Video recommendation method, device, equipment, storage medium and program product

Abstract

The present application discloses a video recommendation method, apparatus, device, storage medium and program product, which belongs to the field of machine learning. The method includes: obtaining a multidimensional video feature vector of a shared video, a multidimensional object feature vector of a first object and a multidimensional video feature vector of a candidate video in a candidate video set; fusing the multidimensional video feature vector and the multidimensional object feature vector of the shared video to obtain an object fusion vector; fusing the multidimensional video feature vectors of the same candidate video to obtain a candidate video fusion vector of the candidate video; performing similarity calculation based on the object fusion vector and the candidate video fusion vector to obtain a similarity value between the object fusion vector and the candidate video fusion vector; and recommending the candidate video to the first object based on the similarity value. The present application improves the accuracy of video recommendation by extracting the multimodal feature vector and the multidimensional object feature vector of the shared video to make a comprehensive judgment on the candidate video.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of machine learning, and in particular to a video recommendation method, apparatus, device, storage medium, and program product. Background Art

[0002] With the development of self-media technology, more and more users can watch videos shared by others at any time. When users watch videos shared by others, the server can use the video recommendation model to recommend some videos related to the shared video from the massive video database, thereby better meeting the user's video viewing needs.

[0003] During the video recommendation process, the server calculates the distance between the user's features and the feature vector corresponding to the ID sequence information of any candidate video in the candidate video set to obtain the similarity value between the user's feature vector and the feature vector corresponding to the ID sequence information. The server repeats the above steps for each video in the candidate video set to obtain the similarity value between the user's feature vector and the feature vectors corresponding to the ID sequence information of multiple candidate videos. Further, based on the ranking of all similarity values, the video corresponding to the similarity value with the highest ranking is recommended to the user.

[0004] However, in related technologies, shared videos often do not appear in the training set or candidate video set of the video recommendation model due to changes in hot spots. That is, the ID sequence information of the shared videos has a low coverage rate in the training set or candidate video set, resulting in a poor correlation between the recommended videos obtained based on the similarity value between the user's feature vector and the feature vector corresponding to the ID sequence information of the candidate video and the shared video, which affects the accuracy of video recommendations. Summary of the Invention

[0005] This application provides a video recommendation method, apparatus, device, storage medium, and program product that can improve the accuracy of video recommendations. The technical solution is as follows:

[0006] According to one aspect of the present application, a video recommendation method is provided, the method comprising:

[0007] Obtaining a multidimensional video feature vector of a shared video, a multidimensional object feature vector of a first object, and multidimensional video feature vectors of candidate videos in a candidate video set, wherein the shared video refers to a video clicked by the first object after receiving a video shared by the second object, and the multidimensional video feature vector includes an identification feature vector and a multimodal feature vector;

[0008] fusing the multidimensional video feature vector of the shared video and the multidimensional object feature vector to obtain an object fusion vector of the first object;

[0009] fusing the multi-dimensional video feature vectors of the same candidate video to obtain a candidate video fusion vector of the candidate video;

[0010] Performing similarity calculation based on the object fusion vector of the first object and the candidate video fusion vector to obtain a similarity value between the object fusion vector and the candidate video fusion vector;

[0011] The candidate video is recommended to the first object according to the similarity value.

[0012] According to one aspect of the present application, a method for training a video recommendation model is provided, the method comprising:

[0013] Obtaining a multidimensional video feature vector of a shared video, a multidimensional object feature vector of a first object, a multidimensional video feature vector of a positive candidate video sample in a candidate video set, and a multidimensional video feature vector of a negative candidate video sample, wherein the shared video refers to a video clicked by the first object after receiving the video shared by the second object, the multidimensional video feature vector includes an identification feature vector and a multimodal feature vector, the positive candidate video sample refers to a video associated with the shared video, and the negative candidate video sample refers to a video not associated with the shared video;

[0014] fusing the multidimensional video feature vector of the shared video and the multidimensional object feature vector to obtain an object fusion vector of the first object;

[0015] fusing the multi-dimensional video feature vectors of the same positive candidate video sample to obtain a positive candidate video sample fusion vector of the positive candidate video sample;

[0016] fusing the multi-dimensional video feature vectors of the same negative candidate video sample to obtain a negative candidate video sample fusion vector of the negative candidate video sample;

[0017] Performing similarity calculation based on the object fusion vector of the first object and the positive candidate video sample fusion vector to obtain a first similarity value between the object fusion vector and the positive candidate video sample fusion vector;

[0018] Performing similarity calculation based on the object fusion vector of the first object and the negative candidate video sample fusion vector to obtain a second similarity value between the object fusion vector and the negative candidate video sample fusion vector;

[0019] Calculating a loss function value based on the first similarity value and the second similarity value;

[0020] The model parameters of the video recommendation model are updated based on the loss function value.

[0021] In a possible implementation, the video recommendation model further includes an attention mechanism network;

[0022] The candidate video set includes one positive candidate video sample and at least one negative candidate video sample, and the video recommendation model includes an attention mechanism network and a feature fusion network;

[0023] The step of fusing the multi-dimensional video feature vectors of the same positive candidate video sample to obtain a positive candidate video sample fusion vector of the positive candidate video sample includes:

[0024] Inputting the identification feature vector and the multimodal feature vector corresponding to the positive candidate video sample into the attention mechanism network for weight calculation to obtain a second weight value corresponding to each feature vector in the positive candidate video sample;

[0025] According to the second weight value, the identification feature vector and the multimodal feature vector corresponding to the positive candidate video sample are input into the feature fusion network for feature fusion to obtain the positive candidate video sample fusion vector.

[0026] In a possible implementation, the candidate video set includes one positive candidate video sample and at least one negative candidate video sample, and the video recommendation model includes an attention mechanism network and a feature fusion network;

[0027] The step of fusing the multi-dimensional video feature vectors of the same negative candidate video sample to obtain a negative candidate video sample fusion vector of the negative candidate video sample includes:

[0028] Inputting the identification feature vector and the multimodal feature vector corresponding to the i-th negative candidate video sample in the candidate video set into the attention mechanism network for weight calculation, obtaining a second weight value corresponding to each feature vector in the i-th negative candidate video sample, where i is a positive integer;

[0029] According to the second weight value, the identification feature vector and the multimodal feature vector corresponding to the i-th negative candidate video sample are input into the feature fusion network for feature fusion to obtain the i-th negative candidate video sample fusion vector.

[0030] In a possible implementation, the video recommendation model includes a multi-layer perceptron network MLP;

[0031] The step of inputting the identification feature vector of the shared video, the multimodal feature vector of the shared video, and the multidimensional object feature vector into the feature fusion network for feature fusion according to the first weight value to obtain the object fusion vector of the first object includes:

[0032] Inputting the identification feature vector of the shared video, the multimodal feature vector of the shared video, and the multidimensional object feature vector into the feature fusion network according to the first weight value to perform feature fusion to obtain an intermediate object fusion vector;

[0033] The intermediate object fusion vector is input into a multi-layer perceptron network (MLP) for nonlinear fitting to obtain the object fusion vector of the first object.

[0034] In one possible implementation, the video recommendation model includes a multi-layer perceptron network (MLP); inputting the identification feature vector and the multimodal feature vector corresponding to the positive candidate video sample into a feature fusion network for feature fusion according to the second weight value to obtain the positive candidate video sample fusion vector includes:

[0035] Inputting the identification feature vector and the multimodal feature vector corresponding to the positive candidate video sample into a feature fusion network for feature fusion according to the second weight value to obtain an intermediate positive candidate video sample fusion vector;

[0036] The intermediate positive candidate video sample fusion vector is input into a multi-layer perceptron network MLP for nonlinear fitting to obtain the positive candidate video sample fusion vector.

[0037] In a possible implementation, the video recommendation model includes a multi-layer perceptron network MLP;

[0038] The step of inputting the identification feature vector and the multimodal feature vector corresponding to the i-th negative candidate video sample into a feature fusion network for feature fusion according to the second weight value to obtain the i-th negative candidate video sample fusion vector includes:

[0039] Inputting the identification feature vector and the multimodal feature vector corresponding to the i-th negative candidate video sample into the feature fusion network for feature fusion according to the second weight value to obtain an intermediate i-th negative candidate video sample fusion vector;

[0040] The intermediate i-th negative candidate video sample fusion vector is input into the multi-layer perceptron network MLP for nonlinear fitting to obtain the i-th negative candidate video sample fusion vector.

[0041] In a possible implementation, the performing similarity calculation on the object fusion vector based on the first object and the positive candidate video sample fusion vector to obtain a first similarity value between the object fusion vector and the positive candidate video sample fusion vector includes:

[0042] A cosine similarity value is calculated on the object fusion vector and the positive candidate video sample fusion vector to obtain a first similarity value between the object fusion vector and the positive candidate video sample fusion vector.

[0043] In a possible implementation, the performing similarity calculation on the object fusion vector based on the first object and the negative candidate video sample fusion vector to obtain a second similarity value between the object fusion vector and the negative candidate video sample fusion vector includes:

[0044] A cosine similarity value is calculated for the object fusion vector and the i-th negative candidate video sample fusion vector in the candidate video set to obtain an i-th similarity value between the object fusion vector and the i-th negative candidate video sample fusion vector, where i is a positive integer.

[0045] In a possible implementation, the method further includes:

[0046] Acquire an object data set and object attribute data of the first object;

[0047] The object data set and the object attribute data are respectively input into a feature extraction network for feature extraction to obtain the object data vector and the object attribute vector of the first object.

[0048] In a possible implementation, the method further includes:

[0049] Obtaining identification information and multimodal information of the shared video, wherein the identification information includes at least one of ID sequence information of the shared video, author information of the shared video, and tag attribute information of the shared video; and the multimodal information of the shared video includes at least two of shared video title segmentation information, shared video title information, shared video cover information, and shared video content information;

[0050] The identification information of the shared video and the multimodal information of the shared video are respectively input into corresponding feature extraction networks for feature extraction to obtain the identification feature vector and the multimodal feature vector of the shared video.

[0051] In a possible implementation, the method further includes:

[0052] Obtain identification information and multimodal information of the positive candidate video sample, the identification information including at least one of ID sequence information of the positive candidate video sample, author information of the positive candidate video sample, and label attribute information of the positive candidate video sample, and the multimodal information of the positive candidate video sample including at least two of positive candidate video sample title segmentation information, positive candidate video sample title information, positive candidate video sample cover information, and positive candidate video sample content information;

[0053] The identification class information of the positive candidate video sample and the multimodal information of the positive candidate video sample are respectively input into corresponding feature extraction networks for feature extraction to obtain the identification class feature vector and the multimodal feature vector of the positive candidate video sample.

[0054] In a possible implementation, the method further includes:

[0055] Obtain identification information and multimodal information of the negative candidate video sample, the identification information including at least one of ID sequence information of the negative candidate video sample, author information of the negative candidate video sample, and label attribute information of the negative candidate video sample, and the multimodal information of the negative candidate video sample including at least two of negative candidate video sample title segmentation information, negative candidate video sample title information, negative candidate video sample cover information, and negative candidate video sample content information;

[0056] The identification class information of the negative candidate video sample and the multimodal information of the negative candidate video sample are respectively input into the corresponding feature extraction network for feature extraction to obtain the identification class feature vector and the multimodal feature vector of the negative candidate video sample.

[0057] In a possible implementation, calculating a loss function value based on the first similarity value and the second similarity value includes:

[0058] Adding the second similarity values ​​to obtain a total similarity value;

[0059] The loss function value is obtained based on a ratio of the first similarity value to the total similarity value.

[0060] According to one aspect of the present application, a video recommendation device is provided, comprising:

[0061] an acquisition module, configured to acquire a multidimensional video feature vector of a shared video, a multidimensional object feature vector of a first object, and multidimensional video feature vectors of candidate videos in a candidate video set, wherein the shared video refers to a video clicked by the first object after receiving a video shared by the second object, and the multidimensional video feature vector includes an identification feature vector and a multimodal feature vector;

[0062] a fusion module, configured to fuse the multidimensional video feature vector of the shared video and the multidimensional object feature vector to obtain an object fusion vector of the first object;

[0063] The fusion module is further configured to fuse the multi-dimensional video feature vectors of the same candidate video to obtain a candidate video fusion vector of the candidate video;

[0064] a calculation module, configured to perform similarity calculation based on the object fusion vector of the first object and the candidate video fusion vector to obtain a similarity value between the object fusion vector and the candidate video fusion vector;

[0065] A recommendation module is configured to recommend the candidate video to the first object based on the similarity value.

[0066] In one possible implementation, the calculation module is also used to input the identification feature vector of the shared video, the multimodal feature vector of the shared video, and the multidimensional object feature vector into the attention mechanism network for weight calculation to obtain the first weight value corresponding to each feature vector.

[0067] In a possible implementation, the fusion module is further used to perform feature fusion on the identification feature vector of the shared video, the multimodal feature vector of the shared video, and the multidimensional object feature vector according to the first weight value to obtain the object fusion vector of the first object.

[0068] In a possible implementation, the fusion module is further used to perform feature fusion on the identification feature vector of the shared video, the multimodal feature vector of the shared video, and the multidimensional object feature vector according to the first weight value to obtain an intermediate object fusion vector; and input the intermediate object fusion vector into a multi-layer perceptron network MLP for nonlinear fitting to obtain the object fusion vector of the first object.

[0069] In one possible implementation, the calculation module is also used to input the identification feature vector and the multimodal feature vector corresponding to the i-th candidate video in the candidate video set into the attention mechanism network for weight calculation, and obtain the second weight value corresponding to each feature vector in the i-th candidate video, where i is a positive integer.

[0070] In a possible implementation, the fusion module is further configured to perform feature fusion on the identification feature vector and the multimodal feature vector corresponding to the i-th candidate video according to a second weight value to obtain a fusion vector of the i-th candidate video.

[0071] In one possible implementation, the fusion module is further used to perform feature fusion on the identification feature vector and the multimodal feature vector corresponding to the i-th candidate video according to the second weight value to obtain an intermediate i-th candidate video fusion vector; and input the intermediate i-th candidate video fusion vector into a multi-layer perceptron network MLP for nonlinear fitting to obtain the i-th candidate video fusion vector.

[0072] In a possible implementation, the calculation module is further used to calculate the cosine similarity value between the object fusion vector and the i-th candidate video fusion vector in the candidate video set to obtain the i-th similarity value between the object fusion vector and the i-th candidate video fusion vector, where i is a positive integer.

[0073] In a possible implementation manner, the acquisition module is further configured to acquire a multi-dimensional object feature vector of the second object.

[0074] The fusion module is further configured to fuse the multidimensional video feature vector of the shared video, the multidimensional object feature vector of the first object, and the multidimensional object feature vector of the second object to obtain an object fusion vector of the first object.

[0075] In one possible implementation, the multidimensional object feature vector of the second object includes an object data vector and an object attribute vector. The acquisition module is further configured to acquire an object data set and object attribute data of the second object; and input the object data set and the object attribute data into a feature extraction network for feature extraction, thereby obtaining the object data vector and the object attribute vector of the second object.

[0076] In one possible implementation, the acquisition module is further configured to acquire a multidimensional object feature vector of the second object, and to acquire a first directional feature vector between the second object and the first object, wherein the first directional feature vector is configured to represent that the shared video is shared from the second object to the first object.

[0077] The fusion module is further used to fuse the multidimensional video feature vector of the shared video, which is shared from the second object to the first object, the multidimensional object feature vector of the first object, the multidimensional object feature vector of the second object and the first directional feature vector to obtain an object fusion vector of the first object.

[0078] In a possible implementation, the acquisition module is further configured to acquire a second directional feature vector of the candidate video in the candidate video set, where the second directional feature vector is used to represent that the candidate video is shared from the third object to the fourth object in the historical sharing record;

[0079] The fusion module is also used to fuse the multidimensional video feature vector of the same candidate video, the multidimensional object feature vector of the third object, the multidimensional object feature vector of the fourth object and the second directional feature vector to obtain a candidate video fusion vector of the candidate video.

[0080] In one possible implementation, the acquisition module is further used to obtain the object data set and object attribute data of the first object; the object data set and the object attribute data are respectively input into the feature extraction network for feature extraction to obtain the object data vector and the object attribute vector of the first object.

[0081] In a possible implementation, the acquisition module is also used to obtain identification information and multimodal information of the shared video, wherein the identification information includes at least one of the ID sequence information of the shared video, the author information of the shared video, and the tag attribute information of the shared video; the multimodal information of the shared video includes at least two of the shared video title segmentation information, the shared video title information, the shared video cover information, and the shared video content information; the identification information of the shared video and the multimodal information of the shared video are respectively input into the corresponding feature extraction network for feature extraction to obtain the identification feature vector and the multimodal feature vector of the shared video.

[0082] In a possible implementation, the acquisition module is further used to obtain identification information and multimodal information of the candidate video, where the identification information includes at least one of the ID sequence information of the candidate video, the author information of the candidate video, and the tag attribute information of the candidate video; the multimodal information of the candidate video includes at least two of the candidate video title segmentation information, the candidate video title information, the candidate video cover information, and the candidate video content information; the identification information of the candidate video and the multimodal information of the candidate video are respectively input into the corresponding feature extraction network for feature extraction to obtain the identification feature vector and the multimodal feature vector of the candidate video.

[0083] According to one aspect of the present application, a video recommendation model training device is provided, the device comprising:

[0084] an acquisition module, configured to acquire a multidimensional video feature vector of a shared video, a multidimensional object feature vector of a first object, a multidimensional video feature vector of a positive candidate video sample in a candidate video set, and a multidimensional video feature vector of a negative candidate video sample, wherein the shared video refers to a video clicked by the first object after receiving a video shared by the second object, the multidimensional video feature vector includes an identification feature vector and a multimodal feature vector, the positive candidate video sample refers to a video associated with the shared video, and the negative candidate video sample refers to a video not associated with the shared video;

[0085] a fusion module, configured to fuse the multidimensional video feature vector of the shared video and the multidimensional object feature vector to obtain an object fusion vector of the first object;

[0086] The fusion module is further configured to fuse the multi-dimensional video feature vectors of the same positive candidate video sample to obtain a positive candidate video sample fusion vector of the positive candidate video sample;

[0087] The fusion module is further configured to fuse the multi-dimensional video feature vectors of the same negative candidate video sample to obtain a negative candidate video sample fusion vector of the negative candidate video sample;

[0088] a calculation module, configured to perform similarity calculation based on the object fusion vector of the first object and the positive candidate video sample fusion vector to obtain a first similarity value between the object fusion vector and the positive candidate video sample fusion vector;

[0089] The calculation module is configured to perform similarity calculation based on the object fusion vector of the first object and the negative candidate video sample fusion vector to obtain a second similarity value between the object fusion vector and the negative candidate video sample fusion vector;

[0090] The calculation module is configured to calculate a loss function value based on the first similarity value and the second similarity value;

[0091] An updating module is used to update the model parameters of the video recommendation model based on the loss function value.

[0092] In one possible implementation, the calculation module is also used to input the identification feature vector of the shared video, the multimodal feature vector of the shared video, and the multidimensional object feature vector into the attention mechanism network for weight calculation to obtain the first weight value corresponding to each feature vector.

[0093] In a possible implementation, the fusion module is further used to input the identification feature vector of the shared video, the multimodal feature vector of the shared video, and the multidimensional object feature vector into the feature fusion network for feature fusion according to the first weight value to obtain the object fusion vector of the first object.

[0094] In a possible implementation, the fusion module is further used to input the identification feature vector of the shared video, the multimodal feature vector of the shared video, and the multidimensional object feature vector into the feature fusion network for feature fusion according to the first weight value to obtain an intermediate object fusion vector; and input the intermediate object fusion vector into the multi-layer perceptron network MLP for nonlinear fitting to obtain the object fusion vector of the first object.

[0095] In one possible implementation, the calculation module is further used to input the identification feature vector and the multimodal feature vector corresponding to the positive candidate video sample into the attention mechanism network for weight calculation to obtain a second weight value corresponding to each feature vector in the positive candidate video sample.

[0096] In a possible implementation, the fusion module is further configured to input the identification feature vector and the multimodal feature vector corresponding to the positive candidate video sample into the feature fusion network for feature fusion according to the second weight value to obtain the positive candidate video sample fusion vector.

[0097] In one possible implementation, the fusion module is further used to input the identification feature vector and the multimodal feature vector corresponding to the positive candidate video sample into a feature fusion network for feature fusion according to the second weight value to obtain an intermediate positive candidate video sample fusion vector; and input the intermediate positive candidate video sample fusion vector into a multi-layer perceptron network MLP for nonlinear fitting to obtain the positive candidate video sample fusion vector.

[0098] In one possible implementation, the calculation module is also used to input the identification feature vector and the multimodal feature vector corresponding to the i-th negative candidate video sample in the candidate video set into the attention mechanism network for weight calculation, and obtain the second weight value corresponding to each feature vector in the i-th negative candidate video sample, where i is a positive integer.

[0099] In one possible implementation, the fusion module is further configured to input the identification feature vector and the multimodal feature vector corresponding to the i-th negative candidate video sample into the feature fusion network for feature fusion according to the second weight value to obtain the i-th negative candidate video sample fusion vector.

[0100] In one possible implementation, the fusion module is further used to input the identification feature vector and the multimodal feature vector corresponding to the i-th negative candidate video sample into the feature fusion network for feature fusion according to the second weight value to obtain the intermediate i-th negative candidate video sample fusion vector; and input the intermediate i-th negative candidate video sample fusion vector into the multi-layer perceptron network MLP for nonlinear fitting to obtain the i-th negative candidate video sample fusion vector.

[0101] In a possible implementation, the calculation module is further configured to calculate a cosine similarity value between the object fusion vector and the positive candidate video sample fusion vector to obtain a first similarity value between the object fusion vector and the positive candidate video sample fusion vector.

[0102] In one possible implementation, the calculation module is further used to calculate the cosine similarity value between the object fusion vector and the i-th negative candidate video sample fusion vector in the candidate video set to obtain the i-th similarity value between the object fusion vector and the i-th negative candidate video sample fusion vector, where i is a positive integer.

[0103] In a possible implementation manner, the acquisition module is further configured to acquire a multi-dimensional object feature vector of the second object.

[0104] The fusion module is further configured to fuse the multidimensional video feature vector of the shared video, the multidimensional object feature vector of the first object, and the multidimensional object feature vector of the second object to obtain an object fusion vector of the first object.

[0105] In one possible implementation, the multidimensional object feature vector of the second object includes an object data vector and an object attribute vector. The acquisition module is further configured to acquire an object data set and object attribute data of the second object; and input the object data set and the object attribute data into a feature extraction network for feature extraction, thereby obtaining the object data vector and the object attribute vector of the second object.

[0106] In a possible implementation, the acquisition module is further configured to acquire the object data set and object attribute data of the first object;

[0107] The object data set and the object attribute data are respectively input into a feature extraction network for feature extraction to obtain the object data vector and the object attribute vector of the first object.

[0108] In a possible implementation, the acquisition module is also used to obtain identification information and multimodal information of the shared video, wherein the identification information includes at least one of the ID sequence information of the shared video, the author information of the shared video, and the tag attribute information of the shared video; the multimodal information of the shared video includes at least two of the shared video title segmentation information, the shared video title information, the shared video cover information, and the shared video content information; the identification information of the shared video and the multimodal information of the shared video are respectively input into the corresponding feature extraction network for feature extraction to obtain the identification feature vector and the multimodal feature vector of the shared video.

[0109] In a possible implementation, the acquisition module is further used to obtain identification information and multimodal information of the positive candidate video sample, wherein the identification information includes at least one of the ID sequence information of the positive candidate video sample, the author information of the positive candidate video sample, and the label attribute information of the positive candidate video sample; the multimodal information of the positive candidate video sample includes at least two of the positive candidate video sample title segmentation information, the positive candidate video sample title information, the positive candidate video sample cover information, and the positive candidate video sample content information; the identification information of the positive candidate video sample and the multimodal information of the positive candidate video sample are respectively input into the corresponding feature extraction network for feature extraction to obtain the identification feature vector and the multimodal feature vector of the positive candidate video sample.

[0110] In a possible implementation, the acquisition module is further used to obtain identification information and multimodal information of the negative candidate video sample, wherein the identification information includes at least one of the ID sequence information of the negative candidate video sample, the author information of the negative candidate video sample, and the label attribute information of the negative candidate video sample; the multimodal information of the negative candidate video sample includes at least two of the negative candidate video sample title segmentation information, the negative candidate video sample title information, the negative candidate video sample cover information, and the negative candidate video sample content information; the identification information of the negative candidate video sample and the multimodal information of the negative candidate video sample are respectively input into the corresponding feature extraction network for feature extraction to obtain the identification feature vector and the multimodal feature vector of the negative candidate video sample.

[0111] According to another aspect of the present application, a computer device is provided, comprising: a processor and a memory, wherein at least one computer program is stored in the memory, and the at least one computer program is loaded and executed by the processor to implement the video recommendation method or the video recommendation model training method as described above.

[0112] According to another aspect of the present application, a computer storage medium is provided, in which at least one computer program is stored. The at least one computer program is loaded and executed by a processor to implement the video recommendation method or the video recommendation model training method described above.

[0113] According to another aspect of the present application, a computer program product is provided, which includes a computer program stored in a computer-readable storage medium; the computer program is read and executed from the computer-readable storage medium by a processor of a computer device, so that the computer device performs the video recommendation method or the video recommendation model training method described above.

[0114] The beneficial effects of the technical solution provided by this application include at least:

[0115] By obtaining the identification feature vector of the shared video, the multimodal feature vector of the shared video, the multidimensional object feature vector of the first object, the identification feature vector of the candidate video and the multimodal feature vector of the candidate video; fusing the identification feature vector of the shared video, the multimodal feature vector of the shared video and the multidimensional object feature vector to obtain an object fusion vector; fusing the identification feature vector of the same candidate video and the multimodal feature vector of the candidate video to obtain a candidate video fusion vector; calculating the similarity value between the object fusion vector and the candidate video fusion vector, and recommending the video corresponding to the top similarity value to the first object according to the sorting of the similarity values. This application improves the accuracy of video recommendation by extracting the multimodal feature vector and the multidimensional object feature vector of the shared video to make a comprehensive judgment on the candidate video. BRIEF DESCRIPTION OF THE DRAWINGS

[0116] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0117] Figure 1 is a schematic diagram of a video recommendation method provided by an exemplary embodiment of the present application;

[0118] Figure 2 is a schematic diagram of the architecture of a computer system provided by an exemplary embodiment of the present application;

[0119] Figure 3 is a flowchart of a video recommendation method provided by an exemplary embodiment of the present application;

[0120] Figure 4 is a flowchart of a video recommendation method provided by an exemplary embodiment of the present application;

[0121] Figure 5 is a schematic diagram of a recommended video display interface provided by an exemplary embodiment of the present application;

[0122] Figure 6 This is a framework diagram of video recommendation model generation and video recommendation provided by an exemplary embodiment of the present application;

[0123] Figure 7 is a flowchart of a method for training a video recommendation model provided by an exemplary embodiment of the present application;

[0124] Figure 8 is a flowchart of a method for training a video recommendation model provided by an exemplary embodiment of the present application;

[0125] Figure 9 is a flowchart of a method for training a video recommendation model provided by an exemplary embodiment of the present application;

[0126] Figure 10 is a flowchart of a video recommendation method provided by an exemplary embodiment of the present application;

[0127] Figure 11 is a block diagram of a video recommendation device provided by an exemplary embodiment of the present application;

[0128] Figure 12 is a block diagram of a training device for a video recommendation model provided by an exemplary embodiment of the present application;

[0129] Figure 13 It is a structural diagram of a computer device provided by an exemplary embodiment of the present application. DETAILED DESCRIPTION

[0130] In order to make the objectives, technical solutions and advantages of this application clearer, the implementation methods of this application will be further described in detail below with reference to the accompanying drawings.

[0131] The embodiment of the present application provides a technical solution for a video recommendation method, such as Figure 1 The schematic diagram of the video recommendation method shown is that the method can be executed by a computer device, which can be a terminal or a server.

[0132] The computer device obtains a multidimensional video feature vector of the shared video, a multidimensional object feature vector of the first object, and multidimensional video feature vectors of candidate videos in the candidate video set.

[0133] Optionally, the shared video refers to a video that the first object clicks after receiving the video shared by the second object.

[0134] Optionally, the multi-dimensional video feature vector includes an identification feature vector and a multimodal feature vector.

[0135] Exemplarily, the computer device obtains the identification information 103 and multimodal information of the shared video; the computer device inputs the identification information 103 of the shared video and the multimodal information of the shared video into the corresponding feature extraction network 107 for feature extraction, and obtains the identification feature vector 110 of the shared video and the multimodal feature vector 127 of the shared video.

[0136] Optionally, the identification information 103 of the shared video includes at least one of ID sequence information of the shared video, author information of the shared video, and tag attribute information of the shared video.

[0137] The multimodal information of the shared video includes at least two of the shared video title segmentation information 104 , the shared video title information 105 , the shared video cover information 106 and the shared video content information.

[0138] For example, the computer device inputs the identification information 103 of the shared video into the corresponding feature extraction network 107 for feature extraction, and obtains the shared video identification feature vector 110; inputs the shared video title segmentation information 104 into the corresponding feature extraction network 107 for feature extraction, and obtains the shared video title segmentation feature vector 111; inputs the shared video title information 105 into the corresponding feature extraction network 107 for feature extraction, and obtains the shared video title feature vector 112; inputs the shared video cover information 106 into the corresponding feature extraction network 107 for feature extraction, and obtains the shared video cover feature vector 113.

[0139] Exemplarily, the computer device obtains identification information and multimodal information of the candidate video, and the computer device inputs the identification information of the candidate video and the multimodal information of the candidate video into the corresponding feature extraction network 107 for feature extraction, thereby obtaining the identification feature vector and multimodal feature vector of the candidate video.

[0140] Optionally, the identification information of the candidate video includes at least one of ID sequence information of the candidate video, author information of the candidate video, and tag attribute information of the candidate video.

[0141] The multimodal information of the candidate video includes at least two of the candidate video title segmentation information, candidate video title information, candidate video cover information and candidate video content information.

[0142] For example, taking the i-th candidate video as an example, the computer device inputs the identification information 114 of the i-th candidate video into the corresponding feature extraction network 107 for feature extraction, and obtains the i-th candidate video identification feature vector 118; the i-th candidate video title segmentation information 115 is input into the corresponding feature extraction network 107 for feature extraction, and obtains the i-th candidate video title segmentation feature vector 119; the i-th candidate video title information 116 is input into the corresponding feature extraction network 107 for feature extraction, and obtains the i-th candidate video title feature vector 120; the i-th candidate video cover information 117 is input into the corresponding feature extraction network 107 for feature extraction, and obtains the i-th candidate video cover feature vector 121.

[0143] Optionally, the multi-dimensional object feature vector 126 of the first object includes the object data vector 108 and the object attribute vector 109 .

[0144] For example, the computer device obtains the object data set 101 and object attribute data 102 of the first object; the computer device inputs the object data set 101 and object attribute data 102 into the feature extraction network 107 for feature extraction, and obtains the object data vector 108 and object attribute vector 109 of the first object.

[0145] The object data set 101 refers to a set of object behavior sequences. For example, the object data set 101 includes at least one of likes, resume, play duration, collection, and forwarding, but is not limited to this. The embodiment of the present application does not make specific limitations on this.

[0146] The object attribute data 102 includes at least one of the object's age, gender, and hobbies, but is not limited thereto. This embodiment of the present application does not make any specific limitation thereto.

[0147] Exemplarily, the computer device inputs the identification feature vector 110 of the shared video, the multimodal feature vector 127 of the shared video, and the multidimensional object feature vector 126 into the attention mechanism network 122 for weight calculation to obtain a first weight value corresponding to each feature vector; the computer device performs feature fusion on the identification feature vector 110 of the shared video, the multimodal feature vector 127 of the shared video, and the multidimensional object feature vector 126 in the feature fusion network 123 according to the first weight value to obtain an intermediate object fusion vector; the computer device inputs the intermediate object fusion vector into the multi-layer perceptron network 124 (Multi-Layer Perception, MLP) for nonlinear fitting to obtain an object fusion vector 125 of the first object.

[0148] Exemplarily, the computer device inputs the i-th candidate video identification class feature vector 118 and the multimodal feature vector 128 in the candidate video set into the attention mechanism network 122 for weight calculation to obtain a second weight value corresponding to each feature vector in the i-th candidate video, where i is a positive integer; the computer device performs feature fusion on the i-th candidate video identification class feature vector 118 and the multimodal feature vector 128 in the feature fusion network 123 according to the second weight value to obtain an intermediate i-th candidate video fusion vector; the computer device inputs the intermediate i-th candidate video fusion vector into the multi-layer perceptron network MLP124 for nonlinear fitting to obtain an i-th candidate video fusion vector 130.

[0149] Exemplarily, the computer device calculates a cosine similarity value between the object fusion vector 125 and the i-th candidate video fusion vector 130 in the candidate video set to obtain an i-th similarity value 129 between the object fusion vector 125 and the i-th candidate video fusion vector 130 .

[0150] The computer device recommends videos corresponding to higher-ranked similarity values ​​to the first object based on the ranking of the similarity values.

[0151] For example, the computer device calculates the cosine similarity value of the object fusion vector 125 and the 100 candidate video fusion vectors in the candidate video set in sequence to obtain 100 similarity values. The computer device recommends the candidate videos corresponding to the top ten similarity values ​​to the first object based on the sorting of the similarity values.

[0152] In summary, the method provided in this embodiment obtains the identification feature vector of the shared video, the multimodal feature vector of the shared video, the multidimensional object feature vector of the first object, the identification feature vector of the candidate video, and the multimodal feature vector of the candidate video; the computer device fuses the identification feature vector of the shared video, the multimodal feature vector of the shared video, and the multidimensional object feature vector to obtain an object fusion vector; the identification feature vector of the same candidate video and the multimodal feature vector of the candidate video are fused to obtain a candidate video fusion vector; the similarity value between the object fusion vector and the candidate video fusion vector is calculated, and the candidate video is recommended to the first object based on the similarity value. This application improves the accuracy of video recommendation by using the multimodal feature vector and multidimensional object feature vector of the shared video to make a comprehensive judgment on the candidate video.

[0153] Figure 2 The schematic diagram of the computer system architecture provided by one embodiment of the present application is shown. The computer system may include: a terminal 100 and a server 200.

[0154] The terminal 100 can be an electronic device such as a mobile phone, a tablet computer, a vehicle-mounted terminal (vehicle computer), a wearable device, a personal computer (PC), an intelligent voice interaction device, a smart home appliance, a vehicle-mounted terminal, an aircraft, an unmanned vending terminal, etc. The terminal 100 can be installed with a client that runs a target application. The target application can be an application that supports video recommendation or other application that provides a video recommendation function, and this application does not limit this. In addition, this application does not limit the form of the target application, including but not limited to an application (Application, App) installed in the terminal 100, a mini-program, etc., and can also be in the form of a web page.

[0155] Server 200 can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud computing services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. Server 200 can be the backend server of the target application described above, used to provide backend services to the client of the target application.

[0156] Cloud technology refers to a hosting technology that unifies hardware, software, and network resources within a wide area network (WAN) or local area network (LAN) to enable data computing, storage, processing, and sharing. Cloud technology is a general term for network technology, information technology, integration technology, management platform technology, and application technology based on the cloud computing business model. It can form a resource pool that can be used on demand with flexibility and convenience. Cloud computing technology will become a crucial support. Backend services for technical network systems, such as video websites, image websites, and more portals, require extensive computing and storage resources. With the rapid development and application of the internet industry, every item will likely have its own unique identification mark and will need to be transmitted to backend systems for logical processing. Data of varying levels will be processed separately, and data from all industries will require a strong system backend, which can only be achieved through cloud computing.

[0157] In some embodiments, the aforementioned servers can also be implemented as nodes in a blockchain system. Blockchain is a novel application model for computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms. Essentially, a blockchain is a decentralized database, a series of data blocks generated using cryptographic methods. Each block contains information about a batch of network transactions, used to verify the validity of the information (to prevent counterfeiting) and generate the next block. A blockchain can include an underlying blockchain platform, a platform product and service layer, and an application service layer.

[0158] The terminal 100 and the server 200 may communicate with each other via a network, such as a wired or wireless network.

[0159] In the video recommendation method or video recommendation model training method provided in the embodiment of the present application, the execution subject of each step can be a computer device, which refers to an electronic device with data calculation, processing and storage capabilities. Figure 2 Taking the implementation environment of the scheme shown as an example, the video recommendation method or the training method of the video recommendation model can be executed by the terminal 100 (such as the client of the target application installed and running in the terminal 100 executes the video recommendation method or the training method of the video recommendation model), or the video recommendation method or the training method of the video recommendation model can be executed by the server 200, or the terminal 100 and the server 200 can interact and cooperate to execute it, and this application does not limit this.

[0160] Figure 3 This is a flowchart of a video recommendation method provided by an exemplary embodiment of the present application. The method can be executed by a computer device, which can be Figure 2 The terminal 100 or the server 200 in the embodiment of the present invention. The method includes:

[0161] Step 302: Obtain a multidimensional video feature vector of the shared video, a multidimensional object feature vector of the first object, and multidimensional video feature vectors of candidate videos in the candidate video set.

[0162] A shared video refers to a video that is clicked by a first subject after being shared by a second subject.

[0163] The shared video can be any video in the local video library of the second object's terminal, or any video shared by the second object from the cloud to the first object, and can include at least one continuous video frame. The first object is the object corresponding to the master computer device, and the second object can be an object corresponding to any computer device.

[0164] Optionally, the shared video is at least one of a video recorded by the second object itself, a video obtained by the second object from a stored database, and a video searched by the second object from the cloud, but is not limited to this. The embodiments of the present application are not limited to this.

[0165] Exemplarily, the multi-dimensional object feature vector refers to a vector feature corresponding to the first object, and the multi-dimensional object feature vector includes an object data vector and an object attribute vector.

[0166] Exemplarily, the multi-dimensional video feature vector includes an identification feature vector and a multimodal feature vector.

[0167] The identification feature vector refers to the feature vector related to the video attributes, and the multimodal feature vector refers to the feature vector related to the semantics of the video content.

[0168] Video attributes refer to video characteristics that are unrelated to the semantics of video content, such as at least one of a video ID and a video author ID, but are not limited thereto and are not limited to these in the embodiments of the present application.

[0169] Video content semantics refers to at least one of the text information, image information, and audio information of the video, but is not limited thereto and is not limited to this in the embodiments of the present application.

[0170] Optionally, the identification information package includes at least one of the ID sequence information of the video, the author information of the video, and the tag attribute information of the video, but is not limited to this. The embodiments of the present application do not limit this.

[0171] The multimodal feature vector refers to the feature vector obtained by extracting the multimodal information of a video.

[0172] Optionally, the multimodal feature vector includes at least two of video title segmentation information, video title information, video cover information and video content information, but is not limited thereto and is not limited in this embodiment of the present application.

[0173] Modality: It is used to represent the source or form of information. Each source or form of information can be called a modality. For example, humans have touch and hearing; information media include voice, video, text, etc.; various sensors include radar, infrared, accelerometers, etc., each of which can be called a modality.

[0174] In an embodiment of the present application, the modality is used to indicate a data set from a video, such as: obtaining shared video title segmentation information, shared video title information, shared video cover information, and shared video content information.

[0175] Step 304: Fusing the multi-dimensional video feature vector and the multi-dimensional object feature vector of the shared video to obtain an object fusion vector of the first object.

[0176] The object fusion vector refers to a feature obtained by fusing the identification feature vector of the shared video, the multimodal feature vector of the shared video, and the multidimensional object feature vector of the first object.

[0177] Optionally, the multimodal feature vector of the shared video includes at least two of the shared video title segmentation feature vector, the shared video title feature vector, the shared video cover feature vector, and the candidate video content feature vector, but is not limited to this. The embodiments of the present application do not limit this.

[0178] For example, a multimodal feature vector of a shared video includes a shared video title word segmentation feature vector, a shared video title feature vector, and a shared video cover feature vector; a multidimensional object feature vector includes an object data vector and an object attribute vector. The computer device fuses the shared video's identification feature vector, the shared video title word segmentation feature vector, the shared video title feature vector, the shared video cover feature vector, the object data vector, and the object attribute vector to generate an object fusion vector.

[0179] Step 306: Fusing the multi-dimensional video feature vectors of the same candidate video to obtain a candidate video fusion vector of the candidate video.

[0180] The candidate video fusion vector refers to a feature obtained by fusing the identification feature vector of the candidate video and the multimodal feature vector of the candidate video.

[0181] Optionally, the multimodal feature vector of the candidate video includes at least two of the candidate video title segmentation feature vector, the candidate video title feature vector, the candidate video cover feature vector, and the candidate video content feature vector, but is not limited to this and is not limited to this in the embodiments of the present application.

[0182] For example, the computer device fuses the candidate video's identification feature vector, the candidate video title word segmentation feature vector, the candidate video title feature vector, the candidate video cover feature vector, and the candidate video content feature vector to obtain a candidate video fusion vector.

[0183] Step 308: performing similarity calculation based on the object fusion vector of the first object and the candidate video fusion vector to obtain a similarity value between the object fusion vector and the candidate video fusion vector.

[0184] The similarity value is used to indicate the degree of association between the shared video and the candidate video. For example, the greater the similarity value, the higher the degree of association between the shared video and the candidate video.

[0185] Exemplarily, the computer device performs similarity calculation based on the object fusion vector of the first object and all or part of the candidate video fusion vectors in the candidate video set to obtain a similarity value between the object fusion vector and the candidate video fusion vectors.

[0186] Step 310: Recommend candidate videos to the first object based on the similarity value.

[0187] The computer device calculates similarity values ​​between the object fusion vector and some or all candidate video fusion vectors in the candidate video set, and recommends videos corresponding to higher-ranked similarity values ​​to the first object based on the order of the similarity values.

[0188] For example, the computer device calculates similarity values ​​between the object fusion vector and the 100 candidate video fusion vectors in the candidate video set in sequence to obtain 100 similarity values. The computer device recommends the candidate videos corresponding to the top ten similarity values ​​to the first object based on the sorting of the similarity values.

[0189] In summary, the method provided in this embodiment obtains the identification feature vector of the shared video, the multimodal feature vector of the shared video, the multidimensional object feature vector of the first object, the identification feature vector of the candidate video, and the multimodal feature vector of the candidate video; the computer device fuses the identification feature vector of the shared video, the multimodal feature vector of the shared video, and the multidimensional object feature vector to obtain an object fusion vector; the identification feature vector of the same candidate video and the multimodal feature vector of the candidate video are fused to obtain a candidate video fusion vector; the similarity value between the object fusion vector and the candidate video fusion vector is calculated, and the candidate video is recommended to the first object based on the similarity value. This application improves the accuracy of video recommendation by using the multimodal feature vector and multidimensional object feature vector of the shared video to make a comprehensive judgment on the candidate video.

[0190] An embodiment of the present application provides a video recommendation model, which includes: a feature extraction network, an attention mechanism network, a feature fusion network, and a multi-layer perceptron network.

[0191] The computer device obtains identification information of the shared video, multimodal information of the shared video, an object data set of the first object, and object attribute data of the first object, and performs feature extraction through a feature extraction network to obtain an identification feature vector of the shared video, a multimodal feature vector, and a multidimensional object feature vector of the first object.

[0192] The computer device obtains identification information and multimodal information of the candidate video, and inputs the identification information of the candidate video and the multimodal information of the candidate video into the corresponding feature extraction network for feature extraction, thereby obtaining an identification feature vector and a multimodal feature vector of the candidate video.

[0193] The computer device inputs the identification feature vector of the shared video, the multimodal feature vector of the shared video, and the multidimensional object feature vector into the attention mechanism network for weight calculation to obtain a first weight value corresponding to each feature vector; the computer device performs feature fusion on the identification feature vector of the shared video, the multimodal feature vector of the shared video, and the multidimensional object feature vector in the feature fusion network according to the first weight value to obtain an intermediate object fusion vector; the computer device inputs the intermediate object fusion vector into the multi-layer perceptron network MLP for nonlinear fitting to obtain an object fusion vector of the first object.

[0194] The computer device inputs the candidate video identification class feature vector and the multimodal feature vector in the candidate video set into the attention mechanism network for weight calculation to obtain a second weight value corresponding to each feature vector in the candidate video; the computer device performs feature fusion on the candidate video identification class feature vector and the multimodal feature vector in the feature fusion network according to the second weight value to obtain an intermediate candidate video fusion vector; the computer device inputs the intermediate candidate video fusion vector into the multi-layer perceptron network MLP for nonlinear fitting to obtain a candidate video fusion vector.

[0195] The computer device calculates a cosine similarity value between the object fusion vector and the candidate video fusion vectors in the candidate video set to obtain a similarity value between the object fusion vector and the candidate video fusion vector.

[0196] The computer device recommends videos corresponding to higher-ranked similarity values ​​to the first object based on the ranking of the similarity values.

[0197] Based on this video recommendation model, the following video recommendation method is provided.

[0198] Figure 4 This is a flowchart of a video recommendation method provided by an exemplary embodiment of the present application. The method can be executed by a computer device, which can be Figure 2 The terminal 100 or the server 200 in the embodiment of the present invention. The method includes:

[0199] Step 402: Obtain a multidimensional video feature vector of the shared video, a multidimensional object feature vector of the first object, and multidimensional video feature vectors of candidate videos in the candidate video set.

[0200] A shared video refers to a video that is clicked by a first subject after being shared by a second subject.

[0201] Exemplarily, the multi-dimensional object feature vector of the first object refers to a vector feature corresponding to the first object, and the multi-dimensional object feature vector includes an object data vector and an object attribute vector.

[0202] The computer device obtains an object data set and object attribute data of the first object; and inputs the object data set and the object attribute data into a feature extraction network for feature extraction, thereby obtaining an object data vector and an object attribute vector of the first object.

[0203] In one possible implementation, the computer device obtains an object data set and object attribute data of the second object; and inputs the object data set and object attribute data into a feature extraction network for feature extraction, thereby obtaining an object data vector and an object attribute vector of the second object.

[0204] An object data set refers to a set of object behavior sequences. For example, an object data set includes at least one of likes, resume, play duration, collection, and forwarding, but is not limited to this. The embodiments of the present application do not make specific limitations on this.

[0205] The object attribute data includes at least one of the object's age, the object's gender, the object's hobbies, and the object's location, but is not limited thereto. The embodiments of the present application do not make specific limitations on this.

[0206] Exemplarily, the multi-dimensional video feature vector of the shared video includes an identification feature vector of the shared video and a multimodal feature vector of the shared video.

[0207] The computer device obtains identification information and multimodal information of the shared video, and inputs the identification information of the shared video and the multimodal information of the shared video into the corresponding feature extraction network for feature extraction, thereby obtaining an identification feature vector and a multimodal feature vector of the shared video.

[0208] Among them, the identification information of the shared video includes at least one of the ID sequence information of the shared video, the author information of the shared video, and the tag attribute information of the shared video; the multimodal information of the shared video includes at least two of the shared video title segmentation information, the shared video title information, the shared video cover information and the shared video content information, but is not limited to this. The embodiments of the present application do not make specific limitations on this.

[0209] For example, the computer device inputs the identification information of the shared video into the corresponding feature extraction network for feature extraction to obtain the shared video identification feature vector; inputs the shared video title word segmentation information into the corresponding feature extraction network for word granularity level feature extraction to obtain the shared video title word segmentation feature vector; inputs the shared video title information into the corresponding feature extraction network, and divides the shared video title information by word and performs feature extraction to obtain the shared video title feature vector; inputs the shared video cover information into the corresponding residual neural network (ResNet), and obtains the shared video cover feature vector by performing feature extraction on the image information.

[0210] Exemplarily, the multi-dimensional video feature vector of the candidate video includes an identification feature vector of the candidate video and a multimodal feature vector of the candidate video.

[0211] The computer device obtains identification information and multimodal information of the candidate video, and inputs the identification information of the candidate video and the multimodal information of the candidate video into the corresponding feature extraction network for feature extraction, thereby obtaining an identification feature vector and a multimodal feature vector of the candidate video.

[0212] Among them, the identification information includes at least one of the ID sequence information of the candidate video, the author information of the candidate video, and the tag attribute information of the candidate video. The multimodal information of the candidate video includes at least two of the candidate video title segmentation information, the candidate video title information, the candidate video cover information and the candidate video content information, but is not limited to this. The embodiments of the present application do not make specific limitations on this.

[0213] For example, the computer device inputs the identification information of the candidate video into the corresponding feature extraction network for feature extraction to obtain the candidate video identification feature vector; inputs the candidate video title segmentation information into the corresponding feature extraction network for word granularity level feature extraction to obtain the candidate video title segmentation feature vector; inputs the candidate video title information into the corresponding feature extraction network, and divides the candidate video title information by word and performs feature extraction to obtain the candidate video title feature vector; inputs the candidate video cover information into the corresponding residual neural network (ResNet), and obtains the candidate video cover feature vector by performing feature extraction on the image information.

[0214] Step 404: Fusing the identification feature vector, the multimodal feature vector, and the multidimensional object feature vector of the shared video to obtain an object fusion vector of the first object.

[0215] The object fusion vector refers to a feature obtained by fusing the identification feature vector of the shared video, the multimodal feature vector of the shared video, and the multidimensional object feature vector of the first object.

[0216] In one possible implementation, a computer device inputs the identification feature vector of the shared video, the multimodal feature vector of the shared video, and the multidimensional object feature vector of the first object into the attention mechanism network for weight calculation to obtain a first weight value corresponding to each feature vector.

[0217] The computer device performs feature fusion on the identification feature vector of the shared video, the multimodal feature vector of the shared video, and the multidimensional object feature vector of the first object according to the first weight value to obtain an object fusion vector of the first object.

[0218] Optionally, the computer device performs feature fusion on the identification feature vector of the shared video, the multimodal feature vector of the shared video, and the multidimensional object feature vector of the first object according to the first weight value to obtain an intermediate object fusion vector; the computer device inputs the intermediate object fusion vector into the multi-layer perceptron network MLP for nonlinear fitting to obtain the object fusion vector of the first object.

[0219] In one possible implementation, the computer device obtains a multidimensional object feature vector of the second object; the computer device fuses the multidimensional video feature vector of the shared video, the multidimensional object feature vector of the first object, and the multidimensional object feature vector of the second object to obtain an object fusion vector of the first object.

[0220] The second object is an object that shares a video with the first object.

[0221] Optionally, the first object is an object corresponding to the master computer device, and the second object may be an object corresponding to any computer device.

[0222] In one possible implementation, a computer device inputs the identification feature vector of the shared video, the multimodal feature vector of the shared video, the multidimensional object feature vector of the first object, and the multidimensional object feature vector of the second object into the attention mechanism network for weight calculation to obtain a first weight value corresponding to each feature vector.

[0223] The computer device performs feature fusion on the identification feature vector of the shared video, the multimodal feature vector of the shared video, the multidimensional object feature vector of the first object, and the multidimensional object feature vector of the second object according to the first weight value to obtain an object fusion vector of the first object.

[0224] Optionally, the computer device performs feature fusion on the identification feature vector of the shared video, the multimodal feature vector of the shared video, the multidimensional object feature vector of the first object, and the multidimensional object feature vector of the second object according to the first weight value to obtain an intermediate object fusion vector; the computer device inputs the intermediate object fusion vector into the multi-layer perceptron network MLP for nonlinear fitting to obtain the object fusion vector of the first object.

[0225] In a possible implementation manner, the computer device obtains a multi-dimensional object feature vector of the second object and obtains a first directivity feature vector between the second object and the first object.

[0226] The first directional feature vector is used to represent that the shared video is shared from the second object to the first object.

[0227] For example, the first directional feature vector can be expressed as: the second object in the first age group shares a video with the first object in the second age group; or, the second object in the first area shares a video with the first object in the second area; but it is not limited to this, and the embodiments of the present application are not limited to this.

[0228] The computer device fuses the multidimensional video feature vector of the shared video, the multidimensional object feature vector of the first object, the multidimensional object feature vector of the second object, and the first directional feature vector to obtain an object fusion vector of the first object.

[0229] For example, the formula for fusing the identification feature vector, multimodal feature vector, and multidimensional object feature vector of the shared video can be expressed as:

[0230]

[0231] Where, output-feature-user represents the object fusion vector, w i represents any feature vector among the identity feature vector, multimodal feature vector and multidimensional object feature vector, F i It represents the first weight value corresponding to each feature vector in the identification feature vector, multimodal feature vector and multidimensional object feature vector, and N represents the total number of identification feature vectors, multimodal feature vectors and multidimensional object feature vectors.

[0232] Step 406: Fusing the identification feature vector and the multimodal feature vector of the same candidate video to obtain a candidate video fusion vector of the candidate video.

[0233] The candidate video fusion vector refers to a feature obtained by fusing the identification feature vector of the candidate video and the multimodal feature vector of the candidate video.

[0234] Optionally, the multimodal feature vector of the candidate video includes at least two of the candidate video title segmentation feature vector, the candidate video title feature vector, the candidate video cover feature vector, and the candidate video content feature vector, but is not limited to this and is not limited to this in the embodiments of the present application.

[0235] For example, the computer device fuses the candidate video's identification feature vector, the candidate video title word segmentation feature vector, the candidate video title feature vector, the candidate video cover feature vector, and the candidate video content feature vector to obtain a candidate video fusion vector.

[0236] In one possible implementation, a computer device inputs the identification feature vector and the multimodal feature vector corresponding to the i-th candidate video in the candidate video set into the attention mechanism network for weight calculation to obtain a second weight value corresponding to each feature vector in the i-th candidate video, where i is a positive integer; the computer device performs feature fusion on the identification feature vector and the multimodal feature vector corresponding to the i-th candidate video according to the second weight value to obtain a fusion vector of the i-th candidate video.

[0237] Optionally, the computer device performs feature fusion on the identification feature vector and the multimodal feature vector corresponding to the i-th candidate video according to the second weight value to obtain an intermediate i-th candidate video fusion vector; and inputs the intermediate i-th candidate video fusion vector into the multi-layer perceptron network MLP for nonlinear fitting to obtain the i-th candidate video fusion vector.

[0238] For example, taking the i-th candidate video as an example, Figure 1 As shown, the computer device inputs the identification information of the i-th candidate video into the corresponding feature extraction network 107 for feature extraction, and obtains the i-th candidate video identification feature vector 118; the i-th candidate video title segmentation information 115 is input into the corresponding feature extraction network 107 for feature extraction, and obtains the i-th candidate video title segmentation feature vector 119; the i-th candidate video title information 116 is input into the corresponding feature extraction network 107 for feature extraction, and obtains the i-th candidate video title feature vector 120; the i-th candidate video cover information 117 is input into the corresponding feature extraction network 107 for feature extraction, and obtains the i-th candidate video cover feature vector 121.

[0239] The computer device inputs the identification feature vector 118, the i-th candidate video title segmentation feature vector 119, the i-th candidate video title feature vector 120, and the i-th candidate video cover feature vector 121 corresponding to the i-th candidate video in the candidate video set into the attention mechanism network 122 for weight calculation to obtain the second weight value corresponding to each feature vector in the i-th candidate video.

[0240] The computer device performs feature fusion on the identification feature vector 118, the i-th candidate video title word segmentation feature vector 119, the i-th candidate video title feature vector 120, and the i-th candidate video cover feature vector 121 corresponding to the i-th candidate video in the feature fusion network 123 according to the second weight value to obtain an intermediate i-th candidate video fusion vector; and inputs the intermediate i-th candidate video fusion vector into the multi-layer perceptron network MLP124 for nonlinear fitting to obtain an i-th candidate video fusion vector 130.

[0241] For example, the formula for fusing the identification feature vector and the multimodal feature vector corresponding to the same candidate video can be expressed as:

[0242]

[0243] Where, output-feature-feed represents the candidate video fusion vector, v i Represents any feature vector in the identity feature vector or multimodal feature vector, P i represents the second weight value corresponding to each feature vector in the identification feature vector and the multimodal feature vector, and M represents the total number of the identification feature vector and the multimodal feature vector.

[0244] In a possible implementation, the computer device obtains a second directional feature vector of a candidate video in the candidate video set.

[0245] The second directional feature vector is used to represent that the candidate video is shared from the third object to the fourth object in the historical sharing record.

[0246] For example, the second directional feature vector can be expressed as: in the historical sharing records of candidate videos in the candidate video set, candidate video A was shared by a third object in the first age group to a fourth object in the second age group within the first time period; or candidate video B was shared by a third object in the first region to a fourth object in the second region within the first time period; but not limited to this, the embodiments of the present application are not limited to this.

[0247] The computer device fuses the multidimensional video feature vector of the same candidate video, the multidimensional object feature vector of the third object, the multidimensional object feature vector of the fourth object, and the second directional feature vector to obtain a candidate video fusion vector of the candidate video.

[0248] Step 408: performing similarity calculation based on the object fusion vector of the first object and the candidate video fusion vector to obtain a similarity value between the object fusion vector and the candidate video fusion vector.

[0249] The similarity value is used to indicate the degree of association between the shared video and the candidate video. For example, the greater the similarity value, the higher the degree of association between the shared video and the candidate video.

[0250] Exemplarily, the computer device calculates a cosine similarity value between the object fusion vector and the i-th candidate video fusion vector in the candidate video set to obtain an i-th similarity value between the object fusion vector and the i-th candidate video fusion vector, where i is a positive integer.

[0251] For example, the calculation formula of the cosine similarity value can be expressed as:

[0252]

[0253] Where y Q Represents the object fusion vector, y D represents the candidate video fusion vector, and R represents the cosine similarity value.

[0254] Step 410: Recommend candidate videos to the first object based on the similarity value.

[0255] The computer device calculates similarity values ​​between the object fusion vector and some or all candidate video fusion vectors in the candidate video set, and recommends videos corresponding to higher-ranked similarity values ​​to the first object based on the order of the similarity values.

[0256] For example, the computer device calculates similarity values ​​between the object fusion vector and the 100 candidate video fusion vectors in the candidate video set in sequence to obtain 100 similarity values. The computer device recommends the candidate videos corresponding to the top ten similarity values ​​to the first object based on the sorting of the similarity values.

[0257] For example, in response to a first subject clicking on a video shared by a second subject, the shared video is played on the user interface, and at the same time, a recommended video display interface is displayed. The recommended video display interface displays recommended videos, and the subject can switch to playing the recommended video on the user interface by clicking on the recommended video.

[0258] Optionally, the recommended video display interface may include at least one user interaction (UI) card, each user interaction card is used to display a recommended video, or the video display interface may also include at least one window, each window is used to display a recommended video, or the video display interface displays a list of recommended videos, or the video display interface displays video information of recommended videos. The embodiment of the present application does not specifically limit the form of displaying recommended videos in the video display interface.

[0259] Optionally, the video information of the recommended video displayed on the video display interface may be at least one of a thumbnail, a web link, and a title of the recommended video. For example, for a certain recommended video, the video information may include a thumbnail, a web link, a title, an author, and a summary of the recommended video. The embodiment of the present application does not specifically limit the content of the video information.

[0260] For example, Figure 5 This is a schematic diagram of the recommended video display interface provided by this embodiment. Figure 5 In response to the first subject clicking on the video shared by the second subject, the shared video 502 is played on the user interface 501, and at the same time, the recommended video display interface 503 is displayed. Based on the similarity values ​​between the object fusion vector of the first subject and the fusion vectors of some or all candidate videos in the candidate video set, the computer device recommends videos corresponding to the top-ranked similarity values ​​to the first subject according to the sorting of the similarity values. That is, at least one window is displayed on the video display interface 503, each window being used to display a recommended video. By clicking on a recommended video, the recommended video is switched to be played on the user interface 501.

[0261] In summary, the method provided in this embodiment obtains the identification feature vector of the shared video, the multimodal feature vector of the shared video, the multidimensional object feature vector of the first object, the identification feature vector of the candidate video, and the multimodal feature vector of the candidate video; the computer device fuses the identification feature vector of the shared video, the multimodal feature vector of the shared video, and the multidimensional object feature vector to obtain an object fusion vector; the identification feature vector of the same candidate video and the multimodal feature vector of the candidate video are fused to obtain a candidate video fusion vector; and the fusion process is described, by calculating the cosine similarity value between the object fusion vector and the candidate video fusion vector, and recommending the candidate video to the first object based on the cosine similarity value. This application improves the accuracy of video recommendation by using the multimodal feature vector and multidimensional object feature vector of the shared video to make a comprehensive judgment on the candidate video.

[0262] The video recommendation method involved in this application can be implemented based on a video recommendation model. The solution includes a video recommendation model generation stage and a video recommendation stage. Figure 6 This is a framework diagram of a video recommendation model generation and video recommendation shown in an exemplary embodiment of the present application. Figure 6As shown, in the video recommendation model generation phase, the video recommendation model generation device 610 obtains a video recommendation model using a pre-set training sample video set and then generates video recommendation results based on the video recommendation model. In the video recommendation phase, the video recommendation device 620 processes the received shared video, the object data set of the first object, and the object attribute data based on the video recommendation model to obtain video recommendation results for the shared video, such as determining the three most relevant recommended videos for the shared video.

[0263] Among them, the above-mentioned video recommendation model generation device 610 and video recommendation device 620 can be computer devices. For example, the computer device can be a fixed computer device such as a personal computer or a server, or the computer device can also be a mobile computer device such as a tablet computer or an e-book reader.

[0264] Optionally, the above-mentioned video recommendation model generation device 610 and video recommendation device 620 can be the same device, or the video recommendation model generation device 610 and the video recommendation device 620 can also be different devices. Moreover, when the video recommendation model generation device 610 and the video recommendation device 620 are different devices, the video recommendation model generation device 610 and the video recommendation device 620 can be the same type of device, such as the video recommendation model generation device 610 and the video recommendation device 620 can both be servers; or the video recommendation model generation device 610 and the video recommendation device 620 can also be different types of devices, such as the video recommendation device 620 can be a personal computer or terminal, and the video recommendation model generation device 610 can be a server, etc. The embodiment of the present application does not limit the specific types of the video recommendation model generation device 610 and the video recommendation device 620.

[0265] The above embodiment illustrates a video recommendation method. Next, a training method for a video recommendation model will be further described.

[0266] Figure 7 This is a flowchart of a method for training a video recommendation model provided by an exemplary embodiment of the present application. The method can be executed by a computer device, which can be Figure 2 The terminal 100 or the server 200 in the embodiment of the present invention. The method includes:

[0267] Step 702: Obtain a multidimensional video feature vector of the shared video, a multidimensional object feature vector of the first object, multidimensional video feature vectors of positive candidate video samples in the candidate video set, and multidimensional video feature vectors of negative candidate video samples.

[0268] A shared video refers to a video that is clicked by a first subject after being shared by a second subject.

[0269] Positive candidate video samples refer to videos that are associated with the shared video, and negative candidate video samples refer to videos that are not associated with the shared video.

[0270] The shared video can be any video in the local video library of the second object's terminal, or any video shared by the second object from the cloud to the first object, and can include at least one continuous video frame. The first object is the object corresponding to the master computer device, and the second object can be an object corresponding to any computer device.

[0271] Optionally, the shared video is at least one of a video recorded by the second object itself, a video obtained by the second object from a stored database, and a video searched by the second object from the cloud, but is not limited to this. The embodiments of the present application are not limited to this.

[0272] Exemplarily, the multi-dimensional object feature vector refers to a vector feature corresponding to the first object, and the multi-dimensional object feature vector includes an object data vector and an object attribute vector.

[0273] Exemplarily, the multi-dimensional video feature vector includes an identification feature vector and a multimodal feature vector.

[0274] The identification feature vector refers to the feature vector obtained by extracting the identification information of the video.

[0275] Optionally, the identification information package includes at least one of the ID sequence information of the video, the author information of the video, and the tag attribute information of the video, but is not limited to this. The embodiments of the present application do not limit this.

[0276] The multimodal feature vector refers to the feature vector obtained by extracting the multimodal information of a video.

[0277] Optionally, the multimodal feature vector includes at least two of video title segmentation information, video title information, video cover information and video content information, but is not limited thereto and is not limited in this embodiment of the present application.

[0278] Modality: It is used to represent the source or form of information. Each source or form of information can be called a modality. For example, humans have touch and hearing; information media include voice, video, text, etc.; various sensors include radar, infrared, accelerometers, etc., each of which can be called a modality.

[0279] In an embodiment of the present application, the modality is used to indicate a data set from a video, such as: obtaining shared video title segmentation information, shared video title information, shared video cover information, and shared video content information.

[0280] Step 704: Fuse the multi-dimensional video feature vector and the multi-dimensional object feature vector of the shared video to obtain an object fusion vector of the first object.

[0281] The object fusion vector refers to a feature obtained by fusing the identification feature vector of the shared video, the multimodal feature vector of the shared video, and the multidimensional object feature vector of the first object.

[0282] Optionally, the multimodal feature vector of the shared video includes at least two of the shared video title segmentation feature vector, the shared video title feature vector, the shared video cover feature vector, and the candidate video content feature vector, but is not limited to this. The embodiments of the present application do not limit this.

[0283] For example, the multimodal feature vector of the shared video includes the shared video title word segmentation feature vector, the shared video title feature vector, and the shared video cover feature vector; the multidimensional object feature vector includes the object data vector and the object attribute vector, and the identification feature vector of the shared video, the shared video title word segmentation feature vector, the shared video title feature vector, the shared video cover feature vector, the object data vector and the object attribute vector are fused to obtain the object fusion vector, but not limited to this, and the embodiments of the present application are not limited to this.

[0284] Step 706: Fusing the multi-dimensional video feature vectors of the same positive candidate video sample to obtain a positive candidate video sample fusion vector, and fusing the multi-dimensional video feature vectors of the same negative candidate video sample to obtain a negative candidate video sample fusion vector.

[0285] The positive candidate video sample fusion vector refers to a feature obtained by fusing the identification feature vector of the positive candidate video sample and the multimodal feature vector of the positive candidate video sample.

[0286] The negative candidate video sample fusion vector refers to a feature obtained by fusing the identification feature vector of the negative candidate video sample and the multimodal feature vector of the negative candidate video sample.

[0287] Exemplarily, the computer device fuses the multi-dimensional video feature vectors of the same positive candidate video sample to obtain a positive candidate video sample fusion vector of the positive candidate video sample. The computer device fuses the multi-dimensional video feature vectors of the same negative candidate video sample to obtain a negative candidate video sample fusion vector of the negative candidate video sample.

[0288] Optionally, the multimodal feature vector of the positive candidate video sample includes at least two of the positive candidate video sample title segmentation feature vector, the positive candidate video sample title feature vector, the positive candidate video sample cover feature vector, and the positive candidate video sample content feature vector, but is not limited to this. The embodiments of the present application do not limit this.

[0289] Optionally, the multimodal feature vector of the negative candidate video sample includes at least two of the negative candidate video sample title segmentation feature vector, the negative candidate video sample title feature vector, the negative candidate video sample cover feature vector, and the negative candidate video sample content feature vector, but is not limited to this. The embodiments of the present application are not limited to this.

[0290] For example, the computer device fuses the identification feature vector of the positive candidate video sample, the title word segmentation feature vector of the positive candidate video sample, the title feature vector of the positive candidate video sample, the cover feature vector of the positive candidate video sample, and the content feature vector of the positive candidate video sample to obtain a positive candidate video sample fusion vector. The computer device fuses the identification feature vector of the negative candidate video sample, the title word segmentation feature vector of the negative candidate video sample, the title feature vector of the negative candidate video sample, the cover feature vector of the negative candidate video sample, and the content feature vector of the negative candidate video sample to obtain a negative candidate video sample fusion vector.

[0291] Step 708: performing similarity calculation based on the object fusion vector of the first object and the positive candidate video sample fusion vector to obtain a first similarity value; performing similarity calculation based on the object fusion vector of the first object and the negative candidate video sample fusion vector to obtain a second similarity value.

[0292] The similarity value is used to indicate the degree of association between the shared video and the candidate video. For example, the greater the similarity value, the higher the degree of association between the shared video and the candidate video.

[0293] Exemplarily, the computer device performs similarity calculation based on the object fusion vector of the first object and the positive candidate video sample fusion vector to obtain a first similarity value between the object fusion vector and the positive candidate video sample fusion vector.

[0294] The computer device performs similarity calculation based on the object fusion vector of the first object and the negative candidate video sample fusion vector to obtain a second similarity value between the object fusion vector and the negative candidate video sample fusion vector.

[0295] Step 710: Calculate a loss function value based on the first similarity value and the second similarity value.

[0296] The computer device calculates a loss function value of the video recommendation model based on the first similarity value and the second similarity value.

[0297] Step 712: Update the model parameters of the video recommendation model based on the loss function value.

[0298] Exemplarily, the computer device updates the model parameters of the video recommendation model based on the loss function value, thereby obtaining a trained video recommendation model.

[0299] Model parameter updating refers to updating the network parameters in the video recommendation model, or updating the network parameters of each network module in the model, or updating the network parameters of each network layer in the model, but is not limited to this, and the embodiments of the present application do not limit this.

[0300] The model parameters of the video recommendation model include at least one of the network parameters of the image feature learning network, the network parameters of the attention mechanism network, the network parameters of the feature fusion network, and the network parameters of the multi-layer perceptron network in the video recommendation model.

[0301] In summary, the method provided by this embodiment obtains the multidimensional video feature vector of the shared video, the multidimensional object feature vector of the first object, the multidimensional video feature vector of the positive candidate video sample in the candidate video set, and the video features of the negative candidate video sample; the computer device fuses the multidimensional video feature vector and the multidimensional object feature vector of the shared video to obtain the object fusion vector of the first object; and fuses the multidimensional video feature vector of the same positive candidate video sample to obtain the positive candidate video sample fusion vector of the positive candidate video sample; the computer device fuses the multidimensional video feature vector of the same negative candidate video sample to obtain the negative candidate video sample fusion vector of the negative candidate video sample; The computer device performs similarity calculation based on the object fusion vector of the first object and the positive candidate video sample fusion vector to obtain a first similarity value between the object fusion vector and the positive candidate video sample fusion vector; the computer device performs similarity calculation based on the object fusion vector of the first object and the negative candidate video sample fusion vector to obtain a second similarity value between the object fusion vector and the negative candidate video sample fusion vector; the computer device calculates a loss function value based on the first similarity value and the second similarity value; the computer device updates the model parameters of the video recommendation model based on the loss function value, so that the trained video recommendation model can have higher video recommendation accuracy, thereby obtaining more accurate recommended videos.

[0302] Figure 8 This is a flowchart of a method for training a video recommendation model provided by an exemplary embodiment of the present application. The method can be executed by a computer device, which can be Figure 2 The terminal 100 or the server 200 in the embodiment of the present invention. The method includes:

[0303] Step 802: Obtain a multidimensional video feature vector of the shared video, a multidimensional object feature vector of the first object, a multidimensional video feature vector of a positive candidate video sample in the candidate video set, and a multidimensional video feature vector of a negative candidate video sample.

[0304] A shared video refers to a video that is clicked by a first subject after being shared by a second subject.

[0305] Exemplarily, the multi-dimensional object feature vector refers to a vector feature corresponding to the first object, and the multi-dimensional object feature vector includes an object data vector and an object attribute vector.

[0306] The computer device obtains an object data set and object attribute data of the first object; and inputs the object data set and the object attribute data into a feature extraction network for feature extraction, thereby obtaining an object data vector and an object attribute vector of the first object.

[0307] In one possible implementation, the computer device obtains an object data set and object attribute data of the second object; and inputs the object data set and object attribute data into a feature extraction network for feature extraction, thereby obtaining an object data vector and an object attribute vector of the second object.

[0308] An object data set refers to a set of object behavior sequences. For example, an object data set includes at least one of likes, resume, play duration, collection, and forwarding, but is not limited to this. The embodiments of the present application do not make specific limitations on this.

[0309] The object attribute data includes at least one of the object's age, the object's gender, the object's hobbies, and the object's location, but is not limited thereto. The embodiments of the present application do not make specific limitations on this.

[0310] Exemplarily, the multi-dimensional video feature vector of the shared video includes an identification feature vector of the shared video and a multimodal feature vector of the shared video.

[0311] The computer device obtains identification information and multimodal information of the shared video, and inputs the identification information of the shared video and the multimodal information of the shared video into the corresponding feature extraction network for feature extraction, thereby obtaining an identification feature vector and a multimodal feature vector of the shared video.

[0312] Among them, the identification information of the shared video includes at least one of the ID sequence information of the shared video, the author information of the shared video, and the tag attribute information of the shared video; the multimodal information of the shared video includes at least two of the shared video title segmentation information, the shared video title information, the shared video cover information and the shared video content information, but is not limited to this. The embodiments of the present application do not make specific limitations on this.

[0313] For example, the computer device inputs the identification information of the shared video into the corresponding feature extraction network for feature extraction to obtain the shared video identification feature vector; inputs the shared video title word segmentation information into the corresponding feature extraction network for word granularity level feature extraction to obtain the shared video title word segmentation feature vector; inputs the shared video title information into the corresponding feature extraction network, and divides the shared video title information by word and performs feature extraction to obtain the shared video title feature vector; inputs the shared video cover information into the corresponding residual neural network (ResNet), and obtains the shared video cover feature vector by performing feature extraction on the image information.

[0314] Exemplarily, the multi-dimensional video feature vector includes an identification feature vector of the video and a multimodal feature vector of the video.

[0315] The computer device obtains the identification class information and multimodal information of the positive candidate video sample, and inputs the identification class information of the positive candidate video sample and the multimodal information of the positive candidate video sample into the corresponding feature extraction network for feature extraction, thereby obtaining the identification class feature vector and the multimodal feature vector of the positive candidate video sample.

[0316] Among them, the identification information includes at least one of the ID sequence information of the positive candidate video sample, the author information of the positive candidate video sample, and the label attribute information of the positive candidate video sample; the multimodal information of the positive candidate video sample includes at least two of the positive candidate video sample title segmentation information, the positive candidate video sample title information, the positive candidate video sample cover information and the positive candidate video sample content information, but is not limited to this, and the embodiments of the present application do not make specific limitations on this.

[0317] The computer device obtains the identification class information and multimodal information of the negative candidate video sample, and inputs the identification class information of the negative candidate video sample and the multimodal information of the negative candidate video sample into the corresponding feature extraction network for feature extraction, thereby obtaining the identification class feature vector and the multimodal feature vector of the negative candidate video sample.

[0318] Among them, the identification information includes at least one of the ID sequence information of the negative candidate video sample, the author information of the negative candidate video sample, and the label attribute information of the negative candidate video sample; the multimodal information of the negative candidate video sample includes at least two of the negative candidate video sample title segmentation information, the negative candidate video sample title information, the negative candidate video sample cover information, and the negative candidate video sample content information, but is not limited to this. The embodiments of the present application do not make specific limitations on this.

[0319] For example, the computer device inputs the identification information of the negative candidate video sample into the corresponding feature extraction network for feature extraction to obtain the identification feature vector of the negative candidate video sample; inputs the title segmentation information of the negative candidate video sample into the corresponding feature extraction network for feature extraction at the word granularity level to obtain the title segmentation feature vector of the negative candidate video sample; inputs the title information of the negative candidate video sample into the corresponding feature extraction network, and divides the title information of the negative candidate video sample by word and performs feature extraction to obtain the title feature vector of the negative candidate video sample; inputs the cover information of the negative candidate video sample into the corresponding residual neural network (ResNet), and obtains the cover feature vector of the negative candidate video sample by performing feature extraction on the image information.

[0320] Step 804: Fusing the identification feature vector, the multimodal feature vector, and the multidimensional object feature vector of the shared video to obtain an object fusion vector of the first object.

[0321] The object fusion vector refers to a feature obtained by fusing the identification feature vector of the shared video, the multimodal feature vector of the shared video, and the multidimensional object feature vector of the first object.

[0322] In one possible implementation, a computer device inputs the identification feature vector of the shared video, the multimodal feature vector of the shared video, and the multidimensional object feature vector into the attention mechanism network for weight calculation to obtain a first weight value corresponding to each feature vector.

[0323] The computer device performs feature fusion on the identification feature vector of the shared video, the multimodal feature vector of the shared video, and the multidimensional object feature vector according to the first weight value to obtain an object fusion vector of the first object.

[0324] Optionally, the computer device inputs the identification feature vector of the shared video, the multimodal feature vector and the multidimensional object feature vector of the shared video into the feature fusion network for feature fusion according to the first weight value to obtain an intermediate object fusion vector; the computer device inputs the intermediate object fusion vector into the multi-layer perceptron network MLP for nonlinear fitting to obtain the object fusion vector of the first object.

[0325] In one possible implementation, the computer device obtains a multidimensional object feature vector of the second object; the computer device fuses the multidimensional video feature vector of the shared video, the multidimensional object feature vector of the first object, and the multidimensional object feature vector of the second object to obtain an object fusion vector of the first object.

[0326] The second object is an object that shares a video with the first object.

[0327] Optionally, the first object is an object corresponding to the master computer device, and the second object may be an object corresponding to any computer device.

[0328] In one possible implementation, a computer device inputs the identification feature vector of the shared video, the multimodal feature vector of the shared video, the multidimensional object feature vector of the first object, and the multidimensional object feature vector of the second object into the attention mechanism network for weight calculation to obtain a first weight value corresponding to each feature vector.

[0329] The computer device performs feature fusion on the identification feature vector of the shared video, the multimodal feature vector of the shared video, the multidimensional object feature vector of the first object, and the multidimensional object feature vector of the second object according to the first weight value to obtain an object fusion vector of the first object.

[0330] Optionally, the computer device performs feature fusion on the identification feature vector of the shared video, the multimodal feature vector of the shared video, the multidimensional object feature vector of the first object, and the multidimensional object feature vector of the second object according to the first weight value to obtain an intermediate object fusion vector; the computer device inputs the intermediate object fusion vector into the multi-layer perceptron network MLP for nonlinear fitting to obtain the object fusion vector of the first object.

[0331] In a possible implementation manner, the computer device obtains a multi-dimensional object feature vector of the second object and obtains a first directivity feature vector between the second object and the first object.

[0332] The first directional feature vector is used to represent that the shared video is shared from the second object to the first object.

[0333] For example, the first directional feature vector can be expressed as: the second object in the first age group shares a video with the first object in the second age group; or, the second object in the first area shares a video with the first object in the second area; but it is not limited to this, and the embodiments of the present application are not limited to this.

[0334] The computer device fuses the multidimensional video feature vector of the shared video, the multidimensional object feature vector of the first object, the multidimensional object feature vector of the second object, and the first directional feature vector to obtain an object fusion vector of the first object.

[0335] For example, the formula for fusing the identification feature vector, multimodal feature vector, and multidimensional object feature vector of the shared video can be expressed as:

[0336]

[0337] Where, output-feature-user represents the object fusion vector, wi represents any feature vector among the identity feature vector, multimodal feature vector and multidimensional object feature vector, F i It represents the first weight value corresponding to each feature vector in the identification feature vector, multimodal feature vector and multidimensional object feature vector, and N represents the total number of identification feature vectors, multimodal feature vectors and multidimensional object feature vectors.

[0338] Step 806: Fusing the identification feature vector and multimodal feature vector of the same positive candidate video sample to obtain a positive candidate video sample fusion vector; fusing the identification feature vector and multimodal feature vector of the same negative candidate video sample to obtain a negative candidate video sample fusion vector.

[0339] The positive candidate video sample fusion vector refers to a feature obtained by fusing the identification feature vector of the positive candidate video sample and the multimodal feature vector of the positive candidate video sample.

[0340] The negative candidate video sample fusion vector refers to a feature obtained by fusing the identification feature vector of the negative candidate video sample and the multimodal feature vector of the negative candidate video sample.

[0341] Optionally, the multimodal feature vector of the positive candidate video sample includes at least two of the positive candidate video sample title segmentation feature vector, the positive candidate video sample title feature vector, the positive candidate video sample cover feature vector, and the positive candidate video sample content feature vector, but is not limited to this. The embodiments of the present application do not limit this.

[0342] Optionally, the multimodal feature vector of the negative candidate video sample includes at least two of the negative candidate video sample title segmentation feature vector, the negative candidate video sample title feature vector, the negative candidate video sample cover feature vector, and the negative candidate video sample content feature vector, but is not limited to this. The embodiments of the present application are not limited to this.

[0343] Exemplarily, the computer device inputs the identification feature vector and the multimodal feature vector corresponding to the positive candidate video sample into the attention mechanism network for weight calculation to obtain the second weight value corresponding to each feature vector in the positive candidate video sample; the computer device inputs the identification feature vector and the multimodal feature vector corresponding to the positive candidate video sample into the feature fusion network for feature fusion according to the second weight value to obtain the positive candidate video sample fusion vector.

[0344] Optionally, the computer device inputs the identification feature vector and the multimodal feature vector corresponding to the positive candidate video sample into the feature fusion network for feature fusion according to the second weight value to obtain an intermediate positive candidate video sample fusion vector; the computer device inputs the intermediate positive candidate video sample fusion vector into the multi-layer perceptron network MLP for nonlinear fitting to obtain the positive candidate video sample fusion vector.

[0345] In one possible implementation, a candidate video set includes a positive candidate video sample and at least one negative candidate video sample. A computer device inputs an identification feature vector and a multimodal feature vector corresponding to an i-th negative candidate video sample in the candidate video set into an attention mechanism network for weight calculation, thereby obtaining a second weight value corresponding to each feature vector in the i-th negative candidate video sample, where i is a positive integer. Based on the second weight value, the identification feature vector and the multimodal feature vector corresponding to the i-th negative candidate video sample are input into a feature fusion network for feature fusion, thereby obtaining a fusion vector for the i-th negative candidate video sample.

[0346] Optionally, the computer device inputs the identification feature vector and the multimodal feature vector corresponding to the i-th negative candidate video sample into the feature fusion network for feature fusion according to the second weight value to obtain the intermediate i-th negative candidate video sample fusion vector; the computer device inputs the intermediate i-th negative candidate video sample fusion vector into the multi-layer perceptron network MLP for nonlinear fitting to obtain the i-th negative candidate video sample fusion vector.

[0347] For example, if the candidate video set includes one positive candidate video sample and nine negative candidate video samples, the computer device inputs the identification feature vector and multimodal feature vector corresponding to the positive candidate video sample into the attention mechanism network for weight calculation, obtaining a second weight value corresponding to each feature vector in the positive candidate video sample. Based on the second weight value, the computer device inputs the identification feature vector and multimodal feature vector corresponding to the positive candidate video sample into the feature fusion network for feature fusion, obtaining a fusion vector of the positive candidate video sample. Similarly, the computer device performs feature fusion on the identification feature vector and multimodal feature vector corresponding to the negative candidate video sample, obtaining a fusion vector of the nine negative candidate video samples.

[0348] For example, the formula for fusing the identification feature vector and the multimodal feature vector corresponding to the candidate video sample can refer to the formula in step 406 and will not be repeated here.

[0349] Step 808: performing similarity calculation based on the object fusion vector of the first object and the positive candidate video sample fusion vector to obtain a first similarity value; performing similarity calculation based on the object fusion vector of the first object and the negative candidate video sample fusion vector to obtain a second similarity value.

[0350] The similarity value is used to indicate the degree of association between the shared video and the candidate video. For example, the greater the similarity value, the higher the degree of association between the shared video and the candidate video.

[0351] Exemplarily, the computer device calculates a cosine similarity value between the object fusion vector and the positive candidate video sample fusion vector to obtain a first similarity value between the object fusion vector and the positive candidate video sample fusion vector.

[0352] The computer device calculates the cosine similarity value between the object fusion vector and the i-th negative candidate video sample fusion vector in the candidate video set to obtain the i-th similarity value between the object fusion vector and the i-th negative candidate video sample fusion vector, where i is a positive integer.

[0353] For example, the calculation formula of the cosine similarity value can be expressed as:

[0354]

[0355] Where y Q Represents the object fusion vector, y D represents the candidate video fusion vector, and R represents the cosine similarity value.

[0356] For example, the candidate video set includes one positive candidate video sample and nine negative candidate video samples. The computer device calculates the cosine similarity value of the object fusion vector of the first object and the fusion vector of the positive candidate video sample to obtain a first similarity value between the object fusion vector and the fusion vector of the positive candidate video sample; the computer device calculates the cosine similarity value of the object fusion vector of the first object and the fusion vectors of the nine negative candidate video samples in turn to obtain nine second similarity values ​​between the object fusion vector and the fusion vectors of the negative candidate video samples.

[0357] Step 810: Calculate a loss function value based on the first similarity value and the second similarity value.

[0358] Exemplarily, the computer device adds the second similarity values ​​to obtain a total similarity value; and the computer device obtains a loss function value based on a ratio of the first similarity value to the total similarity value.

[0359] The calculation formula of the loss function value can be expressed as:

[0360]

[0361] Where P represents the loss function value, γ represents the smoothing factor of the normalized exponential function softmax, Q represents the object fusion vector of the first object, and D + Denotes the fusion vector of the positive candidate video sample, D - Represents the negative candidate video sample fusion vector, R(Q, D+ ) represents the first similarity value, R(Q, D - ) represents the second similarity value, Ds={D + , D - 1, D - 2, D - 2,…,D - N}, where N represents the number of negative candidate video samples that match the positive candidate video samples.

[0362] Step 812: Update the model parameters of the video recommendation model based on the loss function value.

[0363] Model parameter updating refers to updating the network parameters in the video recommendation model, or updating the network parameters of each network module in the model, or updating the network parameters of each network layer in the model, but is not limited to this, and the embodiments of the present application do not limit this.

[0364] The model parameters of the video recommendation model include at least one of a feature extraction network, an attention mechanism network, a feature fusion network, and a multi-layer perceptron network in the video recommendation model.

[0365] In some embodiments, updating the model parameters of the video recommendation model includes updating the network parameters of all network modules in the video recommendation model, or fixing the network parameters of some network modules in the video recommendation model and updating only the network parameters of the remaining network modules. For example, when updating the model parameters of the video recommendation model, the network parameters of the attention mechanism network, the network parameters of the feature fusion network, and the network parameters of the multi-layer perceptron network in the video recommendation model are fixed, and only the network parameters of the feature extraction network are updated.

[0366] Based on the loss function value, the loss function value is used as a training indicator to update the network parameters of the feature extraction network, attention mechanism network, feature fusion network, and multi-layer perceptron network in the video recommendation model until the loss function value converges, thereby obtaining a trained video recommendation model.

[0367] The convergence of the loss function value means that the loss function value no longer changes, or the error difference between two adjacent iterations during the training of the video recommendation model is less than a preset value, or the number of training times of the video recommendation model reaches at least one of the preset times, but is not limited to this. The embodiments of the present application are not limited to this.

[0368] Optionally, the target condition satisfied by training may be that the number of training iterations of the initial model reaches a target number, and the technician may pre-set the number of training iterations. Alternatively, the target condition satisfied by training may be that the loss value meets a target threshold condition, but is not limited thereto and is not limited to this embodiment of the present application.

[0369] In summary, the method provided by this embodiment obtains the multidimensional video feature vector of the shared video, the multidimensional object feature vector of the first object, the multidimensional video feature vector of the positive candidate video sample in the candidate video set, and the video features of the negative candidate video sample; the computer device fuses the multidimensional video feature vector and the multidimensional object feature vector of the shared video to obtain the object fusion vector of the first object; and fuses the multidimensional video feature vector of the same positive candidate video sample to obtain the positive candidate video sample fusion vector of the positive candidate video sample; the computer device fuses the multidimensional video feature vector of the same negative candidate video sample to obtain the negative candidate video sample fusion vector of the negative candidate video sample. The computer device performs similarity calculation based on the object fusion vector of the first object and the positive candidate video sample fusion vector to obtain a first similarity value between the object fusion vector and the positive candidate video sample fusion vector; the computer device performs similarity calculation based on the object fusion vector of the first object and the negative candidate video sample fusion vector to obtain a second similarity value between the object fusion vector and the negative candidate video sample fusion vector; the computer device calculates a loss function value based on the first similarity value and the second similarity value; the computer device updates the model parameters of the video recommendation model based on the loss function value, so that the trained video recommendation model can have higher video recommendation accuracy, thereby obtaining more accurate recommended videos.

[0370] It should be noted that the information (including but not limited to object device information, object personal information, etc.), data (including but not limited to data used for analysis, stored data, displayed data, etc.) and signals involved in this application are all authorized by the object or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions. For example, the object data set, object attribute data, identification information of the shared video, multimodal information of the shared video, identification information of the candidate video and multimodal information of the candidate video involved in this application are all obtained with full authorization.

[0371] Figure 9 This is a flowchart of a video recommendation model training method provided by an exemplary embodiment of the present application. The method can be executed by a computer device, which can be Figure 2 The terminal 100 or the server 200 in the embodiment of the present invention. The method includes:

[0372] Step 901: Obtain a sample.

[0373] The computer device obtains an object data set of the first object, object attribute data, identification information of the shared video, multimodal information of the shared video, identification information of the positive candidate video sample, multimodal information of the positive candidate video sample, identification information of the negative candidate video sample, and multimodal information of the negative candidate video sample.

[0374] Step 902: Feature extraction.

[0375] The computer device inputs the object data set, object attribute data, identification information of the shared video, multimodal information of the shared video, identification information of the positive candidate video sample, multimodal information of the positive candidate video sample, identification information of the negative candidate video sample and multimodal information of the negative candidate video sample into the corresponding feature extraction network for feature extraction, and obtains the multidimensional video feature vector of the shared video, the multidimensional object feature vector of the first object, the multidimensional video feature vector of the positive candidate video sample and the multidimensional video feature vector of the negative candidate video sample.

[0376] Step 903: Model training.

[0377] The computer device fuses the multidimensional video feature vectors and the multidimensional object feature vectors of the shared video to obtain the object fusion vector of the first object; the computer device fuses the multidimensional video feature vectors of the same positive candidate video sample to obtain the positive candidate video sample fusion vector of the positive candidate video sample; the computer device fuses the multidimensional video feature vectors of the same negative candidate video sample to obtain the negative candidate video sample fusion vector of the negative candidate video sample.

[0378] The computer device performs similarity calculation based on the object fusion vector of the first object and the positive candidate video sample fusion vector to obtain a first similarity value between the object fusion vector and the positive candidate video sample fusion vector; the computer device performs similarity calculation based on the object fusion vector of the first object and the negative candidate video sample fusion vector to obtain a second similarity value between the object fusion vector and the negative candidate video sample fusion vector; the computer device calculates a loss function value based on the first similarity value and the second similarity value; and updates the model parameters of the video recommendation model based on the loss function value.

[0379] Step 904: Obtain a trained model.

[0380] Figure 10 This is a flowchart of a video recommendation method provided by an exemplary embodiment of the present application. The method can be executed by a computer device, which can be Figure 2 The terminal 100 or the server 200 in the embodiment of the present invention. The method includes:

[0381] Step 1001: Get the shared video.

[0382] The computer device obtains an object data set of the first object, object attribute data, identification information of the shared video, multimodal information of the shared video, identification information of the candidate video, and multimodal information of the candidate video.

[0383] Step 1002: Feature extraction.

[0384] The computer device inputs the object data set, object attribute data, identification information of the shared video, multimodal information of the shared video, identification information of the candidate video sample and multimodal information of the candidate video into the corresponding feature extraction network for feature extraction, and obtains the multidimensional video feature vector of the shared video, the multidimensional object feature vector of the first object and the multidimensional video feature vector of the candidate video.

[0385] Step 1003: Model calculation.

[0386] The computer device fuses the multidimensional video feature vector and the multidimensional object feature vector of the shared video to obtain the object fusion vector of the first object; the computer device fuses the multidimensional video feature vectors of the same candidate video to obtain the candidate video fusion vector of the candidate video; the computer device performs similarity calculation based on the object fusion vector of the first object and the candidate video fusion vector to obtain the similarity value between the object fusion vector and the candidate video fusion vector.

[0387] Step 1004: Recommend related videos.

[0388] The computer device recommends videos corresponding to higher-ranked similarity values ​​to the first object based on the ranking of the similarity values.

[0389] Figure 11 The following is a block diagram of a video recommendation device provided by an exemplary embodiment of the present application. The device can be implemented as all or part of a computer device through software, hardware, or a combination of both. The device includes:

[0390] An acquisition module 1101 is configured to acquire a multidimensional video feature vector of a shared video, a multidimensional object feature vector of a first subject, and multidimensional video feature vectors of candidate videos in a candidate video set, wherein the shared video is a video clicked by the first subject after receiving a video shared by the second subject, and the multidimensional video feature vector includes an identification feature vector and a multimodal feature vector.

[0391] A fusion module 1102 is configured to fuse the multi-dimensional video feature vector of the shared video and the multi-dimensional object feature vector to obtain an object fusion vector of the first object;

[0392] The fusion module 1102 is further configured to fuse the multi-dimensional video feature vectors of the same candidate video to obtain a candidate video fusion vector of the candidate video;

[0393] A calculation module 1103 is configured to perform similarity calculation based on the object fusion vector of the first object and the candidate video fusion vector to obtain a similarity value between the object fusion vector and the candidate video fusion vector;

[0394] The recommendation module 1104 is configured to recommend the candidate video to the first object based on the similarity value.

[0395] In one possible implementation, the calculation module 1103 is also used to input the identification feature vector of the shared video, the multimodal feature vector of the shared video, and the multidimensional object feature vector into the attention mechanism network for weight calculation to obtain the first weight value corresponding to each feature vector.

[0396] In a possible implementation, the fusion module 1102 is further configured to perform feature fusion on the identification feature vector of the shared video, the multimodal feature vector of the shared video, and the multidimensional object feature vector according to the first weight value to obtain the object fusion vector of the first object.

[0397] In one possible implementation, the fusion module 1102 is further used to perform feature fusion on the identification feature vector of the shared video, the multimodal feature vector of the shared video, and the multidimensional object feature vector according to the first weight value to obtain an intermediate object fusion vector; and input the intermediate object fusion vector into the multi-layer perceptron network MLP for nonlinear fitting to obtain the object fusion vector of the first object.

[0398] In one possible implementation, the calculation module 1103 is further configured to input the identification feature vector and the multimodal feature vector corresponding to the i-th candidate video in the candidate video set into the attention mechanism network for weight calculation, to obtain a second weight value corresponding to each feature vector in the i-th candidate video, where i is a positive integer;

[0399] In a possible implementation, the fusion module 1102 is further configured to perform feature fusion on the identification feature vector and the multimodal feature vector corresponding to the i-th candidate video according to the second weight value to obtain the i-th candidate video fusion vector.

[0400] In one possible implementation, the fusion module 1102 is further used to perform feature fusion on the identification feature vector and the multimodal feature vector corresponding to the i-th candidate video according to the second weight value to obtain an intermediate i-th candidate video fusion vector; and input the intermediate i-th candidate video fusion vector into the multi-layer perceptron network MLP for nonlinear fitting to obtain the i-th candidate video fusion vector.

[0401] In one possible implementation, the calculation module 1103 is further used to calculate the cosine similarity value between the object fusion vector and the i-th candidate video fusion vector in the candidate video set to obtain the i-th similarity value between the object fusion vector and the i-th candidate video fusion vector, where i is a positive integer.

[0402] In one possible implementation, the acquisition module 1101 is also used to obtain the object data set and object attribute data of the first object; the object data set and the object attribute data are respectively input into the feature extraction network for feature extraction to obtain the object data vector and the object attribute vector of the first object.

[0403] In a possible implementation, the acquisition module 1101 is further configured to acquire a multi-dimensional object feature vector of the second object.

[0404] In a possible implementation, the fusion module 1102 is further configured to fuse the multidimensional video feature vector of the shared video, the multidimensional object feature vector of the first object, and the multidimensional object feature vector of the second object to obtain an object fusion vector of the first object.

[0405] In one possible implementation, the multidimensional object feature vector of the second object includes an object data vector and an object attribute vector; the acquisition module 1101 is also used to obtain the object data set and object attribute data of the second object; the object data set and the object attribute data are respectively input into the feature extraction network for feature extraction to obtain the object data vector and the object attribute vector of the second object.

[0406] In one possible implementation, the acquisition module 1101 is further configured to acquire a multidimensional object feature vector of the second object, and acquire a first directional feature vector between the second object and the first object, wherein the first directional feature vector is used to represent that the shared video is shared from the second object to the first object.

[0407] In a possible implementation, the fusion module 1102 is further used to fuse the multidimensional video feature vector of the shared video, the multidimensional object feature vector of the first object, the multidimensional object feature vector of the second object and the first directional feature vector to obtain the object fusion vector of the first object.

[0408] In a possible implementation, the acquisition module 1101 is further configured to acquire a second directional feature vector of the candidate video in the candidate video set, where the second directional feature vector is used to indicate that the candidate video is shared from the third object to the fourth object in the historical sharing record.

[0409] In one possible implementation, the fusion module 1102 is also used to fuse the multidimensional video feature vector of the same candidate video, the multidimensional object feature vector of the third object, the multidimensional object feature vector of the fourth object and the second directional feature vector to obtain a candidate video fusion vector of the candidate video.

[0410] In a possible implementation, the acquisition module 1101 is also used to obtain identification information and multimodal information of the shared video, wherein the identification information includes at least one of the ID sequence information of the shared video, the author information of the shared video, and the tag attribute information of the shared video; the multimodal information of the shared video includes at least two of the shared video title segmentation information, the shared video title information, the shared video cover information, and the shared video content information; the identification information of the shared video and the multimodal information of the shared video are respectively input into the corresponding feature extraction network for feature extraction to obtain the identification feature vector and the multimodal feature vector of the shared video.

[0411] In one possible implementation, the acquisition module 1101 is also used to obtain identification information and multimodal information of the candidate video, where the identification information includes at least one of the ID sequence information of the candidate video, the author information of the candidate video, and the tag attribute information of the candidate video; the multimodal information of the candidate video includes at least two of the candidate video title segmentation information, the candidate video title information, the candidate video cover information, and the candidate video content information; the identification information of the candidate video and the multimodal information of the candidate video are respectively input into the corresponding feature extraction network for feature extraction to obtain the identification feature vector and the multimodal feature vector of the candidate video.

[0412] Figure 12A block diagram of a video recommendation model training device provided by an exemplary embodiment of the present application is shown. The device can be implemented as all or part of a computer device through software, hardware, or a combination of both. The device includes:

[0413] Acquisition module 1201 is used to obtain the multidimensional video feature vector of the shared video, the multidimensional object feature vector of the first object, the multidimensional video feature vector of the positive candidate video sample in the candidate video set, and the multidimensional video feature vector of the negative candidate video sample. The shared video refers to the video clicked by the first object after receiving the video shared by the second object. The multidimensional video feature vector includes an identification feature vector and a multimodal feature vector. The positive candidate video sample refers to a video associated with the shared video, and the negative candidate video sample refers to a video not associated with the shared video.

[0414] The fusion module 1202 is configured to fuse the multi-dimensional video feature vector of the shared video and the multi-dimensional object feature vector to obtain an object fusion vector of the first object.

[0415] The fusion module 1202 is configured to fuse the multi-dimensional video feature vectors of the same positive candidate video sample to obtain a positive candidate video sample fusion vector of the positive candidate video sample.

[0416] The fusion module 1202 is configured to fuse the multi-dimensional video feature vectors of the same negative candidate video sample to obtain a negative candidate video sample fusion vector of the negative candidate video sample.

[0417] The calculation module 1203 is configured to perform similarity calculation based on the object fusion vector of the first object and the positive candidate video sample fusion vector to obtain a first similarity value between the object fusion vector and the positive candidate video sample fusion vector.

[0418] The calculation module 1203 is configured to perform similarity calculation based on the object fusion vector of the first object and the negative candidate video sample fusion vector to obtain a second similarity value between the object fusion vector and the negative candidate video sample fusion vector.

[0419] The calculation module 1203 is configured to calculate a loss function value based on the first similarity value and the second similarity value.

[0420] The updating module 1204 is configured to update the model parameters of the video recommendation model based on the loss function value.

[0421] In one possible implementation, the calculation module 1203 is also used to input the identification feature vector of the shared video, the multimodal feature vector of the shared video, and the multidimensional object feature vector into the attention mechanism network for weight calculation to obtain the first weight value corresponding to each feature vector.

[0422] In one possible implementation, the fusion module 1202 is further used to input the identification feature vector of the shared video, the multimodal feature vector of the shared video, and the multidimensional object feature vector into the feature fusion network for feature fusion according to the first weight value to obtain the object fusion vector of the first object.

[0423] In one possible implementation, the fusion module 1202 is further used to input the identification feature vector of the shared video, the multimodal feature vector of the shared video, and the multidimensional object feature vector into the feature fusion network for feature fusion according to the first weight value to obtain an intermediate object fusion vector; and input the intermediate object fusion vector into the multi-layer perceptron network MLP for nonlinear fitting to obtain the object fusion vector of the first object.

[0424] In one possible implementation, the calculation module 1203 is further used to input the identification feature vector and the multimodal feature vector corresponding to the positive candidate video sample into the attention mechanism network for weight calculation to obtain a second weight value corresponding to each feature vector in the positive candidate video sample.

[0425] In a possible implementation, the fusion module 1202 is further configured to input the identification feature vector and the multimodal feature vector corresponding to the positive candidate video sample into the feature fusion network for feature fusion according to the second weight value to obtain the positive candidate video sample fusion vector.

[0426] In one possible implementation, the fusion module 1202 is further used to input the identification feature vector and the multimodal feature vector corresponding to the positive candidate video sample into the feature fusion network for feature fusion according to the second weight value to obtain an intermediate positive candidate video sample fusion vector; and input the intermediate positive candidate video sample fusion vector into the multi-layer perceptron network MLP for nonlinear fitting to obtain the positive candidate video sample fusion vector.

[0427] In one possible implementation, the calculation module 1203 is also used to input the identification feature vector and the multimodal feature vector corresponding to the i-th negative candidate video sample in the candidate video set into the attention mechanism network for weight calculation, to obtain the second weight value corresponding to each feature vector in the i-th negative candidate video sample, where i is a positive integer.

[0428] In one possible implementation, the fusion module 1202 is further configured to input the identification feature vector and the multimodal feature vector corresponding to the i-th negative candidate video sample into the feature fusion network for feature fusion according to the second weight value to obtain the i-th negative candidate video sample fusion vector.

[0429] In one possible implementation, the fusion module 1202 is further used to input the identification feature vector and the multimodal feature vector corresponding to the i-th negative candidate video sample into the feature fusion network for feature fusion according to the second weight value to obtain the intermediate i-th negative candidate video sample fusion vector; and input the intermediate i-th negative candidate video sample fusion vector into the multi-layer perceptron network MLP for nonlinear fitting to obtain the i-th negative candidate video sample fusion vector.

[0430] In a possible implementation, the calculation module 1203 is further configured to calculate a cosine similarity value between the object fusion vector and the positive candidate video sample fusion vector to obtain a first similarity value between the object fusion vector and the positive candidate video sample fusion vector.

[0431] In one possible implementation, the calculation module 1203 is further used to calculate the cosine similarity value between the object fusion vector and the i-th negative candidate video sample fusion vector in the candidate video set to obtain the i-th similarity value between the object fusion vector and the i-th negative candidate video sample fusion vector, where i is a positive integer.

[0432] In a possible implementation, the acquisition module 1201 is further configured to acquire a multi-dimensional object feature vector of the second object.

[0433] In a possible implementation, the fusion module 1202 is further configured to fuse the multidimensional video feature vector of the shared video, the multidimensional object feature vector of the first object, and the multidimensional object feature vector of the second object to obtain an object fusion vector of the first object.

[0434] In one possible implementation, the multidimensional object feature vector of the second object includes an object data vector and an object attribute vector; the acquisition module 1201 is also used to obtain the object data set and object attribute data of the second object; the object data set and the object attribute data are respectively input into the feature extraction network for feature extraction to obtain the object data vector and the object attribute vector of the second object.

[0435] In a possible implementation, the acquisition module 1201 is further configured to acquire the object data set and object attribute data of the first object;

[0436] The object data set and the object attribute data are respectively input into a feature extraction network for feature extraction to obtain the object data vector and the object attribute vector of the first object.

[0437] In a possible implementation, the acquisition module 1201 is also used to obtain identification information and multimodal information of the shared video, wherein the identification information includes at least one of the ID sequence information of the shared video, the author information of the shared video, and the tag attribute information of the shared video; the multimodal information of the shared video includes at least two of the shared video title segmentation information, the shared video title information, the shared video cover information, and the shared video content information; the identification information of the shared video and the multimodal information of the shared video are respectively input into the corresponding feature extraction network for feature extraction to obtain the identification feature vector and the multimodal feature vector of the shared video.

[0438] In a possible implementation, the acquisition module 1201 is further used to obtain identification information and multimodal information of the positive candidate video sample, where the identification information includes at least one of the ID sequence information of the positive candidate video sample, the author information of the positive candidate video sample, and the label attribute information of the positive candidate video sample; the multimodal information of the positive candidate video sample includes at least two of the positive candidate video sample title segmentation information, the positive candidate video sample title information, the positive candidate video sample cover information, and the positive candidate video sample content information; the identification information of the positive candidate video sample and the multimodal information of the positive candidate video sample are respectively input into the corresponding feature extraction network for feature extraction to obtain the identification feature vector and the multimodal feature vector of the positive candidate video sample.

[0439] In a possible implementation, the acquisition module 1201 is further used to obtain identification information and multimodal information of the negative candidate video sample, where the identification information includes at least one of the ID sequence information of the negative candidate video sample, the author information of the negative candidate video sample, and the label attribute information of the negative candidate video sample; the multimodal information of the negative candidate video sample includes at least two of the negative candidate video sample title segmentation information, the negative candidate video sample title information, the negative candidate video sample cover information, and the negative candidate video sample content information; the identification information of the negative candidate video sample and the multimodal information of the negative candidate video sample are respectively input into the corresponding feature extraction network for feature extraction to obtain the identification feature vector and the multimodal feature vector of the negative candidate video sample.

[0440] In a possible implementation, the calculation module 1203 is further configured to add the second similarity values ​​to obtain a total similarity value; and obtain the loss function value based on a ratio of the first similarity value to the total similarity value.

[0441] Figure 13 The following is a block diagram of a computer device 1300 according to an exemplary embodiment of the present application. The computer device can be implemented as the server in the aforementioned solution of the present application. The image computer device 1300 includes a central processing unit (CPU) 1301, a system memory 1304 including a random access memory (RAM) 1302 and a read-only memory (ROM) 1303, and a system bus 1305 connecting the system memory 1304 and the CPU 1301. The image computer device 1300 also includes a mass storage device 1306 for storing an operating system 1309, application programs 1310, and other program modules 1311.

[0442] The mass storage device 1306 is connected to the central processing unit 1301 via a mass storage controller (not shown) connected to the system bus 1305. The mass storage device 1306 and its associated computer-readable media provide non-volatile storage for the image computing device 1300. In other words, the mass storage device 1306 may include a computer-readable medium (not shown) such as a hard disk or a compact disc read-only memory (CD-ROM) drive.

[0443] Without loss of generality, the computer-readable medium may include computer storage media and communication media. Computer storage media include volatile and non-volatile, removable and non-removable media implemented by any method or technology for storing information such as computer-readable instructions, data structures, program modules or other data. Computer storage media include RAM, ROM, Erasable Programmable Read Only Memory (EPROM), Electronically Erasable Programmable Read-Only Memory (EEPROM), flash memory or other solid-state storage technology, CD-ROM, Digital Versatile Disc (DVD) or other optical storage, tape cassettes, magnetic tape, disk storage or other magnetic storage devices. Of course, those skilled in the art will appreciate that the computer storage media are not limited to the above-mentioned ones. The above-mentioned system memory 1304 and mass storage device 1306 can be collectively referred to as memory.

[0444] According to various embodiments of the present disclosure, the image computer device 1300 may also be connected to a remote computer on a network such as the Internet for operation. That is, the image computer device 1300 may be connected to a network 1308 via a network interface unit 1307 connected to the system bus 1305. Alternatively, the network interface unit 1307 may be used to connect to other types of networks or remote computer systems (not shown).

[0445] The memory also includes at least one computer program, which is stored in the memory. The central processing unit 1301 implements all or part of the steps in the video recommendation method or video recommendation model training method shown in the above embodiments by executing the at least one program.

[0446] An embodiment of the present application also provides a computer device, which includes a processor and a memory, wherein the memory stores at least one program, and the at least one program is loaded and executed by the processor to implement the video recommendation model training method or video recommendation method provided by the above-mentioned method embodiments.

[0447] An embodiment of the present application also provides a computer-readable storage medium, which stores at least one program, and the at least one program is loaded and executed by a processor to implement the video recommendation model training method or video recommendation method provided by the above-mentioned method embodiments.

[0448] An embodiment of the present application also provides a computer program product, which includes a computer program stored in a computer-readable storage medium; the computer program is read and executed from the computer-readable storage medium by a processor of a computer device, so that the computer device executes the video recommendation model training method or video recommendation method provided in the above-mentioned method embodiments.

[0449] It is understandable that in the specific implementation methods of this application, the data involved, historical data, and portraits and other data related to object data processing related to the identity or characteristics of the object, when the above embodiments of this application are applied to specific products or technologies, need to obtain the object's permission or consent, and the collection, use and processing of relevant data need to comply with the relevant laws, regulations and standards of relevant countries and regions.

[0450] It should be understood that the term "plurality" used herein refers to two or more. "And / or" describes a relationship between associated objects, indicating that three possible relationships exist. For example, "A and / or B" can mean: A exists alone, A and B exist simultaneously, or B exists alone. The character " / " generally indicates an "or" relationship between the associated objects.

[0451] Those skilled in the art will understand that all or part of the steps to implement the above embodiments may be accomplished by hardware, or by a program to instruct the relevant hardware, and the program may be stored in a computer-readable storage medium, which may be a read-only memory, a disk, or an optical disk, etc.

[0452] The above description is merely an optional embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent switches, improvements, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. A video recommendation method, characterized in that: The method comprises: Obtaining a multidimensional video feature vector of a shared video, a multidimensional object feature vector of a first object, and multidimensional video feature vectors of candidate videos in a candidate video set, wherein the shared video refers to a video clicked by the first object after receiving a video shared by the second object, and the multidimensional video feature vector includes an identification feature vector and a multimodal feature vector; Obtaining a multidimensional object feature vector of the second object, and obtaining a first directional feature vector between the second object and the first object, wherein the first directional feature vector is used to indicate that the shared video is shared from the second object to the first object; the first directional feature vector is used to indicate that the second object in a first age group shares a video with the first object in a second age group; or, the second object in a first region shares a video with the first object in a second region; fusing the multidimensional video feature vector of the shared video, the multidimensional object feature vector of the first object, the multidimensional object feature vector of the second object, and the first directional feature vector to obtain an object fusion vector of the first object; fusing the multi-dimensional video feature vectors of the same candidate video to obtain a candidate video fusion vector of the candidate video; Performing similarity calculation based on the object fusion vector of the first object and the candidate video fusion vector to obtain a similarity value between the object fusion vector and the candidate video fusion vector; recommending the candidate video to the first object according to the similarity value; The identification feature vector refers to a feature vector related to video attributes, and the multimodal feature vector refers to a feature vector related to the semantics of video content.

2. The method according to claim 1, characterized in that The fusing the multidimensional video feature vector of the shared video, the multidimensional object feature vector of the first object, the multidimensional object feature vector of the second object, and the first directional feature vector to obtain an object fusion vector of the first object includes: Inputting the identification feature vector of the shared video, the multimodal feature vector of the shared video, the multidimensional object feature vector of the first object, the multidimensional object feature vector of the second object, and the first directional feature vector into the attention mechanism network for weight calculation, thereby obtaining a first weight value corresponding to each feature vector; According to the first weight value, the identification feature vector of the shared video, the multimodal feature vector of the shared video, the multidimensional object feature vector of the first object, the multidimensional object feature vector of the second object and the first directional feature vector are feature fused to obtain the object fusion vector of the first object.

3. The method according to claim 2, characterized in that The step of fusing the identification feature vector of the shared video, the multimodal feature vector of the shared video, the multidimensional object feature vector of the first object, the multidimensional object feature vector of the second object, and the first directional feature vector according to the first weight value to obtain the object fusion vector of the first object includes: performing feature fusion on the identification feature vector of the shared video, the multimodal feature vector of the shared video, the multidimensional object feature vector of the first object, the multidimensional object feature vector of the second object, and the first directional feature vector according to the first weight value to obtain an intermediate object fusion vector; The intermediate object fusion vector is input into a multi-layer perceptron network (MLP) for nonlinear fitting to obtain the object fusion vector of the first object.

4. The method according to claim 1, wherein The candidate video set includes at least two candidate videos; The step of fusing the multi-dimensional video feature vectors of the same candidate video to obtain a candidate video fusion vector of the candidate video includes: Inputting the identification feature vector and the multimodal feature vector corresponding to the i-th candidate video in the candidate video set into the attention mechanism network for weight calculation, obtaining a second weight value corresponding to each feature vector in the i-th candidate video, where i is a positive integer; Feature fusion is performed on the identification feature vector and the multimodal feature vector corresponding to the i-th candidate video according to the second weight value to obtain an i-th candidate video fusion vector.

5. The method according to claim 4, characterized in that The step of performing feature fusion on the identification feature vector and the multimodal feature vector corresponding to the i-th candidate video according to the second weight value to obtain the i-th candidate video fusion vector includes: Performing feature fusion on the identification feature vector and the multimodal feature vector corresponding to the i-th candidate video according to the second weight value to obtain an intermediate i-th candidate video fusion vector; The intermediate i-th candidate video fusion vector is input into a multi-layer perceptron network MLP for nonlinear fitting to obtain the i-th candidate video fusion vector.

6. The method according to claim 1, characterized in that The candidate video set includes at least two candidate videos; The performing similarity calculation based on the object fusion vector of the first object and the candidate video fusion vector to obtain a similarity value between the object fusion vector and the candidate video fusion vector includes: A cosine similarity value is calculated for the object fusion vector and an i-th candidate video fusion vector in the candidate video set to obtain an i-th similarity value between the object fusion vector and the i-th candidate video fusion vector, where i is a positive integer.

7. The method according to any one of claims 1 to 6, characterized in that: The multi-dimensional object feature vector of the second object includes an object data vector and an object attribute vector; the method further includes: Acquire an object data set and object attribute data of the second object; The object data set and the object attribute data are respectively input into a feature extraction network for feature extraction to obtain the object data vector and the object attribute vector of the second object.

8. The method according to any one of claims 1 to 6, characterized in that: The method further comprises: Obtaining a second directional feature vector of the candidate video in the candidate video set, where the second directional feature vector is used to represent that the candidate video is shared from a third object to a fourth object in a historical sharing record; The step of fusing the multi-dimensional video feature vectors of the same candidate video to obtain a candidate video fusion vector of the candidate video includes: The multidimensional video feature vector of the same candidate video, the multidimensional object feature vector of the third object, the multidimensional object feature vector of the fourth object and the second directional feature vector are fused to obtain a candidate video fusion vector of the candidate video.

9. The method according to any one of claims 1 to 6, characterized in that: The multi-dimensional object feature vector of the first object includes an object data vector and an object attribute vector; the method further includes: Acquire an object data set and object attribute data of the first object; The object data set and the object attribute data are respectively input into a feature extraction network for feature extraction to obtain the object data vector and the object attribute vector of the first object.

10. The method according to any one of claims 1 to 6, characterized in that: The method further comprises: Obtaining identification information and multimodal information of the shared video, wherein the identification information includes at least one of identification sequence information of the shared video, author information of the shared video, and tag attribute information of the shared video, and the multimodal information of the shared video includes at least two of shared video title segmentation information, shared video title information, shared video cover information, and shared video content information; The identification information of the shared video and the multimodal information of the shared video are respectively input into corresponding feature extraction networks for feature extraction to obtain the identification feature vector and the multimodal feature vector of the shared video.

11. The method according to any one of claims 1 to 6, characterized in that: The method further comprises: Obtaining identification information and multimodal information of the candidate video, wherein the identification information includes at least one of ID sequence information of the candidate video, author information of the candidate video, and tag attribute information of the candidate video, and the multimodal information of the candidate video includes at least two of candidate video title segmentation information, candidate video title information, candidate video cover information, and candidate video content information; The identification information of the candidate video and the multimodal information of the candidate video are respectively input into corresponding feature extraction networks for feature extraction to obtain the identification feature vector and the multimodal feature vector of the candidate video.

12. A video recommendation model training method, characterized in that: The video recommendation model includes a feature fusion network; The method comprises: Obtaining a multidimensional video feature vector of a shared video, a multidimensional object feature vector of a first object, a multidimensional video feature vector of a positive candidate video sample in a candidate video set, and a multidimensional video feature vector of a negative candidate video sample, wherein the shared video refers to a video clicked by the first object after receiving the video shared by the second object, the multidimensional video feature vector includes an identification feature vector and a multimodal feature vector, the positive candidate video sample refers to a video associated with the shared video, and the negative candidate video sample refers to a video not associated with the shared video; Obtaining a multidimensional object feature vector of the second object, and obtaining a first directional feature vector between the second object and the first object, wherein the first directional feature vector is used to indicate that the shared video is shared from the second object to the first object; the first directional feature vector is used to indicate that the second object in a first age group shares a video with the first object in a second age group; or, the second object in a first region shares a video with the first object in a second region; fusing the multidimensional video feature vector of the shared video, the multidimensional object feature vector of the first object, the multidimensional object feature vector of the second object, and the first directional feature vector through the feature fusion network to obtain an object fusion vector of the first object; The multidimensional video feature vectors of the same positive candidate video sample are fused through the feature fusion network to obtain a positive candidate video sample fusion vector of the positive candidate video sample; the multidimensional video feature vectors of the same negative candidate video sample are fused through the feature fusion network to obtain a negative candidate video sample fusion vector of the negative candidate video sample; performing similarity calculation based on the object fusion vector of the first object and the positive candidate video sample fusion vector to obtain a first similarity value between the object fusion vector and the positive candidate video sample fusion vector; performing similarity calculation based on the object fusion vector of the first object and the negative candidate video sample fusion vector to obtain a second similarity value between the object fusion vector and the negative candidate video sample fusion vector; Calculating a loss function value based on the first similarity value and the second similarity value; Updating model parameters of the video recommendation model based on the loss function value; The identification feature vector is a feature vector extracted based on identification information of the video, and the multimodal feature vector is a feature vector extracted based on multimodal information of the video.

13. The method according to claim 12, characterized in that The video recommendation model also includes an attention mechanism network; The step of fusing the multidimensional video feature vector of the shared video, the multidimensional object feature vector of the first object, the multidimensional object feature vector of the second object, and the first directional feature vector through the feature fusion network to obtain an object fusion vector of the first object includes: Inputting the identification feature vector of the shared video, the multimodal feature vector of the shared video, the multidimensional object feature vector of the first object, the multidimensional object feature vector of the second object, and the first directional feature vector into the attention mechanism network for weight calculation, thereby obtaining a first weight value corresponding to each feature vector; According to the first weight value, the identification feature vector of the shared video, the multimodal feature vector of the shared video, the multidimensional object feature vector of the first object, the multidimensional object feature vector of the second object and the first directional feature vector are input into the feature fusion network for feature fusion to obtain the object fusion vector of the first object.

14. A video recommendation device, characterized in that: The device comprises: an acquisition module, configured to acquire a multidimensional video feature vector of a shared video, a multidimensional object feature vector of a first object, and multidimensional video feature vectors of candidate videos in a candidate video set, wherein the shared video refers to a video clicked by the first object after receiving a video shared by the second object, and the multidimensional video feature vector includes an identification feature vector and a multimodal feature vector; The acquisition module is further configured to acquire a multi-dimensional object feature vector of the second object and a first directional feature vector between the second object and the first object, wherein the first directional feature vector is configured to indicate that the shared video is shared from the second object to the first object; the first directional feature vector is configured to indicate that the second object in a first age group shares a video with the first object in a second age group; or the second object in a first region shares a video with the first object in a second region; a fusion module, configured to fuse the multidimensional video feature vector of the shared video, the multidimensional object feature vector of the first object, the multidimensional object feature vector of the second object, and the first directional feature vector to obtain an object fusion vector of the first object; The fusion module is further configured to fuse the multi-dimensional video feature vectors of the same candidate video to obtain a candidate video fusion vector of the candidate video; a calculation module, configured to perform similarity calculation based on the object fusion vector of the first object and the candidate video fusion vector to obtain a similarity value between the object fusion vector and the candidate video fusion vector; A recommendation module is configured to recommend the candidate video to the first object based on the similarity value.

15. A training device for a video recommendation model, characterized in that: The device comprises: an acquisition module, configured to acquire a multidimensional video feature vector of a shared video, a multidimensional object feature vector of a first object, a multidimensional video feature vector of a positive candidate video sample in a candidate video set, and a multidimensional video feature vector of a negative candidate video sample, wherein the shared video refers to a video clicked by the first object after receiving a video shared by the second object, the multidimensional video feature vector includes an identification feature vector and a multimodal feature vector, the positive candidate video sample refers to a video associated with the shared video, and the negative candidate video sample refers to a video not associated with the shared video; The acquisition module is further configured to acquire a multi-dimensional object feature vector of the second object and a first directional feature vector between the second object and the first object, wherein the first directional feature vector is configured to indicate that the shared video is shared from the second object to the first object; the first directional feature vector is configured to indicate that the second object in a first age group shares a video with the first object in a second age group; or the second object in a first region shares a video with the first object in a second region; a fusion module, configured to fuse the multidimensional video feature vector of the shared video, the multidimensional object feature vector of the first object, the multidimensional object feature vector of the second object, and the first directional feature vector to obtain an object fusion vector of the first object; The fusion module is further configured to fuse the multi-dimensional video feature vectors of the same positive candidate video sample to obtain a positive candidate video sample fusion vector of the positive candidate video sample; The fusion module is further configured to fuse the multi-dimensional video feature vectors of the same negative candidate video sample to obtain a negative candidate video sample fusion vector of the negative candidate video sample; a calculation module, configured to perform similarity calculation based on the object fusion vector of the first object and the positive candidate video sample fusion vector to obtain a first similarity value between the object fusion vector and the positive candidate video sample fusion vector; The calculation module is configured to perform similarity calculation based on the object fusion vector of the first object and the negative candidate video sample fusion vector to obtain a second similarity value between the object fusion vector and the negative candidate video sample fusion vector; The calculation module is configured to calculate a loss function value based on the first similarity value and the second similarity value; An updating module is used to update the model parameters of the video recommendation model based on the loss function value.

16. A computer device, characterized in that: The computer device includes: a processor and a memory, wherein at least one computer program is stored in the memory, and the at least one computer program is loaded and executed by the processor to implement the video recommendation method according to any one of claims 1 to 11, or the training method of the video recommendation model according to any one of claims 12 or 13.

17. A computer storage medium, characterized in that The computer-readable storage medium stores at least one computer program, which is loaded and executed by the processor to implement the video recommendation method according to any one of claims 1 to 11, or the training method of the video recommendation model according to any one of claims 12 or 13.

18. A computer program product, characterized in that The computer program product includes a computer program, which is stored in a computer-readable storage medium; the computer program is read and executed from the computer-readable storage medium by a processor of a computer device, so that the computer device performs the video recommendation method according to any one of claims 1 to 11, or the training method of the video recommendation model according to any one of claims 12 or 13.

Images