A content recommendation method and related device

Abstract

This application discloses a content recommendation method and related equipment. The related equipment may include a content recommendation device, an electronic device, a computer program product, and a computer-readable storage medium. This application, after acquiring a content data stream of at least one business scenario and identifying at least one content interaction information in the content data stream, generates at least one content sample of the business scenario based on the content interaction information. It then extracts at least one scene feature data corresponding to the content sample from the content data stream. Next, based on the scene feature data, it trains a preset feature extraction model to obtain at least one object feature. The object feature is written into a preset object feature set to update the preset object feature set. Then, when a content recommendation request for a target object is received, content recommendation is performed on the target object based on the updated object feature set. This scheme can improve the efficiency and accuracy of content recommendation.

Description

Technical Field

[0001] This invention relates to the field of content recommendation technology, specifically to a content recommendation method and related equipment, which may include a content recommendation device, electronic equipment, and computer-readable storage medium. Background Technology

[0002] With the rapid development of internet technology, a massive amount of content needs to be recommended. In the content recommendation process, auxiliary information can be introduced, and the features of this auxiliary information can be used to enrich the feature representations of interactive objects and content objects, thereby improving the accuracy of content recommendation. Currently, in the content recommendation process, these auxiliary features are often extracted during the content recommendation process (i.e., the reasoning / computation process), and then used to perform content recommendations.

[0003] In the process of researching and practicing current technologies, the inventors of this application have found that because a large amount of feature data needs to be processed during the reasoning / computation process, the extraction of additional auxiliary features often consumes more computing resources and storage space. In addition, if there is insufficient auxiliary feature data during the recommendation process, content recommendation will often face the problems of data sparsity and cold start, thus resulting in low efficiency and accuracy of content recommendation. Summary of the Invention

[0004] This application provides a content recommendation method and related equipment. The related equipment may include a content recommendation device, an electronic device, and a computer-readable storage medium, which can improve the efficiency and accuracy of content recommendation.

[0005] A content recommendation method includes:

[0006] Acquire content data streams for at least one business scenario, and identify at least one content interaction information in the content data streams;

[0007] Based on the content interaction information, at least one content sample in the business scenario is generated, and at least one scenario feature data corresponding to the content sample is extracted from the content data stream.

[0008] Based on the scene feature data, a preset feature extraction model is trained to obtain at least one object feature;

[0009] The object features are written into a preset object feature set to update the preset object feature set;

[0010] When a content recommendation request for a target object is received, content recommendation is performed on the target object based on the updated set of object features.

[0011] Accordingly, embodiments of this application provide a content recommendation device, including:

[0012] The acquisition unit is used to acquire at least one content data stream in a business scenario and identify at least one content interaction information in the content data stream.

[0013] The generation unit is used to generate at least one content sample under the business scenario based on the content interaction information, and to extract at least one scenario feature data corresponding to the content sample from the content data stream.

[0014] The training unit is used to train a preset feature extraction model based on the scene feature data to obtain at least one object feature.

[0015] A writing unit is used to write the object features into a preset object feature set to update the preset object feature set;

[0016] The recommendation unit is used to recommend content to the target object based on the updated set of object features when a content recommendation request for the target object is received.

[0017] In some embodiments, the training unit may be specifically used to extract object feature data of at least one interactive object that interacts with the content sample from the scene feature data; extract at least one content feature data of the content sample from the scene feature data; and train the preset feature extraction model based on the object feature data and the content feature data to obtain at least one object feature.

[0018] In some embodiments, the training unit may be specifically used to extract features from the content feature data to obtain at least one content object feature; to extract features from the object feature data to obtain at least one interactive object feature, and to use the interactive object feature as the object feature; to perform offline training on the preset feature extraction model based on the content object feature and the interactive object feature, and to use the trained feature extraction model as the preset feature extraction model.

[0019] In some embodiments, the training unit may specifically be used to extract features from the content feature data to obtain at least one content object feature; extract features from the object feature data to obtain at least one interactive object feature, and use the content object feature and the interactive object feature as the object feature; identify the context information of the content sample in the content data stream, and extract context features from the context information; train the preset feature extraction model based on the context features, content object features, and interactive object features, and use the trained feature extraction model as the preset feature extraction model.

[0020] In some embodiments, the writing unit may be specifically used to obtain the storage space type of the preset object feature set; when the storage space type is a feature database, the object features are written into the preset object feature set through an intermediate database to update the preset object feature set; when the storage space type is a distributed storage space, the object features are written into the preset object feature set through a data writing component to update the preset object feature set.

[0021] In some embodiments, the writing unit may be specifically used to perform deduplication processing on the object features to obtain deduplicated object features; write the deduplicated object features into the intermediate database through a cache queue; and distribute the deduplicated object features to the preset object feature set through the intermediate database to update the preset object feature set.

[0022] In some embodiments, the writing unit may be specifically used to divide the object features into at least one batch of object feature sets according to the generation time of the object features; to select a preset number of consecutive batches of object feature sets from the object feature sets to obtain a set of object features to be deduplicated; and to perform deduplication processing on the object features in the set of object features to be deduplicated to obtain deduplicated object features.

[0023] In some embodiments, the writing unit may be specifically used to obtain the thread state of asynchronous threads in the cache queue; based on the thread state, select at least one target asynchronous thread from the asynchronous threads; and asynchronously write the object features into the intermediate database through the target asynchronous thread.

[0024] In some embodiments, the writing unit may be specifically used to determine at least one current queue parameter of the cache based on the thread state; when the current queue parameter exceeds a preset parameter threshold, an alarm message is generated, and the alarm message is used to indicate that there is an anomaly in the cache queue.

[0025] In some embodiments, the content recommendation device may further include an online training unit, specifically configured to: when the storage space type of the preset object feature set is a feature database, acquire the read / write logs of the preset object feature set, and train at least one preset content recommendation model based on the read / write logs to obtain a trained content recommendation model; when the storage space type of the preset object feature set is a distributed storage space, train at least one preset content recommendation model based on the preset object feature set to obtain a trained content recommendation model.

[0026] In some embodiments, the online training unit is specifically used to obtain the object identifier of the sample object corresponding to at least one preset content recommendation model; based on the object identifier and read / write logs, filter out the object features corresponding to the sample object from the updated object feature set to obtain at least one sample object feature; and train the preset content model based on the sample object feature to obtain the trained content recommendation model.

[0027] In some embodiments, the online training unit can be specifically used to obtain a model training request, the model training request carrying a feature identifier of at least one object feature required by the current model training task; when a general object feature corresponding to the feature identifier exists in the preset object feature set, the general object feature is written into the feature cache space; and at least one preset content recommendation model is trained using the object features in the feature cache space to obtain the trained content recommendation model.

[0028] In some embodiments, the online training unit may be specifically used to obtain the write time of object features in the feature cache space; based on the write time, to filter out at least one object feature in the feature cache space to obtain object features to be eliminated; and to delete the object features to be eliminated in the feature cache space.

[0029] In some embodiments, the recommendation unit may be specifically used to filter out at least one target interaction object feature corresponding to the target object from the updated object feature set; obtain at least one content to be recommended corresponding to the content recommendation request, and extract content features from the content to be recommended; and recommend the content to be recommended to the target object based on the content features and the target interaction object features using the content recommendation model.

[0030] In some embodiments, the recommendation unit may specifically be used to obtain the update time of the target interactive object features in the updated object feature set; when the update time does not exceed a preset time threshold, the content to be recommended is recommended to the target object based on the content features and the target interactive object features using the content recommendation model; when the update time exceeds the preset time threshold, the target interactive object features are updated, and the content to be recommended is recommended to the target object based on the updated interactive object features and the content features using the content recommendation model.

[0031] In some embodiments, the recommendation unit may be specifically used to extract at least one target feature sub-graph from the features of the target interactive object; based on the target feature sub-graph and content features, use the content recommendation model to predict the recommendation parameters of the content to be recommended; according to the recommendation parameters, select at least one target content from the content to be recommended, and recommend the target content to the target object.

[0032] In some embodiments, the recommendation unit may specifically be used to identify target data at the current moment in the content data stream, and filter out the current object feature data of the target interactive object corresponding to the target interactive object feature from the target data; use the preset feature extraction model to extract features from the current scene feature data; when the current interactive object feature of the target interactive object is extracted within a preset time after the current moment, update the target interactive object feature to the current interactive object feature; when the current interactive object feature of the target interactive object is not extracted within the preset time after the current moment, based on the extraction time of the object feature, filter out the historical interactive object feature corresponding to the target interactive object from the object feature, and update the target interactive object feature to the historical interactive object feature.

[0033] Furthermore, embodiments of this application also provide an electronic device, including a processor and a memory, wherein the memory stores an application program, and the processor is used to run the application program in the memory to execute the content recommendation method provided in embodiments of this application.

[0034] Furthermore, embodiments of this application also provide a computer-readable storage medium storing a plurality of instructions adapted for loading by a processor to execute steps in any of the content recommendation methods provided in embodiments of this application.

[0035] Furthermore, embodiments of this application also provide a computer program product, including a computer program or instructions, which, when executed by a processor, implement the steps in the content recommendation method provided in embodiments of this application.

[0036] This application embodiment acquires content data streams from at least one business scenario and identifies at least one content interaction information within the content data stream. Based on the content interaction information, it generates at least one content sample for the business scenario and extracts at least one scenario feature data corresponding to the content sample from the content data stream. Then, based on the scenario feature data, it trains a preset feature extraction model to obtain at least one object feature. The object feature is written into a preset object feature set to update the preset object feature set. Then, when a content recommendation request for a target object is received, content recommendation is performed on the target object based on the updated object feature set. Since this scheme can extract features from the content data stream of at least one business scenario through a pre-trained feature extraction model to capture rich object features and write these object features into a preset object feature set, when a content recommendation request is received, these object features can be directly called as auxiliary information for content recommendation. Therefore, it can improve the efficiency and accuracy of content recommendation. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is a schematic diagram of a scenario illustrating the content recommendation method provided in an embodiment of this application;

[0039] Figure 2 This is a flowchart illustrating the content recommendation method provided in an embodiment of this application;

[0040] Figure 3 This is a schematic diagram illustrating the writing of object features into an intermediate database, as provided in an embodiment of this application.

[0041] Figure 4 This is a schematic diagram illustrating the training of a content recommendation model based on an updated set of object features in a distributed storage space scenario provided by an embodiment of this application.

[0042] Figure 5 This is a schematic diagram illustrating the real-time calculation of features and their writing to the EmbeddingServer in the content recommendation method provided in this application embodiment;

[0043] Figure 6 This is an overall architecture diagram of extracting object features using pre-training and writing them into a preset object feature set, provided in an embodiment of this application.

[0044] Figure 7This is a schematic diagram of the overall process under the behavior sample-driven scheme provided in the embodiments of this application;

[0045] Figure 8 This is a schematic diagram of the overall process under the request-driven scheme provided in the embodiments of this application;

[0046] Figure 9 This is another flowchart illustrating the content recommendation method provided in the embodiments of this application;

[0047] Figure 10 This is a schematic diagram of the structure of the content recommendation device provided in the embodiments of this application;

[0048] Figure 11 This is another structural schematic diagram of the content recommendation device provided in the embodiments of this application;

[0049] Figure 12 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0050] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] This application provides a content recommendation method and related equipment, which may include a content recommendation device, an electronic device, and a computer-readable storage medium. The content recommendation device may be integrated into an electronic device, which may be a server or a terminal, etc.

[0052] The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery network (CDN), and big data and artificial intelligence platforms. The terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, etc., but is not limited to these. The terminal and server can be directly or indirectly connected via wired or wireless communication, which is not limited herein.

[0053] For example, see Figure 1Taking the integration of a content recommendation device into an electronic device as an example, the electronic device can acquire content data streams in at least one business scenario, identify at least one content interaction information in the content data stream, generate at least one content sample in the business scenario based on the content interaction information, extract at least one scenario feature data corresponding to the content sample from the content data stream, train a preset feature extraction model based on the scenario feature data to obtain at least one object feature, write the object feature into a preset object feature set to update the preset object feature set, and then, when a content recommendation request for a target object is received, perform content recommendation for the target object based on the updated object feature set, thereby improving the efficiency and accuracy of content recommendation.

[0054] It is understood that, in the specific embodiments of this application, the data involved includes content data streams of objects, scene feature data, and object features. When the following embodiments of this application are applied to specific products or technologies, permission or consent is required, and the collection, use, and processing of the relevant data must comply with the relevant laws, regulations, and standards of the relevant countries and regions.

[0055] The following sections provide detailed descriptions of each example. It should be noted that the order in which the embodiments are described is not intended to limit the preferred order of the embodiments.

[0056] This embodiment will be described from the perspective of a content recommendation device, which can be integrated into an electronic device, such as a server or a terminal. The terminal can include tablet computers, laptops, personal computers (PCs), wearable devices, virtual reality devices, or other smart devices capable of content recommendation.

[0057] A content recommendation method includes:

[0058] The system acquires content data streams for at least one business scenario and identifies at least one content interaction information in the content data stream. Based on the content interaction information, it generates at least one content sample for the business scenario and extracts at least one scenario feature data corresponding to the content sample from the content data stream. Based on the scenario feature data, it trains a preset feature extraction model to obtain at least one object feature and writes the object feature into a preset object feature set to update the preset object feature set. When a content recommendation request for a target object is received, it recommends content to the target object based on the updated object feature set.

[0059] like Figure 2 As shown, the specific process of this content recommendation method is as follows:

[0060] 101. Obtain content data streams for at least one business scenario and identify at least one content interaction information in the content data streams.

[0061] In this context, a business scenario can be understood as at least one scenario that requires the use of content recommendation, such as at least one scenario under a content interaction platform, or at least one scenario under advertising recommendation, and so on.

[0062] Content data streams can be understood as the data streams generated and interacted with within a content interaction platform. Here, content can be understood as information that can be disseminated or interacted with. Content can be of various types, including modalities such as articles, videos, audio, animations, or images, as well as advertisements containing these modalities, and so on.

[0063] Content interaction information can be understood as information that interacts with content. For example, if the content is an advertisement, content interaction information could include clicking on the advertisement or making a payment, etc.

[0064] There are multiple ways to obtain content data streams in at least one business scenario, as follows:

[0065] For example, content data streams can be obtained from the business server or business database corresponding to the direct business scenario; or, content data streams for at least one business scenario can be obtained from the interaction server corresponding to the content interaction platform; or, data streams uploaded by at least one interactive terminal interacting with the content interaction platform can be received to obtain content data streams for at least one business scenario; or, real-time information on the interaction between at least one interactive terminal and content in at least one business scenario content platform can be received to obtain content data streams, and so on.

[0066] After acquiring content data streams for at least one business scenario, at least one piece of content interaction information can be identified within the content data stream. There are several ways to identify at least one piece of content interaction information within the content data stream. For example, at least one interactive behavior targeting the content can be identified within the content data stream, and the corresponding behavioral information can be extracted from the content data stream to obtain the content interaction information. Alternatively, at least one piece of interactive information targeting the content can be directly identified within the content data stream and used as the content interaction information, and so on.

[0067] 102. Based on content interaction information, generate at least one content sample in the business scenario, and extract at least one scenario feature data corresponding to the content sample from the content data stream.

[0068] In this context, a content sample can be understood as content that interacts with the interactive object. For example, if the content is an advertisement, the advertisement can be considered a content sample when it is clicked or paid for.

[0069] Among them, scenario feature data can be understood as the feature data of content samples in this business scenario. For example, if the content sample is an advertisement, the scenario feature data may include the production time, placement time, click rate, complaint rate, payment rate of the advertisement, or the object identifier, interaction time, interaction terminal or other feature data of the interactive object that interacts with it, etc.

[0070] There are several ways to generate at least one content sample for a business scenario based on content interaction information, as follows:

[0071] For example, at least one interactive behavior can be identified in the interactive information, and the exposure identifier corresponding to the interactive behavior can be determined. The content corresponding to the exposure identifier can be filtered out from a preset content set to obtain at least one content sample for the business scenario.

[0072] After generating at least one content sample for a business scenario based on content interaction information, at least one scenario feature data corresponding to the content sample can be extracted from the content data stream. There are several ways to extract this data. For example, at least one dimension of data related to the content sample can be extracted from the content data stream to obtain at least one scenario feature data. Alternatively, at least one dimension of data related to the content sample can be extracted from the content data stream to obtain content feature data. Then, at least one interactive object interacting with the content sample can be extracted from the content interaction information, and the object attribute information of the interactive object can be obtained. The content feature data and the interactive object attribute information can then be used as scenario feature data, and so on.

[0073] 103. Based on the scene feature data, train the preset feature extraction model to obtain at least one object feature.

[0074] Among them, object features can be understood as auxiliary information used for content recommendation. These object features may include the interaction object features of at least one interactive object that interacts with the content sample, or the content object features of the content sample, and so on.

[0075] The pre-trained feature extraction model can be understood as a pre-trained model used for feature extraction. This pre-trained feature extraction model can include an interactive object feature extraction network and a content object feature extraction network.

[0076] There are several ways to train a preset feature extraction model based on scene feature data to obtain at least one object feature, as follows:

[0077] For example, object feature data of at least one interactive object that interacts with the content sample can be extracted from scene feature data, and at least one content feature data of the content sample can be extracted from scene feature data. Based on the object feature data and content features, a preset feature extraction model can be trained to obtain at least one object feature.

[0078] Object feature data can be understood as feature data related to the interactive object. Content feature data can be understood as feature data related to the content sample. There are multiple ways to train the preset feature extraction model based on object feature data and content feature data. For example, feature extraction can be performed on the content feature data to obtain at least one content object feature, and feature extraction can be performed on the object feature data to obtain at least one interactive object feature. This interactive object feature can then be used as the object feature. Based on the content object feature and the interactive object feature, the preset feature extraction model can be trained offline, and the trained feature extraction model can then be used as the preset feature extraction model.

[0079] There are several ways to extract features from content feature data. For example, a content object feature extraction network in a pre-defined feature extraction model can be used to extract features from content feature data to obtain at least one content object feature.

[0080] There are several ways to extract features from object feature data. For example, an interactive object feature extraction network in a preset feature extraction model can be used to extract features from object feature data, thereby obtaining at least one interactive object feature.

[0081] There are several ways to train a preset feature extraction model offline based on content object features and interaction object features. For example, the feature similarity between content object features and interaction object features can be calculated, the matching loss corresponding to the content sample can be determined based on the feature similarity, and the preset feature extraction model can be converged based on the matching loss to obtain the trained feature extraction model.

[0082] After offline training of the preset feature extraction model, the trained feature extraction model can be used as the preset feature extraction model. When processing scene feature data after the current scene feature data, the trained feature extraction model can be used as the preset feature extraction model to extract features, thereby obtaining at least one object feature corresponding to the scene feature data after the current scene feature data, and so on.

[0083] Optionally, in some embodiments, the process of training a preset feature extraction model based on object feature data and content feature data to obtain at least one object may further include: extracting features from the content feature data to obtain at least one content object feature; extracting features from the object feature data to obtain at least one interactive object feature; using the content object interactive feature and the interactive object feature as object features; identifying the context information of the content sample in the content data stream; extracting the context feature from the context information; training the preset feature extraction model based on the context feature, content object feature, and interactive object feature; and using the trained feature extraction model as the preset feature extraction model.

[0084] The methods for feature extraction from content feature data are described above and will not be repeated here. Similarly, the methods for feature extraction from object feature data are also described above and will not be repeated here.

[0085] After extracting the content object features and interaction object features, these features can be used as object features. Therefore, it can be seen that the extracted object features can include both interaction object features and content object features.

[0086] Contextual information can be understood as the contextual information of a content sample during the interaction process, or as the background information of the content sample during the interaction process. There are several ways to identify the contextual information of a content sample in the content data stream. For example, at least one piece of background information of the content sample during the interaction process can be identified in the content data stream to obtain the contextual information.

[0087] After identifying the contextual information of content samples in the content data stream, contextual features can be extracted from this information. Then, a pre-defined feature extraction model is trained based on these contextual features, content object features, and interaction object features. There are several ways to train the pre-defined feature extraction model. For example, based on the contextual features, the feature similarity between content object features and interaction object features can be calculated, and based on this feature similarity, the matching loss of the content samples can be determined. The pre-defined feature model is then converged based on this matching loss, resulting in the trained feature extraction model.

[0088] After training the preset feature extraction model based on context features, content object features, and interaction object features, the trained feature extraction model can be used as the preset feature extraction model.

[0089] In this solution, multiple feature groups are selected for feature data from the entire scene. By training a pre-trained preset feature extraction model, information from different scenes and multiple data sources can be fused in the pre-training process to capture rich language or feature identifiers. Moreover, the training process can be conducted offline, which can support a larger number of features.

[0090] 104. Write the object features into the preset object feature set to update the preset object feature set.

[0091] The preset object feature set can be understood as a pre-defined set of extracted object features. This preset object feature set can be a feature database or a feature set in a distributed storage space.

[0092] There are several ways to write object features into a preset object feature set, as follows:

[0093] For example, the storage space type of a preset object feature set can be obtained. When the storage space type is a feature database, object features are written to the preset object feature set through an intermediate database to update the preset object feature set. When the storage space type is a distributed storage space, the preset object feature set is written through a data writing component to update the preset object feature set. Specifically, it can be as follows:

[0094] S1. Obtain the storage space type of the preset object feature set.

[0095] In this context, storage space type can be understood as a predefined set of object features serving as the type of storage space. Storage space types can include feature databases and distributed storage spaces. A feature database can be understood as a database that stores object features. A distributed storage space can be understood as a distributed storage space.

[0096] There are several ways to obtain the storage space type of the preset object feature set, as follows:

[0097] For example, one can obtain the attribute information of a preset set of object features, and determine the storage space type of the preset set of object features based on the attribute information; or one can obtain the storage address type of the preset object features, and determine the storage space type of the preset set of object features based on the storage address type, and so on.

[0098] S2. When the storage space type is a feature database, the object features are written to the preset object feature set through the intermediate database to update the preset object feature set.

[0099] The intermediate database can be understood as a database used for data transfer. There are several ways to write object characteristics into a predefined object characteristic set through the intermediate database, including the following:

[0100] For example, deduplication of object features can be performed to obtain deduplicated object features. The deduplicated object features can be written to an intermediate database through a cache queue. The deduplicated object features can then be distributed to a preset object feature set through the intermediate database to update the preset object feature set.

[0101] There are several ways to deduplicate object features. For example, based on the generation time of the object features, the object features can be divided into at least one batch of object feature sets. A preset number of consecutive batches of object feature sets can be selected from the object feature sets to obtain the object feature set to be deduplicated. The object features in the object feature set to be deduplicated are then deduplicated to obtain the deduplicated object features.

[0102] The preset quantity can be a pre-defined number of consecutive batches, such as 2, 3, 4, 5, or any number exceeding 2. There are several ways to filter out the preset number of consecutive batches of object features from the object feature set to obtain the set of object features to be deduplicated. For example, taking a preset quantity of n, where n is any integer greater than 2, the object feature set is sorted based on the generation time. Based on the sorting result, n consecutive batches of object features are selected from the object feature set to obtain the set of object features to be deduplicated.

[0103] After selecting a predetermined number of consecutive batches of object features from the object feature set, the object features in the selected deduplicated object feature set can be deduplicated to obtain the deduplicated object features. There are several ways to deduplicate the object features in the deduplicated object feature set. For example, at least one object corresponding to the same feature identifier or feature field (key) can be selected from the deduplicated object feature set to obtain at least one candidate deduplicated feature set. From the candidate deduplicated feature set, the object features with the shortest generation time from the current time are selected to obtain the retained object features. All object features in the candidate deduplicated feature set except for the retained object features are deleted to obtain the deduplicated object features.

[0104] In this solution, deduplication can be performed based on feature identifiers or key fields before writing object data to the intermediate database. By deduplicating k consecutive batches of data, the amount of data written to the intermediate database can be significantly reduced, alleviating the write pressure on the storage system while ensuring data accuracy and consistency.

[0105] After deduplicating object features, the deduplicated object features can be written to an intermediate database through a cache queue. There are several ways to write deduplicated object features to the intermediate database through a cache queue. For example, the thread status of asynchronous threads in the cache queue can be obtained, and based on the thread status, at least one target asynchronous thread can be selected from the asynchronous threads, and the object features can be asynchronously written to the intermediate database through the target asynchronous thread.

[0106] In order to ensure that the consumption speed of the cache queue can match or even exceed the production speed, this solution adopts a parallel read and write mechanism of multiple asynchronous threads, which can effectively improve the data processing throughput and avoid cache queue blocking or data loss due to untimely data processing.

[0107] After the object features are asynchronously written to the intermediate database via the target asynchronous thread, the deduplicated object features can be distributed to the preset object feature set through the intermediate database to update the preset object feature set.

[0108] Optionally, in some embodiments, after obtaining the thread state of the asynchronous thread in the cache queue, at least one current queue parameter of the cache queue can be determined based on the thread state. When the current queue parameter exceeds a preset parameter threshold, an alarm message is generated to indicate that there is an anomaly in the cache queue.

[0109] The current queue parameters can be understood as the performance parameters of the cache queue at the current moment. For example, they may include queue length, data inflow rate, processing rate, cache utilization, or other performance parameters, etc.

[0110] In this solution, by setting reasonable thresholds, an alarm mechanism can be triggered in a timely manner when one or more current queue parameters in the cache queue reach the set thresholds. This allows for rapid response to potential performance bottlenecks or resource exhaustion issues, ensuring the stable operation of the storage system and enabling timely adjustments.

[0111] In this solution, an intermediate database is used for data transfer when writing object features into the feature database. Taking TDbank (a data storage system) as the intermediate database and UserEedding as the object feature as the interactive object feature, the process of writing object features into the intermediate database can be as follows: Figure 3As shown, for each batch of training, the object features (User Eedding) of that batch can be obtained through the dual-tower structure in the preset feature extraction model. An efficient intermediate cache queue system stores the data generated after training for each batch, optimizing the write operation to involve only memory copying, ensuring that data can be written quickly at millisecond (ms) speeds, thereby reducing the impact of I / O latency on the overall training process. Furthermore, during the process of writing object features to TDBank, an independent operation called uploadToTdbankOP (a data operation process) can be developed. This solution, through asynchronous writing to TFBank, data deduplication, and real-time detection and alarm mechanisms, not only significantly improves the speed and efficiency of data processing but also effectively ensures data quality and system stability. This is particularly important for processing large-scale datasets and supporting the training of complex machine learning tasks. In addition, the introduction of real-time detection and alarm mechanisms gives the system self-protection and early warning capabilities, further enhancing the robustness and reliability of the entire data processing flow.

[0112] S3. When the storage space type is distributed storage space, the object characteristics are written to the preset object characteristic set through the data writing component to update the preset object characteristic set.

[0113] For example, when the storage space type is distributed storage space, the object characteristics can be written to the preset object characteristic set through the data writing component (SDK), thereby updating the preset object characteristic set.

[0114] In this context, the distributed storage space can be understood as a data storage node or space within a distributed computing framework (Parameter Server). The pre-defined set of object features in this solution's distributed storage space can be a shared distributed storage space (Embedding Parameter Server, emb-PST). This shared distributed storage space supports shared access from multiple training tasks. Through these sharing mechanisms, different training tasks can reuse the same embedding data, reducing redundant storage and repeated loading, and improving resource utilization. To ensure data consistency and reduce write conflicts, emb-PST employs a read-only mechanism. This means that during training, the embedded data (object features) will not be modified; all update operations are performed elsewhere, thus ensuring data stability and reliability. Furthermore, for pre-trained object features, there is no need to maintain the optimizer's state; bypass updates can be performed via the SDK, simplifying the update logic and reducing computational and storage overhead. Simultaneously, emb-PST is only responsible for storing embedding information (object features) and does not involve optimizer state management, further improving the system's modularity and maintainability.

[0115] It should be noted that, since the data obtained is a content data stream, the update of the preset object feature set is continuous.

[0116] Optionally, in some embodiments, after writing object features into a preset object feature set to update the preset object feature set, the updated object feature set can be used to train at least one preset content recommendation model, thereby obtaining a trained content recommendation model. There are several ways to train at least one preset content recommendation model using the updated object feature set. For example, when the storage space type of the preset object feature set is a feature database, the read / write logs of the preset object feature set are obtained, and at least one preset content recommendation model is trained based on the read / write logs to obtain a trained content recommendation model. When the storage space type of the preset object feature set is a distributed storage space, at least one preset content recommendation model is trained based on the preset object feature set to obtain a trained content recommendation model. Specifically, it can be as follows:

[0117] (1) When the storage space type of the preset object feature set is a feature database, obtain the read and write logs of the preset object feature set, and train at least one preset content recommendation model based on the read and write logs to obtain the trained content recommendation model.

[0118] The read / write log can be understood as the log data generated in real time as object features are written to and read from a preset set of object features. There are several ways to obtain the read / write log of the preset set of object features, as follows:

[0119] For example, datahub (a metadata management platform) can be used to receive and store log data generated in real time from a preset set of object features, thereby obtaining read and write logs of the preset set of object features.

[0120] After obtaining the read / write logs of a preset object feature set, at least one preset content recommendation model can be trained based on these logs to obtain the trained content recommendation model. There are several ways to train at least one preset content recommendation model based on the read / write logs. For example, one could obtain the object identifiers of the sample objects corresponding to at least one preset content recommendation model; based on the object identifiers and the read / write logs, filter the object features corresponding to the sample objects from the updated object feature set to obtain at least one sample object feature; and then train the preset content recommendation model based on these sample object features to obtain the trained content recommendation model.

[0121] The preset content recommendation model can be understood as the model used for content recommendation. There are various types of preset content recommendation models, such as pCTR (a model for predicting click-through rate), pCVR (a model for predicting click-through conversion rate), pDCVR (a model for predicting deep conversion rate), pLTV (a model for predicting lifetime value), coarse-ranking LTR (a learned ranking model for coarse ranking), recall LTR (a learned ranking model for recall), or other models used for content recommendation, etc.

[0122] There are several ways to train a pre-defined content model based on the features of the sample objects. For example, you can obtain a set of content samples corresponding to the pre-defined content recommendation model, use the pre-defined content recommendation model to extract features from the content samples in the set of content samples to obtain the sample content features, calculate the feature similarity between the sample object features and the sample content features, and determine the recommendation loss of the content sample set based on the feature similarity. Then, you can converge the pre-defined content recommendation model based on the content recommendation loss to obtain the trained content recommendation model.

[0123] It should be noted that the training of the preset content recommendation model will be optimized and updated as new sample object features emerge, thereby adapting to the constantly changing data distribution.

[0124] (2) When the storage space type of the preset object feature set is a distributed storage space, at least one preset content recommendation model is trained based on the preset object feature set to obtain the trained content recommendation model.

[0125] For example, when the storage space type of the preset object feature set is a distributed storage space, a model training request is obtained. This model training request carries the feature identifier of at least one object feature required for the current model training. When there is a general object feature corresponding to the feature identifier in the preset object feature set, the general object feature is written into the feature cache space. The object features in the feature cache space are used to train at least one preset content recommendation model to obtain the trained content recommendation model.

[0126] Specifically, when the storage space type of the preset object feature set is a distributed storage space, the objects in this preset object feature set can serve as shared general object features supporting multiple training tasks. The method of training at least one preset content recommendation model using object features in the feature cache space is similar to the method of training a preset content recommendation model based on sample object features, as detailed above, and will not be repeated here.

[0127] Optionally, in some embodiments, object features in the feature cache space can be phased out in batches to ensure that the cache always contains the latest embedded data (object features). There are several ways to phase out object features in the feature cache space. For example, the write time of the object features in the feature cache space can be obtained, and based on the write time, at least one object feature can be selected from the feature cache space to obtain the object features to be phased out, and then the object features to be phased out can be deleted from the feature cache space.

[0128] In this scheme, to achieve a high cache hit rate, a feature cache space in GPU memory is constructed. This accelerates data access by reducing direct requests to the emb-PST (a pre-defined set of object features in distributed storage), thus achieving a hit rate of over 90%. The caching strategy includes a strict part-time check mechanism, ensuring that only key-value pairs in pre-trained embedding requests that can be found in the emb-PST are cached. Furthermore, the feature cache space employs a periodic batch eviction strategy to ensure it always contains the latest embedding data (object features). This strategy allows the system to update the embedding data (object features) from the emb-PST in a timely manner without impacting training performance, maintaining data timeliness and accuracy.

[0129] In this scheme, the process of writing object features into a preset object feature set in a distributed storage space to update the object feature set, and then training a preset content recommendation model based on the updated object feature set, can be described as follows: Figure 4 As shown, emb-PST supports shared access across multiple training tasks. This sharing mechanism allows different training tasks to reuse the same embedding data, reducing redundant storage and repeated loading, and improving resource utilization. To ensure data consistency and reduce write conflicts, emb-PST employs a read-only mechanism. This means that the embedding data is not modified during training; all update operations are performed elsewhere, ensuring data stability and reliability. Taking training task i as an example, common object features are read from emb-PST in read-only mode and stored in the feature cache space. Non-common object features in job i can be read from other distributed storage spaces. Therefore, it can significantly improve data access speed and processing efficiency during training, while effectively reducing storage and management costs. In particular, by implementing an efficient GPU Memory Cache and a read-only emb-PST mechanism, it ensures efficient execution of large-scale deep learning training tasks and optimized resource utilization. Furthermore, the application of a bypass update strategy further simplifies the system's complexity and improves overall scalability and flexibility.

[0130] 105. When a content recommendation request for a target object is received, content recommendation is performed on the target object based on the updated object feature set.

[0131] For example, when a content recommendation request for a target object is received, at least one target interaction object feature corresponding to the target object can be filtered from the updated object feature set, at least one content to be recommended corresponding to the content recommendation request can be obtained, and content features can be extracted from the content to be recommended. Based on the content features and the target interaction object features, the content recommendation model is used to recommend the content to be recommended to the target object.

[0132] There are several ways to recommend content to the target object using a content recommendation model based on content features and target interaction object features. For example, the update time of the target interaction object features in the updated object feature set can be obtained. When the update time does not exceed a preset time threshold, the content to be recommended is recommended to the target object based on the content features and target interaction object features. When the update time exceeds the preset time threshold, the target interaction object features are updated, and the content recommendation model is used to push the content to be recommended to the target object based on the updated interaction object features and content features.

[0133] If the update time does not exceed a preset time threshold, it indicates that the version difference between the target interaction object features and the content features is not significant or the versions are the same. In this case, there is no need to update the target interaction object features. The content recommendation model can then be used to recommend the content to the target object directly based on the content features and target interaction object features. There are several ways to recommend content to the target object using a content recommendation model based on content features and target interaction object features. For example, at least one target feature subgraph can be extracted from the target interaction object features. Based on the target feature subgraph and content features, the content recommendation model can predict the recommendation parameters of the content to be recommended. Based on the recommendation parameters, at least one target content can be selected from the content to be recommended and then recommended to the target object.

[0134] There are several ways to extract at least one target feature sub-graph from the target interaction object features. For example, the target interaction features can be processed, and the processed features can be further processed through datahub to obtain the target interaction object feature map. At least one target feature sub-graph can then be selected from the target interaction object feature map.

[0135] In this context, recommendation parameters can be understood as parameters indicating the degree of recommendation of content to be recommended to the target object. There are various types of recommendation parameters, such as click-through rate (CTR), click-to-conversion rate (CPC), deep conversion rate (DCPR), lifetime value, or other recommendation parameters. Based on the target feature subgraph and content features, there are several ways to predict the recommendation parameters of the content to be recommended using a content recommendation model. For example, one can obtain the object information of the target object, use a content recommendation model to extract features from the object information to obtain the current object features of the target object, and determine the recommendation parameters for each piece of content to be recommended based on the target feature subgraph, the current object features, and the content features. Alternatively, one can obtain the object information of the target object, use a content recommendation model to extract features from the object information to obtain the current object features of the target object, filter out at least one target content object feature corresponding to the target object from the updated object feature set, and determine the recommendation parameters for each piece of content to be recommended based on the target feature subgraph, the target content object features, the current object features, and the content features, and so on.

[0136] After predicting the recommendation parameters of the content to be recommended using a content recommendation model based on the target feature subgraph and content features, at least one target content can be selected from the list of content to be recommended based on the recommendation parameters. There are several ways to select at least one target content from the list of content to be recommended based on the recommendation parameters. For example, the content to be recommended can be sorted according to the recommendation parameters to obtain the sorting results, and at least one target content can be selected from the list of content to be recommended based on the sorting results. Alternatively, at least one target content can be selected from the list of content to be recommended whose recommendation parameters exceed a preset parameter threshold, and so on.

[0137] After selecting at least one target content from the recommended content based on the recommendation parameters, the target content can be recommended to the target object. There are several ways to recommend target content to the target object. For example, the push address of the target object can be obtained, and the target content can be pushed to the target object based on the push address. Alternatively, the terminal address of the target object's logged-in terminal can be obtained, and the target content can be pushed to the target terminal based on the terminal address, thus recommending the target content to the target object, and so on.

[0138] When the update time exceeds a preset time threshold, it indicates a significant difference between the version of the target interactive object's features and the version of the content features. In this case, the target interactive object's features need to be updated. There are several ways to update the target interactive object's features. For example, the target data at the current moment can be identified in the content data stream. From this target data, the current object feature data corresponding to the target interactive object's features can be filtered out. A preset feature extraction model can then be used to extract features from the current scene data. If the current interactive object's features are extracted within a preset time after the current moment, the target interactive object's features are updated to the current interactive object's features. If the current interactive object's features are not extracted within the preset time after the current moment, based on the extraction time of the object features, the historical interactive object features of the target interactive object are filtered out from the object features, and the target interactive object's features are updated to the historical interactive object features.

[0139] The method of using a preset feature extraction model to extract features from the current scene feature data is similar to the method of training a preset feature extraction model based on scene feature data, as described above, and will not be repeated here.

[0140] When the current interactive object features of the target interactive object are extracted within a preset time after the current moment, it means that the real-time calculation through the preset feature extraction model has been completed, and the target interactive object features can be updated to the current interactive object features.

[0141] If the current interaction object features of the target interaction object are not extracted within a preset time after the current moment, it indicates that the real-time calculation of the preset feature extraction model has not been completed. At this time, based on the extraction time of the object features, the historical object interaction features corresponding to the target interaction object can be selected as the next new interaction object features. For example, the object features can be sorted based on their extraction time, and based on the sorting results, the historical interaction object features of the target interaction object closest to the current moment can be selected. Then, the target interaction object features can be updated with the historical interaction object features.

[0142] In the field of content recommendation, ensuring version consistency between user embeddings (interaction object features / User Embedding) and content embeddings (content object features / Item Embedding) is crucial, as it directly impacts recommendation accuracy and overall system performance. Taking content as advertising as an example, in the traditional behavior sample-driven model, there are imbalances in update frequency and inconsistencies caused by version intervals. The imbalance in update frequency can be understood as user interactions (e.g., impressions, clicks, and conversions) potentially being very sparse, resulting in a low update frequency for user embeddings, possibly only every few days or even longer. In contrast, content, due to its high exposure, has a much higher update frequency, potentially updating daily or even hourly. Inconsistencies caused by version intervals can be understood as a significant version gap between user embeddings (interaction object features) and ad embeddings (content object features / content features), meaning their versions are inconsistent. This inconsistency severely impacts the real-time performance and accuracy of the recommendation system, as the system may make decisions based on outdated user information. This solution upgrades the behavior sample-driven model to a request-driven model, enabling real-time calculation of user embeddings. In content recommendation, the process of real-time feature calculation and writing to the EmbeddingServer (a feature service) can be described as follows: Figure 5As shown, the system requests the required object features (content object features / interaction object features) from the cache of the preset object feature set. If the interaction object features in the cache are too old (i.e., the update time is more than a preset time threshold from the current time), asynchronous real-time computation can be triggered. Furthermore, real-time computation here can be understood as performing real-time computation on the feature subgraph of the interaction object features. Finally, the recall / coarse ranking / fine ranking prediction service requests the cache to obtain the updated current interaction object features. It should be noted that if real-time computation is completed, the latest current interaction object features are obtained; if real-time computation is not completed, the slightly newer current interaction object features are obtained. Regarding real-time computation, in request-driven mode, the user embedding is calculated instantly through the UserTower Network (i.e., the interaction object feature extraction network) trained in real-time in the preset feature extraction model. This method ensures that the version of the user embedding is highly consistent with the version of the ad embedding at the same point in time, greatly reducing the version interval. In addition, real-time computation not only increases the frequency of embedding updates but also ensures the synchronization of user and ad embeddings in time. This enables recommender systems to make decisions based on the latest object behavior and ad features, thereby improving the immediacy and personalization of recommendations. By shifting from a behavior sample-driven model to a request-driven model and adopting a real-time computational user embedding update strategy, the version inconsistency problem between user embeddings and ad embeddings can be effectively solved. This shift not only improves the system's responsiveness and recommendation quality but also enhances the system's adaptability and user satisfaction. Furthermore, the real-time update mechanism provides a solid foundation for further optimization and innovation, such as introducing more sophisticated machine learning models and algorithms to handle real-time data streams.

[0143] After updating the characteristics of the target interactive object, a content recommendation model can be used to recommend the content to the target object based on the updated interactive object characteristics and content characteristics. The method of recommending content to the target object based on the updated interactive object characteristics and content characteristics is similar to the method of recommending content to the target object based on the target interactive object characteristics and content characteristics, as detailed above, and will not be repeated here.

[0144] This solution integrates pre-trained information from different scenarios and multiple data sources to form a unified and powerful feature set. Pre-training enables resource sharing, meaning the same pre-trained model can be reused across multiple tasks or scenarios, avoiding resource waste from repeated training. This resource-sharing mechanism ensures that training costs do not increase significantly with model complexity and data volume, resulting in high economic efficiency. The overall structure of this solution is as follows: Figure 6As shown, the input data is full-scene feature data. Multiple feature groups are selected from the full-scene data and input into the pre-training module. The pre-trained model performs pre-training and fusion to extract object features (Embedding). These extracted object features can be written to a feature library (feature database) or a distributed storage space (PST). After writing the object features to the feature library, downstream applications can input the value into samples for use. After writing the object features to the PST, subsequent use of these features only requires inputting the key into the sample and directly connecting to the PST for training.

[0145] This solution can be completed using either a behavior-sample-driven approach or a request-driven approach, from offline pre-training, training the preset content recommendation model, to using the trained content recommendation model for content recommendation.

[0146] Taking content data streams, including data streams from video interaction platforms and advertising samples, as an example, a behavior sample-driven solution can be as follows: Figure 7As shown, two content data streams are acquired. Based on these data streams, content samples (including video and advertising samples) are generated. These samples are then associated with feature data to obtain full-scene feature data. This full-scene feature data is input into a pre-trained feature extraction model for offline training, thereby extracting object features. The object features extracted from behavior samples can then be used as user embeddings. The extracted object features can be processed and stored through a data transfer mechanism, then stored in X-Stor (a feature database), or directly written to PST via the training endpoint for use in subsequent online services or other processing steps. After pre-training, the preset content recommendation model can be trained and inferred online. Training of the preset content recommendation model begins with real-time data collection via DataHub, followed by sample association. Finally, the associated sample object features are used to train at least one preset content recommendation model, resulting in a trained content recommendation model. After training the preset content recommendation model, content recommendations can be performed based on it. In the content recommendation process, the required object features are first extracted using xstore or PST. Then, the data is processed and input into the content recommendation model to complete the content recommendation for the target object. In the behavior sample-driven approach, the training and inference processes of the model can be driven by specific behavior samples (such as user clicks and purchases), allowing the model to better adapt to actual business needs and environmental changes. Through this refined process design and optimization, it is ensured that data can be processed efficiently, the model can be trained accurately, and effective inference can be performed in real time, whether offline or online.

[0147] Taking content data streams, including the data streams from video interaction platforms and advertising samples, as an example, a request-driven solution can be as follows: Figure 8 As shown, request-driven approaches upgrade the embedding representation compared to behavior-sample-driven approaches, considering user embedding, content embedding, and context embedding to capture triple interaction information. The extracted object features can include user embeddings (interaction object features) and content embeddings (content object features). During content recommendation, both interaction object features and content object features are input as auxiliary information into the content recommendation model. Furthermore, the real-time updates of request-driven approaches ensure that the latest feature information (interaction object features) is obtained with each ad request, thereby improving the response speed and accuracy of content recommendation.

[0148] As can be seen from the above, in this embodiment of the application, after acquiring content data streams in at least one business scenario and identifying at least one content interaction information in the content data streams, at least one content sample in the business scenario is generated based on the content interaction information. At least one scenario feature data corresponding to the content sample is extracted from the content data stream. Then, based on the scenario feature data, a preset feature extraction model is trained to obtain at least one object feature. The object feature is written into a preset object feature set to update the preset object feature set. Then, when a content recommendation request for a target object is received, content recommendation is performed on the target object based on the updated object feature set. Since this scheme can extract features from the content data streams in at least one business scenario through a pre-trained feature extraction model to capture rich object features and write these object features into a preset object feature set, when a content recommendation request is received, these object features can be directly called as auxiliary information for content recommendation. Therefore, the efficiency and accuracy of content recommendation can be improved.

[0149] Based on the method described in the above embodiments, the following examples will provide further detailed explanations.

[0150] In this embodiment, the content recommendation device will be specifically integrated into an electronic device, with the electronic device serving as a server, the feature database as X-Stor, and the distributed storage space as PST, as an example for explanation.

[0151] like Figure 9 As shown, a content recommendation method has the following specific process:

[0152] 201. The server obtains content data streams for at least one business scenario.

[0153] For example, the server can obtain content data streams from the business server or business database corresponding to the direct business scenario, or it can obtain content data streams for at least one business scenario from the interaction server corresponding to the content interaction platform, or it can receive data streams uploaded by at least one interactive terminal interacting with the content interaction platform to obtain content data streams for at least one business scenario, or it can also receive real-time information of at least one interactive terminal interacting with content in at least one business scenario content platform to obtain content data streams, and so on.

[0154] 202. The server identifies at least one content interaction information in the content data stream.

[0155] For example, the server can identify at least one interactive line for the content in the content data stream and extract the corresponding behavioral information from the content data stream to obtain content interaction information. Alternatively, it can directly identify at least one interactive information for the content in the content data stream and use that interactive information as content interaction information, and so on.

[0156] 203. The server generates at least one content sample for the business scenario based on the content interaction information.

[0157] For example, the server can identify at least one interactive behavior in the interactive information, determine the exposure identifier corresponding to the interactive behavior, and filter the content corresponding to the exposure identifier from a preset content set to obtain at least one content sample for the business scenario.

[0158] 204. The server extracts at least one scene feature data corresponding to the content sample from the content data stream.

[0159] For example, the server can extract at least one dimension of data related to the content sample from the content data stream to obtain at least one scene feature data. Alternatively, it can extract at least one dimension of data of the content sample from the content data stream to obtain content feature data, and extract at least one interactive object that interacts with the content sample from the content interaction information, obtain the object attribute information of the interactive object, and use the content feature data and the interactive object attribute information as scene feature data, and so on.

[0160] 205. The server trains a preset feature extraction model based on scene feature data to obtain at least one object feature.

[0161] For example, the server can extract object feature data of at least one interactive object that interacts with the content sample from the scene feature data, and extract at least one content feature data of the content sample from the scene feature data.

[0162] The server can use the interactive object feature extraction network in a pre-defined feature extraction model to extract features from the object feature data, thereby obtaining at least one interactive object feature. This interactive object feature is then used as the object feature.

[0163] The server can calculate the feature similarity between content object features and interaction object features. Based on the feature similarity, it determines the matching loss corresponding to the content sample, and based on the matching loss, it converges the preset feature extraction model to obtain the trained feature extraction model. When processing scene feature data after the current scene feature data, the trained feature extraction model can be used as the preset feature extraction model to extract features, thereby obtaining at least one object feature corresponding to the scene feature data after the current scene feature data, and so on.

[0164] Optionally, in some embodiments, the process of the server training a preset feature extraction model based on object feature data and content feature data to obtain at least one object may further include: the server extracting features from the content feature data to obtain at least one content object feature; extracting features from the object feature data to obtain at least one interactive object feature; and using the content object interactive feature and the interactive object feature as object features, identifying at least one piece of background information of the content sample during the interaction process in the content data stream, thereby obtaining context information. Context features are extracted from the context information; based on the context features, the feature similarity between the content object feature and the interactive object feature is calculated; based on the feature similarity, the matching loss of the content sample is determined; and based on the matching loss, the preset feature model is converged to obtain the trained feature extraction model.

[0165] 206. The server writes the object characteristics into the preset object characteristic set to update the preset object characteristic set.

[0166] For example, the server can obtain attribute information of a preset set of object features and determine the storage space type of the preset set of object features based on the attribute information; or it can obtain the storage address type of the preset object features and determine the storage space type of the preset set of object features based on the storage address type, and so on.

[0167] When the preset object set is X-Stor, the server can divide the object features into at least one batch of object feature sets based on their generation time. The object feature sets are then sorted based on their generation time. Based on the sorting result, n (preset number) consecutive object feature sets are selected from these sets to obtain the object feature set to be deduplicated. From the object feature set to be deduplicated, at least one object corresponding to the same feature identifier or feature field (key) is selected to obtain at least one candidate deduplicated feature set. From the candidate deduplicated feature set, the object feature with the shortest generation time from the current time is selected to obtain the retained object feature. All object features in the candidate deduplicated feature set except for the retained object features are deleted to obtain the deduplicated object features.

[0168] The server can obtain the thread status of asynchronous threads in the cache queue. Based on the thread status, it selects at least one target asynchronous thread and writes the object characteristics asynchronously to the intermediate database through the target asynchronous thread. The intermediate database then distributes the deduplicated object characteristics to a preset object characteristic set to update the preset object characteristic set.

[0169] Optionally, in some embodiments, the server may determine at least one current queue parameter of the cache queue based on the thread state, and generate an alarm message when the current queue parameter exceeds a preset parameter threshold. The alarm message is used to indicate that there is an anomaly in the cache queue.

[0170] When the preset object feature set is PST, the server can write the object features into the preset object feature set through the data writing component (sdk), thereby updating the preset object feature set.

[0171] Optionally, in some embodiments, after the server writes object features into a preset object feature set to update the preset object feature set, it can also use the updated object feature set to train at least one preset content recommendation model, thereby obtaining a trained content recommendation model. The specific training process can be as follows:

[0172] (1) When the preset object feature set is X-Stor

[0173] The server can receive and store real-time log data generated from a preset object feature set through a datahub (a metadata management platform), thereby obtaining the read and write logs of the preset object feature set. It then obtains the object identifier of at least one sample object corresponding to a preset content recommendation model. Based on the object identifier and the read and write logs, it filters out the object features corresponding to the sample object from the updated object feature set, thus obtaining at least one sample object feature.

[0174] The server obtains a set of content samples corresponding to a preset content recommendation model, uses the preset content recommendation model to extract features from the content samples in the set, obtains the sample content features, calculates the feature similarity between the sample object features and the sample content features, and determines the recommendation loss of the content sample set based on the feature similarity. Based on the content recommendation loss, the preset content recommendation model is converged to obtain the trained content recommendation model.

[0175] (2) When the preset object feature set is PST

[0176] For example, the server can obtain a model training request, which carries the feature identifier of at least one object feature required for the current model training. When there is a general object feature corresponding to the feature identifier in the preset object feature set, the general object feature is written into the feature cache space. The at least one preset content recommendation model is trained using the object features in the feature cache space to obtain the trained content recommendation model.

[0177] Optionally, in some embodiments, the server may also obtain the write time of object features in the feature cache space, filter out at least one object feature in the feature cache space based on the write time, obtain object features to be eliminated, and delete the object features to be eliminated in the feature cache space.

[0178] 207. When a content recommendation request for a target object is received, the server recommends content for the target object based on the updated set of object features.

[0179] For example, when receiving a content recommendation request for a target object, the server can filter out at least one target interaction object feature corresponding to the target object from the updated object feature set. It can also obtain at least one piece of content to be recommended corresponding to the content recommendation request and extract content features from that content.

[0180] The server obtains the update time of the target interactive object features in the updated object feature set. When the update time does not exceed the preset time threshold, the server processes the target interactive features and further processes the processed features through datahub to obtain the target interactive object feature map. At least one target feature sub-map is selected from the target interactive object feature map.

[0181] The server obtains the object information of the target object, uses a content recommendation model to extract features from the object information, obtains the current object features of the target object, and determines the recommendation parameters for each piece of content to be recommended based on the target feature subgraph, the current object features, and the content features. Alternatively, the server can obtain the object information of the target object, use a content recommendation model to extract features from the object information, obtain the current object features of the target object, filter out at least one target content object feature corresponding to the target object from the updated object feature set, and determine the recommendation parameters for each piece of content to be recommended based on the target feature subgraph, the target content object features, the current object features, and the content features, and so on.

[0182] The server sorts the content to be recommended according to the recommendation parameters, obtains the sorting results, and filters at least one target content from the content to be recommended. Alternatively, it can also filter at least one target content from the content to be recommended whose recommendation parameters exceed a preset parameter threshold, and so on.

[0183] The server can obtain the push address of the target object and push the target content to the target object based on the push address. Alternatively, it can obtain the terminal address of the target terminal that the target object is logged into and push the target content to the target terminal based on the terminal address, so as to recommend the target content to the target object, and so on.

[0184] When the update time exceeds a preset time threshold, the server can identify the target data at the current moment in the content data stream. It then filters out the current object feature data of the target interactive object corresponding to the target interactive object's features. A preset feature extraction model is used to extract features from the current scene data. If the current interactive object feature is extracted within a preset time after the current moment, the target interactive object feature is updated to the current interactive object feature. If the current interactive object feature is not extracted within the preset time after the current moment, the object features are sorted based on their extraction time. Based on the sorting result, the historical interactive object feature of the target interactive object closest to the current moment is selected from the object features. Then, the target interactive object feature is updated to the historical interactive object feature. Based on the updated interactive object feature and content features, a content recommendation model is used to push the recommendation to the target object.

[0185] As can be seen from the above, in this embodiment, after the server acquires content data streams in at least one business scenario and identifies at least one content interaction information in the content data streams, it generates at least one content sample in the business scenario based on the content interaction information. It then extracts at least one scene feature data corresponding to the content sample from the content data stream. Next, based on the scene feature data, it trains a preset feature extraction model to obtain at least one object feature. This object feature is then written into a preset object feature set to update the preset object feature set. Finally, when a content recommendation request for a target object is received, content recommendation is performed on the target object based on the updated object feature set. Since this scheme can extract features from the content data streams in at least one business scenario using a pre-trained feature extraction model to capture rich object features and write these object features into a preset object feature set, it can directly call these object features as auxiliary information for content recommendation when a content recommendation request is received. Therefore, it can improve the efficiency and accuracy of content recommendation.

[0186] To better implement the above methods, this application also provides a content recommendation device, which can be integrated into an electronic device, such as a server or terminal. The terminal may include a tablet computer, a laptop computer, and / or a personal computer.

[0187] For example, such as Figure 10As shown, the content recommendation device may include an acquisition unit 301, a generation unit 302, a training unit 303, a writing unit 304, and a recommendation unit 305, as follows:

[0188] (1) Obtain unit 301;

[0189] The acquisition unit 301 is used to acquire content data streams in at least one business scenario and identify at least one content interaction information in the content data streams.

[0190] For example, the acquisition unit 301 can be used to acquire at least one content data stream in a business scenario, identify at least one interactive behavior for the content in the content data stream, and extract the behavior information corresponding to the interactive behavior in the content data stream, thereby obtaining content interaction information.

[0191] (2) Generation unit 302;

[0192] The generation unit 302 is used to generate at least one content sample in the business scenario based on content interaction information, and to extract at least one scenario feature data corresponding to the content sample from the content data stream.

[0193] For example, the generation unit 302 can be used to identify at least one interactive behavior in the interactive information, determine the exposure identifier corresponding to the interactive behavior, filter the content corresponding to the exposure identifier in the preset content set to obtain at least one content sample of the business scenario, and extract at least one scene feature data corresponding to the content sample in the content data stream.

[0194] (3) Training Unit 303;

[0195] Training unit 303 is used to train a preset feature extraction model based on scene feature data to obtain at least one object feature.

[0196] For example, training unit 303 can be specifically used to extract object feature data of at least one interactive object that interacts with the content sample from scene feature data, extract at least one content feature data of the content sample from scene feature data, perform feature extraction on the content feature data to obtain at least one content object feature, perform feature extraction on the object feature data to obtain at least one interactive object feature, and use the interactive object feature as the object feature or use the interactive object feature and the content object feature as the object feature, perform offline training on the preset feature extraction model based on the content object feature and the interactive object feature, and use the trained feature extraction model as the preset feature extraction model.

[0197] (4) Write to unit 304;

[0198] The writing unit 304 is used to write object features into a preset object feature set to update the preset object feature set.

[0199] For example, the writing unit 304 can be used to obtain the storage space type of the preset object feature set. When the storage space type is a feature database, the object features are deduplicated to obtain the deduplicated object features. The deduplicated object features are written to the intermediate database through the cache queue. The deduplicated object features are distributed to the preset object feature set through the intermediate database to update the preset object feature set. When the storage space type is a distributed storage space, the preset object feature set is written through the data writing component to update the preset object feature set.

[0200] (5) Recommended Unit 305;

[0201] The recommendation unit 305 is used to recommend content to the target object based on the updated set of object features when a content recommendation request for the target object is received.

[0202] For example, recommendation unit 305 can be used to, when receiving a content recommendation request for a target object, filter at least one target interaction object feature corresponding to the target object from the updated object feature set, obtain at least one content to be recommended corresponding to the content recommendation request, extract content features from the content to be recommended, obtain the update time of the target interaction object feature in the updated object feature set, and when the update time does not exceed a preset time threshold, recommend the content to be recommended to the target object based on the content features and the target interaction object features using a content recommendation model. When the update time exceeds the preset time threshold, update the target interaction object features, and push the content to be recommended to the target object based on the updated interaction object features and content features using a content recommendation model.

[0203] Optionally, in some embodiments, the content recommendation device may further include an online training unit 306, which can, as shown in the example below, train an online training unit 306. Figure 11 As shown, the specific details are as follows:

[0204] The online training unit 306 is used to train the preset content recommendation model based on the updated object features to obtain the trained content recommendation model.

[0205] For example, the online training unit 306 can be used to obtain the read and write logs of the preset object feature set when the storage space type of the preset object feature set is a feature database, and to train at least one preset content recommendation model based on the read and write logs to obtain the trained content recommendation model. When the storage space type of the preset object feature set is a distributed storage space, it can train at least one preset content recommendation model based on the preset object feature set to obtain the trained content recommendation model.

[0206] In practice, each of the above units can be implemented as an independent entity or can be arbitrarily combined to be implemented as the same or several entities. For the specific implementation of each of the above units, please refer to the previous method embodiments, which will not be repeated here.

[0207] As can be seen from the above, in this embodiment of the application, after the acquisition unit 301 acquires content data streams in at least one business scenario and identifies at least one content interaction information in the content data stream, the generation unit 302 generates at least one content sample in the business scenario based on the content interaction information, and extracts at least one scene feature data corresponding to the content sample in the content data stream. Then, the training unit 303 trains a preset feature extraction model based on the scene feature data to obtain at least one object feature. The writing unit 304 writes the object feature into a preset object feature set to update the preset object feature set. Then, when the recommendation unit 305 receives a content recommendation request for a target object, it recommends content to the target object based on the updated object feature set. Since this scheme can extract features from the content data stream in at least one business scenario through a pre-trained feature extraction model to capture rich object features and write these object features into a preset object feature set, when a content recommendation request is received, these object features can be directly called as auxiliary information for content recommendation. Therefore, the efficiency and accuracy of content recommendation can be improved.

[0208] This application also provides an electronic device, such as... Figure 12 As shown, it illustrates a structural schematic diagram of the electronic device involved in the embodiments of this application, specifically:

[0209] The electronic device may include components such as a processor 401 with one or more processing cores, a memory 402 with one or more computer-readable storage media, a power supply 403, and an input unit 404. Those skilled in the art will understand that... Figure 12 The electronic device structure shown does not constitute a limitation on the electronic device and may include more or fewer components than shown, or combine certain components, or have different component arrangements. Wherein:

[0210] The processor 401 is the control center of the electronic device, connecting various parts of the device via various interfaces and lines. It executes software programs and / or modules stored in the memory 402, and calls data stored in the memory 402, to perform various functions and process data. Optionally, the processor 401 may include one or more processing cores; preferably, the processor 401 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user interface, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may not be integrated into the processor 401.

[0211] The memory 402 can be used to store software programs and modules. The processor 401 executes various functional applications and data processing by running the software programs and modules stored in the memory 402. The memory 402 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the electronic device, etc. In addition, the memory 402 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. Accordingly, the memory 402 may also include a memory controller to provide the processor 401 with access to the memory 402.

[0212] The electronic device also includes a power supply 403 that supplies power to the various components. Preferably, the power supply 403 can be logically connected to the processor 401 through a power management system, thereby enabling functions such as charging, discharging, and power consumption management through the power management system. The power supply 403 may also include one or more DC or AC power supplies, recharging systems, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components.

[0213] The electronic device may also include an input unit 404, which can be used to receive input digital or character information, and generate keyboard, mouse, joystick, optical or trackball signal inputs related to user settings and function control.

[0214] Although not shown, the electronic device may also include a display unit, etc., which will not be described in detail here. Specifically, in this embodiment, the processor 401 in the electronic device loads the executable files corresponding to the processes of one or more applications into the memory 402 according to the following instructions, and the processor 401 runs the applications stored in the memory 402 to realize various functions, as follows:

[0215] The system acquires content data streams for at least one business scenario and identifies at least one content interaction information in the content data stream. Based on the content interaction information, it generates at least one content sample for the business scenario and extracts at least one scenario feature data corresponding to the content sample from the content data stream. Based on the scenario feature data, it trains a preset feature extraction model to obtain at least one object feature and writes the object feature into a preset object feature set to update the preset object feature set. When a content recommendation request for a target object is received, it recommends content to the target object based on the updated object feature set.

[0216] For example, an electronic device can acquire at least one content data stream in a business scenario, identify at least one interactive behavior targeting the content in the content data stream, and extract the corresponding behavioral information from the content data stream to obtain content interaction information. It can then identify at least one interactive behavior in the interaction information, determine the exposure identifier corresponding to that behavior, filter the content corresponding to the exposure identifier from a preset content set to obtain at least one content sample for that business scenario, and extract at least one scene feature data corresponding to the content sample from the content data stream. From the scene feature data, it can extract object feature data of at least one interactive object interacting with the content sample, extract at least one content feature data of the content sample from the scene feature data, perform feature extraction on the content feature data to obtain at least one content object feature, perform feature extraction on the object feature data to obtain at least one interactive object feature, and use the interactive object feature as the object feature, or combine the interactive object feature and content object feature as the object feature. Based on the content object feature and interactive object feature, it can perform offline training on a preset feature extraction model, and use the trained feature extraction model as the preset feature extraction model. The system retrieves the storage space type of a preset object feature set. When the storage space type is a feature database, the object features are deduplicated to obtain deduplicated object features. These deduplicated object features are then written to an intermediate database via a cache queue. The intermediate database then distributes the deduplicated object features to the preset object feature set to update the preset object feature set. When the storage space type is a distributed storage space, the preset object feature set is written to via a data writing component to update the preset object feature set. When a content recommendation request for a target object is received, at least one target interaction object feature corresponding to the target object can be filtered from the updated object feature set. At least one piece of content to be recommended corresponding to the content recommendation request is obtained, and its content features are extracted. The update time of the target interaction object feature in the updated object feature set is obtained. If the update time does not exceed a preset time threshold, the content to be recommended is recommended to the target object based on the content features and the target interaction object features using a content recommendation model. If the update time exceeds the preset time threshold, the target interaction object features are updated, and the content to be recommended is pushed to the target object based on the updated interaction object features and content features using a content recommendation model.

[0217] For details on the implementation of each of the above operations, please refer to the previous examples, which will not be repeated here.

[0218] As can be seen from the above, this application embodiment, after acquiring content data streams in at least one business scenario and identifying at least one content interaction information in the content data streams, generates at least one content sample in the business scenario based on the content interaction information, extracts at least one scenario feature data corresponding to the content sample in the content data stream, then trains a preset feature extraction model based on the scenario feature data to obtain at least one object feature, writes the object feature into a preset object feature set to update the preset object feature set, and then, when a content recommendation request for a target object is received, content recommendation is performed on the target object based on the updated object feature set. Since this scheme can extract features from the content data streams in at least one business scenario through a pre-trained feature extraction model to capture rich object features and write these object features into a preset object feature set, when a content recommendation request is received, these object features can be directly called as auxiliary information for content recommendation, thus improving the efficiency and accuracy of content recommendation.

[0219] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be performed by instructions, or by instructions controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor.

[0220] Therefore, embodiments of this application provide a computer-readable storage medium storing a plurality of instructions that can be loaded by a processor to execute steps in any of the content recommendation methods provided in embodiments of this application. For example, the instructions can execute the following steps:

[0221] The system acquires content data streams for at least one business scenario and identifies at least one content interaction information in the content data stream. Based on the content interaction information, it generates at least one content sample for the business scenario and extracts at least one scenario feature data corresponding to the content sample from the content data stream. Based on the scenario feature data, it trains a preset feature extraction model to obtain at least one object feature and writes the object feature into a preset object feature set to update the preset object feature set. When a content recommendation request for a target object is received, it recommends content to the target object based on the updated object feature set.

[0222] For example, acquire content data streams for at least one business scenario, identify at least one interactive behavior targeting the content in the content data stream, and extract the corresponding behavioral information from the content data stream to obtain content interaction information. Identify at least one interactive behavior in the interaction information and determine the exposure identifier corresponding to the interactive behavior. Filter the content corresponding to the exposure identifier from a preset content set to obtain at least one content sample for the business scenario. Extract at least one scenario feature data corresponding to the content sample from the content data stream. Extract object feature data of at least one interactive object interacting with the content sample from the scenario feature data. Extract at least one content feature data of the content sample from the scenario feature data. Perform feature extraction on the content feature data to obtain at least one content object feature. Perform feature extraction on the object feature data to obtain at least one interactive object feature. Use the interactive object feature as the object feature, or use the interactive object feature and content object feature as the object feature. Based on the content object feature and interactive object feature, train a preset feature extraction model offline, and use the trained feature extraction model as the preset feature extraction model. The system retrieves the storage space type of a preset object feature set. When the storage space type is a feature database, the object features are deduplicated to obtain deduplicated object features. These deduplicated object features are then written to an intermediate database via a cache queue. The intermediate database then distributes the deduplicated object features to the preset object feature set to update the preset object feature set. When the storage space type is a distributed storage space, the preset object feature set is written to via a data writing component to update the preset object feature set. When a content recommendation request for a target object is received, at least one target interaction object feature corresponding to the target object can be filtered from the updated object feature set. At least one piece of content to be recommended corresponding to the content recommendation request is obtained, and its content features are extracted. The update time of the target interaction object feature in the updated object feature set is obtained. If the update time does not exceed a preset time threshold, the content to be recommended is recommended to the target object based on the content features and the target interaction object features using a content recommendation model. If the update time exceeds the preset time threshold, the target interaction object features are updated, and the content to be recommended is pushed to the target object based on the updated interaction object features and content features using a content recommendation model.

[0223] For details on the implementation of each of the above operations, please refer to the previous examples, which will not be repeated here.

[0224] The computer-readable storage medium may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0225] Since the instructions stored in the computer-readable storage medium can execute the steps of any of the content recommendation methods provided in the embodiments of this application, the beneficial effects that any of the content recommendation methods provided in the embodiments of this application can achieve can be realized, as detailed in the preceding embodiments, and will not be repeated here.

[0226] According to one aspect of this application, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the electronic device to perform the methods provided in the various alternative implementations of the content recommendation or content distribution aspects described above.

[0227] The foregoing has provided a detailed description of a content recommendation method and related equipment provided in the embodiments of this application. The related equipment may include a content recommendation device, an electronic device, and a computer-readable storage medium. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A content recommendation method, characterized in that, include: Acquire content data streams for at least one business scenario, and identify at least one content interaction information in the content data streams; Based on the content interaction information, at least one content sample in the business scenario is generated, and at least one scenario feature data corresponding to the content sample is extracted from the content data stream. Based on the scene feature data, a preset feature extraction model is trained to obtain at least one object feature; The object features are written into a preset object feature set to update the preset object feature set; When a content recommendation request for a target object is received, content recommendation is performed on the target object based on the updated set of object features.

2. The content recommendation method according to claim 1, characterized in that, The step of training a preset feature extraction model based on the scene feature data to obtain at least one object feature includes: Extract object feature data of at least one interactive object that interacts with the content sample from the scene feature data; Extract at least one content feature data of the content sample from the scene feature data; Based on the object feature data and content feature data, the preset feature extraction model is trained to obtain at least one object feature.

3. The content recommendation method according to claim 2, characterized in that, The step of training the preset feature extraction model based on the object feature data and content feature data to obtain at least one object feature includes: Feature extraction is performed on the content feature data to obtain at least one content object feature; Feature extraction is performed on the object feature data to obtain at least one interactive object feature, and the interactive object feature is used as the object feature; Based on the features of the content object and the features of the interaction object, the preset feature extraction model is trained offline, and the trained feature extraction model is used as the preset feature extraction model.

4. The content recommendation method according to claim 2, characterized in that, The step of training the preset feature extraction model based on the object feature data and content feature data to obtain at least one object feature includes: Feature extraction is performed on the content feature data to obtain at least one content object feature; Feature extraction is performed on the object feature data to obtain at least one interactive object feature, and the content object feature and the interactive object feature are used as the object feature; The context information of the content sample is identified in the content data stream, and context features are extracted from the context information; The preset feature extraction model is trained based on the context features, content object features, and interaction object features, and the trained feature extraction model is used as the preset feature extraction model.

5. The content recommendation method according to claim 1, characterized in that, The step of writing the object features into a preset object feature set to update the preset object feature set includes: Obtain the storage space type of the preset object feature set; When the storage space type is a feature database, the object features are written into the preset object feature set through an intermediate database to update the preset object feature set; When the storage space type is a distributed storage space, the object features are written to the preset object feature set through the data writing component to update the preset object feature set.

6. The content recommendation method according to claim 5, characterized in that, The step of writing the object features into the preset object feature set through an intermediate database to update the preset object feature set includes: The object features are deduplicated to obtain the deduplicated object features; The deduplicated object features are written into the intermediate database through a cache queue; The deduplicated object features are distributed to the preset object feature set through the intermediate database to update the preset object feature set.

7. The content recommendation method according to claim 6, characterized in that, The process of deduplicating the object features to obtain deduplicated object features includes: Based on the generation time of the object features, the object features are divided into at least one batch of object feature sets; A predetermined number of consecutive batches of object features are selected from the object feature set to obtain the object feature set to be deduplicated; The object features in the set of objects to be deduplicated are deduplicated to obtain the deduplicated object features.

8. The content recommendation method according to claim 6, characterized in that, The step of writing the object features into the intermediate database through a cache queue includes: Obtain the thread state of the asynchronous thread in the cache queue; Based on the thread state, at least one target asynchronous thread is selected from the asynchronous threads; The target asynchronous thread asynchronously writes the object characteristics into the intermediate database.

9. The content recommendation method according to claim 8, characterized in that, After obtaining the thread state of the asynchronous thread in the cache queue, the method further includes: Based on the thread state, determine at least one current queue parameter of the cache; When the current queue parameter exceeds a preset parameter threshold, an alarm message is generated to indicate that there is an anomaly in the cache queue.

10. The content recommendation method according to claim 5, characterized in that, After writing the object features into a preset object feature set to update the preset object feature set, the method further includes: When the storage space type of the preset object feature set is a feature database, the read and write logs of the preset object feature set are obtained, and based on the read and write logs, at least one preset content recommendation model is trained to obtain the trained content recommendation model. When the storage space type of the preset object feature set is a distributed storage space, at least one preset content recommendation model is trained based on the preset object feature set to obtain the trained content recommendation model.

11. The content recommendation method according to claim 10, characterized in that, The step of training at least one preset content recommendation model based on the read / write logs to obtain a trained content recommendation model includes: Obtain the object identifier of at least one sample object corresponding to a preset content recommendation model; Based on the object identifier and read / write log, the object features corresponding to the sample object are selected from the updated object feature set to obtain at least one sample object feature; Based on the features of the sample objects, the preset content model is trained to obtain the trained content recommendation model.

12. The content recommendation method according to claim 10, characterized in that, The step of training the preset content recommendation model based on the preset object feature set to obtain the trained content recommendation model includes: Obtain a model training request, wherein the model training request carries a feature identifier of at least one object feature required by the current model training task; When a general object feature corresponding to the feature identifier exists in the preset object feature set, the general object feature is written into the feature cache space; At least one preset content recommendation model is trained using the object features in the feature cache space to obtain the trained content recommendation model.

13. The content recommendation method according to claim 12, characterized in that, Also includes: Obtain the write time of the object features in the feature cache space; Based on the write time, at least one object feature is selected from the feature cache space to obtain object features to be eliminated. The feature to be phased out is deleted from the feature cache space.

14. The content recommendation method according to claim 10, characterized in that, The content recommendation based on the updated object feature set for the target object includes: Filter at least one target interaction object feature corresponding to the target object from the updated object feature set; Obtain at least one content to be recommended corresponding to the content recommendation request, and extract content features from the content to be recommended; Based on the content features and the target interaction object features, the content recommendation model is used to recommend the content to be recommended to the target object.

15. The content recommendation method according to claim 14, characterized in that, The step of recommending the content to be recommended to the target object using the content recommendation model based on the content features and the target interaction object features includes: Obtain the update time of the target interactive object feature in the updated object feature set; When the update time does not exceed a preset time threshold, the content to be recommended is recommended to the target object based on the content features and the target interaction object features using the content recommendation model; When the update time exceeds the preset time threshold, the target interactive object features are updated, and based on the updated interactive object features and the content features, the content recommendation model is used to recommend the content to be recommended to the target object.

16. The content recommendation method according to claim 15, characterized in that, The step of recommending the content to be recommended to the target object using the content recommendation model based on the content features and the target interaction object features includes: Extract at least one target feature sub-graph from the features of the target interactive object; Based on the target feature subgraph and content features, the content recommendation model is used to predict the recommendation parameters of the content to be recommended. Based on the recommendation parameters, at least one target content is selected from the content to be recommended, and the target content is recommended to the target object.

17. The content recommendation method according to claim 15, characterized in that, Updating the features of the target interactive object includes: Identify the target data at the current moment in the content data stream, and filter out the current object feature data of the target interactive object corresponding to the target interactive object feature from the target data; The preset feature extraction model is used to extract features from the current scene feature data; When the current interaction object feature of the target interaction object is extracted within a preset time after the current time, the target interaction object feature is updated to the current interaction object feature; If the current interaction object feature of the target interaction object is not extracted within the preset time after the current time, based on the extraction time of the object feature, the historical interaction object feature corresponding to the target interaction object is filtered out from the object features, and the target interaction object feature is updated to the historical interaction object feature.

18. A content recommendation device, characterized in that, include: The acquisition unit is used to acquire at least one content data stream in a business scenario and identify at least one content interaction information in the content data stream. The generation unit is used to generate at least one content sample under the business scenario based on the content interaction information, and to extract at least one scenario feature data corresponding to the content sample from the content data stream. The training unit is used to train a preset feature extraction model based on the scene feature data to obtain at least one object feature. A writing unit is used to write the object features into a preset object feature set to update the preset object feature set; The recommendation unit is used to recommend content to the target object based on the updated set of object features when a content recommendation request for the target object is received.

19. An electronic device, characterized in that, It includes a processor and a memory, the memory storing an application program, and the processor running the application program within the memory to perform the steps of the content recommendation method according to any one of claims 1 to 17.

20. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a plurality of instructions adapted for loading by a processor to perform the steps of the content recommendation method according to any one of claims 1 to 17.

Images