Training method and device of multimedia resource recommendation model

Abstract

The invention discloses a training method and device for a multimedia resource recommendation model, and the method comprises the steps: obtaining a sample historical multimedia sequence corresponding to a sample object, and obtaining a sample search word and a sample multimedia resource corresponding to the sample search word; according to the sample search word features, screening out a first sample historical sub-sequence from the sample historical multimedia sequence; screening out a second sample historical sub-sequence from the first sample historical sub-sequence according to a sample multi-modal feature of the sample multimedia resource; based on a preset multimedia resource recommendation model, performing multimedia parameter prediction processing on the second sample historical sub-sequence, the sample search word and the sample multimedia resource to obtain a sample prediction multimedia parameter; and training a preset multimedia resource recommendation model based on the sample multimedia parameter tag and the sample prediction multimedia parameter to obtain a target multimedia resource recommendation model. According to the target multimedia resource recommendation model, the accuracy of multimedia resource recommendation can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of information recommendation, and in particular to a multimedia resource recommendation model training method and device. BACKGROUND

[0002] In a recommendation system, user historical behavior has guiding significance for predicting click rate and other parameters. In common e-commerce scenarios, video search scenarios, recommendation scenarios, etc., user historical sequences are introduced so as to fully mine user interest points and better meet user search or recommendation requirements.

[0003] However, in a traditional recommendation system, the number of historical multimedia resources browsed by a user is usually large, user historical sequence modeling is difficult, and usually the focus is on short-term interest, so the long-term interest representation of the user is insufficient, and there are many noises in the user historical sequence, so the prediction effect of the multimedia resource recommendation model obtained by training is not good.

[0004] Therefore, how to improve the accuracy of multimedia resource recommendation is a technical problem to be solved. SUMMARY

[0005] The present application provides a multimedia resource recommendation model training method and device, which can improve the accuracy of multimedia resource recommendation.

[0006] In one aspect, the present application provides a multimedia resource recommendation model training method, which comprises:

[0007] obtaining a sample historical multimedia sequence corresponding to a sample object, and obtaining a sample search term and a sample multimedia resource corresponding to the sample search term; the sample multimedia resource is labeled with a sample multimedia parameter label; the sample historical multimedia sequence includes historical multimedia resources browsed by the sample object at a first time; the sample multimedia resource is a recall multimedia resource corresponding to a sample search term input by the sample object at a second time; the first time is earlier than the second time;

[0008] According to the sample search term feature, a first historical multimedia resource matching the sample search term feature is selected from the sample historical multimedia sequence to obtain a first sample historical sub-sequence; the sample search term feature is a semantic feature corresponding to the sample search term;

[0009] According to the sample multi-modal feature of the sample multimedia resource, a second historical multimedia resource matching the sample multi-modal feature is selected from the first sample historical sub-sequence to obtain a second sample historical sub-sequence;

[0010] Based on the preset multimedia resource recommendation model, the second sample history sub-sequence, the sample search term and the sample multimedia resource are subjected to multimedia parameter prediction processing to obtain sample predicted multimedia parameters corresponding to the sample multimedia resource.

[0011] Based on the sample multimedia parameter label and the sample predicted multimedia parameter, the preset multimedia resource recommendation model is trained to obtain a target multimedia resource recommendation model.

[0012] In another aspect, the present application provides a multimedia resource recommendation method, the method comprising:

[0013] Obtaining a history multimedia sequence corresponding to a target object, and obtaining a target search term and candidate multimedia resources corresponding to the target search term; the history multimedia sequence comprises history multimedia resources browsed by the target object at a first time; the candidate multimedia resources are recall multimedia resources corresponding to the target search term input by the target object at a second time; the first time is earlier than the second time;

[0014] According to the target search term feature, a first history multimedia resource matching the target search term feature is screened from the history multimedia sequence to obtain a first history sub-sequence; the target search term feature is a semantic feature corresponding to the target search term;

[0015] According to the target multi-modal feature of the candidate multimedia resource, a second history multimedia resource matching the target multi-modal feature is screened from the first history sub-sequence to obtain a second history sub-sequence;

[0016] The second history sub-sequence, the target search term, and the candidate multimedia resource are input into a target multimedia resource recommendation model for multimedia parameter prediction processing to obtain target predicted multimedia parameters of the candidate multimedia resource; the target multimedia resource recommendation model is trained according to the training method of the multimedia resource recommendation model as described above;

[0017] Based on the target predicted multimedia parameters, a recommendation order of the candidate multimedia resources is determined, and the candidate multimedia resources are recommended to the target object in the recommendation order.

[0018] In another aspect, the present application provides a training device of a multimedia resource recommendation model, the device comprising:

[0019] The acquisition module is configured to acquire a sample historical multimedia sequence corresponding to a sample object, and acquire a sample search word and a sample multimedia resource corresponding to the sample search word; the sample multimedia resource is labeled with a sample multimedia parameter tag; the sample historical multimedia sequence includes historical multimedia resources browsed by the sample object at a first time; the sample multimedia resource is a recall multimedia resource corresponding to a sample search word input by the sample object at a second time; and the first time is earlier than the second time;

[0020] The first screening module is configured to screen, according to a sample search word feature, a first historical multimedia resource matching the sample search word feature from the sample historical multimedia sequence to obtain a first sample historical sub-sequence; the sample search word feature is a semantic feature corresponding to the sample search word.

[0021] The second screening module is configured to screen, according to a sample multi-modal feature of the sample multimedia resource, a second historical multimedia resource matching the sample multi-modal feature from the first sample historical sub-sequence to obtain a second sample historical sub-sequence.

[0022] The parameter prediction module is configured to perform multimedia parameter prediction processing on the second sample historical sub-sequence, the sample search word and the sample multimedia resource based on a preset multimedia resource recommendation model to obtain a sample predicted multimedia parameter corresponding to the sample multimedia resource.

[0023] The model training module is configured to train the preset multimedia resource recommendation model based on the sample multimedia parameter tag and the sample predicted multimedia parameter to obtain a target multimedia resource recommendation model.

[0024] In another aspect, the present application provides a multimedia resource recommendation device, which comprises:

[0025] The acquisition module is configured to acquire a sample historical multimedia sequence corresponding to a sample object, and acquire a sample search word and a sample multimedia resource corresponding to the sample search word; the sample multimedia resource is labeled with a sample multimedia parameter tag; the sample historical multimedia sequence includes historical multimedia resources browsed by the sample object at a first time; the sample multimedia resource is a recall multimedia resource corresponding to a sample search word input by the sample object at a second time; and the first time is earlier than the second time;

[0026] The first screening module is configured to screen, according to a sample search word feature, a first historical multimedia resource matching the sample search word feature from the sample historical multimedia sequence to obtain a first sample historical sub-sequence; the sample search word feature is a semantic feature corresponding to the sample search word.

[0027] The second screening module is configured to screen a second historical multimedia resource matching the target multi-modal feature from the first historical sub-sequence according to the target multi-modal feature of the candidate multimedia resource, to obtain a second historical sub-sequence.

[0028] The model processing module is configured to input the second historical sub-sequence, the target search term, and the candidate multimedia resource into a target multimedia resource recommendation model for multimedia parameter prediction processing, to obtain a target predicted multimedia parameter of the candidate multimedia resource.

[0029] The recommendation sequence determination module is configured to determine a recommendation sequence of the candidate multimedia resource based on the target predicted multimedia parameter, and recommend the candidate multimedia resource to the target object in the recommendation sequence.

[0030] In another aspect, an electronic device is provided, which includes a processor and a memory, the memory storing at least one instruction or at least one program, the at least one instruction or the at least one program being loaded and executed by the processor to implement the method as described above.

[0031] In another aspect, a computer storage medium is provided, which stores at least one instruction or at least one program, the at least one instruction or the at least one program being loaded and executed by a processor to implement the method as described above.

[0032] In another aspect, a computer program product or a computer program is provided, which includes computer instructions stored in a computer readable storage medium. A processor of a computer device reads the computer instructions from the computer readable storage medium, and the processor executes the computer instructions, so that the computer device performs to implement the method as described above.

[0033] The training method of the multimedia resource recommendation model provided in the application has the following technical effects:

[0034] The embodiment of the present application introduces the historical multimedia sequence of the sample object to represent the long-term interest of the sample object when training the preset multimedia resource recommendation model, so that the model can learn the personalized interest of the sample object. Based on the two-stage screening, the original historical multimedia sequence is first screened in the first stage based on the relevance of the search word and the historical multimedia resource, and then the first-stage screening result is screened in the second stage based on the relevance of the candidate multimedia resource and the first-stage screening result, to obtain the second sample historical sub-sequence, thereby denoising the historical multimedia sequence representing the long-term interest of the user, improving the accuracy of the sub-sequence, which is conducive to realizing more personalized resource content recommendation, and also reducing the resources required for sequence representation. Then, the preset multimedia resource recommendation model is trained based on the second sample historical sub-sequence obtained by the two-stage screening. Compared with the original historical multimedia sequence, the second sample historical sub-sequence has stronger implicit modeling capability, so that the target multimedia resource recommendation model obtained by training can more accurately learn the relationship between the search word, the candidate multimedia resource and the long-term interest of the sample object, so that the target multimedia resource recommendation model can output more accurate prediction results for different users, thereby improving the accuracy of multimedia resource recommendation. BRIEF DESCRIPTION OF DRAWINGS

[0035] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, and the advantages thereof, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0036] Figure 1 is a schematic diagram of an application environment of a multimedia resource recommendation model provided by an embodiment of the present application.

[0037] Figure 2 is a flowchart of a training method of a multimedia resource recommendation model provided by an embodiment of the present application.

[0038] Figure 3 is a flowchart of determining a first sample historical sub-sequence provided by an embodiment of the present application.

[0039] Figure 4 is a flowchart of determining a second sample historical sub-sequence provided by an embodiment of the present application.

[0040] Figure 5 is a flowchart of sample multimedia resource parameter prediction processing provided by an embodiment of the present application.

[0041] Figure 6is a structural schematic diagram of a preset multimedia resource recommendation model provided by the present application.

[0042] Figure 7 is a schematic diagram of a first service scenario and a second service scenario provided by an embodiment of the present application.

[0043] Figure 8 is a flowchart of a process of training a preset multimedia resource recommendation model provided by an embodiment of the present application.

[0044] Figure 9 is a flowchart of a process of recommending a multimedia resource provided by an embodiment of the present application.

[0045] Figure 10 is a scenario schematic diagram of personalized search of a multimedia resource provided by an embodiment of the present application.

[0046] Figure 11 is an architecture schematic diagram of a process of recommending a multimedia resource provided by an embodiment of the present application.

[0047] Figure 12 is a structural schematic diagram of a training device of a multimedia resource recommendation model provided by an embodiment of the present application.

[0048] Figure 13 is a structural schematic diagram of a multimedia resource recommendation device provided by an embodiment of the present application.

[0049] Figure 14 is a hardware structure block diagram of a server of a training method of a multimedia resource recommendation model provided by an embodiment of the present application. DETAILED DESCRIPTION

[0050] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0051] It can be understood that in the specific embodiments of the present application, data related to user information is involved. When the above embodiments of the present application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of related data need to comply with relevant laws, regulations and standards of relevant countries and regions.

[0052] It should be noted that the terms "first", "second" and the like in the description and in the claims of the present application and the above-described accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or server including a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0053] In the embodiments of the present application, the term "module" or "unit" refers to a computer program or a part of a computer program with a predetermined function, and works together with other related parts to achieve a predetermined target, and can be implemented entirely or partially by using software, hardware (such as a processing circuit or a memory) or a combination thereof. Similarly, one processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be a part of an integral module or unit that includes the functions of the module or unit.

[0054] Please refer to Figure 1 , Figure 1 is a schematic diagram of an application environment of a multimedia resource recommendation model provided by an embodiment of the present application. The application environment can at least include a server 01 and a terminal 02.

[0055] In an optional embodiment, the server 01 can be used for model training and model inference. The server 01 can be a stand-alone physical server, a server cluster composed of multiple physical servers or a distributed system, or a cloud server providing cloud computing services.

[0056] In an optional embodiment, the terminal 02 can be used for sequentially displaying multimedia resources and the like according to the recommendation order of the multimedia resources for a user. Specifically, the terminal 02 can include but is not limited to electronic devices such as smart phones, desktop computers, tablet computers, notebook computers, smart speakers, digital assistants, augmented reality (AR) / virtual reality (VR) devices, smart wearable devices, vehicle-mounted terminals, smart televisions, and the like. It can also be a software running on the above electronic devices, such as an application program, an applet, and the like. The operating system running on the electronic device in the embodiments of the present application can include but is not limited to an Android system, an IOS system, linux, windows, and the like.

[0057] The following introduces a training method of a multimedia resource recommendation model of the application, Figure 2 is a flowchart of a training method of a multimedia resource recommendation model provided by an embodiment of the application. The specification provides method operation steps as described in the embodiments or flowcharts, but more or fewer operation steps can be included based on conventional or non-creative labor. The order of steps listed in the embodiments is only one of the many execution orders, and does not represent the only execution order. In actual system or server product execution, the method order shown in the embodiments or the drawings can be executed in sequence or in parallel (for example, in a parallel processor or multi-threaded processing environment). As Figure 2 indicated, the method can include:

[0058] S201: Obtain a sample historical multimedia sequence corresponding to a sample object, and obtain a sample search term and a sample multimedia resource corresponding to the sample search term.

[0059] In an embodiment of the application, the multimedia resource can be in the form of text, picture, video, etc. Specifically, the multimedia resource can also be an advertisement, a commodity, etc. according to the actual application scenario. The specific content of the multimedia resource is not limited in the application.

[0060] The embodiment of the application introduces a historical multimedia sequence to train a multimedia resource recommendation model, which can consider the viewing history of a user in multimedia resource retrieval, thereby constructing a multimedia resource recommendation model that meets the personalized needs of the user.

[0061] It should be noted that the sample multimedia resource is obtained by recall according to the sample search term, and the sample multimedia resource is usually multiple. The specific implementation manner of obtaining the sample multimedia resource by recall is not limited in the application, for example, it can be an inverted index manner, a semantic matching manner or other manners, etc. The embodiment of the application trains the multimedia resource recommendation model on the basis of the sample multimedia resource obtained in the recall stage, thereby ranking the multimedia resource obtained by recall according to the trained multimedia resource recommendation model, and realizing the recommendation of the multimedia resource.

[0062] Among them, the sample multimedia resource is labeled with a sample multimedia parameter label. The sample multimedia parameter label indicates whether the sample multimedia resource is clicked, whether it is converted, and the display duration, etc. The sample multimedia parameter label is used to construct a Click-Through-Rate (CTR) prediction task, a Conversion Rate (CVR) prediction task and a display duration prediction task, thereby training the multimedia resource recommendation model.

[0063] It should be understood that the sample historical multimedia sequence can represent the long-term historical interest of the sample object. The sample historical multimedia sequence includes a historical multimedia resource browsed by the sample object at a first time; the sample multimedia resource is a recall multimedia resource corresponding to a sample search term input by the sample object at a second time; and the first time is earlier than the second time. The first time can be a preset historical time period, for example, 14 days before the second time, and the sample historical multimedia sequence is a sequence formed by the multimedia resources browsed by the sample object in the 14 days before the second time. In addition, the historical multimedia sequence can also be a sequence formed by a preset number of multimedia resources browsed by the sample object before the second time. The preset number can be 2000. It should be noted that the specific value of the preset number can be adjusted according to actual needs.

[0064] It should be noted that when obtaining the sample historical multimedia sequence, the historical multimedia resources with a display duration less than a preset display threshold are first filtered. The display duration can be understood as the playing duration of a video or the staying duration of a text. The preset display threshold can be set according to actual needs, for example, a value such as 2.4 seconds. By filtering out historical multimedia resources with too small display duration, historical multimedia resources browsed in user mis-click, mis-click and other situations can be excluded, thereby improving the accuracy of the sample historical multimedia sequence representing the long-term historical interest of the sample object.

[0065] For the search scenario, since the sample historical multimedia sequence includes a large number of historical multimedia resources, there are many noises in the sample historical multimedia sequence. Therefore, when introducing the long-term historical interest of the user, the present application further screens the original sequence to obtain a sub-sequence related to the current search. Based on the uniqueness of the search scenario, in the embodiment of the present application, a two-stage screening method is used to improve the accuracy of the historical sequence.

[0066] S203: According to the sample search term feature, a first historical multimedia resource matching the sample search term feature is screened from the sample historical multimedia sequence to obtain a first sample historical sub-sequence.

[0067] The sample search term feature is a semantic feature corresponding to the sample search term.

[0068] This step is a one-stage screening process, which coarsely screens according to the sample search term feature to obtain the first sample historical sub-sequence.

[0069] In an implementation manner, the sample search word feature can be an embedding vector of the sample search word, which is obtained by inputting the sample search word into a query subnetwork in a double-tower semantic model for vectorization processing. The double-tower semantic model in the embodiment of the application is pre-trained, and the training process in the prior art can be referred to, which will not be described herein.

[0070] It should be noted that the sample search word feature corresponding to the sample search word can be stored in the feature library for calling.

[0071] S205: According to the sample multi-modal feature of the sample multimedia resource, a second historical multimedia resource matching the sample multi-modal feature is filtered from the first sample history subsequence, to obtain a second sample history subsequence.

[0072] This step is a two-stage filtering process. On the basis of the above-mentioned one-stage filtering, the sample multi-modal feature of the sample multimedia resource is filtered again to obtain the second sample history subsequence.

[0073] In an implementation manner, the sample multi-modal feature of the sample multimedia resource can be a feature obtained after the sample multimedia resource is subjected to semantic understanding. When the sample multimedia is a video, the sample multi-modal feature can be a feature formed by comprehensively considering text features, audio features, and video features. The sample multi-modal feature can be obtained by inputting the sample multimedia resource into a video understanding model for vectorization processing. The video understanding model can be a trained transformer model. The video understanding model in the embodiment of the application is pre-trained, and the training process in the prior art can be referred to, which will not be described herein.

[0074] It should be noted that the sample multi-modal feature of the sample multimedia resource can be stored in the feature library for calling.

[0075] S207: Based on a preset multimedia resource recommendation model, a multimedia parameter prediction processing is performed on the second sample history subsequence, the sample search word, and the sample multimedia resource, to obtain a sample predicted multimedia parameter corresponding to the sample multimedia resource.

[0076] After the filtering in the above steps, the second sample history subsequence is a sequence obtained after the original historical multimedia sequence is denoised, and can be used for training of the multimedia resource recommendation model.

[0077] S209: The preset multimedia resource recommendation model is trained based on the sample multimedia parameter label and the sample predicted multimedia parameter, to obtain a target multimedia resource recommendation model.

[0078] The embodiment of the present application introduces the historical multimedia sequence of the sample object to represent the long-term interest of the sample object when training the preset multimedia resource recommendation model, so that the model can learn the personalized interest of the sample object. Based on two-stage screening, the original historical multimedia sequence is first screened based on the relevance of the search term and the historical multimedia resource, and then based on the relevance of the candidate multimedia resource and the first-stage screening result, the first-stage screening result is screened again to obtain a second sample historical sub-sequence, thereby denoising the historical multimedia sequence representing the long-term interest of the user, improving the accuracy of the sub-sequence, which is conducive to realizing more personalized resource content recommendation, and also reducing the resources required for sequence representation. Then, the second sample historical sub-sequence obtained by the two-stage screening is used to train the preset multimedia resource recommendation model. Compared with the original historical multimedia sequence, the second sample historical sub-sequence has stronger implicit modeling capability, so that the target multimedia resource recommendation model learned by training can more accurately learn the relationship between the search term, the candidate multimedia resource and the long-term interest of the sample object, so that the target multimedia resource recommendation model can output more accurate prediction results for different users, thereby improving the accuracy of multimedia resource recommendation and realizing more personalized resource content recommendation.

[0079] Figure 3 is a flowchart provided by the embodiment of the present application for determining the first sample historical sub-sequence. Figure 3 It can be regarded as a specific example of step S203. In one embodiment, the first historical multimedia resource matching the sample search term feature is filtered from the sample historical multimedia sequence according to the sample search term feature to obtain a first sample historical sub-sequence, comprising:

[0080] S301: obtaining the semantic features of each historical multimedia resource in the sample historical multimedia sequence;

[0081] The semantic feature of the historical multimedia resource can be an embedding vector of the historical multimedia resource. Similar to the sample search term feature, the semantic feature of each historical multimedia resource can be obtained by vectorization processing in the doc sub-network in the double-tower semantic model in advance and stored in the feature library for calling.

[0082] S303: determining the first similarity of each historical multimedia resource according to the similarity of the semantic feature of each historical multimedia resource and the sample search term feature;

[0083] S305: screening the first historical multimedia resource from the historical multimedia resource according to the first similarity, and determining the sequence formed by the first historical multimedia resource as the first sample historical sub-sequence.

[0084] The first similarity of the historical multimedia resource can be obtained by calculating the cosine similarity between the semantic feature of the historical multimedia resource and the sample search term feature. Then, the historical multimedia resources are sorted from high to low according to the first similarity, and the first preset number of historical multimedia resources with the first similarity in front are retained and determined as the first historical multimedia resources, and then a sequence is formed according to the browsing order to obtain a first sample historical sub-sequence. The preset first number can be 50. It should be noted that the specific value of the preset first number can be set according to actual needs.

[0085] The existing technology basically models the long-term historical interest of the user in the recommendation scenario, that is, the user does not need to actively search to obtain the multimedia resource. In the existing technology, the screening of the historical sequence includes two ways of hard search and soft search. However, both of these two ways do not consider the characteristics of the relevance constraints unique to the search scenario. The embodiment of the present application first performs a one-stage screening, and coarsely screens the sample historical multimedia sequence according to the similarity between the semantic feature of the historical multimedia resource and the sample search term feature to obtain the first sample historical sub-sequence, which excludes irrelevant content to the sample search term. This makes the first sample historical sub-sequence be able to specifically represent the long-term interest of the user for the search term, avoiding the introduction of too many irrelevant factors in the subsequent training of the model. For example, when the search term is a cat, irrelevant historical multimedia resources in the historical multimedia sequence, such as news videos, can be excluded, because the news videos do not work for searching for related content about cats. This screening method based on the relevance between the search term and the multimedia resource can reduce the noise generated in the subsequent training of the model, thereby improving the accuracy of the model.

[0086] In one embodiment, the screening of the first historical multimedia resource matching the sample search term feature from the sample historical multimedia sequence to obtain the first sample historical sub-sequence according to the sample search term feature can further include: screening the first historical multimedia resource from the sample historical multimedia sequence to obtain an initial sample historical sub-sequence according to the sample search term feature; performing word segmentation on the sample search term to generate a query element; performing word hit and category matching processing on each first historical multimedia resource in the initial sample historical sub-sequence based on the query element to obtain a matching score of each first historical multimedia resource; and determining a sequence formed by the first historical multimedia resource with a matching score greater than a preset matching threshold as the first sample historical sub-sequence.

[0087] After obtaining the first historical multimedia resources in the above embodiments, the sequence formed by the first historical multimedia resources is recorded as the initial sample historical subsequence according to the browsing order. Then, the initial sample historical subsequence is filtered again using word matching and category matching to obtain the first sample historical subsequence. Specifically, the search term can be segmented to obtain the query element term, and then the category information or title information of each first historical multimedia resource in the initial sample historical subsequence can be traversed to determine whether the term matches the category information or title information of the first historical multimedia resource. The matching score can then be set according to the matching frequency. In one example, the numerical range of the matching score can be 0 to 1, and the preset matching threshold can be set to 0.5. It should be understood that the numerical range of the matching score and the specific value of the preset matching threshold can be set according to actual needs, and this application does not impose specific limitations on them. The above method, based on the filtering of the relevance between the search term and the multimedia resource, further filters by category matching, making the first sample historical subsequence more targeted and accurate for the sample search term.

[0088] It should be noted that the above-mentioned word matching and category matching processing can also be performed directly on each historical multimedia resource in the sample historical multimedia sequence to obtain the matching score of each historical multimedia resource; the sequence formed by historical multimedia resources with a matching score greater than the preset matching threshold is determined as the first sample historical subsequence, thereby improving the generalization ability of the first-stage screening and enhancing the generalization effect of the sequence.

[0089] Based on the first-stage screening described above, this application embodiment performs a second-stage screening. Figure 4 This is a flowchart illustrating the process of determining the second sample historical subsequence provided in an embodiment of this application. Figure 4 This can be seen as a specific example of step S205. In one embodiment, the step of filtering out a second historical multimedia resource that matches the multimodal features of the sample multimedia resource from the first historical subsequence to obtain the second historical subsequence includes:

[0090] S401: Obtain the multimodal features of each first historical multimedia resource in the first sample historical subsequence;

[0091] The multimodal features of the first historical multimedia resources can be understood as the features obtained after semantic understanding of the first historical multimedia resources. Similar to the multimodal features of the samples, the multimodal features of each first historical multimedia resource can be pre-vectorized using a video understanding model and stored in a feature library for later retrieval.

[0092] S403: Determine the second similarity of each first historical multimedia resource based on the similarity between the multimodal features of each first historical multimedia resource and the multimodal features of the sample;

[0093] S405: Based on the second similarity, select the second historical multimedia resource from the first historical multimedia resource;

[0094] The second similarity of the first historical multimedia resources can be obtained by calculating the cosine similarity between the multimodal features of the first historical multimedia resource and the multimodal features of the sample. Then, the first historical multimedia resources are sorted from highest to lowest according to their second similarity, and a predetermined second number of first historical multimedia resources with the highest second similarity are retained and designated as the second historical multimedia resources. The predetermined second number can be 30; it should be noted that the specific value of the predetermined second number can be set according to actual needs.

[0095] S407: The second historical multimedia resource whose similarity to the multimodal features of the sample meets a preset threshold is taken as the target historical multimedia resource, and the sequence formed by the target historical multimedia resource is determined as the second sample historical subsequence.

[0096] Next, the second historical multimedia resources with an absolute value of the second similarity greater than a preset threshold are retained and identified as target historical multimedia resources. These are then arranged in the browsing order to form a sequence, resulting in the second sample historical subsequence. The preset threshold can be 0.7. It should be noted that the preset threshold can be set according to actual needs.

[0097] In this embodiment, a two-stage screening process is performed. First, the first historical sample subsequence is coarsely screened based on the similarity between the multimodal features of the first historical multimedia resource and the multimodal features of the sample. This process excludes content irrelevant to the sample multimodal resources in the resulting second historical multimedia resource. Then, a second-stage fine screening is performed based on similarity, enabling the second historical sample subsequence to more accurately represent the user's long-term interests related to the sample multimedia resources. Furthermore, since the second historical sample subsequence is obtained by screening based on the first historical sample subsequence, it excludes a large amount of content unrelated to the sample search terms and the sample multimedia resources. This two-stage sequence screening method provided in this embodiment can denoise the original sample historical multimedia sequence, improving the accuracy of the historical subsequence and enhancing the implicit modeling capability of the sequence.

[0098] The purpose of the above embodiments is to select historical subsequences of samples for training, which can then be used as input for training the model. Figure 5This is a schematic flowchart of the sample multimedia resource parameter prediction processing provided in this application embodiment. The step of performing multimedia parameter prediction processing on the second sample historical subsequence, the sample search term, and the sample multimedia resource based on a preset multimedia resource recommendation model to obtain the sample predicted multimedia parameters corresponding to the sample multimedia resource includes steps S501 to S507. Steps S501 to S505 can be considered as a preprocessing procedure for the sample multimedia resource to obtain the input to the preset multimedia resource recommendation model. Specifically, Figure 5 include:

[0099] S501: Obtain the sample sequence attribute information corresponding to the second sample historical subsequence;

[0100] The sample sequence attribute information consists of the attribute information of each target historical multimedia resource in the second sample historical subsequence. The sample sequence attribute information may include one or more of the following: the ID number of the target historical multimedia resource, the title segmentation of the target historical multimedia resource, the cover optical character recognition (OCR) segmentation, and the creator information of the target historical multimedia resource.

[0101] S503: Determine the sample relevance information between the sample search term and the sample multimedia resource; the sample relevance information represents the degree of matching between the sample multimedia resource and the sample search term;

[0102] In one implementation, sample relevance information can be determined by the matching score of the query element "term" of the sample search term in the category or title information of the sample multimedia resource. In another implementation, the similarity between the semantic features of the sample search term and the sample multimedia resource can also be used to determine sample relevance information.

[0103] S505: Determine sample sequence matching features based on the similarity between the sample multimodal features and the second sample historical subsequence; the sample sequence matching features characterize the degree of matching between the second sample historical subsequence and the sample multimedia resource;

[0104] In one implementation, step S505 may include steps S5051 to S5053.

[0105] S5051: Arrange the historical multimedia resources in the sample's historical multimedia sequence in the order of browsing, and determine the last historical multimedia resource as the preset multimedia resource; the preset multimedia resource is understood as the multimedia resource most recently viewed by the sample object. When the multimedia resource is a video, the preset multimedia resource is the video last played by the sample object in the sample's historical multimedia sequence. It should be noted that the preset multimedia resource is used to represent the short-term interests of the sample object, and its quantity can be 1.

[0106] S5053: The similarity between the preset multimodal features of the preset multimedia resources and the sample multimodal features is determined as the preset similarity;

[0107] The preset similarity can be the cosine similarity between the preset multimedia resources and the multimodal features corresponding to each sample multimedia resource. The multimodal features of the preset multimedia resources can be pre-vectorized using a video understanding model and stored in a feature library for later retrieval.

[0108] S5055: Based on the similarity between each target multimedia resource in the second sample historical subsequence and the multimodal features of the sample, determine the mean similarity, maximum similarity, and similarity distribution results;

[0109] The similarity between each target multimedia resource and the multimodal features of the sample in the second sample historical subsequence is the second similarity between the multimodal features of the target multimedia resources and the multimodal features of the sample. The average of the second similarities corresponding to each target multimedia resource is taken to obtain the mean similarity; the maximum value of the second similarity for each target multimedia resource is determined as the maximum similarity; the second similarities of each target multimedia resource are then binned, and the number of second similarities in each bin is counted to determine the similarity distribution.

[0110] S5057: Based on the preset similarity, the mean similarity, the maximum similarity, and the similarity distribution result, determine the sample sequence matching features.

[0111] In one implementation, the preset similarity, mean similarity, maximum similarity, and similarity distribution results can be processed by vector mapping, and the resulting embedded vectors can be concatenated to obtain the sample sequence matching features.

[0112] It should be understood that the sample sequence matching features include both the recent interest features of the sample object and the explicit matching features between the second sample historical subsequence and the sample multimedia resources. Compared to the traditional approach of learning the implicit matching features between the second sample historical subsequence and the sample multimedia resources through the fully connected layers of the model, the embodiments of this application construct explicit matching features between the second sample historical subsequence and the sample multimedia resources, thereby enhancing the interaction between the second sample historical subsequence and the sample multimedia resources and increasing the effectiveness of the second sample historical subsequence.

[0113] S507: Input the sample sequence attribute information, the sample object attribute information, the sample relevance information, the sample search term features, the sample multimodal features, and the sample sequence matching features into the preset multimedia resource recommendation model to perform multimedia parameter prediction processing, and obtain the sample predicted multimedia parameters.

[0114] The attribute information of a sample object may include one or more of the following: the sample object's account ID, account name, etc.

[0115] In the above embodiments, by inputting data from multiple data domains into a preset multimedia resource recommendation model for multimedia parameter prediction, the model can learn the features of different data domains and perform feature cross-referencing, making the sample prediction of multimedia parameters more accurate.

[0116] Figure 6 This is a schematic diagram of the structure of the preset multimedia resource recommendation model provided in this application. For example... Figure 6 As shown, the preset multimedia resource recommendation model includes a first feature extraction subnetwork and a second feature extraction subnetwork. The step of inputting the sample sequence attribute information, the sample object attribute information, the sample relevance information, the sample search term features, the sample multimodal features, and the sample sequence matching features into the preset multimedia resource recommendation model for multimedia parameter prediction processing to obtain the sample predicted multimedia parameters includes:

[0117] The sample sequence attribute information, the sample object attribute information, and the sample relevance information are input into the first feature extraction subnetwork for sparse feature extraction processing to obtain the first feature; the sample search term features, the sample multimodal features, and the sample sequence matching features are input into the second feature extraction subnetwork for dense feature extraction processing to obtain the second feature; the first feature and the second feature are concatenated to obtain the sample parameter features of the sample multimedia resource; and the sample predicted multimedia parameters are obtained based on the sample parameter features.

[0118] Since the sample sequence attribute information, sample object attribute information, and sample relevance information are discrete sparse features, the first feature extraction sub-network is used to perform sparse feature extraction processing on the above information to obtain the first feature. Since the sample search term features, sample multimodal features, and sample sequence matching features are continuous dense features, these features can be concatenated, and then the second feature extraction sub-network is used to perform sparse feature extraction processing on the concatenated features to obtain the second feature. For example... Figure 6 As shown, the second feature subnetwork can be a multilayer perceptron. Then, the first and second features are concatenated to obtain the sample parameter features used for prediction.

[0119] By distinguishing the above input content into dense features and sparse features, and processing them separately using the first feature extraction sub-network and the second feature extraction sub-network, the representation ability of the features is improved, making the feature representation more accurate.

[0120] like Figure 6 As shown, the first feature sub-network may include an input representation layer and a task feature extraction layer. The input representation layer is used to embed the information organized by data domain, including sample sequence attribute information, sample object attribute information, and sample correlation information, into vectorized form. The resulting features are then concatenated and used as task features input into the task feature combination layer for feature combination and cross-referencing to obtain the first feature.

[0121] In this embodiment, for sample sequence attribute information, the input representation layer includes a first input sub-network, which embeds different attributes separately and then uses sum pooling to obtain the embedding vector representations of different attributes. These are then concatenated and processed by a multilayer perceptron (MLP) for feature extraction to obtain the first task sub-features. For sample object attribute information, the input representation layer includes a second input sub-network, which embeds different attributes separately and then uses a factorization machine (FM) to obtain the second task sub-features. For sample correlation information, the input representation layer includes a third input sub-network, which embeds the sample correlation information and then uses a factorization machine to obtain the third task sub-features. Finally, the first, second, and third task sub-features are concatenated to obtain the task features.

[0122] It should be noted that the first input sub-network can also be based on an attention mechanism, including attention based on search terms and candidate multimedia resources as keys, to determine the first task sub-features, improve the adaptive weight learning ability of historical sub-sequences, and thus improve the feature representation ability of the first task sub-features.

[0123] Multimedia resource search scenarios are diverse. In this embodiment, the search scenarios for multimedia resources are further broken down to more accurately predict multimedia parameters for different scenarios, thereby achieving more efficient and accurate recommendation of multimedia resources to users. The training process of the preset multimedia resource recommendation model of this application will be explained below using video as an example.

[0124] In this embodiment of the application, the search scenario is divided into a first business scenario and a second business scenario. Figure 7 This is a schematic diagram illustrating the first and second business scenarios provided in this application embodiment. In the first business scenario, sample multimedia resources are displayed as thumbnails and can be played automatically. The second business scenario is the page displayed after the sample multimedia resources are triggered on the first business scenario page. In the second business scenario, the sample multimedia resources are displayed in a swipeable, switchable manner. Figure 7 As shown in (A), after a user submits a search term, the search results are displayed as a two-column list page, showing video covers. When the dwell time meets the condition, the video in the middle of the page automatically plays. Additionally, when a user clicks on one of the videos, a second business scenario is triggered. This second scenario is a single-column display of the clicked video's playback page, allowing the user to immerse themselves in watching the video. Figure 7 As shown in (B), assuming that after a user clicks on video A, they enter the playback page of video A, and in the second business scenario, they can switch to watch other videos by swiping.

[0125] Depend on Figure 7 It is evident that the behavioral patterns in the first and second business scenarios differ significantly. In the first business scenario, multimedia resources can be clicked and viewed; in the second business scenario, multimedia resources can only be viewed. Therefore, to address the differences in behavioral patterns across different scenarios, this embodiment employs a multi-scenario, multi-task structure when constructing the preset multimedia resource recommendation model, establishing corresponding learning tasks based on the click patterns and display duration of sample multimedia resources.

[0126] It should be noted that, in order to predict multimedia parameters under multiple scenarios and tasks, in one implementation, the constructed multimedia resource recommendation model can be as follows: Figure 6As shown, an MMOE (Multi-gate Mixture-of-Experts) structure is introduced into the task feature extraction layer. The core idea of ​​MMOE is to replace the shared underlying network with an expert network. Through multiple gated networks, different expert network weights are learned on multiple expert networks for each task, and then fused to obtain the first feature. After concatenating the first feature with the second feature, the sample parameter features of the sample multimedia resources are obtained. Based on these sample parameter features, a multi-task prediction network is used to predict the multimedia parameters corresponding to each task.

[0127] The sample multimedia resources are labeled with sample multimedia parameter tags. The tag acquisition process in this embodiment includes: acquiring a sample duration tag for the sample multimedia resource, where the sample duration tag represents the display duration of the sample multimedia resource; when the multimedia resource is a video, the display duration is the playback duration; in a first business scenario, acquiring a first sample tag for the sample multimedia resource; the first sample tag represents the click parameter of the sample multimedia resource in the first business scenario; in a second business scenario, acquiring a second sample tag for the sample multimedia resource; the second sample tag indicates whether the display duration of the sample multimedia resource in the second business scenario is greater than a preset duration; the preset duration can be 7 seconds. It should be noted that this application does not limit the specific value of the preset duration and can determine it according to actual needs. That is, when the display duration of the sample multimedia resource after being clicked is greater than or equal to 7 seconds, the sample multimedia resource can be used as a positive sample in the second business scenario, and a corresponding second sample tag can be set. Then, the sample duration tag, the first sample tag, and the second sample tag are determined as the sample multimedia parameter tags of the sample multimedia resource.

[0128] In constructing the preset multimedia resource recommendation model, this application adopts a multi-scenario, multi-task structure. Accordingly, the multimedia parameter labels of the samples are also labeled in multiple scenarios and multiple tasks, enabling the trained multimedia resource recommendation model to simultaneously achieve multi-scenario and multi-task prediction. By utilizing the correlation between different scenarios and different tasks, the sample sparsity problem is reduced, and it is more convenient to obtain the predicted multimedia parameters without training multiple models, thus reducing the processing resources required by the model.

[0129] Correspondingly, the sample predicted multimedia parameters output by the preset multimedia resource recommendation model include a first sample predicted parameter, a second sample predicted parameter, and a sample predicted display duration; the first sample predicted parameter represents the click parameter of the sample multimedia resource predicted in the first business scenario, and the second sample predicted parameter represents the prediction result of whether the display duration of the sample multimedia resource predicted in the second business scenario is greater than the preset duration.

[0130] Figure 8 This is a flowchart illustrating the training of a preset multimedia resource recommendation model provided in an embodiment of this application. In one embodiment, training the preset multimedia resource recommendation model based on the sample multimedia parameter labels and the sample predicted multimedia parameters to obtain a target multimedia resource recommendation model includes:

[0131] S801: Determine the duration loss value based on the difference between the predicted display duration of the sample and the duration label of the sample;

[0132] Since the multimedia resources in the sample will have a display duration parameter label in both the first and second business scenarios, after obtaining the predicted display duration and the sample duration label, the actual display duration in the predicted display duration and the sample duration label are truncated and the logarithm is taken to compress the data scale and make the data more stable. Then, a regression task is established to learn the multimedia parameter of display duration.

[0133] In one implementation, the display progress of the sample multimedia resource can also be used as the sample duration label. When the multimedia resource is a video, the ratio of the playback duration to the total video duration can be used as the sample duration label to eliminate the influence of different total video durations on duration prediction. Correspondingly, the sample predicted multimedia parameters include the sample predicted display progress, and then the duration loss value is determined based on the difference between the sample predicted display progress and the sample duration label.

[0134] S803: Determine the first scene loss based on the difference between the first sample prediction parameters and the first sample label;

[0135] The first sample label includes a first sample click result label and a second sample click result label. The first sample click result label indicates whether the sample multimedia resource was clicked in the first business scenario, thus establishing a click-through rate (CTR) prediction task. If clicked, the sample multimedia resource is used as a positive sample for the CTR prediction task, and a corresponding first sample click result label is set. The second sample click result label indicates whether the click was valid, i.e., whether the display duration of the sample multimedia resource in the first business scenario exceeded a preset duration, thus establishing a conversion rate prediction task. If the display duration exceeded the preset duration, the sample multimedia resource is used as a positive sample for the conversion rate prediction task, and a corresponding second sample click result label is set.

[0136] Correspondingly, the first sample prediction parameters include the first predicted click result and the second predicted click result;

[0137] In one implementation, determining the first scenario loss based on the difference between the first sample prediction parameters and the first sample label may include: determining a first click loss based on the difference between the first predicted click result and the first sample click result label; determining a second click loss based on the difference between the second predicted click result and the second sample click result label; and determining the first scenario loss based on the first click loss and the second click loss. The first business scenario includes two tasks: a CTR task and a CVR task. The first click loss is the loss function of the CTR task, and the second click loss is the loss function of the CVR task. To maintain the dependency between the click-through rate prediction and conversion rate prediction tasks on user behavior, in one implementation, an ESMM structure can also be used to model click-through rate and conversion rate in the same sample space. The CVR model shares embeddings with the CTR model to alleviate the problem of CVR sample sparsity. Simultaneously, the sample selection bias problem is alleviated by modeling the relationship between click-to-conversion rate (CTCVR) and CTR. Correspondingly, the first scenario loss may include the loss function of the CTR task and the loss function of the CTCVR task.

[0138] S805: Determine the second scene loss based on the difference between the second sample prediction parameters and the second sample label;

[0139] In the second business scenario, since multimedia resources with a display duration exceeding the preset duration are used as positive samples for CTR, after obtaining the prediction parameters and labels of the second samples, the cross-entropy loss function can be used to calculate the loss for the second scenario based on the difference between the two. It should be understood that other loss functions can also be used to represent the loss for the second scenario according to actual needs.

[0140] S807: Determine the comprehensive loss value based on the first scene loss, the second scene loss, and the duration loss;

[0141] In one implementation, the first scene loss, the second scene loss, and the duration loss can be added together to obtain the comprehensive loss value. In another implementation, the first scene loss, the second scene loss, and the duration loss can be weighted and summed to obtain the comprehensive loss value. The weights can be set according to the corresponding task priorities of the first scene loss, the second scene loss, and the duration loss, or can be adaptively adjusted during model training.

[0142] S809: Train the preset multimedia resource recommendation model based on the comprehensive loss value until the training conditions are met, and determine the preset multimedia resource recommendation model at the end of training as the target multimedia resource recommendation model.

[0143] As can be seen from the technical solutions provided in the embodiments of this application above, the embodiments of this application introduce historical multimedia sequences of sample objects to represent the long-term interests of sample objects when training a preset multimedia resource recommendation model, enabling the model to learn the personalized interests of sample objects. This application uses a two-stage screening process to denoise the historical multimedia sequences representing users' long-term interests, improving the accuracy of subsequences and facilitating more personalized resource content recommendations, while also reducing the resources required for sequence representation. Compared to the original historical multimedia sequences, the second sample historical subsequence has stronger implicit modeling capabilities; therefore, the trained target multimedia resource recommendation model can more accurately learn the relationship between search terms, candidate multimedia resources, and the long-term interests of sample objects. Furthermore, this application utilizes multimodal vector representation matching to establish explicit matching features between historical sequences and candidate multimedia resources, improving the explicit modeling capability of historical sequence features. Simultaneously, this application distinguishes between the first and second business scenarios and constructs a multi-scenario, multi-task preset multimedia resource recommendation model. Therefore, the trained target multimedia resource recommendation model can output more accurate prediction results for different users, thereby improving the accuracy of multimedia resource recommendations.

[0144] Figure 9 This is a flowchart illustrating the method for recommending multimedia resources provided in this application, including:

[0145] S901: Obtain the historical multimedia sequence corresponding to the target object, and obtain the target search term and the candidate multimedia resources corresponding to the target search term;

[0146] Among them, the historical multimedia sequence includes the historical multimedia resources viewed by the target object at the first time; the candidate multimedia resources are the recalled multimedia resources corresponding to the target search terms entered by the target object at the second time; the first time is earlier than the second time;

[0147] For a detailed description of the implementation of this step, please refer to step S201, which will not be repeated here.

[0148] S903: Based on the target search term features, filter out the first historical multimedia resources that match the target search term features from the historical multimedia sequence to obtain the first historical subsequence;

[0149] Among them, the target search term features are the semantic features corresponding to the target search term;

[0150] For a detailed description of the implementation of this step, please refer to step S203, which will not be repeated here.

[0151] S905: Based on the target multimodal features of the candidate multimedia resources, select the second historical multimedia resources that match the target multimodal features from the first historical subsequence to obtain the second historical subsequence;

[0152] For a detailed implementation of this step, please refer to step S205, which will not be repeated here.

[0153] S907: Input the second historical subsequence, the target search term, and the candidate multimedia resources into the target multimedia resource recommendation model to perform multimedia parameter prediction processing, and obtain the target predicted multimedia parameters of the candidate multimedia resources;

[0154] The target multimedia resource recommendation model is trained according to the training method of the multimedia resource recommendation model in the above embodiment.

[0155] The target predicted multimedia parameters for candidate multimedia resources include predicted display duration, a first prediction parameter, and a second prediction parameter. The first prediction parameter is the click parameter of the candidate multimedia resource predicted in the first business scenario, including a first predicted click result (whether it will be clicked) and a second predicted click result (whether the display duration after clicking is greater than the preset duration). The second prediction parameter is whether the display duration of the candidate multimedia resource predicted in the second business scenario is greater than the preset duration.

[0156] S909: Based on the target predicted multimedia parameters, determine the recommendation order of the candidate multimedia resources, and recommend the candidate multimedia resources to the target object in sequence according to the recommendation order.

[0157] Based on the aforementioned target prediction multimedia parameters, the recommendation order of candidate multimedia resources can be determined according to different scenarios, meeting users' diverse multimedia resource search needs and making content recommendation more reasonable.

[0158] Figure 10 This is a schematic diagram of a scenario for personalized search of multimedia resources provided in an embodiment of this application.

[0159] like Figure 10 As shown in (A), in response to a click on the video display control on the page, the page enters as shown in (A). Figure 10The video display page is shown in (B) above. In the video display page, a click on the search control leads to a video search page. After detecting a search term entered in the search bar control and responding to a click on the search confirmation control, the search results corresponding to the search term are displayed. Traditional solutions often score multimedia resources based on factors such as relevance, click-through rate, and timeliness to objectively rank and recall them. However, this general search method is usually designed to meet the search needs of most people, and when recommending multimedia resources in the search results to users, it often ignores the personalized needs of users, resulting in a poor search experience. The target multimedia resource recommendation model provided in this application can achieve personalized search and improve recommendation accuracy. In one example, such as... Figure 10 As shown in (C), when a user's search term is "national team," the traditional search method would only recommend information related to national sports teams. However, by analyzing the user's browsing history, it was found that the user watched recommended videos related to comics titled "national team" for an extended period. Therefore, it can be inferred that the user's intention in this search was to find videos related to "national team" comics, rather than information related to national sports teams in the traditional sense. Therefore, when recommending multimedia resources from the search results, videos related to "national team" comics should be prioritized to meet personalized needs.

[0160] Figure 11 This is a schematic diagram of the architecture of the multimedia resource recommendation method provided in the embodiments of this application. For example... Figure 11As shown, taking video as an example of multimedia resources, this embodiment utilizes a General Search Unit (GSU) module to perform historical sequence filtering and feature calculation in practical applications. Specifically, after obtaining the user's target search term (Query), the Query Parser module can be called to segment and rewrite the target search term to obtain QP information, which is used for video retrieval. Simultaneously, the GSU module is asynchronously requested. The GSU module calls the session storage module to obtain the user's historical sequence (historical documents), then queries the embedding features of the historical documents in the FKV database, and filters out the top k historical documents with the highest similarity to the Query, forming the first historical subsequence. Based on the QP information, the video proxy module synchronously calls the GSU module. The GSU module uses the QP information to match and score the k historical documents in the first historical subsequence, and queries the attribute features of the k historical documents in the FKV database, storing them in the cache. The GSU module can also filter the k historical documents again based on the matching score, using the filtering result as the first-stage filtering result. The video proxy module calls the video recall module to retrieve multiple candidate videos (target documents) and passes them to the video ranking module for processing. The video ranking module requests the GSU module, which queries the FKV database for the embedding features of the target documents and matches each target document with similar historical sequences, sending these sequences to the video ranking module. Based on the target documents, similar historical sequences, and other features used for prediction, the video ranking module calls the deployed target multimedia resource recommendation model for online inference, thereby ranking the candidate videos and displaying the recommended candidate videos to the client according to the ranking results. Simultaneously, the target documents, similar historical sequences, and other features used for prediction are saved to the HBase database via a message middleware. After receiving the recommended content, the client collects user behavior data through event tracking, forming search behavior logs, which are then reported to the message middleware. The sample concatenation module is responsible for concatenating the features in the HBase database with the corresponding user behavior, defining positive and negative samples, and performing format conversion to form streaming samples. The model receives the samples, performs training and updates, and then deploys the updated model online.

[0161] Figure 12 This is a schematic diagram of the training device structure for the multimedia resource recommendation model provided in this application embodiment. The device 1200 may include:

[0162] The acquisition module 1201 is used to acquire the sample historical multimedia sequence corresponding to the sample object, and to acquire the sample search term and the sample multimedia resource corresponding to the sample search term; the sample multimedia resource is labeled with sample multimedia parameter tags; the sample historical multimedia sequence includes the historical multimedia resources browsed by the sample object at a first time; the sample multimedia resource is the recalled multimedia resource corresponding to the sample search term entered by the sample object at a second time; the first time is earlier than the second time;

[0163] The first filtering module 1202 is used to filter out first historical multimedia resources that match the sample search term features from the sample historical multimedia sequence to obtain a first sample historical subsequence; the sample search term features are the semantic features corresponding to the sample search term.

[0164] The second filtering module 1203 is used to filter out a second historical multimedia resource that matches the multimodal features of the sample multimedia resource from the first sample historical subsequence, thereby obtaining a second sample historical subsequence.

[0165] The parameter prediction module 1204 is used to perform multimedia parameter prediction processing on the second sample historical subsequence, the sample search term and the sample multimedia resource based on a preset multimedia resource recommendation model, so as to obtain the sample predicted multimedia parameters corresponding to the sample multimedia resource.

[0166] The model training module 1205 is used to train the preset multimedia resource recommendation model based on the sample multimedia parameter labels and the sample predicted multimedia parameters to obtain the target multimedia resource recommendation model.

[0167] In some embodiments, the first filtering module may include:

[0168] The semantic feature acquisition submodule is used to acquire the semantic features of each historical multimedia resource in the historical multimedia sequence of the sample;

[0169] The first similarity determination submodule is used to determine the first similarity of each historical multimedia resource based on the similarity between the semantic features of each historical multimedia resource and the features of the sample search terms.

[0170] The first historical multimedia resource determination submodule is used to filter out the first historical multimedia resource from the historical multimedia resources according to the first similarity, and determine the sequence formed by the first historical multimedia resource as the first sample historical subsequence.

[0171] In some embodiments, the second filtering module may include:

[0172] The multimodal feature acquisition submodule is used to acquire the multimodal features of each first historical multimedia resource in the first sample historical subsequence;

[0173] The second similarity determination submodule is used to determine the second similarity of each first historical multimedia resource based on the similarity between the multimodal features of each first historical multimedia resource and the multimodal features of the sample.

[0174] The second historical multimedia resource determination submodule is used to filter out the second historical multimedia resource from the first historical multimedia resource according to the second similarity.

[0175] The second sample historical subsequence determination submodule is used to select the second historical multimedia resource whose similarity to the multimodal features of the sample meets a preset threshold as the target historical multimedia resource, and to determine the sequence formed by the target historical multimedia resource as the second sample historical subsequence.

[0176] In some embodiments, the parameter prediction module may include:

[0177] The sample sequence attribute information acquisition submodule is used to acquire the sample sequence attribute information corresponding to the second sample historical subsequence;

[0178] The sample relevance information determination submodule is used to determine the sample relevance information between the sample search term and the sample multimedia resource; the sample relevance information represents the degree of matching between the sample multimedia resource and the sample search term;

[0179] The sample sequence matching feature determination submodule is used to determine sample sequence matching features based on the similarity between the sample multimodal features and the second sample historical subsequence; the sample sequence matching features characterize the degree of matching between the second sample historical subsequence and the sample multimedia resource;

[0180] The multimedia parameter prediction submodule is used to input the sample sequence attribute information, the sample object attribute information, the sample relevance information, the sample search term features, the sample multimodal features, and the sample sequence matching features into the preset multimedia resource recommendation model to perform multimedia parameter prediction processing and obtain the sample predicted multimedia parameters.

[0181] In some embodiments, the preset multimedia resource recommendation model includes a first feature extraction subnetwork and a second feature extraction subnetwork; the multimedia parameter prediction submodule may include:

[0182] A sparse feature extraction processing unit is used to input the sample sequence attribute information, the sample object attribute information and the sample correlation information into the first feature extraction sub-network for sparse feature extraction processing to obtain the first feature;

[0183] The dense feature extraction processing unit is used to input the sample search term features, the sample multimodal features, and the sample sequence matching features into the second feature extraction sub-network for dense feature extraction processing to obtain the second feature;

[0184] The sample parameter feature determination unit is used to concatenate the first feature and the second feature to obtain the sample parameter features of the sample multimedia resource.

[0185] The sample prediction multimedia parameter determination unit is used to obtain the sample prediction multimedia parameters based on the sample parameter characteristics.

[0186] In some embodiments, the sample sequence matching feature determination submodule may include:

[0187] A preset multimedia resource determination unit is used to arrange each historical multimedia resource in the sample historical multimedia sequence in the order of browsing, and determine the last historical multimedia resource as the preset multimedia resource.

[0188] A preset similarity determination unit is used to determine the similarity between the preset multimodal features of the preset multimedia resource and the sample multimodal features as a preset similarity.

[0189] The similarity feature determination unit is used to determine the mean similarity, maximum similarity, and similarity distribution results based on the similarity between each target multimedia resource in the second sample historical subsequence and the multimodal features of the sample.

[0190] The sample sequence matching feature determination unit is used to determine the sample sequence matching features based on the preset similarity, the mean similarity, the maximum similarity, and the similarity distribution result.

[0191] In some embodiments, the device 1200 may further include:

[0192] The sample duration tag acquisition module is used to acquire the sample duration tag of the sample multimedia resource, wherein the sample duration tag represents the display duration of the sample multimedia resource;

[0193] The first sample tag acquisition module is used to acquire a first sample tag of the sample multimedia resource in a first business scenario; the first sample tag represents the click parameters of the sample multimedia resource in the first business scenario; in the first business scenario, the sample multimedia resource is displayed in the form of a thumbnail;

[0194] The second sample tag acquisition module is used to acquire a second sample tag of the sample multimedia resource in a second business scenario. The second sample tag indicates whether the display duration of the sample multimedia resource in the second business scenario is greater than a preset duration. The second business scenario is the page displayed after the sample multimedia resource is triggered in the first business scenario page. In the second business scenario, the sample multimedia resource is displayed in a swipeable switching format.

[0195] The sample multimedia parameter tag determination module is used to determine the sample duration tag, the first sample tag, and the second sample tag as the sample multimedia parameter tags of the sample multimedia resource.

[0196] In some embodiments, the sample prediction multimedia parameters include a first sample prediction parameter, a second sample prediction parameter, and a sample prediction display duration; the first sample prediction parameter represents the click parameter of the sample multimedia resource predicted in the first business scenario, and the second sample prediction parameter represents the prediction result of whether the display duration of the sample multimedia resource predicted in the second business scenario is greater than the preset duration.

[0197] The model training module may include:

[0198] The duration loss determination submodule is used to determine the duration loss value based on the difference between the predicted display duration of the sample and the duration label of the sample;

[0199] The first scene loss determination submodule is used to determine the first scene loss based on the difference between the first sample prediction parameters and the first sample label;

[0200] The second scene loss determination submodule is used to determine the second scene loss based on the difference between the second sample prediction parameters and the second sample label;

[0201] The comprehensive loss determination submodule is used to determine the comprehensive loss value based on the first scene loss, the second scene loss, and the duration loss;

[0202] The training submodule is used to train the preset multimedia resource recommendation model based on the comprehensive loss value until the training conditions are met, and to determine the preset multimedia resource recommendation model at the end of training as the target multimedia resource recommendation model.

[0203] In some embodiments, the first sample label includes a first sample click result label and a second sample click result label; the first sample click result label indicates whether the sample multimedia resource is clicked in the first business scenario; the second sample click result label indicates whether the display duration of the sample multimedia resource in the first business scenario is greater than the preset duration; the first sample prediction parameter includes a first predicted click result and a second predicted click result; the first scenario loss determination submodule may include:

[0204] The first click loss determination unit is used to determine the first click loss based on the difference between the first predicted click result and the first sample click result label;

[0205] The second click loss determination unit is used to determine the second click loss based on the difference between the second predicted click result and the label of the second sample click result;

[0206] The first scene loss determination unit is used to determine the first scene loss based on the first click loss and the second click loss.

[0207] In some embodiments, the first filtering module may include:

[0208] The initial sample history subsequence determination submodule is used to filter out the first historical multimedia resource from the sample history multimedia sequence based on the features of the sample search terms, so as to obtain the initial sample history subsequence;

[0209] The word segmentation submodule is used to segment the sample search terms and generate query elements;

[0210] The matching score determination submodule is used to perform word matching and category matching processing on each first historical multimedia resource in the initial sample historical subsequence based on the query element, so as to obtain the matching score of each first historical multimedia resource.

[0211] The first sample historical subsequence determination submodule is used to determine the sequence formed by the first historical multimedia resources with a matching score greater than a preset matching threshold as the first sample historical subsequence.

[0212] Figure 13 This is a schematic diagram of the structure of the multimedia resource recommendation device provided in the embodiments of this application. The device 1300 may include:

[0213] The acquisition module 1301 is used to acquire the historical multimedia sequence corresponding to the target object, and to acquire the target search term and the candidate multimedia resources corresponding to the target search term; the historical multimedia sequence includes the historical multimedia resources browsed by the target object at a first time; the candidate multimedia resources are the recalled multimedia resources corresponding to the target search term entered by the target object at a second time; the first time is earlier than the second time;

[0214] The first filtering module 1302 is used to filter out first historical multimedia resources that match the target search term features from the historical multimedia sequence to obtain a first historical subsequence; the target search term features are the semantic features corresponding to the target search term.

[0215] The second filtering module 1303 is used to filter out second historical multimedia resources that match the target multimodal features from the first historical subsequence based on the target multimodal features of the candidate multimedia resources, thereby obtaining a second historical subsequence;

[0216] The model processing module 1304 is used to input the second historical subsequence, the target search term, and the candidate multimedia resources into the target multimedia resource recommendation model for multimedia parameter prediction processing to obtain the target predicted multimedia parameters of the candidate multimedia resources; the target multimedia resource recommendation model is trained according to the training method of the multimedia resource recommendation model described in the above embodiment;

[0217] The recommendation order determination module 1305 is used to determine the recommendation order of the candidate multimedia resources based on the target predicted multimedia parameters, and recommend the candidate multimedia resources to the target object in sequence according to the recommendation order.

[0218] The apparatus and method embodiments described herein are based on the same inventive concept.

[0219] This application provides an electronic device including a processor and a memory. The memory stores at least one instruction or at least one program, which is loaded and executed by the processor to implement the method provided in the above method embodiments.

[0220] Embodiments of this application also provide a computer storage medium, which can be disposed in a terminal to store at least one instruction or at least one program related to the implementation of the method embodiments. The at least one instruction or at least one program is loaded and executed by the processor to implement the method provided in the above-described method embodiments.

[0221] Embodiments of this application also provide a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the method provided in the above-described method embodiments.

[0222] Optionally, in this embodiment, the storage medium may be located at at least one of the multiple network servers in a computer network. Optionally, in this embodiment, the storage medium may include, but is not limited to, various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0223] The memory described in this application embodiment can be used to store software programs and modules. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, applications required for the functions, etc.; the data storage area may store data created according to the use of the device, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. Accordingly, the memory may also include a memory controller to provide the processor with access to the memory.

[0224] The training method for the multimedia resource recommendation model provided in this application can be executed on a mobile terminal, computer terminal, server, or similar computing device. Taking running on a server as an example, Figure 14 This is a hardware structure block diagram of the server for the training method of the multimedia resource recommendation model provided in this application embodiment. For example... Figure 14As shown, the server 1400 can vary significantly due to different configurations or performance. It may include one or more Central Processing Units (CPUs) 1410 (CPUs 1410 may include, but are not limited to, microprocessors (MCUs) or programmable logic devices (FPGAs), a memory 1430 for storing data, and one or more storage media 1420 (e.g., one or more mass storage devices) for storing application programs 1423 or data 1422. The memory 1430 and storage media 1420 may be temporary or persistent storage. The program stored in the storage media 1420 may include one or more modules, each module may include a series of instruction operations on the server. Furthermore, the CPU 1410 may be configured to communicate with the storage media 1420 and execute the series of instruction operations stored in the storage media 1420 on the server 1400. Server 1400 may also include one or more power supplies 1460, one or more wired or wireless network interfaces 1450, one or more input / output interfaces 1440, and / or one or more operating systems 1421, such as Windows Server™, Mac OS X™, Unix™, Linux™, FreeBSD™, etc.

[0225] The input / output interface 1440 can be used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the communication provider of server 1400. In one example, the input / output interface 1440 includes a network interface controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In one example, the input / output interface 1440 can be a radio frequency (RF) module for wireless communication with the Internet.

[0226] Those skilled in the art will understand that Figure 14 The structure shown is for illustrative purposes only and does not limit the structure of the aforementioned electronic device. For example, server 1400 may also include... Figure 14 The more or fewer components shown, or having the same Figure 14 The different configurations shown.

[0227] As can be seen from the embodiments of the training method, recommendation method, apparatus, device, or storage medium of the multimedia resource recommendation model provided in this application, the embodiments of this application introduce historical multimedia sequences of sample objects to represent the long-term interests of sample objects when training the preset multimedia resource recommendation model, so that the model can learn the personalized interests of sample objects. This application is based on a two-stage screening. First, the original historical multimedia sequences are screened based on the relevance between search terms and historical multimedia resources. Then, based on the results of the first-stage screening, the results of the first-stage screening are screened a second time based on the relevance between candidate multimedia resources and the results of the first-stage screening, resulting in a second sample historical subsequence. This denoises the historical multimedia sequences representing the user's long-term interests, improves the accuracy of the subsequence, facilitates more personalized resource content recommendations, and also reduces the resources required for sequence representation. Then, the preset multimedia resource recommendation model is trained based on the second sample historical subsequence obtained from the two-stage screening. Compared to the original historical multimedia sequence, the second sample historical subsequence has a stronger implicit modeling ability. Therefore, the target multimedia resource recommendation model trained can learn more accurately the relationship between search terms, candidate multimedia resources and the long-term interests of sample objects. This enables the target multimedia resource recommendation model to output more accurate prediction results for different users, thereby improving the accuracy of multimedia resource recommendation.

[0228] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, specific embodiments have been described above. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims can be performed in a different order than that shown in the embodiments and still achieve the desired result. Additionally, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0229] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, devices, and storage media are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0230] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer storage medium, such as a read-only memory, a disk, or an optical disk.

[0231] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for training a multimedia resource recommendation model, characterized in that, The method comprises: obtaining a sample historical multimedia sequence corresponding to a sample object, and obtaining a sample search word and a sample multimedia resource corresponding to the sample search word; the sample multimedia resource is labeled with a sample multimedia parameter tag; the sample historical multimedia sequence comprises historical multimedia resources browsed by the sample object at a first time; the sample multimedia resource is a recall multimedia resource corresponding to a sample search word input by the sample object at a second time; the first time is earlier than the second time; screening, according to a sample search word feature, a first historical multimedia resource matching the sample search word feature from the sample historical multimedia sequence to obtain a first sample historical sub-sequence; the sample search word feature is a semantic feature corresponding to the sample search word; screening, according to a sample multi-modal feature of the sample multimedia resource, a second historical multimedia resource matching the sample multi-modal feature from the first sample historical sub-sequence to obtain a second sample historical sub-sequence; performing multimedia parameter prediction processing on the second sample historical sub-sequence, the sample search word and the sample multimedia resource based on a preset multimedia resource recommendation model to obtain a sample predicted multimedia parameter corresponding to the sample multimedia resource; training the preset multimedia resource recommendation model based on the sample multimedia parameter tag and the sample predicted multimedia parameter to obtain a target multimedia resource recommendation model.

2. The method of claim 1, wherein, The method comprises: obtaining a sample historical multimedia sequence corresponding to a sample object, and obtaining a sample search word and a sample multimedia resource corresponding to the sample search word; the sample multimedia resource is labeled with a sample multimedia parameter tag; the sample historical multimedia sequence comprises historical multimedia resources browsed by the sample object at a first time; the sample multimedia resource is a recall multimedia resource corresponding to a sample search word input by the sample object at a second time; the first time is earlier than the second time; screening, according to a sample search word feature, a first historical multimedia resource matching the sample search word feature from the sample historical multimedia sequence to obtain a first sample historical sub-sequence; the sample search word feature is a semantic feature corresponding to the sample search word; screening, according to a sample multi-modal feature of the sample multimedia resource, a second historical multimedia resource matching the sample multi-modal feature from the first sample historical sub-sequence to obtain a second sample historical sub-sequence; 3. The method of claim 1, wherein, performing multimedia parameter prediction processing on the second sample historical sub-sequence, the sample search word and the sample multimedia resource based on a preset multimedia resource recommendation model to obtain a sample predicted multimedia parameter corresponding to the sample multimedia resource; training the preset multimedia resource recommendation model based on the sample multimedia parameter tag and the sample predicted multimedia parameter to obtain a target multimedia resource recommendation model. The method comprises: obtaining a sample historical multimedia sequence corresponding to a sample object, and obtaining a sample search word and a sample multimedia resource corresponding to the sample search word; the sample multimedia resource is labeled with a sample multimedia parameter tag; the sample historical multimedia sequence comprises historical multimedia resources browsed by the sample object at a first time; the sample multimedia resource is a recall multimedia resource corresponding to a sample search word input by the sample object at a second time; the first time is earlier than the second time; screening, according to a sample search word feature, a first historical multimedia resource matching the sample search word feature from the sample historical multimedia sequence to obtain a first sample historical sub-sequence; the sample search word feature is a semantic feature corresponding to the sample search word; screening, according to a sample multi-modal feature of the sample multimedia resource, a second historical multimedia resource matching the sample multi-modal feature from the first sample historical sub-sequence to obtain a second sample historical sub-sequence; performing multimedia parameter prediction processing on the second sample historical sub-sequence, the sample search word and the sample multimedia resource based on a preset multimedia resource recommendation model to obtain a sample predicted multimedia parameter corresponding to the sample multimedia resource; training the preset multimedia resource recommendation model based on the sample multimedia parameter tag and the sample predicted multimedia parameter to obtain a target multimedia resource recommendation model.

4. The method of claim 1, wherein, The preset multimedia resource recommendation model is used for multimedia parameter prediction processing on the second sample historical subsequence, the sample search word, and the sample multimedia resource to obtain a sample predicted multimedia parameter corresponding to the sample multimedia resource, including: obtaining sample sequence attribute information corresponding to the second sample historical subsequence; determining sample relevance information between the sample search word and the sample multimedia resource; the sample relevance information represents a matching degree between the sample multimedia resource and the sample search word; determining sample sequence matching features according to a similarity between the sample multimodal features and the second sample historical subsequence; the sample sequence matching features represent a matching degree between the second sample historical subsequence and the sample multimedia resource; inputting the sample sequence attribute information, the attribute information of the sample object, the sample relevance information, and the sample search word features, the sample multimodal features, and the sample sequence matching features into the preset multimedia resource recommendation model for multimedia parameter prediction processing to obtain the sample predicted multimedia parameter.

5. The method of claim 4, wherein, The preset multimedia resource recommendation model includes a first feature extraction subnetwork and a second feature extraction subnetwork; the sample sequence attribute information, the attribute information of the sample object, and the sample relevance information are input into the first feature extraction subnetwork for sparse feature extraction processing to obtain first features; the sample search word features, the sample multimodal features, and the sample sequence matching features are input into the second feature extraction subnetwork for dense feature extraction processing to obtain second features; the first features and the second features are spliced to obtain sample parameter features of the sample multimedia resource; and the sample predicted multimedia parameter is obtained according to the sample parameter features. The sample sequence matching features are determined according to a similarity between the sample multimodal features and the second sample historical subsequence, including: arranging each historical multimedia resource in the sample historical multimedia sequence in a browsing order, and determining a historical multimedia resource arranged at the end as a preset multimedia resource; determining a similarity between a preset multimodal feature of the preset multimedia resource and the sample multimodal features as a preset similarity; determining a similarity mean value, a similarity maximum value, and a similarity distribution result based on a similarity between each target multimedia resource in the second sample historical subsequence and the sample multimodal features; 6. The method of claim 4, wherein, determining the sample sequence matching features based on the preset similarity, the similarity mean value, the similarity maximum value, and the similarity distribution result. The method further includes: ​ ​ ​ 7. The method of claim 1, wherein, ​ obtaining a sample duration label of the sample multimedia resource, the sample duration label representing a display duration of the sample multimedia resource; in a first business scenario, obtaining a first sample label of the sample multimedia resource; the first sample label representing a click parameter of the sample multimedia resource in the first business scenario; in the first business scenario, the sample multimedia resource is displayed in the form of a thumbnail; in a second business scenario, obtaining a second sample label of the sample multimedia resource; the second sample label representing whether the display duration of the sample multimedia resource in the second business scenario is greater than a preset duration; the second business scenario is a page displayed after the sample multimedia resource is triggered in the first business scenario page, and in the second business scenario, the sample multimedia resource is displayed in the form of a slideable switch; determining the sample duration label, the first sample label and the second sample label as a sample multimedia parameter label of the sample multimedia resource.

8. The method of claim 7, wherein, The sample prediction multimedia parameter includes a first sample prediction parameter, a second sample prediction parameter and a sample prediction display duration; the first sample prediction parameter represents a click parameter of the sample multimedia resource predicted in the first business scenario, and the second sample prediction parameter represents a prediction result of whether the display duration of the sample multimedia resource predicted in the second business scenario is greater than the preset duration; The training of the preset multimedia resource recommendation model based on the sample multimedia parameter label and the sample prediction multimedia parameter to obtain a target multimedia resource recommendation model includes: determining a duration loss value based on the difference between the sample prediction display duration and the sample duration label; determining a first scene loss based on the difference between the first sample prediction parameter and the first sample label; determining a second scene loss based on the difference between the second sample prediction parameter and the second sample label; determining a comprehensive loss value based on the first scene loss, the second scene loss and the duration loss; training the preset multimedia resource recommendation model based on the comprehensive loss value until a training condition is met, and determining the preset multimedia resource recommendation model at the end of training as the target multimedia resource recommendation model.

9. The method of claim 8, wherein, The first sample label includes a first sample click result label and a second sample click result label; the first sample click result label represents whether the sample multimedia resource is clicked in the first business scenario; and the second sample click result label represents whether the display duration of the sample multimedia resource in the first business scenario is greater than the preset duration; The first sample prediction parameter includes a first prediction click result and a second prediction click result; determining a first scene loss based on the difference between the first sample prediction parameter and the first sample label includes: determining a first click loss based on the difference between the first prediction click result and the first sample click result label; determine a second click loss based on a difference between the second predicted click result and the second sample click result label; determine the first scene loss based on the first click loss and the second click loss.

10. The method of claim 1, wherein, The method comprises: obtaining a historical multimedia sequence corresponding to a target object, and obtaining a target search term and a candidate multimedia resource corresponding to the target search term; the historical multimedia sequence comprises historical multimedia resources browsed by the target object at a first time; the candidate multimedia resource is a recall multimedia resource corresponding to a target search term input by the target object at a second time; the first time is earlier than the second time; screening, from the historical multimedia sequence, a first historical multimedia resource matching a target search term feature, to obtain a first historical sub-sequence; the target search term feature is a semantic feature corresponding to the target search term; screening, from the first historical sub-sequence, a second historical multimedia resource matching a target multi-modal feature of the candidate multimedia resource, to obtain a second historical sub-sequence; the target multi-modal feature is a multi-modal feature of the candidate multimedia resource; inputting the second historical sub-sequence, the target search term, and the candidate multimedia resource into a target multimedia resource recommendation model to perform multimedia parameter prediction processing, to obtain a target predicted multimedia parameter of the candidate multimedia resource; the target multimedia resource recommendation model is trained according to the training method of the multimedia resource recommendation model in any one of claims 1-10; 11. A method of recommending multimedia resources, characterized by, determining a recommendation order of the candidate multimedia resource based on the target predicted multimedia parameter, and recommending the candidate multimedia resource to the target object in the recommendation order. The device comprises: an obtaining module configured to obtain a sample historical multimedia sequence corresponding to a sample object, and obtain a sample search term and a sample multimedia resource corresponding to the sample search term; the sample multimedia resource is labeled with a sample multimedia parameter label; the sample historical multimedia sequence comprises historical multimedia resources browsed by the sample object at a first time; the sample multimedia resource is a recall multimedia resource corresponding to a sample search term input by the sample object at a second time; the first time is earlier than the second time; ​ ​ ​ 12.A device for training a multimedia resource recommendation model, characterized in that, ​ ​ The first screening module is configured to screen, according to a sample search term feature, a first historical multimedia resource matching the sample search term feature from the sample historical multimedia sequence to obtain a first sample historical sub-sequence; the sample search term feature is a semantic feature corresponding to the sample search term; The second screening module is configured to screen, according to a sample multi-modal feature of the sample multimedia resource, a second historical multimedia resource matching the sample multi-modal feature from the first sample historical sub-sequence to obtain a second sample historical sub-sequence; The parameter prediction module is configured to perform multimedia parameter prediction processing on the second sample historical sub-sequence, the sample search term and the sample multimedia resource based on a preset multimedia resource recommendation model to obtain a sample predicted multimedia parameter corresponding to the sample multimedia resource; The model training module is configured to train the preset multimedia resource recommendation model based on the sample multimedia parameter label and the sample predicted multimedia parameter to obtain a target multimedia resource recommendation model.

13. A device for recommending a multimedia resource, characterized by, The device comprises: The acquisition module is configured to acquire a historical multimedia sequence corresponding to a target object and a target search term and a candidate multimedia resource corresponding to the target search term; the historical multimedia sequence comprises a historical multimedia resource browsed by the target object at a first time; the candidate multimedia resource is a recall multimedia resource corresponding to a target search term input by the target object at a second time; the first time is earlier than the second time; The first screening module is configured to screen, according to a target search term feature, a first historical multimedia resource matching the target search term feature from the historical multimedia sequence to obtain a first historical sub-sequence; the target search term feature is a semantic feature corresponding to the target search term; The second screening module is configured to screen, according to a target multi-modal feature of the candidate multimedia resource, a second historical multimedia resource matching the target multi-modal feature from the first historical sub-sequence to obtain a second historical sub-sequence; The model processing module is configured to input the second historical sub-sequence, the target search term and the candidate multimedia resource into a target multimedia resource recommendation model for multimedia parameter prediction processing to obtain a target predicted multimedia parameter of the candidate multimedia resource; the target multimedia resource recommendation model is trained according to the training method of the multimedia resource recommendation model in any one of claims 1-10; The recommendation order determination module is configured to determine a recommendation order of the candidate multimedia resource based on the target predicted multimedia parameter and recommend the candidate multimedia resource to the target object in the recommendation order.

14. An electronic device, comprising: The electronic device comprises a processor and a memory, the memory stores at least one instruction or at least one program, the at least one instruction or at least one program is loaded and executed by the processor to implement the training method of the multimedia resource recommendation model in any one of claims 1-10 or the recommendation method of the multimedia resource in claim 11.

15. A computer storage medium, comprising, The computer storage medium stores at least one instruction or at least one program, which is loaded and executed by the processor to implement the training method of the multimedia resource recommendation model according to any one of claims 1-10, or the recommendation method of the multimedia resource according to claim 11.

Images