indicator loss information and the recommendation indicator loss information, so as to obtain a content recommendation model.

15. A content recommendation model processing device, characterized in that, The device includes a processor and a memory, the memory storing at least one instruction or at least one program, the at least one instruction or the at least one program being loaded and executed by the processor to implement the content recommendation model processing method as described in any one of claims 1 to 13.

16. A computer-readable storage medium, characterized in that, The storage medium stores at least one instruction or at least one program segment, which is loaded and executed by a processor to implement the content recommendation model processing method as described in any one of claims 1 to 13.

17. A computer program product, characterized in that, The computer program product includes at least one instruction or at least one program segment, which is loaded and executed by a processor to implement the content recommendation model processing method as described in any one of claims 1 to 13.

Images