Generative recommendation method and device of video, storage medium and electronic equipment
By filtering, deduplicating, and fusing user behavior logs, high-quality video interaction sequences are generated and input into a generative recommendation model. This solves the problem of sparse and complex user behavior in long video platforms and improves the accuracy and resource efficiency of personalized recommendations.
Patent Information
- Application Number
- CN202511808840.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-03
- Publication Date
- 2026-02-17
AI Technical Summary
In long-form video platforms, user behavior is sparse and complex, making it difficult for traditional recommendation systems to identify users' potential interests, resulting in poor recommendation performance.
By acquiring the latest and historical user behavior logs of target users, filtering, deduplication, information fusion, and screening processes are performed to generate high-quality video interaction sequences, which are then input into a generative recommendation model to generate recommended video sequences.
It improves the accuracy and personalization of recommendations, reduces resource consumption, better portrays the evolution of user interests, and solves the problem of sparse and complex user behavior, making it difficult to identify potential interests.
Smart Images
Figure CN121547645A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of multimedia intelligent processing technology, and in particular to a generative recommendation method, apparatus, storage medium, and electronic device for video. Background Technology
[0002] Personalized recommendations are a core capability for improving user viewing efficiency on long-form online video platforms, but this scenario presents significant technical challenges. First, users accumulate vast amounts of viewing history over extended periods, forming long and complex behavioral sequences that result in multi-stage and multi-layered changes in user interests. Second, long-form video content itself is rich in dimensions, involving various knowledge tags such as themes, plots, actors, and styles, making its content system far more complex than that of short videos and product recommendations. Traditional recommendation systems generally employ a "matching-rating" model, but in long-form video scenarios, user behavior is naturally sparse and has a high noise level; short-term viewing and browsing via shortcuts can easily interfere with the recommendation model's interest judgment. Furthermore, traditional methods such as collaborative filtering, by emphasizing similar behaviors, continuously reinforce existing user interests, making it difficult to break through "information cocoons" and uncover deeper or cross-category potential preferences. Summary of the Invention
[0003] This application provides a generative recommendation method, apparatus, storage medium, and electronic device for videos to address the technical problem of sparse and complex user behavior, making it difficult to identify potential interests.
[0004] Firstly, this application provides a generative recommendation method for videos, comprising: upon determining that a target user has generated valid viewing behavior after a target time point, obtaining the latest user behavior log of the target user, wherein the latest user behavior log comprises all valid viewing behavior data generated by the target user after the target time point; obtaining the historical user behavior log of the target user, and performing filtering, deduplication, information fusion, and screening processing on the historical user behavior log to obtain a historical video interaction sequence for each video, wherein the historical user behavior log comprises all viewing behavior data generated by the target user within a target time period before the target time point; fusing the latest user behavior log and all historical video interaction sequences to obtain a latest video interaction sequence for each video; inputting all latest video interaction sequences into a generative recommendation model, so that the generative recommendation model generates a recommended video sequence for the target user based on all latest video interaction sequences, and saving the recommended video sequence to a database.
[0005] Secondly, this application provides a generative recommendation device for videos, comprising: a first acquisition module, configured to acquire the latest user behavior log of the target user when it is determined that the target user has generated effective viewing behavior after a target time point, wherein the latest user behavior log is all effective viewing behavior data generated by the target user after the target time point; a second acquisition module, configured to acquire the historical user behavior log of the target user, and perform filtering, deduplication, information fusion, and screening processing on the historical user behavior log to obtain a historical video interaction sequence for each video, wherein the historical user behavior log is all viewing behavior data generated by the target user within a target time period before the target time point; a fusion module, configured to fuse the latest user behavior log and all historical video interaction sequences to obtain a latest video interaction sequence for each video; and a generation module, configured to input all the latest video interaction sequences into a generative recommendation model, so that the generative recommendation model generates a recommended video sequence for the target user based on all the latest video interaction sequences, and saves the recommended video sequence to a database.
[0006] As an optional example, the second acquisition module includes: an acquisition unit, configured to acquire the viewing duration of each historical user behavior in the historical user behavior log, wherein the historical user behavior includes a user identifier, a video identifier, a viewing start time, and a viewing duration; a first determination unit, configured to determine that the target historical user behavior is an invalid viewing behavior if the viewing duration of the target historical user behavior is less than a preset duration threshold, wherein the target historical user behavior is any historical user behavior in the historical user behavior log; and a first filtering unit, configured to filter out all invalid viewing behaviors in the historical user behavior log to obtain a filtered historical user behavior log.
[0007] As an optional example, the second acquisition module includes: a merging unit, used to merge all historical user behaviors of each video in the filtered historical user behavior log to obtain the corresponding historical video interaction sequence; an accumulation unit, used to accumulate the viewing duration of all historical user behaviors of each video to obtain the total viewing duration of the corresponding historical video interaction sequence; a second determination unit, used to determine the last viewing timestamp among all historical user behaviors of each video as the final timestamp of the corresponding historical video interaction sequence; and a statistics unit, used to count the number of historical user behaviors of each video to obtain the effective viewing count of the corresponding historical video interaction sequence.
[0008] As an optional example, the second acquisition module described above includes: a second filtering unit, used to filter out historical video interaction sequences of each video whose sequence length is less than a preset length threshold or whose effective viewing count is less than a preset number of times threshold.
[0009] As an optional example, the above-mentioned fusion module includes: a processing unit, used to determine the unprocessed historical video interaction sequence as the current historical video interaction sequence, and to perform the following processing on the current historical video interaction sequence: obtain the latest user behavior of the video corresponding to the current historical video interaction sequence from the latest user behavior log, merge it with the current historical video interaction sequence, and update the total viewing time, final timestamp and valid viewing count to obtain the latest video interaction sequence of the video corresponding to the current historical video interaction sequence.
[0010] As an optional example, the above-mentioned generation module includes: an encoding unit, used to organize all the latest video interaction sequences in order of their final timestamps and encode them into an input feature vector that can be accepted by the above-mentioned generative recommendation model; and a generation unit, used to input the above-mentioned input feature vector into the above-mentioned generative recommendation model, so that the above-mentioned generative recommendation model performs contextual modeling based on all the latest video interaction sequences and generates a recommended video sequence that reflects the future interests of the target user.
[0011] As an optional example, the above apparatus further includes: a merging module, used to merge the recommended video sequence with the recall results of other recall sources after saving the recommended video sequence to the database to form a candidate video set; and a sorting module, used to score and sort all candidate videos in the candidate video set to generate a recommended video list for the target user.
[0012] Thirdly, this application provides a storage medium storing a computer program, wherein the computer program is executed by a processor to perform the generative recommendation method for the aforementioned video.
[0013] Fourthly, this application also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to execute the above-described video generative recommendation method through the computer program.
[0014] The technical solutions provided in this application have the following advantages compared with the prior art: This application employs the following method: First, upon determining that a target user has engaged in valid viewing behavior after a target time point, the application obtains the latest user behavior log of the target user, where the latest user behavior log comprises all valid viewing behavior data generated by the target user after the target time point. Second, it obtains the target user's historical user behavior log and performs filtering, deduplication, information fusion, and screening on the historical user behavior log to obtain a historical video interaction sequence for each video, where the historical user behavior log comprises all viewing behavior data generated by the target user within a target time period before the target time point. Third, it fuses the latest user behavior log and all historical video interaction sequences to obtain a latest video interaction sequence for each video. Fourth, it inputs all the latest video interaction sequences into a generative recommendation model, enabling the generative recommendation model to generate recommended video sequences for the target user based on all the latest video interaction sequences, and saves the recommended video sequences to a database. This method, by monitoring new valid viewing behavior of the user after the target time point, triggers incremental recommendation generation when the user's interests change. This approach performs effective viewing filtering, deduplication, information fusion, and sequence selection on historical user behavior logs to construct high-quality long sequences. These sequences are then merged with the latest viewing behavior to form the user's latest video interaction sequence. This sequence is fed into a generative recommendation model to generate candidate video sequences for the target user and stored for online retrieval. This significantly reduces resource overhead while maintaining recommendation accuracy, better characterizes the evolution of user interests, and improves personalized recommendation capabilities. Ultimately, it solves the technical problem of sparse and complex user behavior, making it difficult to identify potential interests. Attached Figure Description
[0015] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.
[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0017] One or more embodiments are illustrated by way of example with reference numerals in the accompanying drawings. These illustrations do not constitute a limitation on the embodiments. Elements with the same reference numerals in the drawings are denoted as similar elements. Unless otherwise stated, the figures in the drawings are not to be limited by scale.
[0018] Figure 1This is a flowchart of an optional generative recommendation method for videos according to an embodiment of this application; Figure 2 This is a flowchart illustrating a specific implementation of an optional generative recommendation method for videos according to an embodiment of this application. Figure 3 This is a schematic diagram of the structure of an optional video generative recommendation device according to an embodiment of this application; Figure 4 This is a schematic diagram of an optional electronic device according to an embodiment of this application. Detailed Implementation
[0019] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0020] The following disclosure provides numerous different embodiments or examples for implementing various structures of this application. To simplify the disclosure, specific examples of components and arrangements are described below. These are merely examples and are not intended to limit the scope of this application. Furthermore, reference numerals and / or letters may be repeated in different examples. Such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.
[0021] According to a first aspect of the embodiments of this application, a generative recommendation method for videos is provided, optionally, as follows: Figure 1 As shown, the above method includes: S102, if it is determined that the target user has generated effective movie-watching behavior after the target time point, obtain the latest user behavior log of the target user, wherein the latest user behavior log is all effective movie-watching behavior data generated by the target user after the target time point; S104, obtain the target user's historical user behavior logs, and perform filtering, deduplication, information fusion and screening on the historical user behavior logs to obtain the historical video interaction sequence of each video. The historical user behavior logs are all the viewing behavior data generated by the target user in the target time period before the target time point. S106, integrate the latest user behavior logs and all historical video interaction sequences to obtain the latest video interaction sequence for each video; S108: Input all the latest video interaction sequences into the generative recommendation model so that the generative recommendation model can generate recommended video sequences for the target user based on all the latest video interaction sequences, and save the recommended video sequences to the database.
[0022] Optionally, in this embodiment, a generative recommendation method for long-form video platforms is proposed. By introducing a sequence generation paradigm and refined data processing strategies, it achieves deep modeling of user interests and the generation of high-quality recommendation results. Specifically, such as... Figure 2 The specific implementation flowchart is shown below, including the following steps: First, the platform continuously monitors new viewing behavior of users and determines whether users have generated new valid viewing records based on a preset time window (e.g., 24 hours). Viewing behavior with a single viewing duration greater than or equal to a preset duration threshold is considered valid viewing behavior. Only when a user has new valid viewing behavior within a certain period of time (or after a target time point, such as a certain time each day) is the user marked as a target user, thus triggering the generative recommendation process in this embodiment. This avoids redundant calculations for all users and reduces system resource consumption.
[0023] After determining that the target user meets the triggering conditions, the latest user behavior log generated within the preset time window or after the target time point is obtained. This log contains only behaviors that meet the definition of valid movie watching. The latest behavior log is used to reflect the user's recent interest dynamics and provide incremental behavior information for recommendation generation.
[0024] Subsequently, historical user behavior logs prior to the target time point are obtained. To ensure the quality of the generative model input, multi-level preprocessing is performed on the historical logs, including: filtering invalid viewing behaviors, deduplicating multiple viewing behaviors of the same video, fusing timestamps and cumulative viewing duration information, and selecting samples based on rules such as minimum sequence length and valid number of video viewings, thereby constructing a historical video interaction sequence with high information density and low noise.
[0025] Next, the latest user behavior logs are fused with historical video interaction sequences to form the target user's latest video interaction sequence. This sequence fully presents the user's historical interest accumulation and recent interest changes, and can serve as a high-quality input context for generative models.
[0026] Finally, the latest video interaction sequence is input into the generative recommendation model. Based on a sequence generation mechanism, the generative model predicts the video sequences a user might watch in the future from the user's complete behavioral context, generating recommended video sequences for the target user. The recommendation results are ultimately written to a database, serving as the "generative recall source" for the online recall layer, participating in ranking and final recommendation display along with other recall results.
[0027] Optionally, in this embodiment, by introducing a "user new behavior monitoring + trigger-based incremental generation" mechanism, redundant calculations for all users are avoided, significantly reducing offline resource consumption and improving recommendation generation efficiency. Combined with strategies such as effective viewing definitions, deduplication and fusion, and sample selection, the quality of user behavior sequences is significantly improved, enabling the generative model to more accurately capture user interest evolution trends and achieve deep interest mining. Ultimately, this not only enhances the personalization and coverage of recommendations but also effectively alleviates the information cocoon problem, providing a resource-efficient and high-performing generative recommendation technology solution for long-form video platforms.
[0028] As an optional example, filtering, deduplication, information fusion, and screening of invalid viewing behaviors are performed on historical user behavior logs to obtain the historical video interaction sequence for each video, including: Obtain the viewing duration of each historical user behavior record from the historical user behavior log. The historical user behavior records include user ID, video ID, viewing start time, and viewing duration. If the viewing duration of the target historical user behavior is less than a preset duration threshold, the target historical user behavior is determined to be an invalid viewing behavior. The target historical user behavior is any historical user behavior in the historical user behavior log. Filter out all invalid movie-watching behaviors from the historical user behavior logs to obtain the filtered historical user behavior logs.
[0029] Optionally, in this embodiment, a high-quality historical video interaction sequence is generated by performing validity determination, deduplication and fusion, and multi-level filtering on the user's historical behavior logs. Specifically, firstly, the viewing duration of each behavior record in the historical user behavior log is obtained. The historical behavior record includes fields such as user identifier, video identifier, viewing start time, and viewing duration. Then, the validity of each behavior is determined according to a preset duration threshold: when the viewing duration of the target historical behavior is less than the threshold, it is determined as an invalid viewing behavior. All behavior records in the historical log that are determined to be invalid viewing are filtered to obtain a filtered historical user behavior log containing only valid viewing records.
[0030] As an optional example, filtering, deduplication, information fusion, and screening of invalid viewing behaviors are performed on historical user behavior logs to obtain the historical video interaction sequence for each video, including: The historical user behaviors of each video in the filtered historical user behavior logs are merged to obtain the corresponding historical video interaction sequence. The total viewing time of all historical user behaviors for each video is summed up to obtain the total viewing time of the corresponding historical video interaction sequence. The last viewing timestamp of all historical user actions for each video is determined as the final timestamp of the corresponding historical video interaction sequence; Count the number of historical user behaviors for each video to obtain the effective viewing counts for the corresponding historical video interaction sequence.
[0031] Optionally, in this embodiment, high-quality historical video interaction sequences are generated by filtering, deduplicating, fusing information, and screening historical user behavior logs, providing reliable input data for generative recommendation models. Specifically, firstly, ineffective viewing behavior is filtered from the historical user behavior logs, removing behaviors with viewing durations below a preset threshold, resulting in filtered historical logs containing only effective viewing records. Subsequently, the filtered historical logs are merged by video dimension, summarizing all historical user behavior records for the same video to form corresponding video historical interaction sequences. During the merging process, the viewing duration of all historical viewing behaviors for each video is accumulated to obtain the total viewing duration of the video's historical interaction sequence. Simultaneously, the timestamp of the last viewing behavior in all historical behaviors for that video is extracted as the final timestamp of the historical interaction sequence. Furthermore, the number of historical user behaviors for each video is counted to obtain the effective viewing counts for subsequent sample screening and weight calculation. Through this series of processes, the historical interaction sequence of each video not only reflects the user's true interests and preferences but also takes into account viewing intensity, recent interests, and behavior coverage, providing complete, accurate, and low-noise historical behavior input for generative recommendation models.
[0032] Optionally, in this embodiment, by refining the processing of historical user behavior, invalid or noisy behaviors are effectively eliminated, multiple viewing information is integrated, and key indicators are statistically analyzed, significantly improving the quality and information density of users' historical interest sequences. This enables generative recommendation models to more accurately capture the evolution trend of user interests, improve the accuracy, reliability, and personalization of recommendation results, and provide a stable and efficient data foundation for subsequent incremental updates and generative recommendations.
[0033] As an optional example, filtering, deduplication, information fusion, and screening of invalid viewing behaviors are performed on historical user behavior logs to obtain the historical video interaction sequence for each video, including: Filter out historical video interaction sequences for each video whose sequence length is less than a preset length threshold or whose effective viewing count is less than a preset number of times threshold.
[0034] Optionally, in this embodiment, a dual screening rule is further applied to the generated historical video interaction sequences: sequences with a length less than a preset length threshold or a number of valid viewings less than a preset number of viewings are filtered out to ensure that the historical interaction sequences used for training and inference have sufficient data volume and representativeness.
[0035] Optionally, through the above multi-layered processing, each retained video history interaction sequence not only reflects the user's real interests, but also takes into account viewing intensity, behavioral coverage and temporal information, providing low-noise, high-information-density input for generative recommendation models.
[0036] Optionally, by filtering low-quality and sparse historical interaction sequences, the quality and effectiveness of data input are improved, and noise interference is avoided in model training and recommendation generation. At the same time, by retaining sequences whose length and number of viewings reach the threshold, the generative recommendation model can be guaranteed to learn based on complete and stable historical interest information, which significantly improves the accuracy, personalization level and user satisfaction of recommendation results, while enhancing the model's ability to mine long-tail content and potential interests.
[0037] As an optional example, information is fused from the latest user behavior logs and all historical video interaction sequences to obtain the latest video interaction sequence for each video, including: The unprocessed historical video interaction sequence is identified as the current historical video interaction sequence, and the following processing is performed on the current historical video interaction sequence: The latest user behavior for the video corresponding to the current historical video interaction sequence is obtained from the latest user behavior log and merged with the current historical video interaction sequence. At the same time, the total viewing time, final timestamp, and effective viewing count are updated to obtain the latest video interaction sequence for the video corresponding to the current historical video interaction sequence.
[0038] Optionally, in this embodiment, a high-quality latest video interaction sequence is generated by fusing the user's latest behavior with historical video interaction sequences to more accurately reflect the evolution of user interests. Specifically, each historical video interaction sequence is processed one by one. The latest viewing behavior of the video corresponding to the current historical video interaction sequence is obtained from the latest user behavior log, and this latest behavior is merged with the current historical video interaction sequence. During the merging process, the total viewing time of the video is updated, the accumulated viewing time is added to the duration of the newly added viewing behavior, and the final timestamp is updated to the time point of the latest viewing behavior. The number of valid viewings is accumulated and counted to obtain the latest video interaction sequence corresponding to each video. By performing the above processing on all historical video interaction sequences, a complete set of the user's latest video interaction sequences is generated, comprehensively integrating the user's historical interests and latest behavioral dynamics. By fusing historical and latest user behavior, the long-term accumulation and recent changes of user interests can be fully reflected, significantly improving the generative recommendation model's ability to understand user behavior.
[0039] As an optional example, all the latest video interaction sequences are input into the generative recommendation model, so that the generative recommendation model generates recommended video sequences for the target user based on all the latest video interaction sequences, including: All the latest video interaction sequences are organized in order of their final timestamps and encoded into input feature vectors that can be accepted by the generative recommendation model; The input feature vector is fed into the generative recommendation model, which then performs contextual modeling based on all the latest video interaction sequences to generate recommended video sequences that reflect the future interests of the target user.
[0040] Optionally, in this embodiment, a deep understanding of user interests and prediction of future preferences are achieved by inputting the user's latest video interaction sequence into the generative recommendation model. Specifically, firstly, all the latest video interaction sequences are sorted according to the final viewing timestamp, presenting the complete evolution of the user's recent interests. Then, the sorted sequences are encoded into input feature vectors that the generative recommendation model can accept. These feature vectors include not only video identifiers, cumulative viewing time, and the latest timestamp, but also video-related information such as topic tags and channel information to enrich the semantic representation of the model input. Next, the encoded input feature vectors are fed into the generative recommendation model. The generative recommendation model performs contextual modeling based on all the latest video interaction sequences, capturing the long-term accumulation and recent dynamics of user interests to generate video sequences that reflect the target user's potential future interests. The generated recommended video sequences can be directly stored in a database for subsequent online recall and ranking layer processing, achieving efficient output of personalized recommendations.
[0041] Optionally, in this embodiment, by fine-grained encoding and generative modeling of the latest video interaction sequences, the evolutionary trends of user interests and potential preferences can be accurately captured, improving the timeliness, accuracy, and personalization of recommendations. Simultaneously, by combining sequence ranking and contextual modeling methods, the recommendation results can cover diverse user interests, achieving deep interest mining and long-tail content recommendation.
[0042] As an optional example, after saving the recommended video sequences to the database, the above method also includes: The recommended video sequences are merged with the recall results from other recall sources to form a candidate video set; All candidate videos in the candidate video set are scored and sorted to generate a recommended video list for the target user.
[0043] Optionally, in this embodiment, after generating the recommended video sequence and saving it to the database, the candidate videos are further integrated and sorted to form the final personalized recommendation list. Specifically, the recommended video sequence output by the generative recommendation model is first merged with candidate video sets from other recall sources in the platform, including collaborative filtering recall, popular recommendations, content-based recall, and other sources, to form a comprehensive candidate video set. Subsequently, all candidate videos in the candidate video set are scored and sorted. The scoring process can be based on multi-dimensional features, including user historical behavior and interest matching degree, video content features, viewing duration preference, time sensitivity, and other available sorting signals. The sorting module will comprehensively consider various features and weights, and generate the final recommended video list for the target user through a unified sorting algorithm, achieving personalization and accuracy of the recommendation results. The sorted recommendation list can be directly used for online display, providing users with interactive recommended content.
[0044] Optionally, in this embodiment, by fusing the generative recommendation results with candidate videos from other recall sources and employing a multi-dimensional scoring and ranking mechanism, not only is the recommendation coverage and diversity improved, but the accuracy and personalization of the recommendations are also enhanced.
[0045] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.
[0046] According to another aspect of the embodiments of this application, a generative recommendation apparatus for videos is also provided, such as... Figure 3 As shown, it includes: The first acquisition module 302 is used to acquire the latest user behavior log of the target user when it is determined that the target user has generated effective movie-watching behavior after the target time point. The latest user behavior log is all effective movie-watching behavior data generated by the target user after the target time point. The second acquisition module 304 is used to acquire the historical user behavior logs of the target user, and to perform filtering, deduplication, information fusion and screening on the historical user behavior logs to obtain the historical video interaction sequence of each video. The historical user behavior logs are all the viewing behavior data generated by the target user in the target time period before the target time point. The fusion module 306 is used to fuse the latest user behavior logs and all historical video interaction sequences to obtain the latest video interaction sequence for each video. The generation module 308 is used to input all the latest video interaction sequences into the generative recommendation model, so that the generative recommendation model can generate recommended video sequences for the target user based on all the latest video interaction sequences, and save the recommended video sequences to the database.
[0047] It should be noted that the first acquisition module 302 in this embodiment can be used to execute step S102 in this application embodiment, the second acquisition module 304 in this embodiment can be used to execute step S104 in this application embodiment, the fusion module 306 in this embodiment can be used to execute step S106 in this application embodiment, and the generation module 308 in this embodiment can be used to execute step S108 in this application embodiment.
[0048] As an optional example, the second acquisition module includes: The acquisition unit is used to acquire the viewing duration of each historical user behavior in the historical user behavior log. The historical user behavior includes user identifier, video identifier, viewing start time, and viewing duration. The first determining unit is used to determine that the target historical user behavior is an invalid viewing behavior when the viewing duration of the target historical user behavior is less than a preset duration threshold, wherein the target historical user behavior is any historical user behavior in the historical user behavior log. The first filtering unit is used to filter out all invalid viewing behaviors in the historical user behavior logs, resulting in filtered historical user behavior logs.
[0049] As an optional example, the second acquisition module includes: The merging unit is used to merge all historical user behaviors of each video in the filtered historical user behavior log to obtain the corresponding historical video interaction sequence. The accumulation unit is used to accumulate the viewing time of all historical user behaviors for each video to obtain the total viewing time of the corresponding historical video interaction sequence; The second determining unit is used to determine the last viewing timestamp of all historical user behaviors of each video as the final timestamp of the corresponding historical video interaction sequence. The statistics unit is used to count the number of historical user behaviors for each video, and to obtain the effective viewing count of the corresponding historical video interaction sequence.
[0050] As an optional example, the second acquisition module includes: The second filtering unit is used to filter out historical video interaction sequences of each video whose sequence length is less than a preset length threshold or whose effective viewing count is less than a preset number of times threshold.
[0051] As an optional example, the fusion module includes: The processing unit is used to determine the unprocessed historical video interaction sequence as the current historical video interaction sequence, and to perform the following processing on the current historical video interaction sequence: The latest user behavior for the video corresponding to the current historical video interaction sequence is obtained from the latest user behavior log and merged with the current historical video interaction sequence. At the same time, the total viewing time, final timestamp, and effective viewing count are updated to obtain the latest video interaction sequence for the video corresponding to the current historical video interaction sequence.
[0052] As an optional example, the generated modules include: The encoding unit is used to organize all the latest video interaction sequences in order of their final timestamps and encode them into an input feature vector that can be accepted by the generative recommendation model. The generation unit is used to input the input feature vector into the generative recommendation model, so that the generative recommendation model can perform contextual modeling based on all the latest video interaction sequences and generate recommended video sequences that reflect the future interests of the target user.
[0053] As an optional example, the above-described apparatus further includes: The merging module is used to merge the recommended video sequences with the recall results of other recall sources after saving the recommended video sequences to the database, forming a candidate video set. The sorting module is used to score and sort all candidate videos in the candidate video set, and generate a recommended video list for the target user.
[0054] For other examples of this embodiment, please refer to the examples above, which will not be repeated here.
[0055] Figure 4This is a schematic diagram of an optional electronic device according to an embodiment of this application, such as... Figure 4 As shown, it includes a processor 402, a communication interface 404, a memory 406, and a communication bus 408. The processor 402, communication interface 404, and memory 406 communicate with each other via the communication bus 408. Memory 406 is used to store computer programs; When processor 402 executes a computer program stored in memory 406, it performs the following steps: If it is determined that the target user has generated effective movie-watching behavior after the target time point, the latest user behavior log of the target user is obtained. The latest user behavior log is all effective movie-watching behavior data generated by the target user after the target time point. Obtain the target user's historical user behavior logs, and perform filtering, deduplication, information fusion, and screening on the historical user behavior logs to obtain the historical video interaction sequence for each video. The historical user behavior logs are all the viewing behavior data generated by the target user within the target time period before the target time point. By fusing the latest user behavior logs and all historical video interaction sequences, the latest video interaction sequence for each video is obtained. All the latest video interaction sequences are input into the generative recommendation model, so that the generative recommendation model can generate recommended video sequences for the target user based on all the latest video interaction sequences, and save the recommended video sequences to the database.
[0056] Optionally, in this embodiment, the communication bus can be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. This communication bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 4 The symbol is represented by a single thick line, but this does not indicate that there is only one bus or one type of bus. The communication interface is used for communication between the aforementioned electronic devices and other devices.
[0057] The memory may include RAM, or non-volatile memory, such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.
[0058] As an example, the memory 406 described above may include, but is not limited to, the first acquisition module 302, the second acquisition module 304, the fusion module 306, and the generation module 308 of the generative recommendation device for the video described above. Furthermore, it may include, but is not limited to, other module units of the generative recommendation device for the video described above, which will not be elaborated upon in this example.
[0059] The processor mentioned above can be a general-purpose processor, including but not limited to: CPU (Central Processing Unit), NP (Network Processor), etc.; it can also be DSP (Digital Signal Processor), ASIC (Application Specific Integrated Circuit), FPGA (Field-Programmable Gate Array) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.
[0060] Optionally, specific examples in this embodiment can refer to the examples described in the above embodiments, and will not be repeated here.
[0061] Those skilled in the art will understand that Figure 4 The structure shown is for illustrative purposes only. The device implementing the above generative recommendation method for videos can be a terminal device, such as a smartphone (e.g., Android phone, iOS phone), tablet computer, PDA, mobile internet device (MID), PAD, etc. Figure 4 This does not limit the structure of the aforementioned electronic devices. For example, the electronic device may also include components that are more... Figure 4 The more or fewer components shown (such as network interfaces, display devices, etc.), or having the same Figure 4 The different configurations shown.
[0062] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing the hardware related to the terminal device. The program can be stored in a computer-readable storage medium, which may include: flash drive, ROM, RAM, disk or optical disk, etc.
[0063] According to another aspect of the embodiments of this application, a computer-readable storage medium is also provided, wherein a computer program is stored in the computer program, which, when executed by a processor, performs the steps in the above-described generative recommendation method for videos.
[0064] Optionally, in this embodiment, those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing the hardware related to the terminal device. The program can be stored in a computer-readable storage medium, which may include: flash drive, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.
[0065] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.
[0066] If the integrated units in the above embodiments are implemented as software functional units and sold or used as independent products, they can be stored in the aforementioned computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause one or more computer devices (which may be personal computers, servers, or network devices, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application.
[0067] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.
[0068] In the several embodiments provided in this application, it should be understood that the disclosed client can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, indirect coupling or communication connection between units or modules, and may be electrical or other forms.
[0069] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.
[0070] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.
[0071] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.
Claims
1. A generative recommendation method of a video, characterized by, Comprise: After determining that the target user has generated valid viewing behavior at the target time point, obtain the latest user behavior log of the target user, wherein the latest user behavior log is all valid viewing behavior data generated by the target user after the target time point; Obtain the historical user behavior log of the target user, and perform filtering, deduplication, information fusion, and screening processing on the non-effective viewing behavior of the historical user behavior log to obtain a historical video interaction sequence of each video, wherein the historical user behavior log is all viewing behavior data generated by the target user within a target time period before the target time point; Fuse the latest user behavior log and all historical video interaction sequences to obtain a latest video interaction sequence of each video; Input all latest video interaction sequences into a generative recommendation model to enable the generative recommendation model to generate a recommended video sequence for the target user according to all latest video interaction sequences, and save the recommended video sequence to a database.
2. The method of claim 1, wherein, The filtering, deduplication, information fusion, and screening processing on the non-effective viewing behavior of the historical user behavior log to obtain a historical video interaction sequence of each video comprises: Obtain the viewing duration of each historical user behavior in the historical user behavior log, wherein the historical user behavior comprises a user identifier, a video identifier, a viewing start time point, and a viewing duration; In a case where the viewing duration of a target historical user behavior is less than a preset duration threshold, determine that the target historical user behavior is a non-effective viewing behavior, wherein the target historical user behavior is any historical user behavior in the historical user behavior log; Filter out all non-effective viewing behaviors in the historical user behavior log to obtain a filtered historical user behavior log.
3. The method of claim 2, wherein, The filtering, deduplication, information fusion, and screening processing on the non-effective viewing behavior of the historical user behavior log to obtain a historical video interaction sequence of each video comprises: Merge all historical user behaviors of each video in the filtered historical user behavior log to obtain a corresponding historical video interaction sequence; Add up the viewing durations of all historical user behaviors of each video to obtain the total viewing duration of the corresponding historical video interaction sequence; Determine the last viewing timestamp in all historical user behaviors of each video as the final timestamp of the corresponding historical video interaction sequence; Statistically count the number of historical user behaviors of each video to obtain the effective viewing frequency of the corresponding historical video interaction sequence.
4. The method of claim 3, wherein, The filtering, deduplication, information fusion, and screening processing on the non-effective viewing behavior of the historical user behavior log to obtain a historical video interaction sequence of each video comprises: Filter out historical video interaction sequences in each video's historical video interaction sequence that have a sequence length less than a preset length threshold or an effective viewing frequency less than a preset frequency threshold.
5. The method of claim 4, wherein, The fusing of the latest user behavior log and all historical video interaction sequences to obtain a latest video interaction sequence of each video comprises: The untreated historical video interaction sequence is determined as a current historical video interaction sequence, and the following processing is performed on the current historical video interaction sequence: The latest user behavior of the video corresponding to the current historical video interaction sequence is obtained from the latest user behavior log, and is merged with the current historical video interaction sequence, while updating the total viewing time, the final timestamp and the valid viewing number, to obtain the latest video interaction sequence of the video corresponding to the current historical video interaction sequence.
6. The method of claim 1, wherein, All the latest video interaction sequences are input into a generative recommendation model, so that the generative recommendation model generates a recommended video sequence for the target user according to all the latest video interaction sequences, including: All the latest video interaction sequences are organized in the order of the final timestamp, and are encoded into an input feature vector that can be received by the generative recommendation model; The input feature vector is input into the generative recommendation model, so that the generative recommendation model performs context modeling based on all the latest video interaction sequences, and generates a recommended video sequence reflecting the future interest of the target user.
7. The method according to any one of claims 1 to 6, characterized in that, After saving the recommended video sequence into the database, the method further includes: The recommended video sequence and the recall result of other recall sources are merged to form a candidate video set; All candidate videos in the candidate video set are scored and sorted to generate a recommended video list of the target user.
8. A generative recommendation apparatus for videos, characterized by, It includes: The first acquisition module is configured to acquire the latest user behavior log of the target user in the case that the target user has valid viewing behavior after the target time point, wherein the latest user behavior log is all valid viewing behavior data of the target user generated after the target time point; The second acquisition module is configured to acquire the historical user behavior log of the target user, and perform filtering, deduplication, information fusion and screening processing on the historical user behavior log to obtain a historical video interaction sequence of each video, wherein the historical user behavior log is all viewing behavior data of the target user generated in a target time period before the target time point; The fusion module is configured to perform information fusion on the latest user behavior log and all historical video interaction sequences to obtain a latest video interaction sequence of each video; The generation module is configured to input all the latest video interaction sequences into a generative recommendation model, so that the generative recommendation model generates a recommended video sequence for the target user according to all the latest video interaction sequences, and saves the recommended video sequence into a database.
9. A computer-readable storage medium storing a computer program, the computer-readable storage medium being characterized by, The computer program is run by the processor to perform the method described in any one of claims 1 to 7.
10. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to execute the method described in any one of claims 1 to 7 by the computer program. The memory stores a computer program, and the processor is configured to execute the method described in any one of claims 1 to 7 by the computer program.