Content seeding strategy generation method based on competitor data collection

Abstract

The application belongs to the technical field of data processing and content operation, and relates to a content seeding strategy generation method based on competitive product data collection; by acquiring content publishing data and user interactive behavior data of a competitive product on multiple platforms, a competitive product content sequence set of unified time sequence is constructed, and a competitive product propagation feature set is formed by extracting propagation growth rate, interactive turning point and theme aggregation features; on this basis, an association relationship chain between contents is established, a propagation path probability sequence is calculated, a time sequence reconstruction of content performance is performed, a coupling relationship between user interest change and content performance is obtained, and a seeding trend prediction curve is generated; further, a nonlinear combination of content element contribution weight and propagation path probability is performed to form a strategy priority distribution diagram, key indicators are extracted and matched with rules, and finally a content seeding strategy scheme is output; the application can improve the description accuracy of the content propagation process and enhance the time sequence adaptation ability and stability of the strategy generation.

Description

Technical Field

[0001] This invention relates to the field of data processing and content operation technology, specifically to a method for generating content seeding strategies based on competitor data collection. Background Technology

[0002] In content seeding scenarios, existing strategies mostly rely on historical data within the platform or coarse-grained competitor analysis, which makes it difficult to cope with the short-term hot topic shifts and sudden changes in user interests in subdivided categories. Especially in cold start or long-tail product promotion, the distribution of competitor content is highly discrete and has low reusability, resulting in time lag and unstable conversion in strategy generation. It is difficult to accurately depict the impact of competitor content on the user's decision-making path, thus causing waste of resources and decreased seeding efficiency. Summary of the Invention

[0003] The purpose of this invention is to provide a content seeding strategy generation method based on competitor data collection, in order to address the shortcomings of the prior art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a method for generating content seeding strategies based on competitor data collection, comprising: Obtain competitor content publishing data and user interaction behavior data on multiple platforms, clean and time-align the data to form a competitor content sequence set that includes publishing time, content type and feedback intensity; Based on the content sequence set of competitors, the dissemination growth rate, interaction turning points and theme clustering characteristics in different time windows are extracted to construct a set of competitor dissemination characteristics. Based on the aforementioned set of competitor propagation characteristics, key influencing nodes are identified and a chain of relationships between content is established. The propagation continuation probability of adjacent content nodes is calculated to form a propagation path probability sequence. Based on the probability sequence of the propagation path, the performance trend of competitor content is reconstructed in time series to generate the coupling relationship between the trajectory of user interest change and the content performance change, and thus obtain the seeding trend prediction curve. Based on the aforementioned trend prediction curve, the contribution weights of different content elements are calculated and nonlinearly combined with the probabilities of the corresponding propagation paths to generate a strategy priority distribution map. Multidimensional feature extraction is performed on the strategy priority distribution map to extract key indicators including content format preference, release time sensitivity and interaction trigger threshold, and then matched with a preset rule base. When any key metric reaches the preset trigger condition, a content seeding strategy plan including content theme combination, release rhythm and channel configuration will be output.

[0005] Preferably, the competitive product content sequence set includes: Collect original publishing data and user interaction behavior data of competitors on multiple content platforms, and generate globally unique content identifiers based on platform identifiers and original content IDs; The original published data and user interaction data are processed to remove outliers, fill in missing values, and align time at the second level to obtain data records arranged according to a unified time base. The number of views, likes, comments, and shares are weighted and summed according to preset weights to obtain the feedback intensity, and then a sequence of competitor content is formed according to the publication time.

[0006] Preferably, constructing a set of competitor communication features includes: The content sequence set of competitors is segmented according to the sliding time window, and the propagation growth rate sequence is obtained based on the difference in the mean of window feedback between adjacent time windows. The interaction inflection point is determined based on the position of the positive and negative changes in the propagation growth sequence and the magnitude of the changes in the feedback intensity before and after it. The changes in the proportion of content types within each time window are statistically analyzed. When the proportion of the same content type increases by more than 0.1% in at least three consecutive time windows, the theme clustering characteristics are determined and aligned with the dissemination growth rate sequence and interaction inflection points to form a set of competitor dissemination characteristics.

[0007] Preferably, key influencing nodes include: Traverse the set of competitor propagation features and filter candidate peak positions where the propagation growth rate is greater than both the previous and next time markers; Calculate the time interval between the candidate peak position and the most recent interaction inflection point. When the time interval does not exceed 0.3 times the length of the adjacent time window, mark the corresponding content as a key influencing node. When multiple candidate peak positions correspond to the same interactive inflection point, the content with the larger propagation growth rate is selected as the key influencing node. When the propagation growth rates are the same, the content with the larger feedback intensity is selected as the key influencing node.

[0008] Preferably, establishing a chain of relationships includes: Starting with the key impact node, search for subsequent content nodes in chronological order within the next 6 to 8 time windows. When adjacent content nodes maintain the same theme clustering feature for no less than 2 consecutive time windows, and the corresponding content type accounts for no less than 0.2%, it is determined that there is a valid association between the two nodes. Subsequent content nodes that meet the criteria for effective association will be connected to key influential nodes in order of time distance from nearest to farthest, forming a chain of associations. When the time distance is the same, content nodes with higher feedback intensity will be prioritized for connection.

[0009] Preferably, the probability sequence for forming a propagation path includes: For each pair of adjacent content nodes in the association chain, the ratio between the feedback strength of the later node and the feedback strength of the previous node is used as the base value. The base value is corrected based on the consistency of the direction of propagation growth rate change between the two nodes to obtain the propagation continuation probability; The cumulative path value is obtained by accumulating the propagation continuation probabilities of each path within the same association chain, and then normalizing the cumulative path value to form a propagation path probability sequence.

[0010] Preferably, the grass planting trend prediction curve includes: The path segments are unfolded in chronological order according to the probability sequence of the propagation path. The feedback intensity of the subsequent node is multiplied by the probability of propagation continuation and then superimposed with the feedback intensity of the preceding node to obtain the reconstructed feedback value. This value is then combined with the path probability to form a reconstructed sequence of content presentation. Based on the content performance, the reconstruction sequence calculates the slope and rate of change of the slope within a continuous time window. When the rate of change of the slope changes in the same direction within at least three adjacent time windows, the interest evolution interval is determined and the user interest change trajectory is extracted. The user interest change trajectory is time-aligned with the content performance reconstruction sequence, and a coupling relationship sequence is formed based on the ratio of the change magnitude of the two. After applying continuity constraints to the coupling relationship sequence, the trend prediction curve of product seeding is obtained.

[0011] Preferably, the generated strategy priority distribution map includes: Based on the direction of change of the coupled values ​​of adjacent trend segments in the grass planting trend prediction curve, the trend segments are divided into an upward trend range, a stable range, and a downward trend range. The frequency of occurrence of content elements and the proportion of feedback intensity are statistically analyzed in each interval, and the two are weighted and accumulated to obtain the initial contribution value. Then, the content elements are assigned an enhancement coefficient based on the increase of the initial contribution value in consecutive periods to form the contribution weight. The contribution weights and corresponding propagation path probabilities are segmented and combined, and then sorted, hierarchically, and normally according to the content element categories to generate a strategy priority distribution map.

[0012] Preferably, the key indicators extracted include: The strategy priority distribution map is divided into sections according to the content element categories. The changes in the proportion of different content formats in each priority level are statistically analyzed. When the proportion of a certain content format increases by more than 0.15 in at least two consecutive priority levels, the content format preference is determined. According to the interval values ​​in the priority distribution chart of the time stamp summary strategy, when the average priority value of a certain time period is higher than the preset multiple of the overall average and the fluctuation range is less than 0.3 times the overall fluctuation benchmark, the release time sensitive interval is determined. The trigger ratio is calculated based on the feedback intensity and range value corresponding to each content. When the trigger ratio of at least 3 consecutive content is higher than the set benchmark value, the minimum value of the corresponding feedback intensity is extracted as the interaction trigger threshold.

[0013] Preferred content seeding strategies include: The content format preference, publication time sensitivity interval and interaction trigger threshold are combined and encoded, and compared with the content format conditions, publication time conditions and feedback intensity conditions in the preset rule base one by one, and the rule entries with a matching similarity of not less than 0.7 are selected. When the matching similarity corresponding to the content format preference is not less than 0.7, or the average priority within the sensitive period of publication time reaches the preset multiple condition, or the ratio of consecutive content triggering reaches the set benchmark value, the key indicator is deemed to have met the preset triggering condition. Based on the matching rules, the system reads the content element categories, priority release time periods, and platform feedback intensity ratios from the high-priority layer, and outputs a content seeding strategy plan that includes content theme combinations, release rhythm, and channel configuration.

[0014] The technical effects and advantages provided by the present invention in the above technical solution are as follows: This invention constructs a competitive content sequence set and discretizes it under a unified time benchmark. This transforms the originally cross-platform, asynchronous content publishing and user interaction behaviors into a continuous signal sequence with a unified time scale, which is physically equivalent to performing temporal resampling and energy normalization on discrete propagation events. Based on this, through propagation speed difference analysis, phase change identification, and continuous evolution analysis of topic proportion, the content propagation process is transformed from single-point statistics into a dynamic process characterization with multi-dimensional coupling characteristics of speed, acceleration, and structural aggregation. Compared to existing techniques that only rank based on popularity or clicks, this invention can identify energy transition locations and structural aggregation trends during propagation, thus solving the technical problem of weak propagation signals and high noise interference in specific scenarios, making it difficult to identify key conversion nodes. Its direct technical effect is to improve the accuracy of judging key moments of propagation and the direction of content aggregation, enabling subsequent strategies to be based on stable and traceable propagation dynamics.

[0015] This invention constructs a probability sequence of propagation paths and performs segment-by-segment weighted accumulation, essentially modeling the path-level energy transfer of content propagation on a timeline. It transforms the influence relationships between content nodes into a chain structure with propagation attenuation and reinforcement characteristics, and combines this with trend coupling constraints to achieve continuous reconstruction of changes in user interests. This process is similar to the superposition and propagation of multi-path signals in a medium and the reconstruction of energy distribution, resulting in a continuous and directionally consistent prediction curve for product seeding trends. Unlike simple time-series prediction or rule-driven methods in existing technologies, this invention uses non-linear coupling between path probability and content contribution, making the formation of strategy priorities dependent on the energy distribution and cumulative effect in the actual propagation path. This effectively solves the problems of delayed user interest and untimely strategy response. This invention can characterize interest evolution trends in advance and output strategy schemes consistent with the propagation structure, improving the time-series matching degree of content delivery and resource utilization efficiency. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0017] Figure 1 This is a flowchart of the content seeding strategy generation method based on competitor data collection according to the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] For examples, please refer to Figure 1 As shown in this embodiment, the content seeding strategy generation method based on competitor data collection includes: We acquire content publishing data and user interaction behavior data of competitors on multiple platforms, clean and align them with time to form a set of competitor content sequences that include publishing time, content type and feedback intensity.

[0020] In one implementation, the construction of the competitor content sequence set is carried out according to the following steps: First, we collected original posting data and corresponding user interaction data from competitors across multiple content platforms. Posting data included at least a unique content identifier, posting timestamp, and content type tag. User interaction data included at least the number of views, likes, comments, and shares. Data from different platforms was normalized using a unified identifier rule, generating a globally unique content identifier by concatenating the platform identifier with the original content number.

[0021] The raw data undergoes cleaning and processing, including outlier removal and missing value imputation. Outliers are identified using statistical distribution boundaries. The mean μ and standard deviation σ for each type of interaction behavior within the statistical period are calculated, and data exceeding μ ± 3σ are identified as outliers and removed. Missing values ​​are imputed using nearest-neighbor interpolation, selecting the two most recent valid data points before and after the timestamp and filling them with a weighted average based on time distance, with the weight inversely proportional to the time interval. Then, time alignment is performed on data from different platforms. Using a unified time baseline, all timestamps are converted into second-level time series and discretized at fixed time intervals Δt, where Δt ranges from 60 to 300 seconds. For any given time interval containing multiple data points, they are aggregated using a weighted average, with the weights determined by the proportion of total interaction for each data point; if a time interval contains no data, it is marked as a placeholder value of 0.

[0022] Based on this, a feedback intensity index is calculated. Feedback intensity is obtained by weighted summation of multiple interactive behaviors. The calculation process is as follows: multiply the number of views, likes, comments, and shares by their respective weights and then sum them. The weights are set to satisfy the following values: views 0.1 to 0.3, likes 0.2 to 0.4, comments 0.2 to 0.5, and shares 0.3 to 0.6, with the sum of all weights being 1. The specific weight values ​​are set based on the behavioral conversion rate statistics of different platforms.

[0023] Finally, the processed data is sorted chronologically to form a competitor content sequence set. Each record in this sequence set contains at least the publication time, content type, and corresponding feedback intensity value, which are used for subsequent dissemination feature extraction and trend analysis.

[0024] Based on the competitor content sequence set, the dissemination growth rate, interaction turning points and theme aggregation characteristics within different time windows are extracted to construct a competitor dissemination feature set.

[0025] In one implementation, the competitor dissemination feature set is constructed based on the competitor content sequence set. The construction process includes segmentation processing, extraction of dissemination growth rate sequence, identification of interaction turning points, determination of theme aggregation features, and time alignment integration.

[0026] First, the competitor content sequence set is segmented according to a preset sliding time window. The time window length is set based on the content release density within the collection period, ranging from 600 to 3600 seconds; when the number of content items per hour is no less than 30, the time window length is 600 to 1200 seconds; when the number of content items per hour is 10 to 29, the time window length is 1200 to 2400 seconds; and when the number of content items per hour is less than 10, the time window length is 2400 to 3600 seconds. The sliding step size is 0.2 to 0.5 of the time window length; when the feedback intensity fluctuation is large, it is 0.2 to 0.3, and when the feedback intensity fluctuation is small, it is 0.4 to 0.5. The feedback intensity fluctuation is obtained by averaging the absolute values ​​of the differences between the feedback intensity of each record and the average feedback intensity.

[0027] For each time window, the total feedback intensity falling within that window is calculated by summing the feedback intensity of each record and then dividing by the number of records within that time window to obtain the window's average feedback value. If no records exist within a time window, the window's average feedback value is recorded as 0, and the window's position is retained. Following chronological order, the window feedback averages of two adjacent time windows are differentially calculated. This is done by subtracting the average feedback value of the previous time window from the average feedback value of the subsequent time window, and then dividing by the time interval between the center moments of the two time windows to obtain the propagation speedup value for that adjacent window. All propagation speedup values ​​are arranged chronologically to form the initial propagation speedup sequence.

[0028] To reduce local jumps caused by sporadic interactions, the initial propagation rate increase sequence is cumulatively smoothed. The cumulative smoothing process uses a weighted approach with adjacent points. For any propagation rate increase value, the current rate increase, the preceding rate increase, and the following rate increase are all included in the calculation. The weight of the current rate increase is 0.5 to 0.7, and the weights of the preceding and following rate increase values ​​are 0.15 to 0.25 respectively, with the sum of the three weights being 1. When a value is at the beginning or end of the sequence and lacks adjacent rate increase values, existing rate increase values ​​are used and their weights are normalized. The result after cumulative smoothing is taken as the propagation rate increase sequence.

[0029] Then, interaction inflection points are identified based on the propagation growth rate sequence. The sign changes of adjacent propagation growth rate values ​​are examined in chronological order. When the propagation growth rate value changes from positive to negative, or from negative to positive, the corresponding time position is taken as a candidate change position. When the propagation growth rate value is equal to 0, the most recent non-zero propagation growth rate value is searched backward and then backward. The candidate change position is determined based on the sign changes of the two non-zero propagation growth rate values.

[0030] For each candidate change location, the magnitude of the change in feedback intensity before and after is calculated. This is done by taking the absolute value of the difference between the mean window feedback of the preceding and following time windows. The overall mean is the average of the absolute values ​​of the differences in the mean window feedback of all adjacent time windows. When the magnitude of the change in feedback intensity before and after a candidate change location is greater than 1.5 to 2 times the overall mean, that candidate change location is identified as an interaction inflection point. The multiple is set based on the number of content items: 1.5 to 1.7 when there are at least 100 content items in the statistical period, 1.7 to 1.9 when there are 50 to 99 content items, and 1.9 to 2 when there are fewer than 50 content items.

[0031] Next, the distribution statistics of content types within each time window are performed. For any given time window, the number of records corresponding to each content type is counted, and the proportion of the same content type within that time window is calculated by dividing the number of records for that content type by the total number of records within that time window; when there are no records within that time window, the proportion of each content type is recorded as 0. Subsequently, the rate of change of proportion between adjacent time windows for the same content type is calculated by subtracting the proportion of the previous time window from the proportion of the subsequent time window. When a content type shows an increase in proportion for at least three consecutive time windows, and each increase is greater than 0.1, this continuous interval is defined as the theme clustering feature corresponding to that content type. If the continuous interval exceeds three time windows, the time window that first meets the condition is taken as the starting position of the theme clustering feature, and the time window before the increase in proportion no longer exceeds 0.1 is taken as the ending position.

[0032] Finally, the propagation growth rate sequence, interaction inflection points, and theme aggregation features are aligned and integrated according to a unified chronological order. During alignment and integration, the center moment of the sliding time window is used as the time marker. The propagation growth rate value, the existence of an interaction inflection point, the corresponding content type, and the theme aggregation feature status under the same time marker are written into the same feature record. When there is no interaction inflection point under a certain time marker, the interaction inflection point status is recorded as not triggered; when there is no theme aggregation feature, the theme aggregation feature status is recorded as empty. After arranging all feature records in chronological order, a set of competitor propagation features is obtained, which is used for subsequent identification of key influencing nodes and establishment of relational chains.

[0033] Based on the set of competitor propagation characteristics, key influencing nodes are identified and a chain of relationships between content is established. The propagation continuation probability of adjacent content nodes is calculated to form a propagation path probability sequence.

[0034] In one implementation, the propagation path probability sequence is formed based on a competitor propagation feature set. Feature records in the competitor propagation feature set are arranged in ascending order by time markers. Each feature record includes a propagation speedup value, interaction inflection point status, content type, and topic aggregation feature status. Content nodes correspond to content records under the specified time markers. The node feedback intensity is the average of the feedback intensities of all content within the same time range in the competitor content sequence set. When multiple pieces of content of the same type exist within the same time range, they are first sorted from highest to lowest feedback intensity, and the content with the highest feedback intensity (top 50%) is retained for subsequent processing to avoid low-response content interfering with the relationship chain.

[0035] First, the competitor's propagation feature set is traversed to identify local peaks in propagation growth rate. Local peaks are determined as follows: the propagation growth rate corresponding to the current time marker is greater than the propagation growth rate corresponding to the previous time marker, and also greater than the propagation growth rate corresponding to the next time marker; feature records located at the beginning or end of the propagation cycle are not included in the local peak determination. To avoid minor fluctuations being misjudged as local peaks, the average absolute change in the total propagation growth rate is calculated. The average absolute change is obtained by summing the absolute values ​​of the differences between adjacent propagation growth rates and dividing by the number of differences. The current time marker is retained as a candidate peak position if the absolute value of the difference between the propagation growth rate of the current time marker and the propagation growth rate of the previous time marker, and the absolute value of the difference between the propagation growth rate of the current time marker and the propagation growth rate of the next time marker, are both not less than 0.8 times the average absolute change. The 0.8 times can be adjusted according to the content release density: 0.8 to 1.0 when there are no fewer than 30 pieces of content per hour, and 1.0 to 1.2 when there are fewer than 30 pieces.

[0036] Next, the candidate peak positions are filtered based on their degree of overlap with the interaction inflection points. For any candidate peak position, the nearest interaction inflection point in time is found, and the interval between their time markers is calculated. If this interval does not exceed 0.3 times the length of the adjacent time window, the content corresponding to the candidate peak position is marked as a key influencing node; the length of the adjacent time window is determined according to the difference between the center times of the two adjacent time windows of the current candidate peak position.

[0037] When the same interactive turning point corresponds to multiple candidate peak positions, the candidate peak position with the larger propagation growth rate is selected as the key influencing node; when the propagation growth rate is the same, the content node with the larger feedback intensity is selected as the key influencing node.

[0038] Establish a relational chain starting from key impact nodes. For each key impact node, search for subsequent content nodes according to their chronological order, with the search scope not exceeding 6 time windows following the key impact node; when the content release interval within the statistical period is long, the search scope can be extended to 8 time windows. Subsequent content nodes need to have a continuous relationship with the key impact node in terms of content type or theme clustering characteristics. The method for determining a continuous relationship is: there must be at least 2 consecutive time windows between adjacent nodes that maintain the same theme clustering characteristics, and the proportion of the corresponding content type in both consecutive time windows must not be less than 0.2%.

[0039] If multiple subsequent content nodes simultaneously meet the conditions, they are added to the association chain in ascending order of time distance; if the time distances are the same, the content node with the higher feedback strength is prioritized. Subsequent content nodes that do not meet the continuity condition are not added to the association chain.

[0040] For each pair of adjacent content nodes in the association chain, calculate the propagation continuation probability. Based on the feedback strength of the previous node and the feedback strength of the subsequent node, if the feedback strength of the current node is greater than 0, divide the feedback strength of the subsequent node by the feedback strength of the previous node to obtain the base value; if the feedback strength of the current node is equal to 0 and the feedback strength of the subsequent node is greater than 0, the base value is 1; if both are 0, the base value is 0. Then, determine the consistency of the propagation growth rate direction between the two nodes. The determination method is to obtain the propagation growth rate direction within one time window after the time mark of the previous node, and the propagation growth rate direction within one time window before the time mark of the subsequent node; if both directions are increasing or both are decreasing, they are considered to be consistent; otherwise, they are considered to be inconsistent. When the directions are consistent, multiply the base value by 1.1 to 1.3 as the correction result; when the directions are inconsistent, multiply the base value by 0.7 to 0.9 as the correction result. The amplification factor is set to 1.2 to 1.3 when the absolute value of the difference in propagation growth rate is greater than the average absolute change, and 1.1 to 1.2 otherwise. The reduction factor is set to 0.7 to 0.8 when the absolute value of the difference in propagation growth rate is greater than the average absolute change, and 0.8 to 0.9 otherwise. A correction result greater than 1 is recorded as 1, and a result less than 0 is recorded as 0, thus obtaining the propagation continuation probability of adjacent content nodes.

[0041] Finally, the propagation continuation probabilities of adjacent content nodes within the same association chain are concatenated according to the node order to obtain the cumulative path value of the association chain.

[0042] The cumulative path value is obtained by multiplying them sequentially. When a relationship chain contains only one key influencing node and there are no subsequent effective connections, the cumulative path value is recorded as the normalized feedback strength of that key influencing node. After calculating the cumulative path value for all relationship chains, the cumulative path value of each relationship chain is divided by the sum of all cumulative path values ​​to obtain the corresponding path probability. The sum of all path probabilities is 1. The path probabilities are arranged from earliest to latest according to the start time of the relationship chain, forming a propagation path probability sequence. If the sum of all cumulative path values ​​is 0, then substitution normalization is performed based on the proportion of feedback strength of each key influencing node; if the feedback strength of all key influencing nodes is also 0, then the path probabilities of each relationship chain are equally distributed according to their number.

[0043] Based on the probability sequence of the propagation path, the performance trend of competitor content is reconstructed in time series, generating the coupling relationship between the trajectory of user interest changes and the changes in content performance, and obtaining the seeding trend prediction curve.

[0044] In one implementation, the trend prediction curve is formed based on a propagation path probability sequence. Each record in the propagation path probability sequence corresponds to a relationship chain and includes the path probability of the relationship chain, adjacent content nodes, the propagation continuation probability of adjacent nodes, the time stamp and feedback intensity corresponding to each content node.

[0045] First, each path segment is processed sequentially according to the time order of the propagation path probability sequence. A path segment represents the propagation connection between two adjacent content nodes. For any path segment, the feedback strength of the preceding node, the feedback strength of the following node, and the propagation continuation probability between them are obtained. The feedback strength of the following node is multiplied by the propagation continuation probability, and then summed with the feedback strength of the preceding node to obtain the weighted feedback value of that path segment. If multiple path segments enter within the same time window, the weighted feedback values ​​are added together and divided by the number of path segments to obtain the reconstruction feedback value for that time window; if no path segments exist in that time window, the reconstruction feedback value is recorded as 0.

[0046] For consecutive path segments within the same relational chain, reconstruction feedback values ​​are calculated sequentially from earliest to latest according to the node timestamps, and the path probability of that relational chain is used as the link weight for correction. The correction method is to multiply the reconstruction feedback value of each time window by the corresponding path probability to obtain the path correction feedback value. After all relational chains have been corrected, the path correction feedback values ​​within the same time window are summed to form the content presentation reconstruction sequence. Each item in the content presentation reconstruction sequence includes the center time of the time window and the corresponding reconstruction feedback value.

[0047] Then, the slope of change within a continuous time window is calculated based on the reconstruction sequence of content performance. The slope of change is obtained by subtracting the reconstruction feedback value of the previous time window from the reconstruction feedback value of the subsequent time window, and then dividing by the time interval between the center moments of the two time windows. The time interval is in seconds, and the calculation result represents the rate of change of content performance per unit time. When the center moments of two adjacent time windows are the same or the time interval is 0, this set of data is not included in the calculation of the slope of change.

[0048] The slope change rate is further calculated. It is obtained by subtracting the previous slope change from the subsequent slope change, and then dividing by the time interval between the corresponding center moments of the two slope changes. A positive slope change rate indicates an increasing rate of change in content expression; a negative slope change rate indicates a decreasing rate of change in content expression; and a slope change rate of 0 indicates that the rate of change remains constant. To reduce the impact of single-point fluctuations, the slope change rate is smoothed using three adjacent points. This is done by adding the current slope change rate, the previous slope change rate, and the next slope change rate, and then dividing by 3. When adjacent values ​​are missing at the beginning or end of the sequence, only the existing values ​​are used to calculate the average.

[0049] After obtaining the smoothed slope change rate, interest evolution intervals are identified chronologically. When the slope change rate is positive or negative for at least three consecutive time windows, this continuous interval is marked as an interest evolution interval. If an interval contains a value of 0, and the slope change rate before and after it is in the same direction, the time window with the 0 value is merged into the same interest evolution interval. If the continuous direction is maintained for more than six time windows, the time window with the largest absolute value of the slope change rate is taken as the interval peak position, and the start and end times, direction, and peak position of the interval are recorded.

[0050] The trajectory of user interest changes is extracted based on interest evolution intervals. For any interest evolution interval, the difference between the reconstruction feedback value of each time window within the interval and the reconstruction feedback value of the starting time window of the interval is used to obtain the amount of interest change. Then, the amount of interest change is divided by the difference between the maximum and minimum reconstruction feedback values ​​within the interval to obtain the normalized interest change value. If the maximum and minimum values ​​are equal, the normalized interest change value is recorded as 0. The normalized interest change values ​​are arranged in chronological order to form the trajectory of user interest changes.

[0051] Subsequently, the user interest change trajectory is time-aligned with the reconstructed sequence of content performance. Alignment is performed using the center moment of the time window as the benchmark, searching for normalized interest change values ​​and reconstructed feedback values ​​at the same center moment. If data for a certain time window is missing from the user interest change trajectory, the two most recent valid normalized interest change values ​​before and after that time window are linearly filled in; the filling method assigns inverse weights according to time distance, with closer distances resulting in larger weights, and the sum of the two weights is 1. If two valid values ​​before and after cannot be obtained, the most recent valid value is used to fill in the gaps.

[0052] After time alignment, the ratio of the magnitude of change between the user interest change trajectory and the reconstructed sequence of content performance within each time window is calculated. The magnitude of change in user interest is the absolute value of the difference between the normalized interest change value of the current time window and the normalized interest change value of the previous time window; the magnitude of change in content performance is the absolute value of the difference between the reconstructed feedback value of the current time window and the reconstructed feedback value of the previous time window. The ratio of magnitude of change is obtained by dividing the magnitude of change in user interest by the magnitude of change in content performance; when the magnitude of change in content performance is 0 and the magnitude of change in user interest is greater than 0, the ratio is 1; when both are 0, the ratio is 0.

[0053] The ratios of change amplitudes across time windows are smoothly connected in chronological order to form a coupling sequence. This smooth connection is achieved using a front-and-back window constraint: if the absolute value of the difference between the current and previous ratios is greater than twice the average of all ratios, the current ratio is adjusted to the average of the previous and current ratios; if the absolute value of the difference is not greater than twice the average of all ratios, the current ratio is retained. After processing all ratios, each item in the coupling sequence includes the center time of the time window and the corresponding coupling value.

[0054] Finally, continuity constraints are applied to the coupling relationship sequence. First, the absolute value of the difference in coupling values ​​within adjacent time windows is calculated, then the average of all absolute values ​​is calculated as the overall mean. When the difference in values ​​within adjacent time windows is less than 0.2 to 0.4 times the overall mean, the adjacent time windows are merged into the same trend segment. The multiple is set based on the fluctuation of the reconstructed sequence of content performance; when the fluctuation of the reconstructed feedback value is higher than 0.5 times its mean, 0.2 to 0.3 is used; when it is not higher than 0.5 times, 0.3 to 0.4 is used. The coupling values ​​of the merged trend segment are the average of all coupling values ​​within the segment, and the trend segment time is from the start time window to the end time window within the segment. The trend segments are connected in chronological order to obtain the seeding trend prediction curve.

[0055] Based on the aforementioned trend prediction curve, the contribution weights of different content elements are calculated and nonlinearly combined with the corresponding propagation path probabilities to generate a strategy priority distribution map.

[0056] In one implementation, the strategy priority distribution map is formed based on the seeding trend prediction curve and the probability sequence of propagation paths. The seeding trend prediction curve includes trend segments arranged in chronological order, each trend segment having a start time, an end time, and a coupling value; the content elements adopt existing content types in the competitor content sequence set and content categories corresponding to thematic clustering characteristics.

[0057] First, the trend prediction curve is segmented and analyzed. Based on the direction of change in the coupling values ​​of adjacent trend segments, the trend is divided into upward, stable, and downward intervals. An upward trend interval is defined as one where the coupling value of a subsequent trend segment is greater than that of the preceding trend segment, and the difference is not less than 0.3 times the average of all adjacent differences. A stable interval is defined as one where the absolute value of the difference is less than 0.3 times the average of all adjacent differences. A downward trend interval is defined as one where the coupling value of a subsequent trend segment is less than that of the preceding trend segment, and the absolute value of the difference is not less than 0.3 times the average of all adjacent differences. The average of all adjacent differences is obtained by summing the absolute values ​​of the differences in the coupling values ​​of each adjacent trend segment and then dividing by the number of adjacent differences. If there are fewer than three trend segments, additional segments are created using time windows in the competitor's dissemination feature set as units, and the coupling values ​​for these additional segments are taken from the coupling values ​​of the corresponding trend segments.

[0058] Within each interval, the frequency of occurrence of content elements and the percentage of feedback intensity are calculated. The frequency of occurrence of a content element is the number of times it appears in the corresponding interval divided by the total number of occurrences of all content elements in that interval. The percentage of feedback intensity is the sum of the feedback intensities corresponding to that content element divided by the sum of the feedback intensities corresponding to all content elements in that interval. If the total feedback intensity within an interval is 0, the percentage of feedback intensity is determined by the frequency of occurrence. The initial contribution value is obtained by weighted accumulation of frequency and percentage of feedback intensity. The calculation process involves multiplying the frequency of occurrence by its frequency weight, multiplying the percentage of feedback intensity by its intensity weight, and then adding the two results. The frequency weight ranges from 0.35 to 0.5, and the intensity weight ranges from 0.5 to 0.65, with the sum of the two weights being 1. When the content publication density is higher than the average publication density of the statistical period, the frequency weight ranges from 0.45 to 0.5; when the content publication density is not higher than the average publication density of the statistical period, the intensity weight ranges from 0.55 to 0.65.

[0059] Differential adjustments are made based on the variation of the initial contribution value across different stages. For the same content element, the corresponding initial contribution values ​​are arranged in chronological order of upward trend, stable period, and downward trend, and the contribution value increase between adjacent stages is calculated. The contribution value increase is the initial contribution value of the later stage minus the initial contribution value of the previous stage; the overall mean is the average of the positive increases of all content elements across all adjacent stages. When the contribution value increase of a content element exceeds 1.2 to 1.6 times the overall mean for at least two consecutive stages, an enhancement coefficient is assigned to that content element. The coefficient setting rules are: 1.2 to 1.35 when the number of sample contents is at least 100, 1.35 to 1.5 when the number of sample contents is 50 to 99, and 1.5 to 1.6 when the number of sample contents is less than 50. The enhancement coefficient is 1.1 to 1.4, and the closer the contribution value increase is to the upper limit of the coefficient, the higher the enhancement coefficient. The contribution weight of a content element is obtained by multiplying the initial contribution value by the enhancement coefficient; for content elements that do not meet the enhancement conditions, the enhancement coefficient is 1.

[0060] The contribution weights of content elements are then combined with their corresponding propagation path probabilities. The corresponding propagation path probability is determined based on the association chain in which the content element appears. When the same content element appears in multiple association chains, the average of the path probabilities of each association chain is taken as the corresponding propagation path probability. When the propagation path probability is higher than 0.6, the contribution weight is incrementally amplified. This is done by first subtracting 0.6 from the propagation path probability to obtain the excess value, then dividing the excess value into increments of 0.1. For each increment, the contribution weight is multiplied by 1.08 to 1.15. When the propagation path probability is between 0.3 and 0.6, linear superposition is maintained. This is done by multiplying the contribution weight by the propagation path probability and then adding 0.3 to 0.5 times the contribution weight. When the propagation path probability is lower than 0.3, suppression is applied. This is done by multiplying the contribution weight by the ratio of the propagation path probability to 0.3, and then multiplying by 0.7 to 0.9. This yields the combined results for each content element.

[0061] The combined results are categorized according to content element types. The combined results within each category are first summed, then divided by the number of content elements in that category to obtain a category priority value. Categories are then stratified from highest to lowest priority value: the top 30% are designated as high priority, those between 30% and 70% as medium priority, and those below 70% as low priority. To facilitate subsequent key indicator extraction, the category priority values ​​are converted to a uniform range of 0 to 1. The conversion process involves subtracting the minimum value from all category priority values ​​and then dividing by the difference between the maximum and minimum values; when the maximum value equals the minimum value, all categories are uniformly assigned a value of 0.5.

[0062] After conversion, the categories, priority levels, and interval values ​​are arranged in chronological order and numerical value to form a strategy priority distribution map.

[0063] The strategy priority distribution map is subjected to multi-dimensional feature extraction, and key indicators including content format preference, release time sensitivity and interaction trigger threshold are extracted and matched with a preset rule base.

[0064] In one implementation, key metrics are extracted and matched based on a strategy priority distribution map. The strategy priority distribution map includes content element categories, priority levels, interval values, time stamps, and feedback intensity, with each data point arranged in chronological order and numerical value.

[0065] First, the strategy priority distribution map is partitioned according to content element categories. Within each partition, the priority level, range value, content format, and feedback intensity for the same content element category are retained. Content formats are taken from the content type tags in the competitor's content sequence set. When multiple content formats exist under the same content element category, the number of each content format in the high-priority, medium-priority, and low-priority layers is counted separately. The priority percentage of a certain content format in a certain priority level is obtained by dividing the number of that content format in that level by the total number of all content formats in that level; when there are no content records in a level, the percentage is recorded as 0. The change range is obtained by subtracting the percentage of the previous priority level from the percentage of the next priority level, with the level order arranged as low-priority, medium-priority, and high-priority.

[0066] When a content format increases by more than 0.15 in at least two consecutive priority levels, that content format is identified as a content format preference. If multiple content formats meet the condition simultaneously, they are sorted from highest to lowest according to the average value of the interval values ​​in the higher priority levels, and the top three are selected as content format preferences; if the average values ​​of the interval values ​​are the same, the one with the higher average feedback intensity is selected.

[0067] Release time-sensitive intervals are extracted based on the distribution characteristics of the strategy priority distribution chart along the time dimension. Using time markers as a baseline, interval values ​​within the same time period are aggregated. The time period length remains consistent with the aforementioned sliding time window length; when the same release time spans multiple time periods, it is aggregated according to the time period into which the release time falls. The average priority of each time period is obtained by summing all interval values ​​within that time period and dividing by the number of interval values. The fluctuation range of each time period is obtained by subtracting the minimum interval value from the maximum interval value within that time period. The overall average is obtained by summing the average priority values ​​of all time periods and dividing by the number of time periods. The overall fluctuation benchmark is obtained by summing the fluctuation ranges of all time periods and dividing by the number of time periods. When the average priority of a certain time period is 1.2 to 1.5 times higher than the overall average, and the fluctuation range of that time period is less than 0.3 times the overall fluctuation benchmark, that time period is determined as a release time-sensitive interval. The multiplier value is set according to the number of time periods. When there are no less than 20 time periods, the multiplier is 1.2 to 1.3; when there are 10 to 19 time periods, the multiplier is 1.3 to 1.4; and when there are fewer than 10 time periods, the multiplier is 1.4 to 1.5.

[0068] The trigger ratio is calculated by combining the feedback intensity and interval values ​​corresponding to each content in the strategy priority distribution chart. For any content, the feedback intensity is first converted to the same interval of 0 to 1. The conversion process is to subtract the minimum feedback intensity of all content from the content's feedback intensity, and then divide by the difference between the maximum and minimum feedback intensity of all content. When the maximum value equals the minimum value, it is uniformly recorded as 0.5. The trigger ratio is obtained by multiplying the converted feedback intensity by the interval value and then dividing by the average of the two. When the average of the two is 0, the trigger ratio is recorded as 0. The trigger ratio is checked in the order of publication time. When the trigger ratio of at least 3 consecutive content items is higher than the set benchmark value of 0.6 to 0.8, the minimum value of the feedback intensity corresponding to the consecutive content items is extracted as the interaction trigger threshold. The set benchmark value is determined based on the dispersion of the feedback intensity, which is the standard deviation of all feedback intensity divided by the average value of the feedback intensity. When the dispersion is not higher than 0.3, it is taken as 0.6 to 0.7, and when the dispersion is higher than 0.3, it is taken as 0.7 to 0.8.

[0069] Content format preferences, release time sensitive intervals, and interaction trigger thresholds are combined and coded, then matched against a preset rule base. The combined code consists of three segments: the first segment is the content format preference number, the second segment is the release time sensitive interval number, and the third segment is the interaction trigger threshold interval number. The content format preference number is determined according to the order of content formats in the high-priority layer; the release time sensitive interval number is determined according to the chronological order of time periods; and the interaction trigger threshold interval number is determined by dividing the range of 0 to 1 into 5 levels. Each rule entry in the preset rule base includes content format conditions, release time conditions, feedback intensity conditions, and strategy action conditions. During each comparison, the matching degree between the three key indicators and the rule entry is calculated. Complete matching within the condition range is recorded as 1; partial overlap is calculated by dividing the overlap length by the indicator length; and no overlap is recorded as 0. The matching similarity is obtained by multiplying the three matching degrees by preset weights and then summing them. The weights for content format, release time, and feedback intensity are 0.3 to 0.4, and the sum of the three weights is 1. When the matching similarity is not less than 0.7, the corresponding rule entry is determined as the matching result; when there are multiple matching results, they are selected from high to low matching similarity.

[0070] When any key metric reaches the preset trigger condition, a content seeding strategy plan including content theme combination, release rhythm and channel configuration will be output.

[0071] In one implementation, after matching the content format preference, the sensitive period for publication time, and the interaction trigger threshold with the preset rule base, trigger judgments are made for the three key indicators. A key indicator is considered to have met the preset trigger condition if the matching similarity corresponding to the content format preference is not less than 0.7, or the average priority within the sensitive period for publication time is 1.2 to 1.5 times higher than the overall average, or the trigger ratio of at least three consecutive pieces of content is 0.6 to 0.8 higher than a set benchmark value. After the preset trigger condition is met, the rule content related to the content element category, priority level, and time period number in the corresponding rule entries is read. The content element categories in the high-priority layer are used as candidate topics, and the content format preference is used as the topic expression form. These are then arranged from high to low according to the interval values ​​in the strategy priority distribution chart to form content topic combinations.

[0072] The content seeding strategy includes content theme combination, release schedule, and channel configuration. The release schedule is determined based on the sensitive time interval for releases, prioritizing these intervals and setting the release interval based on the interaction trigger threshold. When the feedback intensity is below the interaction trigger threshold, the release interval between adjacent content releases is extended; when the feedback intensity reaches or exceeds the interaction trigger threshold, the release interval between adjacent content releases is maintained or shortened. Channel configuration is determined based on the feedback intensity percentage of different platforms within the competitor's content sequence. Platforms with the highest feedback intensity percentage are prioritized, while platforms with a feedback intensity percentage between 30% and 70% are considered supplementary channels. The content format for each channel is determined in conjunction with content format preferences, resulting in a content seeding strategy that includes theme, timing, and channel arrangements.

[0073] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for generating content seeding strategies based on competitor data collection, characterized in that, include: Obtain competitor content publishing data and user interaction behavior data on multiple platforms, clean and time-align the data to form a competitor content sequence set that includes publishing time, content type and feedback intensity; Based on the content sequence set of competitors, the dissemination growth rate, interaction turning points and theme clustering characteristics in different time windows are extracted to construct a set of competitor dissemination characteristics. Based on the aforementioned set of competitor propagation characteristics, key influencing nodes are identified and a chain of relationships between content is established. The propagation continuation probability of adjacent content nodes is calculated to form a propagation path probability sequence. Based on the probability sequence of the propagation path, the performance trend of competitor content is reconstructed in time series to generate the coupling relationship between the trajectory of user interest change and the content performance change, and thus obtain the seeding trend prediction curve. Based on the aforementioned trend prediction curve, the contribution weights of different content elements are calculated and nonlinearly combined with the probabilities of the corresponding propagation paths to generate a strategy priority distribution map. Multidimensional feature extraction is performed on the strategy priority distribution map to extract key indicators including content format preference, release time sensitivity and interaction trigger threshold, and then matched with a preset rule base. When any key metric reaches the preset trigger condition, a content seeding strategy plan including content theme combination, release rhythm and channel configuration will be output.

2. The content seeding strategy generation method based on competitor data collection according to claim 1, characterized in that, The competitive content sequence set includes: Collect original publishing data and user interaction behavior data of competitors on multiple content platforms, and generate globally unique content identifiers based on platform identifiers and original content IDs; The original published data and user interaction data are processed to remove outliers, fill in missing values, and align time at the second level to obtain data records arranged according to a unified time base. The number of views, likes, comments, and shares are weighted and summed according to preset weights to obtain the feedback intensity, and then a sequence of competitor content is formed according to the publication time.

3. The content seeding strategy generation method based on competitor data collection according to claim 1, characterized in that, Constructing a set of competitor communication features includes: The content sequence set of competitors is segmented according to the sliding time window, and the propagation growth rate sequence is obtained based on the difference in the mean of window feedback between adjacent time windows. The interaction inflection point is determined based on the position of the positive and negative changes in the propagation growth sequence and the magnitude of the changes in the feedback intensity before and after it. The changes in the proportion of content types within each time window are statistically analyzed. When the proportion of the same content type increases by more than 0.1% in at least three consecutive time windows, the theme clustering characteristics are determined and aligned with the dissemination growth rate sequence and interaction inflection points to form a set of competitor dissemination characteristics.

4. The content seeding strategy generation method based on competitor data collection according to claim 1, characterized in that, Key impact nodes identified include: Traverse the set of competitor propagation features and filter candidate peak positions where the propagation growth rate is greater than both the previous and next time markers; Calculate the time interval between the candidate peak position and the most recent interaction inflection point. When the time interval does not exceed 0.3 times the length of the adjacent time window, mark the corresponding content as a key influencing node. When multiple candidate peak positions correspond to the same interactive inflection point, the content with the larger propagation growth rate is selected as the key influencing node. When the propagation growth rates are the same, the content with the larger feedback intensity is selected as the key influencing node.

5. The content seeding strategy generation method based on competitor data collection according to claim 4, characterized in that, Establishing a chain of relationships includes: Starting with the key impact node, search for subsequent content nodes in chronological order within the next 6 to 8 time windows. When adjacent content nodes maintain the same theme clustering feature for no less than 2 consecutive time windows, and the corresponding content type accounts for no less than 0.2%, it is determined that there is a valid association between the two nodes. Subsequent content nodes that meet the criteria for effective association will be connected to key influential nodes in order of time distance from nearest to farthest, forming a chain of associations. When the time distance is the same, content nodes with higher feedback intensity will be prioritized for connection.

6. The content seeding strategy generation method based on competitor data collection according to claim 5, characterized in that, The probability sequence of propagation paths includes: For each pair of adjacent content nodes in the association chain, the ratio between the feedback strength of the later node and the feedback strength of the previous node is used as the base value. The base value is corrected based on the consistency of the direction of propagation growth rate change between the two nodes to obtain the propagation continuation probability; The cumulative path value is obtained by accumulating the propagation continuation probabilities of each path within the same association chain, and then normalizing the cumulative path value to form a propagation path probability sequence.

7. The content seeding strategy generation method based on competitor data collection according to claim 1, characterized in that, The predicted trend curve for grass planting includes: The path segments are unfolded in chronological order according to the probability sequence of the propagation path. The feedback intensity of the subsequent node is multiplied by the probability of propagation continuation and then superimposed with the feedback intensity of the preceding node to obtain the reconstructed feedback value. This value is then combined with the path probability to form a reconstructed sequence of content presentation. Based on the content performance, the reconstruction sequence calculates the slope and rate of change of the slope within a continuous time window. When the rate of change of the slope changes in the same direction within at least three adjacent time windows, the interest evolution interval is determined and the user interest change trajectory is extracted. The user interest change trajectory is time-aligned with the content performance reconstruction sequence, and a coupling relationship sequence is formed based on the ratio of the change magnitude of the two. After applying continuity constraints to the coupling relationship sequence, the trend prediction curve of product seeding is obtained.

8. The content seeding strategy generation method based on competitor data collection according to claim 1, characterized in that, The generation strategy priority distribution map includes: Based on the direction of change of the coupled values ​​of adjacent trend segments in the grass planting trend prediction curve, the trend segments are divided into an upward trend range, a stable range, and a downward trend range. The frequency of occurrence of content elements and the proportion of feedback intensity are statistically analyzed in each interval, and the two are weighted and accumulated to obtain the initial contribution value. Then, the content elements are assigned an enhancement coefficient based on the increase of the initial contribution value in consecutive periods to form the contribution weight. The contribution weights and corresponding propagation path probabilities are segmented and combined, and then sorted, hierarchically, and normally according to the content element categories to generate a strategy priority distribution map.

9. The content seeding strategy generation method based on competitor data collection according to claim 1, characterized in that, Key metrics to extract include: The strategy priority distribution map is divided into sections according to the content element categories. The changes in the proportion of different content formats in each priority level are statistically analyzed. When the proportion of a certain content format increases by more than 0.15 in at least two consecutive priority levels, the content format preference is determined. According to the interval values ​​in the priority distribution chart of the time stamp summary strategy, when the average priority value of a certain time period is higher than the preset multiple of the overall average and the fluctuation range is less than 0.3 times the overall fluctuation benchmark, the release time sensitive interval is determined. The trigger ratio is calculated based on the feedback intensity and range value corresponding to each content. When the trigger ratio of at least 3 consecutive content is higher than the set benchmark value, the minimum value of the corresponding feedback intensity is extracted as the interaction trigger threshold.

10. The content seeding strategy generation method based on competitor data collection according to claim 9, characterized in that, The content seeding strategy includes: The content format preference, publication time sensitivity interval and interaction trigger threshold are combined and encoded, and compared with the content format conditions, publication time conditions and feedback intensity conditions in the preset rule base one by one, and the rule entries with a matching similarity of not less than 0.7 are selected. When the matching similarity corresponding to the content format preference is not less than 0.7, or the average priority within the sensitive period of publication time reaches the preset multiple condition, or the ratio of consecutive content triggering reaches the set benchmark value, the key indicator is deemed to have met the preset triggering condition. Based on the matching rules, the system reads the content element categories, priority release time periods, and platform feedback intensity ratios from the high-priority layer, and outputs a content seeding strategy plan that includes content theme combinations, release rhythm, and channel configuration.

Images