Content recommendation and double-tower content recommendation model training method and device

Abstract

The embodiment of the invention provides a content recommendation method and device and a double-tower content recommendation model training method and device, and the content recommendation method comprises the steps: obtaining the content information of to-be-recommended content and the user information of a target user in response to a content recommendation task; the user information and the content information are input into a double-tower content recommendation model, recommended content for the target user is obtained, the double-tower content recommendation model comprises a user tower and a content tower, the user tower is used for extracting user features based on the user information, and the content tower comprises a feature adaptation layer; the feature adaptation layer is used for obtaining target content features corresponding to a task target of the content recommendation task based on the content information, and the recommendation content is obtained by decoding based on the target content features and the user features. The content representation can be dynamically adjusted according to the task target through the feature adaptation layer, so that the same content presents differentiated feature expression under different tasks, and the adaptation precision of the recommendation result to the specific task target is improved on the premise of not changing the double-tower structure.

Description

Technical Field

[0001] The embodiments in this specification relate to the field of artificial intelligence technology, and in particular to a content recommendation method and apparatus for training a dual-tower content recommendation model. Background Technology

[0002] In the era of information overload on the internet, personalized recommendation systems have become a core technology connecting users with massive amounts of content. They are widely used in fields such as search, advertising, and e-commerce, aiming to improve user experience and platform efficiency.

[0003] Current advertising recommendation systems generally adopt multi-objective modeling to meet the conversion needs of different advertisers, but they face severe challenges in the coarse ranking stage: on the one hand, they need to process a large number of candidate contents with extremely low latency; on the other hand, traditional single-tower or multi-head structures are difficult to balance gradient conflicts and computational overhead between multiple objectives, making it difficult to balance prediction accuracy and system performance.

[0004] Therefore, there is an urgent need for a content recommendation method that can balance multi-target push and latency processing. Summary of the Invention

[0005] In view of this, embodiments of this specification provide a content recommendation method. One or more embodiments of this specification also relate to a dual-tower content recommendation model training method, a content recommendation apparatus, a dual-tower content recommendation model training apparatus, a computing device, a computer-readable storage medium, and a computer program product, to address the technical deficiencies existing in the prior art.

[0006] According to a first aspect of the embodiments of this specification, a content recommendation method is provided, comprising:

[0007] In response to content recommendation tasks, obtain content information of the content to be recommended and user information of the target users;

[0008] User information and content information are input into a dual-tower content recommendation model to obtain recommended content for the target user. The dual-tower content recommendation model includes a user tower and a content tower. The user tower is used to extract user features based on user information, and the content tower includes a feature adaptation layer. The feature adaptation layer is used to obtain the target content features corresponding to the task objective of the content recommendation task based on the content information. The recommended content is obtained by decoding the target content features and user features.

[0009] According to a second aspect of the embodiments of this specification, a method for training a dual-tower content recommendation model is provided, comprising:

[0010] In response to the sample recommendation task, obtain sample user information and sample content information;

[0011] The sample user information is input into the user tower of the initial dual-tower recommendation model to obtain the sample user features. The initial dual-tower recommendation model includes a user tower and a content tower, and the content tower includes a feature adaptation layer.

[0012] Based on sample content information, the sample content features corresponding to the task objective of the sample recommendation task are obtained through the feature adaptation layer of the content tower.

[0013] Based on the characteristics of sample users and the characteristics of sample content, a feature adaptation layer is trained to obtain a dual-tower recommendation model.

[0014] According to a third aspect of the embodiments of this specification, a content recommendation apparatus is provided, comprising:

[0015] The first acquisition module is configured to respond to the content recommendation task by acquiring the content information of the content to be recommended and the user information of the target user.

[0016] The recommendation module is configured to input user information and content information into a dual-tower content recommendation model to obtain recommended content for the target user. The dual-tower content recommendation model includes a user tower and a content tower. The user tower is used to extract user features based on user information, and the content tower includes a feature adaptation layer. The feature adaptation layer is used to obtain the target content features corresponding to the task objective of the content recommendation task based on the content information. The recommended content is obtained by decoding the target content features and user features.

[0017] According to a fourth aspect of the embodiments of this specification, a dual-tower content recommendation model training apparatus is provided, comprising:

[0018] The second acquisition module is configured to acquire sample user information and sample content information in response to the sample recommendation task.

[0019] The input module is configured to input sample user information into the user tower of the initial dual-tower recommendation model to obtain sample user features. The initial dual-tower recommendation model includes a user tower and a content tower, and the content tower includes a feature adaptation layer.

[0020] The adaptation module is configured to obtain the sample content features corresponding to the task target of the sample recommendation task through the feature adaptation layer of the content tower based on the sample content information.

[0021] The training module is configured to train a feature adaptation layer based on sample user features and sample content features to obtain a dual-tower recommendation model.

[0022] According to a fifth aspect of the embodiments of this specification, a computing device is provided, comprising:

[0023] Memory and processor;

[0024] The memory is used to store computer programs / instructions, and the processor is used to execute the computer programs / instructions. When the computer programs / instructions are executed by the processor, they implement the steps of the above-described content recommendation method or the dual-tower content recommendation model training method.

[0025] According to a sixth aspect of the embodiments of this specification, a computer-readable storage medium is provided that stores a computer program / instructions, which, when executed by a processor, implement the steps of the above-described content recommendation method or the dual-tower content recommendation model training method.

[0026] According to a seventh aspect of the embodiments of this specification, a computer program product is provided, including a computer program / instructions that, when executed by a processor, implement the steps of the above-described content recommendation method or dual-tower content recommendation model training method.

[0027] One embodiment of this specification implements a response to a content recommendation task, acquiring content information of the content to be recommended and user information of the target user; inputting the user information and content information into a dual-tower content recommendation model to obtain recommended content for the target user. The dual-tower content recommendation model includes a user tower and a content tower. The user tower is used to extract user features based on user information, and the content tower includes a feature adaptation layer. The feature adaptation layer is used to obtain target content features corresponding to the task objective of the content recommendation task based on the content information. The recommended content is obtained by decoding the target content features and user features. The feature adaptation layer generates target content features corresponding to the task objective of the content recommendation task based on the content information, and combines this with user features to decode the recommended content. Because the feature adaptation layer can dynamically adjust the content representation according to the task objective, the same content can present differentiated feature expressions under different tasks, thereby improving the adaptation accuracy of the recommendation results to specific task objectives without changing the dual-tower structure. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of a multi-objective prediction system involving multiple tasks.

[0029] Figure 2 This is a simplified structural diagram of multiple tasks in multi-objective prediction;

[0030] Figure 3 This is a schematic diagram comparing decoupled interest modeling and multi-objective modeling according to one embodiment of this specification;

[0031] Figure 4 This is a flowchart illustrating a content recommendation method provided in one embodiment of this specification;

[0032] Figure 5 This is a flowchart of a dual-tower content pushing model training method provided in one embodiment of this specification;

[0033] Figure 6 This is a flowchart illustrating the processing procedure of a content recommendation method provided in one embodiment of this specification;

[0034] Figure 7 This is a schematic diagram of the structure of a dual-tower content recommendation model provided in one embodiment of this specification;

[0035] Figure 8 This is a schematic diagram of a low-rank adaptation module in a dual-tower content recommendation model provided in one embodiment of this specification;

[0036] Figure 9 This is a schematic diagram of the structure of a content recommendation device provided in one embodiment of this specification;

[0037] Figure 10 This is a schematic diagram of the structure of a dual-tower content recommendation model training device provided in one embodiment of this specification;

[0038] Figure 11 This is a structural block diagram of a computing device provided in one embodiment of this specification. Detailed Implementation

[0039] Many specific details are set forth in the following description to provide a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.

[0040] The terminology used in one or more embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of this specification. The singular forms “a,” “described,” and “the” as used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0041] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this specification, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0042] Furthermore, it should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in one or more embodiments of this specification are all information and data authorized by the user or fully authorized by all parties. Moreover, the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0043] First, the terms and concepts used in one or more embodiments of this specification will be explained.

[0044] Multi-head: In multi-objective prediction scenarios, "multi-head" refers to a model that sets up independent prediction branches for multiple tasks. Each "head" is responsible for a specific task (such as click-through rate, conversion rate, etc.). By sharing low-level features and separating the upper-level modeling, optimization conflicts between different objectives can be mitigated.

[0045] Head: In multi-objective prediction, a "head" is an independent prediction unit in a multi-objective prediction model, corresponding to a single task objective (such as predicting the probability of a user placing an order). Its input usually comes from shared underlying feature representations, and its output is the prediction value for that task. Multiple "heads" together constitute a "multi-head structure".

[0046] Embedding: An embedding vector is a vector that maps sparse IDs to a high-dimensional space. An embedding vector is a dense vector that maps sparse discrete identifiers (such as user IDs and ad IDs) to a continuous high-dimensional semantic space. It is used to represent the latent semantic information of the entity and is the basic representation for feature calculation in deep recommendation and advertising models.

[0047] CVR stands for conversion rate. It refers to the probability that a user will complete a specific target action (such as placing an order, making a payment, or leaving contact information) after clicking on an ad. It is one of the core indicators for evaluating and ranking ad performance and is often used as the optimization target in coarse-grained or fine-grained ranking models.

[0048] MLP stands for Multi-layer Perceptron. A multi-layer perceptron is a feedforward neural network composed of multiple fully connected layers. It extracts high-order abstract representations by performing layer-by-layer nonlinear transformations on the input features. In recommender systems, it is often used for feature encoding in the user tower or content tower.

[0049] Item: In advertising, an "item" refers to a candidate ad unit to be scored and ranked. It is the content-side input object of the recommendation or coarse-grained ranking model, and its features are processed by the content pyramid to generate an embedding vector for matching user interests.

[0050] Lora: Low-rank adaptation, low-rank matrix factorization. Low-rank matrix factorization is a parameter-efficient adaptation technique that introduces a learnable incremental matrix of low-rank factorization as a bypass to the original weight matrix, enabling lightweight task customization of the model output and supporting multi-objective differentiated modeling without modifying the backbone network.

[0051] Large-scale advertising systems, due to strict latency and resource constraints, generally employ a cascaded architecture of recall-coarse ranking-fine ranking to select the ads that best meet user interests from a massive pool of candidates. Coarse ranking, located upstream, needs to deliver results that meet certain accuracy requirements under a larger candidate set and stricter time constraints; a dual-tower coarse ranking model has become a natural choice. On the other hand, advertising operations also face the challenge of multi-objective prediction because different advertisers focus on different conversion behaviors. The relationships between different objectives vary significantly; predicting dozens of different objectives using a single head can lead to conflicting gradient directions among these objectives, causing optimization of one objective to negatively impact others.

[0052] There are numerous optimization objectives for online search advertising, and the physical meanings of these objectives vary greatly. Training multiple models separately for different optimization objectives is unacceptable in terms of resource and maintenance costs.

[0053] Compared to fine-ranking, the coarse-ranking stage features a larger candidate set and shorter inference time. Without addressing performance issues, a conventional fully connected model cannot be used. Typically, a dual-tower model is employed: a fully connected model is computed in real-time for user features, and another fully connected model is computed near-line for item features. The inner product of the two embeddings is used to calculate the final probability value.

[0054] A general solution to the multi-objective gradient conflict problem is a multi-head approach similar to fine-grained ranking models, where the underlying embeddings are shared, and different objectives are estimated separately using independent MLP towers. (See also...) Figure 1 As shown, Figure 1 This is a schematic diagram of a multi-objective prediction structure for multiple tasks, illustrating an architecture that constructs independent dual-tower models for each prediction objective. Inputs of "user information" and "content information" are fed into the corresponding "user tower" and "content tower" for each task, respectively. After generating corresponding features, a matching score is calculated, outputting "Predicted Target 1," "Predicted Target 2," and "Predicted Target 3." This structure achieves accurate prediction for different optimization objectives by independently modeling each target, but it generates dozens of different user embeddings and item membeddings. Online systems only require a single CVR scoring system, resulting in relatively high storage pressure compared to fine-grained ranking.

[0055] Therefore, a multi-head solution is urgently needed for coarse-sorting deployment to resolve multi-objective gradient conflicts, while minimizing the increase in computational power within the constraints of the coarse-sorting dual-tower structure. One implementation method is described below. Figure 2 As shown, Figure 2 This is a simplified structural diagram of multiple tasks in multi-objective prediction, demonstrating a simplified architecture of multi-task shared content towers. Inputting "user information" enters the "user tower" to generate user features, while inputting "content information" enters three parallel "content towers" for processing. Their outputs are merged into "content features," which are then combined with "optimization objectives and constraint types" information to finally generate the "predicted objective." This structure reduces redundant computation and improves resource utilization by sharing the user tower and some content processing paths.

[0056] The specific goal of CVR scoring is determined by the attributes of the advertisement. Filtering out the required user embedding and item embedding based on the advertisement attributes can reduce unnecessary inner product calculations.

[0057] Considering that only the item tower in the coarse-sorted dual-tower architecture has an advertising attribute field, and multi-objective information can be perceived before the embedding is generated, we choose to perform multi-objective modeling only on the item embedding. After generating multiple item embeddings, we select the embedding that will ultimately participate in the CVR calculation through the advertising attribute field.

[0058] Since there are many existing targets (30+), this approach would cause the number of parameters in the item tower to increase by tens of times, affecting current training and performance. Therefore, further optimization is needed to address this issue.

[0059] See this manual. Figure 3 As shown, Figure 3This diagram illustrates a comparison of decoupled interest modeling and multi-objective modeling, provided by one embodiment of this specification. It compares traditional and improved content feature representation methods in a two-dimensional vector space. In the left diagram, user features, along with in-store content features 1, 2, order content features 1, and 2, collectively constitute the vector space. In-store content features 1 and order content features 1 are in similar directions, indicating partial semantic overlap in the "in-store" and "order" tasks. In-store content features 2 and order content features 2 have significantly different directions, reflecting semantic shifts in their expression under different task objectives. In the right diagram, interest content features 1 and in-store content features 1 are in similar directions, indicating high relevance in the "interest" and "in-store" tasks. Interest content features 2 and order content features 2 are in similar directions, reflecting their correlation between the "interest" and "order" tasks. The angles between user features and each content feature reflect the degree of matching between the user and the content under different task objectives. Overall, this reveals how the feature adaptation layer dynamically adjusts content representation according to task objectives to achieve accurate multi-task perception and recommendation.

[0060] The process of calculating similarity in a coarse-sorted dual-tower multi-head algorithm is simplified in two-dimensional space. Typically, there is a high degree of correlation between multiple targets, which is reflected spatially by the close proximity of their corresponding item embeddings. Their center can be considered as the item embedding corresponding to the user's interest target.

[0061] Compared to learning individual item embeddings for each task, multi-target modeling by sharing a single interest item embedding, and then adding a multi-target embedding on top of the interest embedding, results in a shorter and easier learning path. Due to the correlation between multiple targets, modeling the multi-target embedding requires very few parameters. Therefore, multi-head modeling can be performed only in the last layer of the MLP. LoRa techniques can be used to further reduce the number of parameters in the multi-head model.

[0062] For new targets, training from the interest vector requires fewer samples and is more accurate than training from point 0. Theoretically, decoupling interest modeling and multi-target modeling will also improve performance for sparse targets.

[0063] After resolving the multi-objective gradient conflict problem, the multi-head structure allows the prediction of multiple objectives to be unified into a single model, enabling offline training of only one model and saving training resources. Combined with engineering pipeline modifications, it also saves inference resources required for online prediction of multiple models.

[0064] Dual-tower models require the transmission and storage of two embedding vectors, typically 128-dimensional, during inference. Conventional multi-head modeling methods for solving multi-object prediction incur tens of times greater communication and storage overhead in dual-tower models.

[0065] To address the aforementioned issues, this specification applies to the coarse-ranking stage of search advertising. Facing challenges of high latency requirements and a large scoring set, it designs a model structure based on low-rank matrix factorization. This structure achieves accurate prediction of conversion probabilities for multiple different targets with relatively low online inference resources. This structure can effectively model the commonalities and differences between different targets, showing a significant performance improvement compared to the one-head structure, saving online inference resources compared to the multi-head structure, and easily adapting to new targets. It is a relatively universal and effective technical solution.

[0066] This specification provides a content recommendation method. One or more embodiments of this specification also relate to a dual-tower content recommendation model training method, a content recommendation apparatus, a dual-tower content recommendation model training apparatus, a computing device, a computer-readable storage medium, and a computer program product, which will be described in detail in the following embodiments.

[0067] See Figure 4 , Figure 4 This is a flowchart of a content recommendation method provided in one embodiment of this specification, specifically including the following steps 402-404.

[0068] Step 402: In response to the content recommendation task, obtain the content information of the content to be recommended and the user information of the target user.

[0069] The content information for the recommended ads is a dataset describing the attributes of the candidate ads, including static or recently updated feature fields such as ad identifiers, industry categories, creative types, and landing page characteristics. For example, the content information for an e-commerce ad might include the product category "mobile phones," the brand "Brand X," and the advertiser's optimization goal "order conversion."

[0070] The target user's information is a set of features that characterize the current requesting user's interests and state, including the user's historical behavior sequence, demographic attributes, and real-time context (such as search terms and geographic location). For example, user information may include the sequence of recently clicked ad categories "[mobile phone, headphones, charger]", the city "Beijing", and the current device type "iOS".

[0071] In practical applications, when a system responds to a content recommendation task, it first parses the task's contextual metadata to determine its target objective (e.g., "estimated add-to-cart conversion rate"). Then, it retrieves the content information of the advertisements to be recommended from the content database or cache, and obtains the user information of the corresponding target user from the user profile service or real-time logs. The acquisition of content and user information can be performed in parallel, and both are output as structured feature vectors for subsequent processing by the dual-tower model. The source of content information can be an offline-built advertising feature library or a near-line-updated streaming feature pipeline; user information can be obtained by querying the user feature center based on the user ID or by aggregating real-time behavior streams.

[0072] For the step "Responding to the content recommendation task, obtain the content information of the content to be recommended and the user information of the target user", one possible approach is: when the system receives a recommendation request, it extracts the task identifier from the request header, loads the corresponding content feature template and user feature template from the configuration center based on the identifier, and then fills in the specific feature values ​​from each data source according to the template; another possible implementation approach is: the system pre-stores the basic features shared by all tasks in a unified manner, and only dynamically concatenates the task-specific feature fields according to the task target at runtime, thereby reducing the overhead of repeated reading.

[0073] For example, in the coarse-ranking service of a search advertising platform, when the system receives a recommendation request for the task of "car lead generation conversion," it first identifies the task objective as "lead generation" from the request parameters. Then, it retrieves content information of a batch of candidate ads from the ad feature library, including ad ID, industry "car," conversion type set by the advertiser as "form submission," and historical CTR. Simultaneously, based on the user ID in the request, the system retrieves the user's information from the user profile service, including recent search keywords "new energy vehicles," historical clicked ad category "SUV," location "Shanghai," and device type "Android." This content information and user information are organized into a fixed-dimensional feature vector, which serves as input to a dual-tower model to generate task-adaptive recommendation results. The entire acquisition process is completed within milliseconds, meeting the high throughput and low latency requirements of the coarse-ranking stage.

[0074] Step 404: Input user information and content information into the dual-tower content recommendation model to obtain recommended content for the target user. The dual-tower content recommendation model includes a user tower and a content tower. The user tower is used to extract user features based on user information. The content tower includes a feature adaptation layer, which is used to obtain the target content features corresponding to the task objective of the content recommendation task based on the content information. The recommended content is obtained by decoding the target content features and user features.

[0075] The task objective is the type of behavior to be optimized for the content recommendation task, used to instruct the feature adaptation layer to generate recommendation results related to that behavior. For example, the task objective could be different conversion behaviors such as "order conversion", "add-to-cart conversion", or "form submission".

[0076] In practical applications, the system inputs the acquired user information into the user tower of the dual-tower content recommendation model. The user tower performs a non-linear transformation on the user information using a multilayer perceptron, outputting fixed-dimensional user features. Simultaneously, content information is input into the content tower. The content tower first generates basic content features through a shared feature extraction layer, and then the feature adaptation layer dynamically adjusts these basic content features according to the task objective of the current content recommendation task, generating corresponding target content features. Finally, the system performs inner product or similarity calculation between the user features and the target content features to obtain a matching score, and ranks candidate advertisements based on this score, outputting recommended content tailored to the target user. The entire process is completed under a dual-tower architecture, where the user tower can be computed online in real time, and the output of the content tower can be pre-computed near-line, meeting the high throughput and low latency requirements of the coarse-ranking stage.

[0077] For example, in the coarse-ranking service of a search advertising platform, the system receives a request for the task of "converting home appliance orders." User information includes recent searches for "refrigerator," historical clicks on "large home appliances" ads, and the user's city, "Guangzhou." Content information includes a batch of candidate ads, where one ad's content information is the product category "refrigerator," the brand "Brand Y," and the advertiser's optimization goal "place an order." The system inputs the user information into the user tower, processes it through a three-layer MLP, and outputs 128-dimensional user features. The ad's content information is input into the content tower, and after generating basic content features through a shared MLP, the feature adaptation layer activates the corresponding low-rank adaptation parameters based on the task goal "place an order," generating target content features. The system calculates the inner product of the user features and the target content features as a score, and finally sorts the ad and other candidates by score, outputting a recommendation list. The entire process is completed within a dual-tower framework, eliminating the need to maintain an independent model for each task, significantly reducing online resource overhead.

[0078] Furthermore, user information and content information are input into the dual-tower content recommendation model to obtain recommended content for the target user, including: inputting user information into the user tower of the dual-tower content recommendation model for feature extraction to obtain user features; inputting content information into the content tower of the dual-tower content recommendation model for feature extraction and feature adaptation to obtain target content features.

[0079] The target content feature is an ad representation vector that has been processed by the content pyramid and aligned with the task objective of the current content recommendation task. It is generated by adjusting the basic content features through task awareness. For example, the target content features generated for the same ad under the "place an order" task and the features generated under the "add to cart" task exhibit different distributions in the vector space.

[0080] In practical applications, the system inputs user information into the user tower of the dual-tower content recommendation model. The user tower performs layer-by-layer nonlinear transformations on the input features using a multilayer perceptron, ultimately outputting fixed-dimensional user features. Simultaneously, content information is input into the content tower. The content tower first generates basic content features through a shared feature extraction layer (such as a multilayer MLP). Subsequently, a feature adaptation layer dynamically adjusts these basic features according to the current task objective, generating target content features. The user features and target content features are then used to calculate a matching score, thereby determining the recommended content. The computation of the user tower is completed online in real time, while the output of the content tower can be pre-computed and cached near-line, ensuring that the overall inference meets the performance requirements of the coarse-ranking stage.

[0081] For the step "inputting content information into the content tower of the dual-tower content recommendation model for feature extraction and feature adaptation to obtain target content features", one possible approach is: the feature adaptation layer introduces a low-rank adaptation module in the last layer of the MLP, whose parameters are stored independently according to the task objective, and the corresponding parameters are activated through the advertising attribute field; another possible approach is: the feature adaptation layer superimposes a task-specific bias vector on the basic content features, and combines a gating network to adjust the adaptation strength, thereby achieving task-differentiated modeling in a lightweight manner.

[0082] In the embodiments of this specification, by separating the general feature extraction of the user tower and the task-aware adaptation of the content tower, the content representation can dynamically respond to different task objectives, improve the model's ability to distinguish multi-objective transformation behaviors, and avoid the communication and storage overhead caused by the multi-head structure.

[0083] For example, in the coarse-grained ranking service of a search advertising platform, the system receives a request for a "beauty product add-to-cart conversion" task. User information includes the user's recent browsing of "lipstick," historical clicks on "makeup" category ads, and their city "Chengdu." Content information includes a batch of candidate ads, where one ad's content information is the product category "lipstick," the brand "Brand Z," and the advertiser's optimization goal "add to cart." The system inputs the user information into the user tower, processes it through a three-layer MLP, and outputs 128-dimensional user features. The ad's content information is input into the content tower, first processed by a shared MLP to generate basic content features, and then the feature adaptation layer activates the corresponding low-rank parameters based on the task goal "add to cart" to generate target content features. The system scores based on the inner product of user features and target content features, and finally outputs a ranked recommendation list. The entire process is efficiently completed within a dual-tower framework, eliminating the need to maintain independent embeddings for each task, significantly reducing online resource consumption.

[0084] Furthermore, the content tower includes at least one feature extraction layer and a feature adaptation layer located after the at least one feature extraction layer; the content information is input into the content tower of the dual-tower content recommendation model for feature extraction and feature adaptation to obtain target content features, including: processing content information through at least one feature extraction layer to generate basic content features; and through the feature adaptation layer, determining target adjustment parameters corresponding to the task objectives of the content recommendation task, and adjusting the basic content features using the target adjustment parameters to obtain target content features.

[0085] Basic content features are general content representations output by the feature extraction layer in the content pyramid, reflecting the common semantic information of advertisements and independent of specific tasks. For example, basic content features can encode shared features of advertisements such as product attributes, quality scores, and historical click-through rates.

[0086] Target adjustment parameters are learnable parameters associated with a specific task objective, used to fine-tune the basic content features for task awareness. For example, target adjustment parameters for the "place an order" task can enhance feature dimensions related to purchase intent.

[0087] The target content features are the final content representations adjusted by the feature adaptation layer, integrating basic semantics and task-specific information. For example, the target content features generated for the same advertisement under the "add to cart" and "place order" tasks will present different directions in the vector space.

[0088] In practical applications, the system inputs content information into a content pyramid. First, at least one feature extraction layer (such as a multilayer perceptron) performs a non-linear transformation on the original content features, abstracting them layer by layer to generate higher-order, task-independent basic content features. Then, the feature adaptation layer, based on the task objective of the current content recommendation task, retrieves the corresponding target adjustment parameters from a preset parameter set and applies these parameters to the basic content features to generate target content features. The feature extraction layer typically employs a fully connected network structure, with the number of layers and dimensions configured according to system resource and accuracy requirements. The feature adaptation layer is designed as a lightweight module, introducing only a small number of trainable parameters to avoid significantly increasing inference overhead. The entire processing can be completed near-line, and the output target content features are cached for real-time online matching.

[0089] For the step "Determine the target adjustment parameters corresponding to the task target based on the task target of the content recommendation task through the feature adaptation layer", one possible approach is: the target adjustment parameters are stored in the form of a low-rank matrix, each task corresponds to a set of independent decomposition matrices, and are activated by indexing the advertising attribute field; another possible implementation is: the target adjustment parameters exist in the form of a bias vector, each task has its own exclusive bias term, and element-wise scaling or translation is performed on the basic content features.

[0090] In the embodiments of this specification, by separating general feature extraction and task-aware adaptation, the content tower can achieve multi-objective differentiated modeling with extremely low parameter increments while maintaining the sharing of the backbone structure, effectively alleviating gradient conflicts and improving the training stability of sparse targets.

[0091] For example, in the coarse-ranking service of a search advertising platform, the content information of a batch of candidate ads is sent to the content tower for processing. Taking one "smartwatch" ad as an example, its content information includes the product category "wearable devices," the brand "W brand," and the advertiser's optimization goal "place an order." The content tower first maps the original features to basic content features through a three-layer MLP (512→256→128). Subsequently, the feature adaptation layer loads the corresponding low-rank adjustment matrix from the parameter library according to the task goal "place an order," multiplies it with the basic content features, and then superimposes it to generate the target content features. If the same ad is subsequently used for the "add to cart" task, the feature adaptation layer will load another set of parameters to generate different target content features. Both sets of features are derived from the same basic content features, but exhibit semantic shifts due to different task goals, thereby supporting task-customized matching and scoring. The entire process is completed near-line, ensuring that online inference only requires one inner product calculation, meeting the performance requirements of coarse-ranking.

[0092] Furthermore, through the feature adaptation layer, target adjustment parameters corresponding to the task objectives are determined based on the task objectives of the content recommendation task. This includes: obtaining attribute information of the content to be recommended, wherein the attribute information is used to characterize the task objectives; retrieving the task identifier corresponding to the attribute information from a preset task mapping table; and determining the target adjustment parameters based on the task identifier.

[0093] The attribute information of the content to be recommended is a metadata field embedded in the advertising material, used to identify its business intent or optimization direction, and is an explicit carrier of the task objective. For example, the attribute information may include the "conversion event type" field specified by the advertiser when placing the ad, with a value of "purchase", "add_to_cart" or "form_submit".

[0094] Task identifiers are standardized codes used within the system to uniquely associate a set of target adjustment parameters with a specific type of optimization behavior. For example, a task identifier could be "cvr_task_01" representing order conversion, "cvr_task_02" representing add-to-cart conversion, and so on.

[0095] The target adjustment parameters are a set of learnable parameters bound to a specific task identifier, used for task-specific fine-tuning of basic content features. For example, each set of target adjustment parameters can consist of low-rank matrices U and V, stored independently and retrieved by task identifier index.

[0096] In practical applications, the system first obtains the attribute information of the content to be recommended in the feature adaptation layer of the content tower. This information typically comes from ad configuration data or nearline feature pipelines. Then, the system queries a pre-defined task mapping table, which stores the mapping relationship between attribute information and task identifiers in key-value pairs. For example, "conversion event type = purchase" is mapped to the task identifier "cvr_task_01". Finally, based on the retrieved task identifier, the system loads the corresponding target adjustment parameters from the parameter storage unit and applies them to subsequent adjustments to the basic content features. The entire process can be completed synchronously during model forward propagation without additional communication overhead and supports dynamic expansion of new tasks—simply add an entry to the task mapping table and initialize the corresponding parameters.

[0097] For the step "Retrieve the task identifier corresponding to the attribute information from the preset task mapping table", one option is to have the task mapping table reside in the service process as an in-memory hash table, supporting millisecond-level queries; another option is to have the task mapping table loaded into the model input layer as a configuration file, complete the mapping parsing in the feature preprocessing stage, and then pass the task identifier as an auxiliary feature into the feature adaptation layer.

[0098] In the embodiments described in this specification, the task identifier retrieval mechanism driven by attribute information ensures that the target adjustment parameters are strictly aligned with the current content recommendation task, avoids parameter interference in multi-task training, and improves the model's accuracy in perceiving task semantics.

[0099] For example, in the coarse-ranking system of a search advertising platform, the content information of an ad to be recommended includes the attribute field "optimization_event:form_submit", indicating that the advertiser wants to optimize form submission conversion. After reading this attribute information in the feature adaptation layer, the system queries a preset task mapping table, which records the mapping relationship "form_submit→task_id_cv3", thereby obtaining the task identifier "task_id_cv3". Subsequently, the system loads the target adjustment parameters (such as low-rank matrices U3 and V3) bound to "task_id_cv3" from the parameter pool and uses these parameters to adjust the basic content features. If another ad's attribute is "optimization_event:purchase", it is mapped to "task_id_cv1", and the U1 and V1 parameters are loaded. Through this attribute-based dynamic routing mechanism, the system can generate adapted content representations for dozens of different tasks within the same content stack without maintaining multiple independent models, significantly improving resource utilization efficiency and task expansion flexibility.

[0100] Furthermore, after determining the target adjustment parameters corresponding to the task objectives of the content recommendation task through the feature adaptation layer, the process also includes: generating a gating coefficient corresponding to the content information, wherein the gating coefficient is used to characterize the degree of influence of different task objectives on the content features; and multiplying the gating coefficient with the target adjustment parameters to obtain the adjusted target adjustment parameters.

[0101] The gating coefficient is a weighting factor in scalar or vector form, used to quantify the dependence of the current content on the target adjustment parameters under a specific task. For example, for advertisements for high-value goods, the system may generate a higher gating coefficient to enhance the task adjustment effect; for general advertisements, a lower coefficient is generated to preserve the dominance of basic features.

[0102] The adjusted target parameters are the final adaptation parameters modulated by the gating coefficient, used for subsequent weighted adjustments to the basic content features. For example, multiplying the original adjustment parameters by a gating coefficient of 0.8 reduces the perturbation amplitude of the content features.

[0103] In practical applications, after determining the target adjustment parameters, the system further generates corresponding gating coefficients based on content information. This process is typically implemented by a lightweight gated network (gateNN), which receives content information as input, passes through one or two fully connected layers and an activation function (such as sigmoid), and outputs gating coefficients between 0 and 1. Subsequently, the system multiplies these gating coefficients element-wise with the target adjustment parameters (if they are vectors) or scales them overall (if they are scalars) to obtain the adjusted target adjustment parameters. The gating network can share some low-level features with the feature adaptation layer or be designed independently. Its parameters are jointly optimized with the main model during training to ensure that the gating signal is semantically consistent with the task target.

[0104] For the step "generating gating coefficients corresponding to content information", one option is to use the gating network only as input features of the middle layer of the content tower to avoid introducing additional computational overhead. Another option is to integrate the gating coefficients with user-side context (such as user historical preferences) and content information to form a cross-tower gating signal, but this needs to be implemented within the communication range allowed by the dual-tower structure, for example, by indirectly conveying user intent through task identifiers.

[0105] In the embodiments described in this specification, the target adjustment parameters are scaled individually by using gating coefficients, so that the feature adaptation strength matches the content characteristics and task requirements, avoiding the expression bias caused by uniform adjustment and enhancing the model's ability to perceive the importance of the task.

[0106] For example, in a search advertising coarse-grained ranking system, an ad for "high-end laptops" contains information such as high price, a top international brand, and a category of "complete computer systems." The system has already determined the corresponding target adjustment parameters based on its "optimization_event:purchase" attribute in the feature adaptation layer. Subsequently, the gating network, based on this content information, determines that the ad belongs to a high-decision-cost product with a long user conversion path, thus generating a gating coefficient of 0.92, indicating a strong application of task adjustment. In contrast, another ad for "phone cases," due to its low unit price and fast decision-making, has a gating network output coefficient of 0.65, weakening the adjustment effect. The system multiplies each gating coefficient by its corresponding target adjustment parameter to obtain the adjusted parameters, which are then used to generate the final target content features. This dynamic adjustment mechanism allows the model to maintain structural consistency while achieving fine-grained adaptation to different product types and task objectives.

[0107] Furthermore, the basic content features are adjusted using target adjustment parameters to obtain target content features, including: obtaining the target deviation value corresponding to the task objective of the content recommendation task, wherein the target deviation value is a learnable parameter used to correct the content feature expression shift under different task objectives; and multiplying the target deviation value with the basic content features to obtain the target content features.

[0108] The target bias is a learnable parameter tied to a specific task objective. It is used to correct the scale or direction shift of content feature representation under different tasks, and its form is usually a vector with the same dimension as the basic content features. For example, the target bias for the "place an order" task may assign greater weight to the dimension related to high-value goods.

[0109] In practical applications, after obtaining the target adjustment parameters, the system further acquires the target deviation value corresponding to the task objective of the current content recommendation task. This deviation value is assigned according to the task identifier during the model initialization phase and updated through backpropagation during training. Subsequently, the system performs element-wise multiplication (Hadamard product) of the target deviation value with the basic content features to generate the target content features. This multiplicative operation preserves the structural information of the basic features while dynamically adjusting the importance of each dimension through task-specific biases, preventing additive biases from being overwhelmed by the backbone features in high-dimensional space. The target deviation value is usually stored in the form of an embedding table, read by the task identifier index, and its dimensions are strictly aligned with the basic content features to ensure computational compatibility.

[0110] For the step "multiplying the target deviation value with the basic content features", one option is to normalize the target deviation value with Sigmoid or Softmax before multiplying it to constrain its value range and prevent feature explosion; another option is to superimpose the target deviation value with the output of the aforementioned target adjustment parameter and then apply it to the basic content features to form an additive-multiplicative hybrid correction mechanism. However, the embodiment in this specification adopts a pure multiplicative approach to simplify the structure and enhance the stability of gradient propagation.

[0111] In the embodiments described in this specification, by introducing task-specific target bias values ​​to perform multiplicative correction on basic content features, the gradient vanishing problem in multi-objective training is effectively alleviated, especially improving the convergence speed and prediction stability of sparse tasks.

[0112] For example, in the coarse-ranking service of a search advertising platform, the basic content features of an "online education course" advertisement are a 128-dimensional vector, reflecting common information such as course category, instructor rating, and historical completion rate. When this advertisement is used for the "lead generation conversion" task, the system loads the corresponding target bias value based on the task identifier. This bias value has a high value in dimensions related to "long user decision-making cycle" and "high average order value." The system multiplies this bias value element-wise with the basic content features, amplifying the final target content features in key dimensions to better align with the semantic needs of lead generation behavior. If the same advertisement is subsequently used for the "click-through rate" task, another set of bias values ​​is loaded, with higher weights in dimensions such as "title attractiveness" and "cover image quality," thereby generating target content features adapted for click prediction. This multiplicative bias mechanism enables the model to achieve task-aware feature reweighting without increasing network depth, significantly improving the consistency and accuracy of multi-objective prediction.

[0113] Furthermore, after obtaining the recommended content for the target users, the process also includes: obtaining the business metric weights corresponding to the recommended content; adjusting the ranking score of the recommended content based on the business metric weights; and re-determining the recommendation order of the recommended content based on the adjusted ranking score.

[0114] The recommended content consists of a set of candidate ads scored by the dual-tower model, with each ad receiving an initial ranking score. For example, the recommended content might include several e-commerce ads, whose initial scores are calculated by the inner product of user features and target content features.

[0115] Business metric weights are numerical coefficients that reflect the importance of different task objectives in the current business context, and are determined by operational strategies, advertiser bids, or platform revenue models. For example, for high-priced items, the business metric weight for the "order conversion" task might be set to 1.5, while the weight for the "click" task might be set to 0.3.

[0116] The ranking score is the raw matching score generated by the model for the recommended content, used for initial ranking. For example, the ranking score can be the estimated conversion rate (PCVR) or a normalized similarity score.

[0117] The adjusted ranking score is the final score after incorporating the weights of business metrics, and is used to determine the final recommendation order. For example, a task with an original PCVR of 0.02 will have an adjusted score of 0.03 if it corresponds to a weight of 1.5.

[0118] In practical applications, after obtaining the recommended content and its original ranking score for the target user, the system first acquires the business metric weight associated with each recommended content. This weight is typically determined based on the content's task objective, advertiser configuration, or real-time bidding information. Subsequently, the system multiplies the original ranking score by the corresponding business metric weight (or combines them according to a preset fusion formula) to obtain the adjusted ranking score. Finally, the system re-ranks the recommended content based on the adjusted ranking scores and outputs the final recommendation list. Weight information can come from the configuration center, advertiser metadata, or online strategy services, and supports dynamic updates to respond to business changes.

[0119] For the step "adjust the ranking score of recommended content according to the business indicator weight", one possible approach is to use linear weighting, i.e., the adjusted score = the original score × the business indicator weight; another possible approach is to use non-linear fusion, such as the adjusted score = log(1 + original score) × weight + bias term, to alleviate the excessive amplification effect of high-scoring content. However, the embodiment in this specification uses linear multiplication to maintain the simplicity and interpretability of the calculation.

[0120] In the embodiments described in this specification, the original scores of the model are post-processed and adjusted by weighting business indicators, so that the recommendation results are more in line with the actual business value orientation while maintaining the model's predictive ability, thereby improving the overall revenue efficiency of the system.

[0121] For example, in the coarse-ranking output stage of a search advertising platform, the system generates 10 candidate ads and their original ranking scores. One ad, for "high-end smartphones," has an original PCVR of 0.018, and its task objective is "order conversion." Based on the current promotional strategy, the platform assigns a business metric weight of 2.0 to the "order" task, 1.2 to the "add to cart" task, and 0.5 to the "click" task. The system multiplies the ad's original score of 0.018 by 2.0, resulting in an adjusted score of 0.036. Another ad, for "phone cases," has an original PCVR of 0.025, but its task objective is "click," with a weight of 0.5, resulting in an adjusted score of 0.0125. Finally, the system re-ranks the ads according to the adjusted scores, giving higher-value conversion ads higher exposure priority. The entire adjustment process is completed within milliseconds, without requiring modifications to the model structure, allowing for flexible adaptation to different business cycle optimization priorities.

[0122] Furthermore, the dual-tower content recommendation model is trained through the following steps: in response to the sample recommendation task, sample user information and sample content information are obtained; the sample user information is input into the user tower of the initial dual-tower recommendation model to obtain sample user features, wherein the initial dual-tower recommendation model includes a user tower and a content tower, and the content tower includes a feature adaptation layer; based on the sample content information, the sample content features corresponding to the task objective of the sample recommendation task are obtained through the feature adaptation layer of the content tower; based on the sample user features and sample content features, the feature adaptation layer is trained to obtain the dual-tower recommendation model.

[0123] The sample recommendation task consists of labeled data units used for model training, which explicitly specify the type of conversion behavior to be optimized. For example, the sample recommendation task may correspond to specific goals such as "add to cart", "place order", or "submit form".

[0124] Sample user information is a dataset describing user-side features in the training samples, including historical behavior sequences, demographic attributes, and contextual information. For example, sample user information may include the ad categories recently clicked by the user, device type, and geographical location.

[0125] The sample content information is a dataset describing the attributes of advertising materials in the training samples, including product categories, brands, creative features, and task-related metadata. For example, the sample content information may include the advertiser's optimization goals, industry category, landing page quality score, etc.

[0126] The initial dual-tower recommendation model is the model structure before training, containing a user tower, a content tower, and a feature adaptation layer within the content tower. The backbone parameters of the user tower and the content tower are shared across multiple tasks, with only the feature adaptation layer supporting task differentiation. For example, the user tower consists of a three-layer MLP, and the content tower backbone is also a three-layer MLP, ending with a feature adaptation layer.

[0127] Sample user features are fixed-dimensional vectors output by User Tower after processing sample user information, used to represent user interests. For example, sample user features can be 128-dimensional embedding vectors.

[0128] The sample content features are task-specific content representations generated after the content pyramid is processed by the feature adaptation layer. They are obtained by correcting the basic content features with target adjustment parameters. For example, the sample content features generated for the same advertisement under the "place an order" task and the features generated under the "click" task exhibit different distributions in the semantic space.

[0129] In practical applications, the system first responds to a sample recommendation task by obtaining the corresponding sample user information and sample content information from the training dataset. Then, the sample user information is input into the user tower of the initial dual-tower recommendation model, and after processing by a multilayer perceptron, the sample user features are output. Simultaneously, the sample content information is input into the content tower. First, a shared feature extraction layer generates basic content features, and then the feature adaptation layer activates the corresponding target adjustment parameters and target bias values ​​based on the task objective of the current sample recommendation task (e.g., identification through advertising attribute fields), ultimately outputting the sample content features. Finally, the system calculates a predicted score (e.g., inner product) based on the sample user features and sample content features, and calculates a loss function (e.g., cross-entropy) with the sample's true label (e.g., whether it has converted). Backpropagation is used to update only the learnable parameters in the feature adaptation layer (including target adjustment parameters, target bias values, and gating network parameters). The backbone parameters of the user tower and content tower can be frozen or fine-tuned with a very low learning rate during the initial training phase, ensuring that training focuses on improving task adaptation capabilities.

[0130] For the step "training the feature adaptation layer based on sample user features and sample content features", one possible approach is to adopt a multi-task joint training strategy, with all task samples mixed as input, routed to the corresponding parameter branch through task identifier, and the shared optimizer updated uniformly; another possible approach is to perform alternating training by task grouping, with each iteration sampling only a subset of samples from a single task to reduce gradient interference. However, the embodiment in this specification adopts a mixed training method to improve sample utilization efficiency and convergence speed.

[0131] In the embodiments of this specification, by training only the feature adaptation layer and freezing the backbone network, the number of parameters and computational cost of multi-objective training are significantly reduced. At the same time, the destruction of shared features by gradient conflicts between tasks is avoided, so that the model can still effectively learn the differentiated representations of each task under limited resources.

[0132] For example, in the model training process of a search advertising platform, the system loads a batch of training data containing 1 million samples, each sample labeled with a specific task objective (e.g., 30% for "Place an order," 50% for "Add to cart," and 20% for "Leave contact information"). For a "Home Appliance - Refrigerator" ad sample, its sample content information includes "optimization_event:purchase," and the system activates the low-rank parameters and bias vector corresponding to the "Place an order" task in the feature adaptation layer accordingly. The sample user information includes the user's historical clicks on the "Large Appliances" category, which is processed by the user tower to output 128-dimensional user features. The content tower outputs the corresponding task-specific sample content features. The system calculates the inner product of the two as the predicted PCVR and calculates the loss with the actual conversion label, only updating the parameters of the feature adaptation layer. After multiple iterations, the feature adaptation layer learns to generate adapted content representations for different tasks, while the user tower and content tower backbone remain stable, ultimately obtaining a deployable dual-tower recommendation model. The entire training process does not require maintaining multiple independent models, significantly saving GPU memory and training time.

[0133] Furthermore, based on the sample user features and sample content features, a feature adaptation layer is trained, including: calculating the predicted score based on the sample user features and sample content features; calculating the loss value based on the predicted score and the reference label corresponding to the sample recommendation task; and updating the parameters in the feature adaptation layer based on the loss value.

[0134] The reference label is the actual user behavior result that occurred in the sample recommendation task, used for supervised training. For example, if the user eventually completes the order, the reference label is 1; otherwise, it is 0.

[0135] The loss value is a scalar metric that measures the difference between the predicted score and the reference label, and is used to guide parameter updates. For example, the loss value can be calculated using the binary cross-entropy loss function.

[0136] The parameters in the feature adaptation layer include task-specific learnable variables such as target adjustment parameters, target bias values, and gating network weights. Their updates do not affect the backbone structure of the user tower and content tower. For example, each set of parameters is stored independently by task identifier and is activated only when training on the corresponding task samples.

[0137] In practical applications, the system first calculates the predicted score based on the user and content features of the samples, typically using an inner product operation and mapping it to a probability interval via a sigmoid activation function. Then, it calculates the loss value based on this predicted score and the reference label corresponding to the sample recommendation task (e.g., whether a conversion occurred). Common loss functions include binary cross-entropy or Focal Loss to address imbalanced positive and negative samples. Finally, the system calculates the gradient based on this loss value using backpropagation, updating only the parameters in the feature adaptation layer, while the core parameters of the user and content towers remain frozen or fine-tuned with a very low learning rate during this stage. The entire training process supports mixed input of multi-task samples. The system automatically routes to the corresponding parameter branch through task identifiers, ensuring that the gradient only acts on relevant parameters and avoiding interference between tasks.

[0138] For the step "update the parameters in the feature adaptation layer based on the loss value", one possible approach is to use standard stochastic gradient descent (SGD) or Adam optimizer to sum the losses of all tasks and then update the parameters. Another possible approach is to assign different weights to the losses of different tasks (such as weighted amplification for sparse tasks) and then perform joint optimization to alleviate the problem of unbalanced data distribution. However, the embodiments in this specification default to using an unweighted joint training strategy, relying on the parameter isolation mechanism of the feature adaptation layer to naturally achieve task balance.

[0139] In the embodiments of this specification, the feature adaptation layer can independently learn the optimization path of each objective through the task-aware prediction-loss-update closed loop, effectively alleviating gradient conflicts in multiple tasks, while maintaining the stability of the backbone network, thereby improving the overall training efficiency and generalization ability of the model.

[0140] For example, in the training process of a search advertising platform, a sample contains user A, ad B, and the task objective "order conversion," with a reference label of 1 (user completes purchase). The system performs an inner product of the user features (128 dimensions) of user A and the content features of ad B under the "order" task, obtaining an original score of 0.85, which is then predicted to be 0.70 after Sigmoid conversion. The system calculates the binary cross-entropy loss between this predicted value and label 1, obtaining a loss value of 0.36. Subsequently, backpropagation only updates the target adjustment parameters, target bias value, and gating coefficient corresponding to the "order" task, while other task parameters remain unaffected. If the next sample task is "add to cart," another set of parameter branches is activated, and the loss is calculated and updated independently. Through this parameter isolation mechanism, even if the proportion of "leads collection" samples is less than 1%, its corresponding parameters can still converge stably, ultimately enabling the model to obtain reliable prediction capabilities on all tasks.

[0141] Furthermore, before inputting sample user information into the user tower of the initial dual-tower recommendation model to obtain sample user features, the method further includes: constructing the content tower of the initial dual-tower recommendation model, wherein the content tower includes at least one feature extraction layer for extracting semantic information of the content; identifying the target extraction layer located at the top of the feature extraction layer, wherein the target extraction layer is used to generate the final content representation; integrating a low-rank adaptation module into the target extraction layer to obtain a feature adaptation layer, wherein the low-rank adaptation module is used to perform lightweight adjustments to the content representation according to the task objective.

[0142] The feature extraction layer is a neural network layer in the content tower used to abstract semantic information of the content layer by layer, and it is usually composed of fully connected layers. For example, the content tower may contain three MLP layers stacked in sequence, which progressively compress and refine high-order features from the input dimensions of 512→256→128 respectively.

[0143] The target extraction layer is the topmost layer in the feature extraction layer, and its output serves as the basic content representation for subsequent task adaptation. For example, the third layer MLP (128-dimensional output) is the target extraction layer, and its output vector represents the general semantic embedding of the advertisement.

[0144] The low-rank adaptation module is a parameter-efficient fine-tuning structure that introduces a small number of learnable parameters through low-rank matrix factorization to perform task-specific adjustments to the output of the target extraction layer. For example, the low-rank adaptation module can consist of two low-rank matrices. and The product ΔW = UV approximates the original weight increment (r << 128).

[0145] The feature adaptation layer is an enhanced structure that integrates the low-rank adaptation module, used to generate target content features corresponding to the task objective. For example, the feature adaptation layer superimposes ΔW·h (where h is the basic content feature) onto the output of the target extraction layer to form a task-customized representation.

[0146] In practical applications, the system first constructs the content tower of an initial dual-tower recommendation model. This tower contains at least one feature extraction layer for extracting semantic information from the content, typically constructed using a multilayer perceptron. Subsequently, the target extraction layer at the top of these feature extraction layers is identified, as it directly outputs the final content representation and is the optimal intervention point for task adaptation. Finally, a low-rank adaptation module is integrated into this target extraction layer. Specifically, a low-rank transformation branch is added as a bypass to the output path of the target extraction layer. Its parameters are stored independently according to the task objective, and during forward propagation, the corresponding branch is activated based on the task identifier of the current sample. The low-rank transformation result is then added or multiplied with the original output to obtain the feature adaptation layer. The entire integration process does not modify the structure and parameters of the original target extraction layer, only adding a small number of trainable variables, ensuring the stability and ease of transfer of the backbone network.

[0147] For the step "integrating the low-rank adaptation module in the target extraction layer", one option is to add the low-rank adaptation module as a residual to the output of the target extraction layer, i.e., h_adapted = h + UV·h; another option is to apply the low-rank adaptation module to the weight matrix of the target extraction layer itself, i.e., W_total = W_original + UV. However, the former is used in the embodiments of this specification because it is simple to implement, stable in training, and facilitates task parameter isolation.

[0148] In the embodiments of this specification, a feature adaptation layer is constructed by integrating a low-rank adaptation module at the top layer of the content tower, which achieves multi-task differentiated modeling with extremely low parameter increments, avoids the waste of resources caused by full tower replication, and maintains the integrity of the backbone feature extraction capability.

[0149] For example, in the model initialization phase of a search advertising platform, the system constructs a content pyramid consisting of a three-layer MLP (512→256→128), where the third layer is the top-level target extraction layer, outputting 128-dimensional basic content features. Subsequently, the system integrates a low-rank adaptation module after this layer, setting the rank r=8, and assigning independent [features] to each task objective (such as "place an order", "add to cart", "leave contact information"). and Parameter pairs. When processing a sample "optimization_event:purchase", the system loads U1 and V1, calculates Δh = U1V1·h, and uses h_adapted = h + Δh as the final output. Since r = 8 is much smaller than 128, each task only adds about 2KB of parameters, and the total increase for 30 tasks is less than 60KB, while the original content tower parameters are several MB, significantly saving storage and communication overhead. This design allows the model to support dozens of tasks in the same content tower, and adding a new task only requires initializing new low-rank parameters, without retraining the backbone network.

[0150] This specification provides a content recommendation method in one embodiment. In response to a content recommendation task, it obtains content information of the content to be recommended and user information of the target user. The user information and content information are input into a dual-tower content recommendation model to obtain recommended content for the target user. The dual-tower content recommendation model includes a user tower and a content tower. The user tower is used to extract user features based on user information. The content tower includes a feature adaptation layer, which is used to obtain target content features corresponding to the task objective of the content recommendation task based on the content information. The recommended content is obtained by decoding the target content features and user features. The feature adaptation layer generates target content features corresponding to the task objective of the content recommendation task based on the content information, and combines this with user features to decode the recommended content. Because the feature adaptation layer can dynamically adjust the content representation according to the task objective, the same content can present differentiated feature expressions under different tasks, thereby improving the adaptation accuracy of the recommendation results to specific task objectives without changing the dual-tower structure.

[0151] See Figure 5 , Figure 5 This is a flowchart of a dual-tower content pushing model training method provided in one embodiment of this specification, specifically including the following steps 502-508.

[0152] Step 502: In response to the sample recommendation task, obtain sample user information and sample content information.

[0153] Step 504: Input the sample user information into the user tower of the initial dual-tower recommendation model to obtain the sample user features. The initial dual-tower recommendation model includes a user tower and a content tower, and the content tower includes a feature adaptation layer.

[0154] Step 506: Based on the sample content information, obtain the sample content features corresponding to the task objective of the sample recommendation task through the feature adaptation layer of the content tower.

[0155] Step 508: Based on the sample user features and sample content features, train the feature adaptation layer to obtain the dual-tower recommendation model.

[0156] It is understandable that steps 502-508 are the same as the steps for training the dual-tower content recommendation model mentioned above, and will not be repeated here.

[0157] This specification provides an embodiment of a dual-tower content recommendation model training method. In response to a sample recommendation task, it acquires sample user information and sample content information. The sample user information is input into the user tower of an initial dual-tower recommendation model to obtain sample user features. The initial dual-tower recommendation model includes a user tower and a content tower, with the content tower including a feature adaptation layer. Based on the sample content information, the feature adaptation layer of the content tower obtains the sample content features corresponding to the task objective of the sample recommendation task. Based on the sample user features and sample content features, the feature adaptation layer is trained to obtain the dual-tower recommendation model. The feature adaptation layer generates sample content features corresponding to the task objective of the sample recommendation task based on the sample content information, and performs joint training with the sample user features, enabling the feature adaptation layer to learn task-specific content representations. Because the same content generates differentiated sample content features under different tasks, it effectively alleviates multi-objective gradient conflicts, improves the model's prediction accuracy for each task objective, and maintains the dual-tower structure unchanged.

[0158] The following is in conjunction with the appendix Figure 6 Taking the application of the content recommendation method provided in this specification in the advertising recommendation system of a content publishing platform as an example, the content recommendation method will be further explained. Figure 6 This is a flowchart illustrating the processing steps of a content recommendation method provided in one embodiment of this specification, specifically including the following steps.

[0159] Step 602: Add a LoRa module to the last layer of the item tower, MLP. The parameters UV of the LoRa module are independent of each other across multiple tasks. The parameters corresponding to the current sample are selected through the advertising attributes and participate in the LoRa calculation. Finally, the two parts are added together.

[0160] Specifically, in the advertising recommendation system, the final layer of the item pyramid, the MLP (Multilayer Perceptron), is used to generate high-order feature representations of the ads. This step adds a LoRA (Low-Rank Adaptation) module after this layer. Its parameters consist of two low-rank matrices, U and V, and the UV parameters for different task objectives (such as click-through rate and conversion rate) are independent. During training, the system selects the corresponding task's UV parameters from a pre-set parameter set based on the ad attributes of the current sample (such as industry and conversion type), performs matrix multiplication with the original MLP output, and finally adds the LoRA output to the original MLP output to form a task-specific item embedding. This design achieves lightweight adaptation for multiple objectives through task-specific parameter fine-tuning.

[0161] See Figure 7 As shown, Figure 7This is a schematic diagram of a dual-tower content recommendation model provided in one embodiment of this specification, illustrating the dual-tower structure proposed in this scheme. The input "user features" are extracted by a three-layer "multilayer perceptron" (512→256→128). The input "content features" are first processed by a multilayer perceptron with the same structure, and then enter the "low-rank adaptation module (128)". This module receives the "optimization target and constraint type" signals, dynamically adjusts the expression of content features, and finally outputs "content features" and decodes them with user features to obtain "recommended content".

[0162] See Figure 8 As shown, Figure 8 This is a schematic diagram of a low-rank adaptation module in a dual-tower content recommendation model provided in one embodiment of this specification, illustrating the structural principle of the low-rank adaptation module. The left side illustrates "Y = W + U × V", where "W" is the original weight, and "U" and "V" are low-rank matrices used to generate incremental parameters; the right side shows that "U" is composed of "U1, U2, U3", and "V" is composed of "V1, V2, V3", and the parameters of "U" are determined by "optimization objective and constraint type", achieving lightweight adaptation to different tasks and reducing the number of parameters and computational overhead.

[0163] Step 604: The main network part includes a gateNN. The lora part is also multiplied by a gateNN. The output of the lora part is personalized according to the input, which improves the offline AUC.

[0164] Specifically, in the main network of the item tower, both the original MLP output and the LoRA output need to be dynamically adjusted through a gated neural network (gateNN). The gateNN takes user features (such as historical behavior) and ad features (such as attribute labels) as input and outputs a weight coefficient ranging from 0 to 1. This coefficient is multiplied by the original MLP output and the LoRA output respectively, controlling the proportion of their contribution to the final item membedding. For example, when a user has a strong preference for a certain type of ad, the gateNN will increase the weight of the LoRA part, making the features with higher task relevance dominate the output. This mechanism significantly improves the model's responsiveness to personalized user needs, and the offline AUC improvement validates its effectiveness.

[0165] Step 606: Increase the deviation corresponding to each objective to avoid the gradient vanishing problem. Multiply the deviation by the embedding of the final output to improve the offline AUC.

[0166] Specifically, to address the gradient conflict issue in multi-objective training, the system introduces a task bias parameter for each task objective (such as conversion rates across different industries). This bias is directly applied to the output layer of the final itemembedded layer in a learnable form, i.e., the bias value is superimposed on the embedding vector through matrix multiplication. For example, for the "e-commerce conversion" objective, the bias parameter amplifies the feature dimensions related to purchase intent; while for the "information reading" objective, it enhances the dimensions related to content relevance. This bias adjustment not only compensates for the differences in feature distribution between tasks but also alleviates the gradient vanishing problem of sparse objectives through explicit modeling, ultimately further improving the offline AUC.

[0167] Through steps 602-606 above, a multi-objective modeling scheme with high prediction accuracy was designed under a coarse-rank dual-tower architecture. The differences between different objectives are modeled using a low-rank matrix factorization structure. Compared to a single-head structure, this scheme does not increase inference resources during online inference, thus meeting the latency and resource constraints and requirements of actual business applications.

[0168] Corresponding to the above method embodiments, this specification also provides embodiments of a content recommendation device. Figure 9 This is a schematic diagram of the structure of a content recommendation device provided in one embodiment of this specification. Figure 9 As shown, the device includes:

[0169] The first acquisition module 902 is configured to, in response to a content recommendation task, acquire content information of the content to be recommended and user information of the target user.

[0170] The recommendation module 904 is configured to input user information and content information into a dual-tower content recommendation model to obtain recommended content for the target user. The dual-tower content recommendation model includes a user tower and a content tower. The user tower is used to extract user features based on user information. The content tower includes a feature adaptation layer, which is used to obtain the target content features corresponding to the task objective of the content recommendation task based on the content information. The recommended content is obtained by decoding the target content features and user features.

[0171] Optionally, the recommendation module 904 is further configured to input user information into the user tower of the dual-tower content recommendation model for feature extraction to obtain user features; and to input content information into the content tower of the dual-tower content recommendation model for feature extraction and feature adaptation to obtain target content features.

[0172] Optionally, the content tower includes at least one feature extraction layer and a feature adaptation layer located after the at least one feature extraction layer; correspondingly, the recommendation module 904 is further configured to process content information through the at least one feature extraction layer to generate basic content features; through the feature adaptation layer, according to the task objective of the content recommendation task, determine the target adjustment parameters corresponding to the task objective, and use the target adjustment parameters to adjust the basic content features to obtain target content features.

[0173] Optionally, the recommendation module 904 is further configured to obtain attribute information of the content to be recommended, wherein the attribute information is used to characterize the task objective; retrieve the task identifier corresponding to the attribute information from a preset task mapping table; and determine the target adjustment parameters based on the task identifier.

[0174] Optionally, the recommendation module 904 is further configured to generate a gating coefficient corresponding to the content information, wherein the gating coefficient is used to characterize the degree of influence of different task objectives on the content features; the gating coefficient is multiplied by the objective adjustment parameter to obtain the adjusted objective adjustment parameter.

[0175] Optionally, the recommendation module 904 is further configured to obtain a target deviation value corresponding to the task objective of the content recommendation task, wherein the target deviation value is a learnable parameter used to correct the content feature expression shift under different task objectives; and the target content feature is obtained by multiplying the target deviation value with the basic content features.

[0176] Optionally, the content recommendation device also includes a fine-ranking module, configured to obtain the business indicator weights corresponding to the recommended content; adjust the ranking score of the recommended content according to the business indicator weights; and redetermine the recommendation order of the recommended content based on the adjusted ranking score.

[0177] Optionally, the content recommendation device further includes a training module configured to, in response to a sample recommendation task, acquire sample user information and sample content information; input the sample user information into the user tower of an initial dual-tower recommendation model to obtain sample user features, wherein the initial dual-tower recommendation model includes a user tower and a content tower, and the content tower includes a feature adaptation layer; based on the sample content information, obtain the sample content features corresponding to the task objective of the sample recommendation task through the feature adaptation layer of the content tower; and train the feature adaptation layer based on the sample user features and sample content features to obtain the dual-tower recommendation model.

[0178] Optionally, the training module is further configured to calculate a predicted score based on the sample user features and sample content features; calculate a loss value based on the predicted score and the reference label corresponding to the sample recommendation task; and update the parameters in the feature adaptation layer based on the loss value.

[0179] Optionally, the training module is further configured to construct the content tower of the initial dual-tower recommendation model, wherein the content tower includes at least one feature extraction layer for extracting semantic information of the content; identify the target extraction layer at the top of the feature extraction layer, wherein the target extraction layer is used to generate the final content representation; and integrate a low-rank adaptation module in the target extraction layer to obtain a feature adaptation layer, wherein the low-rank adaptation module is used to perform lightweight adjustment of the content representation according to the task objective.

[0180] One embodiment of this specification provides a content recommendation device. A first acquisition module 902 responds to a content recommendation task, acquiring content information of the content to be recommended and user information of the target user. A recommendation module 904 then inputs both information into a dual-tower content recommendation model. The user tower extracts user features based on user information, while the feature adaptation layer in the content tower dynamically generates corresponding target content features based on the content information and the current task objective (such as conversion type)—the same content is mapped to differentiated feature expressions depending on the task objective. The recommendation module 904 further performs inner product or similarity decoding on the target content features and user features, outputting recommended content adapted to the specific task objective. Thus, by leveraging the dynamic adjustment capability of the feature adaptation layer, the device effectively improves the accuracy of recommendation results in adapting to diverse task objectives while maintaining the dual-tower structure.

[0181] The above is an illustrative scheme of a content recommendation device according to this embodiment. It should be noted that the technical solution of this content recommendation device and the technical solution of the content recommendation method described above belong to the same concept. For details not described in detail in the technical solution of the content recommendation device, please refer to the description of the technical solution of the content recommendation method described above.

[0182] Corresponding to the above method embodiments, this specification also provides embodiments of a dual-tower content recommendation model training device. Figure 10 This is a schematic diagram of a dual-tower content recommendation model training device provided in one embodiment of this specification. Figure 10 As shown, the device includes:

[0183] The second acquisition module 1002 is configured to acquire sample user information and sample content information in response to the sample recommendation task.

[0184] The input module 1004 is configured to input sample user information into the user tower of the initial dual-tower recommendation model to obtain sample user features. The initial dual-tower recommendation model includes a user tower and a content tower, and the content tower includes a feature adaptation layer.

[0185] The adaptation module 1006 is configured to obtain the sample content features corresponding to the task target of the sample recommendation task through the feature adaptation layer of the content tower based on the sample content information.

[0186] Training module 1008 is configured to train a feature adaptation layer based on sample user features and sample content features to obtain a dual-tower recommendation model.

[0187] One embodiment of this specification provides a training device for a dual-tower content recommendation model. A second acquisition module 1002 responds to a sample recommendation task, acquiring sample user information and sample content information. An input module 1004 inputs the sample user information into the user tower of the initial dual-tower recommendation model to obtain sample user features. An adaptation module 1006, based on the sample content information, uses a feature adaptation layer in the content tower to generate sample content features corresponding to the task objective of the sample recommendation task. A training module 1008 then jointly trains the feature adaptation layer based on the sample user features and sample content features. Thus, the same content can output differentiated sample content features through the feature adaptation layer under different task objectives, enabling the model to learn task-specific content representations. This effectively alleviates multi-objective gradient conflicts and improves the prediction accuracy of each task objective while maintaining the dual-tower structure.

[0188] The above is a schematic scheme of a dual-tower content recommendation model training device according to this embodiment. It should be noted that the technical solution of this dual-tower content recommendation model training device and the technical solution of the aforementioned dual-tower content recommendation model training method belong to the same concept. Details not described in detail in the technical solution of the dual-tower content recommendation model training device can be found in the description of the technical solution of the aforementioned dual-tower content recommendation model training method.

[0189] Figure 11 This is a structural block diagram of a computing device according to one embodiment of this specification. The components of the computing device 1100 include, but are not limited to, a memory 1110 and a processor 1120. The processor 1120 is connected to the memory 1110 via a bus 1130, and a database 1150 is used to store data.

[0190] The computing device 1100 also includes an access device 1140, which enables the computing device 1100 to communicate via one or more networks 1160. Examples of these networks include Public Switched Telephone Network (PSTN), Local Area Network (LAN), Wide Area Network (WAN), Personal Area Network (PAN), or combinations of communication networks such as the Internet. The access device 1140 may include one or more of any type of wired or wireless network interface (e.g., a network interface controller (NIC)), such as an IEEE 802.11 Wireless Local Area Network (WLAN) wireless interface, a Wi-MAX (Worldwide Interoperability for Microwave Access) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, or a Near Field Communication (NFC) interface.

[0191] In one embodiment of this specification, the aforementioned components of the computing device 1100 and Figure 11 Other components, not shown, can also be connected to each other, for example, via a bus. It should be understood that... Figure 11 The block diagram of the computing device shown is for illustrative purposes only and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.

[0192] The computing device 1100 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or personal computers (PCs). The computing device 1100 can also be a mobile or stationary server.

[0193] The processor 1120 is used to execute the following computer program / instructions, which, when executed by the processor, implement the steps of the above-mentioned content recommendation method or dual-tower content recommendation model training method.

[0194] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. In particular, the computing device embodiments are basically similar to the content recommendation method or dual-tower content recommendation model training method embodiments, so the description is relatively simple. Relevant parts can be referred to the descriptions of the content recommendation method or dual-tower content recommendation model training method embodiments.

[0195] An embodiment of this specification also provides a computer-readable storage medium storing a computer program / instructions that, when executed by a processor, implement the steps of the above-described content recommendation method or the dual-tower content recommendation model training method.

[0196] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. In particular, the computer-readable storage medium embodiments are relatively simple in description because they are fundamentally similar to the content recommendation method or the dual-tower content recommendation model training method embodiments. Relevant details can be found in the descriptions of the content recommendation method or the dual-tower content recommendation model training method embodiments.

[0197] An embodiment of this specification also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the above-described content recommendation method or dual-tower content recommendation model training method.

[0198] The above is an illustrative scheme of a computer program product according to this embodiment. It should be noted that the technical solution of this computer program product belongs to the same concept as the technical solution of the aforementioned content recommendation method or dual-tower content recommendation model training method. Details not described in detail in the technical solution of the computer program product can be found in the description of the technical solution of the aforementioned content recommendation method or dual-tower content recommendation model training method.

[0199] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0200] The computer instructions include computer program code, which may be in the form of source code, object code, executable file, or certain intermediate forms. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium may be appropriately added or removed according to the requirements of patent practice. For example, in some regions, according to patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0201] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.

[0202] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0203] The preferred embodiments disclosed above are merely illustrative of this specification. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the embodiments described herein. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the embodiments, thereby enabling those skilled in the art to better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.

Claims

1. A content recommendation method, characterized in that, include: In response to content recommendation tasks, obtain content information of the content to be recommended and user information of the target users; The user information and the content information are input into a dual-tower content recommendation model to obtain recommended content for the target user. The dual-tower content recommendation model includes a user tower and a content tower. The user tower is used to extract user features based on the user information. The content tower includes a feature adaptation layer, which is used to obtain target content features corresponding to the task objective of the content recommendation task based on the content information. The recommended content is obtained by decoding the target content features and the user features.

2. The method according to claim 1, characterized in that, The step of inputting user information and content information into the dual-tower content recommendation model to obtain recommended content for the target user includes: The user information is input into the user tower of the dual-tower content recommendation model for feature extraction to obtain user features; The content information is input into the content tower of the dual-tower content recommendation model for feature extraction and feature adaptation to obtain the target content features.

3. The method according to claim 2, characterized in that, The content tower includes at least one feature extraction layer and a feature adaptation layer located after the at least one feature extraction layer; the process of inputting the content information into the content tower of the dual-tower content recommendation model for feature extraction and feature adaptation to obtain target content features includes: The content information is processed through the at least one feature extraction layer to generate basic content features; Through the feature adaptation layer, based on the task objective of the content recommendation task, target adjustment parameters corresponding to the task objective are determined, and the basic content features are adjusted using the target adjustment parameters to obtain the target content features.

4. The method according to claim 3, characterized in that, The step of determining target adjustment parameters corresponding to the task target based on the task target of the content recommendation task through the feature adaptation layer includes: Obtain the attribute information of the content to be recommended, wherein the attribute information is used to characterize the task objective; Retrieve the task identifier corresponding to the attribute information from the preset task mapping table; The target adjustment parameters are determined based on the task identifier.

5. The method according to claim 3, characterized in that, After determining the target adjustment parameters corresponding to the task target based on the content recommendation task through the feature adaptation layer, the method further includes: Generate a gating coefficient corresponding to the content information, wherein the gating coefficient is used to characterize the degree of influence of different task objectives on the content features; The adjusted target adjustment parameter is obtained by multiplying the gating coefficient by the target adjustment parameter.

6. The method according to claim 3, characterized in that, The step of adjusting the basic content features using the target adjustment parameters to obtain the target content features includes: Obtain the target deviation value corresponding to the task objective of the content recommendation task, wherein the target deviation value is a learnable parameter used to correct the content feature expression shift under different task objectives; The target content feature is obtained by multiplying the target deviation value by the basic content feature.

7. The method according to claim 1, characterized in that, After obtaining the recommended content for the target user, the method further includes: Obtain the business metric weights corresponding to the recommended content; The ranking score of the recommended content is adjusted according to the weight of the business metrics. The recommendation order of the recommended content is re-determined based on the adjusted ranking scores.

8. The method according to claim 1, characterized in that, The dual-tower content recommendation model was trained through the following steps: In response to the sample recommendation task, obtain sample user information and sample content information; The sample user information is input into the user tower of the initial dual-tower recommendation model to obtain sample user features. The initial dual-tower recommendation model includes a user tower and a content tower, and the content tower includes a feature adaptation layer. Based on the sample content information, the sample content features corresponding to the task objective of the sample recommendation task are obtained through the feature adaptation layer of the content tower. Based on the sample user features and the sample content features, the feature adaptation layer is trained to obtain the dual-tower recommendation model.

9. The method according to claim 8, characterized in that, The step of training the feature adaptation layer based on sample user features and sample content features includes: Calculate the predicted score based on the sample user characteristics and the sample content characteristics; The loss value is calculated based on the predicted score and the reference label corresponding to the sample recommendation task; The parameters in the feature adaptation layer are updated based on the loss value.

10. The method according to claim 8, characterized in that, Before inputting the sample user information into the user tower of the initial dual-tower recommendation model to obtain the sample user features, the method further includes: Construct the content tower of the initial dual-tower recommendation model, wherein the content tower includes at least one feature extraction layer for extracting semantic information of the content; Identify the target extraction layer located at the top of the feature extraction layer, wherein the target extraction layer is used to generate the final content representation; A low-rank adaptation module is integrated into the target extraction layer to obtain a feature adaptation layer, wherein the low-rank adaptation module is used to perform lightweight adjustment of the content representation according to the task objective.

11. A training method for a dual-tower content recommendation model, characterized in that, include: In response to the sample recommendation task, obtain sample user information and sample content information; The sample user information is input into the user tower of the initial dual-tower recommendation model to obtain sample user features. The initial dual-tower recommendation model includes a user tower and a content tower, and the content tower includes a feature adaptation layer. Based on the sample content information, the sample content features corresponding to the task objective of the sample recommendation task are obtained through the feature adaptation layer of the content tower. Based on the sample user features and the sample content features, the feature adaptation layer is trained to obtain the dual-tower recommendation model.

12. A content recommendation device, characterized in that, include: The first acquisition module is configured to respond to the content recommendation task by acquiring the content information of the content to be recommended and the user information of the target user. The recommendation module is configured to input the user information and the content information into a dual-tower content recommendation model to obtain recommended content for the target user. The dual-tower content recommendation model includes a user tower and a content tower. The user tower is used to extract user features based on the user information. The content tower includes a feature adaptation layer, which is used to obtain target content features corresponding to the task objective of the content recommendation task based on the content information. The recommended content is obtained by decoding the target content features and the user features.

13. A dual-tower content recommendation model training device, characterized in that, include: The second acquisition module is configured to acquire sample user information and sample content information in response to the sample recommendation task. The input module is configured to input the sample user information into the user tower of the initial dual-tower recommendation model to obtain sample user features, wherein the initial dual-tower recommendation model includes a user tower and a content tower, and the content tower includes a feature adaptation layer; The adaptation module is configured to obtain the sample content features corresponding to the task target of the sample recommendation task through the feature adaptation layer of the content tower based on the sample content information. The training module is configured to train the feature adaptation layer based on the sample user features and the sample content features to obtain a dual-tower recommendation model.

14. A computing device, characterized in that, include: Memory and processor; The memory is used to store computer programs / instructions, and the processor is used to execute the computer programs / instructions. When the computer programs / instructions are executed by the processor, they implement the steps of the content recommendation method according to any one of claims 1-10 or the dual-tower content recommendation model training method according to claim 11.

15. A computer-readable storage medium, characterized in that, It stores a computer program / instruction that, when executed by a processor, implements the steps of the content recommendation method according to any one of claims 1-10 or the dual-tower content recommendation model training method according to claim 11.

16. A computer program product, characterized in that, Includes a computer program / instruction that, when executed by a processor, implements the steps of the content recommendation method according to any one of claims 1-10 or the dual-tower content recommendation model training method according to claim 11.

Images