A video data processing method and device, computer equipment and a storage medium

Abstract

Embodiments of the present application provide a video data processing method and device, computer equipment and storage medium, which can be applied to cloud technology, artificial intelligence, intelligent transportation, auxiliary driving and various other scenarios; wherein the method comprises: obtaining video feature data of a first video and object feature data of at least one object; calling a video matching model and reference feature data to process the video feature data of the first video and the object feature data of the at least one object, and obtaining a matching result between the first video and each object; wherein the video matching model is obtained by training a first neural network based on training samples and a second neural network, and the reference feature data is determined based on video feature data of a plurality of second videos; determining a target object according to the matching result and sending the first video to a terminal device corresponding to the target object, which can effectively improve the recommendation accuracy of cold start videos and thus improve the video recommendation effect.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a video data processing method, apparatus, computer equipment, and storage medium. Background Technology

[0002] With the rapid development of computer and internet technologies, various video platforms have emerged, especially short video platforms, which have become the type of mobile internet application with the longest user dwell time. Since the production of short videos largely depends on video bloggers, while recommendation technology is provided by the short video platforms, recommendation technology is crucial for them. Typically, short video platform recommendation systems, targeting short videos with a large number of click records, can reasonably infer the target audience of short videos based on video feature data such as identifiers and attributes, thus accurately recommending short videos to the target users. However, with the increasing number of newly released short videos on short video platforms every day, how to recommend these newly added short videos has become a significant problem in video recommendation.

[0003] Newly added short videos lack user interaction records, resulting in missing identifying features. These short videos can be understood as cold-start videos. Existing solutions for recommending cold-start videos typically involve replacing some sample short video identifying features with a 0-mask or random mask during model training to simulate newly added videos. A second solution involves learning the mapping from short video attribute features to identifying features, i.e., deriving the identifying features based on the short video's attribute features and the mapping relationship. However, the first solution requires the model to adapt to two different input distributions: a general distribution of short video identifying features and a distribution with a 0-mask or random mask, making accurate recommendations difficult. Furthermore, the actual identifying features in the second solution are learned and contain errors. The learning of the mapping from short video attribute features to identifying features also introduces errors. These errors accumulate, weakening the accuracy of the mapping and hindering accurate recommendations. Therefore, improving the recommendation performance of cold-start videos has become a pressing issue. Summary of the Invention

[0004] This application provides a video data processing method, apparatus, computer equipment, and storage medium, which can effectively improve the accuracy of cold-start video recommendations, thereby enhancing the video recommendation effect.

[0005] In a first aspect, embodiments of this application provide a video data processing method, the method comprising:

[0006] Obtain video feature data of the first video and object feature data of at least one object.

[0007] The video matching model and reference feature data are used to process the video feature data of the first video and the object feature data of the at least one object to obtain the matching result between the first video and each of the at least one object; wherein, the video matching model is obtained by training the first neural network based on training samples and the second neural network, and the reference feature data is determined based on the video feature data of multiple second videos.

[0008] Based on the matching result, a target object is determined from the at least one object, and the first video is sent to the terminal device corresponding to the target object.

[0009] Optionally, the above training samples are obtained in the following way:

[0010] Get video click history.

[0011] Based on the video click records, at least one second video with click records is obtained from the video set.

[0012] Based on the video click records, obtain the object subset corresponding to each second video from the object set.

[0013] Training samples are determined based on the set of objects and the subset of objects corresponding to each second video.

[0014] Secondly, embodiments of this application provide a video data processing apparatus, the apparatus comprising:

[0015] The acquisition module is used to acquire video feature data of the first video and object feature data of at least one object.

[0016] The processing module is used to call a video matching model and reference feature data to process the video feature data of the first video and the object feature data of the at least one object, so as to obtain the matching result between the first video and each of the at least one object; wherein, the video matching model is obtained by training the first neural network based on training samples and the second neural network, and the reference feature data is determined based on the video feature data of multiple second videos.

[0017] The processing module is further configured to determine a target object from the at least one object based on the matching result.

[0018] The sending module is used to send the first video to the terminal device corresponding to the target object.

[0019] Thirdly, embodiments of this application provide a computer device, which includes a processor, a network interface, and a storage device. The processor, the network interface, and the storage device are interconnected. The network interface is controlled by the processor to send and receive data. The storage device is used to store a computer program, which includes program instructions. The processor is configured to invoke the program instructions to execute the video data processing method described in the first aspect.

[0020] Fourthly, embodiments of this application provide a computer storage medium storing program instructions, which, when executed, are used to implement the video data processing method described in the first aspect.

[0021] Fifthly, embodiments of this application provide a computer program product, including a computer program, which, when executed by a computer processor, is used to implement the video data processing method described in the first aspect.

[0022] This application embodiment can acquire video feature data of a first video and object feature data of at least one object; it calls a video matching model and reference feature data to process the video feature data of the first video and the object feature data of the at least one object to obtain a matching result between the first video and each object; wherein, the video matching model is obtained by training a first neural network based on training samples and a second neural network, and the reference feature data is determined based on the video feature data of multiple second videos; based on the matching result, a target object can be determined from at least one object, and the first video is sent to the terminal device corresponding to the target object. This application generates a general video by utilizing the video feature data of a general video (such as the aforementioned second video) with relatively rich click data. The reference feature data is used to process cold-start videos (such as the first video mentioned above) that lack click data. Since the distribution of the reference feature data is consistent with the distribution of the video feature data of the second video, the video matching model will not be weakened by maintaining two different distributions of input, thus ensuring the accuracy of video matching. Furthermore, this application introduces a knowledge distillation learning framework by guiding the learning of the first neural network through the second neural network. This framework can integrate the knowledge of the second neural network into the first neural network, introducing additional information for the recommendation of cold-start videos. It can be seen that this application can effectively improve the recommendation accuracy of cold-start videos based on the video feature data of general videos and the knowledge distillation learning framework, thereby improving the video recommendation effect. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a schematic diagram of the architecture of a video data processing system provided in an embodiment of this application;

[0025] Figure 2 This is a flowchart illustrating a video data processing method provided in an embodiment of this application;

[0026] Figure 3 This is a flowchart illustrating another video data processing method provided in an embodiment of this application;

[0027] Figure 4a This is a schematic diagram of the structure of a model training framework provided in an embodiment of this application;

[0028] Figure 4b This is a schematic diagram illustrating the principle of a model application provided in an embodiment of this application;

[0029] Figure 5 This is a schematic diagram of the structure of a video data processing device provided in an embodiment of this application;

[0030] Figure 6 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0031] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0032] With the research and advancement of artificial intelligence (AI) technology, AI is being studied and applied in various fields, such as smart homes, smart wearable devices, virtual assistants, smart speakers, smart marketing, autonomous driving, drones, robots, smart healthcare, and smart customer service. It is believed that with the development of technology, AI will be applied in more fields and play an increasingly important role.

[0033] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.

[0034] Machine learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory, among others. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and instructional learning.

[0035] Distillation Learning: The distillation learning framework includes a teacher network and a student network. The teacher network transfers knowledge from the teacher network to the student network by guiding the student network in its learning.

[0036] Frequent itemsets: In a set of data, a combination of elements that frequently appear together is called a frequent item. All frequent items in the data constitute a frequent itemset.

[0037] Embedding: A continuous vector used to represent a video / user, specifically a short video.

[0038] Identity (ID) features: One of the main features of short videos / users, namely, a continuous vector that uniquely represents the identity of a short video / user. For short videos lacking click data (referred to as cold-start videos), such as newly added short videos, the recommendation model cannot learn the identity features of cold-start videos from the click data. This is the cold-start video recommendation problem. The video ID can be a string composed of numbers, letters, etc., and the ID itself is fixed. For example, the ID of a video is the number 100. The identity feature is obtained from the video identity feature matrix based on the video ID. For example, the ID can be represented as a vector, specifically a one-hot encoded vector. The video identity feature is obtained from the video identity feature matrix based on the one-hot vector. For cold-start videos, due to the lack of click data, the recommendation model cannot determine the usable video identity feature matrix based on the scarce click data. Of course, as the video click data becomes richer and more complete, the recommendation model can continuously update the video identity feature matrix based on the click data, thus obtaining accurate identity features from the updated video identity feature matrix based on the video ID.

[0039] Attribute features: These are the collective features of a short video / user other than the identifying features; they are continuous vectors representing the attributes of the short video / user. Since short videos / users may have multiple attributes (such as the title and category of the short video, the user's gender, age, etc.), the attribute features of a short video / user are typically a set of vectors.

[0040] Please see Figure 1 This is a schematic diagram of the architecture of a video data processing system provided in an embodiment of this application. The video data processing system includes a computer device 101 and a terminal device 102, wherein:

[0041] Terminal device 102 can send videos created by an object to computer device 101. The object may be a user, and the video may be a short video. Terminal device 102 can also receive videos sent by computer device 101 and display the received videos to the user through a display interface.

[0042] Computer device 101 can add videos sent by terminal device 102 to a video library to save user-created videos; it can also retrieve videos that match the user from the video library and send the videos that match the user to terminal device 102 so that the user can watch their favorite videos.

[0043] The computer device 101 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The terminal device 102 can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, in-vehicle smart terminal, etc., but is not limited to these. The terminal device 102 and the computer device 101 can be directly or indirectly connected via wired or wireless communication, which is not limited herein. The embodiments of this application can be applied to various scenarios, including but not limited to cloud technology, artificial intelligence, smart transportation, and assisted driving.

[0044] In some feasible implementations, computer device 101 obtains a cold start video (denoted as the first video) from a video library, and obtains video feature data of the first video and object feature data of at least one object. The video feature data may include the identification features of the video and an attribute feature set composed of various attribute features of the video. Similarly, the object feature data may include the identification features of the object and an attribute feature set composed of various attribute features of the object. Since the first video lacks identification features, computer device 101 can generate reference feature data based on the video feature data of multiple second videos. The second videos refer to videos with relatively rich click data in the video library, which can also be called general videos. Computer device 101 calls the video matching model and the reference feature data to process the video feature data of the first video and the object feature data of each object to obtain the matching result between the first video and each object. The matching result may be the predicted click probability. The video matching model is obtained by training the first neural network based on training samples and the second neural network, that is, by training the model through the distillation learning framework. Then, computer device 101 can determine the target object from at least one object according to the matching result and send the first video to the terminal device 102 corresponding to the target object.

[0045] As can be seen, this embodiment utilizes video feature data from videos with abundant click data to generate generalized identifier features (i.e., the aforementioned reference feature data), and uses these generalized identifier features as identifier features for cold-start videos. Since the distribution of the generalized identifier features is consistent with the distribution of identifier features from videos with abundant click data, the video matching model will not suffer reduced accuracy due to maintaining two different input distributions, thereby improving the accuracy of video matching. Furthermore, this embodiment introduces a knowledge distillation learning framework by using a second neural network to guide the learning of the first neural network. This framework integrates the knowledge of the second neural network into the first neural network, introducing additional information for the recommendation of cold-start videos and further improving the recommendation accuracy. Therefore, this application, based on video feature data from general videos and a knowledge distillation learning framework, can effectively improve the recommendation accuracy of cold-start videos, thereby enhancing the video recommendation effect.

[0046] The implementation details of the technical solutions in the embodiments of this application are described in detail below:

[0047] Please see Figure 2 This application's embodiments are based on Figure 1 The diagram shown illustrates a video data processing method provided by the video data processing system. This video data processing method can be applied to... Figure 1 The computer device 101 in the middle may specifically include the following steps:

[0048] 201. Obtain video feature data of the first video and object feature data of at least one object.

[0049] The first video can be a cold start video, lacking click data, i.e., lacking user interaction records, such as a newly added video, specifically a short video submitted by a user after creation. The object can specifically be a user, and this at least one object can refer to all users of a particular short video platform.

[0050] The video feature data includes identifier features and attribute feature sets. Attribute feature sets are collections of multiple attribute features of a video, such as title, category, description, and image information within the video. The object feature data also includes identifier features and attribute feature sets. Attribute feature sets are collections of multiple attribute features of an object, such as age, gender, interest tags, and geographic location.

[0051] In some feasible implementations, for the first video v i Its identifying characteristics can be represented as Its attribute feature set can be represented as The total number of attribute features of the video; for a given object u j Its identifying characteristics can be represented as follows: Its attribute feature set can be represented as The total number of attributes of an object.

[0052] 202. The video matching model and reference feature data are used to process the video feature data of the first video and the object feature data of the at least one object to obtain the matching result between the first video and each of the at least one object; wherein the video matching model is obtained by training the first neural network based on training samples and the second neural network, and the reference feature data is determined based on the video feature data of multiple second videos.

[0053] The second video can be understood as a video in a video collection or video library that contains click data, or it can be understood as a video with relatively rich click data. Such videos can be called general videos. The matching result can be a predicted click probability, that is, the likelihood of each object clicking the first video.

[0054] In practice, since the first video lacks identifying features, the computer device can generate reference feature data based on the video feature data of multiple second videos. Then, it calls the video matching model and the reference feature data to process the video feature data of the first video and the object feature data of each object to obtain the matching result between the first video and each object. Since the distribution of generalized identifying features is consistent with the distribution of identifying features of videos with richer click data, the video matching model will not weaken its accuracy due to maintaining two different distributions of input, thereby improving the accuracy of video matching.

[0055] The video matching model is trained on the first neural network using training samples and the second neural network. In other words, the model is trained using the distillation learning framework, which integrates the knowledge of the second neural network into the first neural network. This introduces additional information for the recommendation of cold-start videos and further improves the recommendation accuracy of cold-start videos.

[0056] In some feasible implementations, the video matching model can be specifically trained periodically, for example, once a day, and the videos added after a set time each day can be used as cold start videos, and the general videos already in the video library before the set time (such as the second video mentioned above) can be used as training samples for training.

[0057] In some feasible implementations, the computer device can generate reference feature data based on the identifier features included in the video feature data of multiple second videos. This reference feature data can be understood as generalized identifier features. Then, the computer device directly uses the reference feature data as the identifier features in the video feature data of the first video. This not only supplements the identifier features of the cold start video and improves its video feature data, but also ensures that the distribution of the identifier features supplemented in the cold start video is consistent with the distribution of the identifier features in the general video. This allows the video matching model to improve accuracy by only maintaining the same distribution of inputs, thereby improving the accuracy of video matching.

[0058] In some feasible implementations, the computer device can acquire multiple second videos with click records in the video set, obtain the identification features of each second video from the video feature data of each second video, and then perform global average pooling on the identification features of multiple second videos to obtain reference feature data. Thus, through global average pooling, reference feature data (i.e., generalized identification features) that are consistent with the distribution of identification features of general videos can be obtained.

[0059] In some feasible implementations, the specific implementation of a computer device calling a video matching model and reference feature data to obtain the matching results between a first video and each of at least one object may include:

[0060] First, the computer device can obtain the first fusion feature of each object based on the identification features and attribute feature set of each object. For example, the identification features of each object are concatenated with the various attribute features included in the attribute feature set, and the concatenation result is the first fusion feature. Similarly, the computer device can obtain the second fusion feature of the first video based on the identification features and attribute feature set of the first video. For example, the identification features of the first video are concatenated with the various attribute features included in the attribute feature set, and the concatenation result is the second fusion feature.

[0061] Then, the computer device can call the video matching model to map the first fusion feature of each object to obtain the third fusion feature of each object, and to map the second fusion feature of the first video to obtain the fourth fusion feature of the first video, thereby mapping the video feature data and the object feature data to a consistent feature space.

[0062] Then, the computer device can call the video matching model to perform matching processing on the third fusion feature and the fourth fusion feature to obtain the matching result between the first video and each of the at least one object. Specifically, the third fusion feature and the fourth fusion feature are feature vectors. The similarity between the third fusion feature and the fourth fusion feature can be calculated by taking the cosine similarity. Then, the similarity is used as the predicted click probability between the first video and the object.

[0063] The first neural network can adopt a dual-tower model, where the first and second fusion features obtained after splicing are respectively transformed into the third and fourth fusion features through a tower-shaped neural network and a highly nonlinear mapping.

[0064] 203. Determine the target object from the at least one object based on the matching result, and send the first video to the terminal device corresponding to the target object.

[0065] In practice, the computer device can determine a target number of objects from at least one set of objects based on the matching results between the first video and each object, and then send the first video to the terminal devices corresponding to each target object. For example, if the target number is 3 and the matching result is the predicted click probability, the computer device can select the top 3 objects with the highest predicted click probabilities as target objects based on the predicted click probabilities between the first video and each object.

[0066] In some feasible implementations, in addition to matching a suitable target object for each first video, a suitable target video can also be matched for each object. For example, for each object (such as a target user), the matching results between the target user and each video in the video set are obtained. For each first video (i.e., a cold start video) in the video set, the reference feature data is used as the identifier feature of the first video, and then input into the video matching model for processing. For each second video (i.e., a general video) in the video set, the identifier feature of each second video is input into the video matching model for processing. The matching results corresponding to each first video and second video in the video set are sorted, and the target video matching the target user is determined from the video set according to the sorting results. The target video can be the first video, the second video, or a combination of the first and second videos.

[0067] It should be noted that the video data processing method provided in this application embodiment is also applicable to other types of videos besides short videos, such as long videos. That is, this application embodiment can also be used in other video recommendation scenarios and can effectively improve the video recommendation effect.

[0068] In this embodiment, the computer device can acquire video feature data of a first video and object feature data of at least one object. Since the first video lacks identifying features, the computer device can generate reference feature data based on the video feature data of multiple second videos. The second videos refer to videos with relatively rich click data in a video library, also known as general videos. The computer device calls a video matching model and the reference feature data to process the video feature data of the first video and the object feature data of each object, obtaining a matching result between the first video and each object. The matching result can be a predicted click probability. The video matching model is obtained by training a first neural network based on training samples and a second neural network, i.e., through distillation. The learning framework is used for model training. Then, the computer device can determine the target object from at least one object based on the matching results and send the first video to the terminal device corresponding to the target object. This generates generalized identifier features using video feature data from videos with rich click data. Since the distribution of the generalized identifier features is consistent with the distribution of the identifier features from videos with rich click data, the video matching model will not suffer from reduced accuracy due to maintaining two different input distributions, thus improving the accuracy of video matching. Furthermore, this application embodiment introduces a knowledge distillation learning framework by guiding the learning of the first neural network through a second neural network. This framework integrates the knowledge of the second neural network into the first neural network, introducing additional information for cold-start video recommendation. Therefore, this application, based on video feature data from general videos and a knowledge distillation learning framework, can effectively improve the recommendation accuracy of cold-start videos, thereby enhancing the video recommendation effect.

[0069] Please see Figure 3 This application's embodiments are based on Figure 1 The diagram shows a flowchart of another video data processing method provided by the video data processing system. This video data processing method can be applied to... Figure 1 The computer device 101 in the middle may specifically include the following steps:

[0070] 301. Obtain training samples, wherein the training samples include a second video, a first object in the object subset corresponding to the second video, and a second object in the object set other than the object subset.

[0071] Specifically, computer devices can use the second video (i.e., the general video) in the video set for training neural networks. The object set can include all objects on a video platform, such as all users on a short video platform. For any second video, the object subset corresponding to the second video is the set of all objects in the object set that have clicked on the second video. Each object that has clicked on the second video can be denoted as the first object. Objects in the object set other than this object subset, that is, objects that have not yet clicked on the second video, can be denoted as the second object.

[0072] It should be noted that since it is uncertain whether the second object will click on the second video in the future, all objects that have not yet clicked on the second video cannot be considered negative samples. Therefore, this application adopts a relatively mild approach, assuming that the first object has a higher click probability than the second object for the second video. Based on this comparative idea, the constructed training samples include three elements: video, object 1, and object 2. For example, the training samples... It can be represented as:

[0073]

[0074] Among them, v i It can refer to any second video, u j It can refer to any first object, u k It can refer to any second object, where V represents the video set; U(i) represents v. i The corresponding subset of objects, that is, the video v clicked in the object set U. i The set consisting of each first object, U\U(i) represents the other objects in set U besides U(i), (v i u j u k ) indicates that for video v i u j Than u k It has a higher click-through rate.

[0075] in, N v and N u These represent the total number of short videos and the total number of users, respectively.

[0076] In some feasible implementations, the specific process of constructing training samples may include:

[0077] The computer device acquires video click records, and based on these video click records, it acquires at least one second video from the video set that contains click records (or click data). Alternatively, it can acquire at least one second video with relatively abundant click records. Based on the video click records, it acquires a subset of objects corresponding to each second video from the object set. Based on the object set and the subset of objects corresponding to each second video, it determines training samples.

[0078] Among them, video click records can be represented as Where, N s v represents the total number of click records. i The corresponding subset of objects U(i) can be represented as U(i) = {u j |(v i u j )∈S}.

[0079] In some feasible implementations, to simulate cold-start videos encountered in real-world applications during training, the identifier features of some second videos in the training samples can be processed, for example, replaced with reference feature data. Specifically, this can include: determining a predetermined number of second videos from at least one second video included in the training samples; replacing the identifier features in the video feature data of the predetermined number of second videos with reference feature data, thereby improving the matching accuracy of the model for cold-start videos during application.

[0080] The preset number can be determined based on the ratio of the number of first videos to the number of second videos. For example, the preset number can be 1% to 2% of all second videos included in the training samples.

[0081] 302. The first neural network is trained based on the training samples and the second neural network to obtain a video matching model; wherein the second neural network is constructed based on the first neural network and business knowledge data, and the business knowledge data is generated based on video click records.

[0082] In practice, computer devices can generate business knowledge data based on video click records. This business knowledge data reflects the correlation between video attributes and user attributes. A second neural network is then constructed based on the first neural network and the business knowledge data. The first neural network acts as the student network, and the second neural network acts as the teacher network. The teacher network is essentially the student network embedded with the business knowledge data. The first and second neural networks share network parameters. The computer device trains a video matching model using training samples and the second neural network through a distillation learning framework. The second neural network guides the learning of the first neural network. This introduction of the knowledge distillation learning framework allows the business knowledge of the second neural network to be integrated into the first neural network, providing additional information for cold-start video recommendations and improving the recommendation accuracy.

[0083] In some feasible implementations, the computer device trains the first neural network based on training samples and the second neural network to obtain the video matching model. Specific implementations may include:

[0084] First, the computer device calls the first neural network to obtain the first predicted matching result between the first object and the second video in the training samples, and the second predicted matching result between the second object and the second video;

[0085] Then, the computer device obtains the third predictive matching result between the first object and the second video and the fourth predictive matching result between the second object and the second video based on the business knowledge data included in the second neural network; at this point, the computer device obtains its respective predictive matching results using the student network and the teacher network.

[0086] Next, the computer device adjusts the network parameters of the first neural network based on the first predicted matching result, the second predicted matching result, the third predicted matching result, the fourth predicted matching result, and the target loss function. When the training process reaches the convergence condition, the first neural network with adjusted network parameters is used as the video matching model.

[0087] In some feasible implementations, during training, the computer device can first determine the total loss and then use the total loss to adjust the network parameters of the first neural network. The process of determining the total loss may include:

[0088] Assuming the target loss function includes a first loss function and a second loss function, the computer device can determine the first loss value of the first neural network based on the prediction results of the first neural network (i.e., the first prediction matching result and the second prediction matching result) and the first loss function; then, based on the prediction results of the first neural network (i.e., the first prediction matching result and the second prediction matching result), the prediction results of the second neural network (i.e., the third prediction matching result and the fourth prediction matching result), and the second loss function, determine the second loss value of the first neural network; then, based on the first loss value, the second loss value, and the balance parameter, determine the total loss value of the first neural network, and the network parameters of the first neural network can be adjusted based on the total loss value.

[0089] The first loss value can be expressed as:

[0090] The second loss value can be expressed as:

[0091] The total loss value can be expressed as:

[0092] in, This is the first predicted match result. This is the second predicted matching result. This is the third predicted matching result. The fourth prediction matching result is given, where ρ is a hyperparameter of the two learning objectives of the student network.

[0093] In some feasible implementations, it is assumed that the first predicted matching result is represented as The third predicted matching result is represented as follows: but

[0094] Among them, f g (i, j) is the business rule function, C is the hyperparameter for adjusting the exponent range, and g represents the rule currently applied, which is used to reward predictions that conform to the positive rule and punish predictions that conform to the negative rule.

[0095] In some feasible implementations, the specific implementation of the computer device obtaining the first predicted matching result between the first object and the second video, and the second predicted matching result between the second object and the second video, can be referred to the relevant description of "obtaining the matching result between the first video and each of at least one object" in step 202 of the above embodiment, which will not be repeated here.

[0096] In some feasible implementations, the specific ways in which the computer device obtains the third predictive matching result between the first object and the second video and the fourth predictive matching result between the second object and the second video based on the business knowledge data included in the second neural network may include:

[0097] The computer device uses a second neural network and business knowledge data to calculate the prediction results between the first object, the second object and the second video respectively. When calculating the prediction results, since the second neural network shares network parameters with the first neural network, the matching result can be predicted directly based on the prediction results of the first neural network (i.e., the first prediction matching result and the second prediction matching result).

[0098] In a specific implementation, the computer device calibrates the first prediction matching result based on the business knowledge data included in the second neural network to obtain a third prediction matching result between the first object and the second video; and calibrates the second prediction matching result based on the business knowledge data included in the second neural network to obtain a fourth prediction matching result between the second object and the second video.

[0099] In some feasible implementations, the computer device calibrates the first predictive matching result based on the business knowledge data included in the second neural network to obtain a third predictive matching result between the first object and the second video. Specific implementation methods may include:

[0100] The computer device obtains attribute data tuples corresponding to the first object and the second video based on the attribute feature set of the first object and the attribute feature set of the second video; determines the first calibration parameter based on the business knowledge data included in the second neural network and the attribute data tuples; and determines the third prediction matching result between the first object and the second video based on the first calibration parameter and the first prediction matching result.

[0101] In some feasible implementations, the business knowledge data may specifically include a first business rule set and a second business rule set, and the specific process of obtaining the aforementioned first calibration parameter may include:

[0102] For any attribute data tuple in the attribute data tuples corresponding to the first object and the second video, if any attribute data tuple matches the first business rule set, the computer device determines the calibration component corresponding to that attribute data tuple as the first component.

[0103] If any attribute data tuple matches the second business rule set, the computer device determines that the calibration component corresponding to that attribute data tuple is the second component;

[0104] If any attribute data tuple does not match either the first business rule set or the second business rule set, the computer device determines that the calibration component corresponding to that attribute data tuple is the third component.

[0105] The first calibration parameter is determined based on the calibration component corresponding to each attribute data tuple.

[0106] The first calibration parameter is the one defined in the above formula.

[0107] Here, it is assumed that the business knowledge data includes the first business rule set R. + Second business rule set R - The aforementioned business rule function f g The specific representation of (i, j) is as follows:

[0108]

[0109] Therefore, the first component can be 1, the second component can be -1, and the third component can be 0, thus rewarding predictions that conform to the positive rule and penalizing predictions that conform to the negative rule.

[0110] Understandably, the method for determining the fourth predicted matching result between the second object and the second video can be as follows:

[0111] The computer device obtains attribute data tuples corresponding to the second object and the second video based on the attribute feature set of the second object and the attribute feature set of the second video; determines the second calibration parameters based on the business knowledge data included in the second neural network and the attribute data tuples; and determines the fourth prediction matching result between the second object and the second video based on the second calibration parameters and the second prediction matching result.

[0112] Similarly, the process of determining the second calibration parameter may specifically include:

[0113] For any attribute data tuple in the attribute data tuples corresponding to the second object and the second video, if any attribute data tuple matches the first business rule set, the computer device determines the calibration component corresponding to that attribute data tuple as the first component.

[0114] If any attribute data tuple matches the second business rule set, the computer device determines that the calibration component corresponding to that attribute data tuple is the second component;

[0115] If any attribute data tuple does not match either the first business rule set or the second business rule set, the computer device determines that the calibration component corresponding to that attribute data tuple is the third component.

[0116] The second calibration parameter is determined based on the calibration component corresponding to each attribute data tuple.

[0117] 303. Obtain video feature data of the first video and object feature data of at least one object.

[0118] 304. The video matching model and reference feature data are used to process the video feature data of the first video and the object feature data of the at least one object to obtain the matching result between the first video and each of the at least one object; wherein the reference feature data is determined based on the video feature data of multiple second videos.

[0119] 305. Determine the target object from the at least one object based on the matching result, and send the first video to the terminal device corresponding to the target object.

[0120] The specific implementation of steps 303 to 305 can be found in the relevant descriptions of steps 201 to 203 in the above embodiments, and will not be repeated here.

[0121] In some feasible implementations, video click records include at least one click data tuple, each click data tuple including an object and a video. The process of constructing business knowledge data can be as follows:

[0122] Assuming the objects included in a click data tuple are denoted as the third object, and the videos included are denoted as the third video, then for each click data tuple containing the third object and the third video, the attribute feature set of the third object is obtained from the object feature data of the third object, and the attribute feature set of the third video is obtained from the video feature data of the third video. Based on the attribute feature sets of the third object and the third video, at least one attribute data tuple is determined. The frequency corresponding to each attribute data tuple is obtained, and business knowledge data is determined from at least one attribute data tuple based on the frequency of each attribute data tuple. For example, the top K most frequent attribute data tuples and the bottom K most frequent attribute data tuples can be used as business knowledge data. Typically, the correlation between short videos and users can be partially mined from their attribute features. For example, users with an age attribute of "15-25 years old" and a gender attribute of "male" are more likely to click on short videos with a category attribute of "basketball" or "games". This can be achieved by statistically analyzing attribute tuples. The frequency of these data can, to some extent, infer the correlation between short videos and user attribute values, which is essentially business knowledge data.

[0123] Among them, the attribute data tuples with the highest frequency can be called the frequent itemsets mentioned above.

[0124] The video click record can be a deduplicated click record. For example, if object 1 clicks video 1 3 times, the corresponding click data tuple can include 3, namely {video 1, object 1}, {video 1, object 1}, and {video 1, object 1}. In order to reduce the amount of data and improve the efficiency of video recommendation, the click data tuple can be deduplicated, that is, only one {video 1, object 1} click data tuple can be added to the video click record.

[0125] In some feasible implementations, to improve the accuracy of business knowledge data, video click records can also include the number of clicks for each click data tuple. That is, for repeated click data tuples, the video click record has only one record, but it can record the actual number of clicks. The record format is {Video i, Object y, Click count x}. For example, for Object 1 and Video 1 mentioned above, the record format for the click data tuples included in the video click record can be {Video 1, Object 1, 3}. Therefore, when obtaining the frequency corresponding to each attribute data tuple, it can be multiplied by the actual number of clicks. Assuming the video click record includes {Video 1, Object 1, 3} and {Video 2, Object 2, 1}, and there is an attribute data tuple {a1, a2} between Video 1 and Object 1, and also between Video 1 and Object 1, then the frequency corresponding to the attribute data tuple {a1, a2} is 1*3 + 1*1 = 4 times.

[0126] In this embodiment, the computer device generates business knowledge data based on video click records. This business knowledge data reflects the correlation between video attributes and user attributes. A second neural network is constructed based on a first neural network and the business knowledge data. The first neural network serves as a student network, and the second neural network serves as a teacher network. The teacher network is essentially the student network embedded with the business knowledge data. The first and second neural networks share network parameters. Model training is performed using a distillation learning framework, which integrates the knowledge of the second neural network into the first neural network, introducing additional information for cold-start video recommendation. Simultaneously, the computer device generates generalized identifier features using video feature data from videos with abundant click data. Since the distribution of the generalized identifier features is consistent with the distribution of the identifier features from videos with abundant click data, the video matching model will not suffer reduced accuracy due to maintaining two different input distributions, thus improving the accuracy of video matching and enhancing the video recommendation effect.

[0127] In some feasible implementations, such as Figure 4a The diagram shown is a schematic representation of a model training framework provided in an embodiment of this application.

[0128] The training framework of the model includes a first neural network 401 and a second neural network 402. The first neural network 401 serves as the student network, and the second neural network 402 is constructed by embedding business knowledge (i.e., the aforementioned business knowledge data) on the basis of the first neural network 401. The second neural network 402 serves as the teacher network.

[0129] The specific training process based on this model training framework may include:

[0130] (1) The first neural network 401 processes the video feature data of the video, the object feature data of user 1, and the object feature data of user 2 in the training samples (video, user 1, user 2) to obtain the prediction results of the student network. The prediction results of the student network include the first prediction result between user 1 and the video and the second prediction result between user 2 and the video.

[0131] (2) The second neural network 402 uses business knowledge to obtain the prediction results of the teacher network based on the prediction results of the first neural network 401. The prediction results of the teacher network include the third prediction result between user 1 and the video, and the fourth prediction result between user 2 and the video.

[0132] (3) Calculate the first loss using the prediction results of the student network, calculate the second loss using the prediction results of the teacher network, and then calculate the total loss of the student network using the first loss and the second loss. Then, the network parameters of the student network can be adjusted using the total loss. When the convergence condition is met, the student network with adjusted network parameters can be used as the video matching model 403 mentioned above. The video matching model 403 can be used to predict the click probability between cold start videos and users, thereby achieving accurate recommendation of cold start videos.

[0133] In some feasible implementations, such as Figure 4b The diagram shown is a schematic diagram illustrating the principle of a model application provided in an embodiment of this application.

[0134] Taking the recommendation of cold-start short videos as an example, the video matching model 403 can embed the attributes of the input cold-start short videos to obtain the attribute feature set in the video feature data of the cold-start short videos, and embed the identifiers and attributes of each input object (such as a user) to obtain the object feature data of the object. Since cold-start short videos lack identifier features, the reference feature data obtained by global average pooling can be used as the identifier features in the video feature data of the cold-start short videos. For example, global average pooling can be performed on the identifier features of multiple general videos to obtain reference feature data; then the video matching model 403 can perform the video matching on the cold-start short videos. The system performs matching processing on frequency feature data and object feature data of each object, such as calculating similarity, to obtain matching results between the cold start short video and each object, such as predicting click probability. Based on the matching results, the system can determine the target object that matches the cold start short video and then recommend the cold start short video to the target object. This not only supplements the cold start short video with identification features and improves its video feature data, but also ensures that the distribution of the supplemented identification features of the cold start short video is consistent with the distribution of the identification features of general videos. This allows the video matching model to improve accuracy by only maintaining the same distribution of inputs.

[0135] It is understood that in the specific implementation of this application, data such as the attributes of the object are involved. When the above embodiments of this application are applied to specific products or technologies, the permission or consent of the object is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0136] Please see Figure 5 This is a schematic diagram of the structure of a video data processing device according to an embodiment of this application. The video data processing device according to this embodiment includes:

[0137] The acquisition module 501 is used to acquire video feature data of the first video and object feature data of at least one object.

[0138] The processing module 502 is used to call a video matching model and reference feature data to process the video feature data of the first video and the object feature data of the at least one object, so as to obtain the matching result between the first video and each of the at least one object; wherein, the video matching model is obtained by training the first neural network based on training samples and the second neural network, and the reference feature data is determined based on the video feature data of multiple second videos.

[0139] The processing module 502 is further configured to determine a target object from the at least one object based on the matching result.

[0140] The sending module 503 is used to send the first video to the terminal device corresponding to the target object.

[0141] Optionally, the video feature data includes identification features and attribute feature sets, and the processing module 502 is specifically used for:

[0142] The reference feature data is used as the identifier feature in the video feature data of the first video, and the reference feature data is determined based on the identifier features included in the video feature data of the plurality of second videos.

[0143] The video matching model is invoked to process the video feature data of the first video and the object feature data of the at least one object to obtain the matching result between the first video and each of the at least one object.

[0144] Optionally, the acquisition module 501 is further configured to acquire multiple second videos in the video set that have click records; and to acquire the identification features of each second video from the video feature data of each second video;

[0145] The processing module 502 is further configured to perform global average pooling on the identifier features of the plurality of second videos to obtain the reference feature data.

[0146] Optionally, the acquisition module 501 is further configured to acquire training samples, the training samples including the second video, a first object in the object subset corresponding to the second video, and a second object in the object set other than the object subset.

[0147] The processing module 502 is further configured to train the first neural network based on the training samples and the second neural network to obtain a video matching model.

[0148] The second neural network is constructed based on the first neural network and business knowledge data, which is generated based on video click records.

[0149] Optionally, the acquisition module 501 is specifically used for:

[0150] Get video click history.

[0151] Based on the video click records, at least one second video with click records is obtained from the video set.

[0152] Based on the video click records, obtain the object subset corresponding to each second video from the object set.

[0153] Training samples are determined based on the set of objects and the subset of objects corresponding to each second video.

[0154] Optionally, the acquisition module 501 is further configured to determine a preset number of second videos from at least one second video included in the training samples.

[0155] The processing module 502 is further configured to replace the identifier features in the video feature data of the preset number of second videos with the reference feature data.

[0156] Optionally, the video click record includes at least one click data tuple, each click data tuple including an object and a video. The acquisition module 501 is further configured to, for each click data tuple including a third object and a third video, obtain the attribute feature set of the third object from the object feature data of the third object, and obtain the attribute feature set of the third video from the video feature data of the third video.

[0157] The processing module 502 is further configured to determine at least one attribute data tuple based on the attribute feature set of the third object and the attribute feature set of the third video.

[0158] The processing module 502 is further configured to obtain the frequency corresponding to each attribute data tuple, and determine business knowledge data from the at least one attribute data tuple based on the frequency corresponding to each attribute data tuple.

[0159] Optionally, the processing module 502 is specifically used for:

[0160] The first neural network is invoked to obtain the first prediction matching result between the first object and the second video, and the second prediction matching result between the second object and the second video.

[0161] Based on the business knowledge data included in the second neural network, a third predictive matching result between the first object and the second video, and a fourth predictive matching result between the second object and the second video are obtained.

[0162] Based on the first prediction matching result, the second prediction matching result, the third prediction matching result, the fourth prediction matching result, and the target loss function, the network parameters of the first neural network are adjusted, and the first neural network with adjusted network parameters is used as the video matching model.

[0163] Optionally, the processing module 502 is specifically used for:

[0164] The first prediction matching result is calibrated based on the business knowledge data included in the second neural network to obtain a third prediction matching result between the first object and the second video.

[0165] The second prediction matching result is calibrated based on the business knowledge data included in the second neural network to obtain a fourth prediction matching result between the second object and the second video.

[0166] Optionally, the processing module 502 is specifically used for:

[0167] Based on the attribute feature set of the first object and the attribute feature set of the second video, obtain the attribute data tuples corresponding to the first object and the second video.

[0168] The first calibration parameter is determined based on the business knowledge data included in the second neural network and the attribute data tuples.

[0169] A third predictive matching result is determined between the first object and the second video based on the first calibration parameters and the first predictive matching result.

[0170] Optionally, the business knowledge data includes a first business rule set and a second business rule set, and the processing module 502 is specifically used for:

[0171] For any attribute data tuple in the attribute data tuples corresponding to the first object and the second video, if the attribute data tuple matches the first business rule set, then the calibration component corresponding to the attribute data tuple is determined as the first component.

[0172] If any attribute data tuple matches the second business rule set, then the calibration component corresponding to any attribute data tuple is determined to be the second component.

[0173] If any attribute data tuple does not match either the first business rule set or the second business rule set, then the calibration component corresponding to any attribute data tuple is determined to be the third component.

[0174] The first calibration parameter is determined based on the calibration components corresponding to each attribute data tuple.

[0175] Optionally, the target loss function includes a first loss function and a second loss function, and the processing module 502 is specifically used for:

[0176] Based on the first prediction matching result, the second prediction matching result, and the first loss function, a first loss value of the first neural network is determined.

[0177] Based on the first prediction matching result, the second prediction matching result, the third prediction matching result, the fourth prediction matching result, and the second loss function, the second loss value of the first neural network is determined.

[0178] The total loss value of the first neural network is determined based on the first loss value, the second loss value, and the balance parameter, and the network parameters of the first neural network are adjusted based on the total loss value.

[0179] It should be noted that the functions of each functional module of the video data processing device in this application embodiment can be specifically implemented according to the methods in the above method embodiments. The specific implementation process can be referred to the relevant descriptions in the above method embodiments, which will not be repeated here.

[0180] Please see Figure 6 This is a schematic diagram of the structure of a computer device according to an embodiment of this application. The computer device according to this embodiment includes a power supply module and other structures, and includes a processor 601, a storage device 602, and a network interface 603. The processor 601, the storage device 602, and the network interface 603 can exchange data.

[0181] The storage device 602 may include volatile memory, such as random-access memory (RAM); the storage device 602 may also include non-volatile memory, such as flash memory, solid-state drive (SSD), etc.; the storage device 602 may also include a combination of the above types of memory.

[0182] The processor 601 may be a central processing unit (CPU). In one embodiment, the processor 601 may also be a graphics processing unit (GPU). The processor 601 may also be a combination of a CPU and a GPU. In one embodiment, the storage device 602 is used to store program instructions. The processor 601 can invoke the program instructions to perform the following operations:

[0183] Obtain video feature data of the first video and object feature data of at least one object.

[0184] The video matching model and reference feature data are used to process the video feature data of the first video and the object feature data of the at least one object to obtain the matching result between the first video and each of the at least one object; wherein, the video matching model is obtained by training the first neural network based on training samples and the second neural network, and the reference feature data is determined based on the video feature data of multiple second videos.

[0185] The target object is determined from the at least one object based on the matching result.

[0186] The first video is sent to the terminal device corresponding to the target object by calling network interface 603.

[0187] Optionally, the video feature data includes identification features and attribute feature sets, and the processor 601 is specifically used for:

[0188] The reference feature data is used as the identifier feature in the video feature data of the first video, and the reference feature data is determined based on the identifier features included in the video feature data of the plurality of second videos.

[0189] The video matching model is invoked to process the video feature data of the first video and the object feature data of the at least one object to obtain the matching result between the first video and each of the at least one object.

[0190] Optionally, the processor 601 is further configured to:

[0191] Retrieve multiple second videos from a video collection that have click records.

[0192] The identification features of each second video are obtained from the video feature data of each second video.

[0193] The identification features of the multiple second videos are subjected to global average pooling to obtain the reference feature data.

[0194] Optionally, the processor 601 is further configured to:

[0195] Obtain training samples, which include the second video, a first object in the object subset corresponding to the second video, and a second object in the object set other than the object subset.

[0196] The first neural network is trained based on the training samples and the second neural network to obtain a video matching model.

[0197] The second neural network is constructed based on the first neural network and business knowledge data, which is generated based on video click records.

[0198] Optionally, the processor 601 is specifically used for:

[0199] Get video click history.

[0200] Based on the video click records, at least one second video with click records is obtained from the video set.

[0201] Based on the video click records, obtain the object subset corresponding to each second video from the object set.

[0202] Training samples are determined based on the set of objects and the subset of objects corresponding to each second video.

[0203] Optionally, the processor 601 is further configured to:

[0204] A predetermined number of second videos are determined from at least one second video included in the training samples.

[0205] Replace the identifier features in the video feature data of the preset number of second videos with the reference feature data.

[0206] Optionally, the video click record includes at least one click data tuple, each click data tuple including an object and a video, and the processor 601 is further configured to:

[0207] For each click data tuple including the third object and the third video, the attribute feature set of the third object is obtained from the object feature data of the third object, and the attribute feature set of the third video is obtained from the video feature data of the third video.

[0208] Based on the attribute feature set of the third object and the attribute feature set of the third video, at least one attribute data tuple is determined.

[0209] Obtain the frequency corresponding to each attribute data tuple, and determine business knowledge data from the at least one attribute data tuple based on the frequency corresponding to each attribute data tuple.

[0210] Optionally, the processor 601 is specifically used for:

[0211] The first neural network is invoked to obtain the first prediction matching result between the first object and the second video, and the second prediction matching result between the second object and the second video.

[0212] Based on the business knowledge data included in the second neural network, a third predictive matching result between the first object and the second video, and a fourth predictive matching result between the second object and the second video are obtained.

[0213] Based on the first prediction matching result, the second prediction matching result, the third prediction matching result, the fourth prediction matching result, and the target loss function, the network parameters of the first neural network are adjusted, and the first neural network with adjusted network parameters is used as the video matching model.

[0214] Optionally, the processor 601 is specifically used for:

[0215] The first prediction matching result is calibrated based on the business knowledge data included in the second neural network to obtain a third prediction matching result between the first object and the second video.

[0216] The second prediction matching result is calibrated based on the business knowledge data included in the second neural network to obtain a fourth prediction matching result between the second object and the second video.

[0217] Optionally, the processor 601 is specifically used for:

[0218] Based on the attribute feature set of the first object and the attribute feature set of the second video, obtain the attribute data tuples corresponding to the first object and the second video.

[0219] The first calibration parameter is determined based on the business knowledge data included in the second neural network and the attribute data tuples.

[0220] A third predictive matching result is determined between the first object and the second video based on the first calibration parameters and the first predictive matching result.

[0221] Optionally, the business knowledge data includes a first business rule set and a second business rule set, and the processor 601 is specifically used for:

[0222] For any attribute data tuple in the attribute data tuples corresponding to the first object and the second video, if the attribute data tuple matches the first business rule set, then the calibration component corresponding to the attribute data tuple is determined as the first component.

[0223] If any attribute data tuple matches the second business rule set, then the calibration component corresponding to any attribute data tuple is determined to be the second component.

[0224] If any attribute data tuple does not match either the first business rule set or the second business rule set, then the calibration component corresponding to any attribute data tuple is determined to be the third component.

[0225] The first calibration parameter is determined based on the calibration components corresponding to each attribute data tuple.

[0226] Optionally, the target loss function includes a first loss function and a second loss function, and the processor 601 is specifically used for:

[0227] Based on the first prediction matching result, the second prediction matching result, and the first loss function, a first loss value of the first neural network is determined.

[0228] Based on the first prediction matching result, the second prediction matching result, the third prediction matching result, the fourth prediction matching result, and the second loss function, the second loss value of the first neural network is determined.

[0229] The total loss value of the first neural network is determined based on the first loss value, the second loss value, and the balance parameter, and the network parameters of the first neural network are adjusted based on the total loss value.

[0230] In specific implementations, the processor 601, storage device 602, and network interface 603 described in the embodiments of this application can execute the embodiments of this application. Figure 2 , 3 The implementation methods described in the related embodiments of the provided video data processing method can also be implemented using the embodiments of this application. Figure 5 The implementation methods described in the relevant embodiments of the provided video data processing device will not be repeated here.

[0231] In the above embodiments, the descriptions of each embodiment have their own emphasis. Parts not described in detail in a certain embodiment can be referred to in the relevant descriptions of other embodiments. The technical solutions of this application, in essence, or the parts that contribute to the prior art, or all or part of the technical solutions, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which can be a computer, server, or network device, specifically a processor in the computer device) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium may include: a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM), and other media capable of storing program code.

[0232] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, it should be understood that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A video data processing method, characterized in that, The method includes: Acquire video feature data of a first video and object feature data of at least one object; the first video includes a cold start video. The video feature data of the first video and the object feature data of the at least one object are processed by calling a video matching model and reference feature data to obtain a matching result between the first video and each of the at least one object; wherein, the reference feature data is determined based on the video feature data of multiple second videos, and the second videos include videos with click data; Based on the matching result, a target object is determined from the at least one object, and the first video is sent to the terminal device corresponding to the target object; The training methods for the video matching model include: Obtain training samples, which include the second video, a first object in the object subset corresponding to the second video, and a second object in the object set other than the object subset. The first neural network is trained based on the training samples and the second neural network to obtain the video matching model; the second neural network is constructed based on the first neural network and business knowledge data; the first neural network serves as the student network, and the second neural network serves as the teacher network, with the first neural network and the second neural network sharing network parameters; The methods for generating the business knowledge data include: For each click data tuple in the video click record, which includes a third object and a third video, the attribute feature set of the third object is obtained from the object feature data of the third object, and the attribute feature set of the third video is obtained from the video feature data of the third video. The video click record includes at least one click data tuple. Based on the attribute feature set of the third object and the attribute feature set of the third video, at least one attribute data tuple is determined; Obtain the frequency corresponding to each attribute data tuple, and determine business knowledge data from the at least one attribute data tuple based on the frequency corresponding to each attribute data tuple.

2. The method according to claim 1, characterized in that, The video feature data includes identifier features and attribute feature sets. The process of calling the video matching model and referencing feature data to process the video feature data of the first video and the object feature data of the at least one object to obtain a matching result between the first video and each of the at least one object includes: The reference feature data is used as the identification feature in the video feature data of the first video, and the reference feature data is determined based on the identification features included in the video feature data of the plurality of second videos; The video matching model is invoked to process the video feature data of the first video and the object feature data of the at least one object to obtain the matching result between the first video and each of the at least one object.

3. The method according to claim 1 or 2, characterized in that, The method further includes: Retrieve multiple second videos from a video collection that contain click records; The identification features of each second video are obtained from the video feature data of each second video; The identification features of the multiple second videos are subjected to global average pooling to obtain the reference feature data.

4. The method according to claim 1, characterized in that, The method further includes: A predetermined number of second videos are determined from at least one second video included in the training samples; Replace the identifier features in the video feature data of the preset number of second videos with the reference feature data.

5. The method according to claim 1 or 4, characterized in that, The step of training the first neural network based on the training samples and the second neural network to obtain the video matching model includes: The first neural network is invoked to obtain a first prediction matching result between the first object and the second video, and a second prediction matching result between the second object and the second video; The third prediction matching result between the first object and the second video and the fourth prediction matching result between the second object and the second video are obtained based on the business knowledge data included in the second neural network. Based on the first prediction matching result, the second prediction matching result, the third prediction matching result, the fourth prediction matching result, and the target loss function, the network parameters of the first neural network are adjusted, and the first neural network with adjusted network parameters is used as the video matching model.

6. The method according to claim 5, characterized in that, The step of obtaining a third predictive matching result between the first object and the second video and a fourth predictive matching result between the second object and the second video based on business knowledge data included in the second neural network includes: The first prediction matching result is calibrated based on the business knowledge data included in the second neural network to obtain a third prediction matching result between the first object and the second video. The second prediction matching result is calibrated based on the business knowledge data included in the second neural network to obtain a fourth prediction matching result between the second object and the second video.

7. The method according to claim 6, characterized in that, The calibration process performed on the first prediction matching result based on the business knowledge data included in the second neural network to obtain a third prediction matching result between the first object and the second video includes: Based on the attribute feature set of the first object and the attribute feature set of the second video, obtain the attribute data tuples corresponding to the first object and the second video; Based on the business knowledge data included in the second neural network and the various attribute data tuples, the first calibration parameter is determined; A third predictive matching result is determined between the first object and the second video based on the first calibration parameters and the first predictive matching result.

8. The method according to claim 7, characterized in that, The business knowledge data includes a first business rule set and a second business rule set. The determination of the first calibration parameter based on the business knowledge data included in the second neural network and the various attribute data tuples includes: For any attribute data tuple in the attribute data tuples corresponding to the first object and the second video, if the attribute data tuple matches the first business rule set, then the calibration component corresponding to the attribute data tuple is determined to be the first component. If any attribute data tuple matches the second business rule set, then the calibration component corresponding to any attribute data tuple is determined to be the second component; If any attribute data tuple does not match either the first business rule set or the second business rule set, then the calibration component corresponding to any attribute data tuple is determined to be the third component. The first calibration parameter is determined based on the calibration components corresponding to each attribute data tuple.

9. The method according to claim 5, characterized in that, The target loss function includes a first loss function and a second loss function. Adjusting the network parameters of the first neural network based on the first prediction matching result, the second prediction matching result, the third prediction matching result, the fourth prediction matching result, and the target loss function includes: Based on the first prediction matching result, the second prediction matching result, and the first loss function, a first loss value of the first neural network is determined. Based on the first prediction matching result, the second prediction matching result, the third prediction matching result, the fourth prediction matching result, and the second loss function, a second loss value for the first neural network is determined; The total loss value of the first neural network is determined based on the first loss value, the second loss value, and the balance parameter, and the network parameters of the first neural network are adjusted based on the total loss value.

10. A video data processing apparatus, characterized in that, The device includes: The acquisition module is used to acquire video feature data of a first video and object feature data of at least one object; the first video includes a cold start video; The processing module is used to call a video matching model and reference feature data to process the video feature data of the first video and the object feature data of the at least one object, so as to obtain the matching result between the first video and each of the at least one object; wherein, the reference feature data is determined based on the video feature data of multiple second videos, and the second videos include videos with click data; The processing module is further configured to determine a target object from the at least one object based on the matching result; The sending module is used to send the first video to the terminal device corresponding to the target object; The training method of the video matching model includes: acquiring training samples, which include the second video, a first object in the object subset corresponding to the second video, and a second object in the object set other than the object subset; training the first neural network based on the training samples and the second neural network to obtain the video matching model; the second neural network is constructed based on the first neural network and business knowledge data; the first neural network serves as a student network, and the second neural network serves as a teacher network, with the first neural network and the second neural network sharing network parameters; The method for generating the business knowledge data includes: for each click data tuple in the video click record that includes a third object and a third video, obtaining the attribute feature set of the third object from the object feature data of the third object, and obtaining the attribute feature set of the third video from the video feature data of the third video, wherein the video click record includes at least one click data tuple; determining at least one attribute data tuple based on the attribute feature set of the third object and the attribute feature set of the third video; obtaining the frequency corresponding to each attribute data tuple, and determining business knowledge data from the at least one attribute data tuple based on the frequency corresponding to each attribute data tuple.

11. A computer device, characterized in that, The computer device includes a processor, a network interface, and a storage device, which are interconnected. The network interface is controlled by the processor to send and receive data. The storage device is used to store a computer program, which includes program instructions. The processor is configured to invoke the program instructions to execute the video data processing method according to any one of claims 1 to 9.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, the computer program including program instructions, which are executed by a processor to perform the video data processing method according to any one of claims 1 to 9.

13. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a computer processor, it implements the video data processing method according to any one of claims 1 to 9.

Images