An AI short video content optimization method based on a knowledge graph

Abstract

The application discloses an AI short video content optimization method based on a knowledge graph, relates to the technical field of short video intelligent editing and knowledge graph application, and specifically comprises the following steps: constructing a short video field basic knowledge graph, collecting data in association by adopting a multilevel burying point technology, and generating a structured data set; updating the graph weight according to the data set increment, adapting the editing parameters by using the updated weight, and generating an optimized short video; pushing the optimized video to a user, collecting new feedback data, and realizing cyclic iteration optimization; the application builds a basic knowledge graph and adopts a multilevel collaborative burying point architecture to form a structured associated data set, improves data utilization and preference matching degree, establishes a dynamic reasoning link to realize intelligent adjustment of editing parameters, meets the demand of personalized presentation and efficient processing, realizes full-link cyclic upgrading through a closed-loop optimization mechanism, continuously optimizes short video content presentation effect, and improves user experience and market competitiveness.

Description

Technical Field

[0001] This invention relates to the field of intelligent editing and knowledge graph application technology for short videos, specifically an AI-based short video content optimization method based on knowledge graphs. Background Technology

[0002] With the rapid development of mobile internet technology, short videos have become a core carrier of information dissemination and user entertainment, and their content consumption is characterized by personalization, immediacy, and high frequency. Users' demands for the content quality and presentation format of short videos are becoming increasingly diverse. Accurately matching user preferences and dynamically optimizing video presentation effects has become crucial for industry competition. Traditional short video optimization relies heavily on data statistical analysis, lacking a systematic understanding of the semantic logic of the short video domain. This makes it difficult to achieve a deep connection between user behavior and video semantics, resulting in a lack of knowledge support in the optimization process and an inability to meet users' refined needs for personalized content. Therefore, an intelligent optimization solution integrating knowledge graphs is urgently needed.

[0003] Existing short video content optimization technologies suffer from several shortcomings: They lack a structured knowledge system specifically designed for the short video domain, failing to systematically construct semantic entities and relationships across dimensions such as content, format, and users. This results in a lack of unified standards for analyzing the correlation between user behavior and video semantics, leading to fragmented data integration. Data collection technologies are often limited, focusing primarily on single dimensions like front-end user behavior or video content, failing to achieve multi-level collaborative collection and timestamp alignment. This results in insufficient accuracy in binding user behavior to the semantic information of corresponding video segments. Weight update mechanisms are static, failing to consider the real-time nature and time decay characteristics of user behavior. Update cycles lack reasonable control, making it difficult to dynamically track changes in user preferences. Editing parameter adjustments rely heavily on fixed rules, lacking a dynamic reasoning link between semantic entities, emotion tags, and editing parameters. This prevents dynamic adaptation based on user preferences and lacks a closed-loop iteration mechanism, hindering continuous improvement in optimization effectiveness.

[0004] In summary, existing technologies fail to effectively integrate knowledge resources and real-time user behavior data in the short video field. They exhibit significant shortcomings in semantic association, data collection, dynamic adaptation, and closed-loop optimization, resulting in low accuracy, real-time performance, and personalization in short video content optimization, making it difficult to meet industry development and user needs. Therefore, constructing a systematic optimization framework based on knowledge graphs to achieve deep integration of user behavior and video semantics, dynamic weight updates, and intelligent adaptation of editing parameters, forming a closed-loop iterative optimization mechanism, is of great significance for improving the quality of short video content and user experience, and has become an urgent technical problem to be solved in this field. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an AI-based short video content optimization method based on knowledge graphs. This method constructs a basic knowledge graph in the short video domain and sets initial weights; it employs multi-level tracking technology to associate and collect user behavior with video semantic information; it uses algorithms to convert user behavior into weight adjustment signals and dynamically updates the graph; it adapts editing parameters according to the updated graph to generate optimized videos; and through a closed-loop optimization mechanism, it collects user feedback for iterative improvement, continuously enhancing the presentation effect and user experience of short video content.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: an AI short video content optimization method based on knowledge graphs, the specific steps of which are as follows: S100. Constructing a basic knowledge graph: Building a basic knowledge graph for the short video domain, clarifying the semantic entity set, relation set, and relation weight set in the knowledge graph, and setting initial weights for each set; the semantic entity set and relation set serve as the benchmark for dataset association and collection; S200, Dataset Association Collection: Multi-level tracking technology is used throughout the short video playback process to synchronously collect real-time user operation behavior and semantic information of corresponding video segments; the collected semantic information is bound to the corresponding semantic entities and relationships in the basic knowledge graph to generate a structured user behavior-video semantic association dataset. S300, Incremental Update of Knowledge Graph Weights: Based on the generated associated dataset, the real-time behavior-weight incremental iteration algorithm is used to transform user behavior into knowledge graph weight adjustment signals, dynamically updating the relation weights of the knowledge graph; and incremental processing logic is used to update the associated content of the knowledge graph, controlling the update cycle of the knowledge graph. S400, Dynamic Adaptation of Editing Parameters: Based on the updated knowledge graph relationship weights, the weight-driven dynamic adaptation algorithm for editing parameters is used to calculate the adaptation parameter values; a dynamic reasoning link of semantic entity-emotion tag-editing parameter is constructed, and the calculated parameters are applied to the real-time playback and batch editing of short videos to generate optimized short videos; S500, Closed-Loop Optimization: Push optimized short videos to users, collect new user behavior feedback data, repeatedly execute the data set association collection, incremental update of graph weights and dynamic adaptation of editing parameters, and complete the loop iteration.

[0007] Furthermore, the basic knowledge graph of the short video domain is a three-layer relational topology, including a semantic entity set, a relational set, and a relational weight set. The semantic entity set includes content-dimensional entities, form-dimensional entities, and user-dimensional entities. Content-dimensional entities cover characters, scenes, dialogue keywords, and background music styles. Form-dimensional entities include shot duration, transition type, subtitle font, and subtitle display duration. User-dimensional entities include user tags and interest categories. The relational set includes mapping relationships between content-dimensional entities and form-dimensional entities, mapping relationships between content-dimensional entities and user-dimensional entities, and mapping relationships between form-dimensional entities. The initial weights of the relational weight set are set according to prior knowledge of the short video domain, with a weight value range of [0,1]. Each initial weight corresponds one-to-one with a mapping relationship in the relational set.

[0008] Furthermore, the multi-level tracking technology adopts a three-layer collaborative tracking architecture consisting of a front-end interaction layer, a video content layer, and a data transmission layer: the front-end interaction layer tracking is embedded in the short video playback page and client, collecting user actions such as clicking, liking, commenting, pausing, dragging the progress bar, and repeating playback in real time, and synchronously recording the action timestamp, action type, and the start and end time positions of the corresponding video segment; the video content layer tracking extracts the semantic entities, audio features, and camera parameters of the corresponding video segment; and the data transmission layer tracking performs timestamp alignment between the user action data and the semantic information of the video segment, driving the synchronous collection of both.

[0009] Furthermore, the process of binding the collected semantic information with the corresponding semantic entities and relationships in the basic knowledge graph to generate a structured user behavior-video semantic association dataset is as follows: extract entity features from the collected video segment semantic information, perform feature matching with the semantic entity set of the basic knowledge graph, and filter out the corresponding matching semantic entities; based on the matched semantic entities, retrieve the corresponding relationships in the relationship set of the basic knowledge graph; and structurally integrate the four types of data—user operation behavior data, video segment semantic information, matching semantic entities, and corresponding relationships—to form a structured user behavior-video semantic association dataset.

[0010] Furthermore, the specific steps of using the real-time behavior-weight increment iteration algorithm to transform user behavior into knowledge graph weight adjustment signals and dynamically update the relation weights of the knowledge graph are as follows: using the real-time behavior-weight increment iteration algorithm, classify and assign values ​​to user behaviors in the associated dataset, and set corresponding behavior coefficients according to the intensity of preference representation; combine the knowledge graph relation weights and time decay factors to calculate the weight increment value of each relation; add the increment value to the original weight, and take the boundary value if it exceeds the [0,1] interval to generate a weight adjustment signal; update the corresponding relation weights according to the adjustment signal.

[0011] Furthermore, the calculation formula for the real-time behavior-weight increment iteration algorithm is as follows: ,in, The updated knowledge graph relationship weights; The original weights of the knowledge graph relationships before the update; User behavior coefficient; This is the time decay factor; This is the lower bound constraint function; The upper limit constraint function transforms various user actions into effective signals that drive knowledge graph weight updates. By dynamically adjusting the correlation strength between semantic entities, the weight distribution of the knowledge graph aligns with the user's real-time preferences, while avoiding excessive interference from historical behaviors on the weights, ensuring that the weights remain within a reasonable range.

[0012] Furthermore, the process of updating the knowledge graph's associated content using incremental processing logic and controlling the knowledge graph's update cycle is as follows: New semantic entities and relationships not yet included in the existing knowledge graph are filtered from the associated dataset; semantic consistency is verified on the new content, and redundant and invalid information is removed; the verified new content is integrated into the existing knowledge graph, while the original structure of the knowledge graph's entities and relationships remains unchanged; simultaneously, a dual-cycle control mechanism is set up, using a short-cycle real-time iterative update of relationship weights and a long-cycle incremental update of the knowledge graph's associated content, balancing the knowledge graph's update efficiency with the accuracy of matching user preferences; the dual-cycle control mechanism includes a short-term weight iteration cycle and a long-term content update cycle, which operate independently yet collaboratively: the short-term weight iteration cycle uses a minute-level frequency to dynamically adjust the weights of relationships in the knowledge graph, quickly responding to users' immediate operational preferences; the long-term content update cycle uses an hour-level frequency to filter, verify, and integrate new semantic entities and relationships in the associated dataset, achieving incremental expansion of the knowledge graph's associated content while avoiding system resource consumption caused by high-frequency content updates.

[0013] Furthermore, the calculation formula for the weight-driven-editing parameter dynamic adaptation algorithm is as follows: ,in, Set the target clipping parameter value; This serves as the baseline value for the editing parameters; The adjustment coefficient is the factor that influences the weighting. The weights of semantic entity associations are averaged. Based on the updated knowledge graph association weights, a mapping relationship between user preferences and short video editing parameters is established. User behavior feedback preferences are transformed into the basis for adapting editing parameters, ensuring the adaptation of video editing style with content semantics and emotional expression. This supports the real-time personalized playback and batch efficient editing needs of short videos, generating optimized short video content that meets user preferences.

[0014] Furthermore, the dynamic reasoning link of semantic entity-emotion tag-editing parameter is a three-level progressive association structure. Specifically, it takes the semantic entities in the updated knowledge graph as the starting point for reasoning, matches the corresponding emotion tags according to the weights of the association relationships between semantic entities, and forms a first-level mapping of semantic entity-emotion tag; takes the emotion tags as the intermediate hub, combines the calculation results of the weight-driven-editing parameter dynamic adaptation algorithm, matches and adapts the editing parameter type and adjustment direction to the emotion expression, and constructs a second-level association of emotion tag-editing parameter; the link has the ability to dynamically adjust in real time, and can synchronously iterate the reasoning logic with the update of the knowledge graph relationship weights, ensuring that the association of semantic entities, emotion tags and editing parameters always fits the user's preferences, and realizes dynamic collaborative reasoning of the three-level link.

[0015] Compared with existing technologies, this knowledge graph-based AI short video content optimization method has the following beneficial effects: I. This invention constructs a foundational knowledge graph in the short video domain, clarifying multi-dimensional semantic entities and their relationships, providing a unified standard for data association and collection. Simultaneously, it employs a multi-level collaborative data tracking architecture to achieve synchronous collection and binding of user actions and video semantic information. Overcoming the shortcomings of traditional single-dimensional data collection, it deeply integrates scattered user behavior data, video content information, and domain knowledge systems to form a structured, associated dataset, giving the data clear semantic logic support. By transforming user behavior into knowledge graph weight adjustment signals, it achieves real-time iterative updates of the knowledge graph, ensuring more timely and accurate capture of user preferences, effectively improving data utilization efficiency and preference matching accuracy.

[0016] Second, this invention establishes a dynamic reasoning link between semantic entities and editing parameters, combined with weight-driven parameter adaptation logic, to achieve intelligent adjustment and dynamic adaptation of editing parameters. This meets the personalized presentation needs of real-time playback scenarios and adapts to the efficient processing flow of batch editing. The link design ensures that the video presentation format corresponds to user preferences and content semantics. Through a closed-loop optimization mechanism, user feedback data is collected and the process is iteratively optimized to achieve a full-link cyclical upgrade from data collection and graph updates to parameter adaptation. This continuously optimizes the presentation effect of short video content, enhances user experience, strengthens the market competitiveness of short video content, provides the industry with a more systematic and intelligent optimization solution, and promotes the production and optimization of short video content.

[0017] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0019] Figure 1 A flowchart illustrating the steps of a knowledge graph-based AI short video content optimization method; Figure 2 A flowchart of an AI short video content optimization method based on knowledge graphs; Figure 3 This is a schematic diagram of data transmission for an AI short video content optimization method based on knowledge graphs. Detailed Implementation

[0020] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0021] Example

[0022] This project constructs a foundational knowledge graph for the beauty tutorial field, employing a three-layer relational topology. It explicitly includes a semantic entity set, a set of relationships, and a set of relationship weights, providing core framework support for subsequent data association, weight updates, and parameter adaptation throughout the entire process. Within the semantic entity set, content-dimensional entities comprehensively cover core beauty content elements such as foundation, eyeshadow, lipstick, eyebrow pencil, blush, makeup steps, concealer techniques, base makeup methods, eye makeup tutorials, and lip makeup pairings. Form-dimensional entities encompass video presentation-related elements such as camera transitions, background music style, speech rate, subtitle size, close-up duration, transition effects, screen brightness, color saturation, voice-over tone, and editing rhythm. User-dimensional entities accurately categorize different user profiles, including beauty gurus, students, working professionals, people with sensitive skin, dry skin users, oily skin users, combination skin users, makeup beginners, and advanced learners. The relationship set system organizes the intrinsic connections between various entities, including mappings between content-dimensional entities and form-dimensional entities (such as the mapping between makeup steps and close-up duration, concealer techniques and camera cuts, foundation makeup methods and screen brightness, and eye makeup tutorials and transition effects). It also covers mappings between content-dimensional entities and user-dimensional entities (such as the mapping between foundation and sensitive skin, lipstick and working professionals, eyebrow pencil and students, and blush and beauty gurus). Furthermore, it includes mappings between form-dimensional entities (such as the mapping between background music style and speaking speed, subtitle size and close-up duration, color saturation and voice-over tone, and editing rhythm and camera cuts). The initial weights of the relationship weight set are set entirely based on prior knowledge in the beauty tutorial field. Each initial weight corresponds one-to-one with a mapping in the relationship set, ensuring that the importance of each relationship in the initial state aligns with industry consensus and basic user needs. This lays a solid foundation for subsequent dynamic adjustments based on user behavior, providing clear directional guidance for the entire optimization process from the outset.

[0023] A three-layer collaborative data collection architecture—front-end interaction layer, video content layer, and data transmission layer—is employed for data acquisition. This collaborative architecture ensures comprehensive, accurate, and synchronized data collection, guaranteeing the generation of high-quality correlated datasets. The front-end interaction layer deeply embeds data collection points into the short beauty tutorial video playback page and various client applications, collecting real-time user actions such as clicks, likes, comments, pauses, progress bar dragging, and replays—all actions that might reflect user preferences. It simultaneously records the timestamps, action types, and the start and end times of corresponding video segments. By fully capturing the details of user interaction with the video, it accurately locates beauty content segments that users are interested in, allowing subsequent analysis to directly focus on the user's core concerns. The video content layer meticulously extracts semantic entities, audio features, and camera parameters from corresponding video segments. Semantic entity extraction clarifies the core beauty content explained in the video; audio features capture auditory information such as background music style and voice-over tone; and camera parameters capture visual details such as close-ups, long shots, and transitions. This extraction of information comprehensively uncovers the core features of the video, providing ample material for matching with the basic knowledge graph. The data transmission layer uses dedicated tracking points to timestamp-align user action data with the semantic information of video clips, driving synchronous data collection and avoiding discrepancies between actions and content caused by asynchronous data collection. This ensures that each user action accurately matches the corresponding video content clip, guaranteeing data relevance and effectiveness. After collection, entity features are extracted from the semantic information of the collected video clips and subjected to comprehensive and detailed feature matching with the semantic entity set of the basic knowledge graph. Matching semantic entities are then selected, establishing a clear association between video content and core entities in the graph. Based on the matched semantic entities, corresponding relationships are retrieved from the basic knowledge graph's relationship set, bridging the gap between video content and entities / relationships. Finally, the four types of data—user action data, video clip semantic information, matched semantic entities, and corresponding relationships—are structurally integrated to form a structured user behavior-video semantic association dataset. This dataset organically merges various types of data, establishing a clear correspondence between user preferences and video content, graph entities, and relationships, providing a structured and analyzable data foundation for subsequent weight updates and parameter adaptation.

[0024] A real-time behavior-weight increment iterative algorithm is used to classify and assign values ​​to user behaviors in the associated dataset. Corresponding behavior coefficients are set according to the strength of preference representation, allowing different types of user behavior to be transformed into quantifiable preference indicators that accurately reflect users' preferences for different beauty content and presentation formats. Combining knowledge graph relationship weights and a time decay factor, the weight increment value of each association is calculated. The inclusion of the time decay factor highlights the influence of recent user behavior, ensuring that weight updates are more aligned with users' current preference status and avoiding interference from outdated behavior. The increment value is superimposed on the original weights to generate a weight adjustment signal. The corresponding association weights are updated based on the adjustment signal, allowing the knowledge graph weights to respond to changes in user behavior in real time, ensuring that the weights always accurately reflect the latest user preference trends. The formula is: ,in, The updated knowledge graph relationship weights; The original weights of the knowledge graph relationships before the update; User behavior coefficient; This is the time decay factor; This is the lower bound constraint function; The upper limit constraint function is used; simultaneously, incremental processing logic is employed to update the associated content of the knowledge graph. This involves precisely filtering new semantic entities and relationships not yet included in the existing knowledge graph from the associated dataset, such as new niche beauty brands, new makeup tools, and innovative makeup steps, as well as the relationships between these entities and existing entities, continuously enriching the coverage of the knowledge graph. Strict semantic consistency checks are performed on the new content, eliminating redundant and invalid information to ensure logical consistency and accuracy between the new content and the existing knowledge graph, preventing invalid information from affecting subsequent optimization processes. The validated new content is integrated into the existing knowledge graph, maintaining the original structure of entities and relationships, ensuring the stability and coherence of the graph structure. A dual-cycle control mechanism is implemented: short-cycle real-time iteration updates of relationship weights allow for rapid response to recent user behavior feedback, enabling timely adjustments to the optimization direction; long-cycle incremental updates of associated content ensure the knowledge graph continuously absorbs new content elements, constantly improving its coverage. This dual-cycle mechanism effectively balances the efficiency of knowledge graph updates with the accuracy of user preference matching, making the optimization process both efficient and accurate.

[0025] Based on the updated knowledge graph relationship weights, the adaptation parameter values ​​are calculated using a weight-driven dynamic adaptation algorithm for editing parameters. The formula is as follows: ,in, Set the target clipping parameter value; This serves as the baseline value for the editing parameters; The adjustment coefficient is the factor that influences the weighting. The algorithm uses the average weight of semantic entity associations to transform abstract weight relationships into concrete and operable editing parameters. This provides clear data support for adjusting editing parameters, ensuring their rationality and relevance. A dynamic reasoning link is constructed between semantic entities, emotion tags, and editing parameters. This link is a three-level progressive association structure, achieving accurate alignment from content semantics to emotional expression and presentation format through layer-by-layer association. Using semantic entities in the updated knowledge graph as the starting point for reasoning (e.g., "beginner," "foundation application steps," "sensitive skin," "concealing techniques"), the algorithm matches corresponding emotion tags (practical, clear, detailed, gentle) based on the weight of the associations between semantic entities. This forms a first-level mapping between semantic entities and emotion tags, directly linking video content to the user's desired emotional experience and ensuring the video accurately conveys emotional value that meets user needs. Using the emotional tags "practical," "clear," "detailed," and "gentle" as the central hub, and combining the calculation results of the weight-driven dynamic adaptation algorithm for editing parameters, this system matches the types and directions of editing parameters that are appropriate for the emotional expression. For example, it extends the close-up duration of foundation application steps, slows down the speaking pace when explaining concealer techniques, increases the size of subtitles for key steps, increases screen brightness, reduces color saturation, and uses smooth transitions. This constructs a secondary association between emotional tags and editing parameters, ensuring that adjustments to editing parameters accurately serve emotional expression and user needs, making the video presentation more aligned with user viewing habits and learning requirements. This chain possesses real-time dynamic adjustment capabilities, synchronously iterating the reasoning logic as the knowledge graph relationship weights are updated, ensuring that editing parameters continuously adjust to changes in user preferences and always maintain a high degree of adaptation to user needs. The calculated parameters are applied to the real-time playback and batch editing of beauty tutorial short videos, generating optimized short videos. Parameter adaptation during real-time playback provides a personalized viewing experience for individual users, while batch editing efficiently produces a large amount of high-quality content that meets the needs of different user groups, significantly improving content production efficiency and user experience.

[0026] Optimized beauty tutorial videos are accurately pushed to the target user group, continuously collecting new user behavioral feedback data, including various user interactions such as liking, commenting, pausing, replaying, sharing, and saving the optimized videos. This comprehensively captures user acceptance of the optimized content and new preference trends. Based on the newly collected behavioral feedback data, the process of dataset association collection, incremental update of graph weights, and dynamic adaptation of editing parameters is repeatedly executed to form a complete closed-loop iterative mechanism, such as... Figure 1As shown, through continuous iterative optimization, shortcomings in the optimization process can be identified in a timely manner, and the weights of relationships and semantic entity associations in the knowledge graph can be continuously corrected. This makes the adaptation of editing parameters more accurate, ensuring that video content remains close to the latest needs and preferences of users. This closed-loop optimization model can prevent content optimization from becoming stagnant, ensuring that beauty tutorial short videos maintain a high degree of adaptability to user needs in the long term, continuously improving user viewing time, interaction frequency, and satisfaction. At the same time, it enables content producers to more accurately grasp user needs and produce more competitive beauty tutorial content, achieving a joint improvement in user experience and content value.

[0027] Example

[0028] A foundational knowledge graph for the outdoor adventure field is constructed, employing a three-layer relational topology structure. This structure clearly defines a set of semantic entities, a set of relational relationships, and a set of relational weights. This provides a robust framework for the entire optimization process of outdoor adventure short videos, ensuring that all subsequent data processing and parameter adjustments revolve around these clearly defined entities and relationships. Within the semantic entity set, content-level entities comprehensively cover core outdoor adventure content such as mountaineering, camping, hiking routes, wilderness survival skills, adventure equipment, mountain trekking, stream exploration, stargazing camping, jungle exploration, and desert exploration. Form-level entities include video presentation elements such as aerial shots, fast-paced editing, environmental sound effects, subtitle colors, long shots, close-ups, slow motion, image contrast, voice-over style, and transition effects. User-level entities accurately categorize different user types, including novice adventurers, experienced adventurers, photography enthusiasts, families, youth groups, middle-aged adventurers, outdoor beginners, and extreme sports enthusiasts. The system of association sets organizes the inherent connections between various entities, including mappings between content-dimensional entities and form-dimensional entities, such as the mapping between hiking routes and aerial shots, wilderness survival skills and fast-paced editing, camping and close-up shooting, mountain trekking and slow-motion use, and stream exploration and environmental sound effects. It also covers mappings between content-dimensional entities and user-dimensional entities, such as the mapping between camping and families, mountaineering and youth groups, wilderness survival skills and experienced players, stargazing camping and photography enthusiasts, and desert exploration and extreme sports enthusiasts. Furthermore, it includes mappings between form-dimensional entities, such as the mapping between environmental sound effects and fast-paced editing, the mapping between subtitle color and long shot duration, the mapping between image contrast and voice-over style, and the mapping between transition effects and aerial shots. The initial weights of the relation weight set are set entirely based on prior knowledge in the field of outdoor adventure. Each initial weight corresponds one-to-one with the mapping relationship in the relation set, ensuring that the initial weights can reflect the general laws and basic user preferences in the field of outdoor adventure. This provides a reliable starting point for subsequent dynamic optimization based on actual user behavior, and ensures that the entire optimization process has a clear direction and basis from the beginning.

[0029] A three-layer collaborative data collection architecture—comprising a front-end interaction layer, a video content layer, and a data transmission layer—is employed for data acquisition. This collaborative architecture ensures comprehensive, accurate, and synchronized data collection, laying a solid foundation for generating high-quality correlated datasets. The front-end interaction layer's data collection is deeply embedded in the outdoor adventure short video playback page and various client applications, capturing real-time user actions such as clicks, likes, comments, pauses, progress bar dragging, and replays—all actions reflecting user interest. It simultaneously records the timestamps, action types, and the start and end times of corresponding video segments. By fully capturing the details of user interaction with the video, it accurately identifies outdoor adventure content segments that users are interested in, allowing subsequent data analysis to focus on the user's core concerns. The video content layer's data collection meticulously extracts semantic entities, audio features, and camera parameters from corresponding video segments. Semantic entity extraction clarifies the core outdoor adventure content presented in the video; audio features capture auditory information such as natural environmental sound effects and dubbing styles; and camera parameters capture visual details such as aerial shots, close-ups, and slow motion. This comprehensive extraction of information fully uncovers the core features of the video, providing ample and effective material for accurate matching with the basic knowledge graph. The data transmission layer uses dedicated event tracking to timestamp-align user action data with the semantic information of video clips, driving synchronous data collection. This avoids mismatches between user actions and video content caused by asynchronous data collection, ensuring that each piece of user action data is accurately associated with its corresponding video content clip, thus guaranteeing data relevance and validity. After collection, entity features are extracted from the semantic information of the collected video clips and subjected to comprehensive and detailed feature matching against the semantic entity set of the basic knowledge graph. Matching semantic entities are then selected, establishing a clear association between the video content and the core entities in the graph. Based on the matched semantic entities, corresponding relationships are retrieved from the basic knowledge graph's relationship set, building a tight connection between the video content, semantic entities, and relationships. Finally, the four types of data—user operation behavior data, video clip semantic information, matching semantic entities, and corresponding relationships—are structured and integrated to form a structured user behavior-video semantic association dataset. This dataset organically integrates various types of scattered data, allowing user preferences to form a clear and explicit correspondence with video content, graph entities, and relationships. This provides structured and analyzable data support for subsequent weight updates and editing parameter adaptation, ensuring that subsequent processes can be carried out efficiently and accurately.

[0030] A real-time behavior-weight incremental iterative algorithm is used to classify and assign values ​​to user behaviors in the associated dataset. Corresponding behavior coefficients are set according to the strength of preference representation, allowing different types of user behavior to be transformed into quantifiable preference indicators. This accurately reflects users' preferences for different outdoor adventure content and presentation formats, avoiding the ambiguity of user preferences. Combining knowledge graph relationship weights and a time decay factor, the incremental weight values ​​of each association are calculated. The inclusion of the time decay factor effectively highlights the influence of recent user behavior, making weight updates more aligned with users' current preference states. This prevents outdated user behavior from interfering with the current optimization direction, ensuring that weights always reflect the latest user needs and trends. The incremental values ​​are superimposed on the original weights to generate a weight adjustment signal. The corresponding association weights are updated based on the adjustment signal, allowing the knowledge graph weights to respond to changes in user behavior in real time, ensuring the timeliness and accuracy of the weights. Simultaneously, an incremental processing logic is employed to update the associated content of the knowledge graph. This involves accurately filtering newly added semantic entities and relationships not yet included in the existing knowledge graph from the associated dataset. Examples include newly added semantic entities such as niche hiking routes, new adventure equipment, and unique campsites, as well as the relationships between these entities and existing entities. This continuously enriches the coverage of the knowledge graph, ensuring it keeps pace with the development trends in the outdoor adventure field. Strict semantic consistency checks are performed on the new content, eliminating redundant and invalid information to guarantee logical consistency and accuracy between the new content and the existing knowledge graph. This prevents invalid information from affecting subsequent optimization processes and ensures the quality of the knowledge graph. The validated new content is then integrated into the existing knowledge graph, maintaining the original structure of entities and relationships. This ensures the stability and coherence of the graph structure, allowing subsequent weight updates and parameter adaptations to be based on a stable graph structure. A dual-cycle control mechanism is implemented: short-cycle real-time iteration updates of relationship weights allow for rapid responses to recent user behavior feedback, enabling timely adjustments to the optimization direction; long-cycle incremental updates of knowledge graph related content ensure that the knowledge graph continuously absorbs new content elements and constantly improves its coverage. This dual-cycle mechanism effectively balances the efficiency of knowledge graph updates with the accuracy of user preference matching, making the optimization process for outdoor adventure short videos both efficient and accurately aligned with user needs.

[0031] Based on the updated knowledge graph relationship weights, a weight-driven dynamic editing parameter adaptation algorithm is used to calculate adaptation parameter values. This algorithm transforms abstract weight relationships into concrete and operable editing parameters, providing clear data support for parameter adjustments, avoiding blind adjustments, and ensuring that parameter adjustments accurately serve user preferences. A dynamic reasoning link between semantic entities, emotion tags, and editing parameters is constructed. This link is a three-level progressive association structure, achieving accurate connection from the semantics of outdoor adventure content to emotional expression and then to the video presentation format through layer-by-layer association. Using semantic entities in the updated knowledge graph as the starting point for reasoning, such as semantic entities like beginner explorers, camping setup, families with children, and stream exploration, corresponding emotion tags such as easy, easy to understand, safe, and fun are matched according to the weights of the association relationships between semantic entities, forming a first-level mapping between semantic entities and emotion tags. This establishes a direct connection between outdoor adventure video content and the emotional experience expected by users, ensuring that the video accurately conveys emotional value that meets user needs and enhances user emotional resonance. Using the emotional tags "easy to understand," "safe," and "fun" as the central hub, and combining the calculation results of a weight-driven dynamic adaptation algorithm for editing parameters, this system matches the types and directions of editing parameters that are appropriate for the emotional expression. For example, it reduces the proportion of fast cuts, increases the font size of subtitles, extends the shot duration of key camping setup steps, enhances the clarity of environmental sound effects, improves image contrast, uses soft transition effects, and increases the dwell time of safety warning subtitles. This constructs a secondary association between emotional tags and editing parameters, ensuring that adjustments to editing parameters accurately serve emotional expression and user needs. This makes the video presentation more suitable for the viewing habits and acceptance abilities of different user groups, especially making it easier for novice adventurers and families with children to understand the video content. This chain has real-time dynamic adjustment capabilities, synchronously iterating the reasoning logic as the knowledge graph relationship weights are updated, ensuring that editing parameters continuously adjust to changes in user preferences and always maintain a high degree of adaptation to user needs. The calculated parameters are applied to the real-time playback and batch editing of outdoor adventure short videos to generate optimized videos. Parameter adaptation during real-time playback allows for a personalized viewing experience for individual users, while batch editing efficiently produces a large amount of high-quality outdoor adventure content that meets the needs of different user groups, such as... Figure 3 As shown, this significantly improves content production efficiency and user experience, allowing more users to find outdoor adventure videos that meet their needs.

[0032] Optimized outdoor adventure short videos are accurately pushed to the target user group, continuously collecting new user behavioral feedback data, including various interactive behaviors such as liking, commenting, pausing, replaying, sharing, saving, and forwarding of the optimized videos. This comprehensively and meticulously captures user acceptance, satisfaction, and new preference trends regarding the optimized content. Based on the newly collected behavioral feedback data, the process of dataset association collection, incremental update of graph weights, and dynamic adaptation of editing parameters is repeatedly executed, forming a complete and efficient closed-loop iterative mechanism, such as... Figure 2 As shown, through continuous iterative optimization, shortcomings and deviations in the optimization process can be identified in a timely manner. This allows for continuous correction of the weights of relationships and semantic entity associations in the knowledge graph, making the editing parameters more accurate and better suited to user needs. This ensures that outdoor adventure short video content remains closely aligned with the latest user demands and preferences. This closed-loop optimization model effectively prevents content optimization from becoming stagnant, ensuring that outdoor adventure short videos maintain a high degree of adaptability to user needs in the long term, continuously increasing user viewing time, interaction frequency, sharing willingness, and satisfaction. Simultaneously, it enables content producers to more accurately grasp user needs and market trends, producing more attractive and competitive outdoor adventure content, achieving a simultaneous improvement in user experience, content value, and market competitiveness.

[0033] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A knowledge graph-based AI short video content optimization method, characterized in that, The specific steps of this method are as follows: S100. Constructing a basic knowledge graph: Building a basic knowledge graph for the short video domain, clarifying the semantic entity set, the set of relationships, and the set of relationship weights in the knowledge graph, and setting initial weights for each set; The set of semantic entities and the set of relationships serve as the benchmark for the collection of dataset associations; S200, Dataset Association Collection: Multi-level tracking technology is used throughout the short video playback process to synchronously collect real-time user operation behavior and semantic information of corresponding video segments; the collected semantic information is bound to the corresponding semantic entities and relationships in the basic knowledge graph to generate a structured user behavior-video semantic association dataset. S300, Incremental Update of Knowledge Graph Weights: Based on the generated associated dataset, the real-time behavior-weight incremental iteration algorithm is used to transform user behavior into knowledge graph weight adjustment signals, and dynamically update the relation weights of the knowledge graph. It also employs incremental processing logic to update the associated content of the knowledge graph, thereby controlling the update cycle of the knowledge graph; S400, Dynamic Adaptation of Editing Parameters: Based on the updated knowledge graph relationship weights, the weight-driven dynamic adaptation algorithm for editing parameters is used to calculate the adaptation parameter values; a dynamic reasoning link of semantic entity-emotion tag-editing parameter is constructed, and the calculated parameters are applied to the real-time playback and batch editing of short videos to generate optimized short videos; S500, Closed-Loop Optimization: Push optimized short videos to users, collect new user behavior feedback data, repeatedly execute the data set association collection, incremental update of graph weights and dynamic adaptation of editing parameters, and complete the loop iteration.

2. The AI ​​short video content optimization method based on knowledge graph according to claim 1, characterized in that, In step S100, the basic knowledge graph of the short video domain is a three-layer relational topology, including a semantic entity set, a relation set, and a relation weight set. The semantic entity set includes content-dimensional entities, form-dimensional entities, and user-dimensional entities. The relation set includes mapping relationships between content-dimensional entities and form-dimensional entities, mapping relationships between content-dimensional entities and user-dimensional entities, and mapping relationships between form-dimensional entities. The initial weights of the relation weight set are set according to prior knowledge of the short video domain, and each initial weight corresponds one-to-one with the mapping relationship in the relation set.

3. The AI ​​short video content optimization method based on knowledge graph as described in claim 1, characterized in that, In step S200, the multi-level tracking technology adopts a three-layer collaborative tracking architecture consisting of a front-end interaction layer, a video content layer, and a data transmission layer: the front-end interaction layer tracking is embedded in the short video playback page and client, collecting user actions such as clicking, liking, commenting, pausing, dragging the progress bar, and repeating playback in real time, and synchronously recording the action timestamp, action type, and the start and end time positions of the corresponding video segment; the video content layer tracking extracts the semantic entities, audio features, and shot parameters of the corresponding video segment; Data transmission layer embedding performs timestamp alignment operations on user operation behavior data and semantic information of video clips, driving the synchronous collection of both.

4. The AI ​​short video content optimization method based on knowledge graph according to claim 1, characterized in that, In step S200, the process of binding the collected semantic information with the corresponding semantic entities and relationships in the basic knowledge graph to generate a structured user behavior-video semantic association dataset is as follows: extract entity features from the collected video segment semantic information, perform feature matching with the semantic entity set of the basic knowledge graph, and filter out the corresponding matching semantic entities; based on the matched semantic entities, retrieve the corresponding relationships in the relationship set of the basic knowledge graph; and structurally integrate the four types of data—user operation behavior data, video segment semantic information, matching semantic entities, and corresponding relationships—to form a structured user behavior-video semantic association dataset.

5. The AI ​​short video content optimization method based on knowledge graph according to claim 1, characterized in that, In step S300, the specific steps of using the real-time behavior-weight increment iteration algorithm to convert user behavior into knowledge graph weight adjustment signals and dynamically update the relation weights of the knowledge graph are as follows: using the real-time behavior-weight increment iteration algorithm, classifying and assigning values ​​to user behaviors in the associated dataset, and setting corresponding behavior coefficients according to the intensity of preference representation; combining the knowledge graph relation weights and the time decay factor to calculate the weight increment value of each relation; superimposing the increment value onto the original weights to generate a weight adjustment signal; and updating the corresponding relation weights according to the adjustment signal.

6. The AI ​​short video content optimization method based on knowledge graph according to claim 5, characterized in that, In step S300, the calculation formula for the real-time behavior-weight increment iteration algorithm is as follows: ,in, The updated knowledge graph relationship weights; The original weights of the knowledge graph relationships before the update; User behavior coefficient; This is the time decay factor; This is the lower bound constraint function; This is the upper limit constraint function.

7. The AI ​​short video content optimization method based on knowledge graph according to claim 1, characterized in that, In step S300, the process of updating the associated content of the knowledge graph using incremental processing logic and controlling the update cycle of the knowledge graph is as follows: new semantic entities and new relationships not included in the existing knowledge graph are selected from the associated dataset; semantic consistency is checked on the new content, and redundant and invalid information is removed; the new content that passes the check is integrated into the existing knowledge graph, while the original structure of the entities and relationships in the knowledge graph remains unchanged; at the same time, a dual-cycle control mechanism is set up, with short-cycle real-time iterative updates of relationship weights and long-cycle incremental updates of the associated content of the knowledge graph, balancing the update efficiency of the knowledge graph with the accuracy of user preference matching.

8. The AI ​​short video content optimization method based on knowledge graph according to claim 1, characterized in that, In step S400, the calculation formula for the weight-driven-editing parameter dynamic adaptation algorithm is as follows: ,in, Set the target clipping parameter value; This serves as the baseline value for the editing parameters; The adjustment coefficient is the factor that influences the weighting. This represents the average weight of semantic entity association relationships.

9. The AI ​​short video content optimization method based on knowledge graph according to claim 1, characterized in that, In step S400, the semantic entity-emotion tag-editing parameter dynamic reasoning link is a three-level progressive association structure, specifically: taking the semantic entities in the updated knowledge graph as the starting point of reasoning, matching the corresponding emotion tags according to the weight of the association relationship between semantic entities to form a first-level mapping of semantic entity-emotion tag; taking the emotion tag as the intermediate hub, combining the calculation results of the weight-driven-editing parameter dynamic adaptation algorithm, matching and adapting the editing parameter type and adjustment direction to the emotion expression to construct a second-level association of emotion tag-editing parameter; The link has the ability to dynamically adjust in real time, and can synchronously iterate the reasoning logic as the weights of the knowledge graph relationships are updated.

Images