Multimedia file recommendation method, recommendation model training method and corresponding device

Abstract

The embodiment of the invention discloses a multimedia file recommendation method, a recommendation model training method and a corresponding device, and the main technical scheme comprises the steps: obtaining a historical behavior sequence and a plurality of candidate multimedia files of a target user, the historical behavior sequence comprising more than one multimedia file having historical interaction with the target user; extracting causal characterization of the historical behavior sequence in at least one mode and causal characterization of each candidate multimedia file in at least one mode, wherein the causal characterization reflects the causal relationship between the candidate multimedia file and the target user preference; determining a user representation of the target user based on the causal representation of the historical behavior sequence, and determining a file representation based on the causal representation of each candidate multimedia file; and selecting at least one of the candidate multimedia files as a recommendation result according to the user representation and the file representation. Through causal representation learning, interference of potential hybrid factors on multimedia recommendation is removed, and recommendation accuracy and user experience are improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a multimedia file recommendation method, a method for training a recommendation model, and corresponding devices. Background Art

[0002] With the rapid development of the internet, multimedia files are experiencing explosive growth. For example, a large number of videos are uploaded every day. To improve user experience and efficiency, major service providers are committed to improving the accuracy of multimedia file recommendations. However, in some scenarios, the accuracy of multimedia file recommendations is difficult to guarantee. For example, after a new multimedia file is uploaded, during the cold start phase, the interaction data between the new multimedia file and the user is extremely sparse, making it difficult to fully learn its representation. This affects its effective recommendation and easily leads to recommendations to inappropriate users, resulting in poor recommendation accuracy. Summary of the Invention

[0003] In view of this, the present application provides a multimedia file recommendation method, a method for training a recommendation model, and corresponding devices to improve the accuracy of multimedia file recommendations.

[0004] This application provides the following solutions:

[0005] According to a first aspect, a method for recommending multimedia files is provided, the method comprising:

[0006] Acquire a historical behavior sequence of a target user and a plurality of candidate multimedia files, wherein the historical behavior sequence includes one or more multimedia files with which the target user has historical interaction behavior;

[0007] Extracting a causal representation of the historical behavior sequence in at least one modality and a causal representation of each candidate multimedia file in at least one modality, wherein the causal representation reflects a causal relationship between the causal representation and the target user's preference;

[0008] Obtaining a user representation of the target user based on the causal representation of the historical behavior sequence in at least one modality, and obtaining a file representation of each candidate multimedia file based on the causal representation of each candidate multimedia file in at least one modality;

[0009] Based on the user representation of the target user and the file representations of the candidate multimedia files, at least one candidate multimedia file is selected from the candidate multimedia files as a recommendation result for the target user.

[0010] According to an achievable manner in an embodiment of the present application, the method further includes: obtaining an attribute representation of the target user and an attribute representation of the historical behavior sequence;

[0011] Obtaining a user representation of the target user based on the causal representation of the historical behavior sequence in at least one modality includes:

[0012] A user representation of the target user is obtained based on the attribute representation of the target user, the attribute representation of the historical behavior sequence, and the causal representation of the historical behavior sequence in at least one modality.

[0013] According to an achievable manner in the embodiment of the present application, the method further includes: obtaining attribute representations of the candidate multimedia files;

[0014] Obtaining the file representation of each candidate multimedia file based on the causal representation of each candidate multimedia file in at least one modality includes:

[0015] Based on the attribute representation of each candidate multimedia file and the causal representation of each candidate multimedia file in at least one modality, a file representation of each candidate multimedia file is obtained.

[0016] According to an achievable method in an embodiment of the present application, obtaining a user representation of the target user based on the attribute representation of the target user, the attribute representation of the historical behavior sequence, and the causal representation of the historical behavior sequence in at least one modality includes:

[0017] Performing a first multi-head attention process using the attribute representation of the target user and the attribute representation of the historical behavior sequence to obtain a first representation of the target user;

[0018] Performing a second multi-head attention process using the attribute representation of the target user and the causal representation of the historical behavior sequence in at least one modality to obtain a second representation of the user;

[0019] The first representation and the second representation are concatenated to obtain a user representation of the target user.

[0020] According to an achievable method in an embodiment of the present application, obtaining the file representation of each candidate multimedia file based on the attribute representation of each candidate multimedia file and the causal representation of each candidate multimedia file in at least one modality includes:

[0021] For each candidate multimedia file, the following steps are performed: weighting processing is performed on the attribute representation of the candidate multimedia file and the causal representation of the candidate multimedia file in at least one modality, and the representations obtained after the weighted processing are spliced ​​to obtain the file representation of the candidate multimedia file; wherein the weight coefficient used in the weighted processing is obtained by the gated network using the heat representation of the candidate multimedia file.

[0022] According to an achievable method in an embodiment of the present application, extracting a causal representation of the historical behavior sequence in at least one modality includes:

[0023] Extracting, for each multimedia file included in the historical behavior sequence, an embedded representation of each multimedia file in at least one modality; inputting the embedded representation of each multimedia file in at least one modality into a causal representation learning module to obtain a causal representation of each multimedia file in at least one modality, wherein the causal representation of the historical behavior sequence in at least one modality is obtained from the causal representation of the multimedia files included in the historical behavior sequence in at least one modality;

[0024] Extracting the causal representation of the candidate multimedia file in at least one modality includes:

[0025] Extracting an embedded representation of the candidate multimedia file in at least one modality; inputting the embedded representation of the candidate multimedia file in at least one modality into the causal representation learning module to obtain a causal representation of the candidate multimedia file in at least one modality.

[0026] According to a second aspect, a method for training a recommendation model is provided, the method comprising:

[0027] Acquire training data comprising a plurality of sample pairs, the sample pairs comprising a positive sample pair and a negative sample pair, the positive sample pair comprising a user sample and a multimedia sample with which the user sample has a history of interaction, and the negative sample pair comprising the user sample and a multimedia sample with which the user sample has no history of interaction;

[0028] The recommendation model is trained using the training data, wherein a historical behavior sequence of the user sample in the sample pair and a multimedia sample in the sample pair are obtained, the historical behavior sequence including one or more multimedia files having historical interactive behaviors with the user sample; the recommendation model extracts a causal representation of the historical behavior sequence in at least one modality and a causal representation of each multimedia sample in at least one modality, the causal representation reflecting the causal relationship between the causal representation and the user preference; the recommendation model obtains a user representation of the user sample based on the causal representation of the historical behavior sequence in at least one modality, and obtains a file representation of the multimedia sample based on the causal representation of the multimedia sample in at least one modality; the recommendation model determines a preference score of the multimedia sample based on the user representation of the user sample and the file representation of the multimedia sample;

[0029] The training objective includes maximizing the difference between the preference scores of the multimedia samples included in the positive sample pair and the preference scores of the multimedia samples included in the negative sample pair in the sample pairs.

[0030] According to an achievable method in an embodiment of the present application, the recommendation model includes: a causal representation learning module, a user representation extraction module, a file representation extraction module and a recommendation module;

[0031] The causal representation learning module extracts a causal representation of the historical behavior sequence in at least one modality and a causal representation of each multimedia sample in at least one modality, wherein the causal representation reflects a causal relationship between the causal representation and the user preference;

[0032] The user representation extraction module obtains the user representation of the user sample based on the attribute representation of the user sample, the attribute representation of the historical behavior sequence, and the causal representation of the historical behavior sequence in at least one modality;

[0033] The file representation extraction module obtains the file representation of the multimedia sample based on the attribute representation of the multimedia sample and the causal representation of the multimedia sample in at least one modality;

[0034] The recommendation module determines a preference score of the multimedia sample based on the user representation of the user sample and the file representation of the multimedia sample.

[0035] According to an achievable method in an embodiment of the present application, the goal of the training also includes: maximizing the mutual information between the causal representations extracted by the causal relationship learning module for the multimedia files contained in the sample pairs under different modalities.

[0036] According to an achievable method in an embodiment of the present application, the causal representation learning module includes a shared representation learning module and a unique representation learning module;

[0037] The shared representation learning module is configured to learn a shared causal representation of the multimedia file in at least one modality based on the embedded representation of the multimedia file in at least one modality input to the causal representation learning module;

[0038] The unique representation learning module is configured to learn a unique causal representation of the input multimedia file in at least one modality based on the embedded representation of the input multimedia file in at least one modality;

[0039] The causal representation of the input multimedia file in at least one modality is obtained from the shared causal representation and the unique causal representation.

[0040] According to an achievable method in an embodiment of the present application, the goal of the training also includes: minimizing the distance between the shared causal representation and the unique causal representation extracted by the causal representation learning module for the multimedia files contained in the sample pair under the same modality, maximizing the mutual information between the shared causal representations extracted by the causal relationship learning module for the multimedia files contained in the sample pair under different modalities, and maximizing the mutual information between the shared causal representation and the unique causal representation extracted by the causal relationship learning module for the multimedia files contained in the sample pair under the same modality.

[0041] According to a third aspect, a multimedia file recommendation device is provided, the device comprising:

[0042] an acquiring unit configured to acquire a historical behavior sequence of a target user and a plurality of candidate multimedia files, wherein the historical behavior sequence includes one or more multimedia files with which the target user has historical interactive behaviors;

[0043] a representation extraction unit configured to extract a causal representation of the historical behavior sequence in at least one modality and a causal representation of each candidate multimedia file in at least one modality, wherein the causal representation reflects a causal relationship between the candidate multimedia file and the target user's preference;

[0044] a representation generating unit configured to obtain a user representation of the target user based on the causal representation of the historical behavior sequence in at least one modality, and to obtain a file representation of each candidate multimedia file based on the causal representation of each candidate multimedia file in at least one modality;

[0045] The recommendation unit is configured to select at least one candidate multimedia file from the candidate multimedia files as a recommendation result for the target user based on the user representation of the target user and the file representations of the candidate multimedia files.

[0046] According to a fourth aspect, a device for training a recommendation model is provided, the device comprising:

[0047] a data acquisition unit configured to acquire training data comprising a plurality of sample pairs, the sample pairs comprising a positive sample pair and a negative sample pair, the positive sample pair comprising a user sample and a multimedia sample with which the user sample has a history of interaction, and the negative sample pair comprising the user sample and a multimedia sample with which the user sample has no history of interaction;

[0048] A training unit is configured to train the recommendation model using the training data, wherein a historical behavior sequence of the user sample in the sample pair and a multimedia sample in the sample pair are obtained, the historical behavior sequence including one or more multimedia files having historical interactive behaviors with the user sample; the recommendation model extracts a causal representation of the historical behavior sequence in at least one modality and a causal representation of each multimedia sample in at least one modality, the causal representation reflecting the causal relationship between the causal representation and the user preference; the recommendation model obtains a user representation of the user sample based on the causal representation of the historical behavior sequence in at least one modality, and obtains a file representation of the multimedia sample based on the causal representation of the multimedia sample in at least one modality; the recommendation model determines a preference score of the multimedia sample based on the user representation of the user sample and the file representation of the multimedia sample;

[0049] The training objective includes maximizing the difference between the preference scores of the multimedia samples included in the positive sample pair and the preference scores of the multimedia samples included in the negative sample pair in the sample pairs.

[0050] According to a fifth aspect, a computer-readable storage medium is provided, on which a computer program is stored. When the program is executed by a processor, the steps of the method described in any one of the first and second aspects are implemented.

[0051] According to a sixth aspect, an electronic device is provided, comprising:

[0052] one or more processors; and

[0053] A memory associated with the one or more processors, the memory being used to store program instructions, wherein the program instructions, when read and executed by the one or more processors, execute the steps of the method described in any one of the first and second aspects above.

[0054] According to a seventh aspect, a computer program product is provided, comprising a computer program, which, when executed by a processor, implements the steps of the method described in any one of the first and second aspects.

[0055] According to the specific embodiments provided in this application, this application discloses the following technical effects:

[0056] 1) In the solution provided by the present application, a user representation is obtained by analyzing and processing the causal representation of the target user's historical behavior sequence in at least one modality, and a file representation is obtained by analyzing and processing the causal representation of each candidate multimedia file in at least one modality. Then, by comparing the user representation with the file representation, multimedia files that are highly consistent with the user's interests can be screened out for recommendation. On the one hand, the present application organically integrates causal relationship learning with multimodal learning, learns causal representation for at least one modality, and extracts modal features that are causally related to user preferences. This can effectively reduce the interference of confounding factors in multimedia files that are irrelevant to user preferences, so that the recommendation results can better reflect the user's true interest preferences and improve the accuracy of recommending multimedia files. On the other hand, the user representation of the target user is obtained by learning the historical behavior sequence composed of multimedia files that have historical interactive behaviors with the target user. Even if the interaction between the candidate multimedia files and the user is extremely sparse during the cold start process, the user's preferences can be reflected through the target user's historical behavior sequence, thereby more accurately achieving the recommendation of multimedia files.

[0057] 2) In the solution provided in this application, when determining the user representation of the target user and the file representation of each candidate multimedia file, the attribute representation and the causal representation under at least one modality are combined to drive the recommendation of multimedia files by combining explicit attributes and implicit causal relationships, thereby improving the accuracy of multimedia file recommendations.

[0058] 3) In the solution provided by this application, dual multi-head attention processing and representation splicing can be used to extract and fuse the user's attribute representation, the attribute representation of the historical behavior sequence, and the multimodal causal representation to generate a user representation. Among them, the first multi-head attention (user attribute + behavior attribute) can be used to extract the first representation representing long-term interests, and the second multi-head attention (user attribute + causal representation) can be used to extract the second representation representing real-time preferences. Splicing the first and second representations can generate a more accurate user representation, thereby fully exploring user preferences.

[0059] 4) The solution provided herein can implement weighted processing using weight coefficients assigned by a gating network, regulating the weights of the attribute representations of candidate multimedia files and the causal representations of the candidate multimedia files in at least one modality based on the file's popularity to generate a file representation. Dynamically calculating weight coefficients based on the gating network and weightedly fusing the file attribute representations with the causal representations can effectively highlight the characteristics of popular files, making the file representations more aligned with user interests and improving the rationality of recommendations.

[0060] 5) In the solution provided in this application, the causal representation learning module is used to mine causal relationships. The causal representation learning module can uniformly process the embedded representations of each multimedia file included in the historical behavior sequence and the candidate multimedia files in at least one modality, and then, based on the embedded representation, it can realize the causal representation of the historical behavior sequence in at least one modality and the extraction of the causal representation of each candidate multimedia file in at least one modality.

[0061] 6) This application also provides a method for training a recommendation model. This method can train a recommendation model based on training data of positive and negative sample pairs with the goal of maximizing the difference in preference scores. The recommendation model trained by this method can construct corresponding user representations and file representations based on the causal representations under at least one modality, and then determine the preference scores of multimedia samples. This method of organically combining causal relationship learning and multimodal learning can learn the factors in each modality that have a causal relationship with user preferences, and effectively separate the non-causal interference factors in each modality, so that the model can accurately recommend multimedia files preferred by users.

[0062] 7) In the solution provided in this application, the causal representation learning module can maximize the cross-modal causal representation mutual information through comparative learning, eliminate the deviation of each modal feature, so as to achieve semantic consistency of causal representation under multimodality, eliminate the recommendation deviation caused by semantic conflict between modalities, and thus improve the effect of recommendation based on the causal representation.

[0063] 8) The solution provided in this application can further coordinate multimodal shared causal representations and unique causal representations. Shared causal representations can be used to learn the latent semantic associations between different modalities and capture cross-modal consistency features, while unique causal representations can be used to learn information unique to each modality and capture key factors reflecting user preferences and content differences. This can further improve the accuracy of multimedia file recommendations and enhance the model's robustness and adaptability in complex application scenarios such as cold starts.

[0064] 9) In the solution provided by this application, the causal representation learning module can minimize the distance between the shared causal representation and the unique causal representation extracted under the same modality to constrain the shared causal representation and the unique causal representation to contain differentiated information. Furthermore, by maximizing the mutual information of cross-modal causal representations, the bias of each modal feature is eliminated. Furthermore, by maximizing the mutual information between the shared causal representation and the unique causal representation under the same modality, the potential semantic associations between modalities are effectively explored. Recommendations based on this causal representation can effectively improve the model's recommendation accuracy for multimedia files such as short videos in cold start scenarios.

[0065] Of course, any product implementing the present application does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0067] Figure 1 A system architecture diagram applicable to the embodiments of the present application;

[0068] Figure 2 A flowchart of a multimedia file recommendation method provided in an embodiment of the present application;

[0069] Figure 3 A schematic diagram of an architecture of a recommendation model provided in an embodiment of the present application;

[0070] Figure 4 Another system architecture diagram applicable to the embodiments of the present application;

[0071] Figure 5 A flowchart of a method for training a recommendation model provided in an embodiment of the present application;

[0072] Figure 6 A schematic diagram of a specific architecture of the recommendation model provided in an embodiment of the present application;

[0073] Figure 7 A schematic diagram of the relationship between the representations provided in the embodiments of the present application;

[0074] Figure 8 A schematic block diagram of a multimedia file recommendation device provided in an embodiment of the present application;

[0075] Figure 9 A schematic block diagram of an apparatus for training a recommendation model provided in an embodiment of the present application;

[0076] Figure 10 A schematic block diagram of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0077] The following will be combined with the accompanying drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field are within the scope of protection of this application.

[0078] The terms used in the embodiments of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The singular forms "a," "an," and "the" used in the embodiments of the present invention and the appended claims are also intended to include plural forms unless the context clearly indicates otherwise.

[0079] It should be understood that the term "and / or" as used herein is merely a description of the relationship between associated objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A exists alone, A and B exist simultaneously, or B exists alone. Furthermore, the character " / " in this document generally indicates that the associated objects are in an "or" relationship.

[0080] The word "if," as used herein, may be interpreted as "at the time of" or "when" or "in response to determining" or "in response to detecting," depending on the context. Similarly, the phrases "if it is determined" or "if (stated condition or event) is detected" may be interpreted as "when it is determined" or "in response to the determination" or "when detecting (stated condition or event)" or "in response to detecting (stated condition or event)," depending on the context.

[0081] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of relevant countries and regions, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0082] Currently, traditional multimedia file recommendation systems typically generate recommendations by analyzing the statistical correlation of historical interactions between users and candidate files. However, this statistical correlation-based approach is susceptible to confounding factors, such as file popularity fluctuations or accidental user clicks, making it difficult for recommendation models to accurately capture users' true interests. This is especially true in cold start scenarios, where interaction data for newly uploaded files is scarce, leading to a significant decrease in recommendation accuracy.

[0083] After long-term analysis and research, the inventors of this application found that multimedia files usually contain rich multimodal information, which can be used to enhance the representation learning of multimedia files. However, it is still easily interfered by confounding factors. The reason is that the existing technology cannot effectively distinguish the causal factors in multimodal information that are truly related to user preferences, thereby affecting the recommendation effect.

[0084] To address the problem in the above-mentioned existing methods of insufficiently modeling the causal relationship between user behavior and multimedia file features, which makes the recommendation system susceptible to false correlations, this application provides a new approach to extract causal features that can directly reflect user preferences from multimodal information, rather than relying solely on statistical correlations. This leads to the technical concept of extracting causal representations under multimodal conditions, eliminating the interference of non-causal factors, and establishing a stable association between user preferences and file features, thereby improving recommendation accuracy in cold start scenarios.

[0085] In order to facilitate the understanding of this application, the system architecture on which this application is based is first described. Figure 1 An exemplary system architecture to which the embodiments of the present application can be applied is shown. Figure 1 As shown in , the system architecture may include: user equipment, and a multimedia file recommendation device located at a multimedia file recommendation application server.

[0086] The user equipment and the server can communicate with each other. The user equipment and the server can be directly or indirectly connected through wired or wireless communication, which is not limited in this application.

[0087] User devices may include, but are not limited to, smart mobile terminals, smart home devices, wearable devices, and personal computers (PCs). Smart mobile devices may include mobile phones, tablets, laptops, PDAs (Personal Digital Assistants), and internet-connected cars. Smart home devices may include smart TVs and smart refrigerators. Wearable devices may include smart watches, smart glasses, virtual reality devices, augmented reality devices, and mixed reality devices (i.e., devices that support both virtual reality and augmented reality).

[0088] Multimedia files, including short videos, can be recommended by application servers. These servers can be standalone servers, server clusters, or even cloud servers. A cloud server, also known as a cloud computing server or cloud host, is a hosting product within the cloud computing service ecosystem. It addresses the management difficulties and limited scalability of traditional physical hosts and virtual private servers (VPS).

[0089] As one of the feasible methods, a user can trigger a multimedia file recommendation request by touching the user device, and the user device sends the multimedia file recommendation request to the multimedia file recommendation device on the server side through the network. The multimedia file recommendation device can obtain the user's corresponding historical behavior sequence and multiple candidate multimedia files, and adopt the method provided in the embodiment of the present application to select at least one candidate multimedia file from each candidate multimedia file as the recommendation result for the target user, and return the recommendation result to the user device through the network, and the user device displays the received recommendation result to the user.

[0090] Apart from Figure 1 In addition to the shown architecture, a computer terminal with strong computing power can also adopt the method provided in the embodiment of the present application to obtain information on the historical behavior sequence of the target user and information on multiple candidate multimedia files, and then select at least one candidate multimedia file from each candidate multimedia file as a recommendation result for the target user.

[0091] It should be understood that Figure 1 The user equipment and server end in the figure are only illustrative. According to the implementation requirements, there can be any number of user equipment and server ends.

[0092] Figure 2 This is a flowchart of a multimedia file recommendation method provided in an embodiment of the present application. The method can be performed by Figure 1 The multimedia file recommendation device in the system shown is executed. Figure 2 As shown in , the method may include the following steps:

[0093] Step 201: Obtain a target user's historical behavior sequence and multiple candidate multimedia files, where the historical behavior sequence includes one or more multimedia files with which the target user has historical interaction behavior.

[0094] Step 202: extracting a causal representation of the historical behavior sequence in at least one modality and a causal representation of each candidate multimedia file in at least one modality, wherein the causal representation reflects the causal relationship between the causal representation and the target user's preference;

[0095] Step 203: obtaining a user representation of the target user based on the causal representation of the historical behavior sequence in at least one modality, and obtaining a file representation of each candidate multimedia file based on the causal representation of each candidate multimedia file in at least one modality;

[0096] Step 204 : Based on the user representation of the target user and the file representations of the candidate multimedia files, at least one candidate multimedia file is selected from the candidate multimedia files as a recommendation result for the target user.

[0097] As can be seen from the above process, in the solution provided by this application, a user representation is obtained by analyzing and processing the causal representation of the target user's historical behavior sequence in at least one modality, and a file representation is obtained by analyzing and processing the causal representation of each candidate multimedia file in at least one modality. Then, by comparing the user representation with the file representation, multimedia files that are highly consistent with the user's interests can be screened out for recommendation. On the one hand, this application organically integrates causal relationship learning with multimodal learning, learns causal representation for at least one modality, and extracts modal features that are causally related to user preferences. This can effectively reduce the interference of confounding factors in multimedia files that are irrelevant to user preferences, so that the recommendation results can better reflect the user's true interest preferences and improve the accuracy of recommended multimedia files. On the other hand, the user representation of the target user is obtained by learning the historical behavior sequence composed of multimedia files that have historical interactive behaviors with the target user. Even if the interaction between the candidate multimedia files and the user is extremely sparse during the cold start process, the user's preferences can be reflected through the target user's historical behavior sequence, thereby more accurately achieving the recommendation of multimedia files.

[0098] The following describes in detail the steps in the above process and the effects that can be further produced in conjunction with the embodiments.

[0099] First, the above step 201, namely, "obtaining a historical behavior sequence of a target user and a plurality of candidate multimedia files, where the historical behavior sequence includes one or more multimedia files with which the target user has historical interactive behaviors," is described in detail with reference to an embodiment.

[0100] In the embodiment of the present application, the historical behavior sequence refers to a sequence composed of multimedia files that have interacted with the target user. Among them, interactive behaviors may include clicks, favorites, or plays. The historical behavior sequence is used to reflect the user's potential interest patterns. Therefore, the method of the present application first collects the user's historical interaction data and constructs a behavior sequence containing multiple multimedia files. This operation is intended to collect the necessary data basis for the recommendation process. It can be understood that the historical behavior sequence contains the target user's previous interaction records with multimedia files, which can reflect the user's interest preferences and behavior patterns. Multiple candidate multimedia files are a collection of suitable recommendation objects from which the recommendation system needs to screen out. These candidate files can come from a multimedia content library and can cover various forms such as video, audio, graphics, and text. By collecting these two parts of data, preparations can be made for subsequent personalized recommendations.

[0101] Next, in conjunction with an embodiment, the above-mentioned step 202, namely, "extracting a causal representation of the historical behavior sequence in at least one modality and a causal representation of each candidate multimedia file in at least one modality, wherein the causal representation reflects the causal relationship between the causal representation and the target user preference" is described in detail.

[0102] In the embodiment of the present application, the recommendation model can be used to recommend multimedia files to target users. The structure of the recommendation model can be as follows: Figure 3 As shown in . Causal representation refers to the features extracted from multimodal information that have a direct causal relationship with user preferences. Specifically, it can be trained using the contrastive learning framework. Figure 4 The causal representation learning module shown in is implemented to eliminate the interference of confounding factors on the recommendation model.

[0103] Step 202 extracts representations that reflect causal relationships with user preferences from historical behavior sequences and candidate multimedia files. "Modality" here refers to the characteristic manifestations of multimedia files in different aspects, such as visual modalities (pictures, images, etc.), textual modalities (titles, descriptions, subtitles, etc.), and audio modalities (voice, music, etc.). Through specialized feature extraction and causal relationships, features that are causally related to the user's preferences in at least one modality are extracted from the historical behavior sequence to form a causal representation of the user's behavior. Simultaneously, a similar operation is performed on each candidate multimedia file to extract features that may be causally related to the user's preferences in at least one modality, thereby obtaining a causal representation of the file. This causal representation can help the system understand the deep causal relationship between user preferences and multimedia file features, rather than just superficial correlations, thereby providing a basis for accurate recommendations.

[0104] Next, in conjunction with an embodiment, the above-mentioned step 203, namely, "obtaining a user representation of the target user based on the causal representation of the historical behavior sequence in at least one modality, and obtaining a file representation of each candidate multimedia file based on the causal representation of each candidate multimedia file in at least one modality" is described in detail.

[0105] In the embodiment of the present application, user representation refers to the expression of user interests formed by integrating the causal features of historical behaviors. Specifically, causal information of different modalities can be integrated through a multi-head attention mechanism to capture stable user preferences; file representation refers to the feature expression of candidate multimedia files in the causal dimension. Specifically, a gated network can be used to dynamically adjust the weights of the causal features of each modality, thereby enhancing the expressive power of key features.

[0106] It is understandable that after obtaining the causal representations of the historical behavior sequence and the candidate multimedia files, the next step is to construct user representations and file representations respectively. The construction of the user representation is to comprehensively integrate the causal representations of the historical behavior sequence in each modality to form a feature vector or feature matrix that can comprehensively and accurately reflect the preferences of the target user. This user representation can be understood as a "user portrait" that condenses key information such as the user's interests, preferences, and behavioral habits in different modalities. Furthermore, the construction of the file representation is to fuse the causal representations of each candidate multimedia file in each modality to obtain a feature representation that can represent the core features of the file and features associated with user preferences. This can be used to display the key characteristics of the file and aspects that may attract users.

[0107] It is understandable that, compared with the prior art, existing recommendation models usually directly use the statistical correlation of multimodal features for prediction, without effectively distinguishing the causal relationship between features and user preferences. For example, when analyzing video recommendation scenarios, traditional methods may use the visual features of high-click-through-rate videos as the basis for recommendation, but ignore that the click behavior may be caused by the popularity of the video rather than the user's real interest. In the embodiment of the present application, this solution can remove the influence of confounding factors such as popularity through causal representation, and only retain modal features that are causally related to user preferences, thereby effectively suppressing the interference of non-causal factors in multimodal information on the recommendation process. Especially in the case of new file recommendations or low-activity user scenarios, causal relationship modeling can accurately identify the core interests of users, avoid over-reliance on statistical biases in historical interaction data, and enhance the generalization ability of the recommendation system in data-sparse scenarios, thereby improving the accuracy of recommendation results and user satisfaction.

[0108] In an embodiment of the present application, in order to make the recommendation results obtained based on user representation and file representation more in line with user needs and to improve the accuracy of the recommendation, the present application can construct user representation by utilizing the attribute representation and causal representation of user behavior, and construct file representation by utilizing the attribute representation and causal representation of candidate multimedia files.

[0109] like Figure 3 As shown in , the recommendation model provided by the embodiment of the present application adopts a "dual-tower" structure, including a "user tower" (i.e., a user representation extraction module) and a "file tower" (i.e., a file representation extraction module). The user representation extraction module is responsible for extracting user representations, while the file representation extraction module is responsible for extracting file representations.

[0110] Furthermore, in addition to the above-mentioned causal representation, attribute representation can be further combined when determining user representation and file representation. As one of the more preferred implementations, the user representation of the target user can be obtained based on the attribute representation of the target user, the attribute representation of the historical behavior sequence, and the causal representation under at least one modality. For example, the attribute representation of the target user and the attribute representation of the historical behavior sequence are obtained; and the file representation extraction module obtains the user representation of the target user based on the attribute representation of the target user, the attribute representation of the historical behavior sequence, and the causal representation under at least one modality.

[0111] The target user's attribute representation can be a representation obtained by mapping the target user's attributes to a preset space, and can be embodied as an ID representation. The attribute representation of the historical behavior sequence can be composed of the attribute representations of each multimedia file included in the historical behavior sequence. The attribute representation of the multimedia file can be a representation obtained by mapping the attributes of the multimedia file to a preset space, and can also be embodied as an ID representation.

[0112] It can be understood that when generating user representation, the attribute representation of the target user and the attribute representation of the historical behavior sequence can be jointly introduced into the calculation process of the user representation. The user attribute representation provides static preference information, the historical behavior attribute representation reflects the dynamic behavior trend, and the causal representation reveals the potential causal relationship between user preferences and multimedia file content. The three work together to build a more comprehensive user interest model.

[0113] It is worth noting that this solution effectively utilizes the complementarity between multi-source heterogeneous data by integrating user attributes, behavioral attributes and multimodal causal representations, alleviates the data sparsity problem, and enhances the robustness of user representation in cold start scenarios.

[0114] In addition to the above methods, user attributes may not be considered, or other types of representations and causal representations of historical behavior sequences in at least one modality may be further considered to determine user representations, which are not listed here one by one.

[0115] In the solution provided by this application, when determining the user representation of the target user and the file representation of each candidate multimedia file, the attribute representation and the causal representation in at least one modality are combined to drive multimedia file recommendations, thereby improving the accuracy of multimedia file recommendations. In this case, the attribute representation of each candidate multimedia file can also be further obtained; based on the attribute representation of each candidate multimedia file and the causal representation in at least one modality, the file representation extraction module obtains the file representation of each candidate multimedia file.

[0116] Among them, it can be understood that the above-mentioned attribute representation can be realized through embedding layer conversion, for example, the attribute information of the target user is input into the embedding layer to obtain the attribute representation of the target user, and for another example, the attribute information of the multimedia file is input into the embedding layer to obtain the attribute representation of the multimedia file. Causal representation refers to the feature that can reflect the causal relationship between user preferences and multimedia files obtained through causal reasoning learning, which can be specifically extracted through a causal representation learning module. Among them, the causal representation learning ability of the causal representation learning module can be obtained by training using a contrastive learning mechanism during the training of the recommendation model, which is used to eliminate the interference of confounding factors on user preferences.

[0117] Through the above technical solution, this application can enhance the robustness of file representation through the complementarity of attribute and causal representation when the popularity of candidate multimedia files differs significantly or there is a cold start problem; when the candidate multimedia files lack interaction data, the attribute representation provides basic semantic information to ensure the feasibility of recommendation; when the candidate files have sufficient behavioral data, the causal representation accurately captures user preference associations, thereby improving the accuracy of recommendation results in both cold start and conventional scenarios.

[0118] In addition to the above methods, it is also possible to ignore the attribute representation of the multimedia file, or further consider other types of representations and the causal representation of the multimedia file in at least one modality to determine the file representation, which are not listed here one by one.

[0119] In order to generate more accurate user representations and fully explore user preferences, this application further proposes a method for generating user representations, including: using the attribute representation of the target user and the attribute representation of the historical behavior sequence to perform a first multi-head attention processing to obtain a first representation of the target user; using the attribute representation of the target user and the causal representation of the historical behavior sequence in at least one modality to perform a second multi-head attention processing to obtain a second representation of the target user; splicing the first representation and the second representation to obtain a user representation of the target user.

[0120] In the solution provided by the present application, dual multi-head attention processing and representation splicing can be used to extract and fuse the user's attribute representation, the attribute representation of the historical behavior sequence, and the multimodal causal representation to generate a user representation. Among them, the first multi-head attention (user attribute + behavior attribute) can be used to extract the first representation representing long-term interests, and the second multi-head attention (user attribute + causal representation) can be used to extract the second representation representing real-time preferences. Splicing the first and second representations can generate a more accurate user representation, thereby fully exploring user preferences.

[0121] In order to make the file representation more in line with user interests and improve the rationality of recommendations, the present application further proposes a method for obtaining the file representation of each candidate multimedia file based on the attribute representation of each candidate multimedia file and the causal representation under at least one modality. The method can be performed for each candidate multimedia file separately: the attribute representation of the candidate multimedia file and the causal representation of the candidate multimedia file under at least one modality are weighted, and the representations obtained after the weighted processing are spliced ​​to obtain the file representation of the candidate multimedia file; wherein, the weight coefficient used in the weighted processing is obtained by the gated network using the heat representation of the candidate multimedia file.

[0122] In the solution provided by this application, weighted processing can be implemented using weight coefficients assigned by a gating network. The weights of the attribute representations of candidate multimedia files and the causal representations of the candidate multimedia files in at least one modality are adjusted based on the popularity of the files to generate file representations. Dynamically calculating weight coefficients based on the gating network and weightedly fusing the file attribute representations with the causal representations can effectively highlight the characteristics of popular files, making the file representations more aligned with user interests and improving the rationality of recommendations.

[0123] Weighted processing involves assigning different weights to attribute representations and causal representations of different modalities, followed by a linear combination. This can be achieved using dot product operations or fully connected layers, dynamically adjusting the contribution of different representations to the final file representation. A gated network is a neural network structure used to generate weight coefficients, implemented using a multi-layer perceptron or attention mechanism. It dynamically adjusts the fusion ratio of attribute and causal representations based on the heat representation. Heat representations are feature vectors that reflect the popularity of candidate multimedia files. These can be generated using statistical indicators such as click-through rate, playback volume, or user interaction frequency, providing an objective popularity reference for weight generation.

[0124] It's worth noting that this solution, through the introduction of a gating network, enables the popularity of cold-start files to directly influence the representation fusion process. For example, for newly uploaded low-popularity files, the system automatically increases the weight of causal representations to capture potential user preferences, while for high-popularity files, it prioritizes the stability of attribute representations. This dynamic adjustment mechanism effectively balances the conflict between data sparsity and feature reliability in cold-start scenarios, reducing recommendation bias caused by insufficient data.

[0125] Next, the above step 204 , namely “selecting at least one candidate multimedia file from the candidate multimedia files as a recommendation result for the target user based on the user representation of the target user and the file representation of each candidate multimedia file”, is described in detail with reference to an embodiment.

[0126] In an embodiment of the present application, step 204 can utilize the user representation and file representation that have been constructed to evaluate the degree of match between each candidate multimedia file and the target user through a specific matching algorithm or similarity calculation method. Based on the matching results, one or more multimedia files are selected from all candidate multimedia files and presented to the user as the final recommendation result. Such recommendation results are based on an understanding of the deep causal relationship between user preferences and file features, and have better accuracy, thereby achieving a more stable recommendation effect and improving user satisfaction.

[0127] Figure 4 Another system architecture diagram applicable to the embodiment of the present application; Figure 4 As shown, in order to realize the causal representation of the historical behavior sequence in at least one modality and the extraction of the causal representation of each candidate multimedia file in at least one modality based on the embedded representation, before performing the multimedia file recommendation task, the recommendation model including the causal representation learning module can be pre-trained by a device for training the recommendation model. The causal representation learning module is used to mine causal relationships. The embedded representations of each multimedia file included in the historical behavior sequence and the candidate multimedia files in at least one modality can be uniformly processed through the causal representation learning module, and then the causal representation of the historical behavior sequence in at least one modality and the extraction of the causal representation of each candidate multimedia file in at least one modality can be realized based on the embedded representation.

[0128] Based on this, as one of the feasible ways, when extracting the causal representation of the historical behavior sequence in at least one modality, the embedded representation of each multimedia file in at least one modality can be extracted for each multimedia file included in the historical behavior sequence; the embedded representation of each multimedia file in at least one modality is input into the causal representation learning module to obtain the causal representation of each multimedia file in at least one modality. The causal representation of the historical behavior sequence in at least one modality is obtained from the causal representation of the multimedia files included in the historical behavior sequence in at least one modality.

[0129] As one of the feasible ways, when extracting the causal representation of the candidate multimedia file in at least one modality, the embedded representation of the candidate multimedia file in at least one modality can be extracted; the embedded representation of the candidate multimedia file in at least one modality is input into the causal representation learning module to obtain the causal representation of the candidate multimedia file in at least one modality.

[0130] Embedded representation refers to converting the raw data of multimedia files into low-dimensional dense vectors through an embedding network. The embedding network is typically a neural network model. For example, a convolutional neural network can be used to extract embedded representations of images, a recurrent neural network can be used to extract embedded representations of text, and a spectrum analysis model can be used to extract embedded representations of audio. The causal representation learning module uses causal inference methods to extract features that are causally related to user preferences from the embedded representation. Modality refers to the different information representation forms of multimedia files, such as visual modality corresponding to image content, text modality corresponding to title or description, and audio modality corresponding to background music or voice content.

[0131] It can be understood that for each multimedia file in the target user's historical behavior sequence, its embedded representations in multiple modalities are first obtained through the embedding network; for example, a video file can be decomposed into three modalities: picture frame sequence, subtitle text, and background music, and visual, text, and audio embedded representations are extracted respectively. Subsequently, these embedded representations are input into the causal representation learning module, and the causal representations in each modality are obtained by removing the confounding factors that have no causal relationship with the user's preferences. The causal representation of the historical behavior sequence is generated by aggregating the causal representations of each multimedia file in the sequence, such as using the weighted summation of the attention mechanism. The causal representation generation process of the candidate multimedia files is similar to that of the files in the historical behavior sequence, and is also achieved through modal decomposition, embedded representation extraction, and causal inference steps.

[0132] It's worth noting that, compared to existing technologies, traditional recommendation systems focus solely on statistical correlations when fusing multimodal features. For example, they directly concatenate or weighted average embedding vectors from different modalities, leading to confounding factors like exposure bias of popular files in cold-start scenarios. This solution explicitly models the causal relationship between user preferences and multimodal features through a causal representation learning module. For example, in video recommendation scenarios, this approach can identify visual elements that users are genuinely interested in, rather than false associations misled by popular tags, thereby reducing the impact of data sparsity on recommendation effectiveness.

[0133] Figure 5 Flowchart of the method for training a recommendation model provided in an embodiment of the present application; Figure 5 As shown, the present application further proposes a method for training a recommendation model, which may include the following steps:

[0134] Step 501: Acquire training data comprising a plurality of sample pairs, wherein the sample pairs include positive sample pairs and negative sample pairs, wherein the positive sample pairs include user samples and multimedia samples that have historical interactions with the user samples, and the negative sample pairs include user samples and multimedia samples that have no historical interactions with the user samples;

[0135] Step 502: Training a recommendation model using training data, wherein a historical behavior sequence of a user sample in a sample pair and a multimedia sample in the sample pair are obtained, wherein the historical behavior sequence includes one or more multimedia files that have had historical interactions with the user sample; the recommendation model extracts a causal representation of the historical behavior sequence in at least one modality and a causal representation of each multimedia sample in at least one modality, wherein the causal representation reflects the causal relationship between the causal representation and the user preference; the recommendation model obtains a user representation of the user sample based on the causal representation of the historical behavior sequence in at least one modality, and obtains a file representation of the multimedia sample based on the causal representation of the multimedia sample in at least one modality; and the recommendation model determines a preference score for the multimedia sample based on the user representation of the user sample and the file representation of the multimedia sample.

[0136] Among them, the training goal includes maximizing the difference between the preference scores of multimedia samples included in the positive sample pair and the preference scores of multimedia samples included in the negative sample pair in the sample pair. The above-mentioned method of training the recommendation model can be based on the training data of the positive and negative sample pairs, with the goal of maximizing the difference in preference scores to train the recommendation model. The recommendation model trained by this method can construct corresponding user representation and file representation based on the causal representation under at least one modality, and then determine the preference score of the multimedia sample. This way of organically combining causal relationship learning and multimodal learning can learn the factors in each modality that have a causal relationship with the user's preference, and realize the effective separation of non-causal interference factors in each modality, so that the model can accurately recommend multimedia files preferred by the user. Among them, it is worth noting that the part of the recommendation model that extracts causal representation can obtain the ability to extract causal representation by synchronously optimizing the corresponding loss function during model training. The relevant loss function will be discussed in detail in the subsequent embodiments.

[0137] In the embodiments of this application, sample pairs include positive and negative sample pairs, referred to as positive sample pairs and negative sample pairs, respectively. Specifically, positive samples can be constructed by using multimedia files containing interactive behaviors in a user's historical click history, and negative samples from multimedia files without interactive behaviors. This feature enhances the model's ability to capture users' true preferences by constructing a comparative learning mechanism between positive and negative samples.

[0138] It is worth noting that compared with existing technologies, traditional recommendation model training relies on statistical correlation and is susceptible to false associations. For example, highly popular files may be misjudged as user preferences due to their exposure advantage. This solution uses causal representation learning to clearly model the causal relationship between user preferences and file features, thereby reducing the impact of exposure on model decisions. In addition, this solution directly enhances the model's adaptability to sparse interactive data through a joint training mechanism of multimodal and causal representation.

[0139] like Figure 3As shown, the recommendation model may include: a causal representation learning module, a user representation extraction module, a file representation extraction module and a recommendation module, wherein the causal representation learning module can maximize the cross-modal causal representation mutual information through comparative learning, eliminate the deviation of each modal feature, so as to achieve semantic consistency of causal representation under multimodality, eliminate the recommendation deviation caused by semantic conflict between modalities, and thus improve the effect of recommendation based on the causal representation.

[0140] The causal representation learning module extracts the causal representation of historical behavior sequences and each multimedia sample in at least one modality.

[0141] The user representation extraction module obtains the user representation of the user sample based on the attribute representation of the user sample, the attribute representation of the historical behavior sequence, and the causal representation.

[0142] The file representation extraction module obtains the file representation based on the attribute representation and causal representation of the multimedia samples.

[0143] The recommendation module determines preference scores of multimedia samples based on user representation and file representation.

[0144] The training objective is to maximize the difference in preference scores between the multimedia samples included in the positive sample pairs and the negative sample pairs.

[0145] For the specific functions of each module, please refer to the relevant records in the previous embodiments.

[0146] Among them, the causal representation learning module refers to a module that can extract representations that have a causal relationship with user preferences from the embedded representations of multimedia files. Specifically, it can be trained using a deep neural network combined with a contrastive learning algorithm to capture the key factors that affect user preferences under different modalities.

[0147] As one possible implementation, the training objective can also include maximizing the mutual information between the causal representations extracted by the causal learning module for each sample pair in different modalities. This training objective can employ a contrastive learning mechanism to enhance the causal representation learning capability of the causal learning module.

[0148] Among them, the above-mentioned interactive behaviors may include but are not limited to clicks, plays or collection behaviors recorded in the user's historical behavior log; mutual information refers to the dependency relationship or shared information between causal representations under different modalities. By maximizing the mutual information of multimedia files contained in sample pairs under different modalities, cross-modal shared causal relationships are captured, thereby filtering out confounding factors unrelated to user preferences; contrastive learning refers to distinguishing relevant and irrelevant features by constructing positive and negative sample pairs for training models.

[0149] In order to effectively improve the model's recommendation accuracy for multimedia files such as short videos in cold start scenarios, in the solution provided in this application, the causal representation learning module can specifically include a shared representation learning module and a unique representation learning module.

[0150] The shared representation learning module is used to learn the shared causal representation of the input multimedia file in at least one modality based on the embedded representation of the multimedia file in at least one modality input to the causal representation learning module.

[0151] The unique representation learning module is used to learn a unique causal representation of the input multimedia file in at least one modality based on the embedded representation of the input multimedia file in at least one modality.

[0152] The causal representation of the input multimedia file in at least one modality is obtained from the shared causal representation and the unique causal representation.

[0153] In this case, the objectives of the above training may also include: minimizing the distance between the shared causal representation and the unique causal representation extracted by the causal representation learning module for the multimedia files contained in the sample pair under the same modality, maximizing the mutual information between the shared causal representations extracted by the causal relationship learning module for the multimedia files contained in the sample pair under different modalities, and maximizing the mutual information between the shared causal representation and the unique causal representation extracted by the causal relationship learning module for the multimedia files contained in the sample pair under the same modality.

[0154] In the embodiment of the present application, it is assumed that the user set and multimedia file collections Select from Select user samples from Select multimedia files that have historical interaction behaviors with the user sample to form a historical behavior sequence Expressed as This user sample from Multimedia file i is selected from u to form a positive sample pair, and the preference score of multimedia file i relative to u can be set to 1.

[0155] The specific sampling and processing process of negative sample pairs may include: Sample one of Then to Corresponding Processing to obtain Make from Select user samples from Select Video Constitute negative sample pairs, multimedia files Relative to The preference score can be set to 0.

[0156] The sample pairs used in training the recommendation model in the above embodiment can be selected from the above positive sample pairs and negative sample pairs.

[0157] For any sample pair, the user and multimedia file are assumed to be represented as u and i respectively. The attribute representations of u and i are represented as and Considering the three modalities of vision, text and speech, the embedding representations of the input corresponding to these modalities of i can be respectively Furthermore, corresponding to the description in the above embodiment, it can be known that the recommendation model outputs a preference score for each given sample pair (u, i) to predict the possibility that user u prefers video i.

[0158] Among them, the embedded representation of multimedia file i in a modality is the basic feature representation of the multimedia file in a specific modality (such as vision, text or speech, etc.), which is obtained through the corresponding embedding network (such as Figure 6 The mid-modal embedding network (including the video embedding network, text embedding network or audio embedding network) is processed and contains the original feature information of the multimedia file under this modality, which is the basis for subsequent feature classification and processing.

[0159] like Figure 6 As shown in , the attribute representations of user sample u and multimedia file i can be extracted through the ID embedding network.

[0160] Assume that the causal representation learning module includes Figure 6 The unique representation learning module and the shared representation learning module shown in , wherein the unique representation learning module is unique to each modality, and the shared representation learning module is shared by all modalities. The unique representation learning module is assumed to include three modality-specific video mapping networks, text mapping networks, and audio mapping networks. Then, for the embedded representations of the three modalities of multimedia file i, namely the video embedding representation, the text embedding representation, and the audio embedding representation, after being processed by the video mapping network, the text mapping network, and the audio mapping network, a unique causal representation in the video modality, a unique causal representation in the text modality, and a unique causal representation in the audio modality are obtained, respectively. The video embedding representation, the text embedding representation, and the audio embedding representation are processed by the shared representation learning module to obtain a shared causal representation in the video modality, a shared causal representation in the text modality, and a shared causal representation in the audio modality.

[0161] Shared causal features of multimedia files within a modality are shared features extracted from the embedded representation of the multimedia file that are causally related to user preferences. These features reflect the association between the multimedia file within that modality and the target user's preferences. They are the features shared across modalities that can trigger user preference responses. For example, for video files, a shared causal feature might be a specific visual style or color combination within the video, which has been frequently clicked and viewed by users in the past and is causally related to their preferences.

[0162] The unique causal features of a multimedia file within a modality are characteristics unique to that modality and causally related to user preferences. These features distinguish the multimedia file from other files and precisely attract the attention of target users. For example, a song's unique melody or arrangement might be its unique causal feature. This melody or arrangement is causally related to the user's preference for that music type, motivating them to choose to listen to the song.

[0163] The causal representation of multimedia file i in at least one modality is obtained from the shared causal representation and the unique causal representation. For example, the shared causal representation and the unique causal representation can be spliced ​​or fused in a manner such as weighted summation.

[0164] In the process of model training, as one of the feasible ways, we can jointly optimize the loss function and The causal representation learning ability of the causal representation learning module is trained in a manner, that is, the training objectives include: minimizing the distance between the shared causal representation and the unique causal representation extracted by the causal representation learning module for the multimedia files contained in the sample pair under the same modality, maximizing the mutual information between the shared causal representations extracted by the causal relationship learning module for the multimedia files contained in the sample pair under different modalities, and maximizing the mutual information between the shared causal representation and the unique causal representation extracted by the causal relationship learning module for the multimedia files contained in the sample pair under the same modality.

[0165] Assume that for the same multimedia file i, the shared causal representation obtained by the shared representation learning module under modality m is The unique causal representation obtained by the unique representation learning module is expressed as: In order to constrain them to contain differentiated information, an orthogonal loss objective function is designed to minimize the square of the two-norm of their dot product:

[0166]

[0167] in, This constraint constrains shared and unique causal representations to contain differentiated information. By minimizing the squared bi-norm of the dot product of the shared and unique causal representations of multimedia file i within the same modality, the two representations are made orthogonal in the feature space, preventing redundant information from being included. This ensures that the shared causal representation focuses on multimodal common features, while the unique causal representation focuses on the unique features of a single modality. Essentially, this minimizes the distance between the shared and unique causal representations extracted from the same modality for each multimedia file in the sample pair using the causal representation learning module.

[0168] Furthermore, the maximization of the mutual information between the shared causal representations extracted from the multimedia files in different modalities contained in the sample pair by the causal relationship learning module can be understood as follows: since the causal features shared by the modalities reflect the potential semantic associations between different modalities, the conditional mutual information can be maximized. Learn a shared causal representation where m1 and m2 represent different modalities. This information-theoretic goal is estimated using contrastive learning, i.e., Compute contrastive learning loss:

[0169]

[0170] Where τ is the temperature hyperparameter. For simplicity, is the shared causal representation of multimedia file i under modality m, represents the shared causal representation of multimedia file i under modality m', denotes the shared causal representation of multimedia file i under modality m' in the kth negative sample pair. N is the number of negative sample pairs corresponding to the positive sample pair. This system is used to learn modality-shared causal representations, maximizing the mutual information between shared causal representations across different modalities and eliminating bias in cross-modal features. It can be understood that for a user's modality, features extracted by the shared representation learning module for the user's other two modalities are considered positive, while features extracted by the shared feature learning module for the same modality for other users are considered negative. Conditional mutual information is estimated using contrastive learning. This allows the shared causal representations of multimedia files across different modalities to be closer in feature space for the same user, thereby extracting shared features that reflect the underlying semantic associations between the different modalities.

[0171] Furthermore, the maximization causal relationship learning module can maximize the mutual information between the shared causal representation and the unique causal representation extracted from the multimedia files contained in the sample pair under the same modality by maximizing the conditional mutual information To learn the unique causal representation under modality m

[0172]

[0173] in, represents the unique causal representation of multimedia file i under modality m in the kth negative sample pair. N is the number of negative sample pairs corresponding to the positive sample pair.

[0174] Through joint optimization The model can jointly learn the shared causal representation and the unique causal representation of each modality. The modality causal representation is obtained by concatenating the shared causal representation and the unique causal representation.

[0175] Continue to see Figure 6 When training the recommendation model, after obtaining the attribute characteristics of user samples and multimedia files through the ID embedding network, and obtaining the unique causal representation and shared causal representation of each multimedia file in each modality through the unique representation learning module, the user representation and file representation are extracted respectively by adopting the "dual tower" structure.

[0176] Assume that the attribute representation of user u’s historical behavior sequence is H u , the causal representation under mode m is First, the user tower, i.e. the user representation extraction module, performs the first multi-head attention processing, i.e. uses the multi-head attention mechanism to model the user's historical behavior sequence and obtain its overall interest representation. Where MHA() represents multi-head attention processing, Q, K, and V are the query, key, and value in the multi-head attention processing respectively. This overall representation is the first representation in the above embodiment. Next, for each modality, its corresponding causal representation is introduced And perform the second multi-head attention processing, that is, construct the user's causal interest representation in this modality through the multi-head attention mechanism: This causal interest representation is the second representation in the above embodiment. u It is concatenated with the causal interest representations under each modality and expressed as The final output of the user tower is the target user's user representation. This representation integrates the user's global behavioral interests and multimodal causal preference information, serving as the final feature input on the user side for subsequent preference score calculation.

[0177] Furthermore, assuming that the attribute representation of the candidate multimedia file i is Its multimodal causal representations are Due to the lack of interaction between videos and users in the cold start phase, It is not accurate enough, so we design a gating network with multimedia file heat features as input, and use its output σ iTo regulate the weight between attribute representation and multimodal causal representation, the final representation output by the file tower, i.e., the file representation extraction module, is A file representation of the candidate multimedia file is obtained.

[0178] Then, the user representation and the file representation of the candidate multimedia file are used to calculate the preference score of the candidate multimedia file for the user sample. Then, the Bayesian personalized ranking criterion is applied to calculate the loss. The recommendation model is optimized to maximize the difference between the preference scores of the multimedia samples included in the positive sample pair and the preference scores of the multimedia samples included in the negative sample pair in the sample pair, thereby obtaining a trained recommendation model.

[0179] As mentioned above, this application can be achieved through the above loss function as well as The combination of ,can optimize the end-to-end recommendation model.

[0180] From the above embodiments, it can be seen that for multimedia files, their file representation is obtained by the shared causal representation, unique causal representation and attribute representation (i.e., non-causal representation) in each modality, while the user representation is obtained by combining the shared causal representation and unique causal representation of the user behavior sequence in addition to the attribute representation (the attribute representation of the multimedia file contained in the user and historical behavior sequence), such as Figure 7 As shown in .

[0181] Figure 8 A schematic block diagram of a multimedia file recommendation device provided in an embodiment of the present application; Figure 8 As shown, the present application provides a multimedia file recommendation device, which may include:

[0182] An acquisition unit 801 is configured to acquire a historical behavior sequence of a target user and a plurality of candidate multimedia files, wherein the historical behavior sequence includes one or more multimedia files with which the target user has historical interaction behaviors;

[0183] The representation extraction unit 802 is configured to extract a causal representation of the historical behavior sequence in at least one modality and a causal representation of each candidate multimedia file in at least one modality, wherein the causal representation reflects the causal relationship between the causal representation and the target user's preference; obtain a user representation of the target user based on the causal representation of the historical behavior sequence in at least one modality; and obtain a file representation of each candidate multimedia file based on the causal representation of each candidate multimedia file in at least one modality;

[0184] The recommendation unit 803 is configured to select at least one candidate multimedia file from the candidate multimedia files as a recommendation result for the target user based on the user representation of the target user and the file representation of each candidate multimedia file.

[0185] Furthermore, the acquisition unit 801 may also be configured to acquire the attribute representation of the target user and the attribute representation of the historical behavior sequence.

[0186] Accordingly, when the representation extraction unit 802 obtains the user representation of the target user based on the causal representation of the historical behavior sequence in at least one modality, it is specifically configured to obtain the user representation of the target user based on the attribute representation of the target user, the attribute representation of the historical behavior sequence, and the causal representation of the historical behavior sequence in at least one modality.

[0187] Furthermore, the acquisition unit 801 may be further configured to acquire an attribute representation of each candidate multimedia file. Accordingly, when the representation extraction unit 802 obtains the file representation of each candidate multimedia file based on the causal representation of each candidate multimedia file in at least one modality, it is specifically configured to obtain the file representation of each candidate multimedia file based on the attribute representation of each candidate multimedia file and the causal representation of each candidate multimedia file in at least one modality.

[0188] As one possible implementation, when the representation extraction unit 802 obtains the user representation of the target user based on the attribute representation of the target user, the attribute representation of the historical behavior sequence, and the causal representation of the historical behavior sequence in at least one modality, it can be specifically configured as follows:

[0189] Using the attribute representation of the target user and the attribute representation of the historical behavior sequence, perform the first multi-head attention processing to obtain the first representation of the target user;

[0190] Using the attribute representation of the target user and the causal representation of the historical behavior sequence in at least one modality, a second multi-head attention process is performed to obtain the second representation of the user;

[0191] The first representation and the second representation are concatenated to obtain a user representation of the target user.

[0192] As one possible implementation, when the representation extraction unit 802 obtains the file representation of each candidate multimedia file based on the attribute representation of each candidate multimedia file and the causal representation of each candidate multimedia file in at least one modality, it can be specifically configured as follows:

[0193] For each candidate multimedia file, the following steps are performed: weighted processing is performed on the attribute representation of the candidate multimedia file and the causal representation of the candidate multimedia file in at least one modality, and the representations obtained after the weighted processing are spliced ​​to obtain the file representation of the candidate multimedia file; wherein the weight coefficient used in the weighted processing is obtained by the gated network using the heat representation of the candidate multimedia file.

[0194] As one of the feasible ways, when extracting the causal representation of the historical behavior sequence in at least one modality, the representation extraction unit 802 can be specifically configured as follows: for each multimedia file included in the historical behavior sequence, extract the embedded representation of each multimedia file in at least one modality; input the embedded representation of each multimedia file in at least one modality into the causal representation learning module to obtain the causal representation of each multimedia file in at least one modality, and the causal representation of the historical behavior sequence in at least one modality is obtained from the causal representation of the multimedia files included in the historical behavior sequence in at least one modality.

[0195] When extracting the causal representation of a candidate multimedia file in at least one modality, the representation extraction unit 802 can be specifically configured to: extract the embedded representation of the candidate multimedia file in at least one modality; input the embedded representation of the candidate multimedia file in at least one modality into the causal representation learning module to obtain the causal representation of the candidate multimedia file in at least one modality.

[0196] The multimedia file recommendation device can be configured to implement the multimedia file recommendation method described in any one of the embodiments and achieve corresponding technical effects, so it will not be described in detail here.

[0197] Figure 9 A schematic block diagram of a device for training a recommendation model provided in an embodiment of the present application; Figure 9 As shown, the present application provides a device for training a recommendation model, which may include:

[0198] A data acquisition unit 901 is configured to acquire training data including a plurality of sample pairs, wherein the sample pairs include positive sample pairs and negative sample pairs, wherein the positive sample pairs include user samples and multimedia samples with which the user samples have historical interactions, and the negative sample pairs include user samples and multimedia samples with which the user samples have no historical interactions;

[0199] A training unit 902 is configured to train a recommendation model using training data, wherein a historical behavior sequence of a user sample in a sample pair and a multimedia sample in the sample pair are obtained, the historical behavior sequence including one or more multimedia files that have historically interacted with the user sample; the recommendation model extracts a causal representation of the historical behavior sequence in at least one modality and a causal representation of each multimedia sample in at least one modality, the causal representation reflecting the causal relationship between the causal representation and the user preference; the recommendation model obtains a user representation of the user sample based on the causal representation of the historical behavior sequence in at least one modality, and obtains a file representation of the multimedia sample based on the causal representation of the multimedia sample in at least one modality; and the recommendation model determines a preference score for the multimedia sample based on the user representation of the user sample and the file representation of the multimedia sample.

[0200] The training objectives include maximizing the difference between the preference scores of the multimedia samples included in the positive sample pairs and the preference scores of the multimedia samples included in the negative sample pairs in the sample pairs.

[0201] The recommendation model may include: a causal representation learning module, a user representation extraction module, a file representation extraction module, and a recommendation module;

[0202] The causal representation learning module extracts the causal representation of the historical behavior sequence in at least one modality and the causal representation of each multimedia sample in at least one modality. The causal representation reflects the causal relationship between the causal representation and the user's preference.

[0203] The user representation extraction module obtains the user representation of the user sample based on the attribute representation of the user sample, the attribute representation of the historical behavior sequence, and the causal representation of the historical behavior sequence in at least one modality;

[0204] The file representation extraction module obtains the file representation of the multimedia sample based on the attribute representation of the multimedia sample and the causal representation of the multimedia sample in at least one modality;

[0205] The recommendation module determines the preference score of the multimedia sample based on the user representation of the user sample and the file representation of the multimedia sample.

[0206] Furthermore, the training goal adopted by the training unit 902 may also include: maximizing the mutual information between the causal representations extracted by the causal relationship learning module in different modalities for the multimedia files contained in the sample pairs.

[0207] As one of the possible implementation methods, the causal representation learning module includes a shared representation learning module and a unique representation learning module;

[0208] The shared representation learning module is configured to learn a shared causal representation of the input multimedia file in at least one modality based on the embedded representation of the multimedia file in at least one modality input to the causal representation learning module;

[0209] The unique representation learning module is used to learn a unique causal representation of the input multimedia file in at least one modality based on the embedded representation of the input multimedia file in at least one modality;

[0210] The causal representation of the input multimedia file in at least one modality is obtained from the shared causal representation and the unique causal representation.

[0211] In this implementation, the training objectives adopted by the training unit 902 may also include: minimizing the distance between the shared causal representation and the unique causal representation extracted by the causal representation learning module for the multimedia files contained in the sample pair under the same modality, maximizing the mutual information between the shared causal representations extracted by the causal relationship learning module for the multimedia files contained in the sample pair under different modalities, and maximizing the mutual information between the shared causal representation and the unique causal representation extracted by the causal relationship learning module for the multimedia files contained in the sample pair under the same modality.

[0212] The device for training a recommendation model can be configured to implement the method for training a recommendation model described in any embodiment and achieve corresponding technical effects, so it will not be described in detail here.

[0213] Each embodiment in this specification is described in a progressive manner. The same or similar parts between the embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments. In particular, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple. For the relevant parts, refer to the partial description of the method embodiment. The device embodiment described above is merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. A person of ordinary skill in the art can understand and implement it without expending creative work.

[0214] In addition, an embodiment of the present application further provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps of any one of the methods in the aforementioned method embodiments are implemented.

[0215] And an electronic device comprising:

[0216] one or more processors; and

[0217] A memory associated with one or more processors, the memory being used to store program instructions, which, when read and executed by one or more processors, execute the steps of any one of the method embodiments described above.

[0218] The present application also provides a computer program product, comprising a computer program, which implements the steps of any one of the method embodiments described above when executed by a processor.

[0219] in, Figure 10The electronic device architecture is shown as an example, and may include a processor 1010, a video display adapter 1011, a disk drive 1012, an input / output interface 1013, a network interface 1014, and a memory 1020. The processor 1010, the video display adapter 1011, the disk drive 1012, the input / output interface 1013, the network interface 1014, and the memory 1020 may be communicatively connected via a communication bus 1030.

[0220] Among them, the processor 1010 can be implemented by a general-purpose CPU, a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, etc., to execute relevant programs to implement the technical solutions provided in this application.

[0221] The memory 1020 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 can store an operating system 1021 for controlling the operation of the electronic device 1000, and a basic input and output system (BIOS) 1022 for controlling the low-level operations of the electronic device 1000. In addition, a web browser 1023, a data storage management system 1024, and a multimedia file recommendation device 800 or a device for training a recommendation model 900, etc. can also be stored. The above-mentioned multimedia file recommendation device 800 or the device for training a recommendation model 900 can be an application program that specifically implements the operations of the aforementioned steps in the embodiment of the present application. In short, when the technical solution provided by the present application is implemented by software or firmware, the relevant program code is stored in the memory 1020 and is called and executed by the processor 1010.

[0222] The input / output interface 1013 is used to connect input / output modules to implement information input and output. The input / output modules can be configured as components within the device (not shown in the figure) or can be externally connected to the device to provide corresponding functions. Input devices may include a keyboard, mouse, touch screen, microphone, various sensors, etc., and output devices may include a display, speaker, vibrator, indicator light, etc.

[0223] The network interface 1014 is used to connect to a communication module (not shown) to enable communication between the device and other devices. The communication module can communicate via a wired method (such as USB, network cable, etc.) or a wireless method (such as mobile network, WIFI, Bluetooth, etc.).

[0224] The bus 1030 comprises a pathway for transmitting information between the various components of the device (eg, the processor 1010 , the video display adapter 1011 , the disk drive 1012 , the input / output interface 1013 , the network interface 1014 , and the memory 1020 ).

[0225] It should be noted that although the above device only shows the processor 1010, the video display adapter 1011, the disk drive 1012, the input / output interface 1013, the network interface 1014, the memory 1020, the bus 1030, etc., in the specific implementation process, the device may also include other components necessary for normal operation. In addition, it will be understood by those skilled in the art that the above device may also include only the components necessary to implement the solution of the present application, and does not necessarily include all the components shown in the figure.

[0226] Through the description of the above embodiments, it can be seen that those skilled in the art can clearly understand that the present application can be implemented by means of software plus a necessary general hardware platform. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a computer program product, which can be stored in a storage medium such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods of each embodiment of the present application or certain parts of the embodiments.

[0227] The above describes the technical solution provided by this application in detail. Specific examples are used herein to illustrate the principles and implementation methods of this application. The description of the above embodiments is only intended to help understand the method and core concept of this application. At the same time, for those skilled in the art, based on the concept of this application, there may be changes in the specific implementation methods and application scope. In summary, the contents of this specification should not be understood as limiting this application.

Claims

1. A multimedia file recommendation method, characterized in that: The method comprises: Acquire a historical behavior sequence of a target user and a plurality of candidate multimedia files, wherein the historical behavior sequence includes one or more multimedia files with which the target user has historical interaction behavior; Extracting a causal representation of the historical behavior sequence in at least one modality and a causal representation of each candidate multimedia file in at least one modality, wherein the causal representation reflects a causal relationship between the causal representation and the target user's preference; Obtaining a user representation of the target user based on the causal representation of the historical behavior sequence in at least one modality, and obtaining a file representation of each candidate multimedia file based on the causal representation of each candidate multimedia file in at least one modality; Based on the user representation of the target user and the file representations of the candidate multimedia files, at least one candidate multimedia file is selected from the candidate multimedia files as a recommendation result for the target user.

2. The method according to claim 1, characterized in that The method further includes: obtaining an attribute representation of the target user and an attribute representation of the historical behavior sequence; Obtaining a user representation of the target user based on the causal representation of the historical behavior sequence in at least one modality includes: A user representation of the target user is obtained based on the attribute representation of the target user, the attribute representation of the historical behavior sequence, and the causal representation of the historical behavior sequence in at least one modality.

3. The method according to claim 1, characterized in that The method further comprises: obtaining attribute representations of each candidate multimedia file; Obtaining the file representation of each candidate multimedia file based on the causal representation of each candidate multimedia file in at least one modality includes: Based on the attribute representation of each candidate multimedia file and the causal representation of each candidate multimedia file in at least one modality, a file representation of each candidate multimedia file is obtained.

4. The method according to claim 2, characterized in that Obtaining a user representation of the target user based on the attribute representation of the target user, the attribute representation of the historical behavior sequence, and the causal representation of the historical behavior sequence in at least one modality includes: Performing a first multi-head attention process using the attribute representation of the target user and the attribute representation of the historical behavior sequence to obtain a first representation of the target user; Performing a second multi-head attention process using the attribute representation of the target user and the causal representation of the historical behavior sequence in at least one modality to obtain a second representation of the user; The first representation and the second representation are concatenated to obtain a user representation of the target user.

5. The method according to claim 3, characterized in that Obtaining the file representation of each candidate multimedia file based on the attribute representation of each candidate multimedia file and the causal representation of each candidate multimedia file in at least one modality includes: For each candidate multimedia file, the following steps are performed: weighting processing is performed on the attribute representation of the candidate multimedia file and the causal representation of the candidate multimedia file in at least one modality, and the representations obtained after the weighted processing are spliced ​​to obtain the file representation of the candidate multimedia file; wherein the weight coefficient used in the weighted processing is obtained by the gated network using the heat representation of the candidate multimedia file.

6. The method according to any one of claims 1 to 5, characterized in that Extracting a causal representation of the historical behavior sequence in at least one modality includes: Extracting, for each multimedia file included in the historical behavior sequence, an embedded representation of each multimedia file in at least one modality; inputting the embedded representation of each multimedia file in at least one modality into a causal representation learning module to obtain a causal representation of each multimedia file in at least one modality, wherein the causal representation of the historical behavior sequence in at least one modality is obtained from the causal representation of the multimedia files included in the historical behavior sequence in at least one modality; Extracting the causal representation of the candidate multimedia file in at least one modality includes: Extracting an embedded representation of the candidate multimedia file in at least one modality; inputting the embedded representation of the candidate multimedia file in at least one modality into the causal representation learning module to obtain a causal representation of the candidate multimedia file in at least one modality.

7. A method for training a recommendation model, characterized in that: The method comprises: Acquire training data comprising a plurality of sample pairs, the sample pairs comprising a positive sample pair and a negative sample pair, the positive sample pair comprising a user sample and a multimedia sample with which the user sample has a history of interaction, and the negative sample pair comprising the user sample and a multimedia sample with which the user sample has no history of interaction; The recommendation model is trained using the training data, wherein a historical behavior sequence of the user sample in the sample pair and a multimedia sample in the sample pair are obtained, the historical behavior sequence including one or more multimedia files having historical interactive behaviors with the user sample; the recommendation model extracts a causal representation of the historical behavior sequence in at least one modality and a causal representation of each multimedia sample in at least one modality, the causal representation reflecting the causal relationship between the causal representation and the user preference; the recommendation model obtains a user representation of the user sample based on the causal representation of the historical behavior sequence in at least one modality, and obtains a file representation of the multimedia sample based on the causal representation of the multimedia sample in at least one modality; the recommendation model determines a preference score of the multimedia sample based on the user representation of the user sample and the file representation of the multimedia sample; The training goal is to maximize the difference between the preference scores of the multimedia samples included in the positive sample pair and the preference scores of the multimedia samples included in the negative sample pair in the sample pairs.

8. The method according to claim 7, characterized in that The recommendation model includes: a causal representation learning module, a user representation extraction module, a file representation extraction module and a recommendation module; The causal representation learning module extracts a causal representation of the historical behavior sequence in at least one modality and a causal representation of each multimedia sample in at least one modality, wherein the causal representation reflects a causal relationship between the causal representation and the user preference; The user representation extraction module obtains the user representation of the user sample based on the attribute representation of the user sample, the attribute representation of the historical behavior sequence, and the causal representation of the historical behavior sequence in at least one modality; The file representation extraction module obtains the file representation of the multimedia sample based on the attribute representation of the multimedia sample and the causal representation of the multimedia sample in at least one modality; The recommendation module determines a preference score of the multimedia sample based on the user representation of the user sample and the file representation of the multimedia sample.

9. The method according to claim 8, characterized in that The training goal also includes: maximizing the mutual information between the causal representations extracted by the causal relationship learning module for the multimedia files contained in the sample pairs under different modalities.

10. The method according to claim 8, characterized in that The causal representation learning module includes a shared representation learning module and a unique representation learning module; The shared representation learning module is configured to learn a shared causal representation of the multimedia file in at least one modality based on the embedded representation of the multimedia file in at least one modality input to the causal representation learning module; The unique representation learning module is configured to learn a unique causal representation of the input multimedia file in at least one modality based on the embedded representation of the input multimedia file in at least one modality; The causal representation of the input multimedia file in at least one modality is obtained from the shared causal representation and the unique causal representation.

11. The method according to claim 10, characterized in that The objectives of the training also include: minimizing the distance between the shared causal representation and the unique causal representation extracted by the causal representation learning module for the multimedia files contained in the sample pair under the same modality, maximizing the mutual information between the shared causal representations extracted by the causal relationship learning module for the multimedia files contained in the sample pair under different modalities, and maximizing the mutual information between the shared causal representation and the unique causal representation extracted by the causal relationship learning module for the multimedia files contained in the sample pair under the same modality.

12. A multimedia file recommendation device, characterized in that: The device comprises: an acquiring unit configured to acquire a historical behavior sequence of a target user and a plurality of candidate multimedia files, wherein the historical behavior sequence includes one or more multimedia files with which the target user has historical interactive behaviors; a representation extraction unit configured to extract a causal representation of the historical behavior sequence in at least one modality and a causal representation of each candidate multimedia file in at least one modality, wherein the causal representation reflects a causal relationship between the causal representation and the target user's preference; obtain a user representation of the target user based on the causal representation of the historical behavior sequence in at least one modality, and obtain a file representation of each candidate multimedia file based on the causal representation of each candidate multimedia file in at least one modality; The recommendation unit is configured to select at least one candidate multimedia file from the candidate multimedia files as a recommendation result for the target user based on the user representation of the target user and the file representations of the candidate multimedia files.

13. A device for training a recommendation model, the device comprising: a data acquisition unit configured to acquire training data comprising a plurality of sample pairs, the sample pairs comprising a positive sample pair and a negative sample pair, the positive sample pair comprising a user sample and a multimedia sample with which the user sample has a history of interaction, and the negative sample pair comprising the user sample and a multimedia sample with which the user sample has no history of interaction; A training unit is configured to train the recommendation model using the training data, wherein a historical behavior sequence of the user sample in the sample pair and a multimedia sample in the sample pair are obtained, the historical behavior sequence including one or more multimedia files having historical interactive behaviors with the user sample; the recommendation model extracts a causal representation of the historical behavior sequence in at least one modality and a causal representation of each multimedia sample in at least one modality, the causal representation reflecting the causal relationship between the causal representation and the user preference; the recommendation model obtains a user representation of the user sample based on the causal representation of the historical behavior sequence in at least one modality, and obtains a file representation of the multimedia sample based on the causal representation of the multimedia sample in at least one modality; the recommendation model determines a preference score of the multimedia sample based on the user representation of the user sample and the file representation of the multimedia sample; The training objective includes maximizing the difference between the preference scores of the multimedia samples included in the positive sample pair and the preference scores of the multimedia samples included in the negative sample pair in the sample pairs.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 11 are implemented.

15. An electronic device, characterized in that: include: one or more processors; as well as A memory associated with the one or more processors, the memory being used to store program instructions, wherein the program instructions, when read and executed by the one or more processors, perform the steps of the method according to any one of claims 1 to 11.

16. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 11 are implemented.

Images