Method and system for analyzing and recommending streaming short video content of short video platform

Abstract

The invention provides a short video platform streaming short video content analysis and recommendation method and system, and belongs to the technical field of internet information, and the method comprises the steps: firstly obtaining a to-be-streaming short video content set and a real-time streaming parameter, carrying out the hierarchical association processing of the to-be-streaming short video content set based on the real-time streaming parameter, and dividing a content hierarchical cluster; performing dynamic adaptive adjustment on the cluster in combination with the real-time interaction track of the user to generate a dynamic adaptive cluster; the dynamic adaptive cluster is associated and matched with a flow throwing resource pool, and a flow throwing recommendation priority sequence is determined; and finally, generating a streaming recommendation instruction based on the streaming recommendation priority sequence and the collaborative streaming parameters, and sending the streaming recommendation instruction to a platform streaming scheduling module, thereby realizing accurate analysis and recommendation of short video streaming content, and improving streaming accuracy and efficiency.

Description

Technical Field

[0001] This invention relates to the field of Internet information technology, and more specifically, to a method and system for analyzing and recommending short video content delivered to short video platforms. Background Technology

[0002] In the current era of rapid development in the short video industry, short video platforms have become important battlegrounds for information dissemination and commercial promotion. Traffic generation, as a key means of increasing the exposure and influence of short videos, directly impacts the revenue of creators and advertisers. However, existing methods for analyzing and recommending traffic generation content on short video platforms have many limitations.

[0003] On the one hand, traditional methods, after acquiring the short video content to be streamed and related parameters, often only perform simple classification processing without fully considering the complexity and dynamism of real-time streaming parameters. For example, insufficient correlation analysis between the time period and target audience tags makes it impossible to accurately segment the streaming content according to different time periods and audience characteristics, resulting in an unreasonable allocation of streaming resources. Some high-quality content may not receive sufficient exposure due to timing or audience mismatch.

[0004] On the other hand, existing methods lack effective utilization of real-time user interaction patterns during content recommendation. Real-time user interactions, such as clicks, pauses, and shares, accurately reflect users' interests and preferences for short video content. However, traditional methods fail to track these behavioral sequences in a timely manner and adjust content relevance weights accordingly. This results in a discrepancy between recommended short video content and actual user needs, hindering accurate recommendations and reducing user satisfaction and platform performance. Furthermore, traditional methods lack a scientific and rational basis for resource matching and priority determination, making it difficult to dynamically adjust recommendation strategies based on resource allocation ratios and user response characteristics. This leads to a lack of flexibility and adaptability in the recommendation process. Summary of the Invention

[0005] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for analyzing and recommending short video content delivered to a short video platform, the method comprising: Obtain the set of short video content to be streamed and the corresponding real-time streaming parameters. The set of short video content to be streamed includes multiple short video clips and content metadata associated with each short video clip. The real-time streaming parameters include the time period range, target audience tags, and resource allocation ratio. Based on the real-time streaming parameters, the set of short video content to be streamed is processed in a hierarchical manner. According to the overlapping characteristics of the time period range and the association characteristics of the target audience tags, the set of short video content to be streamed is divided into multiple content hierarchical clusters, and each content hierarchical cluster corresponds to a set of collaborative streaming parameters. By combining the real-time interaction trajectory of users on short video platforms, the content layer clusters are dynamically adapted and adjusted. By tracking the click behavior sequence, dwell time sequence and forwarding behavior sequence of users on similar content, the association weight of short video segments in each content layer cluster is corrected, and a dynamic adaptation cluster adapted to user behavior is generated. The dynamic adaptation clusters are associated and matched with the platform's traffic delivery resource pool. Based on the resource allocation ratio and the user response characteristics of the dynamic adaptation clusters, the traffic delivery recommendation priority sequence corresponding to each dynamic adaptation cluster is determined. Based on the priority sequence of traffic recommendations and the collaborative traffic recommendation parameters of each content layer cluster, a traffic recommendation instruction is generated and sent to the platform's traffic scheduling module.

[0006] In another aspect, embodiments of the present invention also provide a short video platform content analysis and recommendation system, including a processor and a machine-readable storage medium connected to the processor. The machine-readable storage medium is used to store programs, instructions or code, and the processor is used to execute the programs, instructions or code in the machine-readable storage medium to implement the above-mentioned method.

[0007] Based on the above, this embodiment of the invention obtains a set of short video content to be streamed and its corresponding real-time streaming parameters. Based on these parameters, it performs hierarchical association processing, dividing the content into hierarchical clusters according to the overlapping characteristics of the streaming time period and the association characteristics of the target audience tags, thus improving the rationality of streaming resource allocation. The content hierarchical clusters are dynamically adapted and adjusted by combining the real-time interaction trajectory of short video platform users. By tracking user clicks, dwell times, and forwarding sequences, the association weights of short video segments are corrected in real time, generating a dynamically adapted cluster that matches user behavior. This achieves accurate matching between streaming content and user needs, improving user acceptance and satisfaction with recommended content. The dynamically adapted clusters are then associated and matched with the platform's streaming resource pool. A streaming recommendation priority sequence is determined based on resource allocation ratios and user response characteristics. Finally, a streaming recommendation instruction is generated based on the streaming recommendation priority sequence and collaborative streaming parameters and sent to the platform's streaming scheduling module, effectively improving the accuracy and efficiency of short video platform streaming. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the execution flow of the short video platform content analysis and recommendation method provided in the embodiments of the present invention.

[0009] Figure 2 This is a schematic diagram of exemplary hardware and software components of the short video platform content analysis and recommendation system provided in this embodiment of the invention. Detailed Implementation

[0010] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating a method for analyzing and recommending short video content on a short video platform according to an embodiment of the present invention. The following is a detailed description of this method.

[0011] Step S110: Obtain the set of short video content to be streamed and the corresponding real-time streaming parameters. The set of short video content to be streamed includes multiple short video clips and content metadata associated with each short video clip. The real-time streaming parameters include the time period range, target audience tags, and resource allocation ratio.

[0012] This embodiment uses an e-commerce short video platform as an application scenario. This platform needs to analyze and recommend a batch of product promotion short videos to be streamed. The collection of short video content to be streamed covers promotional videos for various products, such as short video clips for clothing, electronic products, and home furnishings.

[0013] Each short video clip has associated metadata, which includes the video theme, such as showcasing a dress, introducing the features of a mobile phone model, or presenting a usage scenario of a sofa brand; video duration, i.e., the length of each short video from start to finish; information such as the brand, specifications, and price range of the products in the video; and the video shooting style, such as live-action shooting or animated demonstration.

[0014] The ad placement time range in the real-time ad delivery parameters is set by the platform based on past user activity and may include multiple time slots on weekdays and weekends, aiming to reach more potential users. Target audience tags are determined based on the product's target audience characteristics, such as age range, gender, consumption habits, and interests. For example, short videos targeting dresses might include young women and those interested in fashion and styling; short videos targeting mobile phones might include people interested in digital products and those with a certain level of purchasing power. Resource allocation percentage refers to the proportion of total ad delivery resources allocated to different categories of short videos, such as allocating a certain percentage to clothing short videos and another percentage to electronics short videos.

[0015] Step S120: Based on the real-time streaming parameters, perform hierarchical association processing on the set of short video content to be streamed. According to the overlapping characteristics of the streaming time range and the association characteristics of the target audience tags, divide the set of short video content to be streamed into multiple content hierarchical clusters. Each content hierarchical cluster corresponds to a set of collaborative streaming parameters.

[0016] In e-commerce short video platforms, the short video content to be streamed needs to be layered and associated according to real-time streaming parameters to form different content clusters. This process allows short videos with similar streaming characteristics to be grouped together, facilitating subsequent streaming management and recommendation.

[0017] Step S121: Extract the delivery time range from the real-time delivery parameters, decompose the delivery time range of each short video segment into multiple consecutive time sub-intervals, and calculate the proportion of the number of overlaps between any two short video segments in the time sub-intervals to the total number of sub-intervals, as the delivery time overlap feature.

[0018] After extracting the time period range from the real-time streaming parameters, to more accurately analyze the overlap of different short video clips' streaming times, it is necessary to decompose the streaming time period range of each short video clip into multiple consecutive time sub-intervals. The time interval for decomposition can be determined based on the platform's user activity patterns, such as using a fixed duration as a sub-interval, to ensure a detailed reflection of the time period distribution.

[0019] For each short video segment, its delivery time is divided into several consecutive and non-overlapping time sub-intervals according to a defined time interval. For example, if the delivery time of a short video is several hours in the morning and afternoon of a certain weekday, it is broken down into multiple sub-intervals in the morning and multiple sub-intervals in the afternoon.

[0020] To calculate the overlap in delivery time between any two short video clips, first identify all time sub-intervals that can be decomposed from these two short video clips, and then count the number of overlapping time sub-intervals. The total number of sub-intervals is the sum of the number of time sub-intervals for each of the two short video clips. Finally, divide the number of overlapping time sub-intervals by the total number of sub-intervals to obtain the ratio, which represents the overlap in delivery time between the two short video clips.

[0021] Step S122: Analyze the target audience tags, decompose each tag into multiple tag dimensions, each tag dimension contains multiple tag values, calculate the proportion of the number of tag values ​​that different short video clips share on the same tag dimension to the total number of tag values ​​in that tag dimension, and use this as the association feature of the target audience tags.

[0022] When parsing target audience tags, it is necessary to break them down into multiple tag dimensions. These tag dimensions can include age, gender, consumption level, interest category, etc. Each tag dimension contains multiple tag values. For example, the age dimension may contain multiple consecutive age ranges, the gender dimension includes male, female, etc., the consumption level dimension includes low, medium, and high levels, and the interest category dimension includes clothing, digital products, home furnishings, etc.

[0023] For different short video clips, their respective target audience tags are obtained, and these tags are mapped to various tag dimensions. Then, for each shared tag dimension, the number of tag values ​​that the two short video clips share in that dimension is counted. Simultaneously, the total number of tag values ​​under that tag dimension is determined. Finally, the ratio of the shared tag value count to the total number of tag values ​​is the association feature value for that tag dimension. By combining all the association feature values ​​for the same tag dimensions, the association features of the target audience tags of the two short video clips can be obtained.

[0024] Step S123: Construct a correlation evaluation matrix based on the overlapping features of the delivery time period and the correlation features of the target audience tags. Each element in the matrix corresponds to the comprehensive correlation degree of two short video segments. The comprehensive correlation degree is the weighted combination result of the overlapping features of the delivery time period and the correlation features of the target audience tags.

[0025] After obtaining the overlap characteristics of the delivery time periods and the correlation characteristics of the target audience tags, it is necessary to construct a correlation evaluation matrix. The rows and columns of the correlation evaluation matrix correspond to each short video segment in the short video content set to be delivered.

[0026] For each element in the correlation evaluation matrix, corresponding to the position of two short video clips, its value is the comprehensive correlation between the two short video clips. The comprehensive correlation is calculated by weighting the overlap characteristics of the ad placement time and the correlation characteristics of the target audience tags. The weights need to be determined based on the platform's ad placement strategy; the importance of ad placement time and target audience tags may differ in different ad placement scenarios, therefore different weight values ​​are assigned. The comprehensive correlation is obtained by multiplying the overlap characteristics of the ad placement time by their corresponding weights, and then adding the correlation characteristics of the target audience tags multiplied by their corresponding weights.

[0027] Step S124: A density-based clustering algorithm is used to cluster the association evaluation matrix, and short video clips with a comprehensive association degree higher than the clustering threshold are grouped into the same content hierarchical cluster, forming multiple initial content hierarchical clusters.

[0028] Step S1241: Initialize clustering parameters, set the clustering radius and the minimum number of included segments. The clustering radius is the comprehensive correlation threshold for determining whether two short video segments can be classified into the same content hierarchical cluster. The minimum number of included segments is the minimum number of short video segments that a content hierarchical cluster should contain.

[0029] Before performing clustering operations, clustering parameters need to be initialized. The cluster radius is a key parameter, determining the degree of correlation between two short video clips to be considered part of the same cluster. The minimum number of clips required for an effective content-layered cluster is specified, avoiding overly fragmented clusters.

[0030] The setting of the above clustering parameters needs to take into account the size and characteristics of the short video content set to be streamed. If there are a large number of short videos, the clustering radius and the minimum number of segments to be included may need to be adjusted accordingly to ensure the rationality of the clustering results.

[0031] Step S1242: Randomly select a short video segment that has not been clustered from the association evaluation matrix as the initial cluster center.

[0032] In the association evaluation matrix, all short video clips are in an unclustered state. At this point, one short video clip is randomly selected as the initial cluster center, which will serve as a reference for subsequently determining whether other short video clips belong to this cluster.

[0033] Step S1243: Calculate the overall correlation between the initial cluster center and all other un-clustered short video segments, and include short video segments with an overall correlation higher than the cluster radius into the current content hierarchical cluster.

[0034] Using the initial cluster center as a benchmark, the overall correlation between that cluster center and all other un-clustered short video segments is obtained from the correlation evaluation matrix. Each overall correlation score is compared with the cluster radius. If the overall correlation score is higher than the cluster radius, the corresponding short video segment is included in the currently forming content hierarchical cluster.

[0035] Step S1244: Determine whether the number of short video segments contained in the current content layering cluster has reached the minimum number of segments. If it has, then the content layering cluster is determined to be a valid content layering cluster. If it has not, then abandon the initial clustering center and select a new initial clustering center for clustering.

[0036] After incorporating eligible short video clips into the current content-layered cluster, the number of short video clips contained in the cluster is counted. If the number reaches the preset minimum number of clips, the cluster is considered valid and can be identified as a content-layered cluster. If the number does not reach the minimum, the cluster formed by the initial cluster center does not meet the requirements, and the initial cluster center needs to be abandoned. A new un-clustered short video clip is then selected from the association evaluation matrix as the new initial cluster center, and the above clustering process is repeated.

[0037] Step S1245: Repeat the above process of selecting initial cluster centers and including segments until all short video segments are clustered or labeled as noise segments.

[0038] The process of repeatedly selecting initial cluster centers, calculating the overall correlation, including eligible short video clips, and determining cluster validity is repeated. During this process, short video clips already classified into valid content hierarchical clusters are no longer included in subsequent clustering. This continues until all short video clips are either classified into a valid content hierarchical cluster or marked as noise clips because they fail to meet the clustering criteria.

[0039] Step S1246: Perform secondary clustering on the short video segments marked as noise segments, reduce the clustering radius and perform the clustering operation again. If it is still impossible to form an effective content hierarchical cluster, then the noise segments are individually grouped into one or more small content hierarchical clusters.

[0040] For short video clips marked as noise segments, secondary clustering is required to make full use of resources. During secondary clustering, the cluster radius is appropriately reduced to allow more noise segments to be grouped together. The clustering process is repeated as described above. If a cluster that meets the minimum number of segments can be formed, it is considered a valid content hierarchical cluster. If a valid cluster cannot be formed after secondary clustering, these noise segments are grouped into one or more smaller content hierarchical clusters. These smaller clusters may be subject to special streaming strategies in subsequent processing.

[0041] Step S1247: Deduplication is performed on all the initial content layered clusters. If any short video segment belongs to multiple content layered clusters at the same time, it is assigned to the content layered cluster with the highest correlation based on the comprehensive correlation between the short video segment and the center of each content layered cluster.

[0042] During clustering, a short video clip may be included in multiple content hierarchy clusters simultaneously, necessitating deduplication. For each short video clip belonging to multiple clusters, its overall correlation with the center of each cluster is calculated. Then, the short video clip is assigned to the cluster with the highest overall correlation to ensure that each short video clip belongs to only one content hierarchy cluster.

[0043] Step S125: The real-time delivery parameters of each initial content layer cluster are collaboratively integrated. The common part of the delivery time range of each short video segment is taken as the delivery time of the content layer cluster. The merged set of target audience tags is taken as the target audience tags of the content layer cluster. The average value of the resource allocation ratio is taken as the resource ratio of the content layer cluster, thereby generating the collaborative delivery parameters corresponding to each group of content layer clusters.

[0044] For each initial content-tiered cluster, the real-time streaming parameters of the short video segments it contains need to be collaboratively integrated to determine the collaborative streaming parameters for that cluster. Regarding the streaming time period, the common portion of the streaming time range for all short video segments within the cluster is identified; this common portion constitutes the streaming time period for that content-tiered cluster, ensuring that all short videos within the cluster can be streamed within this time period.

[0045] Regarding target audience tags, the target audience tags of all short video clips within the cluster are merged, duplicate tags are removed, and a set containing all relevant tags is formed as the target audience tag for the content-layered cluster, so that the streaming can cover the target audience of all short videos in the cluster.

[0046] Regarding resource allocation ratios, the average resource allocation ratio of all short video segments within the cluster is calculated and used as the resource allocation ratio for that content-layered cluster, in order to rationally allocate streaming resources. Through the above integration, collaborative streaming parameters corresponding to each content-layered cluster are generated.

[0047] Step S126: Verify the validity of the generated content layered clusters, calculate the average comprehensive correlation within each content layered cluster, and if the average comprehensive correlation is lower than the verification threshold, split the content layered cluster into multiple sub-content layered clusters and re-integrate the collaborative delivery parameters until the average comprehensive correlation of all content layered clusters is higher than the verification threshold.

[0048] Step S1261: Calculate the average of the comprehensive correlation between all short video segments within each content layer cluster, and use it as the average comprehensive correlation of the content layer cluster.

[0049] For each content-layered cluster, obtain the overall correlation score among all short video segments within the cluster. Then, sum these overall correlation scores and divide by the total number of overall correlation scores to obtain the average value, which is the mean overall correlation score of the cluster. This mean overall correlation score reflects the degree of correlation between short video segments within the cluster.

[0050] Step S1262: Compare the average comprehensive correlation score with the preset verification threshold. If the average comprehensive correlation score is higher than or equal to the verification threshold, the content layered cluster is determined to be valid and no adjustment is required. If the average comprehensive correlation score is lower than the verification threshold, the content layered cluster is determined to be invalid and needs to be split.

[0051] A preset verification threshold is set, which is determined based on the platform's requirements for the tightness of association between content layer clusters. The average overall association degree of each content layer cluster is compared with the verification threshold. If the average value is higher than or equal to the threshold, it means that the cluster is tightly associated and is valid, requiring no adjustment; if the average value is lower than the threshold, it means that the cluster is not tightly associated and is invalid, requiring splitting.

[0052] Step S1263: Split the invalid content layered cluster, select the short video segment with the lowest comprehensive correlation in the content layered cluster as the split point, remove the short video segment from the content layered cluster to form a new sub-content layered cluster, and the remaining short video segments form a sub-content layered cluster.

[0053] For ineffective content-layered clusters, splitting is necessary to improve cluster effectiveness. First, identify the short video segment with the lowest overall correlation within the cluster and use it as the splitting point. Remove the short video segment corresponding to this splitting point from the original cluster, forming a new sub-content-layered cluster. The remaining short video segments in the original cluster then form another sub-content-layered cluster.

[0054] Step S1264: Recalculate the average comprehensive correlation of the two sub-content layered clusters after splitting. If the average comprehensive correlation of any sub-content layered cluster is still lower than the verification threshold, repeat the above splitting process until the average comprehensive correlation of all sub-content layered clusters is higher than the verification threshold. Then, re-integrate the collaborative flow parameters of the split sub-content layered clusters to generate the collaborative flow parameters corresponding to each sub-content layered cluster, and record the identifiers of all valid content layered clusters and their corresponding collaborative flow parameters.

[0055] For the two sub-content hierarchical clusters formed after splitting, recalculate their respective average comprehensive correlation scores. If the average score of any sub-cluster is still lower than the verification threshold, continue splitting that sub-cluster using the above splitting method. Repeat this process until the average comprehensive correlation score of all sub-content hierarchical clusters is higher than the verification threshold.

[0056] Next, the collaborative delivery parameters of these effective sub-content tiered clusters were re-integrated to determine the delivery time period, target audience tags, and resource allocation ratio for each sub-cluster. The identifiers of all effective content tiered clusters and their corresponding collaborative delivery parameters were recorded for future use.

[0057] Step S130: Combine the real-time interaction trajectory of users on the short video platform to dynamically adapt and adjust the content layer clusters. By tracking the click behavior sequence, dwell time sequence and forwarding behavior sequence of users on similar content, the association weight of short video segments in each content layer cluster is corrected to generate a dynamic adaptation cluster that adapts to user behavior.

[0058] In e-commerce short video platforms, users' real-time interaction patterns can reflect their interests and preferences for different short videos. Dynamically adapting and adjusting content tiers based on these interaction patterns can make the tiers better suited to users' actual needs, thereby improving ad delivery effectiveness.

[0059] Step S131: Call the platform user behavior database interface to obtain the user's real-time interaction trajectory. The click behavior sequence includes the user's click time and number of clicks on each short video segment. The dwell time sequence includes the start and end times of the user watching each short video segment. The forwarding behavior sequence includes the time and forwarding channel of the user forwarding each short video segment.

[0060] The platform will establish a dedicated user behavior database to store user interaction information with short videos. By calling the database's API, the platform can obtain the user's real-time interaction trajectory. The click behavior sequence records in detail the specific time and number of times a user clicks on each short video segment, allowing the platform to understand which short videos the user is more interested in.

[0061] The dwell time sequence records the moments when users start and end watching each short video segment. These two moments can be used to calculate the user's dwell time, reflecting the duration of the user's interest in the short video content. The forwarding behavior sequence records the moments when users forward each short video segment and the forwarding channels used, such as social media platforms and chat software. This reflects the dissemination effect of the short videos and the degree of user acceptance.

[0062] Step S132: Perform time-series feature analysis on the click behavior sequence, count the click frequency of different short video segments in the same content layer cluster within a unit time, and use the ratio of the click frequency of each short video segment to the total click frequency in the content layer cluster as the initial click weight.

[0063] When performing time-series feature analysis on click behavior sequences, it is necessary to determine an appropriate unit of time, such as per minute or per hour. Within this unit of time, the click frequency of each short video segment in the same content hierarchical cluster is counted, that is, the number of times the user clicks on each short video within this unit of time.

[0064] Then, the total click frequency of all short video segments within the content-layered cluster is calculated. For each short video segment, its click frequency is divided by the total click frequency, and the resulting ratio is the initial click weight of that short video segment. This initial click weight reflects the relative click popularity of the short video by users per unit time.

[0065] Step S133: Perform statistical feature processing on the dwell time sequence, calculate the ratio of the dwell time of each short video segment to the total duration of the short video segment as the dwell ratio, normalize the dwell ratio of all short video segments in the same content layer cluster, and obtain the dwell weight of each short video segment.

[0066] For the dwell time sequence, the dwell time of the user watching each short video segment is first calculated based on the start and end times of viewing. Then, the dwell time is divided by the total duration of the short video segment to obtain the dwell percentage, which reflects the user's completeness of watching the short video content.

[0067] Within the same content-layered cluster, the dwell time percentage of all short video segments is normalized. The purpose of normalization is to eliminate the impact of differences in the total duration of different short video segments, ensuring that dwell time percentages are compared on the same scale. The result is the dwell time weight for each short video segment; a higher dwell time weight indicates that the user is more interested in the content of that short video and watches it more completely.

[0068] Step S134: Perform channel feature analysis on the forwarding behavior sequence, set channel weight coefficients for different channels according to the user coverage scale of the forwarding channels, calculate the sum of the products of the number of forwards of each short video segment in the same content layer cluster and the corresponding channel weight coefficient, and use the ratio of the sum of the products to the total sum of the products in the content layer cluster as the forwarding weight.

[0069] Different forwarding channels have different user reach scales, and channels with larger reach are likely to have better dissemination effects. Therefore, it is necessary to assign a channel weight coefficient to each channel based on its user reach scale. For example, a higher channel weight coefficient should be assigned to a social platform channel that reaches tens of millions of users, while a lower channel weight coefficient should be assigned to a niche community channel that only reaches a few million users. The setting of the channel weight coefficient needs to be based on the platform's analysis of the historical dissemination data of each channel to ensure that it accurately reflects the dissemination effectiveness of different channels.

[0070] Within the same content-layered cluster, for each short video segment, the number of times it is forwarded on different forwarding channels is counted. Then, the number of forwards for each forwarding channel is multiplied by the channel weight coefficient corresponding to that channel to obtain the weighted forwarding volume for that channel. The weighted forwarding volumes of all channels are summed to obtain the sum of the products of the number of forwards for that short video segment and the corresponding channel weight coefficient.

[0071] Simultaneously, the sum of the aforementioned products of all short video segments within the content-layered cluster is calculated, i.e., the total sum of products. For each short video segment, its sum of products is divided by the total sum of products; the resulting ratio is the forwarding weight of that short video segment. This forwarding weight reflects the relative influence of the short video's spread through forwarding channels.

[0072] Step S135: Construct a behavior weight combination model based on the initial click weight, dwell weight, and forwarding weight. The output of the behavior weight combination model is the comprehensive behavior weight of each short video segment. The comprehensive behavior weight is the weighted sum of the initial click weight, dwell weight, and forwarding weight.

[0073] The construction of the behavioral weight combination model requires determining the weight coefficients for initial click weight, dwell time weight, and forwarding weight. These weight coefficients are determined based on the platform's emphasis on different user behaviors. If the platform believes that click behavior better reflects the user's immediate interest, it can assign a higher coefficient to the initial click weight; if it believes that dwell time better reflects the user's deep appreciation of the content, it can increase the dwell time weight coefficient; if it emphasizes the dissemination effect of short videos, it can increase the forwarding weight coefficient.

[0074] The initial click weight, dwell weight, and share weight are multiplied by their respective weight coefficients. These three products are then summed to obtain the overall behavioral weight for each short video segment. This overall behavioral weight comprehensively considers various user interactions and provides a more complete reflection of user preferences for short video segments.

[0075] Step S136: Correct the association weight of short video segments within the content-layered cluster according to the comprehensive behavior weight, and multiply the association weight with the comprehensive behavior weight to obtain the corrected association weight.

[0076] The initial association weights of short video clips within a content-layered cluster are determined based on factors such as the delivery time and target audience tags. To make these association weights more closely reflect users' real-time behavioral preferences, they need to be adjusted using comprehensive behavioral weights.

[0077] The specific operation involves multiplying the initial association weight of each short video segment with its corresponding comprehensive behavior weight. This method increases the association weight of short video segments with high user preference (high comprehensive behavior weight) within the cluster, while correspondingly decreasing the association weight of short video segments with low user preference. This allows the corrected association weight to more accurately reflect the degree of association of short video segments under the current user behavior trend.

[0078] Step S137: Reorder the short video segments in the content layering cluster according to the corrected association weights, take the top K short video segments with the highest weights in the sorting results as core segments, and the rest as auxiliary segments to generate a dynamic adaptation cluster that adapts to user behavior.

[0079] Based on the revised association weights, all short video clips within the content-layered cluster are reordered from highest to lowest. After sorting, the top K short video clips with the highest weights are selected as core clips; these core clips represent the content users are most likely to be interested in. The remaining short video clips serve as auxiliary clips, supplementing and assisting the core clips during their streaming.

[0080] By using the above method, the original content layered cluster is adjusted into a dynamic adaptation cluster that includes core segments and auxiliary segments. This dynamic adaptation cluster can adapt to the user's interactive behavior in real time.

[0081] Step S140: Associate and match the dynamic adaptation clusters with the platform's traffic delivery resource pool, and determine the traffic delivery recommendation priority sequence corresponding to each dynamic adaptation cluster based on the resource allocation ratio and the user response characteristics of the dynamic adaptation clusters.

[0082] In e-commerce short video platforms, the ad placement resource pool contains various resources that can be used to advertise short videos, such as traffic quotas and time slots. Associating and matching dynamically adapted clusters with the ad placement resource pool ensures that ad placement resources are used efficiently, and at the same time, determines ad placement priorities based on the characteristics of each cluster, improving the overall effectiveness of ad placement.

[0083] Step S141: Obtain resource information from the platform's traffic pool. The resource information includes the available traffic amount, traffic distribution characteristics at different times, and traffic coverage capabilities for different user tags.

[0084] The resource information in the platform's traffic pool is the foundation for matching and association. The available traffic quota refers to the total number of exposures or display opportunities the platform can provide within a set time period. The traffic distribution characteristics across different time periods reflect the platform's user activity and available traffic volume during those periods; for example, traffic peaks during certain times and traffic troughs during others.

[0085] The traffic coverage capability of different user tags indicates the number of users the platform can reach under a specific user tag. For example, how large a user base can the platform cover for the "young women" tag? This resource information needs to be updated in real time to ensure the accuracy of the matching.

[0086] Step S142: Analyze the user response characteristics of the dynamic adaptation cluster. The user response characteristics include the historical click rate, historical dwell time, and historical forwarding rate of short video clips within the dynamic adaptation cluster.

[0087] Step S1421: Obtain historical delivery data of each short video segment in the dynamic adaptation cluster from the platform's historical data center. The historical delivery data includes the number of exposures, the number of clicks, the total dwell time, and the number of reposts.

[0088] The platform's historical data center stores all past delivery records for short video clips. Historical delivery data for each short video clip within the dynamic adaptation cluster is extracted from this data. This data details the short video's exposure count (number of times it was seen by users), click count (number of times users clicked to watch), total dwell time (cumulative time all users watched the short video), and forward count (number of times users forwarded the short video to others).

[0089] Step S1422: Calculate the historical click-through rate, which is the ratio of the number of clicks to the number of exposures for each short video segment.

[0090] For each short video clip, the historical click-through rate (CTR) is calculated by dividing its number of clicks by its number of impressions. The historical CTR reflects the short video's ability to attract user clicks in past campaigns; a higher CTR indicates that the short video has garnered more user attention.

[0091] Step S1423: Calculate the historical dwell time, which is the ratio of the total dwell time of each short video segment to the number of clicks.

[0092] Dividing the total dwell time of a short video clip by the number of clicks yields the historical dwell time of that clip. This metric reflects the average time users spend watching short video content after clicking to view it, indicating the duration of user interest in the content. The longer the dwell time, the higher the user's level of approval of the content.

[0093] Step S1424: Calculate the historical forwarding rate, which is the ratio of the number of times each short video segment is forwarded to the number of times it is clicked.

[0094] Dividing the number of times a short video clip is forwarded by the number of clicks gives the historical forwarding rate. The historical forwarding rate reflects users' approval of the short video content and their willingness to share it after watching it; the higher the forwarding rate, the greater the short video's potential for dissemination.

[0095] Step S1425: The historical click-through rate, historical dwell time, and historical forwarding rate of all short video segments within the same dynamic adaptation cluster are averaged to obtain the average historical click-through rate, average historical dwell time, and average historical forwarding rate at the dynamic adaptation cluster level.

[0096] Within the same dynamic adaptation cluster, the historical click-through rates of all short video clips are summed and then divided by the number of short video clips to obtain the average historical click-through rate of the cluster. Using the same method, the average historical dwell time and average historical forwarding rate are calculated. These cluster-level user response characteristics can comprehensively reflect the performance of the dynamic adaptation cluster in historical campaigns.

[0097] Step S1426: Use the average historical click-through rate, average historical dwell time, and average historical forwarding rate as user response features for the dynamic adaptation cluster.

[0098] The average historical click-through rate, average historical dwell time, and average historical forwarding rate obtained from the above calculations together constitute the user response characteristics of the dynamic adaptation cluster. These characteristics are important bases for evaluating the value of cluster traffic delivery.

[0099] Step S1427: Perform outlier detection on the user response features. If the difference between any feature value and the corresponding feature value of other segments in the dynamic adaptation cluster is greater than the set difference threshold, then the outlier value is corrected by replacing the outlier value with the median of the feature values ​​of other segments in the dynamic adaptation cluster.

[0100] During the calculation of user response features, there may be instances where the feature values ​​of individual short video segments differ significantly from those of other segments within the cluster. These outliers can affect the accuracy of the overall cluster features. Therefore, outlier detection is necessary.

[0101] A difference threshold is set, which is determined based on the platform's tolerance for data fluctuations. For each feature value, it is compared with the corresponding feature value of other segments within the cluster. If the difference is greater than the set difference threshold, the feature value is determined to be an outlier.

[0102] The correction of outliers is replaced by the median of the corresponding feature values ​​of other segments within the cluster. The median is more resistant to the influence of extreme values ​​than the mean, and can ensure that the corrected feature values ​​are more consistent with the overall feature level of the cluster.

[0103] Step S143: The matching degree of the collaborative traffic delivery parameters of the dynamic adaptation cluster and the resource information of the traffic delivery resource pool is calculated. The matching degree includes the matching degree of the delivery time period and traffic distribution characteristics, the matching degree of the target audience tags and traffic coverage capabilities, and the matching degree of the resource allocation ratio and the available traffic amount.

[0104] When calculating the matching degree between the delivery time period and traffic distribution characteristics, we first analyze the delivery time period of the dynamic adaptation cluster and the traffic distribution characteristics of the traffic resource pool during that time period. If the traffic distribution of the resource pool is at a peak during the delivery time period, it indicates a high degree of matching; if the traffic distribution is at a trough during the delivery time period, the degree of matching is low. This degree of matching can be quantified by calculating the ratio of the traffic share during the delivery time period to the share of that time period in the total delivery time.

[0105] When calculating the matching degree between target audience tags and traffic coverage capabilities, the target audience tags of the dynamic adaptation cluster and the traffic coverage capability of the resource pool for that tag are compared. If the resource pool has a large coverage of the target audience tag and a high degree of overlap with the cluster's target audience, the matching degree is high. This can be measured by calculating the ratio of the number of overlapping target audiences to the total number of target audiences in the cluster.

[0106] When calculating the matching degree between the resource allocation ratio and the available traffic quota, the traffic demand corresponding to the resource allocation ratio of the dynamically adapted cluster is compared with the available traffic quota of the traffic delivery resource pool. If the traffic demand is within the available traffic quota and the ratio is reasonable, the matching degree is high. This can be determined by calculating the ratio of the traffic corresponding to the resource allocation ratio to the available traffic quota.

[0107] Step S144: Construct a traffic evaluation model based on matching degree and user response characteristics. The output of the traffic evaluation model is the traffic evaluation score of each dynamic adaptation cluster. The traffic evaluation score is a weighted combination of matching degree and user response characteristics.

[0108] The construction of a traffic evaluation model requires assigning weights to various indicators in matching degree and user response characteristics. The weights are determined based on the platform's traffic goals. If the platform focuses more on the efficiency of traffic resource utilization, the weight of matching degree can be increased; if it focuses more on user feedback, the weight of user response characteristics can be increased.

[0109] The matching degree between the ad placement time and traffic distribution characteristics, the matching degree between the target audience tags and traffic coverage, and the matching degree between the resource allocation ratio and the available ad placement amount are each multiplied by their respective weights. Then, the average historical click-through rate, average historical dwell time, and average historical forwarding rate are multiplied by their respective weights. Finally, all these weighted results are summed to obtain the ad placement value score for each dynamic adaptation cluster. This ad placement value score comprehensively reflects the ad placement potential and value of the cluster.

[0110] Step S145: Sort each dynamic adaptation cluster according to the traffic value score to generate a preliminary traffic recommendation priority sequence.

[0111] All dynamically adapted clusters are sorted from highest to lowest according to their traffic value scores. The sorting result is the initial traffic recommendation priority sequence, with clusters having higher initial scores having higher traffic recommendation priority.

[0112] Step S146: Adjust the initial traffic recommendation priority sequence. Considering the resource competition relationship between dynamic adaptation clusters, if the overlap ratio between the delivery time periods of two dynamic adaptation clusters is greater than the set overlap ratio and the overlap ratio of the target audience tags exceeds the preset dimension threshold, then reduce the priority of the dynamic adaptation cluster with the lowest traffic value score.

[0113] There may be resource competition between dynamically adapted clusters, especially clusters with overlapping delivery times and target audience tags. To address this, we can set an overlap ratio and a preset dimension threshold. The overlap ratio measures the degree of overlap between the delivery times of two clusters, while the preset dimension threshold measures the percentage of overlapping target audience tag dimensions.

[0114] For clusters in the initial ranking, check the overlap ratio of their ad placement periods and the percentage of overlapping target audience tags between any two clusters. If both exceed the set values, it indicates strong resource competition between the two clusters. To avoid resource waste and mutual negative impact on ad performance, reduce the priority of clusters with lower ad performance value scores.

[0115] Step S147: The adjusted traffic recommendation priority sequence is correlated and verified with the resource allocation ratio. If the recommended traffic ratio corresponding to the priority of any dynamic adaptation cluster exceeds its resource allocation ratio, the priority of that dynamic adaptation cluster is lowered until the recommended traffic ratio of all dynamic adaptation clusters does not exceed its resource allocation ratio. Finally, the traffic recommendation priority sequence corresponding to each dynamic adaptation cluster is determined.

[0116] Each dynamically adapted cluster has a corresponding resource allocation percentage, which defines the maximum proportion of total traffic resources it can occupy. The recommended traffic percentage corresponding to the priority of each cluster in the adjusted traffic recommendation priority sequence is compared with the resource allocation percentage of that cluster.

[0117] If the recommended traffic percentage exceeds the resource allocation percentage, it indicates that the cluster's resource demand exceeds the preset allocation, and its priority needs to be lowered to reduce the recommended traffic percentage. Repeat this process until the recommended traffic percentage of all dynamically adapted clusters is within its resource allocation percentage range. The resulting sequence is the final determined traffic delivery recommendation priority sequence.

[0118] Step S150: Based on the traffic recommendation priority sequence and the collaborative traffic recommendation parameters of each content layer cluster, generate a traffic recommendation instruction and send the traffic recommendation instruction to the platform traffic scheduling module.

[0119] In e-commerce short video platforms, the ad delivery recommendation instruction is crucial information guiding the platform's actual ad delivery operations. Generating this instruction based on the previously determined ad delivery recommendation priority sequence and collaborative ad delivery parameters ensures that ad delivery operations are carried out in an orderly and precise manner.

[0120] Step S151: Analyze the priority sequence of traffic recommendations and determine the delivery order of each dynamic adaptation cluster according to the order of priority from high to low.

[0121] The priority sequence for streaming recommendations is analyzed to determine the position of each dynamic adaptation cluster within the sequence. The clusters are then arranged in descending order of priority; this arrangement determines the streaming order for each dynamic adaptation cluster. During actual streaming, short video clips within each dynamic adaptation cluster are delivered sequentially according to this order, ensuring that higher-priority clusters receive streaming resources first.

[0122] Step S152: Based on the content layering cluster's delivery time period in the collaborative delivery parameters of each content layering cluster, the delivery time period is subdivided into multiple consecutive delivery sub-time periods. The duration of each delivery sub-time period is determined based on the resource allocation ratio and traffic distribution characteristics of the implemented content layering cluster.

[0123] Step S1521: Extract the start and end times of the content layered cluster delivery time period and calculate the total duration of the delivery time period.

[0124] The start and end times of the delivery period are obtained from the collaborative delivery parameters of the content layering cluster. The total duration of the delivery period is calculated by the difference between these two times. The total duration is the basis for segmenting the delivery period.

[0125] Step S1522: Obtain the traffic distribution characteristics of the platform's traffic resource pool during the content tiered cluster delivery period. The traffic distribution characteristics are the traffic proportions in different time periods.

[0126] The platform's traffic pool records the traffic percentage at different times within the content tiered cluster delivery period, such as the percentage of traffic per hour relative to the total traffic during the delivery period. These traffic distribution characteristics reflect the traffic fluctuations within that time period.

[0127] Step S1523: Determine the traffic allocation ratio for each delivery sub-period based on the resource allocation ratio and traffic distribution characteristics of the content layered cluster.

[0128] By combining the resource allocation ratio of the content-layered cluster (i.e., the proportion of the cluster in the total traffic resources) with the traffic distribution characteristics of the traffic resource pool, the traffic allocation ratio for each sub-period of the campaign is determined. During periods with a higher traffic share, the traffic allocation ratio for that sub-period can be appropriately increased to fully utilize peak traffic; during periods with a lower traffic share, the traffic allocation ratio should be reduced accordingly to avoid resource waste.

[0129] Step S1524: Allocate the total duration of the campaign period to each campaign sub-period according to the traffic allocation ratio, and associate and store the subdivided campaign sub-periods with the corresponding traffic allocation ratio.

[0130] The total duration of each campaign segment is allocated based on the traffic distribution ratio for each segment. For example, if a segment has a certain traffic distribution ratio and a certain total duration, the duration of that segment is the total duration multiplied by that ratio. The start and end times of each segment, along with its corresponding traffic distribution ratio, are stored together for easy execution of subsequent campaign traffic operations.

[0131] Step S153: Extract the target audience tags of each content layer cluster from the collaborative delivery parameters of each content layer cluster, and convert the target audience tags into audience targeting codes. The audience targeting codes include interest category codes, region codes, and device type codes.

[0132] The target audience tags for content-layered clusters contain various information describing user characteristics. These tags need to be converted into audience targeting codes that the platform's traffic delivery system can recognize. Interest category codes correspond to users' interests and hobbies, such as clothing interests, digital product interests, etc., and each interest category has a unique code identifier.

[0133] The region code is based on the user's location; different provinces and cities have corresponding codes. The device type code distinguishes the user's device, such as mobile phone, tablet, computer, etc., each with its own specific code. Through this conversion, the target audience tags can be accurately interpreted by the targeting system, enabling precise audience targeting.

[0134] Step S154: Integrate the delivery order, time segmentation results, and audience targeting codes to generate the main content of the traffic recommendation instruction. The main content also includes the short video segment identifiers and recommended traffic percentages corresponding to each content layer cluster.

[0135] The delivery order, detailed information for each delivery sub-period (start time, end time, traffic allocation ratio), and audience targeting codes are integrated to form the main content of the delivery recommendation instruction. Simultaneously, short video clip identifiers corresponding to each content layer cluster are added to the main content to clearly identify the specific content being delivered; a recommended traffic ratio is also included to specify the proportion of traffic each cluster can obtain during the delivery process.

[0136] The integration of this information ensures that the recommended traffic delivery instructions contain all the necessary traffic delivery details and can comprehensively guide traffic delivery operations.

[0137] Step S155: Send the flow recommendation instruction to the flow scheduling module, and record the sending time and instruction identifier of the flow recommendation instruction to track the execution status of the flow recommendation instruction.

[0138] The generated traffic recommendation instruction is sent to the traffic scheduling module via the platform's internal communication interface. The traffic scheduling module will then perform specific traffic allocation arrangements and execution based on this instruction.

[0139] Simultaneously with sending a command, the sending time and a unique command identifier are recorded. This command identifier can be used to query the execution status of the command later. The command identifier is automatically generated by the platform system and includes a timestamp of command generation, a content layer cluster identifier fragment, and a random checksum, ensuring that each command identifier is unique within the platform.

[0140] The platform can establish an instruction tracking database specifically for storing tracking information related to traffic recommendation instructions. Once a traffic recommendation instruction is sent, the instruction identifier, sending time, corresponding dynamic adaptation cluster information, and collaborative traffic recommendation parameters can be synchronously written into this database.

[0141] During the execution of streaming recommendation instructions, the streaming scheduling module can provide real-time feedback on the execution progress to the instruction tracking database. For example, when a streaming task for a certain time period begins, the start time can be recorded and associated with the corresponding instruction identifier; when the streaming task for that time period ends, information such as the end time, the actual number of short video clips delivered, and the actual amount of data used can be recorded.

[0142] For user interaction data generated during the campaign, such as new clicks, changes in dwell time, and sharing activity, the campaign scheduling module will periodically summarize and associate them with instruction identifiers and store them in the database.

[0143] Platform administrators can enter a command identifier through the command query interface to retrieve all corresponding tracking information from the command tracking database. The information is displayed in a timeline format, showing the command's sending time, execution status at each stage, user interaction data, and other content, providing a clear view of the command's entire lifecycle.

[0144] If an interruption, delay, or other abnormal situation occurs during the execution of a streaming recommendation command, the streaming scheduling module will immediately generate an exception log, recording in detail the time of the exception, the type of exception, and the possible cause. The exception log will then be associated with the corresponding command identifier and stored in the database. Simultaneously, an exception alert can be automatically sent to pre-set management personnel terminals. The alert will include the command identifier and a brief description of the exception, allowing management personnel to intervene promptly.

[0145] After the instruction is executed, an execution summary report can be automatically generated based on the records in the instruction tracking database. The report will compare the planned deployment parameters in the instruction with the actual execution parameters, such as the deviation between the planned and actual deployment periods, and the difference between the planned traffic quota and the actual usage quota.

[0146] To ensure the security and traceability of instruction tracking data, the database employs a scheduled backup mechanism, backing up daily tracking data to an off-site storage server. Simultaneously, strict access control is implemented for the database; only authorized personnel can perform operations such as querying and exporting, and all operations are recorded in the operation log for easy auditing and tracing.

[0147] Figure 2 The illustration shows exemplary hardware and software components of a short video platform streaming short video content analysis and recommendation system 100 that can implement the ideas of this application, according to some embodiments of this application. For example, processor 120 can be used in the short video platform streaming short video content analysis and recommendation system 100 and to perform the functions in this application.

[0148] The short video platform content analysis and recommendation system 100 can be a general-purpose server or a special-purpose server; both can be used to implement the short video platform content analysis and recommendation method of this application. Although only one server is shown in this application, for convenience, the functions described in this application can be implemented in a distributed manner on multiple similar platforms to balance the load.

[0149] For example, the short video platform content analysis and recommendation system 100 may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and various forms of storage media 140, such as a disk, ROM, or RAM, or any combination thereof. Exemplarily, the short video platform content analysis and recommendation system 100 may also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The methods of this application can be implemented according to these program instructions. The short video platform content analysis and recommendation system 100 also includes an I / O interface 150 between the computer and other input / output devices.

[0150] For ease of explanation, only one processor is described in the short video platform content analysis and recommendation system 100. However, it should be noted that the short video platform content analysis and recommendation system 100 in this application may also include multiple processors. Therefore, the steps performed by one processor described in this application may also be performed jointly or individually by multiple processors. For example, if the processor of the short video platform content analysis and recommendation system 100 performs steps A and B, it should be understood that steps A and B may also be performed jointly by two different processors or individually by one processor. For example, the first processor performs step A, the second processor performs step B, or the first processor and the second processor jointly perform steps A and B.

[0151] Furthermore, this embodiment of the invention also provides a readable storage medium, which has computer-executable instructions pre-set in it. When the processor executes the computer-executable instructions, the above-mentioned method for analyzing and recommending short video content streamed on short video platforms is implemented.

[0152] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A method for analyzing and recommending short video content delivered to a short video platform, characterized in that, The method includes: Obtain the set of short video content to be streamed and the corresponding real-time streaming parameters. The set of short video content to be streamed includes multiple short video clips and content metadata associated with each short video clip. The real-time streaming parameters include the time period range, target audience tags, and resource allocation ratio. Based on the real-time streaming parameters, the set of short video content to be streamed is processed in a hierarchical manner. According to the overlapping characteristics of the time period range and the association characteristics of the target audience tags, the set of short video content to be streamed is divided into multiple content hierarchical clusters, and each content hierarchical cluster corresponds to a set of collaborative streaming parameters. By combining the real-time interaction trajectory of users on short video platforms, the content layer clusters are dynamically adapted and adjusted. By tracking the click behavior sequence, dwell time sequence and forwarding behavior sequence of users on similar content, the association weight of short video segments in each content layer cluster is corrected, and a dynamic adaptation cluster adapted to user behavior is generated. The dynamic adaptation clusters are associated and matched with the platform's traffic delivery resource pool. Based on the resource allocation ratio and the user response characteristics of the dynamic adaptation clusters, the traffic delivery recommendation priority sequence corresponding to each dynamic adaptation cluster is determined. Based on the priority sequence of traffic recommendations and the collaborative traffic recommendation parameters of each content layer cluster, a traffic recommendation instruction is generated and sent to the platform's traffic scheduling module.

2. The method for analyzing and recommending short video content on a short video platform according to claim 1, characterized in that, The process involves performing hierarchical association processing on the set of short video content to be streamed based on the real-time streaming parameters. According to the overlapping characteristics of the streaming time periods and the association characteristics of the target audience tags, the set of short video content to be streamed is divided into multiple content hierarchical clusters, including: Extract the delivery time range from the real-time delivery parameters, decompose the delivery time range of each short video segment into multiple consecutive time sub-intervals, and calculate the proportion of the number of overlaps between any two short video segments in the time sub-intervals to the total number of sub-intervals, as the delivery time overlap feature; The target audience tags are analyzed, and each tag is broken down into multiple tag dimensions. Each tag dimension contains multiple tag values. The proportion of the number of tag values ​​that different short video clips share on the same tag dimension is calculated to the total number of tag values ​​on that tag dimension, which is used as the association feature of the target audience tags. Based on the overlapping characteristics of the delivery time period and the correlation characteristics of the target audience tags, a correlation evaluation matrix is ​​constructed. Each element in the matrix corresponds to the comprehensive correlation degree of two short video segments. The comprehensive correlation degree is the weighted combination result of the overlapping characteristics of the delivery time period and the correlation characteristics of the target audience tags. A density-based clustering algorithm is used to cluster the association evaluation matrix, and short video clips with a comprehensive association degree higher than the clustering threshold are grouped into the same content hierarchical cluster, forming multiple initial content hierarchical clusters. The real-time delivery parameters of each initial content layer cluster are collaboratively integrated. The common part of the delivery time range of each short video segment is taken as the delivery time of the content layer cluster. The merged set of target audience tags is taken as the target audience tags of the content layer cluster. The average value of the resource allocation ratio is taken as the resource ratio of the content layer cluster, thereby generating the collaborative delivery parameters corresponding to each group of content layer clusters. The generated content layered clusters are validated. The average comprehensive correlation score within each content layered cluster is calculated. If the average comprehensive correlation score is lower than the validation threshold, the content layered cluster is split into multiple sub-content layered clusters, and the collaborative delivery parameters are re-integrated until the average comprehensive correlation score of all content layered clusters is higher than the validation threshold.

3. The method for analyzing and recommending short video content on a short video platform according to claim 2, characterized in that, The density-based clustering algorithm is used to cluster the association evaluation matrix, grouping short video clips with a comprehensive association degree higher than the clustering threshold into the same content hierarchical cluster, forming multiple initial content hierarchical clusters, including: Initialize clustering parameters, set cluster radius and minimum number of included segments. The cluster radius is a comprehensive correlation threshold for determining whether two short video segments can be classified into the same content hierarchical cluster. The minimum number of included segments is the minimum number of short video segments that a content hierarchical cluster should contain. Randomly select an unclustered short video segment from the association evaluation matrix as the initial cluster center; Calculate the overall correlation between the initial cluster center and all other unclustered short video segments, and include short video segments with an overall correlation higher than the cluster radius into the current content hierarchical cluster; Determine whether the number of short video segments contained in the current content layer cluster has reached the minimum number of segments. If it has, the content layer cluster is determined to be a valid content layer cluster. If it has not, the initial cluster center is abandoned and a new initial cluster center is selected for clustering. Repeat the above process of selecting initial cluster centers and including segments until all short video segments are clustered or labeled as noise segments; The short video segments marked as noise segments are clustered a second time, and the clustering radius is reduced and the clustering operation is repeated. If an effective content hierarchical cluster still cannot be formed, the noise segments are individually grouped into one or more small content hierarchical clusters. All initial content layered clusters are deduplicated. If any short video clip belongs to multiple content layered clusters, it is assigned to the content layered cluster with the highest correlation based on the comprehensive correlation between the short video clip and the center of each content layered cluster.

4. The method for analyzing and recommending short video content on a short video platform according to claim 2, characterized in that, The validity of the generated content-layered clusters is verified by calculating the average comprehensive correlation score within each content-layered cluster. If the average comprehensive correlation score is lower than the verification threshold, the content-layered cluster is split into multiple sub-content-layered clusters, and the collaborative delivery parameters are re-integrated until the average comprehensive correlation score of all content-layered clusters is higher than the verification threshold. This includes: Calculate the average of the overall correlation between all short video segments within each content layer cluster, and use it as the average overall correlation of that content layer cluster; The average comprehensive correlation score is compared with the preset verification threshold. If the average comprehensive correlation score is higher than or equal to the verification threshold, the content layered cluster is determined to be valid and no adjustment is required. If the average comprehensive correlation score is lower than the verification threshold, the content layered cluster is determined to be invalid and needs to be split. The invalid content layered cluster is split, and the short video segment with the lowest comprehensive correlation within the content layered cluster is selected as the split point. This short video segment is removed from the content layered cluster to form a new sub-content layered cluster. The remaining short video segments form a sub-content layered cluster. Recalculate the average comprehensive correlation of the two sub-content layered clusters after splitting. If the average comprehensive correlation of any sub-content layered cluster is still lower than the verification threshold, repeat the above splitting process until the average comprehensive correlation of all sub-content layered clusters is higher than the verification threshold. Then, re-integrate the collaborative flow parameters of the split sub-content layered clusters to generate the collaborative flow parameters corresponding to each sub-content layered cluster, and record the identifiers of all valid content layered clusters and their corresponding collaborative flow parameters.

5. The method for analyzing and recommending short video content on a short video platform according to claim 1, characterized in that, The method involves dynamically adapting and adjusting the content layering clusters based on real-time user interaction patterns on short video platforms. This is achieved by tracking user click sequences, dwell time sequences, and forwarding sequences for similar content, correcting the association weights of short video segments within each content layering cluster, and generating dynamically adapted clusters that match user behavior. The platform's user behavior database interface is called to obtain the user's real-time interaction trajectory. The click behavior sequence includes the user's click time and number of clicks on each short video segment. The dwell time sequence includes the start and end times of the user watching each short video segment. The forwarding behavior sequence includes the time and forwarding channel of the user forwarding each short video segment. The click behavior sequence is subjected to temporal feature analysis. The click frequency of different short video segments in the same content layer cluster is counted within a unit time. The ratio of the click frequency of each short video segment to the total click frequency in the content layer cluster is used as the initial click weight. The dwell time sequence is subjected to statistical feature processing. The ratio of the dwell time of each short video segment to the total duration of the short video segment is calculated as the dwell ratio. The dwell ratios of all short video segments in the same content layer cluster are normalized to obtain the dwell weight of each short video segment. Channel feature analysis is performed on the forwarding behavior sequence. Channel weight coefficients are set for different channels according to the user coverage scale of the forwarding channels. The sum of the products of the number of forwards of each short video segment in the same content layer cluster and the corresponding channel weight coefficient is calculated. The ratio of the sum of the products to the total sum of the products in the content layer cluster is used as the forwarding weight. A behavior weight combination model is constructed based on the initial click weight, dwell weight, and forwarding weight. The output of the behavior weight combination model is the comprehensive behavior weight of each short video segment. The comprehensive behavior weight is the weighted sum of the initial click weight, dwell weight, and forwarding weight. The association weights of short video segments within the content-layered cluster are corrected based on the comprehensive behavioral weights. The association weights are then multiplied by the comprehensive behavioral weights to obtain the corrected association weights. Based on the corrected association weights, the short video segments within the content-layered cluster are reordered. The top K short video segments with the highest weights in the sorting results are used as core segments, and the rest are used as auxiliary segments, generating a dynamic adaptation cluster that adapts to user behavior.

6. The method for analyzing and recommending short video content on a short video platform according to claim 1, characterized in that, The step of associating and matching the dynamic adaptation clusters with the platform's traffic delivery resource pool, and determining the traffic delivery recommendation priority sequence corresponding to each dynamic adaptation cluster based on the resource allocation ratio and user response characteristics of the dynamic adaptation clusters, includes: Obtain resource information from the platform's traffic pool, including the available traffic amount, traffic distribution characteristics at different times, and traffic coverage capabilities for different user tags. Analyze the user response characteristics of the dynamic adaptation cluster, which include the historical click rate, historical dwell time, and historical forwarding rate of short video clips within the dynamic adaptation cluster; The matching degree is calculated by matching the collaborative traffic delivery parameters of the dynamic adaptation cluster with the resource information of the traffic delivery resource pool. The matching degree includes the matching degree of the delivery time period and traffic distribution characteristics, the matching degree of the target audience tags and traffic coverage capacity, and the matching degree of the resource allocation ratio and the available traffic amount. A traffic evaluation model is constructed based on the matching degree and user response characteristics. The output of the traffic evaluation model is the traffic evaluation value score of each dynamic adaptation cluster. The traffic evaluation value score is a weighted combination of the matching degree and user response characteristics. Based on the traffic delivery value score, each dynamically adapted cluster is sorted to generate a preliminary traffic delivery recommendation priority sequence; The initial priority sequence of ad delivery recommendations is adjusted. Taking into account the resource competition relationship between dynamic adaptation clusters, if the overlap ratio between the ad delivery time periods of two dynamic adaptation clusters is greater than the set overlap ratio and the overlap ratio of the target audience tags exceeds the preset dimension threshold, the priority of the dynamic adaptation cluster with the lowest ad delivery value score is reduced. The adjusted traffic recommendation priority sequence is correlated and verified with the resource allocation ratio. If the recommended traffic ratio corresponding to the priority of any dynamic adaptation cluster exceeds its resource allocation ratio, the priority of that dynamic adaptation cluster is lowered until the recommended traffic ratio of all dynamic adaptation clusters does not exceed their resource allocation ratio, and finally the traffic recommendation priority sequence corresponding to each dynamic adaptation cluster is determined.

7. The method for analyzing and recommending short video content on a short video platform according to claim 6, characterized in that, The analysis of the user response characteristics of the dynamically adapted cluster includes: The platform’s historical data center obtains historical delivery data for each short video segment within the dynamic adaptation cluster. The historical delivery data includes exposure counts, click counts, total dwell time, and forward counts. Calculate the historical click-through rate (CTR), which is the ratio of the number of clicks to the number of impressions for each short video segment. Calculate the historical dwell time, which is the ratio of the total dwell time of each short video segment to the number of clicks; Calculate the historical forwarding rate, which is the ratio of the number of times each short video clip is forwarded to the number of clicks; The historical click-through rate, historical dwell time, and historical forwarding rate of all short video clips within the same dynamic adaptation cluster are averaged to obtain the average historical click-through rate, average historical dwell time, and average historical forwarding rate at the dynamic adaptation cluster level. Average historical click-through rate, average historical dwell time, and average historical forwarding rate are used as user response characteristics for dynamically adaptable clusters. Anomaly detection is performed on user response features. If the difference between any feature value and the corresponding feature value of other segments in the dynamic adaptation cluster is greater than a set difference threshold, the anomaly value is corrected by replacing it with the median of the feature values ​​of other segments in the dynamic adaptation cluster.

8. The method for analyzing and recommending short video content on a short video platform according to claim 1, characterized in that, The process of generating a traffic recommendation instruction based on the traffic recommendation priority sequence and the collaborative traffic recommendation parameters of each content layer cluster, and sending the traffic recommendation instruction to the platform traffic scheduling module, includes: The delivery recommendation priority sequence is analyzed, and the delivery order of each dynamic adaptation cluster is determined according to the order of priority from high to low. Based on the content layer cluster's delivery time period in the collaborative delivery parameters of each content layer cluster, the delivery time period is subdivided into multiple consecutive delivery sub-time periods. The duration of each delivery sub-time period is determined based on the resource allocation ratio and traffic distribution characteristics of the implemented content layer cluster. Extract the target audience tags of each content layer cluster from the collaborative delivery parameters of each content layer cluster, and convert the target audience tags into audience targeting codes. The audience targeting codes include interest category codes, region codes, and device type codes. The main content of the traffic recommendation instruction is generated by integrating the order of delivery, the results of time segmentation, and the audience targeting code. The main content also includes the short video segment identifiers and the recommended traffic proportions corresponding to each content layer cluster. The flow recommendation instruction is sent to the flow scheduling module, and the sending time and instruction identifier of the flow recommendation instruction are recorded to track the execution status of the flow recommendation instruction.

9. The method for analyzing and recommending short video content on a short video platform according to claim 8, characterized in that, The delivery time period is subdivided into multiple consecutive delivery sub-time periods based on the delivery time period of the content layer cluster in the collaborative delivery parameters of each content layer cluster, including: Extract the start and end times of the content layered cluster delivery time period, and calculate the total duration of the delivery time period; Obtain the traffic distribution characteristics of the platform's traffic resource pool during the content-layered cluster delivery period, wherein the traffic distribution characteristics are the traffic proportions in different time periods; Based on the resource allocation ratio and traffic distribution characteristics of the content-layered cluster, determine the traffic allocation ratio for each delivery sub-period. The total duration of the campaign period is allocated to each sub-campaign period according to the traffic allocation ratio, and the sub-campaign periods are associated with and stored with the corresponding traffic allocation ratio.

10. A short video platform content analysis and recommendation system, characterized in that, The short video platform content analysis and recommendation system includes a processor and a memory, the memory and the processor are connected, the memory is used to store programs, instructions or code, and the processor is used to execute the programs, instructions or code in the memory to implement the short video platform content analysis and recommendation method according to any one of claims 1-9.

Images