On-demand content recommendation method and device, storage medium and electronic equipment

Abstract

The invention provides an on-demand content recommendation method and device, a storage medium and electronic equipment, and is applied to the technical field of artificial intelligence. According to the method, the live broadcast stream of the user in the current live broadcast channel is cut into the time sequence video slices in real time, and the slices are input into the pre-trained live broadcast identification model to obtain the current live broadcast content and state. And on the basis of the identification result, the on-demand program matched with the current live broadcast content can be positioned by using the constructed on-demand knowledge graph. And then, inputting the current channel identifier and the matched on-demand program identifier into an on-demand recommendation model so as to output on-demand content and an optimal user recommendation opportunity. And when the current live broadcast state is matched with the recommendation opportunity, pushing the on-demand content. According to the method and the device, the on-demand content can be accurately recommended at the key moment of the live broadcast by accurately identifying the live broadcast content and the state, dynamically matching the on-demand program, generating the on-demand content through the recommendation model and predicting the optimal pushing time, so that the conversion rate of the on-demand program is effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method, device, storage medium and electronic device for recommending on-demand content. Background Art

[0002] Current TV viewing trends show a high proportion of IPTV (Internet Protocol Television) users who exclusively watch live broadcasts, offering significant potential for converting these users into on-demand consumption. Data shows that approximately 70% of IPTV users consume content exclusively within live channels, with no on-demand viewing, indicating ample room for the transition from live broadcasts to on-demand consumption.

[0003] However, the particularity of live broadcast scenarios makes it difficult to effectively apply traditional content recommendation strategies, and users often cannot obtain relevant on-demand information in a timely manner.

[0004] Therefore, how to push personalized on-demand recommendations at the right time becomes the key to improving the overall on-demand rate of the platform. Summary of the Invention

[0005] In view of the above problems, the present invention provides a method, device, storage medium and electronic device for recommending on-demand content that overcomes or at least partially solves the above problems. The technical solution is as follows:

[0006] A method for recommending on-demand content, comprising:

[0007] Cut the live stream of the user end under the current live channel into time-series video slices in real time;

[0008] Inputting the time-series video slices into a pre-trained live broadcast recognition model to obtain a live broadcast recognition result output by the live broadcast recognition model, wherein the live broadcast recognition result includes the current live broadcast content and current live broadcast status of the live broadcast stream;

[0009] Using a pre-built on-demand knowledge graph, locate a first on-demand program that matches the current live content;

[0010] Inputting the channel identifier of the current live channel and the program identifier of the first on-demand program into a pre-trained on-demand recommendation model, and obtaining the first on-demand content and the corresponding user recommendation timing output by the on-demand recommendation model;

[0011] adding the first on-demand content and the user recommendation timing as on-demand recommendation data to the on-demand recommendation ranking result associated with the user terminal;

[0012] When the current live broadcast state matches the user recommendation timing, the first on-demand content is pushed to the user terminal, so that when the user triggers the first on-demand content displayed on the user terminal, the first on-demand program is played.

[0013] A device for recommending on-demand content includes: a live stream real-time cutting unit, a live recognition result obtaining unit, an on-demand program matching unit, an on-demand content recommendation opportunity obtaining unit, an on-demand recommendation data updating unit and an on-demand content recommendation unit.

[0014] The live stream real-time cutting unit is used to cut the live stream of the user terminal under the current live channel into time-series video slices in real time;

[0015] The live broadcast recognition result obtaining unit is configured to input the time-series video slice into a pre-trained live broadcast recognition model to obtain a live broadcast recognition result output by the live broadcast recognition model, wherein the live broadcast recognition result includes the current live broadcast content and current live broadcast status of the live broadcast stream;

[0016] The on-demand program matching unit is configured to locate a first on-demand program that matches the current live content using a pre-built on-demand knowledge graph;

[0017] The on-demand content recommendation opportunity obtaining unit is configured to input the channel identifier of the current live channel and the program identifier of the first on-demand program into a pre-trained on-demand recommendation model, and obtain the first on-demand content and the corresponding user recommendation opportunity output by the on-demand recommendation model;

[0018] The on-demand recommendation data updating unit is configured to add the first on-demand content and the user recommendation opportunity as on-demand recommendation data to the on-demand recommendation ranking result associated with the user terminal;

[0019] The on-demand content recommendation unit is configured to push the first on-demand content to the user terminal when the current live broadcast status matches the user recommendation timing, so as to jump to play the first on-demand program when the user triggers the first on-demand content displayed on the user terminal.

[0020] A computer-readable storage medium stores a program, which implements the on-demand content recommendation method when executed by a processor.

[0021] An electronic device comprises at least one processor, and at least one memory and a bus connected to the processor; wherein the processor and the memory communicate with each other via the bus; and the processor is used to call program instructions in the memory to execute the on-demand content recommendation method.

[0022] By means of the above technical solution, the present invention provides a method, device, storage medium and electronic device for recommending on-demand content, which cuts the live stream of the user terminal under the current live channel into time-series video slices in real time; inputs the time-series video slices into a pre-trained live recognition model to obtain the live recognition results output by the live recognition model, wherein the live recognition results include the current live content and the current live status of the live stream; uses a pre-built on-demand knowledge graph to locate the first on-demand program that matches the current live content; inputs the channel identifier of the current live channel and the program identifier of the first on-demand program into the pre-trained on-demand recommendation model to obtain the first on-demand content and the corresponding user recommendation timing output by the on-demand recommendation model; adds the first on-demand content and the user recommendation timing as on-demand recommendation data to the on-demand recommendation ranking result associated with the user terminal; and pushes the first on-demand content to the user terminal when the current live status matches the user recommendation timing, so that when the user triggers the first on-demand content displayed on the user terminal, the first on-demand program is played. The present invention accurately identifies live broadcast content and live broadcast status, dynamically matches and associates on-demand programs, and then uses an on-demand recommendation model to generate on-demand content and predict the optimal push timing. It can accurately recommend on-demand content at key moments of live broadcast, thereby effectively improving the conversion rate of on-demand programs.

[0023] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are specifically listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Various other advantages and benefits will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiment below. The accompanying drawings are for illustration purposes only and are not to be considered as limiting the present invention. The same reference symbols are used throughout the drawings to represent the same components. In the drawings:

[0025] Figure 1 A schematic diagram showing a flow chart of an implementation of a method for recommending on-demand content provided by an embodiment of the present invention;

[0026] Figure 2 A schematic diagram showing a flow chart of a first specific implementation of a method for recommending on-demand content provided by an embodiment of the present invention;

[0027] Figure 3 A schematic diagram showing a flow chart of a training process of a live broadcast status recognition model provided by an embodiment of the present invention;

[0028] Figure 4 A schematic diagram showing a flow chart of a training process of a live broadcast status recognition model provided by an embodiment of the present invention;

[0029] Figure 5 A schematic diagram showing a process flow of constructing an on-demand knowledge graph provided by an embodiment of the present invention;

[0030] Figure 6 A flow chart showing a second specific implementation of the on-demand content recommendation method provided by an embodiment of the present invention is shown;

[0031] Figure 7 A logical block diagram of a system for recommending on-demand content provided by an embodiment of the present invention is shown;

[0032] Figure 8 A schematic diagram showing the structure of an on-demand content recommendation device provided by an embodiment of the present invention is shown;

[0033] Figure 9 A schematic structural diagram of an electronic device provided by an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0034] Exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0035] Current TV viewing trends show a high proportion of IPTV (Internet Protocol Television) users who exclusively watch live streaming. This presents significant potential for guiding these users towards on-demand consumption. The unique characteristics of live streaming scenarios dictate the limitations of conventional recommendation methods. To effectively guide users towards on-demand consumption, targeted messaging recommendations must be delivered. This can include various methods, such as pop-ups and text messages.

[0036] The key to live streaming recommendations lies in accurately identifying the current live content and its status to minimize disruption to users. This requires controlling the frequency of recommendations and achieving higher accuracy within a limited number of recommendations, thereby pushing content to users at the right time. For example, when a user is watching the ninth episode of the TV series "ABCD," the system needs to determine the best time to recommend a content. While the conventional approach might be to recommend after the current episode has finished playing, triggering the recommendation at the end of the ninth episode may be more effective, as it won't disrupt the user's viewing experience and will more easily capture their attention.

[0037] Current program guides typically only provide coarse-grained broadcast time information, so the system must analyze live streams in real time to accurately identify the current playback status. Furthermore, because live and on-demand versions may differ, the system must also analyze the content attributes of the live stream, such as dialogue characteristics, to accurately match it with on-demand content. Furthermore, if certain live content lacks corresponding media assets in the on-demand library, the system must analyze the characters, scenes, costumes, and dialogue in the live stream to identify similar on-demand content or activities for recommendation.

[0038] After achieving accurate live stream identification, the next step for the recommendation system is to provide more compelling content. For example, based on viewing patterns for the ninth episode of "ABCD," a standard recommendation might simply suggest watching the entire series. However, to change user viewing habits, the recommended content must be more engaging. By analyzing the plot's trajectory, the system can add captivating captions to the recommendations, such as "Visiting Chiang's mansion with clever banter, encountering old grudges and angrily rebuking the ugly deeds," to heighten user anticipation.

[0039] Based on this, in an embodiment of the present invention, a method for recommending on-demand content is provided. First, the live stream of the user on the current live channel is cut into time-series video slices in real time, and these slices are input into a pre-trained live recognition model to obtain the current live content and status. Based on the recognition results, the constructed on-demand knowledge graph can be used to locate the on-demand program that matches the current live content. Next, the current channel identifier and the matching on-demand program identifier are input into the on-demand recommendation model to output the on-demand content and the best user recommendation time. When the current live status matches the recommendation time, the on-demand content is pushed. It can be seen that the present invention can accurately recommend on-demand content at critical moments of live broadcasts by accurately identifying live content and status, dynamically matching on-demand programs, and then generating on-demand content and predicting the best push time through a recommendation model, thereby effectively improving the conversion rate of on-demand programs.

[0040] like Figure 1 FIG. 1 is a flow chart of an implementation of a method for recommending on-demand content provided by an embodiment of the present invention. The method may include:

[0041] S100: Cut the live stream of the user terminal under the current live channel into time-series video slices in real time.

[0042] The user end refers to the device or interface through which users receive and interact with content. In digital media and online services, this can include devices such as televisions, smartphones, tablets, and personal computers, as well as related software applications or web interfaces.

[0043] The current live broadcast channel refers to the channel where the live broadcast program that the user is watching is located. Each live broadcast channel usually has a specific content theme, and the user can watch the corresponding live broadcast content by selecting different channels.

[0044] Live streaming refers to the real-time transmission of audio and video data streams, which users receive and watch live programs over the Internet. Audio and video data streams are dynamic and contain real-time information about the program being played, including video images, audio, and subtitles.

[0045] Among them, time-series video slicing is to divide the live stream into several continuous small segments in chronological order, and each slice represents a specific time period in the live stream.

[0046] To facilitate understanding, let's use an example: Suppose a user is watching a live TV show on "XX Satellite TV," and the live stream contains the program's video and audio information. During this live broadcast, the live stream is monitored in real time and sliced ​​into time-series video slices every five minutes. For example, from the start of the live broadcast to the fifth minute, slice 1 is generated; from the fifth minute to the tenth minute, slice 2 is generated; and so on, until the end of the live broadcast. Each slice contains the program content for that time period, facilitating subsequent analysis, such as identifying the beginning and end of the show, or commercial breaks, to provide users with precise content recommendations.

[0047] S110: Input the time-series video slices into a pre-trained live broadcast recognition model to obtain a live broadcast recognition result output by the live broadcast recognition model, wherein the live broadcast recognition result includes the current live broadcast content and current live broadcast status of the live broadcast stream.

[0048] The live broadcast recognition model is a trained machine learning model designed to analyze and identify various information and statuses in real-time live streams. Based on a large amount of training data, the live broadcast recognition model can extract features from video and audio to accurately identify content and classify status.

[0049] Among them, the live broadcast recognition result is the output generated by the live broadcast recognition model after analyzing the input time-series video slices, including the current live broadcast content and current live broadcast status of the live broadcast stream.

[0050] The current live broadcast status refers to the status at a certain moment, which can be any of the playback links including the opening, ending, advertisement and main film.

[0051] The current live content refers to the specific content information recognized by the model at a certain moment in the live stream, including at least one of the lines, stars, content categories, and character costumes.

[0052] To facilitate understanding, let's use an example: Suppose a user is watching a live TV series. A time-series video slice (for example, from the 15th to 20th minute of the live broadcast) is extracted from the live stream. This slice is fed into a trained live broadcast recognition model. The model analyzes and recognizes the following information: the current live broadcast content includes "female lead Xiao Li is talking to male lead Xiao Zhang about their upcoming wedding" and "series category is romance drama." It also identifies the current live broadcast status as "main film."

[0053] S120: Utilize the pre-built on-demand knowledge graph to locate the first on-demand program that matches the current live broadcast content.

[0054] The on-demand knowledge graph is a structured data representation designed to organize and store information related to on-demand content, including relationships and attributes such as film and television works, actors, directors, categories, and plots. The on-demand knowledge graph can be built based on the open-source Neo4j framework. It facilitates rapid retrieval and recommendation of on-demand programs related to the user's current viewing experience, improving user satisfaction and viewing experience.

[0055] The first on-demand program refers to the first recommended content in the on-demand knowledge graph that matches the current live content. The first on-demand program is usually a program that is highly related to the current viewing theme, character, or plot.

[0056] To facilitate understanding, let's use an example: Suppose a user is watching the fifth episode of the TV series "AABB" live. The current content involves the development of the love story between the female lead Xiao Li and the male lead Xiao Zhang. After extracting this information from the live broadcast recognition results, it is matched using the on-demand knowledge graph. The nodes and relationships in the graph reveal that "the sixth episode of the TV series "AABB"" is highly relevant to the current content. Therefore, "the sixth episode of the TV series "AABB"" can be selected as the first on-demand content, providing the user with a natural viewing continuation, further increasing their engagement and satisfaction.

[0057] S130: Input the channel identifier of the current live channel and the program identifier of the first on-demand program into a pre-trained on-demand recommendation model to obtain the first on-demand content and the corresponding user recommendation timing output by the on-demand recommendation model.

[0058] Among them, the channel identifier refers to an identifier used to uniquely identify a specific live broadcast channel, which is used to identify and distinguish different live broadcast content sources.

[0059] Among them, the on-demand recommendation model is an algorithm model based on machine learning technology, which is used to output the corresponding on-demand content and its most suitable user recommendation time based on the channel ID and the program ID of the on-demand program.

[0060] The user recommendation timing refers to the time when the on-demand recommendation model identifies the best time for on-demand content and pushes the first on-demand content to the user. The user recommendation timing can be the beginning, end, advertisement, or main film.

[0061] To facilitate understanding, let's use an example: Suppose a user is watching a live TV series. The current channel's channel ID is "drama_channel," and the first on-demand content is the TV series "CCDD." The program IDs "drama_channel" and "CCDD" are input into the on-demand recommendation model for processing. The on-demand recommendation model analyzes the viewing history of this channel and finds that most users are most interested in new content at the end of the credits. Therefore, based on the recalled content template, it generates the first on-demand content and identifies the corresponding user recommendation timing as "end of credits."

[0062] When training an on-demand recommendation model, it's important to comprehensively consider the effective combination of recommended content and timing, as well as the coordination between various content types, to enhance the appeal of recommendations. For example, if "FFGG" is currently playing on a live broadcast channel and the series is experiencing a limited-time free promotion on the platform, the recommended content should include this promotional information. This can be achieved through a recall content format, such as "{FFGG on-demand|Platform limited-time free promotion, HN TV, 2:00 PM-3:00 PM end credits timing}," which indicates that during a specific time period on HN TV, "FFGG" will be recommended at the end credits, informing viewers of the limited-time promotion. To construct this recall content combination template, template rules allow for the free combination of information such as content, resource location, and timing, generating all possible recall content combinations. The feature data for model training should be constructed around the combination of content combination and timing. Therefore, key features such as the live broadcast program, the recommended content combination, and its corresponding timing should be recorded in the tracking information. Furthermore, based on the analysis of live content slices, user profile information can be enriched, for example, identifying a user's preference for historical dramas and assigning relevant tags. For model training, conventional machine learning algorithms, such as LightGBM or neural networks, can be used to train the model by inputting these feature data to optimize recommendation results. Ultimately, the on-demand recommendation model will be able to output the user recommendation timing corresponding to the first on-demand content, thereby achieving accurate on-demand recommendations.

[0063] S140: Add the first on-demand content and the user recommendation timing as on-demand recommendation data to the on-demand recommendation ranking result associated with the user terminal.

[0064] The on-demand recommendation ranking result refers to a list of on-demand content related to the current live channel after being processed by a recommendation algorithm.

[0065] S150: When the current live broadcast state matches the user recommendation timing, push the first on-demand content to the user terminal, so that when the user triggers the first on-demand content displayed on the user terminal, the first on-demand program is played.

[0066] Specifically, an embodiment of the present invention can push the first on-demand content to the user terminal when the current live broadcast status is consistent with the user recommendation timing, so as to display the first on-demand content in the form of a pop-up window on the user terminal. After the user clicks on the first on-demand content, the user terminal jumps to the playback interface of the first on-demand program.

[0067] The present invention provides a method for recommending on-demand content, which includes: cutting a live stream of a user terminal under a current live channel into time-series video slices in real time; inputting the time-series video slices into a pre-trained live recognition model to obtain a live recognition result output by the live recognition model, wherein the live recognition result includes the current live content and the current live status of the live stream; using a pre-built on-demand knowledge graph to locate a first on-demand program that matches the current live content; inputting the channel identifier of the current live channel and the program identifier of the first on-demand program into a pre-trained on-demand recommendation model to obtain the first on-demand content and the corresponding user recommendation timing output by the on-demand recommendation model; adding the first on-demand content and the user recommendation timing as on-demand recommendation data to an on-demand recommendation ranking result associated with the user terminal; and pushing the first on-demand content to the user terminal when the current live status matches the user recommendation timing, so that when the user triggers the first on-demand content displayed on the user terminal, the first on-demand program is jumped to play. The present invention accurately identifies live broadcast content and live broadcast status, dynamically matches and associates on-demand programs, and then uses an on-demand recommendation model to generate on-demand content and predict the optimal push timing. It can accurately recommend on-demand content at key moments of live broadcast, thereby effectively improving the conversion rate of on-demand programs.

[0068] Optionally, the live broadcast recognition model consists of a live broadcast status recognition model and a live broadcast content recognition model. Figure 1 The method shown, such as Figure 2 As shown in FIG. 1 , a flowchart of a first specific implementation of the on-demand content recommendation method provided by an embodiment of the present invention is shown. Step S110 may specifically include:

[0069] S200: Input the time-series video slices into a pre-trained live broadcast status recognition model to obtain the current live broadcast status of the live broadcast stream output by the live broadcast status recognition model.

[0070] The live broadcast status recognition model, built using machine learning and deep learning technologies, is used to analyze and determine the different playback stages of live broadcast content (such as the opening, main film, commercials, and end credits) in real time. The model processes multimodal video information, including visual, temporal, and auxiliary features, and classifies it according to preset status rules, thereby supporting content recommendations and optimizing the user experience.

[0071] S210: Input the time-series video slices into a pre-trained live content recognition model to obtain the current live content of the live stream output by the live content recognition model.

[0072] The live content recognition model, built using machine learning and deep learning technologies, is used to identify and analyze specific content features presented in live broadcasts, such as lines, celebrities, content categories, and character costumes. The live content recognition model processes multimodal video information, including visual, temporal, and auxiliary features, effectively classifying and describing live streams, thereby providing users with relevant content information and recommendations.

[0073] By inputting time-series video slices into a live broadcast status recognition model and a live broadcast content recognition model, the embodiment of the present invention can quickly obtain the current live broadcast status and live broadcast content, providing timely and effective data support for real-time on-demand content recommendation and improving user experience.

[0074] Optional, based on Figure 2 The method shown, such as Figure 3 As shown, a flowchart of the training process of the live broadcast status recognition model provided by an embodiment of the present invention is shown. The training process of the live broadcast status recognition model may include:

[0075] S300: Marking a first live video clip based on preset status rules to obtain a first video material marked with a live status, wherein the preset status rules include a title marking rule, an end marking rule, an advertisement marking rule, and a main film marking rule.

[0076] Among them, the preset status rules are standardized rules used to mark the different states of live streams, covering the playback links including the beginning, end, advertisements and main film.

[0077] The first live video clip refers to a video material of a specific period of time extracted from the live broadcast, which is used to train the live broadcast state recognition model. The first video material refers to the annotated first live video clip, which contains annotation information corresponding to the preset state rule.

[0078] Among them, the title marking rules refer to the standards used to identify and mark the beginning of a live broadcast program, including the program name, title animation, and copyright information.

[0079] Among them, the end credits marking rules refer to the standards used to identify and mark the end of a live broadcast program, including credits, thank you subtitles, and ending songs.

[0080] Among them, advertising labeling rules refer to the rules used to identify and mark the advertising parts in live broadcasts, including brand logos, advertising slogans, and product displays.

[0081] Among them, the main film marking rules refer to the standards used to identify and mark the main content parts of the live broadcast, usually including plot content, etc.

[0082] S310: Perform multimodal feature extraction on the first video material to obtain a first visual feature, a first temporal feature, and a first auxiliary feature of the video material.

[0083] Among them, the first visual feature is feature information related to the image extracted from the first video material, which is usually obtained through a deep learning model (such as a convolutional neural network) and may include visual information including the color, shape and texture of the image.

[0084] Among them, the first temporal feature is information related to the time series extracted from the first video material, which is usually obtained through a recurrent neural network (such as a recursive neural network) or a Transformer to help the model understand the temporal relationship between video frames and enhance the accuracy of classification.

[0085] The first auxiliary features refer to other information extracted in addition to visual and temporal features, such as text features and audio features.

[0086] S320: Concatenate the first visual feature, the first temporal feature, and the first auxiliary feature into a first multimodal input vector.

[0087] S330: Input the first multimodal input vector into the live broadcast status recognition model for training, optimize the live broadcast status classification task through the cross entropy loss function, and obtain a trained live broadcast status recognition model.

[0088] The cross-entropy loss function is a loss function used for model training, primarily for classification tasks, to evaluate the difference between the predicted probability distribution and the actual label. This embodiment of the present invention can evaluate the live status recognition model using metrics such as accuracy, F1 score, and AUC-ROC curve to obtain a live status recognition model with good live status classification.

[0089] The embodiment of the present invention ensures the standardization and accuracy of the data by annotating the first live video clip based on preset status rules, thereby obtaining a clearly annotated first video material. Next, multimodal feature extraction is performed on the video material to enrich the input information of the model, enabling it to fully understand the multidimensional attributes of the live broadcast status. Subsequently, these features are spliced ​​into a first multimodal input vector, which is input into the live broadcast status recognition model and optimized using the cross-entropy loss function, so that the accuracy and robustness of the live broadcast status recognition model in the live broadcast status classification task are significantly improved, thereby not only improving the real-time classification capability of the live broadcast status, but also helping to provide a solid data foundation for personalized recommendations.

[0090] Optional, based on Figure 2 The method shown, such as Figure 4 As shown, a flowchart of the training process of the live broadcast status recognition model provided by an embodiment of the present invention is provided. The training process of the live broadcast content recognition model may include:

[0091] S400: Annotate the second live video clip based on preset content rules to obtain a second video material with annotated live content, wherein the preset content rules include dialogue annotation rules, star annotation rules, content category annotation rules, and character costume annotation rules.

[0092] Among them, the preset content rules are standardized rules used to mark different content in live streams, covering lines, stars, content categories and character costumes.

[0093] The second live video clip refers to a video material of a specific period of time extracted from the live broadcast, which is used to train the live content recognition model. The second video material refers to the annotated second live video clip, which contains annotation information corresponding to the preset content rules.

[0094] Among them, the dialogue annotation rules refer to the specifications used to annotate dialogues or monologues in live broadcasts.

[0095] Among them, celebrity tagging rules refer to the rules used to identify and tag well-known figures or celebrities appearing in live broadcasts.

[0096] Among them, content category labeling rules refer to the standards used to classify live broadcast content, such as entertainment, games, education, sports and other different types.

[0097] Among them, the character costume annotation rules refer to the specifications used to identify and annotate the character costumes or appearance features in live broadcasts.

[0098] S410: Perform multimodal feature extraction on the second video material to obtain a second visual feature, a second temporal feature, and a second auxiliary feature of the video material.

[0099] Among them, the second visual feature is feature information related to the image extracted from the second video material, which is usually obtained through a deep learning model (such as a convolutional neural network) and can include visual information including the color, shape and texture of the image.

[0100] Among them, the second temporal feature is information related to the time series extracted from the second video material, which is usually obtained through a recurrent neural network (such as a recursive neural network) or a Transformer to help the model understand the temporal relationship between video frames and enhance the accuracy of classification.

[0101] Among them, the second auxiliary features refer to other information extracted in addition to visual and temporal features, such as text features and audio features.

[0102] S420: Concatenate the second visual feature, the second temporal feature, and the second auxiliary feature into a second multimodal input vector.

[0103] S430: Input the second multimodal input vector into the live content recognition model for training, optimize the live content recognition task through the cross entropy loss function, and obtain a trained live content recognition model.

[0104] The embodiment of the present invention can evaluate the live content recognition model through indicators such as accuracy, F1 score, AUC-ROC curve, etc., to obtain a live content recognition model with good live content recognition.

[0105] The embodiment of the present invention ensures the standardization and accuracy of the data by annotating the second live video clip based on preset content rules, thereby obtaining a clearly annotated second video material. Next, multimodal feature extraction is performed on the video material to enrich the input information of the model, enabling it to fully understand the multidimensional attributes of the live content. Subsequently, these features are spliced ​​into a second multimodal input vector, which is input into the live content recognition model and optimized using the cross-entropy loss function, so that the accuracy and robustness of the live content recognition model in the live content recognition task are significantly improved, thereby not only improving the real-time recognition capability of live content, but also helping to provide a solid data foundation for personalized recommendations.

[0106] Optional, based on Figure 1 The method shown, such as Figure 5 As shown, a flowchart of the process of constructing an on-demand knowledge graph provided by an embodiment of the present invention is provided. The process of constructing an on-demand knowledge graph may include:

[0107] S500. Define the graph structure of the on-demand knowledge graph: create episode nodes, scene nodes, and live content nodes; define a first node relationship between episode nodes and scene nodes; and define a second node relationship between scene nodes and live content nodes.

[0108] An episode node is a node that represents an episode of a specific TV series or program in the on-demand knowledge graph. Each episode node represents an independent episode and contains important information about the episode.

[0109] A scene node is a node in the knowledge graph that represents a specific scene in a TV series. Each scene node represents a specific scene in the TV series.

[0110] Among them, the live content node refers to the node that represents specific live content in the on-demand knowledge graph, which usually includes real-time information related to the drama.

[0111] The first node relationship is used to represent one or more scenes that an episode may contain.

[0112] The second node relationship is used to indicate that a scene may contain multiple live content information.

[0113] S510: Configure node attributes of the set node, scene node, and live content node.

[0114] Among them, the node attributes of the episode node include the episode number, title and season; the node attributes of the scene node include the scene number, start timestamp and end timestamp; the node attributes of the live content node include text content, associated role and line timestamp.

[0115] S520. Parse the script text or subtitle file, extract the set entity, scene entity, and live content entity according to the graph structure, and inject the set entity into the corresponding set node, the scene entity into the corresponding scene node, and the live content entity into the corresponding live content node.

[0116] To facilitate understanding, let's illustrate this with an example: suppose when parsing the script text of a TV series, the title and episode number of the series are first extracted as episode entities, and then this information is injected into the corresponding episode node; then, the scene description is identified from the script, such as "In a cafe, character A talks with character B", the scene number and start and end timestamps are extracted, and this information is injected into the corresponding scene node as a scene entity; finally, the dialogue content in the scene is analyzed, the character's lines and related character information are extracted to form a live content entity, which is then injected into the corresponding live content node.

[0117] S530: Associating the scene node and the live content node according to the dialogue timestamp.

[0118] S540: Establish a temporal relationship between scene nodes according to the time axis sequence.

[0119] The embodiment of the present invention makes it possible to establish an accurate match between live content and on-demand recommendations through a systematic on-demand knowledge graph construction process, thereby improving the user's viewing experience and the personalization of content recommendations, and ultimately providing users with accurate on-demand recommendations.

[0120] Optional, based on Figure 1 The method shown, such as Figure 6 As shown in FIG. , a flow chart of a second specific implementation of the on-demand content recommendation method provided by an embodiment of the present invention is provided. Before step S100, the method may further include:

[0121] S600: In response to a power-on event of the user terminal, obtain a live broadcast program list of the current live broadcast channel, wherein the live broadcast program list of the current day includes at least one live broadcast program.

[0122] Among them, the power-on event refers to the event triggered when the user turns on the user terminal, indicating that the user starts using the device and enters the viewing mode.

[0123] The daily live program list refers to a list of live programs scheduled for a specific date on the current live channel. It typically includes program titles, broadcast times, and channel information, and is used to display the live content available for viewing that day. For example, the daily live program list might be shown in Table 1.

[0124]

[0125] Live broadcasts are programs that are broadcast in real time at a specific time. They can include news, variety shows, sports events, TV dramas, and other types of programs.

[0126] S610: Calculate the similarity between each live program in the day's live program list and each second on-demand program in the on-demand library using a text matching algorithm, and select the N second on-demand programs with the highest similarity as the initial recall results, where N is the preset frequency control number.

[0127] Among them, the on-demand library is a database containing various on-demand programs (such as movies, TV series, variety shows and documentaries, etc.).

[0128] The second on-demand program refers to a related on-demand program that can be recommended to the user while the user is watching the live program.

[0129] The preset frequency control number indicates the number of on-demand content recommendations a user can receive within a specified time period. The specific value of the preset frequency control number can be adjusted based on factors such as user viewing habits and system performance to optimize the recommendation effect.

[0130] S620: Calculate the initial recommendation timing of the initial recall result based on the broadcast time information of each live program in the live program list, and generate an initial recall result set.

[0131] The initial recall result set includes multiple initial recall results and their corresponding initial recommendation timings.

[0132] The initial recommendation time refers to the recommended time point automatically generated for each initial recall result after the user terminal is turned on. The initial recommendation time can be defaulted to the end of the corresponding live program.

[0133] S630: Use the initial recall result set as initialization data for the on-demand recommendation sorting result, so as to constrain the frequency control range of subsequent real-time recommendation of on-demand content.

[0134] Specifically, this embodiment of the present invention uses the initial recall result set as the initial data for sorting on-demand recommendations. This means that subsequent real-time recommendations will be based on this set as a starting point to limit the frequency and scope of recommendations. By setting the on-demand programs and recommendation timings in the initial result set, the number of recommendations a user receives within a certain period of time can be effectively controlled, thereby avoiding information overload while ensuring the relevance and effectiveness of recommendations, thereby improving the user experience.

[0135] This embodiment of the present invention calculates the similarity between live programs and programs in the on-demand library and selects the N most similar on-demand programs as the initial recall results. This not only provides a preliminary basis for recommendation but also ensures the diversity and relevance of the recommendations. Furthermore, by using the broadcast time information of the live program list to calculate the initial recommendation timing, it optimizes the selection of recommendation timing, providing a reliable data foundation for subsequent real-time recommendations and enhancing the user experience. By constraining the frequency range of recommendations, it ensures that users receive the most relevant on-demand content at the appropriate time, thereby improving overall viewing satisfaction.

[0136] Optional, in the above Figure 6 On the basis of one or more corresponding embodiments, in another optional embodiment provided by the embodiment of the present invention, when the current live broadcast state matches the user recommendation timing, pushing the first on-demand content to the user terminal may specifically include:

[0137] When the current live broadcast status matches the user recommended timing, the current frequency control times are checked; if the current frequency control times are not greater than the preset frequency control times, the first on-demand content is pushed to the user terminal and the current frequency control times are updated.

[0138] Specifically, when the current live broadcast status matches the user recommendation timing, the embodiment of the present invention first checks the current frequency control count to ensure that the recommended content push complies with the preset frequency control limit. If the frequency control count does not exceed the preset frequency control count, the first on-demand content is pushed to the user terminal and the current frequency control count is updated accordingly. This avoids overly frequent recommendations, ensures that users receive relevant on-demand content at the appropriate time, and improves the user experience.

[0139] like Figure 7 As shown in the figure, the logical principle block diagram of the on-demand content recommendation system provided by an embodiment of the present invention includes three major modules: live broadcast recognition, recommendation engine, and recommendation trigger. Live broadcast recognition is divided into three steps: live broadcast status recognition model training, live broadcast content recognition model training, and online recognition. It aims to identify the current live broadcast status (such as advertisements, end credits, etc.) and content attributes (such as lines, celebrities, etc.) through data annotation and model training. Online recognition generates recognition results by cutting the live broadcast stream in real time and analyzing the slices to support subsequent recommendations. The recommendation engine focuses on the combination of refined recommendations, recommendation frequency, and timing, using the on-demand knowledge graph to improve the accuracy of content recall, and designing a recommendation model to combine content and activity information. Through the recommendation engine, the system implements recall, sorting, and re-arrangement to ensure the relevance and timeliness of recommendations. The recommendation trigger responds in real time while the user is watching the live broadcast, making recommendation corrections and optimizing the recommendation frequency, thereby improving the user's viewing experience.

[0140] Although the operations are depicted in a particular order, this should not be understood as requiring that the operations be performed in the particular order shown or in a sequential order.Multitasking and parallel processing may be advantageous under certain circumstances.

[0141] It should be understood that the various steps described in the method embodiments of the present invention may be performed in different orders and / or in parallel. In addition, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present invention is not limited in this respect.

[0142] Corresponding to the above method embodiment, the embodiment of the present invention further provides an on-demand content recommendation device, the structure of which is as follows: Figure 8 As shown, it may include: a live stream real-time cutting unit 10, a live broadcast identification result obtaining unit 20, an on-demand program matching unit 30, an on-demand content recommendation opportunity obtaining unit 40, an on-demand recommendation data updating unit 50 and an on-demand content recommendation unit 60.

[0143] The live stream real-time cutting unit 10 is used to cut the live stream of the user terminal under the current live channel into time-series video slices in real time.

[0144] The live broadcast recognition result obtaining unit 20 is used to input the time-series video slices into a pre-trained live broadcast recognition model to obtain the live broadcast recognition result output by the live broadcast recognition model, wherein the live broadcast recognition result includes the current live broadcast content and current live broadcast status of the live broadcast stream.

[0145] The on-demand program matching unit 30 is used to locate the first on-demand program that matches the current live content by using the pre-built on-demand knowledge graph.

[0146] The on-demand content recommendation timing obtaining unit 40 is used to input the channel identifier of the current live channel and the program identifier of the first on-demand program into a pre-trained on-demand recommendation model to obtain the first on-demand content and the corresponding user recommendation timing output by the on-demand recommendation model.

[0147] The on-demand recommendation data updating unit 50 is configured to add the first on-demand content and the user recommendation timing as on-demand recommendation data to the on-demand recommendation ranking result associated with the user terminal.

[0148] The on-demand content recommendation unit 60 is configured to push the first on-demand content to the user terminal when the current live broadcast state matches the user recommendation timing, so as to jump to play the first on-demand program when the user triggers the first on-demand content displayed on the user terminal.

[0149] Optionally, the live broadcast recognition model is composed of a live broadcast status recognition model and a live broadcast content recognition model. The live broadcast recognition result acquisition unit 20 is specifically used to input the time-series video slices into a pre-trained live broadcast status recognition model to obtain the current live broadcast status of the live broadcast stream output by the live broadcast status recognition model; and input the time-series video slices into a pre-trained live broadcast content recognition model to obtain the current live broadcast content of the live broadcast stream output by the live broadcast content recognition model.

[0150] Optionally, the on-demand content recommendation device may further include: a live broadcast status recognition model training unit.

[0151] The live broadcast status recognition model training unit is used to label the first live broadcast video clip based on preset status rules to obtain a first video material labeled with the live broadcast status, wherein the preset status rules include a title labeling rule, an end labeling rule, an advertisement labeling rule and a main film labeling rule; multimodal feature extraction is performed on the first video material to obtain a first visual feature, a first temporal feature and a first auxiliary feature of the video material; the first visual feature, the first temporal feature and the first auxiliary feature are spliced ​​into a first multimodal input vector; the first multimodal input vector is input into the live broadcast status recognition model for training, and the live broadcast status classification task is optimized through the cross entropy loss function to obtain a trained live broadcast status recognition model.

[0152] Optionally, the on-demand content recommendation device may further include: a live content recognition model training unit.

[0153] The live broadcast content recognition model training unit is used to annotate a second live broadcast video clip based on preset content rules to obtain a second video material with annotated live broadcast content, wherein the preset content rules include dialogue annotation rules, star annotation rules, content category annotation rules and character costume annotation rules; multimodal feature extraction is performed on the second video material to obtain a second visual feature, a second temporal feature and a second auxiliary feature of the video material; the second visual feature, the second temporal feature and the second auxiliary feature are spliced ​​into a second multimodal input vector; the second multimodal input vector is input into the live broadcast content recognition model for training, and the live broadcast content recognition task is optimized through the cross-entropy loss function to obtain a trained live broadcast content recognition model.

[0154] Optionally, the on-demand content recommendation device may further include: an on-demand knowledge graph construction unit.

[0155] The on-demand knowledge graph construction unit is used to define the graph structure of the on-demand knowledge graph: create episode nodes, scene nodes and live content nodes; define the first node relationship between episode nodes and scene nodes; define the second node relationship between scene nodes and live content nodes.

[0156] Configure the node properties of the episode node, scene node, and live content node, where the node properties of the live content node include the dialogue timestamp; parse the script text or subtitle file, extract the episode entity, scene entity, and live content entity according to the graph structure, and inject the episode entity into the corresponding episode node, the scene entity into the corresponding scene node, and the live content entity into the corresponding live content node; associate the scene node and the live content node according to the dialogue timestamp; establish the temporal relationship between the scene nodes according to the timeline order.

[0157] Optionally, the on-demand content recommendation device may further include: a recommendation initialization unit.

[0158] The recommendation initialization unit is used for obtaining the current live broadcast program list of the current live broadcast channel in response to the power-on event of the user terminal before the live broadcast stream real-time cutting unit 10 cuts the live broadcast stream of the user terminal under the current live broadcast channel into time-series video slices in real time, wherein the current live broadcast program list includes at least one live program; calculating the similarity between each live program in the current live broadcast program list and each second on-demand program in the on-demand library through a text matching algorithm, and selecting the N second on-demand programs with the highest similarity as the initial recall result, wherein N is the preset frequency control number; calculating the initial recommendation timing of the initial recall result based on the broadcast time information of each live program in the live broadcast program list, and generating an initial recall result set; using the initial recall result set as the initialization data of the on-demand recommendation sorting result, and constraining the frequency control range of the subsequent real-time recommendation of on-demand content.

[0159] Optionally, the on-demand content recommendation unit 60 is specifically configured to check the current frequency control times when the current live broadcast status matches the user recommendation timing; if the current frequency control times are not greater than the preset frequency control times, push the first on-demand content to the user terminal and update the current frequency control times.

[0160] The present invention provides a device for recommending on-demand content, which is used to: cut the live stream of a user terminal under the current live channel into time-series video slices in real time; input the time-series video slices into a pre-trained live recognition model to obtain a live recognition result output by the live recognition model, wherein the live recognition result includes the current live content and the current live status of the live stream; use a pre-built on-demand knowledge graph to locate a first on-demand program that matches the current live content; input the channel identifier of the current live channel and the program identifier of the first on-demand program into the pre-trained on-demand recommendation model to obtain the first on-demand content and the corresponding user recommendation timing output by the on-demand recommendation model; add the first on-demand content and the user recommendation timing as on-demand recommendation data to the on-demand recommendation ranking result associated with the user terminal; and push the first on-demand content to the user terminal when the current live status matches the user recommendation timing, so that when the user triggers the first on-demand content displayed on the user terminal, the first on-demand program is played. The present invention accurately identifies live broadcast content and live broadcast status, dynamically matches and associates on-demand programs, and then uses an on-demand recommendation model to generate on-demand content and predict the optimal push timing. It can accurately recommend on-demand content at key moments of live broadcast, thereby effectively improving the conversion rate of on-demand programs.

[0161] Regarding the apparatus in the above embodiment, the specific manner in which each unit performs operations has been described in detail in the embodiment of the method, and will not be elaborated on here.

[0162] The on-demand content recommendation device includes a processor and a memory. The above-mentioned live stream real-time cutting unit 10, live broadcast identification result obtaining unit 20, on-demand program matching unit 30, on-demand content recommendation opportunity obtaining unit 40, on-demand recommendation data updating unit 50 and on-demand content recommendation unit 60 are all stored in the memory as program units, and the processor executes the above-mentioned program units stored in the memory to realize corresponding functions.

[0163] The processor contains a core, which retrieves the corresponding program unit from memory. One or more cores can be configured, and by adjusting core parameters, they accurately identify live content and status, dynamically matching on-demand programs. A recommendation model then generates on-demand content and predicts the optimal timing for push notifications. This allows for precise recommendations at key moments during live broadcasts, effectively increasing conversion rates for on-demand programs.

[0164] An embodiment of the present invention provides a computer-readable storage medium having a program stored thereon. When the program is executed by a processor, the on-demand content recommendation method is implemented.

[0165] An embodiment of the present invention provides a processor, which is used to run a program, wherein the on-demand content recommendation method is executed when the program is run.

[0166] like Figure 9 As shown, an embodiment of the present invention provides an electronic device 1000, which includes at least one processor 1001, at least one memory 1002 connected to the processor 1001, and a bus 1003. The processor 1001 and the memory 1002 communicate with each other via the bus 1003. The processor 1001 is configured to invoke program instructions stored in the memory 1002 to execute the above-described on-demand content recommendation method. The electronic device herein may be a server, a PC, a PAD, a mobile phone, or the like.

[0167] The present invention also provides a computer program product, which, when executed on an electronic device, is adapted to execute the program steps of initializing the on-demand content recommendation method.

[0168] The present invention is described with reference to flowcharts and / or block diagrams of methods, apparatuses, electronic devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable device to produce a machine, so that the instructions executed by the processor of the computer or other programmable device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1A device that provides the functions specified in a block or multiple blocks.

[0169] In a typical configuration, an electronic device includes one or more processors (CPUs), a memory, and a bus. The electronic device may also include an input / output interface, a network interface, and the like.

[0170] Memory may include non-permanent memory in a computer-readable medium, random access memory (RAM), and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory includes at least one memory chip. Memory is an example of a computer-readable medium.

[0171] Computer-readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. The information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory media such as modulated data signals and carrier waves.

[0172] 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 the present invention 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.

[0173] It is understandable that before using the technical solutions disclosed in the various embodiments of this disclosure, the type, scope of use, usage scenarios, etc. of the personal information involved in this disclosure should be informed to the user and the user's authorization should be obtained in an appropriate manner in accordance with relevant laws and regulations.

[0174] In the description of the present invention, it should be understood that if the terms "up", "down", "front", "back", "left" and "right" are used to indicate directions or positional relationships, they are based on the directions or positional relationships shown in the accompanying drawings. They are only used to facilitate the description of the present invention and simplify the description, and do not indicate or imply that the positions or elements referred to must have a specific direction, be constructed and operate in a specific direction. Therefore, they should not be understood as limitations of the present invention.

[0175] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. It should also be noted that the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, commodity, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, commodity, or device comprising the element.

[0176] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROMs, optical storage, etc.) containing computer-usable program code.

[0177] The above are merely embodiments of the present invention and are not intended to limit the present invention. It will be apparent to those skilled in the art that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are intended to be within the scope of the present invention.

Claims

1. A method for recommending on-demand content, characterized in that: include: Cut the live stream of the user end under the current live channel into time-series video slices in real time; Inputting the time-series video slices into a pre-trained live broadcast recognition model to obtain a live broadcast recognition result output by the live broadcast recognition model, wherein the live broadcast recognition result includes the current live broadcast content and current live broadcast status of the live broadcast stream; Using a pre-built on-demand knowledge graph, locate a first on-demand program that matches the current live content; Inputting the channel identifier of the current live channel and the program identifier of the first on-demand program into a pre-trained on-demand recommendation model, and obtaining the first on-demand content and the corresponding user recommendation timing output by the on-demand recommendation model; adding the first on-demand content and the user recommendation timing as on-demand recommendation data to the on-demand recommendation ranking result associated with the user terminal; When the current live broadcast state matches the user recommendation timing, the first on-demand content is pushed to the user terminal, so that when the user triggers the first on-demand content displayed on the user terminal, the first on-demand program is played.

2. The method according to claim 1, characterized in that The live broadcast recognition model is composed of a live broadcast status recognition model and a live broadcast content recognition model. The time-series video slices are input into a pre-trained live broadcast recognition model to obtain a live broadcast recognition result output by the live broadcast recognition model, including: Inputting the time-series video slices into the pre-trained live broadcast status recognition model to obtain the current live broadcast status of the live broadcast stream output by the live broadcast status recognition model; The time-series video slices are input into the pre-trained live content recognition model to obtain the current live content of the live stream output by the live content recognition model.

3. The method according to claim 2, characterized in that The training process of the live broadcast status recognition model includes: Marking the first live video clip based on preset status rules to obtain a first video material marked with a live status, wherein the preset status rules include a title marking rule, an end marking rule, an advertisement marking rule, and a main film marking rule; Performing multimodal feature extraction on the first video material to obtain a first visual feature, a first temporal feature, and a first auxiliary feature of the video material; splicing the first visual feature, the first temporal feature, and the first auxiliary feature into a first multimodal input vector; The first multimodal input vector is input into the live broadcast status recognition model for training, and the live broadcast status classification task is optimized through the cross entropy loss function to obtain the trained live broadcast status recognition model.

4. The method according to claim 2, characterized in that The training process of the live content recognition model includes: annotating the second live video clip based on preset content rules to obtain a second video material annotated with live content, wherein the preset content rules include dialogue annotation rules, celebrity annotation rules, content category annotation rules, and character costume annotation rules; performing multimodal feature extraction on the second video material to obtain a second visual feature, a second temporal feature, and a second auxiliary feature of the video material; splicing the second visual feature, the second temporal feature, and the second auxiliary feature into a second multimodal input vector; The second multimodal input vector is input into the live content recognition model for training, and the live content recognition task is optimized through the cross entropy loss function to obtain the trained live content recognition model.

5. The method according to claim 1, wherein The process of constructing the on-demand knowledge graph includes: Defining the graph structure of the on-demand knowledge graph: creating an episode node, a scene node, and a live content node; defining a first node relationship between the episode node and the scene node; defining a second node relationship between the scene node and the live content node; Configuring node attributes of the episode node, the scene node, and the live content node, wherein the node attributes of the live content node include a line timestamp; Parsing the script text or subtitle file, extracting the set entity, the scene entity, and the live content entity according to the graph structure, and injecting the set entity into the corresponding set node, injecting the scene entity into the corresponding scene node, and injecting the live content entity into the corresponding live content node; Associating the scene node with the live content node according to the dialogue timestamp; The temporal relationship between the scene nodes is established according to the time axis sequence.

6. The method according to claim 1, wherein Before the live stream of the user terminal under the current live channel is cut into time-sequential video slices in real time, the method further includes: In response to a power-on event of the user terminal, obtaining a live program list of the current live channel, wherein the live program list includes at least one live program; Calculating the similarity between each live program in the live program list of the day and each second on-demand program in the on-demand library using a text matching algorithm, and selecting the N second on-demand programs with the highest similarity as initial recall results, where N is a preset frequency control number; Calculating the initial recommendation timing of the initial recall result based on the broadcast time information of each live program in the live program list, and generating an initial recall result set; The initial recall result set is used as initialization data for the on-demand recommendation sorting result, and is used to constrain the frequency control range of subsequent real-time recommendation of on-demand content.

7. The method according to claim 6, characterized in that When the current live broadcast state matches the user recommended timing, pushing the first on-demand content to the user terminal includes: When the current live broadcast state matches the user recommended timing, checking the current frequency control times; When the current frequency control number is not greater than the preset frequency control number, the first on-demand content is pushed to the user terminal, and the current frequency control number is updated.

8. A device for recommending on-demand content, characterized in that: include: Live stream real-time cutting unit, live recognition result obtaining unit, on-demand program matching unit, on-demand content recommendation opportunity obtaining unit, on-demand recommendation data updating unit and on-demand content recommendation unit, The live stream real-time cutting unit is used to cut the live stream of the user terminal under the current live channel into time-series video slices in real time; The live broadcast recognition result obtaining unit is configured to input the time-series video slice into a pre-trained live broadcast recognition model to obtain a live broadcast recognition result output by the live broadcast recognition model, wherein the live broadcast recognition result includes the current live broadcast content and current live broadcast status of the live broadcast stream; The on-demand program matching unit is configured to locate a first on-demand program that matches the current live content using a pre-built on-demand knowledge graph; The on-demand content recommendation opportunity obtaining unit is configured to input the channel identifier of the current live channel and the program identifier of the first on-demand program into a pre-trained on-demand recommendation model, and obtain the first on-demand content and the corresponding user recommendation opportunity output by the on-demand recommendation model; The on-demand recommendation data updating unit is configured to add the first on-demand content and the user recommendation opportunity as on-demand recommendation data to the on-demand recommendation ranking result associated with the user terminal; The on-demand content recommendation unit is configured to push the first on-demand content to the user terminal when the current live broadcast status matches the user recommendation timing, so as to jump to play the first on-demand program when the user triggers the first on-demand content displayed on the user terminal.

9. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, the on-demand content recommendation method according to any one of claims 1 to 7 is implemented.

10. An electronic device, characterized in that: The electronic device includes at least one processor, and at least one memory and bus connected to the processor; wherein the processor and the memory communicate with each other through the bus; the processor is used to call program instructions in the memory to execute the on-demand content recommendation method as described in any one of claims 1 to 7.

Images