A method, device and medium for ad delivery based on video interest point prediction

By recording user video playback behavior and extracting multi-dimensional operation sequence feature vectors to generate an interest dust cloud set, the ad delivery strategy is dynamically controlled, solving the problem of inaccurate ad insertion timing in existing technologies and achieving efficient and accurate ad delivery and optimized user experience.

CN123123620APending Publication Date: 2026-11-17BEIJING KUAIDU TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611073371.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-20
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In existing technologies, the timing of video ad insertion cannot be precisely and dynamically adjusted based on the user's micro-interest in the video content, and it is difficult to control the timing of ad insertion at the frame level, resulting in a poor user experience.

Method used

By recording user video playback behavior, extracting multi-dimensional operation sequence feature vectors, quantifying user interest intensity, and generating an interest dust cloud set, we can dynamically control advertising delivery strategies and achieve precise ad insertion by combining frame-level playback progress monitoring.

Benefits of technology

It enables precise ad delivery based on user interests, improving ad delivery efficiency and user experience, reducing ad insertion interference, and increasing ad click-through rate and conversion rate.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123123620A_ABST
    Figure CN123123620A_ABST
Patent Text Reader

Abstract

This invention discloses an advertising delivery method, device, and medium based on video interest point prediction, belonging to the field of video data processing and user behavior analysis technology. It addresses the technical problems in existing technologies where the timing of video ad insertion cannot be precisely and dynamically adjusted based on users' micro-interests in video content, and it is difficult to dynamically control ad insertion timing with frame-level precision. The method includes: recording video operation behaviors as player operation types; extracting multi-dimensional operation sequence features from repeated viewing behaviors in the player operation event stream; quantifying the interest intensity of single repeated viewing behaviors based on basic interest intensity, time decay factor, and social enhancement factor; mapping single repeated viewing behaviors to topic tags based on the granularity coefficient of repeated video segments; and dynamically controlling ad insertion for target videos based on high-interaction plot nodes.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the fields of video data processing and user behavior analysis, and in particular to an advertising delivery method, device and medium based on video interest point prediction. Background Technology

[0002] With the rapid development of fragmented video content such as electronic short dramas and short videos, how to reasonably insert advertisements during video playback, so as to ensure the exposure effect of advertisements without damaging the user's viewing experience, has become an important technical problem for video platforms.

[0003] Currently, video ads are inserted primarily using the following methods: The first method is the fixed-interval insertion method, which inserts an advertisement at preset intervals during video playback. This method completely ignores the user's viewing status and changes in interest, suddenly inserting advertisements when the user is immersed in an exciting storyline, which is very likely to arouse user resentment and even cause the user to quit midway.

[0004] The second method is scene transition insertion, which involves inserting ads when a significant change in the video frame is detected. While this method reduces the disruption of the storyline caused by ads to some extent, it still cannot differentiate between users' varying interests in different segments.

[0005] The third method is interest determination based on overall viewing time. This method determines a user's level of interest in a video by counting the number of times they watch it completely or the total viewing time. This approach treats the video as a whole and cannot pinpoint which specific plot points a user is interested in.

[0006] However, a common shortcoming of existing technologies is the lack of refined modeling of the micro-behavior of repeated viewing, particularly neglecting the operational sequence characteristics inherent in the repeated viewing behavior itself. This includes how users rewind, the speed of rewinding, whether they adjust the playback speed after rewinding, and whether they linger repeatedly at specific locations. Moreover, these operational trajectories themselves carry far richer interest information than simply whether or not the user rewatches the video. Furthermore, existing technologies cannot precisely and dynamically adjust the timing of video ad insertion based on the user's micro-interest in the video content, and it is difficult to dynamically control the ad insertion timing with frame-level precision. Summary of the Invention

[0007] This application provides an advertising delivery method, device, and medium based on video interest point prediction to solve the following technical problem: In the prior art, the timing of video advertising insertion cannot be accurately and dynamically adjusted according to the user's micro-interest in the video content, and it is difficult to dynamically control the timing of advertising insertion at the frame level.

[0008] The embodiments of this application adopt the following technical solutions: On one hand, this application provides an advertising delivery method based on video interest point prediction, including: recording user video operation behaviors as player operation types according to the user's player operation timeline in the video playback client, generating a player operation event stream; extracting multi-dimensional operation sequence features based on single repeated viewing behaviors from the repeated viewing behaviors in the player operation event stream, obtaining a multi-dimensional operation sequence feature vector corresponding to each repeated viewing behavior; and using the multi-dimensional operation sequence feature vector, measuring the interest intensity of single repeated viewing behaviors in terms of basic interest intensity, time decay factor, and social enhancement factor. The process involves several steps: First, a multidimensional interest signal vector is obtained from the single repeated viewing behavior. Based on this vector, a topic tag mapping process based on the granularity coefficient of the repeated video segments is performed. Then, based on the dust contribution value of the repeated video segments, an interest dust cloud set representing the user's repeated viewing interest preferences is generated. Based on the interest dust cloud set and the corresponding dust cloud turbulence state, dynamic ad insertion control is applied to the target video for high-interaction plot node types, resulting in a dynamic ad delivery strategy. After the dynamic ad delivery strategy is executed, the user's actual interaction behavior data in the target video is used for feedback optimization based on the dust contribution value.

[0009] This application's embodiments record the user's operation timeline while playing videos, providing detailed event recording of user behavior. This allows for a more accurate understanding of the user's specific interests and preferences regarding video content, providing precise user behavior data for targeted advertising. It can also identify users' continued interest in repeated viewings, providing a basis for consistency and continuity of user behavior in subsequent advertising. Furthermore, it can map topic tags based on interest intensity vectors, generating an interest dust cloud set, visually representing user interests and providing an intuitive reference for ad targeting. Simultaneously, through continuous learning and adjustment, the advertising strategy better aligns with actual user responses, improving the efficiency and effectiveness of advertising.

[0010] In one feasible implementation, the repeated viewing behavior in the player operation event stream is processed by extracting multi-dimensional operation sequence features based on a single repeated viewing behavior to obtain a multi-dimensional operation sequence feature vector corresponding to each repeated viewing behavior. Specifically, this includes: performing playback progress rewind judgment on the player operation event stream based on a fixed time point to determine the repeated viewing behavior and its corresponding repetition count; calculating the rewind distance of the operation event stream corresponding to the repeated viewing behavior at the current playback position to determine a rewind span feature representing spatial span; calculating the dragging speed of the operation event stream corresponding to the repeated viewing behavior to determine a rewind speed feature representing operation urgency; and further processing the operation event stream corresponding to the repeated viewing behavior... The process involves repeatedly calculating the amplitude range of repeated video segments to determine the hovering frequency feature, which represents the positioning accuracy of the repeated video segments. The operation event stream corresponding to the repeated viewing behavior is then monitored for the pause duration of the repeated video segments to obtain the dwell gaze duration feature, which represents the gaze depth. The playback speed of the repeated video segments is recorded within the operation event stream corresponding to the repeated viewing behavior to obtain the slow-motion magnification sequence feature, which represents the rhythm control. Finally, the backswing span feature, backswing speed feature, hovering frequency feature, dwell gaze duration feature, and slow-motion magnification sequence feature in a single repeated viewing behavior are normalized into a sequence feature vector, and integrated to obtain the multi-dimensional operation sequence feature vector corresponding to each repeated viewing behavior.

[0011] The embodiments of this application help understand how users interact with video content, thereby better designing user experiences and personalized recommendations. They can also provide detailed information about the depth of user interests, helping to improve the accuracy of content recommendations and ad placement. Furthermore, they can help identify the degree of user attention to different parts of a video, providing a basis for video editing and ad insertion. They also help understand users' sustained attention to specific content, providing guidance for content optimization and ad targeting. Simultaneously, they ensure that feature vectors are on the same scale, facilitating subsequent data analysis and model training, and improving the accuracy of prediction and recommendation.

[0012] In one feasible implementation, the interest intensity of a single repeated viewing behavior is quantified using the multidimensional operation sequence feature vector, considering basic interest intensity, time decay factor, and social enhancement factor, to obtain a multidimensional interest signal vector for the single repeated viewing behavior. Specifically, this includes: calculating the basic interest intensity of the single repeated viewing behavior by weighting each operation sequence feature in the multidimensional operation sequence feature vector according to the difference in contribution of different operation features to interest intensity; calculating the time decay factor of the single repeated viewing behavior by performing an inverse proportional decay calculation on the time variable from the time of occurrence of the behavior to the current time based on the absolute time of occurrence of the single repeated viewing behavior; identifying whether a social sharing event exists in the single repeated viewing behavior through adjacent time windows, and determining the social enhancement factor used to represent the social sharing behavior; and performing a fusion calculation of the basic interest intensity, the time decay factor, and the social enhancement factor to determine the multidimensional interest signal vector of the single repeated viewing behavior; wherein, the multidimensional interest signal vector is the final interest weight of the user based on repeated video segments under the single repeated viewing behavior.

[0013] This application's embodiments, through weighted fusion calculations, can quantify the contribution of different operational features to interest intensity, providing a precise measure of user interest intensity and helping to more accurately predict user preferences. Furthermore, based on a defined time decay factor, it can consider the impact of time factors on user interests, making interest intensity assessment more consistent with dynamically changing user interests. Simultaneously, by leveraging the influence of social networks, it enhances the understanding of user interests and improves the social relevance of the recommendation system.

[0014] In one feasible implementation, based on the multidimensional interest signal vector, the single repeated viewing behavior is processed by topic tag mapping based on the granularity coefficient of repeated video segments, and an interest dust cloud set representing the user's repeated viewing interest preferences is generated based on the dust contribution value of the repeated video segments. Specifically, this includes: obtaining the target time span length of the repeated video segments corresponding to the single repeated viewing behavior; performing multi-level granularity classification processing on the target time span length based on different video duration spans according to a time span range threshold to obtain granularity coefficients; wherein the granularity coefficients include: microgranularity coefficients, medium granularity coefficients, and macro granularity coefficients; extracting information on keyword information, scene type information, and emotional tone information from the repeated video segments using dialogue subtitle features, scene annotation features, and audio features; and processing the keyword information, scene type information, and emotional tone information according to a preset plot node type library. The scene type information and the emotional tone information are subjected to tag matching processing, and the tag with the highest matching degree is mapped to the topic tag of the repeated video segment; based on all repeated video segments of the user in a preset time window, the multidimensional interest signal vector corresponding to each repeated video segment is fused and calculated with the granularity coefficient to determine the dust contribution value of each repeated video segment; the dust contribution value is distributed according to the topic tag corresponding to each repeated video segment; if a repeated video segment corresponds to multiple topic tags, the dust contribution value is equally distributed to each corresponding topic tag; the topic tags and corresponding dust contribution values ​​of all repeated video segments are accumulated and aggregated to obtain the interest dust cloud set; wherein, each topic tag is a dust particle in the interest dust cloud, and the corresponding accumulated dust contribution value represents the density of the dust particle.

[0015] This application's embodiments map topic tags based on the granularity coefficients of repeated video clips, enabling more refined analysis of video content and identification of user interests at different granularity levels. Simultaneously, by utilizing introduced granular, medium-granular, and macro-granularity coefficients, multi-level granularity classification is provided, allowing adjustment of the analysis granularity according to different needs, thus improving the applicability and accuracy of the analysis results. Furthermore, by combining dialogue subtitles, scene annotations, and audio feature extraction of keywords, scene types, and emotional tone, multi-dimensional feature extraction is employed to provide a more comprehensive video content analysis. Additionally, an interest dust cloud set is generated based on dust contribution values, visually presenting the distribution and density of user interests to help understand overall user interest preferences. It can also more deeply identify user interest preferences for repeated video clips, supporting personalized recommendations.

[0016] In one feasible implementation, after generating an interest dust cloud set representing a user's repeated viewing interest preferences, the method further includes: monitoring the increment of dust contribution value for each topic tag within a unit time window; if the increment of dust contribution value is three standard deviations of the historical average daily increment, then the topic tag is identified as a topic tag in a turbulent state; wherein, the turbulent state indicates that the user's interest distribution is in a non-steady state; prioritizing the topic tags in the turbulent state as candidate nodes, and determining them as priority topic tags; accelerating the decay rate of interest time corresponding to the priority topic tags to simulate the rapid decline of sudden user interest, and obtaining the interest decay rate; based on the priority topic tags and the interest decay rate, determining the interest dust cloud corresponding to the topic tags in the turbulent state as the dust cloud turbulent state; wherein, topic tags in the turbulent state are used to capture sudden changes in user interest, and topic tags in the normal state are used to capture the user's steady-state interest preferences.

[0017] In one feasible implementation, based on the interest dust cloud set and the corresponding dust cloud turbulence state, dynamic ad insertion control is applied to the target video for high-interaction plot node types to obtain a dynamic ad delivery strategy. Specifically, this includes: extracting target theme tags and timestamps from each video segment in the target video according to a preset plot node type library; comparing the target theme tags of each video segment with the theme tags in the interest dust cloud set, and calculating the cumulative dust contribution value corresponding to the target theme tags in the interest dust cloud for each video segment; determining whether each video segment is in the dust cloud turbulence state, identifying the interest dust clouds of video segments in the dust cloud turbulence state, and applying an additional turbulence enhancement coefficient to the cumulative dust contribution value in the interest dust clouds in the dust cloud turbulence state to obtain an enhanced cumulative dust contribution value; and predicting target video segments of high-interaction plot node types based on the cumulative dust contribution value and the enhanced cumulative dust contribution value. The system first identifies the target video segment as a candidate node. Based on the frame-level playback progress monitoring mechanism of the target video, if the current playback position of the target video is not within the progress range of the candidate node, the existing ad insertion strategy is executed. If the current playback position of the target video is within the progress range of the candidate node, active suppression control of ad insertion is performed during the playback of the video segment corresponding to the candidate node, resulting in an ad insertion suppression strategy. After the video segment corresponding to the candidate node finishes playing, starting from the non-center area of ​​the video playback window, the transparency of the ad information window to be inserted is gradually reduced, and the current position of the ad information window is controlled to move closer to the center area in a step-like manner until it smoothly transitions into a complete ad information window in the center area, resulting in an ad insertion suppression recovery strategy. Based on the existing ad insertion strategy, the ad insertion suppression strategy, and the ad insertion suppression recovery strategy, the dynamic ad delivery strategy is generated.

[0018] This application's embodiments, by analyzing the interest dust cloud set and turbulence state in the target video and combining it with the plot node type, can more accurately predict highly interactive plot nodes, thereby achieving more precise ad placement. Furthermore, it can intelligently control the timing of ad insertion based on the video playback progress and the progress range of candidate nodes, avoiding frequent ad interruptions during user viewing and improving user experience. By applying a turbulence enhancement coefficient to video segments in dust cloud turbulence state, ad information is made more prominent at key nodes, increasing ad exposure and interaction rates. Moreover, the ad insertion strategy can be adjusted in real time based on video playback progress and user interaction, making ad placement more flexible and efficient. This achieves precise ad placement and optimized user experience, increasing ad click-through rates and conversion rates, thereby increasing ad revenue. Finally, this dynamic ad placement strategy is also applicable to different types of video content and can be adjusted according to actual needs, exhibiting strong adaptability and versatility.

[0019] In one feasible implementation, after the dynamic advertising delivery strategy is executed, the user's actual interaction behavior data in the target video is optimized for feedback regarding the dust contribution value. Specifically, this includes: after the dynamic advertising delivery strategy is executed, if the user's actual interaction behavior data at a candidate node shows repeated viewing behavior, then the cumulative dust contribution value in the candidate node's topic tag and corresponding interest dust cloud is enhanced and controlled; and the turbulence state detection threshold of the corresponding topic tag is lowered; if the user's actual interaction behavior data at a candidate node does not show repeated viewing behavior, then the cumulative dust contribution value in the candidate node's topic tag and corresponding interest dust cloud is attenuated and controlled; and the turbulence state detection threshold of the corresponding topic tag is increased; based on the positive and negative feedback of the cumulative dust contribution value and the turbulence state, the feedback optimization of the interest dust cloud and the dust cloud turbulence state is completed.

[0020] In one feasible implementation, based on the user's player operation timeline in the video playback client, the user's video operation behavior is processed by recording player operation type events to generate a player operation event stream. Specifically, this includes: collecting each playback operation event performed by the user on the video player in the video playback client through an event-driven mechanism; wherein the event types of the playback operation events include at least: playback event, pause event, progress dragging event, speed adjustment event, full-screen switching event, and share event; generating the player operation timeline for each user based on the time sequence of each playback operation event; and sorting all playback operation events according to the player operation timeline to generate the player operation event stream related to the video operation behavior.

[0021] Secondly, embodiments of this application also provide an advertising delivery device based on video point of interest prediction, the device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to execute an advertising delivery method based on video point of interest prediction as described in any of the above embodiments.

[0022] Thirdly, embodiments of this application also provide a non-volatile computer storage medium, which is a non-volatile computer-readable storage medium storing at least one program. Each program includes instructions, which, when executed by a terminal, cause the terminal to execute an advertising delivery method based on video interest point prediction as described in any of the above embodiments.

[0023] This application provides an advertising delivery method, device, and medium based on video interest point prediction. Compared with the prior art, the embodiments of this application have the following beneficial technical effects: 1. By extracting five orthogonal features—rewind span, rewind speed, number of hovers, duration of gaze, and slow-motion magnification sequence—the same repeated viewing behavior is quantified into a five-dimensional feature vector, thereby enabling precise differentiation of different interest states.

[0024] 2. By detecting the increase in abnormal dust values ​​within a unit time window, the system identifies the user's interest turbulence state and dynamically adjusts the prediction priority and decay rate, so that the predicted candidate nodes have both stability and sensitivity.

[0025] 3. By combining the phased and gradual presentation of ad preview icons with the start of playback from the middle of the ad, the integration of ad insertion and user experience is made compatible and unified.

[0026] 4. All user behavior data is passively collected during the natural interaction between the user and the player. No additional rating operation is required from the user, nor is it necessary to deploy hardware such as facial recognition cameras. The data is highly authentic, the deployment cost is low, and user privacy is guaranteed.

[0027] 5. By comparing the prediction results with actual user behavior and feeding them back to the dust cloud and turbulence detection parameters, the system achieves self-optimization and continuous evolution, and the prediction accuracy continues to improve with the increase of usage time. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 A flowchart of an advertising delivery method based on video interest point prediction is provided for an embodiment of this application; Figure 2 This is a schematic diagram of an advertising delivery device based on video point of interest prediction, provided as an embodiment of this application. Detailed Implementation

[0029] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0030] It should be noted that this application provides a video interest point prediction and advertising control method based on user repeated viewing behavior sequence feature modeling. First, the original player operation event stream is collected to provide a data foundation; then, repeated viewing behavior and its complete operation sequence features are identified and extracted from the operation event stream; next, the operation sequence features are quantized into multi-dimensional interest signal vectors; and the quantized behavior signals are associated and mapped with the semantic tags of the video content and aggregated into a computable structured interest model; then, based on this model, frame-level precision prediction is performed and advertising control decisions are output; finally, a feedback loop is formed by real-time monitoring of the user's actual behavior at the prediction node.

[0031] This application provides an advertising delivery method based on video interest point prediction, such as... Figure 1 As shown, the ad delivery method based on video interest point prediction specifically includes steps S101-S106: S101. Based on the user's player operation timeline in the video playback client, perform player operation type event recording processing on the user's video operation behavior and generate a player operation event stream.

[0032] Specifically, an event-driven mechanism is first used to collect every playback operation event performed by the user on the video player. These playback operation events include at least the following types: playback events, pause events, progress dragging events, speed adjustment events, full-screen switching events, and share events.

[0033] Furthermore, based on the temporal order of each playback operation event, a player operation timeline is generated for each user. Then, according to the player operation timeline, all playback operation events are sorted into event streams to generate a player operation event stream related to video operation behavior.

[0034] In one implementation, while a user watches an electronic short drama through a video playback client, the system collects every user operation on the video player in real time using an event-driven approach. Unlike traditional methods that only record repeated viewings, this application records the complete operation event stream. Each playback operation event includes at least the following fields: {user identifier, video identifier, event type, playback timestamp when the event occurred, additional event parameters, absolute time of event occurrence}. All the above playback operation events are arranged in chronological order to form a player operation timeline for each user. Finally, based on the player operation timeline, all playback operation events are sorted into an event stream to generate a player operation event stream related to video operation behavior.

[0035] As a possible implementation method, the event types include, but are not limited to: Playback event: the user clicks the play button, with the additional parameter being the current playback position; Pause event: the user clicks the pause button, with the additional parameter being the playback position when paused; Progress drag event: the user drags the progress bar, with the additional parameters being the start and end positions of the drag; Speed ​​adjustment event: the user adjusts the playback speed, with the additional parameter being the speed values ​​before and after the adjustment; Fullscreen switch event: the user switches to fullscreen mode, with the additional parameter being the switching direction; Sharing event: the user shares on social media, with the additional parameters being the sharing channel and the sharing time.

[0036] S102. Extract the multi-dimensional operation sequence features based on a single repeated viewing behavior from the player operation event stream to obtain the multi-dimensional operation sequence feature vector corresponding to each repeated viewing behavior.

[0037] Specifically, the playback progress is first rewound based on a fixed time point in the player operation event stream to determine the repeated viewing behavior and the corresponding number of repetitions.

[0038] In one embodiment, when the system detects the following operation mode, it is determined as a repeated viewing behavior: the user performs a progress dragging event at time point T1, rewinds the playback position from the current progress P_current to an earlier progress P_target (P_target < P_current), and the time span between P_target and P_current is greater than or equal to 1 second. The system records the target segment of this repeated viewing as the interval [P_target, P_current].

[0039] Further, the rewind distance of the current playback position is calculated for the operation event stream corresponding to the repeated viewing behavior, so as to determine the rewind span feature for representing the spatial span.

[0040] As a feasible implementation, for the rewind span feature (R): the distance the user rewinds from the current playback position, the calculation formula is R=P_current-P_target. The rewind span reflects the volume of content that the user wants to re-experience. A larger span indicates that the user is willing to spend more time re-experiencing a relatively long segment of content; a smaller span indicates that the user focuses on a certain extremely short moment. This is a continuous value feature.

[0041] Further, the dragging speed related to video progress dragging is calculated for the operation event stream corresponding to the repeated viewing behavior, so as to determine the rewind speed feature for representing the operation eagerness.

[0042] As a feasible implementation, for the rewind speed (V) feature: the dragging speed when the user performs the progress dragging operation, the calculation formula is V=R / T_drag, where T_drag is the time elapsed from when the user starts dragging to releasing the progress bar. The rewind speed reflects the eagerness degree of the user's operation. Fast rewind (high V value) means that the user has a strong immediate urge to rewatch the target segment; slow rewind (low V value) may mean that the user is also looking for a certain uncertain position during the rewinding process. And this is a continuous value feature.

[0043] Further, the repeated advance and retreat calculation of the amplitude range related to the repeated video segment is performed on the operation event stream corresponding to the repeated viewing behavior, so as to determine the wandering times feature for representing the positioning accuracy of the repeated video segment.

[0044] As a feasible implementation, regarding the hover count (W) feature: after the rewind operation of a single repeated viewing behavior is completed and before the user continues to play forward normally, the system monitors whether the user repeatedly makes small forward and backward adjustments within the target segment [P_target, P_current]. Each time a minor adjustment is detected (e.g., the user makes a progress adjustment of less than 3 seconds within the target segment), the hover count is incremented by 1. The hover count reflects the user's need for precise positioning at a specific smaller position within the target segment; the user may not be interested in the entire [P_target, P_current] interval, but rather in a specific instant within the interval, but the initial rewind did not accurately locate that instant, thus requiring repeated fine-tuning to find it; therefore, it is a non-negative integer feature.

[0045] Furthermore, the pause duration of repeated video segments is monitored in the operation event stream corresponding to the repeated viewing behavior to obtain the dwell gaze duration feature used to represent the gaze depth.

[0046] As a feasible implementation method, regarding the dwell gaze duration (G) feature: after the user completes rewind and begins replaying the target segment, the system monitors the user's pausing behavior during the playback of the target segment. If the user performs a pause operation at a certain position in the target segment and the pause duration exceeds 2 seconds, then the pause duration is added to the dwell gaze duration. Dwell gaze duration reflects the user's deep gaze at a specific scene or line during repeated viewing; that is, the user is not only willing to rewind and rewatch, but also willing to stop and examine it carefully during rewatching. This is also a continuous value feature.

[0047] Furthermore, the playback speed of the repeated video segments is recorded in the operation event stream corresponding to the repeated viewing behavior to obtain the slow-motion magnification sequence feature used to represent rhythm control.

[0048] As a feasible implementation method, regarding the slow-motion magnification sequence (S) feature: during the user's repeated viewing of the target segment, the system records each time the user adjusts the playback speed, forming a time series of slow-motion magnification. For example, if the user first plays at 0.75x speed for 3 seconds, then adjusts to 0.5x speed for 2 seconds, and then returns to 1.0x speed, the slow-motion magnification sequence is [(0.75, 3s), (0.5, 2s)]. The slow-motion magnification sequence reflects the user's differentiated attention to different sub-segments during repeated viewing; that is, the lower the magnification and the longer the duration, the more carefully the user wants to savor the content of that sub-segment.

[0049] Furthermore, the features of backslide span, backslide speed, number of hovers, duration of gaze, and slow-motion magnification sequence in a single repeated viewing behavior are normalized into sequence feature vectors, and integrated to obtain the multi-dimensional operation sequence feature vector corresponding to each repeated viewing behavior.

[0050] In one embodiment, since the total duration of different videos varies, the absolute values ​​of features such as the rewind span R are not directly comparable between different videos. Therefore, the system normalizes each feature: R is divided by the total duration of the current video to obtain the normalized rewind span R_norm; V is divided by the maximum draggable speed of the video player to obtain the normalized rewind speed V_norm; W and G are already relative values ​​and do not require additional normalization; S is normalized by weighted summation of the reciprocal of the slow-motion magnification (i.e., 1 / magnification) multiplied by the duration. Finally, based on the five normalized features, the same repeated viewing behavior is characterized from five completely different and mutually orthogonal dimensions: spatial span, operation urgency, positioning accuracy, gaze depth, and rhythm control, forming a complete five-dimensional operation sequence feature vector, i.e., the multi-dimensional operation sequence feature vector: F=(R, V, W, G, S).

[0051] S103. By using the multidimensional operation sequence feature vector, the interest intensity of a single repeated viewing behavior is quantified by considering the basic interest intensity, time decay factor, and social enhancement factor, resulting in a multidimensional interest signal vector for the single repeated viewing behavior.

[0052] Specifically, based on the differences in the contribution of different operational features to the intensity of interest, the weights of each operational sequence feature in the multidimensional operational sequence feature vector need to be fused and calculated to obtain the basic intensity of interest for a single repeated viewing behavior.

[0053] In one embodiment, the base interest intensity can be calculated by fusing the feature vectors of the five-dimensional operation sequence: I_base = α·R_norm + β·V_norm + γ·W + δ·G_norm + ε·S_norm; where α, β, γ, δ, and ε are preset weight coefficients, for example: α = 0.20, β = 0.25, γ = 0.20, δ = 0.20, ε = 0.15. The values ​​of each weight coefficient reflect the differences in the contribution of different operation features to the interest intensity. That is, the backoff speed V_norm is considered to be the most direct indicator of impulse intensity, and therefore is given the highest weight.

[0054] Furthermore, based on the absolute time of occurrence of a single repeated viewing behavior, the time variable between the time of occurrence and the current time is calculated inversely proportionally to determine the time decay factor of the single repeated viewing behavior.

[0055] In one embodiment, the time decay factor can be calculated based on the absolute time of a single repeated viewing behavior. The decay model uses inverse proportional decay instead of exponential decay, and the time decay factor is: T_decay = 1 / (1 + λ·Δt); where Δt is the number of days between the single behavior and the current time, and λ is a preset decay coefficient (λ = 0.03). It should be noted that the advantage of inverse proportional decay over exponential decay is that it does not assign excessive weight to recent behaviors, nor does it allow the weight of earlier behaviors to approach zero. This allows the interest-dust cloud to respond more smoothly to changes in user interests, avoiding drastic fluctuations in prediction results caused by a user's occasional intensive behavior on a particular day.

[0056] Furthermore, by using adjacent time windows of single repeated viewing behavior, the existence of social sharing events in single repeated viewing behavior is identified and determined, thereby identifying social enhancement factors used to represent social sharing behavior.

[0057] In one embodiment, if a social sharing event occurs within a preset time window before and after a single repeated viewing behavior (e.g., from the first 5 minutes to the last 5 minutes), then the social enhancement factor E_social = 1.8; otherwise, E_social = 1.0.

[0058] Furthermore, the basic interest intensity, time decay factor, and social enhancement factor are fused and calculated to determine the multidimensional interest signal vector for a single repeated viewing behavior. This multidimensional interest signal vector represents the final interest weight of the user based on repeated video segments under a single repeated viewing behavior.

[0059] In one embodiment, the final interest weight of a single repeated viewing behavior, which is the multidimensional interest signal vector, can be expressed as: Weight = I_base × T_decay × E_social.

[0060] S104. Based on the multidimensional interest signal vector, perform topic tag mapping processing on the single repeated viewing behavior based on the granularity coefficient of repeated video segments, and generate an interest dust cloud set to represent the user's repeated viewing interest preference based on the dust contribution value of repeated video segments.

[0061] Specifically, the target time span length of the repeated video segments corresponding to a single repeated viewing behavior is first obtained. Then, based on the time span range threshold, the target time span length is subjected to multi-level granularity classification processing under different video duration spans to obtain granularity coefficients. Among them, the granularity coefficients include: microgranularity coefficients, medium granularity coefficients, and macro granularity coefficients.

[0062] In one embodiment, for each repeated viewing action, the system can classify it into one of the following three granularity levels based on the time span length of its target segment [P_target, P_current], and assign a corresponding granularity coefficient: 1) Micro: Time span ≤ 3 seconds, corresponding to the single-shot level. Granularity coefficient = 1.0. 2) Meso: Time span > 3 seconds and ≤ 10 seconds, corresponding to the single line of dialogue or a single continuous action level. Granularity coefficient = 2.0. 3) Macro: Time span > 10 seconds, corresponding to the entire dialogue or complete scene level. Granularity coefficient = 3.5.

[0063] Furthermore, it is necessary to extract relevant keyword information, scene type information, and emotional tone information from repeated video clips using dialogue subtitle features, scene annotation features, and audio features. Then, based on a pre-defined plot node type library, the keyword information, scene type information, and emotional tone information are subjected to tag matching, and the tag with the highest matching degree is mapped to the theme tag of the repeated video clip.

[0064] In one implementation example, the system can also extract semantic theme tags for each repeating video segment through video dialogue subtitles, scene annotation information, and audio feature analysis. The theme tags are derived from a pre-built plot node type library, including but not limited to: misunderstanding resolution, identity revelation, intense fight, emotional confession, crisis outbreak, plot twist, farewell scene, reunion moment, suspense revelation, climactic showdown, and other plot node types. Furthermore, in mapping theme tags, keywords (extracted from dialogue), scene type (obtained from visual analysis), and emotional tone (obtained from audio analysis) within the segment are first extracted. These three are then matched against the plot node type library, and the 1-2 tags with the highest matching degree are selected as the theme tags for the repeating video segment.

[0065] Furthermore, based on all the repeated video clips of the user within the preset time window, it is also necessary to fuse and calculate the multidimensional interest signal vector and granularity coefficient corresponding to each repeated video clip to determine the dust contribution value of each repeated video clip.

[0066] Furthermore, the dust contribution value is allocated according to the topic tag corresponding to each repeated video segment; if a repeated video segment corresponds to multiple topic tags, the dust contribution value is equally allocated to each corresponding topic tag.

[0067] Furthermore, the topic tags and corresponding dust contribution values ​​of all repeated video clips are aggregated to obtain an interest dust cloud set. Each topic tag represents a dust particle within the interest dust cloud, and the corresponding cumulative dust contribution value represents the density of the dust particle.

[0068] In one embodiment, the system first aggregates all repeated viewing records of a user within a preset time window (e.g., the past 60 days), that is, aggregates all repeated video clips; then, it constructs the user's interest dust cloud. For each repeated video clip, its final interest weight (multidimensional interest signal vector) is multiplied by its corresponding granularity coefficient to obtain the dust contribution value for that record: Dust_contribution = Weight × granularity coefficient. Then, this dust contribution value is allocated according to the topic tags corresponding to the record. If a record corresponds to multiple topic tags, the dust contribution value is equally distributed among the tags. Afterwards, the topic tags and corresponding dust contribution values ​​of all repeated video clips are cumulatively aggregated. After aggregation, the user's interest dust cloud presents as a set of topic tags and their corresponding cumulative dust values, i.e., the interest dust cloud set. Each topic tag can be regarded as a dust particle in the dust cloud, and its cumulative dust value represents the density of that particle. The higher the density of the topic tag, the stronger the user's preference for that type of plot node.

[0069] Furthermore, after obtaining the interest dust cloud, it is also necessary to monitor the increase in dust contribution value for each topic tag within a unit time window. If the increase in dust contribution value is three standard deviations of the historical average daily increase, then the topic tag is identified as being in a turbulent state. Here, turbulence refers to a non-steady-state distribution of user interests.

[0070] In one embodiment, turbulence as defined in this application refers to a user repeatedly viewing the same topic tag with an abnormally high frequency within a short period of time (e.g., within 24 hours). The specific detection method is as follows: the system monitors the dust value increment of each topic tag within a unit time window. If the dust value increment of a topic tag within 24 hours exceeds three times the standard deviation of the historical average daily increment for that topic tag, then the topic tag is determined to have experienced turbulence, i.e., it is identified as a topic tag in a turbulent state.

[0071] It should be noted that this application's detection of the turbulent state of topic tags can identify sudden changes in user interests. That is, a user may suddenly experience a strong peak of interest in a particular type of plot point. This peak may be triggered by the user's recent life experiences, social influences, or accidental content exposure. In the turbulent state, the distribution of user interests is non-steady, and conventional predictions based on historical data may fail.

[0072] Furthermore, the topic tags in the turbulent state are given priority by candidate nodes and identified as priority topic tags. Then, the interest decay rate corresponding to the priority topic tags is accelerated to simulate the rapid decline of users' sudden interest and obtain the interest decay rate.

[0073] Furthermore, based on priority topic tags and interest decay rates, the interest dust clouds corresponding to topic tags in a turbulent state are identified as being in a turbulent dust cloud state. Topics in a turbulent state are used to capture sudden changes in user interest, while topics in a normal state are used to capture users' steady-state interest preferences.

[0074] In one embodiment, when the system detects that a certain topic tag is in a turbulent state, the system performs the following two operations: 1) Temporarily increase the priority of the corresponding topic tag in later predictions: In subsequent video ad insertion control, the candidate node corresponding to this topic tag will receive a higher avoidance priority (i.e., earlier and more decisive suppression of ad insertion). 2) Activate turbulence decay control: After turbulence occurs, the dust contribution value of this topic tag has its time decay rate temporarily accelerated (the decay coefficient is temporarily increased to twice the normal value) to simulate the rapid decay characteristics of sudden interest and prevent sudden interest from having an excessive impact on long-term prediction. In other words, by utilizing the turbulence state detection and response mechanism, this application can simultaneously capture users' stable interest preferences and sudden interest changes, achieving a clever balance between prediction stability and sensitivity.

[0075] S105. Based on the interest dust cloud set and the corresponding dust cloud turbulence state, dynamically insert advertisements related to highly interactive plot nodes into the target video to obtain a dynamic advertisement delivery strategy.

[0076] Specifically, it is also necessary to use a pre-defined plot node type library to extract the target topic tags and timestamps of each video segment in the target video. After coal combustion, the target topic tags of each video segment are compared with the topic tags in the interest dust cloud set, and the cumulative dust contribution value of the target topic tag in each video segment in the interest dust cloud is calculated.

[0077] Furthermore, each video segment is further processed to determine whether it is in a dust cloud turbulent state. The dust clouds of interest in the video segments that are in a dust cloud turbulent state are identified, and the cumulative dust contribution value in the dust clouds of interest in a dust cloud turbulent state is further processed with a turbulence enhancement coefficient to obtain the enhanced cumulative dust contribution value.

[0078] Furthermore, based on the cumulative dust contribution value and the enhanced cumulative dust contribution value, target video segments in the high-interaction plot node type are predicted, and the target video segments are identified as candidate nodes.

[0079] In one embodiment, when a user is about to watch a new electronic short drama (target video), the system needs to preprocess each video segment of the target video, extracting the topic tag, start timestamp, and end timestamp of each segment (with frame-level precision, i.e., 1 / 30th of a second). Then, the system compares the topic tags of each video segment in the target video with the topic tags in the user's interest dust cloud, calculating the basic interest matching degree of each video segment, i.e., the cumulative dust value contribution corresponding to the topic tag of that video segment in the user's interest dust cloud. Simultaneously, the system also checks whether the topic tag of the video segment is in a turbulent state. If so, a turbulence enhancement coefficient (e.g., 1.5 times) is applied to the basic interest matching degree of the segment to further enhance the cumulative dust contribution value in the interest dust cloud that is in a turbulent state, resulting in an enhanced cumulative dust contribution value. Finally, video segments whose final interest matching degree, represented by the cumulative dust contribution value, exceeds a preset threshold are identified and marked as candidate nodes where users may generate high interactive behavior. In other words, the cumulative dust contribution value and the enhanced cumulative dust contribution value are used to predict target video segments of the target video with high interactive plot node type. Finally, the video segment with the highest confidence is predicted and then the target video segment is determined as a candidate node.

[0080] Furthermore, a frame-level playback progress monitoring mechanism based on the target video is also needed. If the current playback position of the target video is not within the progress range of the candidate node, the existing ad insertion strategy is executed. If the current playback position of the target video is within the progress range of the candidate node, active ad insertion suppression control is applied during the playback of the video segment corresponding to the candidate node, resulting in an ad insertion suppression strategy.

[0081] In one embodiment, during the actual playback of the target video by the user, the system monitors the playback progress with frame-level precision (checking once every 1 / 30th of a second). When the playback progress is about to enter a segment that has been marked as a candidate node, the system executes the following advertising control logic: If the current playback position is not within the range of any candidate node, the system executes the normal advertising insertion strategy, i.e., executes the existing advertising insertion strategy. If the current playback position is within the range of candidate nodes, the system actively suppresses advertising insertion, i.e., no advertising is inserted during the playback of the video segment corresponding to the candidate node, and the normal advertising insertion strategy is restored after the segment has finished playing, i.e., the advertising insertion suppression strategy is implemented. This application's frame-level control of playback progress enables advertising insertion decisions with precision down to the single frame, avoiding false negatives due to insufficient precision, i.e., preventing the incorrect insertion or omission of advertisements at the boundaries of candidate nodes.

[0082] Furthermore, after the video segment corresponding to the candidate node has finished playing, starting from the non-central area of ​​the video playback window, the transparency of the advertisement window to be inserted needs to be gradually reduced, and the current position of the advertisement window needs to be controlled to move closer to the central area in a step-like manner until it smoothly transitions into a complete advertisement window in the central area, thus obtaining the suppression and recovery advertisement insertion strategy. Finally, based on the existing advertisement insertion strategy, the suppression advertisement insertion strategy, and the suppression and recovery advertisement insertion strategy, a dynamic advertisement delivery strategy can be generated.

[0083] In one embodiment, when the candidate node finishes playing and the system needs to resume ad insertion, this application employs a preemptive smooth recovery mechanism instead of simple immediate insertion. In other words, traditional methods insert ads immediately after the candidate node ends, which causes users to be abruptly interrupted by ads after experiencing a highly interesting segment, resulting in a strong psychological letdown; the user's newly aroused enthusiasm is forcibly cut short by the ad, a worse experience than inserting ads during low-interest segments. This application's preemptive smooth recovery mechanism begins pre-recovery operations within 5 seconds of the candidate node ending: 1) From the first second to the third second: The system displays a tiny ad preview icon with extremely low transparency (90% transparency) in a non-central area of ​​the video playback screen (such as the lower right corner). This icon contains no dynamic content and is merely a static brand logo. At this time, the user can hardly perceive the existence of the icon.

[0084] 2) Between the 3rd and 5th seconds: The system gradually reduces the transparency of the ad preview icon to 60% and slightly increases its size, but still keeps it in the off-center area.

[0085] 3) At the 5th second after the candidate node ends: the system smoothly transitions the ad preview icon into the full ad content window, and the ad starts playing in the middle of the ad content rather than at the beginning, that is, it starts playing from the 5th or 10th second of the ad, rather than from the 0th second.

[0086] As a feasible implementation method, the preemptive smooth recovery mechanism of the advertising information window in this application can gradually present the preview icon in stages, allowing users to receive a weak visual signal that an ad is about to appear before the ad officially appears. This prepares the user's mind and significantly reduces the startling and resistant feeling when the ad suddenly appears. Furthermore, the ad starts playing from the middle section, avoiding the strong brand exposure at the beginning. While sacrificing some ad exposure, this results in a natural transition of user emotions; users don't feel their emotions are abruptly interrupted, but rather perceive a natural transition to a different section. Thus, the overall goal of zero ad interference at high-interest nodes and no loss of ad exposure opportunities is achieved synergistically.

[0087] S106. After the dynamic advertising delivery strategy is executed, the actual interactive behavior data of users in the target video will be used to optimize the dust contribution value.

[0088] Specifically, after the dynamic advertising delivery strategy is implemented, if users exhibit repeated viewing behavior in the actual interaction data of candidate nodes, the topic tags of the candidate nodes and the cumulative dust contribution value in the corresponding interest dust cloud will be subject to enhanced control. Furthermore, the turbulence state detection threshold for the corresponding topic tags will be lowered.

[0089] In one embodiment, after a user finishes watching a new video, the system collects the user's actual interactive behavior data in the video (e.g., whether the user engaged in repeated viewing, slow-motion, or sharing at the predicted candidate nodes), and compares this data with the predicted results of the candidate nodes to generate feedback data. If the user's actual interactive behavior data at the candidate node shows repeated viewing behavior, i.e., positive feedback (prediction hit): the user did indeed engage in repeated viewing or other interactive behavior at the candidate node. The system then slightly enhances the cumulative dust value of the topic tag corresponding to the candidate node in the dust cloud (e.g., enhancement coefficient 1.05), while fine-tuning the turbulence state detection threshold, slightly lowering the turbulence state detection threshold for that topic tag to make it more sensitive to similar future signals.

[0090] Furthermore, if the user's actual interaction data at the candidate node does not show repeated viewing behavior, then the cumulative dust contribution value in the candidate node's topic tag and corresponding interest dust cloud will be attenuated. The turbulence state detection threshold for the corresponding topic tag will also be increased.

[0091] In one embodiment, if the user's actual interaction data at a candidate node does not show repeated viewing behavior, i.e., negative feedback (prediction miss): the user did not generate any interaction behavior at the candidate node, the system slightly attenuates the cumulative dust value of the topic tag corresponding to the candidate node in the dust cloud (e.g., attenuation coefficient 0.95), while increasing the turbulence state detection threshold of the topic tag to reduce the false alarm rate.

[0092] As a feasible implementation method, positive and negative feedback between the cumulative dust contribution value and the turbulence state is used to optimize the dust cloud of interest and its turbulence state. In other words, the above feedback mechanism enables the detection parameters of the dust cloud of interest and its turbulence state to continuously self-calibrate during long-term use, gradually improving the prediction accuracy. More importantly, the feedback data itself is also incorporated into the operation event stream and becomes the input data for the next iteration.

[0093] In addition, embodiments of this application also provide an advertising delivery device based on video interest point prediction, such as... Figure 2 As shown, the advertising delivery device 200 based on video interest point prediction specifically includes: At least one processor 201; and a memory 202 communicatively connected to the at least one processor 201; wherein the memory 202 stores instructions executable by the at least one processor 201 to enable the at least one processor 201 to execute: Based on the user's player operation timeline in the video playback client, the user's video operation behavior is processed by recording player operation type events and generating a player operation event stream. The repeated viewing behavior in the player operation event stream is processed by extracting multi-dimensional operation sequence features based on a single repeated viewing behavior, so as to obtain the multi-dimensional operation sequence feature vector corresponding to each repeated viewing behavior. By using multidimensional operation sequence feature vectors, the interest intensity of single repeated viewing behavior is quantified by considering basic interest intensity, time decay factor, and social enhancement factor, resulting in a multidimensional interest signal vector for single repeated viewing behavior. Based on the multidimensional interest signal vector, the single repeated viewing behavior is processed by topic tag mapping based on the granularity coefficient of repeated video segments, and based on the dust contribution value of repeated video segments, an interest dust cloud set representing the user's repeated viewing interest preference is generated. Based on the interest dust cloud set and the corresponding dust cloud turbulence state, dynamic ad insertion control is applied to the target video for high-interaction plot nodes to obtain a dynamic ad delivery strategy. After the dynamic ad delivery strategy is implemented, the actual interactive behavior data of users in the target video will be used to optimize the dust contribution value.

[0094] This application's embodiments record the user's operation timeline while playing videos, providing detailed event recording of user behavior. This allows for a more accurate understanding of the user's specific interests and preferences regarding video content, providing precise user behavior data for targeted advertising. It can also identify users' continued interest in repeated viewings, providing a basis for consistency and continuity of user behavior in subsequent advertising. Furthermore, it can map topic tags based on interest intensity vectors, generating an interest dust cloud set, visually representing user interests and providing an intuitive reference for ad targeting. Simultaneously, through continuous learning and adjustment, the advertising strategy better aligns with actual user responses, improving the efficiency and effectiveness of advertising.

[0095] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device and medium embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the description of the method embodiments.

[0096] The devices and media provided in this application are one-to-one with the methods. Therefore, the devices and media also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the devices and media will not be repeated here.

[0097] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0098] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations 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, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0099] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0100] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0101] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0102] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0103] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, 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 technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0104] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0105] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of this specification.

Claims

1. An advertising delivery method based on video interest point prediction, characterized in that, The method includes: Based on the user's player operation timeline in the video playback client, the user's video operation behavior is processed by recording player operation type events and generating a player operation event stream. The repeated viewing behavior in the player operation event stream is processed by extracting multi-dimensional operation sequence features based on a single repeated viewing behavior to obtain the multi-dimensional operation sequence feature vector corresponding to each repeated viewing behavior. The multidimensional operation sequence feature vector is used to quantify the interest intensity of a single repeated viewing behavior by considering the basic interest intensity, time decay factor, and social enhancement factor, thereby obtaining a multidimensional interest signal vector for the single repeated viewing behavior. Based on the multidimensional interest signal vector, the single repeated viewing behavior is processed by topic tag mapping based on the granularity coefficient of repeated video segments, and an interest dust cloud set representing the user's repeated viewing interest preference is generated based on the dust contribution value of repeated video segments. Based on the set of interest dust clouds and the corresponding dust cloud turbulence state, dynamic ad insertion control is performed on the target video for high-interaction plot node types to obtain a dynamic ad delivery strategy. After the dynamic advertising delivery strategy is executed, the actual interactive behavior data of users in the target video will be used to optimize the dust contribution value.

2. The advertising delivery method based on video interest point prediction according to claim 1, characterized in that, The repeated viewing behavior in the player operation event stream is processed by extracting multi-dimensional operation sequence features based on a single repeated viewing behavior to obtain a multi-dimensional operation sequence feature vector corresponding to each repeated viewing behavior, specifically including: The playback progress is rewound based on a fixed time point in the player operation event stream to determine the repeated viewing behavior and the corresponding number of repetitions; The back-away distance of the operation event stream corresponding to the repeated viewing behavior is calculated at the current playback position to determine the back-away span feature used to represent the spatial span. The dragging speed of the operation event stream corresponding to the repeated viewing behavior is calculated to determine the rewind speed feature used to represent the urgency of the operation. The operation event stream corresponding to the repeated viewing behavior is subjected to repeated advance and retreat calculations on the amplitude range of the repeated video segments to determine the hovering number feature used to represent the positioning accuracy of the repeated video segments; The pause duration of repeated video segments is monitored in the operation event stream corresponding to the repeated viewing behavior to obtain the dwell gaze duration feature used to represent the gaze depth. The playback speed of the repeated video segments is recorded in the operation event stream corresponding to the repeated viewing behavior to obtain the slow motion magnification sequence feature used to represent rhythm control. The features of backslide span, backslide speed, number of hovers, duration of gaze, and slow-motion magnification sequence in a single repeated viewing behavior are normalized into sequence feature vectors, and then integrated to obtain the multidimensional operation sequence feature vector corresponding to each repeated viewing behavior.

3. The advertising delivery method based on video interest point prediction according to claim 1, characterized in that, Using the multidimensional operation sequence feature vector, the interest intensity of a single repeated viewing behavior is quantified based on the basic interest intensity, time decay factor, and social enhancement factor, resulting in a multidimensional interest signal vector for the single repeated viewing behavior, specifically including: Based on the differences in the contribution of different operational features to the intensity of interest, the weights of each operational sequence feature in the multidimensional operational sequence feature vector are fused and calculated to obtain the basic intensity of interest for a single repeated viewing behavior. Based on the absolute time of the occurrence of the single repeated viewing behavior, the time variable of the time from the time of the behavior occurrence to the current time is calculated inversely proportionally to determine the time decay factor of the single repeated viewing behavior. By using the adjacent time windows of the single repeated viewing behavior, the existence of a social sharing event is identified and determined, and the social enhancement factor used to represent the social sharing behavior is determined. The basic interest intensity, the time decay factor, and the social enhancement factor are fused and calculated to determine the multidimensional interest signal vector of the single repeated viewing behavior; wherein, the multidimensional interest signal vector is the final interest weight of the user based on repeated video segments under the single repeated viewing behavior.

4. The advertising delivery method based on video interest point prediction according to claim 1, characterized in that, Based on the multidimensional interest signal vector, the single repeated viewing behavior is processed by topic tag mapping based on the granularity coefficient of repeated video segments, and an interest dust cloud set representing the user's repeated viewing interest preferences is generated based on the dust contribution value of the repeated video segments, specifically including: Obtain the target time span length of the repeated video segment corresponding to a single repeated viewing behavior; Based on the time span range threshold, the target time span length is subjected to multi-level granularity classification processing based on different video duration spans to obtain granularity coefficients; wherein, the granularity coefficients include: microgranularity coefficients, medium granularity coefficients, and macro granularity coefficients; By using dialogue subtitle features, scene annotation features, and audio features, information such as keyword information, scene type information, and emotional tone information is extracted from the repeated video segments. Based on a preset plot node type library, the keyword information, scene type information, and emotional tone information are subjected to tag matching processing, and the tag with the highest matching degree is mapped as the theme tag of the repeated video clip; Based on all repeated video segments of the user within a preset time window, the multidimensional interest signal vector corresponding to each repeated video segment is fused with the granularity coefficient to determine the dust contribution value of each repeated video segment. The dust contribution value is allocated according to the topic tag corresponding to each repeated video segment; if a repeated video segment corresponds to multiple topic tags, the dust contribution value is equally allocated to each corresponding topic tag. The interest dust cloud set is obtained by accumulating and aggregating several topic tags and corresponding dust contribution values ​​of all repeated video clips; wherein each topic tag is a dust particle in the interest dust cloud, and the corresponding cumulative dust contribution value represents the density of the dust particle.

5. The advertising delivery method based on video interest point prediction according to claim 4, characterized in that, After generating an interest dust cloud set to represent a user's repeated viewing interest preferences, the method further includes: Monitor the increase in dust contribution value for each topic tag within a unit time window; if the increase in dust contribution value is three standard deviations of the historical average daily increase, then the topic tag is identified as being in a turbulent state; wherein, the turbulent state is when the user's interest distribution is in a non-steady state; Topics in turbulent flow are prioritized and designated as priority topics. The interest decay rate corresponding to the priority topic tags is accelerated to simulate the rapid decline of users' sudden interest and obtain the interest decay rate. Based on the preferred topic tags and the interest decay rate, the interest dust cloud corresponding to the topic tags in the turbulent state is determined as the dust cloud turbulent state; wherein, topic tags in the turbulent state are used to capture sudden changes in user interest, and topic tags in the normal state are used to capture user steady-state interest preferences.

6. The advertising delivery method based on video interest point prediction according to claim 1, characterized in that, Based on the aforementioned set of interest dust clouds and the corresponding dust cloud turbulence states, dynamic ad insertion control is applied to the target video for high-interaction storyline nodes, resulting in a dynamic ad delivery strategy, specifically including: Based on a preset library of plot node types, extract the target theme tags and timestamps of each video segment in the target video; The target topic tags of each video segment are compared with the topic tags in the interest dust cloud set, and the cumulative dust contribution value of the target topic tags in each video segment in the interest dust cloud is calculated. Each video segment is processed to determine whether it is in the turbulent state of the dust cloud. The dust cloud of interest in the video segment in the turbulent state is determined. The cumulative dust contribution value in the dust cloud of interest in the turbulent state is further processed by the turbulence enhancement coefficient to obtain the enhanced cumulative dust contribution value. Based on the cumulative dust contribution value and the enhanced cumulative dust contribution value, target video segments of the high-interaction plot node type are predicted, and the target video segments are identified as candidate nodes. Based on the frame-level playback progress monitoring mechanism of the target video, if the current playback position of the target video is not within the progress range of the candidate node, the existing ad insertion strategy will be executed. If the current playback position of the target video is within the progress range of the candidate node, then active suppression control of ad insertion will be performed during the playback of the video segment corresponding to the candidate node to obtain an ad insertion suppression strategy. After the video segment corresponding to the candidate node finishes playing, starting from the non-center area of ​​the video playback window, the transparency of the advertisement information window to be inserted is gradually reduced, and the current position of the advertisement information window is controlled to move closer to the center area in a step-like manner until it smoothly transitions into a complete advertisement information window in the center area, thus obtaining the suppression and recovery advertisement insertion strategy. The dynamic ad delivery strategy is generated based on the existing ad insertion strategy, the ad insertion suppression strategy, and the ad insertion suppression and recovery strategy.

7. The advertising delivery method based on video interest point prediction according to claim 1, characterized in that, After the dynamic ad delivery strategy is executed, the user's actual interaction data in the target video is used to optimize the dust contribution value, specifically including: After the dynamic advertising delivery strategy is executed, if the user's actual interaction behavior data at the candidate node shows repeated viewing behavior, the topic tags of the candidate node and the cumulative dust contribution value in the corresponding interest dust cloud will be enhanced and controlled; and the turbulence state detection threshold of the corresponding topic tag will be reduced. If the user's actual interaction data at the candidate node does not show repeated viewing behavior, then the topic tag of the candidate node and the cumulative dust contribution value in the corresponding interest dust cloud will be attenuated; and the turbulence state detection threshold of the corresponding topic tag will be increased. Based on the positive and negative feedback of the cumulative dust contribution value and the turbulence state, feedback optimization of the dust cloud of interest and the turbulence state of the dust cloud is completed.

8. The advertising delivery method based on video interest point prediction according to claim 1, characterized in that, Based on the user's player operation timeline in the video playback client, the user's video operation behavior is processed into player operation type event records, generating a player operation event stream, specifically including: Through an event-driven mechanism, every playback operation event performed by the user on the video player in the video playback terminal is collected; wherein, the event types of the playback operation events include at least: playback event, pause event, progress dragging event, speed adjustment event, full-screen switching event, and share event; Based on the temporal order of each playback operation event, a player operation timeline for each user is generated. Based on the player operation timeline, all playback operation events are sorted into event streams to generate the player operation event stream related to video operation behavior.

9. An advertising delivery device based on video point of interest prediction, characterized in that, The device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor to enable the at least one processor to perform an advertising delivery method based on video interest point prediction according to any one of claims 1-8.

10. A non-volatile computer storage medium, characterized in that, The storage medium is a non-volatile computer-readable storage medium storing at least one program, each program including instructions that, when executed by a terminal, cause the terminal to perform an advertising delivery method based on video point of interest prediction according to any one of claims 1-8.