Recommended object recall model training method and recommended content recall method

Abstract

The invention relates to a recommended object recall model training method and a recommended content recall method. The method comprises the following steps: acquiring a recommendation request object and a recommendation content sequence; according to the arrangement sequence of the at least two pieces of recommendation content in the recommendation content sequence, performing duplicate removal and sorting on recommendation objects indicated by the at least two pieces of recommendation content to obtain a recommendation object sequence; constructing training samples according to recommended objects in the recommended object sequence, wherein each training sample comprises object attribute information of at least one recommended object in the recommended object sequence; for each training sample, determining a sample mark of the training sample according to an arrangement sequence of at least one recommendation object indicated by object attribute information in the training sample in a recommendation object sequence; and according to the request object data of the recommended request object, the at least one training sample and the respective sample mark of the at least one training sample, training to obtain a recommended object recall model. By adopting the method, the recall accuracy of the recommended content can be improved.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a method, apparatus, computer device, storage medium and computer program product for training a recommendation object recall model, and a method, apparatus, computer device, storage medium and computer program product for recalling recommendation content. Background Art

[0002] With the development of computer technology, recommendation systems have emerged. Recommendation systems can be used to find content that users are interested in from a large amount of content and recommend content to users when user needs are unclear. Recommendation systems can be widely used in various Internet products.

[0003] In traditional technology, the recommendation system adopts a funnel-type architecture of "recall-coarse ranking-fine ranking". In the recall stage, the recommended content is recalled. By scoring and sorting each candidate recommended content, the recommended content is selected as the recall result and given to the coarse ranking and fine ranking. The candidate recommended content usually indicates that there is a recommended object.

[0004] However, traditional methods can usually only recall candidate recommendation content at the granularity, which leads to low recall accuracy of recommended content in the recall phase, and thus low recommendation accuracy of the recommendation system. Summary of the Invention

[0005] Based on this, it is necessary to provide a recommendation object recall model training method, device, computer equipment, computer-readable storage medium and computer program product that can support improving the recall accuracy of recommended content in the recall link, as well as a recommendation content recall method, device, computer equipment, computer-readable storage medium and computer program product that can improve the recall accuracy of recommended content, in response to the above technical problems.

[0006] In a first aspect, the present application provides a method for training a recommendation object recall model. The method comprises:

[0007] Acquire historical recommendation records; the historical recommendation records include a recommendation request object and a recommended content sequence generated for the recommendation request object; the recommended content sequence includes at least two recommended contents;

[0008] Deduplicating and sorting the recommended objects indicated by the at least two recommended contents according to the arrangement order of the at least two recommended contents in the recommended content sequence to obtain a recommended object sequence;

[0009] constructing at least one training sample according to the recommended objects in the recommended object sequence, each of the training samples comprising object attribute information of at least one recommended object in the recommended object sequence;

[0010] For each of the training samples, determining a sample label of the training sample according to an arrangement order of at least one recommended object indicated by the object attribute information in the training sample in the recommended object sequence;

[0011] A recommendation object recall model is trained based on the request object data of the recommendation request object, the at least one training sample, and the sample label of each of the at least one training sample.

[0012] In a second aspect, the present application also provides a device for training a recommendation object recall model. The device comprises:

[0013] A recommendation record acquisition module, configured to acquire historical recommendation records; the historical recommendation records include a recommendation request object and a recommendation content sequence generated for the recommendation request object; the recommendation content sequence includes at least two recommended contents;

[0014] a recommendation object sequence generation module, configured to remove duplicates and sort the recommendation objects indicated by the at least two recommended contents according to the arrangement order of the at least two recommended contents in the recommendation content sequence, to obtain a recommendation object sequence;

[0015] a training sample construction module, configured to construct at least one training sample based on the recommended objects in the recommended object sequence, each of the training samples comprising object attribute information of at least one recommended object in the recommended object sequence;

[0016] a sample label generation module, configured to determine, for each of the training samples, a sample label of the training sample according to an arrangement order of at least one recommended object indicated by object attribute information in the training sample in the recommended object sequence;

[0017] The training module is used to train and obtain a recommendation object recall model based on the request object data of the recommendation request object, the at least one training sample and the sample label of each of the at least one training sample.

[0018] In a third aspect, the present application further provides a computer device. The computer device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are performed:

[0019] Acquire historical recommendation records; the historical recommendation records include a recommendation request object and a recommended content sequence generated for the recommendation request object; the recommended content sequence includes at least two recommended contents;

[0020] Deduplicating and sorting the recommended objects indicated by the at least two recommended contents according to the arrangement order of the at least two recommended contents in the recommended content sequence to obtain a recommended object sequence;

[0021] constructing at least one training sample according to the recommended objects in the recommended object sequence, each of the training samples comprising object attribute information of at least one recommended object in the recommended object sequence;

[0022] For each of the training samples, determining a sample label of the training sample according to an arrangement order of at least one recommended object indicated by the object attribute information in the training sample in the recommended object sequence;

[0023] A recommendation object recall model is trained based on the request object data of the recommendation request object, the at least one training sample, and the sample label of each of the at least one training sample.

[0024] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the following steps:

[0025] Acquire historical recommendation records; the historical recommendation records include a recommendation request object and a recommended content sequence generated for the recommendation request object; the recommended content sequence includes at least two recommended contents;

[0026] Deduplicating and sorting the recommended objects indicated by the at least two recommended contents according to the arrangement order of the at least two recommended contents in the recommended content sequence to obtain a recommended object sequence;

[0027] constructing at least one training sample according to the recommended objects in the recommended object sequence, each of the training samples comprising object attribute information of at least one recommended object in the recommended object sequence;

[0028] For each of the training samples, determining a sample label of the training sample according to an arrangement order of at least one recommended object indicated by the object attribute information in the training sample in the recommended object sequence;

[0029] A recommendation object recall model is trained based on the request object data of the recommendation request object, the at least one training sample, and the sample label of each of the at least one training sample.

[0030] In a fifth aspect, the present application further provides a computer program product. The computer program product includes a computer program that, when executed by a processor, implements the following steps:

[0031] Acquire historical recommendation records; the historical recommendation records include a recommendation request object and a recommended content sequence generated for the recommendation request object; the recommended content sequence includes at least two recommended contents;

[0032] Deduplicating and sorting the recommended objects indicated by the at least two recommended contents according to the arrangement order of the at least two recommended contents in the recommended content sequence to obtain a recommended object sequence;

[0033] constructing at least one training sample according to the recommended objects in the recommended object sequence, each of the training samples comprising object attribute information of at least one recommended object in the recommended object sequence;

[0034] For each of the training samples, determining a sample label of the training sample according to an arrangement order of at least one recommended object indicated by the object attribute information in the training sample in the recommended object sequence;

[0035] A recommendation object recall model is trained based on the request object data of the recommendation request object, the at least one training sample, and the sample label of each of the at least one training sample.

[0036] In a sixth aspect, the present application provides a method for recalling recommended content. The method comprises:

[0037] In response to a recommended content acquisition request, determining a target request object corresponding to the recommended content acquisition request, and acquiring request object data of the target request object and object attribute information of a plurality of candidate recommended objects;

[0038] For each candidate recommended object, inputting the request object data of the target request object and the object attribute information of the candidate recommended object into a recommended object recall model to obtain a recall probability of the candidate recommended object; the recommended object recall model is obtained by the recommended object recall model training method described above;

[0039] Selecting at least one recall recommendation object from the multiple candidate recommendation objects according to the recall probability of each candidate recommendation object;

[0040] For each of the recalled recommendation objects, candidate recommendation content indicating the recalled recommendation object is recalled as recommended content.

[0041] In a seventh aspect, the present application further provides a device for recalling recommended content. The device comprises:

[0042] a request response module, configured to respond to a recommendation content acquisition request, determine a target request object corresponding to the recommendation content acquisition request, and acquire request object data of the target request object and object attribute information of a plurality of candidate recommendation objects;

[0043] a recall probability generation module configured to input the request object data of the target request object and the object attribute information of the candidate recommendation object into a recommendation object recall model for each candidate recommendation object, thereby obtaining a recall probability of the candidate recommendation object; the recommendation object recall model is obtained by the recommendation object recall model training method described above;

[0044] A recommendation object selection module, configured to select at least one recall recommendation object from the plurality of candidate recommendation objects according to the recall probability of each candidate recommendation object;

[0045] The recommended content determination module is configured to recall, for each of the recalled recommendation objects, the multimedia content indicating the recalled recommendation object as the recommended content.

[0046] In an eighth aspect, the present application further provides a computer device. The computer device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0047] In response to a recommended content acquisition request, determining a target request object corresponding to the recommended content acquisition request, and acquiring request object data of the target request object and object attribute information of a plurality of candidate recommended objects;

[0048] For each candidate recommended object, inputting the request object data of the target request object and the object attribute information of the candidate recommended object into a recommended object recall model to obtain a recall probability of the candidate recommended object; the recommended object recall model is obtained by the recommended object recall model training method described above;

[0049] Selecting at least one recall recommendation object from the multiple candidate recommendation objects according to the recall probability of each candidate recommendation object;

[0050] For each of the recalled recommendation objects, candidate recommendation content indicating the recalled recommendation object is recalled as recommended content.

[0051] In a ninth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the following steps:

[0052] In response to a recommended content acquisition request, determining a target request object corresponding to the recommended content acquisition request, and acquiring request object data of the target request object and object attribute information of a plurality of candidate recommended objects;

[0053] For each candidate recommended object, inputting the request object data of the target request object and the object attribute information of the candidate recommended object into a recommended object recall model to obtain a recall probability of the candidate recommended object; the recommended object recall model is obtained by the recommended object recall model training method described above;

[0054] Selecting at least one recall recommendation object from the multiple candidate recommendation objects according to the recall probability of each candidate recommendation object;

[0055] For each of the recalled recommendation objects, candidate recommendation content indicating the recalled recommendation object is recalled as recommended content.

[0056] In a tenth aspect, the present application further provides a computer program product. The computer program product includes a computer program that, when executed by a processor, implements the following steps:

[0057] In response to a recommended content acquisition request, determining a target request object corresponding to the recommended content acquisition request, and acquiring request object data of the target request object and object attribute information of a plurality of candidate recommended objects;

[0058] For each candidate recommended object, inputting the request object data of the target request object and the object attribute information of the candidate recommended object into a recommended object recall model to obtain a recall probability of the candidate recommended object; the recommended object recall model is obtained by the recommended object recall model training method described above;

[0059] Selecting at least one recall recommendation object from the multiple candidate recommendation objects according to the recall probability of each candidate recommendation object;

[0060] For each of the recalled recommendation objects, candidate recommendation content indicating the recalled recommendation object is recalled as recommended content.

[0061] The above-mentioned recommendation object recall model training method, device, computer equipment, storage medium and computer program product can obtain the recommendation request object and the recommendation content sequence generated for the recommendation request object by obtaining historical recommendation records, and then can deduplicate and sort the recommendation objects indicated by at least two recommendation contents according to the arrangement order of at least two recommendation contents in the recommendation content sequence to obtain a recommendation object sequence. On this basis, at least one training sample can be constructed using the recommendation objects in the recommendation object sequence to realize the construction of the training sample. Each training sample contains the object attribute information of at least one recommendation object in the recommendation object sequence. For each training sample, the sample label of the training sample is determined by determining the arrangement order of at least one recommendation object in the recommendation object sequence indicated by the object attribute information in the training sample. At least one training sample can be constructed using The arrangement order of the recommended objects realizes the accurate labeling of the training samples, and then the model training can be carried out using the request object data of the recommendation request object, at least one training sample and the sample labels of at least one training sample. The recommendation object recall model is obtained by training. In the whole process, by first using the recommendation content sequence to generate the recommendation object sequence, and then using the object attribute information of the recommended object in the recommendation object sequence to train the recommendation object recall model, it is possible to obtain a recommendation object recall model that uses the recommendation object as the driving factor for recall on the basis of improving the consistency of the sorting of the recommended objects in the recall link and the recommendation content sequence, and can support improving the recall accuracy of the recommended content in the recall link. Furthermore, by first recalling the recommended object and then recalling the recommended content indicated by the recommended object, the recall accuracy of the recommended content in the recall link can be improved on the basis of making full use of the object attribute information of the recommended object.

[0062] The above-mentioned recommended content recall method, apparatus, computer equipment, storage medium and computer program product can, in response to a recommended content acquisition request, determine the target request object corresponding to the recommended content acquisition request, and then, based on obtaining the request object data of the target request object and the object attribute information of multiple candidate recommended objects, for each candidate recommended object, by inputting the request object data of the target request object and the object attribute information of the candidate recommended object into the recommended object recall model, to obtain the recall probability of the candidate recommended object, so that according to the recall probability of each candidate recommended object, at least one recalled recommended object can be selected from multiple candidate recommended objects, and for each recalled recommended object, the candidate recommended content indicating the recalled recommended object is recalled as the recommended content. The entire process can improve the recall accuracy of the recommended content in the recall link by first recalling the recommended object and then recalling the recommended content indicated by the recommended object, based on the full use of the object attribute information of the recommended object. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 A diagram illustrating an application environment of a method for training a recommendation object recall model in one embodiment;

[0064] Figure 2 A flowchart of a method for training a recommendation object recall model in one embodiment is shown;

[0065] Figure 3 A schematic diagram showing an arrangement of recommendation objects indicated by at least two recommendation contents according to an embodiment;

[0066] Figure 4 A schematic diagram showing another embodiment of arranging recommendation objects indicated by at least two recommendation contents;

[0067] Figure 5 A schematic diagram of determining a target ranking of recommended objects in one embodiment;

[0068] Figure 6 A schematic diagram of sorting recommended objects in one embodiment;

[0069] Figure 7 Schematic diagram of the model structure of the initial recall model in one embodiment;

[0070] Figure 8 Schematic diagram of the model structure of the initial recall model in another embodiment;

[0071] Figure 9 A diagram illustrating an application environment of a method for recalling recommended content in one embodiment;

[0072] Figure 10 This is a diagram of an application scenario of a method for recalling recommended content in one embodiment;

[0073] Figure 11 Schematic diagram of a flow chart of a method for recalling recommended content in one embodiment;

[0074] Figure 12 A schematic diagram of the architecture of an advertisement recommendation system in one embodiment;

[0075] Figure 13 This is a logic diagram of a product-first, advertisement-later recall mode in one embodiment;

[0076] Figure 14 A schematic diagram of obtaining a product sequence by finely sorting advertisements in one embodiment;

[0077] Figure 15 Schematic diagram of the model structure of a product recall model in one embodiment;

[0078] Figure 16 is a structural block diagram of a recommended object recall model training device in one embodiment;

[0079] Figure 17 is a structural block diagram of a recommended content recall device in one embodiment;

[0080] Figure 18 FIG. 1 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0081] This application relates to the field of artificial intelligence technology. Artificial Intelligence (AI) is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use knowledge to achieve optimal results. In other words, AI is a comprehensive technology in computer science that attempts to understand the essence of intelligence and produce a new type of intelligent machine that can respond in a manner similar to human intelligence. AI is the study of the design principles and implementation methods of various intelligent machines, giving them the capabilities of perception, reasoning, and decision-making.

[0082] Artificial intelligence technology is a comprehensive discipline covering a wide range of fields, encompassing both hardware and software technologies. Basic AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interaction systems, and mechatronics. AI software technologies primarily encompass computer vision, speech processing, natural language processing, machine learning / deep learning, autonomous driving, and smart transportation. This application primarily addresses machine learning.

[0083] Machine Learning (ML) is a multi-disciplinary interdisciplinary subject involving probability theory, statistics, approximation theory, convex analysis, algorithmic complexity theory and other disciplines. It specializes in studying how computers simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their own performance. Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications are spread across all areas of artificial intelligence. Machine learning and deep learning generally include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and learning by teaching techniques. In order to make the purpose, technical solutions and advantages of this application clearer, the application is further described in detail below in conjunction with the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain this application and are not used to limit this application.

[0084] The recommendation object recall model training method provided in the embodiment of the present application can be applied to Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store data that server 104 needs to process. The data storage system can be integrated with server 104, or it can be located in the cloud or on another server. After the recommendation request subject initiates a recommendation request using 102, server 104 generates a recommended content sequence for the recommendation request subject, pushes recommended content based on the recommended content sequence to terminal 102, and generates a historical recommendation record for the recommendation request subject and the recommended content sequence generated for the recommendation request subject. When training the recommendation object recall model, the server 104 will obtain historical recommendation records, which include recommendation request objects and recommendation content sequences generated for the recommendation request objects; the recommendation content sequence includes at least two recommendation contents, and according to the arrangement order of the at least two recommendation contents in the recommendation content sequence, the recommendation objects indicated by the at least two recommendation contents are deduplicated and sorted to obtain a recommendation object sequence, and at least one training sample is constructed according to the recommendation objects in the recommendation object sequence, and each training sample contains object attribute information of at least one recommendation object in the recommendation object sequence. For each training sample, the sample label of the training sample is determined according to the arrangement order of at least one recommended object indicated by the object attribute information in the training sample in the recommendation object sequence, and the recommendation object recall model is trained based on the request object data of the recommendation request object, at least one training sample and the sample label of each of the at least one training samples.

[0085] Terminal 102 may include, but is not limited to, various desktop computers, laptops, smartphones, tablet computers, IoT devices, portable wearable devices, and aircraft. IoT devices may include smart speakers, smart TVs, smart air conditioners, and smart car devices. Portable wearable devices may include smart watches, smart bracelets, and head-mounted devices. Server 104 may be implemented as a standalone server or a server cluster consisting of multiple servers.

[0086] In one embodiment, Figure 2 As shown, a method for training a recommendation object recall model is provided. The method can be executed by a terminal or a server alone, or by a terminal and a server in collaboration. In the embodiment of the present application, the method is applied to a server as an example for illustration, and includes the following steps:

[0087] Step 202: Obtain historical recommendation records; the historical recommendation records include a recommendation request object and a recommended content sequence generated for the recommendation request object; the recommended content sequence includes at least two recommended contents.

[0088] Here, historical recommendation records refer to completed content recommendation records. Specifically, historical recommendation records are used to record the recommendation request subject and the recommended content sequence generated for the recommendation request subject after the recommended content sequence is generated for the recommendation request subject. It is understood that each time a recommendation request subject initiates a recommendation request, a historical recommendation record is generated for the recommendation request subject. A recommendation request subject refers to the subject requesting a content recommendation. For example, a recommendation request subject can specifically refer to the user requesting a content recommendation.

[0089] The recommended content sequence refers to a sequence of content generated for a recommendation request and recommended for the requesting party. The recommended content sequence includes at least two recommended pieces of content, which are arranged in a specific order within the recommended content sequence. For example, the recommended content sequence may be the output of a trained ranking model. In this case, the at least two pieces of recommended content within the recommended content sequence are ranked from high to low according to the ratings output by the ranking model. Recommended content refers to content recommended for the requesting party, and may include, for example, videos, audio, news, advertisements, etc.

[0090] Specifically, when it is necessary to train the recommendation object recall model, the server will obtain historical recommendation records, which include recommendation request objects and a recommendation content sequence generated for the recommendation request objects, where the recommendation content sequence includes at least two recommended contents.

[0091] Step 204 : De-duplicate and sort the recommended objects indicated by the at least two recommended contents according to the arrangement order of the at least two recommended contents in the recommended content sequence to obtain a recommended object sequence.

[0092] The recommended object refers to the object indicated by the recommended content, that is, the object recommended by the recommended content. For example, if the recommended content is an advertisement, the recommended object can be the product placed in the advertisement. For another example, if the recommended content is information, the recommended object can be the object involved in the information, specifically a person, place, etc. The recommended object sequence refers to the sequence of recommended objects. The recommended objects are arranged in a certain order in the recommended object sequence. It is understood that the order of the recommended objects in the recommended object sequence is associated with the recommended content indicating the recommended objects.

[0093] Specifically, the server will sort the recommended objects indicated by at least two recommended contents according to the arrangement order of the at least two recommended contents in the recommended content sequence, and obtain the arrangement order of the recommended objects indicated by at least two recommended contents. Then, based on the arrangement order of the recommended objects indicated by at least two recommended contents, the server will deduplicate and sort the recommended objects indicated by at least two recommended contents to obtain a recommended object sequence.

[0094] Step 206: construct at least one training sample based on the recommended objects in the recommended object sequence, where each training sample contains object attribute information of at least one recommended object in the recommended object sequence.

[0095] The training samples are samples used to train the recommendation object recall model, and each training sample contains object attribute information for at least one recommended object in the recommended object sequence. The object attribute information for a recommended object refers to information used to describe the characteristics of the recommended object. For example, when the recommended object is a product, the object attribute information may specifically include a product identifier, product name, product category, etc., which are used to describe the characteristics of the product. The product identifier is used to uniquely identify the product and can be configured according to the actual application scenario. For example, the product identifier can specifically be a string used to uniquely identify the product. For another example, when the recommended object is a location, the object attribute information may specifically include a location identifier, location name, and the region to which the location belongs, which are used to describe the characteristics of the location. The location identifier is used to uniquely identify the location and can be configured according to the actual application scenario. The location name is used to describe the location and can be a proprietary name of the location, such as "XX Park". The region to which the location belongs can refer to the administrative region to which the location belongs, such as "XX Province" or "XX City".

[0096] Specifically, the server constructs at least one training sample based on the learning-to-rank algorithm used when training the recommended object recall model and the recommended objects in the recommended object sequence. Each training sample contains object attribute information for at least one recommended object in the recommended object sequence. It is understood that different learning-to-rank algorithms require different training samples, and the learning-to-rank algorithm used can be configured based on the actual application scenario. For example, the learning-to-rank algorithm can be one of pointwise learning to rank, pairwise learning to rank, and listwise learning to rank.

[0097] Among them, single-point ranking learning is used to learn the correlation between the recommendation request object and a single recommended object in the recommendation object sequence, that is, if the ranking learning algorithm adopted is single-point ranking learning, the training sample includes the object attribute information of a single recommended object in the recommendation object sequence. Paired ranking learning is used to learn the partial order relationship between the recommendation request object and the pairwise recommended objects in the recommendation object sequence, that is, if the ranking learning algorithm adopted is paired ranking learning, the training sample includes the object attribute information of the two recommended objects in the recommendation object pair, where the recommendation object pair is obtained by combining any two recommended objects in the recommendation object sequence. List ranking learning is used to model the order relationship between the recommendation request object and all recommended objects in the recommendation object sequence, that is, if the ranking learning algorithm adopted is list ranking learning, the training sample includes the object attribute information of all recommended objects in the recommendation object sequence.

[0098] Step 208 : For each training sample, determine a sample label of the training sample according to the arrangement order of at least one recommended object indicated by the object attribute information in the training sample in the recommended object sequence.

[0099] Among them, sample labeling is mainly used to label training samples as a reference for training the recommendation object recall model, and is used to adjust the network parameters of the trained model during the training of the recommendation object recall model.

[0100] Specifically, for each training sample, the server will determine the sample label of the training sample based on the arrangement order of at least one recommended object indicated by the object attribute information in the training sample in the recommended object sequence. In specific applications, different ranking learning algorithms are used, and the methods of determining sample labels are also different. If the ranking learning algorithm used is single-point ranking learning, since its training sample is the object attribute information of a single recommended object, it is sufficient to determine the sample label based on the arrangement order of the single recommended object in the recommended object sequence. If the ranking learning algorithm used is pairing ranking learning, since its training sample is the object attribute information of two recommended objects, it is sufficient to determine the sample label based on the arrangement order of the two recommended objects in the recommended object sequence. If the ranking learning algorithm used is list ranking learning, since its training sample is the object attribute information of all recommended objects in the recommended object sequence, it is necessary to determine the sample label based on the arrangement order of all recommended objects in the recommended object sequence.

[0101] Step 210 : training a recommendation object recall model based on the request object data of the recommendation request object, at least one training sample, and the sample label of each of the at least one training sample.

[0102] Among them, the request object data of the recommendation request object refers to the data used to describe the characteristics of the recommendation request object. For example, when the recommendation request object is a user, the request object data can specifically be data used to describe the characteristics of the user. For example, when the recommendation request object is a user, the request object data can specifically be object attribute characteristics, object behavior statistical characteristics, context characteristics, etc. Among them, the object attribute characteristics refer to the characteristics used to characterize the attributes of the recommendation request object. For example, the object attribute characteristics can specifically be the place where the object belongs, etc. The object behavior statistical characteristics refer to the characteristics used to characterize the browsing, purchasing, clicking and other behaviors of the recommendation request object on the candidate recommendation objects within the statistical time period. For example, the object behavior statistical characteristics can specifically be the recommendation object identifier of the candidate recommendation object browsed, purchased or clicked by the recommendation request object. The context feature refers to the feature used to characterize the recommended content display area browsed by the recommendation request object. For example, the context feature can specifically refer to the identifier of the application that displays the browsed recommended content display area.

[0103] For example, if the candidate recommendation object is a commodity, the object behavior statistical feature refers to the feature used to characterize the browsing, purchasing, clicking and other behaviors of the recommendation request object on the commodity during the statistical time period. Specifically, it can be the identifier of the commodity browsed, purchased or clicked by the recommendation request object, that is, the commodity identifier. Among them, the statistical time period can be configured according to actual statistical needs. For example, the statistical time period can be one week, one month, etc. Contextual features refer to the features used to characterize the advertising display area browsed by the recommendation request object, that is, the advertising position features. Specifically, it can be the identifier of the application that displays the advertising display area.

[0104] Specifically, the server will obtain an initial recall model, and then train the initial recall model based on the request object data of the recommendation request object, at least one training sample, and the sample label of at least one training sample to obtain a recommendation object recall model. In a specific application, the initial recall model is used to predict the recall probability of the recommended object based on the request object data of the recommendation request object and the object attribute information of the recommended object, and output the recall prediction probability. For each training sample, by inputting the object attribute information of at least one recommended object contained in the training sample into the initial recall model, the corresponding recall prediction probability of at least one recommended object can be obtained. Based on this, the server can determine the recall prediction result of the training sample, and thus can adjust the network parameters of the initial recall model according to the recall prediction result and sample label of at least one training sample to obtain a recommendation object recall model.

[0105] The above-mentioned recommendation object recall model training method can obtain the recommendation request object and the recommendation content sequence generated for the recommendation request object by obtaining historical recommendation records, and then can deduplicate and sort the recommendation objects indicated by at least two recommendation contents according to the arrangement order of at least two recommendation contents in the recommendation content sequence to obtain a recommendation object sequence. On this basis, at least one training sample can be constructed using the recommendation objects in the recommendation object sequence to realize the construction of the training sample. Each training sample contains the object attribute information of at least one recommendation object in the recommendation object sequence. For each training sample, the sample label of the training sample is determined by determining the arrangement order of at least one recommendation object in the recommendation object sequence indicated by the object attribute information in the training sample. The arrangement order of at least one recommendation object can be used to realize Accurately label the training samples, and then use the request object data of the recommendation request object, at least one training sample and the sample labels of at least one training sample to perform model training, and obtain the recommendation object recall model through training. The whole process first uses the recommendation content sequence to generate the recommendation object sequence, and then uses the object attribute information of the recommended object in the recommendation object sequence to train the recommendation object recall model. On the basis of improving the consistency of the sorting of the recommended objects in the recall link and the recommendation content sequence, a recommendation object recall model that uses the recommendation object as the driving factor for recall can be obtained, which can support improving the recall accuracy of the recommended content in the recall link. Furthermore, by first recalling the recommended object and then recalling the recommended content indicated by the recommended object, the recall accuracy of the recommended content in the recall link can be improved on the basis of fully utilizing the object attribute information of the recommended object.

[0106] In one embodiment, according to the arrangement order of the at least two recommended contents in the recommended content sequence, the recommended objects indicated by the at least two recommended contents are deduplicated and sorted, and the recommended object sequence obtained includes:

[0107] sorting the recommended objects indicated by the at least two recommended contents according to the arrangement order of the at least two recommended contents in the recommended content sequence to obtain the arrangement order of the recommended objects indicated by the at least two recommended contents;

[0108] Determining a target ranking of each recommended object according to the arrangement order of the recommended objects indicated by the at least two recommended contents;

[0109] The recommended objects are sorted according to the target order of each recommended object to obtain a recommended object sequence.

[0110] The target order rank is used to identify the priority of the recommended object. It is understood that the higher the target order rank, the higher the priority of the recommended object. It should be noted that the target order rank of each recommended object is unique.

[0111] Specifically, for each recommended object, the server uses the order of the recommended content indicating the recommended object in the recommended content sequence as the recommended object's ranking, thereby obtaining the ranking order of the recommended objects indicated by each of the at least two recommended contents. Since different recommended contents may indicate the same recommended object, after obtaining the ranking order of the recommended objects indicated by each of the at least two recommended contents, the server deduplicates the recommended objects based on the ranking order of the recommended objects indicated by the at least two recommended contents, determines a unique target ranking order for each recommended object, and then sorts the recommended objects according to their target ranking order to obtain a recommended object sequence.

[0112] In specific applications, such as Figure 3 As shown, assuming that the recommended content sequence includes 11 recommended contents arranged in sequence, when sorting the recommended objects indicated by each of the 11 recommended contents, for each recommended object, the server will use the arrangement order of the recommended contents indicating the recommended object in the recommended content sequence as the order of the recommended object, and obtain the arrangement order of the recommended objects indicated by each of at least two recommended contents.

[0113] It should be noted that not all recommended content in the recommended content sequence indicates a recommended object. If a recommended content does not indicate a recommended object, then when sorting the recommended objects indicated by at least two recommended contents, the recommended content that does not indicate a recommended object will be ignored. Figure 4 As shown, if the second and seventh recommended contents do not indicate a recommended object, then when sorting the recommended objects indicated by the at least two recommended contents, these two recommended contents will be ignored.

[0114] In this embodiment, by utilizing the arrangement order of at least two recommended contents in the recommended content sequence, the recommended objects indicated by each of the at least two recommended contents are sorted. The arrangement order of the recommended objects indicated by the recommended contents can be used to determine the arrangement order of the recommended objects. Based on this, the target sequence position of each recommended object can be determined. According to the target sequence position of each recommended object, the recommended objects can be sorted to obtain a recommended object sequence. This method of sorting recommended objects maintains the consistency of the recommended content sorting in the recommended content sequence and the recommended object sorting in the recommended object sequence, can improve the consistency of the recommended object sorting in the recall link and the recommended content sequence, is conducive to obtaining a recommended object recall model that uses recommended objects as the driving factor for recall, and can support improving the accuracy of recommended content recall in the recall link.

[0115] In one embodiment, determining the target ranking of each recommended object according to the ranking order of the recommended objects indicated by the at least two recommended contents includes:

[0116] Determining at least one order rank for each recommended object according to the order of the recommended objects indicated by the at least two recommended contents;

[0117] For each recommended object, a target sequence rank of the recommended object is selected from at least one sequence rank of the recommended object.

[0118] The order rank of a recommended object refers to the order in which the recommended object is ranked within the order of the recommended objects indicated by at least two recommended contents. It is understood that if each recommended content indicates a different recommended object, each recommended object will have only one order rank. If at least two recommended contents indicate the same recommended object, then this same recommended object will have at least two order rank positions, i.e., the number of order rank positions will be the same as the number of recommended contents indicating the recommended object.

[0119] Specifically, the server can determine at least one sequence position for each recommended object based on the arrangement order of the recommended objects indicated by at least two recommended contents, and then sort at least one sequence position of each recommended object, and select the highest sequence position from at least one sequence position of the recommended object as the target sequence position of the recommended object.

[0120] In specific applications, such as Figure 5 As shown, assuming that the recommended content sequence includes 11 recommended contents arranged in sequence, wherein the second and seventh recommended contents do not indicate a recommended object, wherein the first, sixth, and eighth recommended contents may indicate the same recommended object 4, the third and fourth recommended contents may indicate the same recommended object 2, the fifth, ninth, and eleventh recommended contents may indicate the same recommended object 1, and the tenth recommended content may indicate a recommended object 3 that is different from all other recommended contents, then after sorting the recommended objects indicated by the at least two recommended contents according to their order in the recommended content sequence and obtaining the order of the recommended objects indicated by the at least two recommended contents, for recommended objects such as recommended object 1, recommended object 2, and recommended object 3 that have at least two order positions, it is necessary to sort the at least two order positions of the recommended objects and select the highest order position from the at least two order positions of the recommended objects as the target order position of the recommended objects. For recommended objects such as recommended object 4 that have only one order position, this order position can be directly used as the target order position of the recommended objects.

[0121] Specifically, for recommended object 1, by sorting its sequence positions (fifth, ninth, eleventh), it can be determined that the target sequence position of recommended object 1 is fifth. For recommended object 2, by sorting its sequence positions (third and fourth), it can be determined that the target sequence position of recommended object 2 is third. For recommended object 4, by sorting its sequence positions (first, sixth, eighth), it can be determined that the target sequence position of recommended object 4 is first. As for recommended object 3, since it has only one sequence position (tenth), this sequence position can be directly used as the target sequence position of recommended object 3, and the target sequence position of each recommended object can be obtained as follows Figure 5 shown.

[0122] Furthermore, the recommended objects are sorted according to the target order of each recommended object (recommended object 4 is the first, recommended object 2 is the third, recommended object 1 is the fifth, and recommended object 3 is the tenth). The recommended object sequence can be as follows: Figure 6 As shown, the arrangement order of the four recommended objects in the recommended object sequence is recommended object 4-recommended object 2-recommended object 1-recommended object 3, among which recommended object 4 is ranked 1 in the recommended object sequence, recommended object 2 is ranked 2 in the recommended object sequence, recommended object 1 is ranked 3 in the recommended object sequence, and recommended object 3 is ranked 4 in the recommended object sequence.

[0123] In this embodiment, by utilizing the arrangement order of the recommended objects indicated by at least two recommended contents, it is possible to determine at least one sequence position of each recommended object, and then for each recommended object, the target sequence position of the recommended object can be selected from at least one sequence position of the recommended object, so as to achieve accurate sorting of the recommended objects and obtain a recommended object sequence on the basis of maintaining consistency with the arrangement of the recommended contents in the recommended content sequence.

[0124] In one embodiment, training a recommendation object recall model based on the request object data of the recommendation request object, at least one training sample, and a sample label of each of the at least one training sample includes:

[0125] For each training sample, inputting the request object data of the recommended request object into the initial recall model, and inputting the object attribute information of at least one recommended object contained in the training sample into the initial recall model respectively, to obtain the corresponding recall prediction probability of the at least one recommended object;

[0126] Obtaining a recall prediction result of the training sample based on the recall prediction probability corresponding to at least one recommended object;

[0127] According to the recall prediction results and sample labels of at least one training sample, the network parameters of the initial recall model are adjusted to obtain a recommended object recall model.

[0128] The initial recall model is used to predict the recall probability of the recommended object based on the request object data of the recommendation request object and the object attribute information of the recommended object, and output the recall prediction probability.

[0129] Specifically, for each training sample, the server will input the request object data of the recommendation request object into the initial recall model, and input the object attribute information of at least one recommended object contained in the training sample into the initial recall model respectively, to obtain the corresponding recall prediction probability of at least one recommended object. On this basis, the server can obtain the recall prediction result of the training sample based on the corresponding recall prediction probability of at least one recommended object, and then calculate the corresponding loss function value of at least one training sample according to the respective recall prediction result and sample label of at least one training sample, and use the corresponding loss function value of at least one training sample to adjust the network parameters of the initial recall model to obtain the recommended object recall model.

[0130] In a specific application, for each recommended object in at least one recommended object included in the training sample, the server will input the request object data of the recommended request object and the object attribute information of the recommended object into the initial recall model to obtain the corresponding recall prediction probability of the recommended object.

[0131] In a specific application, if the training sample contains the object attribute information of a single recommended object in the recommended object sequence, the server will directly use the corresponding recall prediction probability of the single recommended object as the recall prediction result of the training sample. If the training sample contains the object attribute information of two recommended objects in the recommended object sequence, the server will determine the probability that the first recommended object i has a higher recommendation priority than the second recommended object j based on the corresponding recall prediction probabilities of the two recommended objects, and use the determined probability as the recall prediction result of the training sample. If the training sample contains the object attribute information of all recommended objects in the recommended object sequence, the server will calculate the probability of the recommended object sequence based on the corresponding recall prediction probabilities of all recommended objects, and use the obtained probability of the recommended object sequence as the recall result of the training sample.

[0132] In specific applications, for each training sample, based on the recall prediction result of the training sample and the sample label, the server calculates the corresponding loss function value of the training sample. After obtaining the loss function value corresponding to at least one training sample, the server calculates the average of the loss function values ​​of at least one training sample. Based on the average loss function value, the server adjusts the network parameters of the initial recall model through backpropagation to obtain the recommended object recall model.

[0133] In a specific application, since different ranking learning algorithms use different training samples, different loss functions are constructed for different ranking learning algorithms when calculating the loss function value. The loss function constructed by each ranking learning algorithm can be configured according to the actual application scenario. In a specific application, taking single-point ranking learning as an example, the constructed loss function can be cross-entropy loss, and the specific loss function form can be: , where loss represents the loss function, y i is the sample label. When the training sample is a positive sample, y i =1, when the training sample is a negative sample, y i =0, p i is the recall prediction result of the training sample, that is, the recall prediction probability of a single recommended object, that is, the probability that the training sample is a positive sample.

[0134] In a specific application, taking paired ranking learning as an example, the constructed loss function can also be cross entropy loss. The specific loss function form can be: , where loss represents the loss function, y ij represents the sample label of the training sample including the first recommended object i and the second recommended object j, p ij is the recall prediction result of the training sample. ij For example, when the priority of the first recommended object i and the second recommended object j is the same, y ij =0, when the priority of the first recommended object i is greater than that of the second recommended object j, y ij =1, when the priority of the first recommended object i is lower than that of the second recommended object j, y ij =-1.

[0135] In a specific application, after adjusting the network parameters of the initial recall model by back propagation, the server can obtain other recommendation records other than the historical recommendation records used for training as a verification set, and evaluate the recall ability of the initial recall model after the network parameters are adjusted. If the recall ability meets the recall requirements, the initial recall model after the network parameters are adjusted is used as the recommended object recall model. Otherwise, the initial recall model is continuously trained until its recall ability meets the recall requirements, thereby obtaining the recommended object recall model. Among them, the server can use the recall precision ranking consistency evaluation index to evaluate the recall ability of the initial recall model after the network parameters are adjusted. The recall requirement can be a specific index threshold, that is, as long as the index value of the calculated recall precision ranking consistency evaluation index is greater than the index threshold, the recall ability can be considered to meet the recall requirements. It can be understood that in this embodiment, the adopted recall precision ranking consistency evaluation index can be configured according to the actual application scenario.

[0136] In this embodiment, by inputting the request object data of the recommendation request object into the initial recall model, and inputting the object attribute information of at least one recommended object contained in the training sample into the initial recall model respectively, the recall prediction of at least one recommended object can be realized, and the corresponding recall prediction probability of at least one recommended object can be obtained. Then, the corresponding recall prediction probability of at least one recommended object can be used to determine the recall prediction result of the training sample, so that the network parameters of the initial recall model can be adjusted using the respective recall prediction results and sample labels of at least one training sample to obtain the recall model of the recommended object.

[0137] In one embodiment, the request object data of the recommendation request object is input into the initial recall model, and the object attribute information of at least one recommended object included in the training sample is respectively input into the initial recall model to obtain the corresponding recall prediction probability of at least one recommended object.

[0138] Inputting the request object data of the recommended request object into the request object feature extraction network in the initial recall model to obtain a request object feature vector;

[0139] For each recommended object in at least one recommended object included in the training sample, inputting object attribute information of the recommended object into a recommended object feature extraction network in the initial recall model to obtain a recommended object feature vector for the recommended object;

[0140] The request object feature vector and the recommended object feature vector of the recommended object are input into the interoperation layer in the initial recall model to obtain the recall prediction probability of the recommended object.

[0141] Among them, the initial recall model includes a request object feature extraction network, a recommendation object feature extraction network and an interoperation layer. The request object feature extraction network is used to perform feature extraction based on the request object data to obtain a request object feature vector. The recommendation object feature extraction network is used to perform feature extraction based on object attribute information to obtain a recommendation object feature vector. The interoperation layer is used to calculate the similarity between the request object feature vector and the recommendation object feature vector of the recommendation request to obtain the recall prediction probability of the recommended object.

[0142] Specifically, when using the initial recall model to predict the recall probability of the recommended object, the server will input the request object data of the recommended request object into the request object feature extraction network in the initial recall model to obtain the request object feature vector. For each recommended object in at least one recommended object included in the training sample, the object attribute information of the recommended object will be input into the recommended object feature extraction network in the initial recall model to obtain the recommended object feature vector of the recommended object. The request object feature vector and the recommended object feature vector of the recommended object will be input into the interoperation layer in the initial recall model to obtain the recall prediction probability of the recommended object.

[0143] In specific applications, such as Figure 7 As shown, the initial recall model can be specifically a dual-tower model, in which the request object feature extraction network is a tower for recommending request objects, and the recommended object feature extraction network is a tower for recommending objects. The parameters of the two towers are not shared, and the request object feature vector and the recommended object feature vector are output respectively. Then, the request object feature vector and the recommended object feature vector are used as inputs of the interoperation layer to calculate their similarity.

[0144] The request object feature extraction network can include an encoding layer and multiple fully connected layers. The encoding layer encodes the input request object data to obtain a high-dimensional dense encoding vector. The multiple fully connected layers reduce the dimensionality of the high-dimensional dense encoding vector output by the encoding layer to obtain the request object feature vector. The recommended object feature extraction network can include an encoding layer and multiple fully connected layers. The encoding layer encodes the input object attribute information to obtain a high-dimensional dense encoding vector. The multiple fully connected layers reduce the dimensionality of the high-dimensional dense encoding vector output by the encoding layer to obtain the recommended object feature vector. There are also various methods for calculating similarity, such as using simple dot products, cosine similarity calculations, or more complex MLP (Multilayer Perceptron) structures.

[0145] In a specific application, taking the case where the number of fully connected layers of the request object feature extraction network and the recommendation object feature extraction network are both 4, and the similarity is calculated by dot product operation, the model structure of the initial recall model can be as follows: Figure 8 shown.

[0146] In this embodiment, by inputting the request object data of the recommended request object into the request object feature extraction network, feature extraction of the request object data can be achieved to obtain a request object feature vector. By inputting the object attribute information of the recommended object into the recommendation object feature extraction network, feature extraction of the object attribute information can be achieved to obtain a recommendation object feature vector. By inputting the request object feature vector and the recommendation object feature vector into the interoperability layer, recall prediction of the recommended object can be achieved through feature vector interaction to obtain a recall prediction probability of the recommended object.

[0147] In one embodiment, at least one training sample is constructed based on the recommended objects in the recommended object sequence, and each training sample contains object attribute information of at least one recommended object in the recommended object sequence, including:

[0148] For each recommended object in the recommended object sequence, object attribute information of the recommended object is obtained, and the object attribute information of the recommended object is used as a training sample.

[0149] Specifically, if the adopted ranking learning algorithm is single-point ranking learning, for each recommended object in the recommended object sequence, the server will obtain object attribute information of the recommended object and use the object attribute information of the recommended object as a training sample.

[0150] In this embodiment, the construction of training samples can be achieved by obtaining object attribute information of each recommended object.

[0151] In one embodiment, determining the sample label of the training sample according to the arrangement order of at least one recommended object indicated by the object attribute information in the training sample in the recommended object sequence includes:

[0152] Determining the sample type of the training sample according to the arrangement order of the recommended objects indicated by the object attribute information in the training sample in the recommended object sequence;

[0153] Based on the sample type of the training sample, a sample tag of the training sample is generated; the sample tag is used to identify the sample type of the training sample.

[0154] The sample type is used to distinguish different types of training samples. For example, the sample type can be a positive sample or a negative sample.

[0155] Specifically, if the ranking learning algorithm employed is single-point ranking learning, the server will determine the sample type of the training sample, i.e., whether the training sample is a positive sample or a negative sample, based on the order of the recommended objects in the recommended object sequence indicated by the object attribute information in the training sample. Based on the sample type of the training sample, the server will generate a sample label for the training sample, which is used to identify the sample type of the training sample. For example, the sample label can specifically be a string used to identify the sample type of the training sample, with different sample types corresponding to different sample labels. For example, the sample label of a positive sample can be 1, and the sample label of a negative sample can be 0.

[0156] In this embodiment, the arrangement order of the recommended objects in the recommended object sequence indicated by the object attribute information in the training sample can be used to determine the sample type of the training sample, and then based on the sample type of the training sample, a sample label for identifying the sample type of the training sample can be generated to achieve accurate labeling of the training sample.

[0157] In one embodiment, determining the sample type of the training sample according to the arrangement order of the recommended objects indicated by the object attribute information in the training sample in the recommended object sequence includes:

[0158] Obtaining a ranking range of recommended objects corresponding to at least one sample type;

[0159] Comparing the recommended object ranking range corresponding to at least one sample type and the arrangement order of the recommended objects indicated by the object attribute information in the training sample in the recommended object sequence, determining a target ranking range to which the arrangement order of the recommended objects indicated by the object attribute information in the training sample in the recommended object sequence belongs;

[0160] The sample type corresponding to the target rank range is used as the sample type of the training sample.

[0161] Specifically, each sample type has a corresponding recommended object ranking range. When determining the sample type of the training sample, the server will obtain the recommended object ranking range corresponding to at least one sample type, compare the recommended object ranking range corresponding to at least one sample type, and the arrangement order of the recommended objects indicated by the object attribute information in the training sample in the recommended object sequence, so as to determine the target ranking range to which the arrangement order of the recommended objects indicated by the object attribute information in the training sample in the recommended object sequence belongs, and use the sample type corresponding to the target ranking range as the sample type of the training sample.

[0162] In a specific application, the recommended object ranking range corresponding to at least one sample type can be pre-configured according to the actual application scenario. For example, at least one sample type can specifically include positive samples and negative samples, wherein the recommended object ranking range corresponding to the positive samples can be 1 to N, and the recommended object ranking range corresponding to the negative samples can be N+1 to M, wherein M is the number of recommended objects in the recommended object sequence, and N is a positive integer greater than or equal to 1, and N+1 is a positive integer less than or equal to M. For a specific application, for the convenience of calculation, N=x%*M can be set, where x can be configured according to the actual application scenario. For example, x can be 50, then N=M / 2.

[0163] In this embodiment, by obtaining the recommended object ranking range corresponding to at least one sample type, the recommended object ranking range corresponding to at least one sample type can be used to determine the target ranking range to which the arrangement order of the recommended object in the recommended object sequence belongs, and then the sample type corresponding to the target ranking range can be used to determine the sample type of the training sample.

[0164] In one embodiment, at least one training sample is constructed based on the recommended objects in the recommended object sequence, and each training sample contains object attribute information of at least one recommended object in the recommended object sequence, including:

[0165] Combine any two recommended objects in the recommended object sequence to obtain multiple recommended object pairs;

[0166] For each recommended object pair, object attribute information of the two recommended objects in the recommended object pair is obtained, and the object attribute information of the two recommended objects in the recommended object pair is used as a training sample.

[0167] Specifically, if the ranking learning algorithm adopted is pairing ranking learning, the server will combine any two recommended objects in the recommended object sequence to obtain multiple recommended object pairs, and then for each recommended object pair, obtain the object attribute information of the two recommended objects in the recommended object pair, and use the object attribute information of the two recommended objects in the recommended object pair as training samples.

[0168] In this embodiment, the construction of training samples can be achieved by first constructing a recommended object pair and then obtaining object attribute information of the two recommended objects in the recommended object pair.

[0169] In one embodiment, determining the sample label of the training sample according to the arrangement order of at least one recommended object indicated by the object attribute information in the training sample in the recommended object sequence includes:

[0170] determining a recommendation object priority between two recommended objects in a recommended object pair according to an arrangement order of the two recommended objects indicated by the object attribute information in the training sample in the recommended object sequence;

[0171] Based on the priority of the recommended object, sample labels of the training samples are generated; the sample labels are used to identify the priority of the recommended object.

[0172] Specifically, if the learning-by-rank algorithm employed is pairwise learning-by-rank, the server determines the priority of the two recommended objects in a pair based on the order in which the two recommended objects are ranked in the recommended object sequence, as indicated by the object attribute information in the training sample. It is understood that the priority of the recommended object ranked earlier in the recommended object sequence is higher than the priority of the recommended object ranked later. Upon determining the priority of the recommended objects, the server generates a sample label for the training sample based on the priority of the recommended objects, used to identify the priority of the recommended objects.

[0173] In specific applications, the sample tag can be a string used to identify the priority of the recommended object, with different recommendation object priorities represented by different strings. For example, in a recommendation object pair, where the two recommended objects are the first recommended object i and the second recommended object j, respectively, when the first recommended object i and the second recommended object j have the same recommendation object priority, the sample tag can be 0; when the first recommended object i has a higher priority than the second recommended object j, the sample tag can be 1; and when the first recommended object i has a lower priority than the second recommended object j, the sample tag can be -1.

[0174] In this embodiment, the arrangement order of the two recommended objects in the recommended object sequence indicated by the object attribute information in the training sample can be used to determine the recommended object priority between the two recommended objects in the recommended object pair, and then the recommended object priority can be used to generate a sample label for identifying the recommended object priority, which can achieve accurate labeling of the training sample when the training sample is the object attribute information of the two recommended objects in the recommended object pair.

[0175] In one embodiment, at least one training sample is constructed based on the recommended objects in the recommended object sequence, and each training sample contains object attribute information of at least one recommended object in the recommended object sequence, including:

[0176] Get the object attribute information of all recommended objects in the recommended object sequence;

[0177] The object attribute information of all recommended objects in the recommended object sequence is used as training samples.

[0178] Specifically, if the adopted ranking learning algorithm is list ranking learning, the server will obtain the object attribute information of all recommended objects in the recommended object sequence, and use the object attribute information of all recommended objects in the recommended object sequence as training samples.

[0179] In this embodiment, the construction of training samples can be achieved by obtaining object attribute information of all recommended objects.

[0180] In one embodiment, determining the sample label of the training sample according to the arrangement order of at least one recommended object indicated by the object attribute information in the training sample in the recommended object sequence includes:

[0181] The sample labels of the training samples are generated according to the arrangement order of all recommended objects in the recommended object sequence indicated by the object attribute information in the training samples; the sample labels are used to identify the arrangement order of all recommended objects in the recommended object sequence.

[0182] Specifically, if the list-based learning-to-rank algorithm is used, the server will directly generate sample labels for the training samples based on the order of all recommended objects in the recommended object sequence as indicated by the object attribute information in the training samples. The sample labels are used to identify the order of all recommended objects in the recommended object sequence. It is understood that the sample labels can specifically be the order of all recommended objects in the recommended object sequence.

[0183] In this embodiment, the arrangement order of all recommended objects in the recommended object sequence indicated by the object attribute information in the training sample is used to generate sample labels for identifying the arrangement order of all recommended objects in the recommended object sequence. This can achieve accurate labeling of the training samples when the training samples are the object attribute information of all recommended objects.

[0184] The recommended content recall method provided in the embodiment of the present application can be applied to Figure 9In the application environment shown. Among them, the terminal 902 communicates with the server 904 through the network. The data storage system can store data that the server 904 needs to process. The data storage system can be integrated on the server 904, or it can be placed on the cloud or other servers. In response to the recommended content acquisition request initiated by the target request object using the terminal 902, the server 904 determines the target request object corresponding to the recommended content acquisition request, obtains the request object data of the target request object and the object attribute information of multiple candidate recommended objects, and for each candidate recommended object, inputs the request object data of the target request object and the object attribute information of the candidate recommended objects into the recommended object recall model to obtain the recall probability of the candidate recommended object. The recommended object recall model is obtained by the above-mentioned recommended object recall model training method. According to the recall probability of each candidate recommended object, at least one recalled recommended object is selected from the multiple candidate recommended objects. For each recalled recommended object, the candidate recommended content indicating the recalled recommended object is recalled as the recommended content, the recalled recommended content is roughly sorted and finely sorted, the target recommended content is determined, and the target recommended content is pushed to the terminal 902 used by the target request object.

[0185] In one embodiment, Figure 10 As shown, the terminal 902 is installed with an application 1 that can display recommended content and other applications (such as Figure 10 Application 2 and application 3 are shown), in response to the target request object triggering the opening operation of application 1 on terminal 902, terminal 902 will initiate a recommended content acquisition request to server 904, so that server 904 responds to the recommended content acquisition request, first recalls the recommended content, then performs rough and fine sorting on the recalled recommended content, determines the target recommended content, and pushes the target recommended content to terminal 902 used by the target request object. After receiving the target recommended content, terminal 902 will display the target recommended content in the recommended content display area of ​​application 1.

[0186] Terminal 902 may include, but is not limited to, various desktop computers, laptops, smartphones, tablet computers, IoT devices, portable wearable devices, and aircraft. IoT devices may include smart speakers, smart TVs, smart air conditioners, and smart car devices. Portable wearable devices may include smart watches, smart bracelets, and head-mounted devices. Server 904 may be implemented as a standalone server or a server cluster consisting of multiple servers.

[0187] In one embodiment, Figure 11 As shown, a method for recalling recommended content is provided. The method can be executed by a terminal or a server alone, or by a terminal and a server in collaboration. In the embodiment of the present application, the method is applied to a server as an example for explanation, and includes the following steps:

[0188] Step 1102 : In response to the recommended content acquisition request, determine the target request object corresponding to the recommended content acquisition request, and acquire the request object data of the target request object and object attribute information of multiple candidate recommended objects.

[0189] The recommended content acquisition request is a request for triggering content recommendations for a target request object, which refers to the object initiating the recommended content acquisition request. For example, the target request object can be the user initiating the recommended content acquisition request. The recommended content acquisition request can be generated based on an operation of the target request object. For example, if the target request object is a user, when the user opens an application, content may be recommended for the user on the application homepage. Therefore, when the user performs the operation of opening the application, a recommended content acquisition request can be generated and sent to the server via the user's terminal.

[0190] Among them, the request object data of the target request object refers to the data used to describe the characteristics of the target request object. For example, when the target request object is a user, the request object data can specifically be data used to describe the characteristics of the user. For example, when the target request object is a user, the request object data can specifically be object attribute characteristics, object behavior statistical characteristics, context characteristics, etc. Among them, the object attribute characteristics refer to the characteristics used to characterize the attributes of the target request object. For example, the object attribute characteristics can specifically be the place where the object belongs, etc. The object behavior statistical characteristics refer to the characteristics used to characterize the browsing, purchasing, clicking and other behaviors of the target request object on the candidate recommendation objects within the statistical time period. For example, the object behavior statistical characteristics can specifically be the recommended object identifier of the candidate recommendation object browsed, purchased or clicked by the target request object. The context feature refers to the feature used to characterize the recommended content display area browsed by the target request object. For example, the context feature can specifically refer to the identifier of the application that displays the browsed recommended content display area.

[0191] For example, if the candidate recommendation object is a commodity, the object behavior statistical feature refers to the feature used to characterize the browsing, purchasing, clicking, and other behaviors of the target request object on the commodity during the statistical time period. Specifically, it can be the identifier of the commodity browsed, purchased, or clicked by the target request object, that is, the commodity identifier. Among them, the statistical time period can be configured according to actual statistical needs. For example, the statistical time period can be specifically one week, one month, etc. The context feature refers to the feature used to characterize the advertising display area browsed by the target request object, that is, the advertising position feature. Specifically, it can be the identifier of the application that displays the advertising display area.

[0192] Among them, candidate recommendation objects refer to objects that can be recommended as candidates to the target request object. For example, candidate recommendation objects can specifically refer to products that can be recommended as candidates to the target request object, or they can also refer to people, places, etc. that can be recommended as candidates to the target request object. The object attribute information of the candidate recommendation object refers to information used to describe the characteristics of the candidate recommendation object. For example, when the candidate recommendation object is a product, the object attribute information can specifically be the product identifier, product name, product category, etc. used to describe the characteristics of the product.

[0193] Specifically, the target requesting object may initiate a recommended content acquisition request via a terminal. In response to the recommended content acquisition request, the server will determine the target requesting object corresponding to the recommended content acquisition request and obtain the target requesting object's request object data and object attribute information of multiple candidate recommendation objects. In specific applications, the target requesting object's request object data and the object attribute information of multiple candidate recommendation objects may be pre-stored in a database. After determining the target requesting object, the server may directly obtain the target requesting object's request object data and the object attribute information of multiple candidate recommendation objects from the database.

[0194] Step 1104: For each candidate recommended object, the request object data of the target request object and the object attribute information of the candidate recommended object are input into the recommended object recall model to obtain the recall probability of the candidate recommended object; the recommended object recall model is obtained by the above-mentioned recommended object recall model training method.

[0195] Specifically, for each candidate recommended object, the server will input the request object data of the target request object and the object attribute information of the candidate recommended object into the recommended object recall model to obtain the recall probability of the candidate recommended object, wherein the recommended object recall model is obtained through the above-mentioned recommended object recall model training method.

[0196] In specific applications, the recommended object recall model includes a request object feature extraction network, a recommended object feature extraction network and an interoperation layer. The request object data of the target request object and the object attribute information of the candidate recommended object are input into the recommended object recall model. The request object feature extraction network performs feature extraction based on the request object data of the target request object, obtains the request object feature vector and outputs it to the interoperation layer. The recommended object feature extraction network performs feature extraction based on the object attribute information of the candidate recommended object, obtains the recommended object feature vector and outputs it to the interoperation layer. The interoperation layer calculates the similarity between the request object feature vector and the recommended object feature vector to obtain the recall probability of the candidate recommended object.

[0197] Step 1106 : Select at least one recall recommendation object from the multiple candidate recommendation objects based on the recall probability of each candidate recommendation object.

[0198] Specifically, the server will sort multiple candidate recommendation objects based on the recall probability of each candidate recommendation object, and select at least one recall recommendation object from the multiple candidate recommendation objects. The number of recall recommendation objects selected can be configured according to the actual application scenario. For example, the number of recall recommendation objects selected can be N. After sorting the multiple candidate recommendation objects, the server will select the candidate recommendation objects ranked in the top N from the multiple candidate recommendation objects as the recall recommendation objects.

[0199] Step 1108 : For each recall recommendation object, recall the candidate recommendation content indicating the recall recommendation object as the recommended content.

[0200] Specifically, when at least one recalled recommendation object is selected, for each recalled recommendation object, the server will recall the candidate recommended content indicating the recalled recommendation object as the recommended content, so as to recall the recommended content by first recalling the recommended object and then recalling the recommended content based on the recommended object.

[0201] The above-mentioned recommended content recall method, in response to the recommended content acquisition request, can determine the target request object corresponding to the recommended content acquisition request, and then, based on obtaining the request object data of the target request object and the object attribute information of multiple candidate recommended objects, for each candidate recommended object, by inputting the request object data of the target request object and the object attribute information of the candidate recommended object into the recommended object recall model, to obtain the recall probability of the candidate recommended object, so that according to the recall probability of each candidate recommended object, at least one recalled recommended object can be selected from the multiple candidate recommended objects, and for each recalled recommended object, the candidate recommended content indicating the recalled recommended object is recalled as the recommended content. The whole process can improve the recall accuracy of the recommended content in the recall link on the basis of fully utilizing the object attribute information of the recommended object by first recalling the recommended object and then recalling the recommended content indicated by the recommended object.

[0202] It should be noted that the inventors believe that in traditional methods, recommendation systems usually use the method of recalling recommended content in the recall phase. By scoring and ranking each candidate recommendation content, the recommended content is selected as the recall result and given to the coarse ranking and fine ranking. The candidate recommendation content usually indicates that there are recommended objects. It is understandable that because traditional methods can usually only recall at the granularity of candidate recommendation content, the object attribute information of the recommended object can only be added as the underlying feature in the recommendation content recall. The object attribute information of the recommended object has little impact on the result of the recommendation content ranking. As a result, some high-quality recommended objects will not have the opportunity to enter the downstream link due to the failure of the recommended content to be selected. This will lead to low recall accuracy of the recommended content in the recall phase, and thus low recommendation accuracy of the recommendation system.

[0203] Based on this, the present application provides a new recommendation object recall model, which adopts a recall mode of recommending objects first and then recommending content, and uses the recommended objects as the driving factor to recall recommended objects. The recommended content is pulled through the recalled recommended objects, thereby improving the utilization of the object attribute information of the recommended objects in the recall link, and solving the problem that high-quality recommended objects have no chance to enter the downstream due to failure in the recommended content.

[0204] In one embodiment, the present invention's recommendation object recall model training method and recommendation content recall method are applied to the advertisement recall phase of an advertisement recommendation system. The recommendation object is a product, the object attribute information of the recommendation object is product information, and the recommended content is an advertisement. The recommendation request object and target request object are users.

[0205] like Figure 12 As shown, the advertising recommendation system adopts a funnel-type architecture of recall-coarse ranking-fine ranking. It first recalls tens of thousands of advertisements from an advertising library that stores millions of advertisements, then performs coarse ranking on the tens of thousands of advertisements to obtain hundreds of advertisements, and then performs fine ranking on the hundreds of advertisements to obtain individual advertisements, and recommends individual advertisements to users. In traditional methods, the advertising recall model is usually used to score and sort in the recall stage, and high-quality advertisements are selected and sent to the downstream coarse ranking and fine ranking. However, the inventor believes that in the advertising recommendation system, for advertisements with products, advertisers usually need to provide detailed product information (such as product name, product brand, etc.). When such advertisements are placed, the advertisers will explicitly bind the products they place. However, the current ad recall model is based on ad granularity. The rich product information provided by advertisers can only be added to the ad recall model as underlying features. Product information has little impact on the results of ad ranking, which may cause some high-quality products to have no chance to enter the downstream links due to ad failure. This leads to low ad recall accuracy in the ad recall link, and thus low recommendation accuracy of the ad recommendation system.

[0206] Based on this, the present application provides a new product recall model, which adopts a recall mode of products first and advertisements later, uses products as the driving factor for product recall, and uses the recalled products to pull advertisements, thereby improving the utilization of product information in the recall link and solving the problem that high-quality products have no chance to enter the downstream due to failed advertisements. It can be understood that the recall mode of products first and advertisements later proposed in the present application can improve the consistency of product sorting between the recall link of the advertising system and the refined ranking module, so that the recall module can more efficiently select products that are considered high-quality by refined ranking, thereby recalling the high-quality product advertisements corresponding to the products.

[0207] In one embodiment, Figure 13 As shown, the recall logic for product advertisements in this application is upgraded from the traditional method of recalling advertisements to a recall mode of recalling products first and advertisements later. That is, compared with the traditional method of directly recalling advertisements based on ordinary advertisements and product advertisements, this application first recalls products from the product collection of product advertisements, and then pulls advertisements through the recalled products to obtain the advertisement collection of the recalled products to achieve advertisement recall. Among them, ordinary advertisements refer to advertisements that do not contain products. This application does not involve improvements to the advertising recall method for these ordinary advertisements, so it will not be explained here.

[0208] In one embodiment, the product recall model in this application is mainly learned using a ranking learning algorithm. The product recall model training mainly includes three parts: sample construction, model design, and loss function design. Each part is described below:

[0209] 1. Sample Construction

[0210] The product recall model learns the product sequence in the precise ranking queue. The product sequence is constructed based on the advertisement sequence. For different advertisements of the same product, the position corresponding to the advertisement with the highest precise ranking in the advertisement sequence is retained, and then the products are re-sorted according to this position to obtain the product sequence.

[0211] Specifically, the server will obtain historical recommendation records, which include the user and the order of advertisements in the refined ranking queue generated for the user (i.e., the recommended content sequence). The advertisement sequence includes at least two advertisements (i.e., recommended content). After obtaining the advertisement sequence, the server will sort the products indicated by the at least two advertisements (i.e., the recommended objects) according to the order of the at least two advertisements in the advertisement sequence, and obtain the order of the advertisements indicated by the at least two advertisements. Based on the order of the products indicated by the at least two advertisements, the server will determine at least one order position for each product. For each advertisement, the server will select the target order position (i.e., the highest order position) from the at least one order position of the advertisement, and sort the products according to the target order position of each product to obtain the product sequence (i.e., the recommended object sequence).

[0212] In specific applications, Figure 14 The order of products is obtained by using the refined advertisement order shown in the figure as an example. Figure 14 As shown, product ads ad1, ad6, and ad8 advertise product c4; ad3 and ad4 advertise product c2; ad5, ad9, and ad11 advertise product c1; and ad10 advertises product c3. Ad2 and ad7 are standard ads (standard ads lack products). The precision rankings for product ads ad1, ad6, and ad8 are 1, 6, and 8, respectively. The highest precision ranking is 1, so the target ranking for product c4 is 1. Similarly, product c2 is 3, product c1 is 5, and product c3 is 10. Then, sorting the products according to the target ranking yields the following rankings: c4, c2, c1, c3.

[0213] 2. Model Design

[0214] The product recall model adopts the dual-tower structure used in the recall link of the advertising recommendation system, including the user tower and the product tower.

[0215] (1) Input features

[0216] The input features of the user tower are the user's request object data, which can specifically be object attribute features, object behavior statistical features, context features, etc. Among them, object attribute features refer to features used to characterize the user's attributes, which can specifically be the object's location, etc. Object behavior statistical features refer to features used to characterize the user's browsing, purchasing, clicking, and other behaviors on products within the statistical time period. For example, object behavior statistical features can specifically be the identifiers of products browsed, purchased, or clicked by users. Context features refer to features used to characterize the advertising display areas that users have browsed. For example, context features can specifically be the identifiers of applications that display the browsed advertising display areas. Product tower input features are product features (i.e., object attribute information) of products, such as the product identifier, product name, product category, etc.

[0217] (2) Model structure

[0218] The model structure of the product recall model can be as follows Figure 15 As shown in the figure, the user tower input features and the product tower input features are respectively mapped into a high-dimensional dense feature vector through the encoding layer, and then passed through multiple fully connected layers (such as Figure 15 The four fully connected layers are shown in Figure 1), and the feature vectors of the requested object and the recommended object with reduced dimensions are output. Finally, the similarity between the feature vectors of the requested object and the recommended object is calculated through the interoperability layer to obtain the recall prediction probability of the product (e.g. Figure 15 As described above, the dot product can be used to calculate the similarity. It is understandable that when the product recall model is well trained, in the advertisement recall phase, the product recall model outputs the recall probability of the product.

[0219] (3) Loss function design

[0220] Specifically, the server constructs at least one training sample based on the products in the product sequence. Each training sample contains the product features of at least one product in the product sequence. For each training sample, the server determines the sample label of the training sample based on the arrangement order of at least one product in the product sequence indicated by the product features in the training sample. The product recall model is trained based on the user's request object data, at least one training sample, and the sample labels of at least one training sample.

[0221] In a specific application, when training to obtain a product recall model, for each training sample, the server will input the user's request object data into the request object feature extraction network in the initial recall model to obtain a request object feature vector. For each of at least one product included in the training sample, the server will input the object attribute information of the targeted product into the product feature extraction network in the initial recall model to obtain a product feature vector for the targeted product. The request object feature vector and the product feature vector of the targeted product will be input into the interoperability layer in the initial recall model to obtain a recall prediction probability for the targeted product. On this basis, the server will obtain a recall prediction result for the training sample based on the corresponding recall prediction probability of the at least one product. Based on the recall prediction result and sample label of the at least one training sample, the server will calculate the corresponding loss function value of the at least one training sample. The server will use the corresponding loss function value of the at least one training sample to adjust the network parameters of the initial recall model to obtain a product recall model.

[0222] In specific applications, the product recall model in this application adopts a ranking learning algorithm. It can be understood that different ranking learning algorithms require different training samples to be constructed, the method of determining sample labels, and the method of calculating loss function values.

[0223] The following uses the single-point ranking learning method as an example to explain the construction of training samples, the method of determining sample labels, and the steps of training and obtaining a product recall model. If the single-point ranking learning method is used, for each product in the product sequence, the server will obtain the product features of the product and use the product features of the product as training samples. Based on the order of the products in the product sequence, the sample type of the training sample is determined. Based on the sample type of the training sample, the sample label of the training sample is generated. For example, the sample type can be a positive sample or a negative sample. The server can first set the product ranking range corresponding to the positive sample and the product ranking range corresponding to the negative sample, and then determine the sample type of the training sample based on this product ranking range corresponding to the positive sample and the product ranking range corresponding to the negative sample.

[0224] In a specific application, the server can set the product ranking range corresponding to the positive sample to 1 to N, and the product ranking range corresponding to the negative sample to N+1 and beyond. By comparing the product ranking ranges corresponding to the positive and negative samples, and the arrangement order of the products corresponding to the product features in the training samples in the product sequence, the target ranking range to which the product belongs can be determined, and then the sample type corresponding to the target ranking range can be used as the sample type of the training sample.

[0225] In a specific application, taking single-point ranking learning as an example, the constructed loss function can be cross entropy loss, and the specific loss function form can be: , where loss represents the loss function, y i is the sample label. When the training sample is a positive sample, y i =1, when the training sample is a negative sample, y i =0, p i is the recall prediction result of the training sample, that is, the recall prediction probability of a single product, that is, the probability that the training sample is a positive sample.

[0226] In one embodiment, after training the product recall model, the product recall model can be applied to advertising recommendations. In response to an advertisement acquisition request initiated by a user at a terminal, the server determines the user corresponding to the advertisement acquisition request, obtains the user's request object data and the product features of multiple candidate products, and for each candidate product, inputs the user's request object data and the product features of the candidate product into the product recall model to obtain the recall probability of the candidate product. The product recall model is obtained through the above-mentioned product recall model training method. Based on the recall probability of each candidate product, at least one recalled product is selected from the multiple candidate products. For each recalled product, the candidate advertisement indicating the recalled product is used as an advertisement recall. The recalled advertisements are roughly and finely sorted, the target advertisement is determined, and the target advertisement is pushed to the terminal used by the user.

[0227] Among them, in response to the user triggering the opening operation of the application on the terminal, the terminal will initiate an advertising acquisition request to the server, so that the server will respond to the advertising acquisition request, first recall the advertisements, and then perform rough and fine sorting on the recalled advertisements, determine the target advertisements, and push the target advertisements to the terminal used by the user. After receiving the target advertisements, the terminal will display the target advertisements in the advertisement display area of ​​the application.

[0228] It should be noted that the recommendation object recall model training method and the recommendation content recall method of this application are applied to advertisement recall, and a new product advertisement recall method of first product and then advertisement is proposed. At the same time, an advertisement recall method based on fitting the precise ranking of advertisements is proposed. The recall and precise ranking consistency indicators are greatly improved, as shown in Table 1.

[0229] Table 1. Consistency evaluation experiment of recall sorting with different recall methods

[0230]

[0231] The following is an explanation of the recall sorting consistency evaluation indicators:

[0232] (1) GAUC: The first-ranked ad in the recommendation request is used as a positive example, and the non-first-ranked ad is used as a negative example to calculate the AUC (Area Under ROC Curve). AUC represents the probability that the probability of predicting a positive sample as 1 is greater than the probability of predicting a negative sample as 1 when randomly selecting a positive and negative sample pair (a positive and negative sample pair includes a positive sample and a negative sample). The ranking ability is evaluated by aggregation based on the recommendation request. The calculation formula of GAUC is as follows:

[0233] ;

[0234] Among them, GAUC is used to measure the ranking ability of the product recall model for different recommendation requests, AUCi represents the AUC corresponding to the i-th recommendation request, and #pv represents the number of recommendation requests.

[0235] (2) Recall@N_K: The recall ratio of the top K ranked ads (i.e., the top K in the ranked ads sequence) to the top N recalled ads (i.e., the top N recalled ads output by the product recall model). Aggregated by request, the recall capability of the top N recalled ads to the top K ranked ads is evaluated, with K limited to less than N.

[0236] ;

[0237] Among them, Recall represents the recall capability of the top N ranked items relative to the top K ranked items, #pv represents the number of recommendation requests, and i represents the i-th recommendation request.

[0238] It should be understood that, although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence according to the instructions of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above can include multiple steps or multiple stages, and these steps or stages are not necessarily executed and completed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0239] Based on the same inventive concept, the embodiments of the present application also provide a recommendation object recall model training device for implementing the recommendation object recall model training method involved above, and a recommendation content recall device for implementing the recommendation content recall method involved above. The implementation solution provided by the device is similar to the implementation solution described in the above method, so the specific limitations in the embodiments of one or more recommendation object recall model training devices and recommendation content recall devices provided below can be referred to the limitations of the recommendation object recall model training method and recommendation content recall method respectively above, and will not be repeated here.

[0240] In one embodiment, Figure 16 As shown, a recommended object recall model training device is provided, comprising: a recommended record acquisition module 1602, a recommended object sequence generation module 1604, a training sample construction module 1606, a sample label generation module 1608 and a training module 1610, wherein:

[0241] The recommendation record acquisition module 1602 is used to acquire historical recommendation records; the historical recommendation records include a recommendation request object and a recommended content sequence generated for the recommendation request object; the recommended content sequence includes at least two recommended contents;

[0242] The recommended object sequence generation module 1604 is configured to remove duplicates and sort the recommended objects indicated by the at least two recommended contents according to the order in which the at least two recommended contents are arranged in the recommended content sequence, thereby obtaining a recommended object sequence.

[0243] A training sample construction module 1606 is configured to construct at least one training sample based on the recommended objects in the recommended object sequence, wherein each training sample includes object attribute information of at least one recommended object in the recommended object sequence;

[0244] A sample label generation module 1608 is configured to determine, for each training sample, a sample label of the training sample according to the order of arrangement of at least one recommended object in the recommended object sequence indicated by the object attribute information in the training sample;

[0245] The training module 1610 is configured to train and obtain a recommendation object recall model based on the request object data of the recommendation request object, at least one training sample, and a sample label of each of the at least one training sample.

[0246] The above-mentioned recommendation object recall model training device can obtain the recommendation request object and the recommendation content sequence generated for the recommendation request object by obtaining historical recommendation records, and then can deduplicate and sort the recommendation objects indicated by at least two recommendation contents according to the arrangement order of at least two recommendation contents in the recommendation content sequence to obtain a recommendation object sequence. On this basis, at least one training sample can be constructed using the recommendation objects in the recommendation object sequence to realize the construction of the training sample. Each training sample contains the object attribute information of at least one recommendation object in the recommendation object sequence. For each training sample, the sample label of the training sample is determined by determining the arrangement order of at least one recommendation object in the recommendation object sequence indicated by the object attribute information in the training sample. The arrangement order of at least one recommendation object can be used to realize Accurately label the training samples, and then use the request object data of the recommendation request object, at least one training sample and the sample labels of at least one training sample to perform model training, and obtain the recommendation object recall model through training. The whole process first uses the recommendation content sequence to generate the recommendation object sequence, and then uses the object attribute information of the recommended object in the recommendation object sequence to train the recommendation object recall model. On the basis of improving the consistency of the sorting of the recommended objects in the recall link and the recommendation content sequence, a recommendation object recall model that uses the recommendation object as the driving factor for recall can be obtained, which can support improving the recall accuracy of the recommended content in the recall link. Furthermore, by first recalling the recommended object and then recalling the recommended content indicated by the recommended object, the recall accuracy of the recommended content in the recall link can be improved on the basis of fully utilizing the object attribute information of the recommended object.

[0247] In one embodiment, the recommended object sequence generation module is further used to sort the recommended objects indicated by at least two recommended contents according to the arrangement order of the at least two recommended contents in the recommended content sequence, to obtain the arrangement order of the recommended objects indicated by at least two recommended contents, to determine the target sequence position of each recommended object according to the arrangement order of the recommended objects indicated by at least two recommended contents, to sort the recommended objects according to the target sequence position of each recommended object, to obtain a recommended object sequence.

[0248] In one embodiment, the recommended object sequence generation module is also used to determine at least one sequence position for each recommended object based on the arrangement order of the recommended objects indicated by at least two recommended contents, and for each recommended object, select a target sequence position for the recommended object from the at least one sequence position of the recommended object.

[0249] In one embodiment, the training module is also used to input the request object data of the recommended request object into the initial recall model for each training sample, and input the object attribute information of at least one recommended object contained in the training sample into the initial recall model respectively, to obtain the corresponding recall prediction probability of at least one recommended object, and based on the corresponding recall prediction probability of at least one recommended object, to obtain the recall prediction result of the training sample, and according to the recall prediction result and sample label of at least one training sample, to adjust the network parameters of the initial recall model to obtain the recommended object recall model.

[0250] In one embodiment, the training module is also used to input the request object data of the recommended request object into the request object feature extraction network in the initial recall model to obtain the request object feature vector, and for each recommended object in at least one recommended object included in the training sample, input the object attribute information of the recommended object into the recommendation object feature extraction network in the initial recall model to obtain the recommended object feature vector of the recommended object, and input the request object feature vector and the recommended object feature vector of the recommended object into the interoperability layer in the initial recall model to obtain the recall prediction probability of the recommended object.

[0251] In one embodiment, the training sample construction module is further configured to obtain object attribute information of each recommended object in the recommended object sequence, and use the object attribute information of the recommended object as a training sample.

[0252] In one embodiment, the sample label generation module is further used to determine the sample type of the training sample according to the arrangement order of the recommended objects in the recommended object sequence indicated by the object attribute information in the training sample, and generate the sample label of the training sample based on the sample type of the training sample; the sample label is used to identify the sample type of the training sample.

[0253] In one embodiment, the sample labeling generation module is further used to obtain the recommended object ranking range corresponding to at least one sample type, compare the recommended object ranking range corresponding to at least one sample type, and the arrangement order of the recommended objects indicated by the object attribute information in the training sample in the recommended object sequence, determine the target ranking range to which the arrangement order of the recommended objects indicated by the object attribute information in the training sample in the recommended object sequence belongs, and use the sample type corresponding to the target ranking range as the sample type of the training sample.

[0254] In one embodiment, the training sample construction module is also used to combine any two recommended objects in the recommended object sequence to obtain multiple recommended object pairs. For each recommended object pair, the object attribute information of the two recommended objects in the recommended object pair is obtained, and the object attribute information of the two recommended objects in the recommended object pair is used as a training sample.

[0255] In one embodiment, the sample label generation module is also used to determine the recommendation object priority between two recommended objects in a recommended object pair based on the arrangement order of the two recommended objects indicated by the object attribute information in the training sample in the recommended object sequence, and generate a sample label for the training sample based on the recommended object priority; the sample label is used to identify the recommendation object priority.

[0256] In one embodiment, the training sample construction module is further configured to obtain object attribute information of all recommended objects in the recommended object sequence, and use the object attribute information of all recommended objects in the recommended object sequence as training samples.

[0257] In one embodiment, the sample label generation module is further used to generate sample labels for the training samples based on the arrangement order of all recommended objects in the recommended object sequence indicated by the object attribute information in the training samples; the sample labels are used to identify the arrangement order of all recommended objects in the recommended object sequence.

[0258] In one embodiment, Figure 17 As shown, a recommendation object recall model training device is provided, including: a request response module 1702, a recall probability generation module 1704, a recommendation object selection module 1706 and a recommendation content determination module 1708, wherein:

[0259] The request response module 1702 is configured to respond to the recommended content acquisition request, determine the target request object corresponding to the recommended content acquisition request, and acquire the request object data of the target request object and object attribute information of multiple candidate recommended objects;

[0260] The recall probability generation module 1704 is configured to input the target object data and the object attribute information of the candidate recommended object into a recommended object recall model for each candidate recommended object, thereby obtaining the recall probability of the candidate recommended object. The recommended object recall model is obtained using the recommended object recall model training method described above.

[0261] A recommendation object selection module 1706 is configured to select at least one recall recommendation object from a plurality of candidate recommendation objects based on the recall probability of each candidate recommendation object;

[0262] The recommended content determination module 1708 is configured to recall, for each recall recommendation object, the multimedia content indicating the recall recommendation object as the recommended content.

[0263] The above-mentioned recommended content recall device, in response to the recommended content acquisition request, can determine the target request object corresponding to the recommended content acquisition request, and then, based on the obtained request object data of the target request object and the object attribute information of multiple candidate recommended objects, for each candidate recommended object, by inputting the request object data of the target request object and the object attribute information of the candidate recommended object into the recommended object recall model, to obtain the recall probability of the candidate recommended object, so that according to the recall probability of each candidate recommended object, at least one recalled recommended object can be selected from the multiple candidate recommended objects, and for each recalled recommended object, the candidate recommended content indicating the recalled recommended object is recalled as the recommended content. The whole process can improve the recall accuracy of the recommended content in the recall link by first recalling the recommended object and then recalling the recommended content indicated by the recommended object, on the basis of fully utilizing the object attribute information of the recommended object.

[0264] Each module in the aforementioned recommended object recall model training device and recommended content recall device can be implemented in whole or in part through software, hardware, or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in hardware form, or can be stored in a computer device memory in software form, so that the processor can call and execute the corresponding operations of each module.

[0265] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Figure 18 As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O) and a communication interface. The processor, memory and input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database of the computer device is used to store historical recommendation record data. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a recommendation object recall model training method and a recommendation content recall method are implemented.

[0266] Those skilled in the art will understand that Figure 18The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0267] In one embodiment, a computer device is further provided, including a memory and a processor. The memory stores a computer program, and the processor implements the steps in the above method embodiments when executing the computer program.

[0268] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0269] In one embodiment, a computer program product is provided, including a computer program, which implements the steps in the above method embodiments when executed by a processor.

[0270] It should be noted that the requested information (including but not limited to requested data) and data (including but not limited to data used for analysis, storage, and display) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use, and processing of the relevant data must comply with relevant regulations. Furthermore, when used for advertising recommendations, users can reject pushed ads or easily reject ad push information.

[0271] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned embodiments. In particular, any reference to memory, database, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The databases involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchains. The processors involved in the various embodiments provided herein may be, but are not limited to, general-purpose processors, central processing units (CPUs), graphics processing units (GPUs), digital signal processors (DSPs), programmable logic devices (PLDs), data processing logic devices based on quantum computing, and the like.

[0272] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0273] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.

Claims

1. A method for training a recommendation object recall model, characterized in that: The method comprises: Acquire historical recommendation records; the historical recommendation records include a recommendation request object and a recommended content sequence generated for the recommendation request object; the recommended content sequence includes at least two recommended contents; Deduplicating and sorting the recommended objects indicated by the at least two recommended contents according to the arrangement order of the at least two recommended contents in the recommended content sequence to obtain a recommended object sequence; constructing at least one training sample according to the recommended objects in the recommended object sequence, each of the training samples comprising object attribute information of at least one recommended object in the recommended object sequence; For each of the training samples, determining a sample label of the training sample according to an arrangement order of at least one recommended object indicated by the object attribute information in the training sample in the recommended object sequence; A recommendation object recall model is trained based on the request object data of the recommendation request object, the at least one training sample, and the sample label of each of the at least one training sample.

2. The method according to claim 1, characterized in that The step of removing duplicates and sorting the recommended objects indicated by the at least two recommended contents according to the arrangement order of the at least two recommended contents in the recommended content sequence to obtain the recommended object sequence includes: sorting the recommended objects indicated by the at least two recommended contents according to the arrangement order of the at least two recommended contents in the recommended content sequence to obtain the arrangement order of the recommended objects indicated by the at least two recommended contents; Determining a target ranking of each of the recommended objects according to the arrangement order of the recommended objects indicated by the at least two recommended contents; The recommended objects are sorted according to the target order of each recommended object to obtain a recommended object sequence.

3. The method according to claim 2, characterized in that The determining, based on the arrangement order of the recommended objects indicated by the at least two recommended contents, a target order ranking of each recommended object comprises: Determining at least one order rank of each of the recommended objects according to the arrangement order of the recommended objects indicated by the at least two recommended contents; For each of the recommended objects, a target sequence rank of the recommended object is selected from at least one sequence rank of the recommended objects.

4. The method according to claim 1, wherein The training of a recommendation object recall model based on the request object data of the recommendation request object, the at least one training sample, and a sample label of each of the at least one training sample includes: For each of the training samples, inputting the request object data of the recommendation request object into an initial recall model, and inputting the object attribute information of at least one recommended object included in the training sample into the initial recall model respectively, to obtain the corresponding recall prediction probability of the at least one recommended object; Obtaining a recall prediction result of the training sample based on the recall prediction probability corresponding to each of the at least one recommended object; According to the recall prediction result and sample label of each of the at least one training sample, the network parameters of the initial recall model are adjusted to obtain a recommended object recall model.

5. The method according to claim 4, characterized in that Inputting the request object data of the recommendation request object into the initial recall model, and inputting the object attribute information of at least one recommended object included in the training sample into the initial recall model respectively, to obtain the corresponding recall prediction probability of the at least one recommended object includes: Inputting the request object data of the recommended request object into the request object feature extraction network in the initial recall model to obtain a request object feature vector; For each recommended object among at least one recommended object included in the training sample, inputting object attribute information of the recommended object into a recommended object feature extraction network in the initial recall model to obtain a recommended object feature vector for the recommended object; The requested object feature vector and the recommended object feature vector of the targeted recommended object are input into the interoperation layer of the initial recall model to obtain the predicted recall probability of the targeted recommended object.

6. The method according to any one of claims 1 to 5, characterized in that The step of constructing at least one training sample based on the recommended objects in the recommended object sequence, wherein each training sample contains object attribute information of at least one recommended object in the recommended object sequence, comprises: For each recommended object in the recommended object sequence, object attribute information of the recommended object is obtained, and the object attribute information of the recommended object is used as a training sample.

7. The method according to claim 6, characterized in that Determining the sample label of the training sample according to the arrangement order of at least one recommended object indicated by the object attribute information in the training sample in the recommended object sequence includes: determining a sample type of the training sample according to an arrangement order of the recommended objects indicated by the object attribute information in the training sample in the recommended object sequence; Based on the sample type of the training sample, a sample label of the training sample is generated; the sample label is used to identify the sample type of the training sample.

8. The method according to claim 7, characterized in that The determining the sample type of the training sample according to the arrangement order of the recommended objects indicated by the object attribute information in the training sample in the recommended object sequence includes: Obtaining a ranking range of recommended objects corresponding to at least one sample type; Comparing the recommended object ranking range corresponding to each of the at least one sample type and the arrangement order of the recommended objects indicated by the object attribute information in the training sample in the recommended object sequence, determining a target ranking range to which the arrangement order of the recommended objects indicated by the object attribute information in the training sample in the recommended object sequence belongs; The sample type corresponding to the target ranking range is used as the sample type of the training sample.

9. The method according to any one of claims 1 to 5, characterized in that The step of constructing at least one training sample based on the recommended objects in the recommended object sequence, wherein each training sample contains object attribute information of at least one recommended object in the recommended object sequence, comprises: Combining any two recommended objects in the recommended object sequence to obtain multiple recommended object pairs; For each of the recommended object pairs, object attribute information of the two recommended objects in the recommended object pair is obtained, and the object attribute information of the two recommended objects in the recommended object pair is used as a training sample.

10. The method according to claim 9, characterized in that Determining the sample label of the training sample according to the arrangement order of at least one recommended object indicated by the object attribute information in the training sample in the recommended object sequence includes: determining a recommendation object priority between two recommended objects in the recommended object pair according to an arrangement order of the two recommended objects indicated by the object attribute information in the training sample in the recommended object sequence; Based on the recommendation object priority, a sample label of the training sample is generated; the sample label is used to identify the recommendation object priority.

11. The method according to any one of claims 1 to 5, characterized in that: The step of constructing at least one training sample based on the recommended objects in the recommended object sequence, wherein each training sample contains object attribute information of at least one recommended object in the recommended object sequence, comprises: Obtaining object attribute information of all recommended objects in the recommended object sequence; The object attribute information of all recommended objects in the recommended object sequence is used as training samples.

12. The method according to claim 11, characterized in that Determining the sample label of the training sample according to the arrangement order of at least one recommended object indicated by the object attribute information in the training sample in the recommended object sequence includes: A sample label of the training sample is generated according to the arrangement order of all recommended objects in the recommended object sequence indicated by the object attribute information in the training sample; the sample label is used to identify the arrangement order of all recommended objects in the recommended object sequence.

13. A method for recalling recommended content, characterized in that: The method comprises: In response to a recommended content acquisition request, determining a target request object corresponding to the recommended content acquisition request, and acquiring request object data of the target request object and object attribute information of a plurality of candidate recommended objects; For each candidate recommended object, the request object data of the target request object and the object attribute information of the candidate recommended object are input into a recommended object recall model to obtain a recall probability of the candidate recommended object; the recommended object recall model is obtained by the recommended object recall model training method according to any one of claims 1 to 12; Selecting at least one recall recommendation object from the multiple candidate recommendation objects according to the recall probability of each candidate recommendation object; For each of the recalled recommendation objects, candidate recommendation content indicating the recalled recommendation object is recalled as recommended content.

14. A recommendation object recall model training device, characterized in that: The device comprises: A recommendation record acquisition module, configured to acquire historical recommendation records; the historical recommendation records include a recommendation request object and a recommendation content sequence generated for the recommendation request object; the recommendation content sequence includes at least two recommended contents; a recommendation object sequence generation module, configured to remove duplicates and sort the recommendation objects indicated by the at least two recommended contents according to the arrangement order of the at least two recommended contents in the recommendation content sequence, to obtain a recommendation object sequence; a training sample construction module, configured to construct at least one training sample based on the recommended objects in the recommended object sequence, each of the training samples comprising object attribute information of at least one recommended object in the recommended object sequence; a sample label generation module, configured to determine, for each of the training samples, a sample label of the training sample according to an arrangement order of at least one recommended object indicated by object attribute information in the training sample in the recommended object sequence; The training module is used to train and obtain a recommendation object recall model based on the request object data of the recommendation request object, the at least one training sample and the sample label of each of the at least one training sample.

15. A recommended content recall device, characterized in that: The device comprises: a request response module, configured to respond to a recommendation content acquisition request, determine a target request object corresponding to the recommendation content acquisition request, and acquire request object data of the target request object and object attribute information of a plurality of candidate recommendation objects; a recall probability generation module, configured to input the request object data of the target request object and the object attribute information of the candidate recommendation object into a recommendation object recall model for each candidate recommendation object, to obtain a recall probability of the candidate recommendation object; the recommendation object recall model is obtained by the recommendation object recall model training method according to any one of claims 1 to 12; A recommendation object selection module, configured to select at least one recall recommendation object from the plurality of candidate recommendation objects according to the recall probability of each candidate recommendation object; The recommended content determination module is configured to recall, for each of the recalled recommendation objects, the multimedia content indicating the recalled recommendation object as the recommended content.

16. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 13 are implemented.

17. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 13 are implemented.

18. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 13 are implemented.

Images