Video recommendation method and device, equipment and storage medium

Abstract

The embodiment of the invention provides a video recommendation method and device, equipment and a storage medium. The method specifically comprises the steps of obtaining interaction behavior data of a user for playing a video; under the condition that the interactive behavior data indicates that the negative feedback behavior occurs, performing feature extraction on the played video to obtain at least one disgust feature; determining the aversion score of the user for the aversion feature according to the basic aversion intensity, the continuous strike weight and the time attenuation factor corresponding to the aversion feature and the time interval between the occurrence time of the negative feedback behavior and the current time; and obtaining candidate videos in a pre-generated video recommendation list, and updating the candidate videos in the video recommendation list according to the aversion scores corresponding to the aversion features included in the candidate videos and a preset score threshold.

Description

Video recommendation methods, devices, equipment and storage media Technical Field

[0001] This invention relates to the field of Internet data mining and recommendation technology, and in particular to a video recommendation method, apparatus, device and storage medium. Background Technology

[0002] Currently, video recommendation typically employs methods based on models such as collaborative filtering, DeepFM, or Wide&Deep. The core technology primarily involves calculating the probability of user preferences based on click-through rate (CTR) prediction, thereby generating a personalized recommendation list. This approach is essentially a "positive-driven" mechanism, where the main goal is to continuously identify and reinforce existing user interests to maximize positive interaction metrics such as clicks and dwell time.

[0003] However, this "positive-driven" design has the following significant technical limitations: First, the recommendation mechanism overfits users' historical preferences, leading to a severe information cocoon effect and recommending a large amount of highly similar content, thus narrowing the user's field of vision; second, the negative feedback mechanism is singular and inefficient, mainly relying on users actively clicking the "not interested" button, and this operation usually only applies to a single video, failing to achieve generalized identification and blocking of content with similar negative characteristics; finally, existing algorithms lack effective management of user aesthetic fatigue, making it difficult to distinguish between permanent aversion and temporary boredom, resulting in a decline in user experience due to receiving a large amount of repetitive content in a short period, increasing the risk of churn. Summary of the Invention

[0004] The main objective of this invention is to provide a video recommendation method, apparatus, device, and storage medium, which aims to solve the problems of existing video recommendation technologies, such as the solidification of user information cocoons, difficulty in effectively identifying negative intentions and aesthetic fatigue, and imbalance of content ecosystem diversity, caused by reliance on "positive-driven" design and overfitting to short-term feedback.

[0005] In a first aspect, embodiments of this disclosure provide a video recommendation method, comprising: acquiring user interaction behavior data on a played video; when the interaction behavior data indicates the occurrence of negative feedback behavior, extracting features from the played video to obtain at least one aversion feature; the aversion feature includes at least one of the following: audio features, visual features, and semantic features; determining the user's aversion score for the aversion feature based on the basic aversion intensity, combo weight, time decay factor, and the time interval between the occurrence time of the negative feedback behavior and the current time; wherein, the basic aversion intensity is determined based on the behavior type of the negative feedback behavior; the combo weight is determined based on the number of consecutive occurrences of the negative feedback behavior within a preset time window; the time decay factor is configured based on a preset attribute category of the aversion feature; the preset attribute category is used to determine whether the aversion feature is a rapidly decaying first feature representing the user's aesthetic fatigue, or a slowly decaying second feature representing the user's persistent rejection; acquiring candidate videos from a pre-generated video recommendation list, and updating the candidate videos in the video recommendation list based on the aversion score corresponding to the aversion feature included in the candidate videos, and a preset score threshold.

[0006] Secondly, embodiments of this disclosure provide a video recommendation device, comprising: a first acquisition module, configured to acquire user interaction behavior data regarding a played video; a second acquisition module, configured to extract features from the played video to obtain at least one aversion feature when the interaction behavior data indicates the occurrence of negative feedback behavior; the aversion feature includes at least one of the following: audio features, visual features, and semantic features; and a determination module, configured to determine the user's aversion score for the aversion feature based on the base aversion intensity, combo weight, time decay factor corresponding to the aversion feature, and the time interval between the occurrence time of the negative feedback behavior and the current time; wherein the base aversion intensity... The degree is determined based on the behavior type of the negative feedback behavior; the combo weight is determined based on the number of consecutive occurrences of the negative feedback behavior within a preset time window; the time decay factor is configured based on the preset attribute category of the aversion feature; the preset attribute category is used to determine whether the aversion feature is a first feature that decays rapidly and represents the user's aesthetic fatigue, or a second feature that decays slowly and represents the user's persistent rejection; the update module is used to obtain candidate videos in a pre-generated video recommendation list, and update the candidate videos in the video recommendation list according to the aversion score corresponding to the aversion feature included in the candidate videos, and a preset score threshold.

[0007] Thirdly, embodiments of this disclosure provide an electronic device, including: a processor; and a memory configured to store computer-executable instructions, which, when executed, cause the processor to perform the steps of the method described in the first aspect above.

[0008] Fourthly, embodiments of this disclosure provide a computer-readable storage medium for storing computer-executable instructions that, when executed by a processor, implement the steps of the method described in the first aspect above.

[0009] Fifthly, embodiments of this disclosure provide a computer program product, the computer program product including a computer program, which, when executed by a processor, implements the steps of the method described in the first aspect above.

[0010] The at least one technical solution provided by the embodiments of the present invention can achieve the following technical effects: In the embodiments of the present invention, by capturing negative feedback in user interaction behavior, atomic feature extraction can be performed on the video content that triggers the feedback, and the user's aversion score to various features can be dynamically calculated based on behavior type, frequency of occurrence, feature attributes, and time interval. The embodiments of the present invention can transform the traditional one-time, static, and overfitting recommendation mechanism into a precise filtering process that can cyclically perceive, dynamically decay, and distinguish aversion types, thereby effectively solving the problems of homogenized recommended content, decreased user experience, and user churn caused by the information cocoon effect, inefficient negative feedback mechanism, and lack of aesthetic fatigue management. At the same time, by prioritizing the avoidance of user-averse content, the embodiments of the present invention fundamentally optimize the diversity of the recommendation ecosystem and significantly improve the level of long-term user retention and intelligent video recommendation. Attached Figure Description

[0011] The accompanying drawings, which are included to provide a further understanding of this application and constitute a part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 is a schematic flowchart of a video recommendation method according to an embodiment of the present invention; Figure 2 is a schematic flowchart of a video recommendation method according to another embodiment of the present invention; Figure 3 is a schematic diagram of the module composition of a video recommendation device 300 according to an embodiment of the present invention; Figure 4 is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0013] The technical solutions provided by the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0014] Please refer to Figure 1. Figure 1 is a flowchart of a video recommendation method provided in an embodiment of the present invention. As shown in Figure 1, the method includes the following steps: Step 102: Obtain user interaction behavior data for playing videos.

[0015] Step 104: When the interaction behavior data indicates that a negative feedback behavior has occurred, perform feature extraction on the playing video to obtain at least one aversion feature; the aversion feature includes at least one of the following: audio features, visual features, and semantic features.

[0016] Step 106: Determine the user's aversion score for the aversion feature based on the base aversion intensity, combo weight, time decay factor, and the time interval between the occurrence time of the negative feedback behavior and the current time. The base aversion intensity is determined based on the behavior type of the negative feedback behavior; the combo weight is determined based on the number of consecutive occurrences of the negative feedback behavior within a preset time window; the time decay factor is configured based on the preset attribute category of the aversion feature; the preset attribute category is used to determine whether the aversion feature is a rapidly decaying first feature representing user aesthetic fatigue, or a slowly decaying second feature representing user persistent rejection.

[0017] Step 108: Obtain candidate videos from the pre-generated video recommendation list, and update the candidate videos in the video recommendation list according to the aversion scores corresponding to the aversion features included in the candidate videos and the preset score threshold.

[0018] In one embodiment of the present invention, interactive behavior data generated by users during the use of short video applications can be acquired. This interactive behavior data can include explicit negative feedback actions actively triggered by the user, such as clear "disinterested" or "report," as well as a series of subtle, implicit negative feedback signals. These implicit negative feedback signals represent the user's subconscious rejection response and are crucial for constructing an accurate profile of aversion.

[0019] For example, the system can precisely time the interval between the start of a video playback and the user's swipe-to-skip action. When this interval is less than a preset first duration threshold, such as 800 milliseconds, a "visual escape" behavior can be identified, strongly suggesting that the user has developed an immediate aversion to the video's opening content or style. Simultaneously, volume adjustment events during playback can be monitored. If the user lowers the volume by more than a preset threshold, such as reducing it by more than 70% of the original volume, or directly sets the volume to zero (mute), this can be recorded as an "auditory rejection" behavior. Furthermore, a short-term behavioral context window, such as 30 seconds, can be maintained. If, within this window, the user consecutively skips multiple videos identified as having the same or highly similar characteristics—such as belonging to the same creator, using the same popular background music, or having a similar visual style—a "negative feedback combo" behavior can be identified, indicating that the user has developed a concentrated and strong aversion to a certain type of content. This multi-dimensional, fine-grained interactive behavior data is sent to the backend processing system in real time, constituting the raw input for the user's negative preferences.

[0020] In one embodiment of the present invention, after acquiring the interaction behavior, if the interaction behavior data indicates that any of the above-mentioned negative feedback behaviors have occurred, an atomic feature extraction process can be immediately initiated for the currently playing video that triggered the behavior. The goal of this process is not to understand the complete semantics of the video, but to deconstruct it into a series of measurable and generalizable "aversion features", thereby realizing the transformation from "blocking a single video" to "avoiding a type of content".

[0021] The extracted atomized features can include visual, audio, and semantic features. When extracting visual features, keyframes can be extracted from the video stream based on a fixed time interval, such as one frame per second, or a scene change detection algorithm based on inter-frame differences. For each keyframe, it can be converted to a color space containing independent saturation components, such as HSV (Hue, Saturation, Value) or HSL (Hue, Saturation, Lightness), and its saturation channel histogram distribution can be calculated. Subsequently, the statistical data from this histogram can be normalized to generate a fixed-dimensional numerical vector, serving as a visual feature characterizing the color saturation of the video image. This visual feature effectively quantifies the "color intensity" or "tacky" feel of the video image and is a key indicator for identifying specific video visual styles.

[0022] Simultaneously, pre-trained text detection models, such as the deep learning-based CTPN (Detecting Text in Natural Image with Connectionist Text Proposal Network) model, can be used to scan keyframes, identify the location and bounding boxes of all text regions in the image, and then calculate the ratio of the total area of ​​all text regions to the total area of ​​the image. This ratio can serve as a visual feature representing text coverage density, demonstrating high discriminative power for identifying "marketing accounts" or "news flash" videos filled with excessive text titles and exaggerated subtitles. Furthermore, pre-trained face recognition models can be invoked to extract facial feature vectors appearing in the image and compare them with a known creator database to identify specific internet celebrities or public figures that users may dislike.

[0023] When extracting audio features, Mel-frequency cepstral coefficients (MFCCs) can be extracted from the video's audio stream. These MFCCs are widely used in audio recognition and can simulate human hearing. The extracted MFCC feature sequence can then be input into a pre-trained audio event classification model. This model is a deep neural network classifier whose output layer corresponds to several predefined audio category labels. In this embodiment, these categories are specifically designed to capture offensive audio patterns, such as "harsh, shrill laughter," "specific, repetitive popular background music clips," and "pops with sudden, sharp increases in volume." The classification labels output by the model can serve as the audio features of the video, thus characterizing which sound elements in the video the user found offensive.

[0024] When extracting text features, multi-source text information from videos can be aggregated, including titles filled in by the video uploader, text dynamically extracted from the video footage using optical character recognition technology, and high-frequency bullet screen text generated during playback. For these aggregated texts, two optional, complementary analysis paths can be provided. The first analysis path is based on a pre-trained sentiment recognition model, such as a fine-tuned BERT (Bidirectional Encoder Representations from Transformers) model based on the Transformer architecture. After encoding the input text, this model can directly output a sentiment polarity classification label, such as "extremely negative," "moderately negative," or "neutral," or a continuous negative sentiment intensity score. The output sentiment polarity classification label or negative sentiment intensity score can be used as semantic features. The second analysis path is based on matching a carefully constructed and regularly updated pre-defined negative sentiment lexicon. This lexicon can contain words that clearly express aversion, such as "disgusting," "vulgar," "script," and "deceptive," as well as their common variations. After segmenting and matching the aggregated text, successfully matched keywords can be used as semantic features. These two analytical approaches can be used independently or their results can be combined to collectively characterize the negative emotions evoked by the video content at the semantic level.

[0025] In one embodiment of the present invention, after obtaining one or more of the aforementioned aversion features, the user's aversion score for the aversion feature can be calculated based on the base aversion intensity, combo weight, time decay factor, and the time interval between the occurrence time of the negative feedback behavior and the current time. This calculation process comprehensively considers three dimensions: behavior intensity, emotional accumulation, and time decay.

[0026] First, a base aversion intensity can be determined based on the specific type of negative feedback behavior that triggered this feature extraction. This is a configurable scalar value. For example, the base intensity for explicit negative feedback behaviors such as actively clicking the "not interested" button can be set to 1.0; the base intensity for strong implicit feedback such as "swiping instantly" can be set to 0.8; the base intensity for "mute" or "significantly lower the volume" can be set to 0.6; and the base intensity for general quick skipping can be set to 0.4. This processing operation can effectively distinguish the differences in the degree of aversion implied by different behaviors.

[0027] Secondly, a combo weight can be introduced to quantify the escalation of emotions caused by a user's concentrated aversion to similar features within a short period. Specifically, a sliding time window, such as 30 seconds, can be maintained, and the number of times a user continuously generates negative feedback behavior towards videos with the same aversion features can be counted within this window. The combo weight is not linearly related to the number of times it is generated, but rather uses a non-linear growth function, for example, set to... ,in, This refers to the number of consecutive negative feedback behaviors. This means that if a user says "no" to the same type of content three times in a row, the cumulative negative emotional weight is much higher than the simple sum of three independent events, more realistically reflecting the user's psychological change from "somewhat annoyed" to "unbearable".

[0028] Then, a time decay factor can be configured for each aversion feature based on its preset attribute category. In specific configuration, features can be divided into two main categories: The first category represents primary features that indicate "user aesthetic fatigue," such as a temporarily over-popular background music track or a prevalent visual filter style. These features often cause users to become annoyed due to short-term overexposure, but may be reintroduced over time. Therefore, they are configured with a larger time decay factor, such as a time decay factor value of 0.1, so that their corresponding aversion scores will decrease quickly. The second category represents secondary features that indicate "persistent user rejection," such as violent or disturbing visual elements, or content themes that users clearly and persistently dislike. These features are configured with a very small decay factor, such as a time decay factor value of 0.0001, so that the aversion score remains high for a long period, achieving near-permanent blocking.

[0029] After obtaining the basic aversion intensity, combo weight, and time decay factor mentioned above, the user can be calculated based on the current moment. For features At any moment Dynamic aversion score Its calculation logic is based on a combo enhancement model that takes into account time decay, which can be expressed as follows: the aversion score is equal to the product of the base aversion intensity, the combo weight, and the decay composite value, where the decay composite value is calculated by an exponential function with the natural constant as the base and the negative of the product of the time decay factor and the time interval as the exponent.

[0030] Specifically, the score is determined by the basic aversion intensity. Combo weight Time decay factor And the time interval from the moment the negative feedback behavior occurs to the current calculation moment. (Usually measured in days) jointly determined. The cascading effect is achieved through... To enhance the foundation strength, This is an adjustable addition factor, with a default value of 0.5. The time decay effect is determined by the natural constant. negative exponential function This is reflected in the formula. The formula ensures that the aversion score decays naturally over time, and the rate of decay is... This decision effectively distinguishes between "temporarily getting tired of it" and "always hating it".

[0031] In one embodiment of the present invention, when it is necessary to generate or update a video recommendation list for a user, a pre-generated initial candidate video list based on the user's positive interests can be obtained first. For each candidate video in the list, all aversion features it contains can be identified based on the above operations, and the current user's real-time aversion score for these features can be queried. Subsequently, a preset update strategy can be applied to update the recommendation list.

[0032] In one example, the preset update strategy could include a hard filtering strategy: a high first score threshold, such as 0.8, could be set. If any of the aversion features contained in a candidate video has a user aversion score exceeding this first score threshold, it can be assumed that the user strongly dislikes the video, and it will be directly removed from the recommendation list to ensure that the user never sees such content.

[0033] In another example, the pre-defined update strategy may include a soft de-weighting strategy: for candidate videos that have passed the hard filtering check (i.e., all features' aversion scores do not exceed a first score threshold), a comprehensive aversion score can be calculated to quantify their overall "aversion risk." In one embodiment, the comprehensive aversion score can be simply defined as the sum of the aversion scores corresponding to all features of the video. Then, based on this comprehensive aversion score, the ranking weight of the candidate video in the list can be reduced. For example, a penalty term proportional to the comprehensive aversion score can be subtracted from the original click-through rate prediction score. In this way, even if two videos have similar positive interest scores, the video containing more features that users slightly dislike and having a higher comprehensive aversion score will have its final ranking lower.

[0034] In one embodiment of the present invention, a dynamic aversion feature verification and attenuation adjustment mechanism may be included to address the "over-blocking" problem caused by misjudgment or changes in interests. Specifically, the attenuation of all active aversion scores can be continuously monitored. When it is found that the score of a certain aversion feature has fallen below a second score threshold lower than a first score threshold due to time decay, it can be determined that the feature may no longer be a point of aversion for the user, or the initial judgment may have been biased. At this time, a trial can be proactively performed: one or more videos containing the feature, but with other features that are more positive or neutral, can be inserted into the video recommendation list with a lower exposure probability.

[0035] Then, user interaction with these trial videos can be closely monitored. If a user watches the trial video and does not exhibit any of the aforementioned negative feedback behaviors, such as swiping immediately, muting, or clicking "not interested," the validation is successful. At this point, the time decay factor for this aversion trait can be increased. For example, doubling the value could increase the decay rate, thus accelerating the "unbanning" of the feature from the user's aversion list; conversely, if the user again produces significant negative feedback to the test video, the validation fails. In this case, the aversion score corresponding to the feature can be immediately reset to a higher initial value, such as the historical high, and its time decay factor can be significantly reduced. For example, the aversion rate can be reduced to one-tenth of its original value to implement stricter and more persistent blocking of that feature over a longer period. This process enhances the self-correction and dynamic adaptation capabilities, allowing the aversion model to evolve in line with changes in users' actual interests, effectively balancing the conflict between blocking inappropriate content and maintaining recommendation diversity.

[0036] In one example, as shown in Figure 2, user interaction with the video can be monitored first. Specific behaviors such as "slide-by" (skipping within a very short time), "mute," or continuous "skipping" while watching the video can be identified as potential negative feedback signals. This raw behavioral data can be sent to the negative signal capture module. This module performs preliminary filtering and formatting, transforming chaotic user actions into standardized negative events that can be processed by the subsequent engine.

[0037] These events can then be passed to the core feature extraction engine. This engine can concurrently perform in-depth deconstruction of the original video content that triggers negative feedback from multiple dimensions. Specifically, in the visual dimension, the engine can extract the visual features of the image and map them into a structured visual vector space to quantify attributes such as color saturation and composition style; in the auditory dimension, the engine analyzes the audio stream to extract audio features and projects them into the audio vector space to identify specific background music, sound effects, or acoustic patterns; in the textual semantic dimension, the engine extracts semantic features from titles, subtitles, and bullet comments to form a tag word cloud space to capture the keywords and sentiment of the content. These three feature spaces together constitute a description of the multimodal aversion characteristics of the video content.

[0038] After obtaining the aforementioned visual, audio, and semantic features, they can be fed into a dynamic aversion score calculation model. This model, based on a pre-defined algorithm, comprehensively considers factors such as behavior type, combo effect, and time decay factor to calculate a dynamically updated aversion score for each aversion feature. Once the aversion scores are obtained, they can be persistently stored in a user negative profile database for accumulating and updating the user's long-term aversion preference profile. Furthermore, these scores can be used immediately to intervene in the current recommendation process.

[0039] Then, the recommendation candidate pool, which is the set of videos initially filtered based on users' positive interests, can be compared with the real-time aversion scores retrieved from the negative profile database. This comparison process can be completed in the reverse re-ranking filtering module. The core logic of this module is "removing potential problems before recommending": that is, candidate videos can be hard-filtered based on aversion scores, directly removing content with high aversion features, or soft-weighted, i.e., reducing their ranking weight. After this series of processes prioritizing the avoidance of negative preferences, an optimized final list of recommended videos to avoid can be output, aiming to minimize the possibility of users encountering averse content, thereby improving user satisfaction and long-term user experience of the recommendation system.

[0040] In this embodiment of the invention, negative feedback from user interaction behavior can be captured, and atomic features can be extracted from the video content that triggers the feedback. Based on behavior type, frequency of occurrence, feature attributes, and time interval, the user's aversion score for various features can be dynamically calculated. This embodiment transforms the traditional one-off, static, and overfitting recommendation mechanism into a cyclically perceptive, dynamically decaying, and aggression-distinguishing precise filtering process. This effectively solves the problems of homogenized recommended content, decreased user experience, and user churn caused by the information cocoon effect, inefficient negative feedback mechanisms, and lack of aesthetic fatigue management. Furthermore, by prioritizing the avoidance of user-averse content, this embodiment fundamentally optimizes the diversity of the recommendation ecosystem, significantly improving long-term user retention and the intelligence level of video recommendations.

[0041] The video recommendation device 300 shown in Figure 3 can implement the method of the embodiment in Figure 1 and achieve the same technical effect. For details, please refer to the above description of the video recommendation in the embodiment shown in Figure 1, which will not be repeated here. The video recommendation device 300 includes: a first acquisition module 301, used to acquire user interaction behavior data related to a played video; a second acquisition module 302, used to extract features from the played video to obtain at least one aversion feature when the interaction behavior data indicates a negative feedback behavior; the aversion feature includes at least one of the following: audio features, visual features, and semantic features; and a determination module 303, used to determine the user's aversion score for the aversion feature based on the basic aversion intensity, combo weight, time decay factor, and the time interval between the occurrence time of the negative feedback behavior and the current time; wherein, the basic aversion intensity is a base... The negative feedback behavior is determined by the behavior type; the combo weight is determined based on the number of consecutive occurrences of the negative feedback behavior within a preset time window; the time decay factor is configured based on the preset attribute category of the aversion feature; the preset attribute category is used to determine whether the aversion feature is a first feature that decays rapidly and represents the user's aesthetic fatigue, or a second feature that decays slowly and represents the user's persistent rejection; the update module 304 is used to obtain candidate videos in a pre-generated video recommendation list, and update the candidate videos in the video recommendation list according to the aversion score corresponding to the aversion feature included in the candidate videos, and a preset score threshold.

[0042] Optionally, the negative feedback behavior includes explicit negative feedback behavior and implicit negative feedback behavior; the explicit negative feedback behavior includes the user's operation representing disinterest in the played video; the implicit negative feedback behavior includes at least one of the following: the viewing time of the played video is less than a preset duration threshold; during the playback of the played video, the video volume is reduced by an amount exceeding a preset amplitude threshold, or the video volume is reduced to zero; within a preset time period, multiple videos of the same type as the played video are skipped.

[0043] Optionally, the determining module 303 is configured to: determine the basic service intensity based on the behavior type of the negative feedback behavior; wherein, when the behavior type of the negative feedback behavior is an explicit negative feedback behavior, a first preset basic value is determined as the basic aversion intensity corresponding to the aversion feature, and when the behavior type of the negative feedback behavior is an implicit negative feedback behavior, a second preset basic value is determined as the basic aversion intensity corresponding to the aversion feature; the second preset basic value is less than the first preset basic value; determine the combo weight based on the number of consecutive occurrences of the negative feedback behavior within a preset time window; wherein, the higher the number of consecutive occurrences, the larger the value of the combo weight, and the greater the increase in the connection weight; determine the time decay factor based on the preset attribute category of the aversion feature; wherein, the time decay factor corresponding to the first feature is greater than the time decay factor corresponding to the second feature; calculate the aversion score based on the basic aversion intensity, the combo weight, the time decay factor, and the time interval; wherein, the larger the time interval, the smaller the aversion score.

[0044] Optionally, the determining module 303 is further configured to: determine the aversion score by multiplying the basic aversion intensity, the combo weight, and the decay comprehensive value; wherein the decay comprehensive value is calculated by an exponential function with the natural constant as the base and the negative of the product of the time decay factor and the time interval as the exponent.

[0045] Optionally, the feature extraction of the played video to obtain at least one aversion feature includes at least one of the following: extracting keyframes from the played video according to fixed time intervals or scene changes, and using the histogram of the saturation components of the keyframes in a specific color space as the visual feature characterizing the visual style of the video; or, using a text detection model to identify text regions in the image, and using the ratio of the text region area to the total image area as the visual feature characterizing the text coverage of the video; extracting Mel-frequency cepstral coefficients from the audio stream of the played video, and inputting the Mel-frequency cepstral coefficients into the pre-training... In the trained audio event classification model, the audio classification labels output by the audio event classification model, which are used to identify harsh sounds, specific background music, or sudden changes in volume, are determined as the audio features; video text extraction is performed on the playing video, and the video text is recognized based on a pre-trained sentiment recognition model, and the recognized negative sentiment labels are determined as semantic features, or the video text is matched based on a preset negative sentiment word library, and the matched negative sentiment words are determined as semantic features; the video text includes at least one of the following: the title of the playing video, on-screen text, and high-frequency bullet screen text.

[0046] Optionally, the update module 304 is configured to: remove the candidate video from the video recommendation list if the aversion score corresponding to any of the aversion features contained in the candidate video is greater than a preset first score threshold; and determine a comprehensive aversion score based on the sum of the aversion scores corresponding to each of the aversion features contained in the candidate video if none of the aversion scores corresponding to the aversion features contained in the candidate video are greater than the first score threshold, and reduce the ranking weight of the candidate video in the video recommendation list based on the comprehensive aversion score.

[0047] Optionally, the apparatus further includes (not shown in claim 3): a recommendation module 305, configured to, after updating the candidate videos in the video recommendation list according to the aversion score corresponding to the aversion feature included in the candidate videos and a preset score threshold, recommend trial videos containing the aversion feature to the user when it is detected that the user's aversion score for the aversion feature is lower than a second score threshold due to time decay; and an adjustment module 306, configured to, if the user does not produce negative feedback behavior towards the trial video, increase the time decay factor; and if the user produces negative feedback behavior towards the trial video, reset the aversion score to the historical highest value and decrease the time decay factor.

[0048] In this embodiment of the invention, negative feedback from user interaction behavior can be captured, and atomic features can be extracted from the video content that triggers the feedback. Based on behavior type, frequency of occurrence, feature attributes, and time interval, the user's aversion score for various features can be dynamically calculated. This embodiment transforms the traditional one-off, static, and overfitting recommendation mechanism into a cyclically perceptive, dynamically decaying, and aggression-distinguishing precise filtering process. This effectively solves the problems of homogenized recommended content, decreased user experience, and user churn caused by the information cocoon effect, inefficient negative feedback mechanisms, and lack of aesthetic fatigue management. Furthermore, by prioritizing the avoidance of user-averse content, this embodiment fundamentally optimizes the diversity of the recommendation ecosystem, significantly improving long-term user retention and the intelligence level of video recommendations.

[0049] Figure 4 is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Referring to Figure 4, at the hardware level, the electronic device includes a processor, and optionally also includes an internal bus, a network interface, and a memory. The memory may include RAM, such as high-speed random-access memory (RAM), or non-volatile memory, such as at least one disk storage device. Of course, the electronic device may also include other hardware required for other services.

[0050] The processor, network interface, and memory can be interconnected via an internal bus, which can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. This bus can be categorized as an address bus, data bus, control bus, etc. For ease of illustration, only a single bidirectional arrow is used in Figure 4, but this does not imply that there is only one bus or one type of bus.

[0051] Memory is used to store programs. Specifically, programs may include program code, which includes computer operation instructions. Memory may include main memory and non-volatile memory, and provides instructions and data to the processor.

[0052] The processor reads the corresponding computer program from non-volatile memory into memory and then runs it, forming a discontinuous transfer configuration device at the logical level. The processor executes a program stored in memory and is specifically configured to perform the following operations: acquire user interaction behavior data related to a played video; when the interaction behavior data indicates the occurrence of negative feedback behavior, extract features from the played video to obtain at least one aversion feature; the aversion feature includes at least one of the following: audio features, visual features, and semantic features; determine the user's aversion score for the aversion feature based on the base aversion intensity, combo weight, time decay factor, and the time interval between the occurrence time of the negative feedback behavior and the current time; wherein the base aversion intensity is determined based on the behavior type of the negative feedback behavior; the combo weight is determined based on the number of consecutive occurrences of the negative feedback behavior within a preset time window; the time decay factor is configured based on a preset attribute category of the aversion feature; the preset attribute category is used to determine whether the aversion feature is a rapidly decaying first feature representing user aesthetic fatigue or a slowly decaying second feature representing user persistent rejection; acquire candidate videos from a pre-generated video recommendation list, and update the candidate videos in the video recommendation list based on the aversion scores corresponding to the aversion features included in the candidate videos and a preset score threshold.

[0053] The video recommendation method disclosed in the embodiment shown in Figure 1 of this application can be applied to a processor or implemented by a processor. The processor may be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method can be completed by integrated logic circuits in the processor's hardware or by instructions in software form. The processor can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in one or more embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in one or more embodiments of this application can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software module can reside in a mature storage medium in the field, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method.

[0054] The electronic device can also perform the video recommendation method described in Figure 1, which will not be repeated here.

[0055] This application also proposes a computer-readable storage medium that stores one or more programs, the programs including instructions that, when executed by a portable electronic device including multiple applications, enable the portable electronic device to perform the method of the embodiment shown in FIG1, which will not be described in detail here.

[0056] This application also proposes a computer program product that is stored in a storage medium and executed by at least one processor to implement the method of the embodiment shown in FIG1, which will not be described in detail here.

[0057] Of course, in addition to software implementation, the electronic device of this application does not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. In other words, the execution subject of the following processing flow is not limited to each logic unit, but can also be hardware or logic devices.

[0058] In summary, the above description is merely a preferred embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments of this application should be included within the scope of protection of one or more embodiments of this application.

[0059] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, a computer can be, for example, a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email device, game console, tablet computer, wearable device, or any combination of these devices.

[0060] Computer-readable media, including both permanent and non-permanent, removable and non-removable media, can store information using any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory 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 in the embodiments of this application, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0061] 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.

[0062] 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 system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

Claims

1. A video recommendation method, characterized in that, include: The system acquires user interaction data related to a played video; when the interaction data indicates a negative feedback behavior, it extracts features from the played video to obtain at least one aversion feature; the aversion feature includes at least one of the following: audio features, visual features, and semantic features; based on the base aversion intensity, combo weight, time decay factor, and the time interval between the occurrence time of the negative feedback behavior and the current time corresponding to the aversion feature, it determines the user's aversion score for the aversion feature; wherein, the base aversion intensity is determined based on the behavior type of the negative feedback behavior; the combo weight is determined based on the number of consecutive occurrences of the negative feedback behavior within a preset time window; the time decay factor is configured based on a preset attribute category of the aversion feature; the preset attribute category is used to determine whether the aversion feature is a rapidly decaying first feature representing the user's aesthetic fatigue, or a slowly decaying second feature representing the user's persistent aversion; it acquires candidate videos from a pre-generated video recommendation list, and updates the candidate videos in the video recommendation list based on the aversion scores corresponding to the aversion features included in the candidate videos and a preset score threshold.

2. The method according to claim 1, characterized in that, The negative feedback behavior includes explicit negative feedback behavior and implicit negative feedback behavior; the explicit negative feedback behavior includes the user's operation on the played video that indicates disinterest; the implicit negative feedback behavior includes at least one of the following: the viewing time of the played video is less than a preset duration threshold; during the playback of the played video, the video volume is reduced by an amount exceeding a preset amplitude threshold, or the video volume is reduced to zero; within a preset time period, multiple videos of the same type as the played video are skipped.

3. The method according to claim 2, characterized in that, The step of determining the user's aversion score for the aversion feature based on the base aversion intensity, combo weight, time decay factor, and the time interval between the occurrence time of the negative feedback behavior and the current time includes: determining the base service intensity based on the behavior type of the negative feedback behavior; wherein, when the behavior type of the negative feedback behavior is an explicit negative feedback behavior, a first preset base value is determined as the base aversion intensity corresponding to the aversion feature, and when the behavior type of the negative feedback behavior is an implicit negative feedback behavior, a second preset base value is determined as the base aversion intensity corresponding to the aversion feature; The second preset base value is less than the first preset base value; the combo weight is determined based on the number of consecutive occurrences of the negative feedback behavior within a preset time window; wherein, the higher the number of consecutive occurrences, the larger the value of the combo weight, and the greater the increase in the connection weight; the time decay factor is determined based on the preset attribute category of the aversion feature; wherein, the time decay factor corresponding to the first feature is greater than the time decay factor corresponding to the second feature; the aversion score is calculated based on the base aversion intensity, the combo weight, the time decay factor, and the time interval; wherein, the larger the time interval, the smaller the aversion score.

4. The method according to claim 3, characterized in that, The step of calculating the aversion score based on the base aversion intensity, the combo weight, the time decay factor, and the time interval includes: determining the aversion score by multiplying the base aversion intensity, the combo weight, and the decay composite value; wherein the decay composite value is calculated by an exponential function with the natural constant as the base and the negative of the product of the time decay factor and the time interval as the exponent.

5. The method according to claim 1, characterized in that, The feature extraction of the played video to obtain at least one aversion feature includes at least one of the following: extracting keyframes from the played video according to a fixed time interval or scene change, and using the histogram of the saturation components of the keyframes in a specific color space as the visual feature characterizing the visual style of the video. Alternatively, a text detection model can be used to identify text regions in the image, and the ratio of the text region area to the total image area can be used as the visual feature representing the text coverage of the video; Mel frequency cepstral coefficients can be extracted from the audio stream of the playing video, and the Mel frequency cepstral coefficients can be input into a pre-trained audio event classification model, and the audio classification labels output by the audio event classification model, which are used to identify harsh sounds, specific background music, or sudden changes in volume, can be determined as the audio features; The video text is extracted from the playing video. The video text is recognized based on a pre-trained sentiment recognition model, and the identified negative sentiment tags are determined as semantic features. Alternatively, the video text is matched based on a preset negative sentiment word library, and the matched negative sentiment words are determined as semantic features. The video text includes at least one of the following: the title of the playing video, the on-screen text, and the high-frequency bullet screen text.

6. The method according to claim 1, characterized in that, The step of updating candidate videos in the video recommendation list based on the aversion scores corresponding to the aversion features included in the candidate videos and a preset score threshold includes: removing the candidate video from the video recommendation list if the aversion score corresponding to any of the aversion features included in the candidate video is greater than a preset first score threshold; and determining a comprehensive aversion score based on the sum of the aversion scores corresponding to the aversion features included in the candidate video if none of the aversion scores corresponding to the aversion features included in the candidate video are greater than the first score threshold, and reducing the ranking weight of the candidate video in the video recommendation list based on the comprehensive aversion score.

7. The method according to claim 1, characterized in that, After updating the candidate videos in the video recommendation list based on the aversion score corresponding to the aversion feature included in the candidate videos and a preset score threshold, the method further includes: when it is detected that the user's aversion score for the aversion feature is lower than a second score threshold due to time decay, recommending a trial video containing the aversion feature to the user; if the user does not produce negative feedback behavior to the trial video, increasing the time decay factor; if the user produces negative feedback behavior to the trial video, resetting the aversion score to the historical highest value and decreasing the time decay factor.

8. A video recommendation device, characterized in that, include: The first acquisition module is used to acquire user interaction data related to playing videos; The second acquisition module is used to extract features from the playing video when the interaction behavior data indicates the occurrence of negative feedback behavior, to obtain at least one aversion feature; the aversion feature includes at least one of the following: audio features, visual features, and semantic features; the determination module is used to determine the user's aversion score for the aversion feature based on the basic aversion intensity, combo weight, time decay factor corresponding to the aversion feature, and the time interval between the occurrence time of the negative feedback behavior and the current time; wherein, the basic aversion intensity is determined based on the behavior type of the negative feedback behavior; the combo weight is determined based on the number of consecutive occurrences of the negative feedback behavior within a preset time window; the time decay factor is configured based on a preset attribute category of the aversion feature; the preset attribute category is used to determine whether the aversion feature is a rapidly decaying first feature representing the user's aesthetic fatigue, or a slowly decaying second feature representing the user's persistent rejection; An update module is used to obtain candidate videos from a pre-generated video recommendation list and update the candidate videos in the video recommendation list according to the aversion scores corresponding to the aversion features included in the candidate videos and a preset score threshold.

9. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store computer-executable instructions that, when executed by a processor, implement the steps of the method described in any one of claims 1 to 7.

Images