Content recommendation method, model training method and related products

Abstract

The invention discloses a content recommendation method, a model training method and a related product, and the content recommendation method comprises the steps: obtaining a historical behavior label aggregation vector corresponding to a first object when a content recommendation request of the first object is received; performing interest label prediction processing on the first object according to the historical behavior label aggregation vector and an interest label vector matrix to obtain a predicted interest label; the prediction interest tag comprises a preferred prediction interest tag and an alternative prediction interest tag; generating a content recommendation candidate set for the content recommendation request according to the preferred prediction interest tag, performing sorting optimization processing on the content recommendation candidate set according to the alternative prediction interest tag to obtain an optimized content recommendation set, and sending the optimized content recommendation set to a first terminal corresponding to the first object, the optimized content recommendation set is displayed by the first terminal in sequence.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of natural language processing, and particularly relates to a content recommendation method, a model training method and related products. BACKGROUND

[0002] In a content recommendation system, the content recommended to a user is usually accompanied by different labeled representations. These labels can help the system better understand the user's interests and thus provide more personalized recommendations. In the prior art, a recommendation system can collect historical behavior data of a user, determine content labels that the user is interested in from the content labels browsed by the user, as historical interest labels of the user, and then recommend content corresponding to the content labels consistent with the historical interest labels to the user in the next content recommendation. This improves the interest of the user in the recommended content to a certain extent, but also leads to the recommended content being limited to the historical interest labels of the user, lacking diversity. SUMMARY

[0003] The present application provides a content recommendation method, a model training method and related products, wherein the related products include a content recommendation device, a model training device, an electronic device, a computer readable storage medium and a computer program product.

[0004] In a first aspect, a content recommendation method is provided, comprising:

[0005] When a content recommendation request of a first object is received, a historical behavior label aggregation vector corresponding to the first object is obtained;

[0006] According to the historical behavior label aggregation vector and an interest label vector matrix, interest label prediction processing is performed on the first object to obtain predicted interest labels; the interest label vector matrix is determined based on a label vector representation model; the label vector representation model is obtained by training an initial vector representation model based on preference interest labels of at least two sample objects;

[0007] According to the first preferred predicted interest label, a content recommendation candidate set for the content recommendation request is generated, the content recommendation candidate set is subjected to sorting optimization processing according to the alternative predicted interest labels, an optimized content recommendation set is obtained, and the optimized content recommendation set is sent to a first terminal corresponding to the first object, so that the first terminal displays the optimized content recommendation set in sequence.

[0008] According to any one of the embodiments of the present application, the preference interest label is determined according to historical content interaction behaviors of the sample object on a content library; each content label corresponding to each content in the content library belongs to a content label bag; the interest label vector matrix is used to represent a vector representation of each content label in the content label bag; and the predicted interest label includes a preferred predicted interest label and an alternative predicted interest label.

[0009] According to any one of the embodiments of the present application, the interest label prediction processing on the first object based on the historical behavior label aggregation vector and the interest label vector matrix to obtain a predicted interest label includes:

[0010] performing vector neighbor search processing on the historical behavior label aggregation vector based on the interest label vector matrix to obtain a neighbor search label sequence corresponding to the historical behavior label aggregation vector; in the neighbor search label sequence, a neighbor search label at a position closer to the front has a closer distance between a neighbor search label vector corresponding to the neighbor search label and the historical behavior label aggregation vector;

[0011] determining a first neighbor search label in the neighbor search label sequence as a preferred predicted interest label;

[0012] from the remaining neighbor search label sequence, obtaining neighbor search labels with a preset interest generalization number as alternative predicted interest labels; the remaining neighbor search label sequence refers to a neighbor search label sequence excluding the preferred predicted interest label;

[0013] determining the preferred predicted interest label and the alternative predicted interest label as the predicted interest label.

[0014] According to any one of the embodiments of the present application, the vector neighbor search processing on the historical behavior label aggregation vector based on the interest label vector matrix to obtain a neighbor search label sequence corresponding to the historical behavior label aggregation vector includes:

[0015] respectively determining a vector Euclidean distance between each interest label vector in the interest label vector matrix and the historical behavior label aggregation vector;

[0016] performing sorting processing on interest labels corresponding to each interest label vector according to the vector Euclidean distance to obtain a sorted interest label sequence; a vector Euclidean distance corresponding to a kth interest label in the sorted interest label sequence is smaller than a vector Euclidean distance corresponding to a k+1th interest label, and k is a positive integer;

[0017] from the sorted interest label sequence, according to a preset search number N, sequentially obtaining N interest labels, and determining the N interest labels as a neighbor search label sequence corresponding to the historical behavior label aggregation vector, and N is a positive integer.

[0018] In combination with any of the embodiments of the present application, the preset interest generalization quantity includes a preset historical interest quantity S and a preset extended interest quantity H, S and H are positive integers;

[0019] The obtaining of the neighbor search label with the preset interest generalization quantity from the remaining neighbor search label sequence as the candidate preset interest label includes:

[0020] The historical behavior label set corresponding to the first object is obtained;

[0021] The historical neighbor search label sequence is obtained by sequentially obtaining the neighbor search label same as the historical behavior label in the historical behavior label set from the neighbor search label sequence;

[0022] The S neighbor search labels are sequentially obtained from the historical nearest neighbor search label sequence, and the S neighbor search labels are determined as historical neighbor search labels;

[0023] The extended neighbor search label sequence is obtained by sequentially obtaining the neighbor search label different from the historical behavior label in the historical behavior label set from the neighbor search label sequence;

[0024] The H neighbor search labels are sequentially obtained from the extended neighbor search label sequence, and the H neighbor search labels are determined as extended neighbor search labels;

[0025] The historical neighbor search label and the extended neighbor search label are taken as candidate prediction interest labels.

[0026] In combination with any of the embodiments of the present application, the sorting and optimization processing of the content recommendation candidate set according to the candidate prediction interest label to obtain an optimized content recommendation set includes:

[0027] The label hit rate between the content label corresponding to each recommendation content in the content recommendation candidate set and the candidate prediction interest label is determined respectively;

[0028] The position adjustment processing of each recommendation content in the content recommendation candidate set is performed according to the label hit rate corresponding to each recommendation content in the content recommendation candidate set, to obtain an optimized content recommendation set; in the optimized content recommendation set, the label hit rate corresponding to the recommendation content in the front position is higher than the label hit rate corresponding to the recommendation content in the rear position.

[0029] In combination with any of the embodiments of the present application, the obtaining of the historical behavior label aggregation vector corresponding to the first object includes:

[0030] The historical behavior label set corresponding to the first object is obtained;

[0031] obtaining a historical behavior label vector corresponding to a historical behavior label in the set of historical behavior labels from a matrix of interest label vectors;

[0032] performing cumulative average processing on the historical behavior label vector to obtain a historical behavior label aggregation vector corresponding to the first object.

[0033] In a second aspect, a model training method is provided, including:

[0034] determining a preferred interest label of at least two sample objects according to historical content interaction behaviors of the at least two sample objects on a content library; each content in the content library corresponds to a content label belonging to a content label bag; the preferred interest label belongs to the content label bag;

[0035] constructing an initial vector representation model for the content label bag; the initial vector representation model contains a default initial vector representation matrix; the vector representation matrix is used to represent a vector representation corresponding to each content label in the content label bag;

[0036] training the initial vector representation model according to the preferred interest label of the at least two sample objects to obtain a label vector representation model, and determining a vector representation matrix contained in the label vector representation model as an interest label vector matrix; the interest label vector matrix is used to perform interest label prediction processing on a first object.

[0037] According to any one of the embodiments of the present application, the at least two sample objects include a target sample object;

[0038] determining a preferred interest label of at least two sample objects according to historical content interaction behaviors of the at least two sample objects on a content library, including:

[0039] determining a content dimension interaction score of the target sample object on the content library according to historical content interaction behaviors of the target sample object on the content library;

[0040] dividing the content library according to a content label bag to obtain a label dimension content library corresponding to each label in the content label bag;

[0041] determining a label dimension interaction score of the target sample object on each label dimension content library according to historical content interaction behaviors of the target sample object on each label dimension content library;

[0042] performing self-smoothing processing on the label dimension interaction score of each label dimension content library according to the content dimension interaction score to obtain a label preference score of the target sample object on each label in the content label bag;

[0043] According to the label preference score, the labels in the content label bag-of-words are ranked to obtain a ranked label sequence, and from the ranked label sequence, a preset number of labels are obtained in sequence as the preferred interest labels of the target sample object.

[0044] According to any one of the embodiments of the present application, the initial vector representation model is trained according to the preferred interest labels of the at least two sample objects to obtain a label vector representation model, including:

[0045] The preferred interest labels of the at least two sample objects are respectively subjected to label division processing to obtain predicted preferred interest labels and real preferred interest labels of the at least two sample objects;

[0046] The predicted preferred interest labels of the at least two sample objects are input into the initial vector representation model to obtain predicted preferred interest labels of the at least two sample objects;

[0047] According to the predicted preferred interest labels and the real preferred interest labels, a model loss degree corresponding to the initial vector representation model is determined, and the initial vector representation model is subjected to model parameter adjustment processing based on the model loss degree to obtain a new vector representation model;

[0048] If the model loss degree is greater than a model loss threshold, the predicted preferred interest labels of the at least two sample objects are continuously input into the new vector representation model;

[0049] If the model loss degree is less than or equal to the model loss threshold, the new vector representation model is determined as a label vector representation model.

[0050] In a third aspect, a content recommendation device is provided, including:

[0051] An acquisition unit is configured to acquire a historical behavior label aggregation vector corresponding to a first object when receiving a content recommendation request of the first object.

[0052] A prediction unit is configured to perform interest label prediction processing on the first object according to the historical behavior label aggregation vector and an interest label vector matrix to obtain a predicted interest label; the interest label vector matrix is determined based on a label vector representation model; and the label vector representation model is obtained by training an initial vector representation model based on preferred interest labels of at least two sample objects.

[0053] The sorting optimization unit is configured to generate a content recommendation candidate set for the content recommendation request according to the preferred predicted interest label, perform sorting optimization processing on the content recommendation candidate set according to the alternative predicted interest label, obtain an optimized content recommendation set, and send the optimized content recommendation set to the first terminal corresponding to the first object, so that the first terminal sequentially displays the optimized content recommendation set.

[0054] According to any one of the embodiments of the present application, the preference interest label is determined according to the historical content interaction behavior of the sample object on the content library; each content label in the content library belongs to a content label bag; the interest label vector matrix is used to represent the vector representation of each content label in the content label bag; and the predicted interest label includes a preferred predicted interest label and an alternative predicted interest label.

[0055] According to any one of the embodiments of the present application, the interest label prediction unit is specifically configured to:

[0056] perform vector neighbor search processing on the historical behavior label aggregated vector based on the interest label vector matrix to obtain a neighbor search label sequence corresponding to the historical behavior label aggregated vector; in the neighbor search label sequence, the neighbor search label at a more front position has a closer distance between the neighbor search label vector corresponding to the neighbor search label and the historical behavior label aggregated vector;

[0057] determine the first neighbor search label in the neighbor search label sequence as the preferred predicted interest label;

[0058] obtain neighbor search labels with a preset interest generalization number from the remaining neighbor search label sequence as alternative predicted interest labels; the remaining neighbor search label sequence refers to the neighbor search label sequence excluding the preferred predicted interest label;

[0059] determine the preferred predicted interest label and the alternative predicted interest label as the predicted interest label.

[0060] According to any one of the embodiments of the present application, the interest label vector matrix is used to represent the vector representation of each content label in the content label bag; the predicted interest label includes a preferred predicted interest label and an alternative predicted interest label.

[0061] determine the vector Euclidean distance between each interest label vector in the interest label vector matrix and the historical behavior label aggregated vector, respectively;

[0062] According to the vector Euclidean distance, each interest label corresponding to the interest label vector is sorted to obtain a sorted interest label sequence; a vector Euclidean distance corresponding to a kth interest label in the sorted interest label sequence is less than a vector Euclidean distance corresponding to a (k+1)th interest label, k is a positive integer;

[0063] From the sorted interest label sequence, N interest labels are sequentially obtained according to a preset search number N, the N interest labels are determined as a neighbor search label sequence corresponding to the historical behavior label aggregation vector, N is a positive integer.

[0064] According to any one of the embodiments of the present application, the preset interest generalization number includes a preset historical interest number S and a preset extended interest number H, S and H are positive integers;

[0065] The neighbor search label with the preset interest generalization number is obtained from the remaining neighbor search label sequence as a candidate preset interest label, and the prediction unit is specifically configured to:

[0066] Obtain a historical behavior label set corresponding to the first object;

[0067] From the neighbor search label sequence, a neighbor search label that is the same as a historical behavior label in the historical behavior label set is sequentially obtained to obtain a historical neighbor search label sequence;

[0068] From the historical nearest neighbor search label sequence, S neighbor search labels are sequentially obtained, and the S neighbor search labels are determined as historical neighbor search labels;

[0069] From the neighbor search label sequence, a neighbor search label that is not the same as a historical behavior label in the historical behavior label set is sequentially obtained to obtain an extended neighbor search label sequence;

[0070] From the extended neighbor search label sequence, H neighbor search labels are sequentially obtained, and the H neighbor search labels are determined as extended neighbor search labels;

[0071] The historical neighbor search label and the extended neighbor search label are used as candidate prediction interest labels.

[0072] According to any one of the embodiments of the present application, the content recommendation candidate set is sorted and optimized according to the candidate prediction interest label to obtain an optimized content recommendation set, and the sorting optimization unit is specifically configured to:

[0073] Determine a label hit rate between a content label corresponding to each recommendation content in the content recommendation candidate set and the candidate prediction interest label, respectively;

[0074] According to a label hit rate corresponding to each recommended content in the content recommendation candidate set, each recommended content in the content recommendation candidate set is subjected to position adjustment processing, and an optimized content recommendation set is obtained; in the optimized content recommendation set, a label hit rate corresponding to a recommended content in a front position is higher than a label hit rate corresponding to a recommended content in a rear position.

[0075] According to any one of the embodiments of the present application, the acquisition unit is specifically configured to:

[0076] Acquire a historical behavior label set corresponding to the first object;

[0077] Acquire a historical behavior label vector corresponding to a historical behavior label in the historical behavior label set from an interest label vector matrix;

[0078] Perform cumulative average processing on the historical behavior label vector to obtain a historical behavior label aggregation vector corresponding to the first object.

[0079] In a fourth aspect, a model training apparatus is provided, comprising:

[0080] A determination unit configured to determine a preferred interest label of at least two sample objects according to historical content interaction behaviors of the at least two sample objects on a content library; each content in the content library corresponds to a content label belonging to a content label bag; the preferred interest label belongs to the content label bag;

[0081] A construction unit configured to construct an initial vector representation model for the content label bag; the initial vector representation model contains a default initial vector representation matrix; the vector representation matrix is used to represent a vector representation corresponding to each content label in the content label bag;

[0082] A training unit configured to train the initial vector representation model according to the preferred interest labels of the at least two sample objects to obtain a label vector representation model, and determine a vector representation matrix contained in the label vector representation model as an interest label vector matrix; the interest label vector matrix is used to perform interest label prediction processing on a first object.

[0083] According to any one of the embodiments of the present application, the at least two sample objects include a target sample object;

[0084] The determination unit is specifically configured to:

[0085] determine a content dimension interaction score of the target sample object to the content library according to historical content interaction behaviors of the target sample object to the content library;

[0086] divide the content library according to content label bags to obtain a label dimension content library corresponding to each label in the content label bags;

[0087] determine a label dimension interaction score of the target sample object to each label dimension content library according to historical content interaction behaviors of the target sample object to each label dimension content library;

[0088] perform self-smoothing processing on the label dimension interaction score of each label dimension content library according to the content dimension interaction score to obtain a label preference score of the target sample object to each label in the content label bags;

[0089] sort the labels in the content label bags according to the label preference score to obtain a sorted label sequence, and obtain, in sequence, a preset number of labels from the sorted label sequence as the preference interest labels of the target sample object.

[0090] According to any one of the embodiments of the present application, the training unit is specifically configured to:

[0091] perform label division processing on the preference interest labels of the at least two sample objects respectively to obtain predicted preference interest labels and real preference interest labels of the at least two sample objects;

[0092] input the predicted preference interest labels of the at least two sample objects into the initial vector representation model to obtain predicted preference interest labels of the at least two sample objects;

[0093] determine a model loss degree corresponding to the initial vector representation model according to the predicted preference interest labels and the real preference interest labels, and perform model parameter adjustment processing on the initial vector representation model based on the model loss degree to obtain a new vector representation model;

[0094] if the model loss degree is greater than a model loss threshold, continue to input the predicted preference interest labels of the at least two sample objects into the new vector representation model;

[0095] if the model loss degree is less than or equal to the model loss threshold, determine the new vector representation model as a label vector representation model.

[0096] In a fifth aspect, an electronic device is provided, comprising: a processor and a memory, the memory being configured to store computer program code comprising computer instructions which, when executed by the processor, cause the electronic device to perform the method of the first aspect and any one of its implementations, or the method of the second aspect.

[0097] In a sixth aspect, another electronic device is provided, comprising: a processor, a sending apparatus, an input apparatus, an output apparatus and a memory, the memory being configured to store computer program code comprising computer instructions which, when executed by the processor, cause the electronic device to perform the method of the first aspect and any one of its implementations, or the method of the second aspect.

[0098] In a seventh aspect, a computer readable storage medium is provided, the computer readable storage medium having stored thereon a computer program comprising program instructions that, when executed by a processor, cause the processor to perform the method of the first aspect and any one of its implementations, or the method of the second aspect.

[0099] In an eighth aspect, a computer program product is provided, the computer program product comprising a computer program or instructions which, when run on a computer, cause the computer to perform the method of the first aspect and any one of its implementations, or the method of the second aspect.

[0100] It should be understood that the general description above and the following detailed description are only exemplary and explanatory, rather than restrictive of the application.

[0101] In the present application, the content recommendation device can obtain a historical behavior label aggregation vector corresponding to the first object when receiving a content recommendation request of the first object. Then, the first object is subjected to interest label prediction processing according to the historical behavior label aggregation vector and an interest label vector matrix, to obtain a predicted interest label. The interest label vector matrix is determined based on a label vector representation model. The label vector representation model is obtained by training an initial vector representation model based on the preferred interest labels of at least two sample objects. The predicted interest label includes a preferred predicted interest label and an alternative predicted interest label. Finally, a content recommendation candidate set for the content recommendation request is generated according to the preferred predicted interest label, the content recommendation candidate set is subjected to sorting optimization processing according to the alternative predicted interest label, to obtain an optimized content recommendation set, and the optimized content recommendation set is sent to a first terminal corresponding to the first object, so that the first terminal displays the optimized content recommendation set in sequence. The present application embodiment can generate an interest label vector matrix for a content label bag based on the preferred interest labels of at least two sample objects. The closer the vector representation of the content label in the matrix, the more likely the user is interested in it. After determining the historical behavior label aggregation vector of the user, the interest label vector matrix and the historical behavior label aggregation vector can be used to explore the interest of the user, to obtain a plurality of predicted interest labels that the user may be interested in. These predicted interest labels will no longer be limited to the content labels that the user has browsed in the past, and may also be content labels that the user has not browsed. In this way, the diversity of the content recommended to the user can be enhanced without affecting the user's interest. BRIEF DESCRIPTION OF DRAWINGS

[0102] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the background art, the drawings required to be used in the embodiments of the present application or the background art will be described below.

[0103] The drawings herein are incorporated into the specification and form part of the specification, which show embodiments consistent with the present application, and together with the specification, serve to illustrate the technical solutions of the present application.

[0104] Figure 1 A network architecture schematic diagram is provided for the embodiments of the present application.

[0105] Figure 2 An application scenario schematic diagram of a content recommendation method is provided for the embodiments of the present application.

[0106] Figure 3 A flowchart of a content recommendation method is provided for the embodiments of the present application.

[0107] Figure 4 A flowchart of a model training method is provided for the embodiments of the present application.

[0108] Figure 5 A schematic diagram of initial vector representation model training is provided for an embodiment of the present application.

[0109] Figure 6 A structural schematic diagram of a content recommendation device is provided for an embodiment of the present application.

[0110] Figure 7 A structural schematic diagram of a model training device is provided for an embodiment of the present application.

[0111] Figure 8 A hardware structural schematic diagram of an electronic device is provided for an embodiment of the present application. DETAILED DESCRIPTION

[0112] In order to enable persons skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by persons skilled in the art without creative work fall within the scope of protection of the present application.

[0113] The terms “first”, “second”, and the like in the specification and claims of the present application and the above-mentioned drawings are used to distinguish different objects, and are not used to describe a specific order. In addition, the terms “include” and “have” and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but can optionally include steps or units not listed, or can optionally include other steps or units inherent to the process, method, product or device.

[0114] In this document, the term “embodiment” means that a specific feature, structure, or characteristic described in connection with the embodiment can be included in at least one embodiment of the present application. The phrase appears in various places in the specification does not necessarily refer to the same embodiment, nor is it mutually exclusive or alternative to other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0115] The embodiments of the present application will be described below with reference to the drawings in the embodiments of the present application.

[0116] Please refer to Figure 1 , Figure 1 A structural schematic diagram of a network architecture is provided for an embodiment of the present application. As shown in Figure 1As shown, the network architecture can include a server 2000 and a terminal cluster. The terminal cluster can specifically include one or more terminals, and the number of terminals in the terminal cluster will not be limited here. As shown in the figure, Figure 1 As shown, the plurality of terminals can specifically include a terminal 3000a, a terminal 3000b, a terminal 3000c, …, and a terminal 3000n. The terminal 3000a, the terminal 3000b, the terminal 3000c, …, and the terminal 3000n can be directly or indirectly connected to the server 2000 through wired or wireless communication, so that each terminal can interact with the server 2000 through the network connection. It can be understood that the above network connection is not limited to the connection mode, and can be directly or indirectly connected through wired communication, or directly or indirectly connected through wireless communication, or connected through other connection modes, which are not limited in the present application.

[0117] Each terminal in the terminal cluster can include a smartphone, a tablet computer, a notebook computer, a desktop computer, a smart voice interaction device, a smart home appliance (for example, a smart television), a wearable device, a vehicle-mounted terminal, an aircraft, and other smart terminals with data processing functions. It should be understood that, Figure 1 As shown, each terminal in the terminal cluster can be installed with an application client. When the application client runs in each terminal, it can respectively interact with the server 2000 as described above. Figure 1 As shown, the server 2000 can be an application server corresponding to the application client. The server 2000 can be a physical server, a server cluster or a distributed system composed of multiple physical servers, a cloud server providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms, etc. Basic cloud computing services.

[0118] As shown, the server 2000 can be an application server corresponding to the application client. The server 2000 can be a physical server, a server cluster or a distributed system composed of multiple physical servers, a cloud server providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms, etc. Basic cloud computing services.

[0119] As shown, the server 2000 can be an application server corresponding to the application client. The server 2000 can be a physical server, a server cluster or a distributed system composed of multiple physical servers, a cloud server providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms, etc. Basic cloud computing services. Figure 1 As shown, the server 2000 can be an application server corresponding to the application client. The server 2000 can be a physical server, a server cluster or a distributed system composed of multiple physical servers, a cloud server providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms, etc. Basic cloud computing services.

[0120] For ease of understanding, the embodiments of the present application can be in the multiple terminals shown in Figure 1 select one terminal as the first terminal. For example, the embodiments of the present application can select the terminal 3000a shown in Figure 1 as the first terminal, and the application client with data processing function can be installed in the first terminal. At this time, the first terminal can realize data interaction between the application client and the server 2000. In addition, the user corresponding to the first terminal can be referred to as the first object.

[0121] As shown in Figure 1 , the first terminal can generate a content recommendation request through the running application client, and then send the content recommendation request to the server 2000. After receiving the content recommendation request, the server 2000 can obtain the historical behavior label aggregation vector corresponding to the first object; then perform interest label prediction processing on the first object according to the historical behavior label aggregation vector and the interest label vector matrix, to obtain the predicted interest label; wherein the interest label vector matrix is determined based on the label vector representation model; the label vector representation model is obtained by training the initial vector representation model based on the preference interest labels of at least two sample objects; the predicted interest label includes the preferred predicted interest label and the alternative predicted interest label; finally, a content recommendation candidate set for the content recommendation request is generated according to the preferred predicted interest label, the content recommendation candidate set is sorted and optimized according to the alternative predicted interest label, to obtain the optimized content recommendation set, and the optimized content recommendation set is sent to the first terminal corresponding to the first object, so that the first terminal displays the optimized content recommendation set in sequence.

[0122] For ease of understanding, further, please refer to Figure 2 , Figure 2 application scenario of the content recommendation method provided by the embodiments of the present application. As shown in Figure 2 , the server 20a can be the server 2000 in the above-mentioned Figure 1 corresponding embodiments, and as shown in Figure 2 , the terminal 20b can be the first terminal in the above-mentioned Figure 1 corresponding embodiments, and the application client can be installed in the terminal 20b, and the user corresponding to the terminal 20b can be the object 20c. For ease of understanding, the embodiments of the present application take the content recommendation method executed by the server 20a as an example for description.

[0123] As shown in Figure 2 , assuming that the application corresponding to the application client is a note application, and the object 20c wants to pass the time by browsing some notes through the application client, the object 20c can generate a content recommendation request through the application client in the terminal 20b, and the terminal 20b can send the content recommendation request to the server 20a (as shown in S211a in Figure 2 ).

[0124] As shown in Figure 2 , after receiving the content recommendation request, the server 20a can obtain the historical behavior tag set 201a corresponding to the object 20c, which can contain some tags corresponding to the notes browsed by the object 20c in the historical time period, such as the historical behavior tag set 201a containing the two content tags of “avatar sharing interaction” and “anime fan fiction creation”. Then, the server 20a can generate the historical behavior tag aggregation vector 202a corresponding to the historical behavior tag set 201a (as shown in S212a in Figure 2 ), which is used to comprehensively reflect the historical content tags of interest of the object 20c. Then, the server 20a can perform interest tag prediction processing on the object 20c based on the historical behavior tag aggregation vector 202a and the interest tag vector matrix 203a to obtain the predicted interest tag set 204a (as shown in S213a in Figure 2 ). The interest tag vector matrix 203a is a weight matrix in a label vector representation model, which is obtained by training an initial vector representation model based on the preference interest tags of at least two sample objects. As shown in Figure 2 , the predicted interest tag set 204a can contain content tags that the object 20c is likely to be interested in, such as “game fan fiction creation”, “anime sharing interaction”, “rhythm game”, etc. Then, the server 20a can generate an optimized content recommendation set 205a according to the predicted interest tag set 204a (as shown in S214a in Figure 2 ), and return the optimized content recommendation set 205a to the terminal 20b (as shown in S215a in Figure 3 ), and the terminal 20b will display the notes A1, A2, …, An in order according to the order of the notes in the optimized content recommendation set 205a. The specific process of the interest tag prediction processing in step S213a and the generation of the optimized content recommendation set in step S214a can be referred to the specific description in the embodiments described below. Figure 3

[0125] It can be seen that the content recommendation method provided by the embodiments of the present application can generalize and expand the content tags of interest of the user to a certain extent, avoid information cocooning, and avoid excessive narrowing of the content recommended to the user.

[0126] ​Further, the embodiment of the present application provides a content recommendation method and a model training method. The execution subject of the content recommendation method is a content recommendation device. The content recommendation device can be any electronic device capable of executing the technical solutions disclosed in the method embodiments of the present application. Optionally, the content recommendation device can be one of a computer and a server. It should be understood that the method embodiments of the present application can also be implemented by a processor executing computer program codes. The content recommendation method provided by the embodiment of the present application is used to determine the content recommended for a first object by using an interest label vector matrix in a label vector representation model, wherein the label vector representation model is obtained according to the following model training method. In the embodiment of the present application, the content includes text, video, image, audio, and graphic note.

[0127] Please refer to Figure 3 , Figure 4 A flowchart of a content recommendation method provided by the embodiment of the present application is shown in FIG. 1.

[0128] 101, when receiving a content recommendation request of a first object, obtaining a historical behavior label aggregated vector corresponding to the first object.

[0129] Specifically, the first object can be any user using a content recommendation system, and the content recommendation system can be a video recommendation system, a note recommendation system, an image recommendation system, etc. The content recommendation request is a request for obtaining content recommendation sent by the first object to the content recommendation system. The content can be an image, text, a note, a video, etc. The content recommendation system can label each content with a corresponding content label. The content label is a labeled representation used to describe the theme or attribute of the content, which can help better understand the features and theme of the content. For example, a graphic teaching note introducing sweet and sour spare ribs can be labeled with “food” and “cooking skills”. It can be understood that one content can correspond to one or more content labels.

[0130] Specifically, the historical behavior label aggregated vector can be a vector obtained by aggregating historical behavior label vectors corresponding to historical behavior labels of the first object, and is used to reflect the situation of the content labels that the first object is interested in. The historical behavior label of the first object refers to the content label corresponding to the content that the first object browsed in a historical time period and meets a specific browsing behavior condition. The historical time period can be set by the user, such as one week, one month, etc. The specific browsing behavior condition can be set based on the actual situation, such as the browsing time of the user on the content exceeding a certain browsing time threshold.

[0131] Specifically, a feasible implementation process of obtaining the historical behavior label aggregation vector corresponding to the first object can be: obtaining a historical behavior label set corresponding to the first object; obtaining a historical behavior label vector corresponding to a historical behavior label in the historical behavior label set from the interest label vector matrix; and performing cumulative average processing on the historical behavior label vector to obtain the historical behavior label aggregation vector corresponding to the first object.

[0132] 102. performing interest label prediction processing on the first object based on the historical behavior label aggregation vector and the interest label vector matrix to obtain a predicted interest label.

[0133] Specifically, the interest label vector matrix is determined based on a label vector representation model; the label vector representation model is obtained by training an initial vector representation model based on preference interest labels of at least two sample objects; the preference interest labels are determined according to historical content interaction behaviors of the sample objects on a content library; each content label corresponding to each content in the content library belongs to a content label bag; the interest label vector matrix is used to represent a vector representation corresponding to each content label in the content label bag; and the predicted interest label includes a preferred predicted interest label and an alternative predicted interest label.

[0134] Specifically, a feasible implementation process of the interest label prediction processing on the first object according to the historical behavior label aggregation vector and the interest label vector matrix can be as follows: performing vector neighbor search processing on the historical behavior label aggregation vector based on the interest label vector matrix to obtain a neighbor search label sequence corresponding to the historical behavior label aggregation vector; in the neighbor search label sequence, the neighbor search label at a position closer to the front has a neighbor search label vector closer to the historical behavior label aggregation vector; determining the first neighbor search label in the neighbor search label sequence as a first-priority predicted interest label; obtaining neighbor search labels with a preset interest generalization number from the remaining neighbor search label sequence as alternative predicted interest labels; the remaining neighbor search label sequence refers to the neighbor search label sequence excluding the first-priority predicted interest label; and determining the first-priority predicted interest label and the alternative predicted interest labels as the predicted interest labels. The distance between vectors can be Euclidean distance or geometric distance. The preset interest generalization number can be set based on actual conditions, for example, the preset interest generalization number can be 3. For ease of understanding, assuming that the neighbor search label sequence obtained for the historical behavior label aggregation vector A1 is [furniture lighting, wedding preparation, bedroom furniture, calligraphy and painting decoration, living room furniture, new house decoration], the distance between the neighbor search label vector corresponding to “furniture lighting” and the historical behavior label aggregation vector A1 will be smaller than the distance between the neighbor search label vector corresponding to “bedroom furniture” and the historical behavior label aggregation vector A1. At this time, “furniture lighting” can be taken as the first-priority predicted interest label, and if the preset interest generalization number is 3, the alternative predicted interest labels obtained are “wedding preparation”, “bedroom furniture”, and “calligraphy and painting decoration”.

[0135] Specifically, a feasible implementation process of the vector neighbor search processing on the historical behavior label aggregation vector based on the interest label vector matrix to obtain the neighbor search label sequence corresponding to the historical behavior label aggregation vector can be as follows: determining the vector Euclidean distance between each interest label vector in the interest label vector matrix and the historical behavior label aggregation vector respectively; performing sorting processing on the interest labels corresponding to each interest label vector according to the vector Euclidean distance to obtain a sorted interest label sequence; the vector Euclidean distance corresponding to the kth interest label in the sorted interest label sequence is smaller than the vector Euclidean distance corresponding to the k+1th interest label, k being a positive integer; and obtaining N interest labels in sequence from the sorted interest label sequence according to a preset search number N, and determining the N interest labels as the neighbor search label sequence corresponding to the historical behavior label aggregation vector, N being a positive integer. The vector Euclidean distance can also be referred to as the Euclidean distance of a vector. In mathematics, the Euclidean distance is the “ordinary” (i.e., straight-line) distance between two points in Euclidean space. For ease of understanding, the determination of the vector Euclidean distance can be represented by the following formula 1:

[0136]

[0137] wherein, denotes and the Euclidean matrix between vectors vector

[0138] Specifically, the preset interest generalization quantity includes a preset historical interest quantity S and a preset extended interest quantity H, both S and H are positive integers; a feasible implementation process of obtaining the neighbor search label with the preset interest generalization quantity from the remaining neighbor search label sequence as the candidate preset interest label can be: obtaining a historical behavior label set corresponding to the first object; sequentially obtaining the neighbor search label same as the historical behavior label in the historical behavior label set from the neighbor search label sequence to obtain a historical neighbor search label sequence; sequentially obtaining S neighbor search labels from the historical neighbor search label sequence, and determining the S neighbor search labels as historical neighbor search labels; sequentially obtaining the neighbor search label different from the historical behavior label in the historical behavior label set from the neighbor search label sequence to obtain an extended neighbor search label sequence; sequentially obtaining H neighbor search labels from the extended neighbor search label sequence, and determining the H neighbor search labels as extended neighbor search labels; and taking the historical neighbor search labels and the extended neighbor search labels as the candidate prediction interest labels. For ease of understanding, assuming that the historical behavior label set corresponding to the first object is [European drama, overseas education], the neighbor search label sequence includes [minority language learning, immigration information, postgraduate education, European drama, overseas education], assuming that the preset historical interest quantity is 1 and the preset extended interest quantity is 2, then the final obtained candidate prediction interest labels are [minority language learning, immigration information, European drama].

[0139] 103. Generating a content recommendation candidate set for the content recommendation request according to the preferred prediction interest label, performing sorting optimization processing on the content recommendation candidate set according to the candidate prediction interest label to obtain an optimized content recommendation set, and sending the optimized content recommendation set to the first terminal corresponding to the first object, so that the first terminal sequentially displays the optimized content recommendation set.

[0140] Specifically, the sorting optimization processing can also be called fine sorting processing, and the purpose is to finely adjust the position of the recommended content in the content recommendation candidate set, to ensure that the recommended content with high quality and easy interest of the user is seen by the user first, thereby improving the user experience.

[0141] Specifically, a feasible implementation process of the sorting optimization of the content recommendation candidate set according to the alternative predicted interest label to obtain the optimized content recommendation set can be: respectively determining the label hit rate between the content label corresponding to each recommendation content in the content recommendation candidate set and the alternative predicted interest label; performing position adjustment processing on each recommendation content in the content recommendation candidate set according to the label hit rate corresponding to each recommendation content in the content recommendation candidate set to obtain the optimized content recommendation set; and in the optimized content recommendation set, the label hit rate corresponding to the recommendation content in the front position is higher than the label hit rate corresponding to the recommendation content in the rear position. That is, when performing the position adjustment processing, the position of the recommendation content with a high label hit rate is arranged in the front, thereby enhancing the attraction to the user.

[0142] The label hit rate can also be referred to as a label coincidence rate. The label hit rate refers to the ratio of the number of the same labels between the content label corresponding to the recommendation content and the alternative predicted interest label to the number of the alternative predicted interest label. Assuming that the content recommendation candidate set contains the target recommendation content, a feasible implementation process of the above-mentioned respectively determining the label hit rate between the content label corresponding to each recommendation content in the content recommendation candidate set and the alternative predicted interest label can be: counting the number of the same labels between the content label corresponding to the target recommendation content and the alternative predicted interest label as the label hit number; and dividing the label hit number by the total number of the labels of the alternative predicted interest label to obtain the label hit rate between the content label corresponding to the target recommendation content and the alternative predicted interest label.

[0143] In the embodiment of the present application, when the content recommendation device receives the content recommendation request of the first object, the historical behavior label aggregation vector corresponding to the first object can be obtained. Then, the interest label prediction processing is performed on the first object according to the historical behavior label aggregation vector and the interest label vector matrix, and the predicted interest label is obtained. The interest label vector matrix is determined based on the label vector representation model. The label vector representation model is obtained by training the initial vector representation model based on the preference interest labels of at least two sample objects. The preference interest labels are determined according to the historical content interaction behaviors of the sample objects on the content library. Each content label corresponding to each content in the content library belongs to the content label bag. The interest label vector matrix is used to represent the vector representation corresponding to each content label in the content label bag. The predicted interest label includes the preferred predicted interest label and the alternative predicted interest label. Finally, the content recommendation candidate set for the content recommendation request is generated according to the preferred predicted interest label, the content recommendation candidate set is sorted and optimized according to the alternative predicted interest label, and the optimized content recommendation set is obtained. The optimized content recommendation set is sent to the first terminal corresponding to the first object, so that the first terminal displays the optimized content recommendation set in sequence. The embodiment of the present application can generate the interest label vector matrix for the content label bag based on the preference interest labels of at least two sample objects. The closer the vector representation of the content labels in the matrix, the more likely the user is interested in them. After determining the historical behavior label aggregation vector of the user, the interest label vector matrix and the historical behavior label aggregation vector can be used to explore the interest of the user, and a plurality of predicted interest labels that the user may be interested in are obtained. These predicted interest labels will no longer be limited to the content labels that the user has browsed, and may also be content labels that the user has not browsed. In this way, the diversity of the content recommended to the user can be enhanced without affecting the user's interest.

[0144] The embodiment of the present application also provides a model training method. The execution subject of the model training method is a model training device. The model training device can be any electronic device that can execute the technical solutions disclosed in the embodiments of the present application. Optionally, the model training device can be one of a computer and a server. The model training method is used to train a label vector representation model. Please refer to Figure 4 , Figure 5 The flowchart of the model training method provided in the embodiment of the present application is shown.

[0145] 201. Determine the preference interest labels of at least two sample objects according to the historical content interaction behaviors of the sample objects on the content library. Each content label corresponding to each content in the content library belongs to the content label bag. The preference interest labels belong to the content label bag.

[0146] Specifically, at least two sample objects include a target sample object, and a feasible implementation process of determining the preference interest label of the at least two sample objects according to the historical content interaction behavior of the at least two sample objects on the content library can be determined. Taking the target sample object as an example, the process can be: determining the content dimension interaction score of the target sample object on the content library according to the historical content interaction behavior of the target sample object on the content library; dividing the content library according to the content label bag to obtain a label dimension content library corresponding to each label in the content label bag; determining the label dimension interaction score of each label dimension content library according to the historical content interaction behavior of the target sample object on each label dimension content library; performing self-smoothing processing on the label dimension interaction score of each label dimension content library according to the content dimension interaction score to obtain a label preference score of each label in the content label bag for the target sample object; and sorting the labels in the content label bag according to the label preference score to obtain a sorted label sequence, and obtaining labels with a preset preference number from the sorted label sequence as the preference interest labels of the target sample object.

[0147] Among them, the sample object refers to a user who has used the content recommendation system. The content library refers to a collection of contents published in the content recommendation system. The historical content interaction behavior refers to a series of recordable and analyzable behavior data generated when the user interacts with each content in the content library in the past, such as reading duration, number of likes, comment content, number of comments, etc. The content label bag refers to a collection of all content labels, that is, the content labels corresponding to the contents in the content library all belong to the content label bag.

[0148] Specifically, the above-mentioned feasible implementation process of determining the content dimension interaction score of the target sample object on the content library according to the historical content interaction behavior of the target sample object on the content library can be represented by the process described in formula 2:

[0149]

[0150] Among them, engage_juhe refers to the content dimension interaction score of the target sample object on the content library, which is determined based on the historical content interaction behavior of the target sample object on the content library, that is, click, imp, read_duration, fav, cmt, share, follow, lik, hide, and quick_down in formula 2 can be determined by the historical content interaction behavior of the target sample object on the content library.

[0151] Wherein, click / imp refers to a click rate (Click-Through Rate, CTR), that is, the number of clicks / exposure, which can be used to measure the initial attractiveness of the content, and a high click rate indicates that the title or cover of the content and other elements can effectively attract users to click;

[0152] Wherein, max(0.1, min((read_duration) / (60*imp), 1)) can be understood as a reading duration factor, which can be used to measure the user stickiness of the content, and the higher the ratio, the longer the average reading duration of the user, and the more attractive the content is to the user (such as in-depth articles, high-quality videos), wherein read_duration is the total reading duration (seconds), and 60*imp can be understood as the theoretical maximum possible reading duration (assuming that each exposed user reads for 60 seconds); for example, if a certain content is exposed 100 times (imp=100), the total reading duration is 3000 seconds (50 minutes), and the ratio is 3000 / (60*100)=0.5, then the reading duration factor is 0.5; if a certain content is exposed 1000 times (imp=1000), the total reading duration is 300 seconds (50 minutes), and the ratio is 300 / (60*1000)=0.005, then the reading duration factor is 0.1;

[0153] Wherein, max(fav+cmt+share+follow+0.25*lik-hide+1.0, 1.0) can be understood as an interaction factor, which can be used to measure the social communication and user recognition of the content. Wherein, fav represents the number of collections, cmt represents the number of comments, share represents the number of shares, follow represents the number of attentions, lik represents the number of likes, and hide represents the number of hidden;

[0154] Wherein, (1.0-(quick_down) / imp) / imp can be understood as a quick down factor, which can be used to measure the instant matching degree of the content. If the quick down rate is low (such as the numerator approaches 1) under high exposure, it indicates that the content matches the user's demand; if the quick down rate is high (such as the user slides after reading the title), the factor tends to 0, which greatly reduces the content score, wherein quick_down is the quick down quantity, that is, the number of times of sliding within a short time after exposure, which reflects the matching degree of the content and the user's expectation, and imp is the exposure quantity.

[0155] Specifically, the above-mentioned feasible implementation process of determining the label dimension interaction score of the target sample object to each label dimension content library according to the historical content interaction behavior of the target sample object to each label dimension content library can be determined by the process described in formula 3:

[0156]

[0157] wherein engage_d refers to the label dimension interaction score of the target sample object to the label dimension content library d, which is determined based on the historical content interaction behavior of the target sample object to the label dimension content library d, i.e., click_d, imp_d, read_duration_d, fav_d, cmt_d, share_d, follow_d, lik_d, hide_d, quick_down_d in formula 2 can be determined by the historical content interaction behavior of the target sample object to the label dimension content library d.

[0158] wherein the meanings of click_d / imp_d, max(0.1, min((read_duration) / (60*imp_d), 1)), max(fav_d+cmt_d+share_d+follow_d+0.25*lik_d-hide_d+1.0, 1.0) and (1.0-(quick_down_d) / (imp_d)) / (imp_d) can refer to the above formula 2, which will not be described herein.

[0159] Further, a feasible implementation process of performing self-smoothing processing on the label dimension interaction score of each label dimension content library according to the content dimension interaction score to obtain the label preference score of each label in the content label bag of the target sample object can be represented by the process described in formula 4:

[0160]

[0161] wherein engage refers to the label preference score of the target sample object to the label corresponding to the label dimension content library d, engage_juhe refers to the content dimension interaction score of the target sample object to the content library, engage_d refers to the label dimension interaction score of the target sample object to the label dimension content library d, pow(imp, 0.25) can be understood as an exposure size factor, imp_zx refers to an exposure quality factor, and click_zx refers to a click quality factor. It can be understood that by formula 4, the performance of the target sample object can be self-smoothed to prevent the value from being too large or too small, and the value domain is scaled to obtain a standardized label preference score.

[0162] 202, constructing an initial vector representation model for the content label bag; the initial vector representation model contains a default initial vector matrix as a vector representation matrix; the vector representation matrix is used to represent the vector representation corresponding to each content label in the content label bag.

[0163] Specifically, the initial vector representation model can be a language model that can vectorize the labels, such as a Condition Bag of Words (CBOW) or Skip-gram model. Among them, the vector representation matrix contained in the initial vector representation model generally refers to the weight matrix in the hidden layer of the model, which mainly functions to map each label in the content label bag to a low-dimensional dense word vector. It can be understood that the vector representation matrix contained in the initial vector representation model belongs to a part of the model, so when the initial vector representation model is trained, the vector representation matrix contained in the initial vector representation model will also be adjusted, and when the initial vector representation model training is completed to obtain the label vector representation model, the vector representation matrix contained in the label vector representation model can be directly used to vectorize the labels in the content label bag.

[0164] 203. training the initial vector representation model according to the preference interest labels of the at least two sample objects to obtain a label vector representation model, and determining a vector representation matrix contained in the label vector representation model as an interest label vector matrix; the interest label vector matrix is used for interest label prediction processing of the first object.

[0165] Specifically, the initial vector representation model is trained based on the preference interest labels of at least two sample objects, and a sample set containing multiple labels can be constructed first. For example, the preference interest label set of sample object A is ["technology news", "artificial intelligence", "photography skills"], and the preference interest label set of sample object B is ["tourist guide", "outdoor camping", "food store exploration"], which form a differentiated interest label combination. When training the initial vector representation model (such as CBOW or Skip-gram of Word2Vec), these label sequences are used as input, and the prediction relationship between the context and the center label is used to optimize the model parameters, such as predicting the center label "artificial intelligence" using "technology news" and "photography skills" of sample object A, or inferring the probability distribution of "outdoor camping" through "tourist guide" and "food store exploration" of user B.

[0166] Specifically, taking the Skip-gram model as the initial vector representation model as an example, a feasible implementation process for training the initial vector representation model based on the preference and interest labels of at least two sample objects to obtain the label vector representation model can be as follows: Perform label segmentation on the preference and interest labels of at least two sample objects to obtain the predicted preference and interest labels and the true preference and interest labels of at least two sample objects; input the predicted preference and interest labels of at least two sample objects into the initial vector representation model to obtain the predicted preference and interest labels of at least two sample objects; determine the model loss degree corresponding to the initial vector representation model based on the predicted preference and interest labels and the true preference and interest labels; perform model parameter tuning on the initial vector representation model based on the model loss degree to obtain a new vector representation model; if the model loss degree is greater than the model loss threshold, continue to input the predicted preference and interest labels of at least two sample objects into the new vector representation model; if the model loss degree is less than or equal to the model loss threshold, then determine the new vector representation model as the label vector representation model.

[0167] To better understand the model training process described above, please refer to the following: Figure 5 , Figure 5 This is a schematic diagram illustrating the training of an initial vector representation model provided in an embodiment of this application. For example... Figure 5 As shown, the entire initial vector representation model consists of an input layer, hidden layers, and an output layer. The input to the input layer is X. k Its dimension is Vdim, meaning the dimension of the input layer vector is V. The input to the hidden layer is h. i The input layer, with dimensions Ndim, receives signals from the input layer and performs computational processing. The output layer, with dimensions C × Vdim, consists of C sub-output layers, each with dimension Vdim, and outputs ydim. 1j y 2j , ..., y cj Additionally, such as Figure 5 As shown, the weight matrix from the input layer to the hidden layer is W. V×N , used to represent the connection weights between the V neurons in the input layer and the N neurons in the hidden layer. The weight matrix from the hidden layer to the output layer is W'. N×V , used to represent the connection weights between the N neurons in the hidden layer and the V neurons in the output layer. It can be understood that the weight matrix W... V×N That is, the vector representation matrix described in this application.

[0168] like Figure 6As shown, when the initial vector representation model is trained, the encoding vector corresponding to the to-be-predicted preference interest label can be input to the input layer, the input layer can map the input encoding vector to the hidden layer by multiplying the input encoding vector with the weight matrix of the hidden layer to obtain a low-dimensional label vector representation, and then the hidden layer can multiply the obtained low-dimensional label vector representation with the weight matrix of the output layer, and process it through the Softmax activation function to obtain the probability distribution of each label in the content label bag as the context label of the to-be-predicted preference interest label, so that the label with a high probability can be determined as the predicted preference interest label. Then, the model loss degree can be calculated based on the coincidence degree of the predicted preference interest label and the real preference interest label determined in advance. The model loss degree can be calculated by using a negative log-likelihood loss function or other loss functions, which are not limited in the present application. After the model loss degree is determined, the initial vector representation model can be back propagated, that is, according to the calculated loss, the gradient of the loss function with respect to each layer weight and word vector is calculated by using the chain rule, and then the word vector and the weight parameter of the neural network are updated according to the calculated gradient by using an optimization algorithm (such as stochastic gradient descent (SGD)), so that the loss function value is continuously reduced. When the loss function value is less than or equal to the model loss threshold, the new vector representation model is determined as the label vector representation model.

[0169] In the embodiment of the present application, the model training apparatus can determine the preference interest labels of the at least two sample objects according to the historical content interaction behaviors of the at least two sample objects on the content library, then construct an initial vector representation model for the content label bag, and finally train the initial vector representation model according to the preference interest labels of the at least two sample objects to obtain a label vector representation model, and determine the vector representation matrix contained in the label vector representation model as an interest label vector matrix. Wherein, the content labels corresponding to each content in the content library belong to the content label bag; the preference interest label belongs to the content label bag; the vector representation matrix contained in the initial vector representation model is a default initial vector matrix; the vector representation matrix is used to represent the vector representation corresponding to each content label in the content label bag; and the interest label vector matrix is used for interest label prediction processing of the first object. Through the method provided in the embodiment of the present application, the content labels corresponding to the content can be tokenized as the expression of user interest, so that the exploration of interest is regarded as a generation problem of a text sequence. By analyzing the historical content interaction behaviors of the sample objects, the interest label vector matrix corresponding to the content label can be obtained, and through the interest label vector matrix, the first object's most likely preferred predicted interest label and alternative predicted interest label can be quickly predicted, which effectively optimizes the user interest exploration path, reduces the overfitting risk of the end-to-end model while using the historical behavior information of the user, and avoids the user from falling into the information cocoon of local interest.

[0170] Those skilled in the art can understand that the sequence of writing each step in the above method of the specific embodiment does not mean a strict execution sequence and does not constitute any limitation on the implementation process. The specific execution sequence of each step should be determined by its function and possible internal logic.

[0171] The above describes the method of the embodiment of the application in detail. The device of the embodiment of the application is provided below.

[0172] Please refer to Figure 6 , Figure 7 The structure diagram of a content recommendation device provided by the embodiment of the application is shown in FIG. 1. The content recommendation device 1 comprises an acquisition unit 11, a prediction unit 12, and a ranking optimization unit 13, and specifically comprises:

[0173] The acquisition unit 11 is configured to acquire a historical behavior label aggregated vector corresponding to a first object when receiving a content recommendation request of the first object.

[0174] The prediction unit 12 is configured to perform interest label prediction processing on the first object according to the historical behavior label aggregated vector and an interest label vector matrix to obtain a predicted interest label. The interest label vector matrix is determined based on a label vector representation model. The label vector representation model is obtained by training an initial vector representation model based on preference interest labels of at least two sample objects.

[0175] The ranking optimization unit 13 is configured to generate a content recommendation candidate set for the content recommendation request according to the preferred predicted interest label, perform ranking optimization processing on the content recommendation candidate set according to the alternative predicted interest label, obtain an optimized content recommendation set, and send the optimized content recommendation set to a first terminal corresponding to the first object, so that the first terminal displays the optimized content recommendation set in sequence.

[0176] According to any one of the embodiments of the application, the preference interest label is determined according to historical content interaction behaviors of the sample object on a content library. Each content label corresponding to each content in the content library belongs to a content label bag. The interest label vector matrix is used to represent a vector representation corresponding to each content label in the content label bag. The predicted interest label comprises a preferred predicted interest label and an alternative predicted interest label.

[0177] According to any one of the embodiments of the application, the prediction unit 12 is specifically configured to perform interest label prediction processing on the first object according to the historical behavior label aggregated vector and the interest label vector matrix to obtain a predicted interest label.

[0178] perform vector neighbor search processing on the historical behavior label aggregation vector based on the interest label vector matrix to obtain a neighbor search label sequence corresponding to the historical behavior label aggregation vector; in the neighbor search label sequence, a neighbor search label at a position closer to the front has a closer distance between a neighbor search label vector corresponding to the neighbor search label and the historical behavior label aggregation vector;

[0179] determine a first neighbor search label in the neighbor search label sequence as a first preferred prediction interest label;

[0180] obtain neighbor search labels with a preset interest generalization number from a remaining neighbor search label sequence as alternative prediction interest labels; the remaining neighbor search label sequence refers to a neighbor search label sequence excluding the first preferred prediction interest label;

[0181] determine the first preferred prediction interest label and the alternative prediction interest labels as prediction interest labels.

[0182] In combination with any of the embodiments of the present application, the prediction unit 12 is specifically configured to perform vector neighbor search processing on the historical behavior label aggregation vector based on the interest label vector matrix to obtain a neighbor search label sequence corresponding to the historical behavior label aggregation vector.

[0183] determine a vector Euclidean distance between each interest label vector in the interest label vector matrix and the historical behavior label aggregation vector, respectively;

[0184] perform sorting processing on interest labels corresponding to each interest label vector according to the vector Euclidean distance to obtain a sorted interest label sequence; a vector Euclidean distance corresponding to a kth interest label in the sorted interest label sequence is smaller than a vector Euclidean distance corresponding to a (k+1)th interest label, k being a positive integer;

[0185] obtain N interest labels in sequence according to a preset search number N from the sorted interest label sequence, and determine the N interest labels as a neighbor search label sequence corresponding to the historical behavior label aggregation vector, N being a positive integer.

[0186] In combination with any of the embodiments of the present application, the preset interest generalization number includes a preset historical interest number S and a preset expansion interest number H, S and H both being positive integers;

[0187] The prediction unit 12 is specifically configured to obtain neighbor search labels with a preset interest generalization number from a remaining neighbor search label sequence as alternative prediction interest labels.

[0188] obtain a historical behavior label set corresponding to the first object;

[0189] obtain, from the neighbor search label sequence, neighbor search labels that are different from the historical behavior labels in the historical behavior label set, to obtain an extended neighbor search label sequence;

[0190] obtain, from the historical nearest neighbor search label sequence, S neighbor search labels in sequence, and determine the S neighbor search labels as historical neighbor search labels;

[0191] obtain, from the neighbor search label sequence, neighbor search labels that are different from the historical behavior labels in the historical behavior label set, to obtain an extended neighbor search label sequence;

[0192] obtain, from the extended neighbor search label sequence, H neighbor search labels in sequence, and determine the H neighbor search labels as extended neighbor search labels;

[0193] use the historical neighbor search labels and the extended neighbor search labels as candidate predicted interest labels.

[0194] According to any one of the embodiments of the present application, the sorting optimization unit 13 is specifically configured to:

[0195] determine the label hit rate between the content label corresponding to each recommended content in the content recommendation candidate set and the candidate predicted interest label;

[0196] perform position adjustment processing on each recommended content in the content recommendation candidate set according to the label hit rate corresponding to each recommended content in the content recommendation candidate set, to obtain an optimized content recommendation set; in the optimized content recommendation set, the label hit rate corresponding to the recommended content in the front position is higher than the label hit rate corresponding to the recommended content in the rear position.

[0197] According to any one of the embodiments of the present application, the obtaining unit 11 is specifically configured to:

[0198] obtain a historical behavior label set corresponding to the first object;

[0199] obtain a historical behavior label vector corresponding to the historical behavior label in the historical behavior label set from an interest label vector matrix;

[0200] perform cumulative average processing on the historical behavior label vector to obtain a historical behavior label aggregation vector corresponding to the first object.

[0201] In the embodiment of the present application, when the content recommendation device receives the content recommendation request of the first object, the historical behavior label aggregation vector corresponding to the first object can be obtained. Then, the interest label prediction processing is performed on the first object according to the historical behavior label aggregation vector and the interest label vector matrix, and the predicted interest label is obtained. The interest label vector matrix is determined based on the label vector representation model. The label vector representation model is obtained by training the initial vector representation model based on the preferred interest labels of at least two sample objects. The predicted interest label includes the preferred predicted interest label and the alternative predicted interest label. Finally, the content recommendation candidate set for the content recommendation request is generated according to the preferred predicted interest label, the content recommendation candidate set is sorted and optimized according to the alternative predicted interest label, the optimized content recommendation set is obtained, and the optimized content recommendation set is sent to the first terminal corresponding to the first object, so that the first terminal displays the optimized content recommendation set in sequence. The embodiment of the present application can generate the interest label vector matrix for the content label bag based on the preferred interest labels of at least two sample objects. The closer the vector representation of the content label in the matrix, the more likely the user is interested in it. After determining the historical behavior label aggregation vector of the user, the interest label vector matrix and the historical behavior label aggregation vector can be used to explore the interest of the user, and a plurality of predicted interest labels that the user may be interested in are obtained. These predicted interest labels will no longer be limited to the content labels that the user has browsed, and may also be content labels that the user has not browsed. In this way, the diversity of the content recommended to the user can be enhanced without affecting the user's interest.

[0202] Please refer to Figure 7 , Figure 8 The structure schematic diagram of a model training device provided in the embodiment of the present application. The model training device 2 comprises a determination unit 21, a construction unit 22 and a training unit 23, specifically:

[0203] The determination unit 21 is configured to determine the preferred interest labels of at least two sample objects according to the historical content interaction behaviors of the at least two sample objects on the content library. Each content in the content library corresponds to a content label belonging to a content label bag. The preferred interest labels belong to the content label bag.

[0204] The construction unit 22 is configured to construct an initial vector representation model for the content label bag. The initial vector representation model includes a default initial vector matrix. The vector representation matrix is used to represent the vector representation of each content label in the content label bag.

[0205] The training unit 23 is configured to train the initial vector representation model according to the preference interest labels of the at least two sample objects, to obtain a label vector representation model, and to determine a vector representation matrix contained in the label vector representation model as an interest label vector matrix; and the interest label vector matrix is used for interest label prediction processing on the first object.

[0206] According to any one of the embodiments of the present application, the at least two sample objects include a target sample object.

[0207] According to the historical content interaction behavior of the at least two sample objects on the content library, the determination unit 21 is specifically configured to:

[0208] According to the historical content interaction behavior of the target sample object on the content library, the determination unit 21 is specifically configured to determine a content dimension interaction score of the target sample object on the content library.

[0209] According to the content label bag, the content library is divided into a plurality of label dimension content libraries corresponding to each label in the content label bag.

[0210] According to the historical content interaction behavior of the target sample object on each label dimension content library, the determination unit 21 is specifically configured to determine a label dimension interaction score of the target sample object on each label dimension content library.

[0211] According to the content dimension interaction score, the label dimension interaction score of each label dimension content library is self-smoothed to obtain a label preference score of each label in the content label bag for the target sample object.

[0212] According to the label preference score, the labels in the content label bag are sorted to obtain a sorted label sequence, and from the sorted label sequence, a label with a preset preference number is obtained in sequence as the preference interest label of the target sample object.

[0213] According to any one of the embodiments of the present application, the training unit 23 is specifically configured to:

[0214] The preference interest labels of the at least two sample objects are respectively subjected to label division processing to obtain predicted preference interest labels and real preference interest labels of the at least two sample objects.

[0215] The predicted preference interest labels of the at least two sample objects are input into the initial vector representation model to obtain the predicted preference interest labels of the at least two sample objects.

[0216] According to the predicted preference interest label and the real preference interest label, a model loss degree corresponding to the initial vector representation model is determined, and the initial vector representation model is subjected to model parameter adjustment processing based on the model loss degree, to obtain a new vector representation model;

[0217] If the model loss degree is greater than a model loss threshold, the at least two sample objects are continuously input into the new vector representation model for prediction of the preference interest label.

[0218] If the model loss degree is less than or equal to the model loss threshold, the new vector representation model is determined as a label vector representation model.

[0219] In some embodiments, the apparatus provided by the embodiments of the present application has functions or includes modules that can be used to execute the methods described in the above method embodiments, and the specific implementation can refer to the description of the above method embodiments. For brevity, details are not repeated here.

[0220] Figure 8 A hardware structure schematic diagram of an electronic device provided by an embodiment of the present application is provided. The electronic device 4 includes a processor 41 and a memory 42. Optionally, the electronic device 4 further includes an input device 43 and an output device 44. The processor 41, the memory 42, the input device 43 and the output device 44 are coupled through a connector, which includes various interfaces, transmission lines or buses, etc. The present application does not limit this. It should be understood that in various embodiments of the present application, coupling means mutual contact in a specific way, including direct connection or indirect connection through other devices, for example, various interfaces, transmission lines, buses, etc.

[0221] The processor 41 can include one or more processors, for example, including one or more central processing units (CPUs). In the case of a CPU, the CPU can be a single-core CPU or a multi-core CPU. Optionally, the processor 41 can be a processor group composed of multiple CPUs, and the multiple processors are coupled with each other through one or more buses. Optionally, the processor can also be other types of processors, etc. The present application does not limit this.

[0222] The memory 42 can be used to store computer program instructions, and various computer program codes including program codes for implementing the schemes of the present application. Optionally, the memory includes, but is not limited to, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read only memory (EPROM), or a compact disc read-only memory (CD-ROM), which is used for storing relevant instructions and data.

[0223] The input device 43 is configured to input data and / or signals, and the output device 44 is configured to output data and / or signals. The input device 43 and the output device 44 can be independent devices, or can be an integral device.

[0224] It can be understood that, in the embodiments of the present application, the memory 42 can be used to store not only relevant instructions, but also relevant data, and the embodiments of the present application do not limit the data stored in the memory.

[0225] It can be understood that, ​ Only a simplified design of an electronic device is shown. In actual applications, the electronic device can also include other necessary elements, including but not limited to any number of input / output devices, processors, memories, etc., and all electronic devices that can implement the embodiments of the present application are within the protection scope of the present application.

[0226] Those skilled in the art can understand that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be realized in electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0227] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described system, device and unit can refer to the corresponding processes in the foregoing method embodiments, which will not be described here. Those skilled in the art can also clearly understand that each embodiment of the present application describes different focuses, and for the convenience and brevity of description, the same or similar parts can not be described in different embodiments. Therefore, the parts not described or not described in detail in a certain embodiment can refer to the description in other embodiments.

[0228] In several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented in other manners. For example, the described device embodiments are merely schematic. The division of the units is merely logical function division. There can be another division manner for the actual implementation, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections can be indirect couplings or communication connections through some interfaces, devices or units, and can be in electrical, mechanical or other forms.

[0229] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e., can be located in one place, or can be distributed on multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the embodiment.

[0230] In addition, each functional unit in the various embodiments of the present application can be integrated into one processing unit, or each unit can exist physically, or two or more units can be integrated into one unit.

[0231] In the above embodiments, all or part of the embodiments can be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part of the embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted by the computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center through a wired (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.) manner. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. integrated with one or more available media. The available media can be a magnetic medium (such as a floppy disk, a hard disk, a magnetic tape), an optical medium (such as a digital versatile disc (DVD)), or a semiconductor medium (such as a solid state disk (SSD)), etc.

[0232] Those skilled in the art can understand that all or part of the processes in the above-mentioned method embodiments can be implemented by a computer program instructing relevant hardware to complete, the program can be stored in a computer-readable storage medium, and the program can include the processes of the above-mentioned method embodiments when executed. The aforementioned storage medium includes read-only memory (ROM) or random access memory (RAM), a magnetic disc or an optical disc, and various media that can store program codes.

Claims

1. A content recommendation method characterized by, The method comprises the steps of: When receiving a content recommendation request of a first object, obtaining a historical behavior label aggregation vector corresponding to the first object; According to the historical behavior label aggregation vector and the interest label vector matrix, interest label prediction processing is performed on the first object to obtain a predicted interest label; the interest label vector matrix is determined based on a label vector representation model; the label vector representation model is obtained after training based on the preference interest labels of at least two sample objects; According to the predicted interest label, a content recommendation candidate set for the content recommendation request is generated, the content recommendation candidate set is sorted and optimized according to the alternative predicted interest label, an optimized content recommendation set is obtained, and the optimized content recommendation set is sent to a first terminal corresponding to the first object, so that the first terminal displays the optimized content recommendation set in sequence.

2. The method of claim 1, wherein, The preference interest label is determined according to the historical content interaction behavior of the sample object on the content library; each content label corresponding to each content in the content library belongs to a content label bag; the interest label vector matrix is used to represent the vector representation of each content label in the content label bag.

3. The method of claim 1, wherein, The method comprises the steps of: Based on the interest label vector matrix, vector neighbor search processing is performed on the historical behavior label aggregation vector to obtain a neighbor search label sequence corresponding to the historical behavior label aggregation vector; in the neighbor search label sequence, the neighbor search label at the earlier position has a closer distance between the neighbor search label vector corresponding to the neighbor search label and the historical behavior label aggregation vector; The first neighbor search label in the neighbor search label sequence is determined as a preferred predicted interest label; From the remaining neighbor search label sequence, a neighbor search label with a preset interest generalization number is obtained as an alternative predicted interest label; the remaining neighbor search label sequence refers to the neighbor search label sequence excluding the preferred predicted interest label; The preferred predicted interest label and the alternative predicted interest label are determined as the predicted interest label.

4. The method of claim 3, wherein, The method comprises the steps of: The vector Euclidean distance between each interest label vector in the interest label vector matrix and the historical behavior label aggregation vector is determined respectively; According to the vector Euclidean distance, the interest labels corresponding to each interest label vector are sorted to obtain a sorted interest label sequence; the vector Euclidean distance corresponding to the kth interest label in the sorted interest label sequence is smaller than the vector Euclidean distance corresponding to the k+1th interest label, and k is a positive integer; From the sorted interest label sequence, N interest labels are obtained in sequence according to a preset search number N, and the N interest labels are determined as the neighbor search label sequence corresponding to the historical behavior label aggregation vector, and N is a positive integer.

5. The method of claim 2, wherein, The preset interest generalization quantity includes a preset historical interest quantity S and a preset extended interest quantity H, both S and H are positive integers; The preset interest generalization quantity includes a preset historical interest quantity S and a preset extended interest quantity H, both S and H are positive integers; The historical behavior label set corresponding to the first object is acquired; A historical nearest neighbor search label sequence is obtained by sequentially acquiring, from the neighbor search label sequence, neighbor search labels same as the historical behavior labels in the historical behavior label set; S neighbor search labels are sequentially acquired from the historical nearest neighbor search label sequence, and the S neighbor search labels are determined as historical neighbor search labels; An extended neighbor search label sequence is obtained by sequentially acquiring, from the neighbor search label sequence, neighbor search labels different from the historical behavior labels in the historical behavior label set; H neighbor search labels are sequentially acquired from the extended neighbor search label sequence, and the H neighbor search labels are determined as extended neighbor search labels; The historical neighbor search labels and the extended neighbor search labels are determined as candidate prediction interest labels.

6. The method of claim 1, wherein, The candidate content recommendation set is sorted and optimized according to the candidate prediction interest labels, to obtain an optimized content recommendation set. The label hit rates between the content labels corresponding to each recommendation content in the candidate content recommendation set and the candidate prediction interest labels are respectively determined. The position of each recommendation content in the candidate content recommendation set is adjusted according to the label hit rate corresponding to the recommendation content, to obtain an optimized content recommendation set; in the optimized content recommendation set, the label hit rate corresponding to the recommendation content in the front position is higher than the label hit rate corresponding to the recommendation content in the rear position.

7. The method of claim 1, wherein, The historical behavior label aggregation vector corresponding to the first object is acquired, including: The historical behavior label set corresponding to the first object is acquired; The historical behavior label vector corresponding to the historical behavior label in the historical behavior label set is acquired from the interest label vector matrix; The historical behavior label vector is accumulated and averaged, to obtain the historical behavior label aggregation vector corresponding to the first object.

8. A model training method, comprising: The historical content interaction behavior of at least two sample objects to a content library is determined, to obtain the preferred interest labels of the at least two sample objects; Each content in the content library corresponds to a content label, and the content label belongs to a content label bag; the preferred interest label belongs to the content label bag; An initial vector representation model for the content label bag is constructed; the vector representation matrix contained in the initial vector representation model is a default initial vector matrix; the vector representation matrix is used to represent the vector representation corresponding to each content label in the content label bag; The initial vector representation model is trained according to the preferred interest labels of the at least two sample objects, to obtain a label vector representation model; the vector representation matrix contained in the label vector representation model is determined as an interest label vector matrix; the interest label vector matrix is used to perform interest label prediction processing on a first object. The at least two sample objects include a target sample object.

9. The method of claim 8, wherein, ​ According to the historical content interaction behavior of the at least two sample objects on the content library, the preference interest label of the at least two sample objects is determined, comprising: According to the historical content interaction behavior of the target sample object on the content library, the content dimension interaction score of the target sample object on the content library is determined; According to the content label bag, the content library is divided, and the label dimension content library corresponding to each label in the content label bag is obtained; According to the historical content interaction behavior of the target sample object on each label dimension content library, the label dimension interaction score of the target sample object on each label dimension content library is determined; According to the content dimension interaction score, the label dimension interaction score of each label dimension content library is self-smoothed to obtain the label preference score of each label in the content label bag for the target sample object; According to the label preference score, the labels in the content label bag are sorted to obtain a sorted label sequence, and from the sorted label sequence, a label with a preset preference number is obtained in sequence as the preference interest label of the target sample object.

10. The method of claim 8, wherein, According to the at least two sample objects, the initial vector representation model is trained to obtain a label vector representation model, comprising: The preference interest label of the at least two sample objects is respectively processed by label division to obtain the predicted preference interest label and the real preference interest label of the at least two sample objects; The predicted preference interest label of the at least two sample objects is input into the initial vector representation model to obtain the predicted preference interest label of the at least two sample objects; According to the predicted preference interest label and the real preference interest label, the model loss degree corresponding to the initial vector representation model is determined, and the initial vector representation model is processed by model parameter adjustment based on the model loss degree to obtain a new vector representation model; If the model loss degree is greater than the model loss threshold, the predicted preference interest label of the at least two sample objects is continuously input into the new vector representation model; If the model loss degree is less than or equal to the model loss threshold, the new vector representation model is determined as the label vector representation model.

11. A content recommendation apparatus characterized by comprising: Comprising: The acquisition unit is used for acquiring the historical behavior label aggregation vector corresponding to the first object when receiving the content recommendation request of the first object; The prediction unit is used for performing interest label prediction processing on the first object according to the historical behavior label aggregation vector and the interest label vector matrix to obtain a predicted interest label; the interest label vector matrix is determined based on a label vector representation model; the label vector representation model is obtained by training an initial vector representation model based on the preference interest label of at least two sample objects. The sorting optimization unit is configured to generate a content recommendation candidate set for the content recommendation request according to the preferred predicted interest label, perform sorting optimization processing on the content recommendation candidate set according to the alternative predicted interest label, obtain an optimized content recommendation set, and send the optimized content recommendation set to a first terminal corresponding to the first object, so that the first terminal sequentially displays the optimized content recommendation set.

12. A model training apparatus, comprising: The method comprises: The determining unit is configured to determine a preferred interest label of at least two sample objects according to historical content interaction behaviors of the at least two sample objects on a content library; Each content label corresponding to each content in the content library belongs to a content label bag; the preferred interest label belongs to the content label bag; The constructing unit is configured to construct an initial vector representation model for the content label bag; the initial vector representation model contains a default initial vector matrix as a vector representation matrix; the vector representation matrix is used to represent a vector representation corresponding to each content label in the content label bag. The training unit is configured to train the initial vector representation model according to the preferred interest label of the at least two sample objects, obtain a label vector representation model, and determine a vector representation matrix contained in the label vector representation model as an interest label vector matrix; the interest label vector matrix is used to perform interest label prediction processing on a first object.

13. An electronic device, comprising: The processor and the memory are configured to store computer program codes, the computer program codes comprise computer instructions, and the electronic device executes the method in any one of claims 1 to 7 when the processor executes the computer instructions; the electronic device executes the method in any one of claims 8 to 10 when the processor executes the computer instructions. The computer readable storage medium stores a computer program, the computer program comprises program instructions, and the processor executes the method in any one of claims 1 to 7 when the program instructions are executed by the processor; the processor executes the method in any one of claims 8 to 10 when the program instructions are executed by the processor.

14. A computer-readable storage medium, characterized in that, The computer program product comprises a computer program or instructions; the computer executes the method in any one of claims 1 to 7 when the computer program or instructions run on the computer; the computer executes the method in any one of claims 8 to 10 when the computer program or instructions run on the computer.

15. A computer program product, characterised in that, ​

Images