Machine learning-based short video platform precision diversion method and system

Abstract

The present application relates to the technical field of internet information processing and machine learning, in particular to a short video platform accurate flow method and system based on machine learning; the method comprises the following steps: collecting user short video browsing behavior data on the terminal side and extracting video content features of the current video, inputting the features into a feature extraction model to generate current interest features, and determining interest change information based on the current interest features and reference interest features; determining behavior effectiveness information and link congestion information according to user browsing integrity, interaction state, backslide state and link state, and selecting an upload strategy; after receiving an event package on the cloud side, reconstructing user comprehensive interest features in combination with user historical interest features, hot spot change conditions and interest change information, determining a hierarchical processing mode, generating a distribution content decision package and delivering it to the terminal side for display.

Description

Technical Field

[0001] This invention relates to the fields of internet information processing and machine learning technology, specifically to a method and system for precise traffic acquisition on short video platforms based on machine learning. Background Technology

[0002] Existing short video platform traffic acquisition systems typically collect and upload behavioral data from the client side, relying on the cloud side to complete interest identification and content distribution.

[0003] However, existing solutions rely excessively on cloud-based intent modeling and sorting of high-dimensional features, leading to uploaded data volumes exceeding preset thresholds and processing links exceeding defined limits. This is particularly problematic when network bandwidth is limited or users are continuously and rapidly scrolling, easily causing upload delays due to changes in interest, delayed cloud-based result returns, and missed display conditions.

[0004] Furthermore, given the high level of interaction in short videos, existing solutions still employ a fixed full-feature processing and unified distribution mechanism. This increases the additional computing power consumption on the device side, resulting in a mismatch between the distributed traffic-driving content and the user's instantaneous browsing state, ultimately causing the traffic to fail to reach the preset targets. Summary of the Invention

[0005] The purpose of this invention is to provide a machine learning-based method and system for precise traffic acquisition on short video platforms, addressing the following technical problems: existing short video platform traffic acquisition systems have technical deficiencies in capturing transient interests and matching traffic acquisition content in real time when users are continuously and rapidly scrolling. This invention proposes a machine learning-based method and system for precise content distribution that can accurately perceive transient intentions on the device side and dynamically optimize the collaborative delivery strategy between the device and the cloud. This invention is implemented using the following technical solution: The device collects user short video browsing behavior data and extracts video content features of the current video to obtain device-collected data; the device inputs the device-collected data into a feature extraction model to generate current interest features, and determines interest change information based on the current interest features and reference interest features; The client determines the validity of the behavior and the congestion of the link based on the user's browsing integrity, interaction status, scrolling status and link status, and selects to upload the full change packet, upload the minimum event packet or write it to the client-side cache queue accordingly. After receiving an event packet containing user identifier, session identifier, timestamp, and interest change information, the cloud side obtains the user's historical interest characteristics and determines the hotspot changes based on the exposure, clicks, or conversion changes of candidate content in the cloud side's pre-loaded candidate set within the statistics window. The cloud side then reconstructs the user's comprehensive interest characteristics by combining the user's historical interest characteristics, hotspot changes, and interest change information. The cloud side determines the hierarchical processing method based on the user's comprehensive interest characteristics, link status, and interest change information, generates a content distribution decision package, and sends it to the terminal side for display.

[0006] Preferably, the user's short video browsing behavior data includes swiping speed, dwell time, click status, swipe back status, interaction quality status, audio interaction status, and entry into the product page status; the interaction quality status includes at least one of like status, comment status, favorite status, forward status, and follow status; the audio interaction status includes audio on status, mute status, volume adjustment status, or audio repeat playback status; the video content features include keyframe visual features, speech rhythm features, optical flow change features, and text tag features.

[0007] Preferably, the edge device inputs the edge-collected data into the feature extraction model deployed on the edge device to generate the current interest feature vector; the edge device saves the reference interest feature vector corresponding to the last valid feature update, and generates an interest change vector and interest change magnitude based on the current interest feature vector and the reference interest feature vector; when the currently acquired behavioral data meets the preset validity conditions, and the edge device receives a cloud-side processing confirmation mark, an edge device display feedback signal, or a timeout confirmation signal, the edge device updates the reference interest feature vector with the current interest feature vector.

[0008] As a preferred option, the corresponding upload strategy is selected, including: determining the user's browsing completeness based on the dwell time and video duration; and generating a behavior effectiveness score based on the user's browsing completeness, dwell time, interaction quality status, and scrollback status. The endpoint generates a link congestion metric based on the upload queue length, round-trip latency, and packet loss rate. The endpoint selects a corresponding upload strategy based on a comparison of the interest change magnitude with the upload threshold issued by the cloud, the behavior effectiveness score, and the link congestion metric with the congestion threshold issued by the cloud. The upload threshold and the congestion threshold are generated and issued by the cloud, and the endpoint synchronously updates the upload threshold and the congestion threshold based on the threshold adjustment signal issued by the cloud.

[0009] Preferably, when multiple interest change vectors in the edge cache queue satisfy the direction consistency rule within a continuous observation window, the edge merges the multiple interest change vectors and uploads them; the direction consistency rule includes that the cosine similarity of the angle between the vectors is greater than the direction consistency threshold, the direction consistency threshold is a preset scalar value, or that the positive and negative polarities of multiple interest change vectors are the same on the same feature dimension; The cloud side returns a processing confirmation flag, a mode flag, and a reference correction suggestion for the uploaded event packet. The mode flag is used to indicate whether the end side should subsequently use full change packet upload, minimum event packet upload, or local caching. The reference correction suggestion is used to indicate whether the end side should perform incremental correction on the reference interest feature vector. The device generates local matching results from the local preloaded candidate set based on the current interest feature vector. If no processing confirmation flag is received within the timeout period, the device continues to display the local matching results and marks the event as a sample that the cloud side did not participate in. The event is then removed from the training sample set or parameter adjustment sample set.

[0010] As a preferred option, the user's comprehensive interest characteristics are reconstructed, including: the cloud side obtains the received event sequence consisting of event identifier, session identifier and timestamp from the cloud side cache based on the user identifier and session identifier; the cloud side determines the current session time window based on the start and end time of the current session, and determines the cross-session tolerance window based on the preset time interval between adjacent sessions; Event packets that fall within the current session time window and whose event identifiers are not duplicated are rearranged according to their timestamps; event packets with duplicate event identifiers, or event packets with inconsistent session identifiers and timestamps exceeding the cross-session tolerance window are discarded; the cloud side converts interest change information into a global interest change vector with the same dimension as the feature vector corresponding to the candidate content; Video bounce correction features are generated based on the degree of video playback truncation, user historical interest characteristics, and hotspot changes. User historical interest characteristics, global interest change vector, and video bounce correction features are combined according to the non-negative weights determined by the cloud side to obtain user comprehensive interest characteristics. Recall filtering conditions, edge-side reference correction amount, and threshold adjustment signals are generated based on user comprehensive interest characteristics.

[0011] Preferably, the hierarchical processing method includes a first-level method, a second-level method, and a third-level method, respectively corresponding to the local processing method on the edge, the coarse ranking processing method on the cloud, and the fine ranking processing method on the cloud; the edge pre-caches the preloaded candidate set issued by the cloud; in the first-level method, the edge performs local matching from the preloaded candidate set and displays it. In the two-level mode, the local matching results are first displayed on the client side, and the cloud side performs coarse matching within a limited candidate range. The limited candidate range is the set of candidate traffic-driving content obtained by filtering according to the recall filtering conditions, the upper limit of the number of candidates, and the preset risk coefficient. When the coarse matching results of the cloud side are returned within a preset response time and the ranking score of the coarse matching results of the cloud side is greater than the ranking score of the local matching results of the client side, the local matching results of the client side are overridden. In the three-level mode, the client-side reserves a placeholder area to be displayed or displays a pre-display result at the target display position. The cloud-side performs fine-grained matching on the complete set of candidate traffic-driving content and generates a sorted list. If no response is received within a limited time, the display conditions of the second-level mode are met, and the result of the second-level mode is downgraded to the first-level mode. The cloud-side sorts the candidate traffic-driving content according to the sorting coefficient and the risk coefficient, and determines the hierarchical processing mode according to the mode switching threshold.

[0012] Preferably, the content distribution decision package includes a decision method marker, a set of candidate traffic-driving content identifiers, a display position strategy, a display intensity strategy, and an effective time limit; after receiving the content distribution decision package, the client checks the current page status, decoding load, gesture status, and local candidate display status. The display conditions include that the current page has not been switched, the decoding load is lower than the preset load threshold, the gesture state is a non-continuous swipe state, and the local candidate display state is idle. The client generates a display score based on the sorting score of the candidate traffic-driving content, the display intensity strategy, and the current page state. When the display conditions are met and the display score is greater than or equal to the preset display threshold, the display is executed. When the display conditions are not met or the display score is less than the display threshold, the display is not executed. When the terminal detects that the number of swipes within a preset time window is greater than or equal to a preset swipe count threshold, or the swipe speed is greater than or equal to a preset swipe speed threshold, it determines that it is in a continuous swipe state and writes the distribution content decision package into a delay queue; when the number of swipes is less than the swipe count threshold and the swipe speed is less than the swipe speed threshold, it determines that it is in a discontinuous swipe state; when the effective time limit expires or the page context changes, the distribution content decision package is invalidated.

[0013] Preferably, the client generates and sends back the display result after the display is executed. The display result includes whether the display was executed, the execution time, whether it was covered, whether it was clicked, whether it was closed, whether it was redirected, and whether it remained after redirection. The cloud side distinguishes between cases of not being displayed in time, cases of display without response, and cases of being covered by results with higher processing levels based on the display result. The cloud side adjusts the display threshold, mode switching threshold, sorting coefficient, and risk coefficient based on the distinction results.

[0014] Preferably, a machine learning-based precise traffic acquisition system for short video platforms for executing the method described in this invention includes a terminal side, a cloud side, and a terminal-cloud communication module; the terminal-cloud communication module is used to upload the full change package or minimum event package generated on the terminal side to the cloud side, and to send the distribution content decision package generated on the cloud side to the terminal side. The endpoint includes an endpoint observation data management module, an endpoint interest generation module, a change upload module, and an endpoint display module. The endpoint observation data management module is used to receive behavioral data and video content features. The endpoint interest generation module is used to generate current interest features and interest change information. The change upload module is used to determine whether to upload the full change package, upload the minimum event package, or cache it locally. The endpoint display module is used to control the display timing based on the distributed content decision package and send back the display results. The cloud side includes a user historical data management module, a candidate content data management module, a cloud side reconstruction module, a hierarchical processing module, a content sorting module, a parameter adjustment module, and a cloud side distribution module; The user history data management module stores users' historical interest characteristics; the candidate content data management module stores candidate traffic-driving content identifiers, content vectors, estimated conversion rates, timeliness factors, and risk coefficients; the cloud-side reconstruction module generates comprehensive user interest characteristics; the hierarchical processing module determines the first-level, second-level, or third-level methods; the content sorting module generates a set of candidate traffic-driving content; the parameter adjustment module adjusts thresholds, sorting coefficients, and risk coefficients based on the display results; and the cloud-side distribution module generates and distributes content decision packages. Among them, the estimated conversion rate is used to characterize the target conversion attribute of the candidate traffic-generating content, the timeliness factor is used to characterize the time decay attribute of the candidate traffic-generating content, and the risk coefficient is used to characterize the risk attribute of the candidate traffic-generating content.

[0015] Compared with the prior art, the present invention has the following advantages: This invention generates interest change information on the client side and dynamically selects upload strategies in conjunction with link congestion information. Combined with the hierarchical processing method on the cloud side, it effectively reduces the amount of data uploaded and alleviates the problem of cloud response lag. This solves the defects of missing display opportunities and the disconnect between traffic-driving content and instant browsing status when users scroll quickly, and achieves accurate traffic-driving with low latency. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0017] Figure 1 A flowchart illustrating the precise traffic acquisition method for short video platforms based on machine learning, as provided in this embodiment of the invention.

[0018] Figure 2 A schematic diagram of the module of the machine learning-based precise traffic acquisition system for short video platforms provided in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings.

[0020] This invention provides a method for precise traffic acquisition on short video platforms based on machine learning; this embodiment collects user short video browsing behavior data on the client side and extracts the video content features of the current video; The client inputs the above data into the feature extraction model to generate current interest features, and determines interest change information based on the current interest features and reference interest features; the client determines the validity of behavior and link congestion information based on the user's browsing integrity and other statuses, and selects the appropriate upload strategy accordingly. After receiving the event packet, the cloud side reconstructs the user's comprehensive interest characteristics by combining the user's historical interest characteristics and hotspot changes. Finally, the cloud side determines the hierarchical processing method, generates the content distribution decision package, and sends it to the terminal side for display. Through the above steps, low-latency content delivery is achieved under the conditions of preset terminal computing power limitations and network bandwidth limitations. Reference Figure 1 Machine learning-based methods for precise user acquisition on short video platforms include: Step 1: Collect user short video browsing behavior data on the device side and extract the video content features of the current video to obtain the device-side collected data; Step 2: The edge inputs the data collected at the edge into the feature extraction model to generate the current interest feature vector, and determines the interest change information based on the current interest feature vector and the pre-saved reference interest feature vector. Step 3: The endpoint determines the validity information of the behavior and the link congestion information based on the data collected on the endpoint, and selects the corresponding upload strategy accordingly. The upload strategy includes uploading the full change packet, uploading the minimum event packet, or writing to the endpoint cache queue. Step 4: After receiving the event packet containing user identifier, session identifier, timestamp and interest change information, the cloud side obtains the user's historical interest characteristics and determines the hotspot changes based on the exposure, click or conversion changes of candidate content in the cloud side's pre-loaded candidate set in the statistics window. The user's comprehensive interest characteristics are reconstructed by combining the user's historical interest characteristics, hotspot changes and interest change information. Step 5: The cloud side determines the hierarchical processing method based on the user's comprehensive interest characteristics, link status, and interest change information, generates a content distribution decision package, and sends it to the terminal side for display.

[0021] Based on the above embodiments, the steps of edge-side data acquisition and interest feature generation are further defined; User short video browsing behavior data includes swiping speed, dwell time, click status, scroll back status, interaction quality status, audio interaction status, and entry into product page status; interaction quality status includes at least one of the following: like status, comment status, favorite status, share status, and follow status; audio interaction status includes audio on status, mute status, volume adjustment status, or audio repeat playback status; video content features include keyframe visual features, speech rhythm features, optical flow change features, and text tag features.

[0022] Acquire short video browsing behavior data of users on the terminal side, and extract features from the current video according to the preset feature extraction logic to obtain the terminal side collected data; among them, behavior data refers to the time sequence signal that represents the interaction state between the user and the terminal interface, such as the pixel movement rate in the touch event trajectory representing the swipe speed, dwell time, click state, swipe back state, interaction quality state, audio interaction state, and entry into the product page state. Interaction quality states include at least one of the following: like state, comment state, etc., used to characterize the user's explicit approval of the current content; audio interaction states include audio on state, mute state, etc., used to characterize the user's auditory attention to the current content; video content features refer to lightweight descriptors that reflect the multimodal semantics of the video, such as keyframe visual features, speech rhythm features, optical flow change features, and text tag features. The specific feature extraction logic is as follows: key frame visual features are generated by extracting one frame every preset number of frames and reducing the resolution to a preset size, and directly calculating the mean and variance combination of each channel of the pixel (red, green, and blue); speech rhythm features are generated by statistically analyzing the short-time zero-crossing rate and energy peak frequency of the audio signal within a fixed time window. Optical flow variation features are generated by calculating the average pixel displacement between adjacent downsampled frames; text tag features are generated by mapping the preset classification tags inherent in the video metadata to multi-dimensional one-hot vectors; this embodiment of the invention extracts lightweight multimodal summary features to obtain shallow semantic and behavioral action features, reducing forward inference computational power consumption and responding within a millisecond window; this embodiment of the invention collects transient intents from the sensing end based on lightweight features, shortens processing time, and overcomes the defect of missing the lead-in window due to excessive time consumption of high-dimensional feature extraction; The edge device inputs the data collected on the edge device into the feature extraction model deployed on the edge device to generate the current interest feature vector; the edge device saves the reference interest feature vector corresponding to the last valid feature update, and generates an interest change vector and interest change magnitude based on the current interest feature vector and the reference interest feature vector; when the currently acquired behavioral data meets the preset validity conditions, and the edge device receives a cloud processing confirmation mark, an edge device display feedback signal or a timeout confirmation signal, the edge device updates the reference interest feature vector with the current interest feature vector.

[0023] The system acquires behavioral data and video content features, processes the data according to fusion logic, and generates a current interest feature vector. Specifically, the current interest feature vector is generated using a fusion formula. in, This is the current interest feature vector. For behavioral data vectors, For video content feature vectors, The behavior mapping weight matrix, The weight matrix is ​​used for mapping multimodal summarization. For bias vectors, It is a nonlinear transformation function; The obtained dimension is behavioral data vector With dimension Video content feature vector Combined, using dimensional configuration as Behavior mapping weight matrix Characterize the intent weights corresponding to different behavioral states; utilize dimension configuration as Multimodal summarization mapping weight matrix Map content features to a dimension of The intention space, combined with dimensions of the same bias vector This ensures that the output dimensions of each component are consistent after matrix multiplication, through a nonlinear transformation function. Process it; Among them, the above weight matrix , and bias vector The parameters that constitute the shallow two-layer fully connected neural network model are obtained by offline training and convergence of a sample set based on historical user browsing behavior-content conversion on the cloud side, and are periodically distributed and deployed on the edge side. During training, the sample set is obtained by combining the behavioral features and content features of historical product clicks or purchase conversions as positive samples, and those ignored by swiping as negative samples; the behavioral features and content features of historical product clicks or target page jumps are combined as positive samples. The loss function is calculated using the following formula: in, For the summation operator, For positive sample set, For negative sample set, The cosine similarity function is used. This is the intent mapping vector corresponding to the behavioral data. This is the intent mapping vector corresponding to the content features. These are the preset boundary hyperparameters. This is a function to find the maximum value. Obtain the reference interest feature vector corresponding to the previous valid observation. Based on the difference logic, compare the current interest feature vector with the reference interest feature vector to generate the interest change vector and the corresponding interest change magnitude. In specific implementation, the interest change vector is obtained by subtracting the reference interest feature vector from the current interest feature vector, which is the corresponding difference vector. The magnitude of interest change is calculated as the Euclidean distance of the interest change vector, i.e., the L2 norm; the feature residuals are dimensionally consistent and scalarized in the metric space for subsequent numerical comparison based on a single threshold; where the interest change vector refers to the residual vector that represents the instantaneous offset direction and magnitude of the current transient intention relative to the existing stable intention. When the currently acquired behavioral data meets the preset validity conditions, and the edge receives a cloud-side processing confirmation flag, an edge display feedback signal, or a timeout confirmation signal, the edge updates the reference interest feature vector with the current interest feature vector; in specific implementation, the updated reference interest feature vector is generated through update logic: Using adaptive smoothing coefficient The reference interest feature vector corresponding to the previous valid observation and the current interest feature vector Weighted fusion is performed to obtain the updated reference interest feature vector. This is used to smoothly absorb new observation features after the observation has been confirmed to be valid; Adaptive smoothing coefficient lie in The interval, whose specific values ​​are determined by the current interest feature vector. Reference interest feature vector corresponding to the previous valid observation The cosine similarity between them is dynamically determined, and the calculation formula is: in, This is the reference interest feature vector corresponding to the previous valid observation. A default positive real number penalty term is used to prevent division by zero anomalies. This represents the vector dot product operation. The L2 norm of a vector. The function aims to maximize the similarity coefficient, ensuring that when the similarity is greater than the similarity threshold, the smoothing coefficient approaches 1 to maintain the historical baseline, and when the similarity is less than the similarity threshold, the smoothing coefficient decreases accordingly, thereby accelerating the absorption of new observation features. This scheme suppresses noise data generated by rapid sliding by using a comparison-then-update control method. During the first effective observation, due to the lack of historical features, the current interest feature vector is directly assigned to the reference interest feature vector for initialization. The embodiments of this invention represent relative intentions based on residual comparison and delayed update logic, preventing invalid feature writing and overcoming the defect of intention recognition distortion caused by the accumulation of noise during rapid sliding. In the case of continuous and rapid scrolling on short video platforms, if the time taken for a single feature extraction exceeds the time spent in rapid scrolling, it will lead to feature extraction lag and failure to output effective intent. This embodiment adopts a multimodal summarization and interest feature generation mechanism to control the time taken for feature extraction and interest change calculation within a preset threshold, thereby capturing transient intent shifts within a short time window. Based on the above embodiments, the residual compression upload and cloud-side high-dimensional reconstruction steps are further defined; Select the corresponding upload strategy, including: User browsing integrity is determined based on dwell time and video duration; a behavior effectiveness score is generated based on user browsing integrity, dwell time, interaction quality status, and scrollback status; a link congestion index is generated based on the client-side upload queue length, round-trip latency, and packet loss rate; the client-side selects the corresponding upload strategy based on the comparison results of the interest change amplitude and the upload threshold issued by the cloud side, the behavior effectiveness score, and the comparison results of the link congestion index and the congestion threshold issued by the cloud side; wherein, the upload threshold and congestion threshold are generated and issued by the cloud side, and the client-side synchronously updates the upload threshold and congestion threshold according to the threshold adjustment signal issued by the cloud side.

[0024] The system acquires dwell time and video duration, calculates the percentage of dwell time to total video duration based on their ratio, evaluates user browsing status, and determines user browsing completeness. It also acquires user browsing completeness, dwell time, interaction quality status, and scrollback status, and enhances these indicators based on specific weighted fusion logic to generate a behavior effectiveness score. Positive weight coefficients are pre-configured for user browsing completeness, dwell time, interaction quality status, and slip-back status. Among them, the slip-back status and interaction quality status, which represent the first interest confirmation, are set with higher weight coefficients than dwell time. By multiplying the quantitative values ​​or Boolean state values ​​corresponding to each indicator (e.g., 1 for occurrence and 0 for non-occurrence) with their corresponding weight coefficients and summing them, a specific behavior effectiveness score is generated. Among them, the behavior effectiveness score refers to the quantitative index that characterizes the information completeness and credibility of the current observation segment; the system obtains the end-side upload queue length, round-trip latency and packet loss rate, and quantifies the current link status according to the network status assessment rules to generate the link congestion index; the system sets normalization factors for the end-side upload queue length, round-trip latency and packet loss rate respectively, maps them uniformly to a fixed interval, and performs linear weighted summation based on the cloud-side preset congestion weight configuration to obtain the final link congestion index; The system acquires the magnitude of interest changes, the effectiveness score of behavior, and the link congestion index. Based on the threshold judgment logic, it comprehensively analyzes the above indicators to determine the upload strategy: upload the full change packet, upload the minimum event packet, or write it to the end-side cache queue, i.e., execute the local cache operation. Determine the intensity of the change in intent. If the magnitude of the change in interest is greater than or equal to the upload threshold, determine the network status. If the link congestion index is lower than the congestion threshold, select to upload the full change packet. If the link congestion index is greater than or equal to the congestion threshold, select to upload the minimum event packet. Conversely, when the change in interest is less than the upload threshold, the current data is directly written to the end-side cache queue as local cache data, regardless of the congestion state of the link, for local caching. The specific mechanism for generating upload thresholds and congestion thresholds from the cloud side is as follows: The cloud side constructs and maintains a network congestion control state machine based on an exponentially weighted moving average algorithm; In practice, the cloud side continuously collects the current bandwidth load rate and the actual packet loss rate within the historical evaluation window of the global server cluster at fixed time intervals. The current collected data and the historical smoothed value are attenuated and fused by the exponential weighted moving average algorithm to filter out network jitter noise caused by transient burst traffic and calculate the smoothed bandwidth load rate and smoothed packet loss rate. When the smooth bandwidth load rate or smooth packet loss rate exceeds the preset security threshold, the cloud side increases the upload threshold and congestion threshold sent to the terminal side by a preset penalty multiplier; if the smooth bandwidth load rate or smooth packet loss rate is lower than the preset security threshold for several consecutive periods, the threshold is decreased by a fixed linear constant step size; the embodiments of the present invention communicate based on a dynamic upload strategy, reduce the amount of data reported, and overcome the defect of delayed cloud response caused by network congestion. When multiple interest change vectors in the edge cache queue satisfy the direction consistency rule within a continuous observation window, the edge merges the multiple interest change vectors and uploads them. The direction consistency rule includes that the cosine similarity of the angle between vectors is greater than the direction consistency threshold, which is a preset scalar value, or that multiple interest change vectors have the same positive and negative polarity on the same feature dimension; the cloud side returns a processing confirmation flag, a pattern flag, and a reference correction suggestion for the uploaded event packet. The pattern flag is used to indicate whether the end side should subsequently use full change packet upload, minimum event packet upload, or local caching. The reference correction suggestion is used to indicate whether the end side should perform incremental correction on the reference interest feature vector. The endpoint calculates the dot product score between the current interest feature vector and the feature vectors corresponding to each candidate content in the cloud-side preloaded candidate set. The endpoint selects the top preset number of candidate contents with the highest dot product scores to generate the endpoint local matching result. If no processing confirmation mark is received within the timeout period, the endpoint local matching result is used to continue displaying the data, and the event is marked as a sample not participated in by the cloud side. The event is then removed from the training sample set or parameter adjustment sample set.

[0025] Multiple interest change vectors are obtained from the cache queue on the client side. According to the direction consistency rule, interest offsets smaller than a preset amplitude are accumulated and judged. When the condition is met, multiple interest change vectors are merged and uploaded. The direction consistency rule refers to the logic of determining whether a user has formed a stable same-direction interest offset in multiple consecutive videos. For example, the vector angle is less than the direction consistency threshold, or multiple interest change vectors have the same positive and negative polarity on the same feature dimension. The cloud side returns a processing confirmation flag, a mode flag, and a reference correction suggestion to the uploaded event packet. The specific execution logic of the endpoint using the reference correction suggestion to incrementally correct the local reference interest feature vector is as follows: the endpoint parses the correction compensation feature and correction decay rate in the reference correction suggestion, and adds the current reference interest feature vector and the correction compensation feature according to the correction decay rate to perform vector weighted addition, thereby integrating the global intent compensation from the cloud side perspective locally. The edge generates local matching results from the local preloaded candidate set based on the current interest feature vector. If no processing confirmation mark is received within the timeout period, the edge local matching results are used to continue to be displayed, and the event is marked as a sample that the cloud side did not participate in. The event is then removed from the training sample set or parameter adjustment sample set. Upon determining a timeout, the endpoint generates a failure synchronization instruction carrying the event identifier and writes it to the endpoint's local cache queue. When the network is restored and the event is uploaded again, the terminal will attach the failed synchronization instruction to the data packet and synchronize it to the cloud. After receiving it, the cloud will perform a dual state check based on the event identifier. If it finds that the corresponding event has been received and temporarily stored by the cloud, it will perform a backtracking removal operation on the delayed event in the cloud dataset to maintain the consistency of the terminal-cloud collaborative training samples. This solution accumulates continuous weak intents based on the direction consistency rule, overcoming the defect of missed intent detection due to excessively high single trigger threshold; Reconstructing comprehensive user interest characteristics, including: The cloud side retrieves the sequence of received events, consisting of event identifier, session identifier, and timestamp, from the cloud-side cache based on the user identifier and session identifier. The cloud side determines the current session time window based on the start and end times of the current session and determines the cross-session tolerance window based on the preset time interval between adjacent sessions. Event packets that fall within the current session time window and whose event identifiers are not duplicated are rearranged according to their timestamps; event packets with duplicate event identifiers, or event packets with inconsistent session identifiers and timestamps exceeding the cross-session tolerance window are discarded; the cloud side converts interest change information into a global interest change vector with the same dimension as the feature vector corresponding to the candidate content; Video bounce correction features are generated based on the degree of video playback truncation, user historical interest characteristics, and hotspot changes. User historical interest characteristics, global interest change vector, and video bounce correction features are combined according to the non-negative weights determined by the cloud side to obtain user comprehensive interest characteristics. Recall filtering conditions, edge-side reference correction amount, and threshold adjustment signals are generated based on user comprehensive interest characteristics.

[0026] The user identifier and session identifier are obtained. According to the time sequence alignment rules, the received event sequence is matched and extracted from the cloud-side cache to obtain the event sequence composed of event identifier, session identifier and timestamp. The cloud side determines the current session time window based on the start and end time of the current session, and determines the cross-session tolerance window in combination with the preset time interval of adjacent sessions. The event packets are then rearranged or discarded in a refined manner: valid event packets that fall into the current session time window and whose event identifiers are not duplicated are strictly rearranged according to the timestamp. Discard abnormal event packets with duplicate event identifiers or inconsistent session identifiers and timestamps exceeding the cross-session tolerance window; obtain interest change information, video playback truncation degree, user historical interest characteristics and hotspot changes, and dynamically balance the above characteristics according to the reconstruction calculation logic to generate comprehensive user interest characteristics; Through formula User historical interest characteristics Transform into a global interest change vector with a global dimension And video bounce correction features generated based on changes in trending topics. Process it; The specific rule for transforming to the global dimension is as follows: the low-dimensional interest change vector uploaded from the client is linearly mapped through a pre-initialized feature projection matrix, expanding it to the same dimension as the candidate content features to obtain... The video jumps out of the correction feature. The truncation scalar is calculated by multiplying the video playback truncation degree (set to 1 minus the ratio of the actual playback time to the total video time) by the hotspot content center vector corresponding to the hotspot changes in the same dimension as the candidate content. To ensure the user's historical interest characteristics Global interest change vector Video jump-out correction features Assuming all features are strictly aligned to the same global high-dimensional content feature space, to prevent dimensional imbalance during weighting due to differences in absolute numerical ranges among the feature vectors, the following measures are taken: , , Perform L2 normalization operations on each to ensure that they are uniformly mapped onto the unit hypersphere; Based on this, the three will be weighted using non-negative methods. , , The weighted vectors are summed and combined, and the result is normalized again using the second norm to obtain the final comprehensive user interest features. ; to ensure that the metric space remains consistent during the reconstruction of the high-dimensional intention space; The aforementioned non-negative weights are used to dynamically balance historical priors, current residual observations, and truncation compensation; in practical applications, these non-negative weights satisfy... The system sets the basic logarithmic weight score assigned to the user's historical prior knowledge as follows: The non-negative weight allocation process is as follows: the system sets the basic logarithmic weight score assigned to the user's historical prior knowledge as follows: ,in The physical time interval between the current session and the user's last active session. The preset time decay constant ensures that the longer the time interval, the greater the exponential shrinkage of the influence of historical memory; Log-weighted scores assigned to the current residual observations Fixed as a system baseline scalar, such as 1.0, serving as a stable anchor point for real-time intent on the edge; the logarithmic weight score assigned for truncation compensation. The dynamic calculation is as follows: the discrete difference rate of change of the exposure click rate of candidate content within adjacent statistical time windows, that is, the discrete difference rate of change of conversion rate, multiplied by a preset positive incentive hyperparameter to characterize the magnitude of change in the popularity of sudden hot topics. The three log-weighted scores obtained Input the normalized exponential function for calculation: The final output is the normalized non-negative weights. ;in, For the summation operator, For the natural constant An exponential function with base 0; The specific rules for generating recall filtering conditions, edge reference correction amount and threshold adjustment signal based on user comprehensive interest features are as follows: calculate the cosine similarity between the generated user comprehensive interest features and the preset center vector of each content category in the system, and add the categories with similarity lower than the preset truncation threshold to the recall filtering conditions to directly block their corresponding candidate content. Calculate the difference vector between the reconstructed user comprehensive interest features and the user's historical interest features. When the Euclidean distance of the difference vector is greater than a preset deviation value, the difference vector is scaled proportionally and sent as a reference correction amount on the end side. If the Euclidean distance of the difference vector continues to exceed the preset deviation value, it indicates that the user is in an active probing period, and a threshold adjustment signal for adjusting the high-end upload threshold is generated. This embodiment uses dynamic weight combination based on hotspot changes and video bounce correction features to achieve intent reconstruction, correct model bias, and overcome the defect of intent inference bias caused by lack of truncation perception mechanism. In this embodiment, based on the above embodiments, the hierarchical triggering decision and end-side display control steps are further defined; The hierarchical processing methods include Level 1, Level 2, and Level 3 methods, which correspond to local processing on the device side, coarse ranking processing on the cloud side, and fine ranking processing on the cloud side, respectively; the device side pre-caches the preloaded candidate set issued by the cloud side; In the first-level mode, the client performs local matching and displays the results from the pre-loaded candidate set. In the second-level mode, the client first displays the local matching results, and the cloud performs coarse matching within a limited candidate range. The limited candidate range is the set of candidate traffic-driving content filtered according to the recall filtering conditions, the upper limit of the number of candidates, and the preset risk coefficient. When the cloud coarse matching results are returned within the preset response time and the ranking score of the cloud coarse matching results is greater than the ranking score of the client local matching results, the client local matching results are overridden. In the three-level approach, the client-side reserves a placeholder area for display or displays a pre-display result at the target display position. The cloud-side performs fine-tuning and matching on the complete set of candidate traffic-driving content and generates a sorted list. If no response is received within a limited time, the result is downgraded to the second-level approach if the display conditions are met; otherwise, the result is downgraded to the first-level approach. The cloud-side sorts the candidate traffic-driving content according to the sorting coefficient and risk coefficient, and determines the graded processing method according to the method switching threshold.

[0027] The cloud side acquires link status, user attention rating, and interest change information. It then compares the specific decision tree with the system-set switching threshold to dynamically determine the appropriate hierarchical processing method. Based on structured encapsulation logic, it packages the determined hierarchical processing method and candidate traffic-driving content into a distribution content decision package. The decision tree logic is as follows: when the link congestion index is greater than or equal to the severe congestion threshold, or when the current computing power load on the cloud side is greater than the full load threshold, the first-level method is directly downgraded; when the link congestion index is between the mild congestion threshold and the severe congestion threshold, the second-level method is adopted; when the link congestion index is less than the mild congestion threshold, i.e. the network is unobstructed, and the magnitude of interest change is greater than the fine ranking trigger threshold, the third-level method is adopted. Among them, the graded processing method refers to the dynamic matching depth adapted to different network and computing power states. For example, the first-level method corresponds to the shallow and fast matching on the edge, the second-level method corresponds to the limited candidate coarse ranking on the cloud side, and the third-level method corresponds to the complete candidate fine ranking. In this process, the cloud side will comprehensively score and rank the candidate traffic-driving content based on the ranking coefficient and the risk coefficient. The specific comprehensive scoring calculation logic is as follows: subtract the product of the dynamic penalty factor and the current accumulated risk coefficient of the content from the basic ranking coefficient score of the candidate traffic-driving content, thereby effectively suppressing the final score and ranking of content that frequently causes negative feedback. In the second-level mode, when the cloud-side coarse matching result returns within the preset response time and meets the replacement conditions, it covers the local matching result on the terminal side; in the third-level mode, if it does not return within the timeout period, it will be downgraded and rolled back; this embodiment dynamically allocates processing resources based on the hierarchical processing method to overcome the defect of missing the display opportunity due to the lack of hierarchical control; The content distribution decision package includes decision method markers, candidate traffic-driving content identifier set, display position strategy, display intensity strategy, and effective time limit; after receiving the content distribution decision package, the client checks the current page status, decoding load, gesture status, and local candidate display status. The display conditions include that the current page has not been switched, the decoding load is lower than the preset load threshold, the gesture state is a non-continuous swipe state, and the local candidate display state is idle. The client generates a display score based on the sorting score of the candidate traffic-driving content, the display intensity strategy, and the current page state. When the display conditions are met and the display score is greater than or equal to the preset display threshold, the display is executed. When the display conditions are not met or the display score is less than the display threshold, the display is not executed. When the client detects that the number of swipes within a preset time window is greater than or equal to a preset swipe count threshold, or the swipe speed is greater than or equal to a preset swipe speed threshold, it determines that it is in a continuous swipe state and writes the content distribution decision package into the delay queue; when the number of swipes is less than the swipe count threshold and the swipe speed is less than the swipe speed threshold, it determines that it is in a discontinuous swipe state; when the effective time limit expires or the page context changes, the content distribution decision package is invalidated.

[0028] The system acquires the current page state, decoding load, gesture state, and local candidate display state. Based on the display condition judgment logic, it evaluates the candidate traffic-driving content, generates a display score, and executes display control. In this process, the logic for generating the display score is a specific rule accumulation mechanism: extracting the ranking score from the distribution content decision package as the base score. If the display intensity strategy in the decision package is marked as strong blocking, a fixed strong prompt bonus will be added; if the current page state is immersive browsing state or the decoding load is higher than 80% of the load threshold, a preset disturbance penalty score will be deducted from the base score, and the final display score will be output after linear calculation. The display conditions include that the current page has not been switched, the decoding load is lower than the preset load threshold, and the gesture state is a non-continuous swipe state. When the terminal detects that the number of swipes within the preset time window is greater than or equal to the preset number of swipes threshold, or the swipe speed is greater than or equal to the preset swipe speed threshold, it is determined to be in a continuous swipe state, and the content distribution decision package is written into the delay queue. The aforementioned page context switching includes the user exiting the current short video feed container; when the user swipes within the feed, the decision package in the delay queue is retained until the gesture ends, meeting the display conditions and the effective time limit has not expired; this embodiment controls the display timing based on multi-dimensional state checks, overcoming the defect of increased negative interactions due to the lack of a display timing perception mechanism; After the client executes the display, it generates and sends back the display results, which include whether the display was executed, the execution time, whether it was covered, whether it was clicked, whether it was closed, whether it was redirected, and whether it remained after redirection. The cloud side distinguishes between cases of not being displayed in time, cases of displaying without response, and cases of being covered by results with higher processing levels based on the display results. The cloud side adjusts the display threshold, method switching threshold, sorting coefficient, and risk coefficient based on the distinction results.

[0029] The system acquires data such as whether the display was executed and the execution time. Based on the result feedback mechanism, it analyzes the display results to distinguish between cases of not being displayed in a timely manner, displaying without response, and being covered by results with higher processing levels. It also adaptively adjusts the display threshold, method switching threshold, ranking coefficient, and risk coefficient. Among them, the risk coefficient refers to a quantitative penalty indicator that represents the probability of candidate traffic-driving content causing negative user interactions. When a feedback signal indicating no response or closure is received consecutively, the risk coefficient is increased by multiplying the previous step by a preset coefficient greater than 1. The S-shaped activation function is used to calculate the final risk coefficient, which constrains the risk value within the interval [0,1]. This embodiment dynamically adjusts parameters based on the feedback of the display results, overcoming the defect that updates fail due to a lack of real display confirmation; Reference Figure 2 A machine learning-based precise traffic acquisition system for short video platforms for performing the methods of the embodiments includes an end-side, a cloud-side, and an end-cloud communication module; The edge-cloud communication module is used to upload the full change package or minimum event package generated on the edge to the cloud, and to send the distribution content decision package generated on the cloud to the edge. The edge-side includes an edge-side observation data management module, an edge-side interest generation module, a change upload module, and an edge-side display module; The edge observation data management module is used to receive behavioral data and video content features; the edge interest generation module is used to generate current interest features and interest change information; the change upload module is used to determine whether to upload the full change package, upload the minimum event package, or cache it locally; and the edge display module is used to control the display timing based on the distributed content decision package and send back the display results. The cloud side includes a user historical data management module, a candidate content data management module, a cloud side reconstruction module, a hierarchical processing module, a content sorting module, a parameter adjustment module, and a cloud side distribution module; The user history data management module stores users' historical interest characteristics; the candidate content data management module stores candidate traffic-driving content identifiers, content vectors, estimated conversion rates, timeliness factors, and risk coefficients; the cloud-side reconstruction module generates comprehensive user interest characteristics; the hierarchical processing module determines the first-level, second-level, or third-level methods; the content sorting module generates a set of candidate traffic-driving content; the parameter adjustment module adjusts thresholds, sorting coefficients, and risk coefficients based on the display results; and the cloud-side distribution module generates and distributes content decision packages. Among them, the estimated conversion rate is used to characterize the target conversion attribute of the candidate traffic-generating content, the timeliness factor is used to characterize the time decay attribute of the candidate traffic-generating content, and the risk coefficient is used to characterize the risk attribute of the candidate traffic-generating content.

[0030] This invention also provides a machine learning-based precise traffic acquisition system for short video platforms. This system is deployed in a short video distribution architecture and includes a terminal side, a cloud side, and a terminal-cloud communication module. The terminal side includes a terminal side observation data management module, a terminal side interest generation module, a change upload module, and a terminal side display module. The cloud side includes a user historical data management module, a candidate content data management module, a cloud-side reconstruction module, a hierarchical processing module, a content sorting module, a parameter adjustment module, and a cloud-side distribution module. In actual business scenarios, such as e-commerce short video shopping guide applications, when a user scrolls through a video, new data flows in from the end-side observation data management module. After the end-side interest generation module extracts the residuals, the change upload module dynamically triggers event packets to upload based on the current network bandwidth. After receiving the event packet, the cloud-side reconstruction module restores the high-dimensional intent and combines it with the estimated conversion rate in the candidate content data management module, such as the estimated conversion probability, and the timeliness factor, such as the content lifecycle timestamp. The content sorting module generates product cards or jump entrances, which are then packaged into a decision package and sent to the terminal display module. The on-device display module captures the non-continuous swiping state when the user stops swiping and triggers the display, and sends the click or close behavior back to the parameter adjustment module to update the threshold and risk coefficient.

[0031] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make several improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A machine learning-based method for precise traffic acquisition on short video platforms, characterized in that: include: Step 1: Collect user short video browsing behavior data on the device side and extract the video content features of the current video to obtain the device-side collected data; Step 2: The edge side inputs the edge-collected data into the feature extraction model to generate the current interest feature vector, and determines the interest change information based on the current interest feature vector and the pre-saved reference interest feature vector; Step 3: The endpoint determines the behavior validity information and link congestion information based on the data collected at the endpoint, and selects the corresponding upload strategy accordingly. The upload strategy includes uploading the full change packet, uploading the minimum event packet, or writing to the endpoint cache queue. Step 4: After receiving the event packet containing user identifier, session identifier, timestamp and interest change information, the cloud side obtains the user's historical interest characteristics and determines the hotspot changes based on the exposure, click or conversion changes of candidate content in the cloud side's pre-loaded candidate set in the statistics window. The user's comprehensive interest characteristics are reconstructed by combining the user's historical interest characteristics, hotspot changes and interest change information. Step 5: The cloud side determines the hierarchical processing method based on the user's comprehensive interest characteristics, link status, and interest change information, generates a content distribution decision package, and sends it to the terminal side for display.

2. The method for precise traffic acquisition on short video platforms based on machine learning according to claim 1, characterized in that, The user's short video browsing behavior data includes swiping speed, dwell time, click status, swipe back status, interaction quality status, audio interaction status, and entry into the product page status; the interaction quality status includes at least one of the following: like status, comment status, favorite status, share status, and follow status; the audio interaction status includes audio on status, mute status, volume adjustment status, or audio repeat playback status; the video content features include keyframe visual features, speech rhythm features, optical flow change features, and text tag features.

3. The method for precise traffic acquisition on short video platforms based on machine learning according to claim 1, characterized in that, The edge device inputs the edge-collected data into the feature extraction model deployed on the edge device to generate the current interest feature vector; the edge device saves the reference interest feature vector corresponding to the last valid feature update, and generates an interest change vector and interest change magnitude based on the current interest feature vector and the reference interest feature vector; when the currently acquired behavioral data meets the preset validity conditions, and the edge device receives a cloud-side processing confirmation mark, an edge device display feedback signal, or a timeout confirmation signal, the edge device updates the reference interest feature vector with the current interest feature vector.

4. The method for precise traffic acquisition on short video platforms based on machine learning according to claim 3, characterized in that, Select the corresponding upload strategy, including: User browsing completeness is determined based on dwell time and video duration; a behavior effectiveness score is generated based on user browsing completeness, dwell time, interaction quality status, and scrollback status; a link congestion index is generated based on the end-side upload queue length, round-trip latency, and packet loss rate; the end-side selects a corresponding upload strategy based on the comparison result of the interest change amplitude and the upload threshold issued by the cloud side, the behavior effectiveness score, and the comparison result of the link congestion index and the congestion threshold issued by the cloud side; wherein, the upload threshold and the congestion threshold are generated and issued by the cloud side, and the end-side synchronously updates the upload threshold and the congestion threshold according to the threshold adjustment signal issued by the cloud side.

5. The method for precise traffic acquisition on short video platforms based on machine learning according to claim 4, characterized in that, When multiple interest change vectors in the edge-side cache queue satisfy the direction consistency rule within a continuous observation window, the edge merges the multiple interest change vectors and uploads them. The direction consistency rule includes the cosine similarity of the angle between vectors being greater than the direction consistency threshold, where the direction consistency threshold is a preset scalar value, or multiple interest change vectors having the same positive and negative polarity on the same feature dimension; the cloud side returns a processing confirmation flag, a mode flag, and a reference correction suggestion for the uploaded event packet, where the mode flag is used to indicate whether the end side should subsequently use full change packet upload, minimum event packet upload, or local caching, and the reference correction suggestion is used to indicate whether the end side should perform incremental correction on the reference interest feature vector; The endpoint calculates the dot product score between the current interest feature vector and the feature vectors corresponding to each candidate content in the cloud-side preloaded candidate set. The endpoint selects the top preset number of candidate contents with the highest dot product scores to generate the endpoint local matching result. If no processing confirmation flag is received within the timeout period, the endpoint local matching result is used to continue displaying the data, and the event is marked as a sample not participated in by the cloud side. The event is then removed from the training sample set or parameter adjustment sample set.

6. The method for precise traffic acquisition on short video platforms based on machine learning according to claim 1, characterized in that, Reconstructing comprehensive user interest characteristics, including: The cloud side retrieves the sequence of received events, consisting of event identifier, session identifier, and timestamp, from the cloud-side cache based on the user identifier and session identifier. The cloud side determines the current session time window based on the start and end times of the current session and determines the cross-session tolerance window based on the preset time interval between adjacent sessions. Event packets that fall within the current session time window and whose event identifiers are not duplicated are rearranged according to their timestamps; event packets with duplicate event identifiers, or event packets with inconsistent session identifiers and timestamps exceeding the cross-session tolerance window are discarded; the cloud side converts interest change information into a global interest change vector with the same dimension as the feature vector corresponding to the candidate content; Video bounce correction features are generated based on the degree of video playback truncation, user historical interest characteristics, and hotspot changes. User historical interest characteristics, global interest change vector, and video bounce correction features are combined according to the non-negative weights determined by the cloud side to obtain user comprehensive interest characteristics. Recall filtering conditions, edge-side reference correction amount, and threshold adjustment signals are generated based on user comprehensive interest characteristics.

7. The method for precise traffic acquisition on short video platforms based on machine learning according to claim 6, characterized in that, The hierarchical processing method includes a first-level method, a second-level method, and a third-level method, respectively corresponding to the local processing method on the edge side, the coarse ranking processing method on the cloud side, and the fine ranking processing method on the cloud side; the edge side pre-caches the preloaded candidate set issued by the cloud side; In the first-level mode, the endpoint performs local matching and displays the pre-loaded candidate set. In the two-level mode, the local matching results are first displayed on the client side, and the cloud side performs coarse matching within a limited candidate range. The limited candidate range is the set of candidate traffic-driving content obtained by filtering according to the recall filtering conditions, the upper limit of the number of candidates, and the preset risk coefficient. When the coarse matching results of the cloud side are returned within a preset response time and the ranking score of the coarse matching results of the cloud side is greater than the ranking score of the local matching results of the client side, the local matching results of the client side are overridden. In the three-level mode, the client-side reserves a placeholder area to be displayed or displays a pre-display result at the target display position. The cloud-side performs fine-grained matching on the complete set of candidate traffic-driving content and generates a sorted list. If no response is received within a limited time, the display conditions of the second-level mode are met, and the result of the second-level mode is downgraded to the first-level mode. The cloud-side sorts the candidate traffic-driving content according to the sorting coefficient and the risk coefficient, and determines the hierarchical processing mode according to the mode switching threshold.

8. The method for precise traffic acquisition on short video platforms based on machine learning according to claim 7, characterized in that, The content distribution decision package includes a decision method marker, a set of candidate traffic-driving content identifiers, a display position strategy, a display intensity strategy, and an effective time limit; after receiving the content distribution decision package, the client checks the current page status, decoding load, gesture status, and local candidate display status. The display conditions include that the current page has not been switched, the decoding load is lower than the preset load threshold, the gesture state is a non-continuous swipe state and the local candidate display state is idle; the client generates a display score based on the sorting score of the candidate traffic-driving content, the display intensity strategy and the current page state, and executes the display when the display conditions are met and the display score is greater than or equal to the preset display threshold. If the display conditions are not met or the display score is less than the display threshold, no display will be performed. When the terminal detects that the number of swipes within a preset time window is greater than or equal to a preset swipe count threshold, or the swipe speed is greater than or equal to a preset swipe speed threshold, it determines that it is in a continuous swipe state and writes the distribution content decision package into a delay queue; when the number of swipes is less than the swipe count threshold and the swipe speed is less than the swipe speed threshold, it determines that it is in a discontinuous swipe state; when the effective time limit expires or the page context changes, the distribution content decision package is invalidated.

9. The method for precise traffic acquisition on short video platforms based on machine learning according to claim 8, characterized in that, The client generates and sends back the display result after the display is executed. The display result includes whether the display was executed, the execution time, whether it was covered, whether it was clicked, whether it was closed, whether it was redirected, and whether it remained after redirection. The cloud side distinguishes between cases where the display is not timely, cases where the displayed result is unresponsive, and cases where the result is covered by a higher processing level, based on the display results. The cloud side adjusts the display threshold, mode switching threshold, sorting coefficient, and risk coefficient based on the distinction results.

10. A machine learning-based precise traffic acquisition system for short video platforms for performing the method according to any one of claims 1 to 9, characterized in that, Includes edge-side, cloud-side, and edge-cloud communication modules; The edge-cloud communication module is used to upload the full change package or minimum event package generated on the edge to the cloud, and to send the distribution content decision package generated on the cloud to the edge. The endpoint includes an endpoint observation data management module, an endpoint interest generation module, a change upload module, and an endpoint display module; The edge observation data management module is used to receive behavioral data and video content features; the edge interest generation module is used to generate current interest features and interest change information; the change upload module is used to determine whether to upload the full change package, upload the minimum event package, or cache it locally; and the edge display module is used to control the display timing based on the distributed content decision package and send back the display results. The cloud side includes a user historical data management module, a candidate content data management module, a cloud side reconstruction module, a hierarchical processing module, a content sorting module, a parameter adjustment module, and a cloud side distribution module; The user history data management module stores users' historical interest characteristics; the candidate content data management module stores candidate traffic-driving content identifiers, content vectors, estimated conversion rates, timeliness factors, and risk coefficients; the cloud-side reconstruction module generates comprehensive user interest characteristics; the hierarchical processing module determines the first-level, second-level, or third-level methods; the content sorting module generates a set of candidate traffic-driving content; the parameter adjustment module adjusts thresholds, sorting coefficients, and risk coefficients based on the display results; and the cloud-side distribution module generates and distributes content decision packages. Among them, the estimated conversion rate is used to characterize the target conversion attribute of the candidate traffic-generating content, the timeliness factor is used to characterize the time decay attribute of the candidate traffic-generating content, and the risk coefficient is used to characterize the risk attribute of the candidate traffic-generating content.

Images