Culture circle applet multi-module dynamic association recommendation method and system

Abstract

This invention relates to the field of recommender system technology and discloses a multi-module dynamic association recommendation method and system for a cultural circle mini-program. The method includes: calculating a transition progress coefficient based on the remaining time value of a child role node to generate an enhanced role graph; using the transition progress coefficient to perform linear interpolation fusion of the current role feature vector and the behavior migration pattern vector to generate a transitional period hybrid role feature vector; generating a transitional period perceived family demand fusion vector through attention-weighted aggregation; mining high-frequency cross-module consumption link patterns based on prefix trees, modulating the confidence of the link patterns belonging to different school age stages, and combining similarity scores to output stage-perceived multi-module dynamic association recommendation results.

Description

Technical Field

[0001] This invention relates to the field of recommendation system technology, specifically to a method and system for dynamic association recommendation of multiple modules in a cultural circle mini-program. Background Technology

[0002] Cultural mini-programs typically offer services to families, covering multiple modules such as education, parent-child activities, and cultural consumption. As children in these families progress through different school stages, their needs for these services also gradually change across these stages. Existing recommendation methods usually generate recommendations based on users' static role identifiers or historical behavior records. However, when a child's role is in a transitional phase between school stages, static role identifiers cannot reflect the gradual change in needs, causing recommendations to lag behind the actual changes in user requirements.

[0003] Furthermore, the consumption behavior of family users across multiple service modules typically exhibits a chain-like continuity, meaning that consumption in one module often triggers subsequent demands for services in other modules. Existing recommendation methods mostly trigger recommendations independently based on single state transition events, failing to perceive the user's current stage in the cross-module consumption chain, nor can they adjust the preference patterns of different stages based on the progress of school-age transition. Therefore, a dynamic multi-module association recommendation method and system for cultural circle mini-programs is needed, capable of perceiving the progress of school-age transition and supporting the identification of cross-module consumption chain stages. Summary of the Invention

[0004] In response to the problems and needs mentioned above, this invention proposes a multi-module dynamic association recommendation method and system for cultural circle mini-programs. It solves the above problems and brings other technical effects by adopting the following technical features.

[0005] On one hand, this invention proposes a multi-module dynamic association recommendation method for cultural circle mini-programs, which includes at least the following steps: obtaining the birth date and current school stage label of each child role node in the family role graph; calculating the remaining time value of each child role node from the next school stage transition node based on the school age transition timetable; when the remaining time value is less than a preset transition period threshold, calculating the difference between the transition period threshold and the remaining time value and dividing it by the transition period threshold to obtain a transition progress coefficient; attaching the transition period marker, the target school stage label, and the transition progress coefficient to the corresponding child role node to generate an enhanced role graph; and using the transition progress... The degree coefficient is used as the interpolation weight. Linear interpolation is performed to fuse the current role feature vector with the typical transitional behavior migration pattern vector of the peer group to generate a transitional hybrid role feature vector. The transitional hybrid role feature vector replaces the original feature vector of the corresponding role node in the enhanced role graph. The transition progress coefficient is encoded and concatenated to the query vector of the transitional role node. Attention-weighted aggregation is performed on all role nodes to generate a transitional perceived family demand fusion vector. The matching score between the feature vector of each candidate service item and the family demand fusion vector is calculated. The results are sorted in descending order of the matching score and the multi-module dynamic association recommendation results are output.

[0006] Further, the linear interpolation fusion method is as follows: multiply the current role feature vector by a value minus the transition progress coefficient, multiply the transition period typical behavior migration pattern vector by the transition progress coefficient, and add the two to obtain the transition period hybrid role feature vector; wherein, the transition period typical behavior migration pattern vector is generated by: retrieving a set of peer users according to the target learning stage label and age range, obtaining the access records of each module service category of the peer user set within a preset time window before and after the transition period, calculating the usage frequency before and after the transition period according to the module identifier and service category label, and concatenating the frequency change values ​​of all modules to form the transition period typical behavior migration pattern vector.

[0007] Furthermore, the step of encoding and concatenating the transition progress coefficient to the query vector of the transition period role node includes: inputting the transition progress coefficient into a fully connected layer, the fully connected layer performing a linear transformation on the transition progress coefficient and mapping it through an activation function to output a fixed-dimensional vector consistent with the dimension of the query vector, concatenating the fixed-dimensional vector to the end of the query vector of the corresponding role node to form an extended query vector, and using the extended query vector to participate in the attention weighted aggregation.

[0008] Furthermore, the candidate service item feature vector is the encoded representation of each service item in terms of service category, applicable age group, and module affiliation dimension, wherein the classification attributes of service category, applicable age group, and module affiliation are converted into numerical vectors through one-hot encoding; the matching score is calculated using the cosine similarity algorithm.

[0009] Furthermore, after generating the transition-perceived family demand fusion vector, the method further includes: extracting cross-module service consumption sequences from the historical order data of all users according to the user dimension and a preset time window; replacing the service items in each sequence with a combination code of their module identifier and service category label; applying a prefix tree generation algorithm to the combined coded consumption sequences of all users to generate a prefix tree structure; traversing all paths in the prefix tree structure; filtering subsequences whose frequency exceeds a preset support threshold to generate a high-frequency cross-module consumption link pattern library; querying the association mapping table between service category labels and school age stages based on the service category labels contained in each link pattern; labeling each link pattern with a school age stage affiliation label; and generating a stage-labeled consumption link pattern library.

[0010] Furthermore, the school age stage attribution label includes three types: source school stage label, target school stage label, and cross-school stage label; wherein, when all service category labels in the link mode belong to the service category corresponding to the source school stage, it is labeled as the source school stage label; when all belong to the service category corresponding to the target school stage, it is labeled as the target school stage label; and when it contains service categories of both the source school stage and the target school stage, it is labeled as the cross-school stage label.

[0011] Furthermore, it also includes: acquiring records of state migration events completed by the current family within a preset backtracking time window; arranging the combined encoding of the module identifier and service category label corresponding to each event in chronological order to form a consumption link prefix sequence; performing prefix matching between the consumption link prefix sequence and the consumption link pattern library labeled with the stage to obtain a set of successfully matched link patterns and the original link confidence and school age stage affiliation label of each pattern; performing stage modulation on the original link confidence of each matched link pattern based on the transition progress coefficient: for link patterns labeled with the target school stage label, multiplying its original link confidence by one plus the product of the enhancement adjustment coefficient and the transition progress coefficient; for link patterns labeled with the source school stage label, multiplying its original link confidence by one minus the product of the attenuation adjustment coefficient and the transition progress coefficient; for link patterns labeled with the cross-school stage label, keeping its original link confidence unchanged; wherein the values ​​of the enhancement adjustment coefficient and the attenuation adjustment coefficient are greater than zero and less than or equal to one.

[0012] Furthermore, it also includes: for each unfinished link in a matching link pattern, retrieving a set of service items matching the service category label of that link from the candidate service items of the corresponding module; calculating the similarity score between the feature vector of each candidate service item and the fusion vector of the family demand using a cosine similarity algorithm, multiplying the similarity score by the modulated link confidence of the link pattern to generate a stage-aware multi-module dynamic association recommendation score; for candidate service items belonging to the target learning stage association module, multiplying their stage-aware multi-module dynamic association recommendation score by one plus the transition progress coefficient to obtain the final score; and sorting and outputting the transition period-aware multi-module dynamic association recommendation results in descending order of the final score.

[0013] Furthermore, it also includes: grouping by link mode, arranging by link time sequence within each group, selecting the service item combination corresponding to the link mode with the highest confidence after modulation as the main package, and using the service items of the second-best link mode as the alternative packages; performing module diversity interval insertion processing in the descending order of the candidate service item sequence, when two adjacent service items belong to the same module, retrieving the different module service items with the highest final score from the subsequent sequence and inserting them between the two, until the module identifiers of adjacent positions are all different or the sequence has been traversed.

[0014] On the other hand, this invention proposes a multi-module dynamic association recommendation system for cultural circle mini-programs, used to execute the aforementioned multi-module dynamic association recommendation method for cultural circle mini-programs, comprising at least: a transition period status acquisition module, used to acquire the birth date and current school stage label of each child role node in the family role graph, calculate the remaining time value of each child role node from the next school stage transition node based on the school age transition timetable, calculate the transition progress coefficient when the remaining time value is less than a preset transition period threshold, and generate an enhanced role graph; and a feature fusion module, used to use the transition progress coefficient as an interpolation weight to compare the current role feature vector with that of peers. The transitional period typical behavior migration pattern vector of the group is linearly interpolated and fused to generate a transitional period hybrid role feature vector; the demand aggregation module is used to replace the original feature vector of the corresponding role node in the enhanced role graph with the transitional period hybrid role feature vector, encode and concatenate the transition progress coefficient to the query vector of the transitional period role node, and perform attention-weighted aggregation on all role nodes to generate a transitional period perceived family demand fusion vector; the recommendation output module is used to calculate the matching score between the feature vector of each candidate service item and the family demand fusion vector, sort them in descending order of the matching score and output the multi-module dynamic association recommendation results.

[0015] This invention calculates a transition progress coefficient based on the remaining time value of the child role node, and uses this coefficient to perform linear interpolation and fusion of the current role feature vector and the behavioral migration pattern vector of the peer group, so that the mixed role feature vector during the transition period can change continuously with the transition progress. By encoding and concatenating the transition progress coefficient to the query vector of the transition period role node and then performing attention-weighted aggregation, the generated family demand fusion vector can reflect the mixed state of the coexistence of new and old module demands during the transition period, solving the technical problem that recommendation results under static role identification lag behind actual demand changes.

[0016] Meanwhile, this invention mines high-frequency cross-module consumption link patterns from the full user consumption sequence using a prefix tree generation algorithm, and labels each pattern with a school age stage affiliation tag, enabling the system to identify the user's current consumption link stage. Based on this, a transition progress coefficient is used to enhance or attenuate the confidence of link patterns belonging to different school age stages, allowing the recommendation results to dynamically adjust cross-module association preferences according to the school age transition progress. This solves the technical problem that recommendations triggered independently by a single state transition event cannot perceive the consumption link stage and cannot adjust cross-module association preferences according to the transition progress, achieving the technical effect that the recommendation results possess both link coherence and the ability to perceive the school age transition progress. Attached Figure Description

[0017] Figure 1 This is a flowchart of the multi-module dynamic association recommendation method for cultural circle mini-programs provided in this embodiment of the invention; Figure 2 This is a schematic diagram comparing the feature vectors of service categories of each module provided in the embodiments of the present invention; Figure 3 This is a schematic diagram illustrating the ranking of candidate service item matching scores provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the frequency distribution of the consumption link pattern library path provided in the embodiments of the present invention; Figure 5 This is a schematic diagram comparing the link confidence stage before and after modulation, provided in an embodiment of the present invention. Figure 6 This is a schematic diagram illustrating the final recommendation score composition analysis provided in an embodiment of the present invention; Figure 7 This is a schematic diagram illustrating the influence of the transition progress coefficient α on feature vector interpolation provided in an embodiment of the present invention; Figure 8 This is a schematic diagram illustrating the relationship between remaining time and transition progress coefficient provided in an embodiment of the present invention; Figure 9 This is a schematic diagram illustrating the trend of the confidence modulation coefficients of the target and source learning segments as a function of α, provided in an embodiment of the present invention. Detailed Implementation

[0018] like Figure 1 As shown, a multi-module dynamic association recommendation method for cultural circle mini-programs according to an embodiment of the present invention includes the following steps: Step 1: Obtain the transitional state information of child character nodes and generate an enhanced character graph; Retrieve the birth date and current school stage label for each child role node in the family role graph. Based on a preset school transition schedule, calculate the remaining time value between each child role node and the next school stage transition node. When the remaining time value is less than a preset transition period threshold, generate a transition period marker and a target school stage label for that child role node.

[0019] The transition schedule coefficient is calculated based on the remaining time value, as follows: ; in, For the first The transition progress coefficient of each child character node. The node number for the child character. This represents the remaining time until the next learning stage transition point. This is a preset transition period threshold, and both have the same unit. The range of values ​​is The closer This indicates that the closer to the transition point in the learning stage, the better.

[0020] The transition period marker, target grade level label, and transition progress coefficient are attached to the corresponding child role node to form an enhanced role graph with transition period status.

[0021] It should be noted that the above school age transition timetable is a pre-configured mapping table that records the standard school entry age and entry month corresponding to each school stage label. Remaining time value The calculation method is as follows: Calculate the target transition date based on the birth date and the standard enrollment age for the target educational stage, then subtract the current date and take the difference. Transition period threshold. It can be set to a fixed duration in months, with a value range of 6 to 12 months.

[0022] Step 2: Integrate current character features with transitional behavior migration patterns to generate a transitional hybrid character feature vector; For role nodes marked with a transitional status, cross-module behavior distribution data of peer users matching the target learning stage label are retrieved from the historical behavior database. The frequency of service category usage for each module within the peer group before and after the transitional period is statistically analyzed, and typical behavioral migration pattern vectors during the transitional period are extracted.

[0023] Step 2 may specifically include: Step 201: Search for the set of users of the same age group by target learning stage label and age range, and obtain the access records of each module service category of the set of users of the same age group within the preset time window before and after the transition period.

[0024] Step 202: Calculate the usage frequency before and after the transition period based on module identifier and service category label, calculate the frequency change value, and concatenate the frequency change values ​​of all modules into a typical behavior migration pattern vector for the transition period. .in, For the first The typical behavior migration pattern vector corresponding to each transitional role node.

[0025] Step 203: Use a linear interpolation algorithm with transition progress coefficients. As interpolation weights, for the current character feature vector Transitional period typical behavior transfer pattern vector The feature vectors of the transitional hybrid roles are fused together. The calculation method is as follows: ; in, For the first The current role feature vector of each transitional role node. This represents the vector of typical transitional behavior patterns. The transition progress coefficient calculated in step 1, This is the generated feature vector for the transitional hybrid role. When near During the transition period, the hybrid role feature vector is dominated by the current role's features; when near During this period, the feature vector of the mixed role in the transition period is mainly composed of migration pattern features.

[0026] It should be noted that the selection criteria for the aforementioned peer group include: consistent target grade level labels, a birth date difference between the current transitional role node and the target grade level birth date within a preset range, and completion of the corresponding grade level transition. By limiting these conditions, behavioral data of groups that have already undergone the same grade level transition can be obtained as a statistical source for migration patterns.

[0027] Step 3: Aggregate the updated role node features to generate a transition period-aware family needs fusion vector; Replace the original feature vector of the corresponding character node in the enhanced character graph with the transitional hybrid character feature vector. Perform attention-weighted aggregation on all updated character node feature vectors.

[0028] Step 3 may specifically include: Step 301: For role nodes in the transition period, additionally concatenate the transition progress coefficient encoding into their attention query vector. It should be noted that the transition progress coefficient encoding is generated by: converting the scalar... Input fully connected layer, fully connected layer to Perform a linear transformation and map it through an activation function to output a fixed-dimensional vector with the same dimension as the query vector. Then, append this fixed-dimensional vector to the end of the query vector of the corresponding role node to form an extended query vector.

[0029] Step 302: Perform attention-weighted aggregation on all updated role nodes. The input to attention-weighted aggregation is the feature vector and query vector of each role node, and the output is a family demand fusion vector. For transitional role nodes, an extended query vector with transition progress coefficient encoding is concatenated to enable the attention weight allocation to reflect the gradual change in demand during the transition period, generating a transition-perceived family demand fusion vector.

[0030] Step 4: Calculate the matching score between the candidate service item and the transition period perceived family demand fusion vector, and output the multi-module dynamic association recommendation result; The system retrieves a set of service items from the candidate service items of each module that are associated with both the target grade level label and the current grade level label of the transitional role node. Using the cosine similarity algorithm, and taking the feature vector of each candidate service item and the fusion vector of family needs perceived during the transitional period as input, it calculates the matching score between each candidate service item and family needs. The candidate service items are then sorted in descending order of matching score, and the dynamic association recommendation results across multiple modules are output.

[0031] It should be noted that the feature vectors of the aforementioned candidate service items are coded representations of each service item in dimensions such as service category, applicable age group, and module affiliation, and are pre-extracted and stored in the platform's service database. Among them, the categorical attributes such as service category, applicable age group, and module affiliation are converted into numerical vectors through one-hot encoding before participating in subsequent calculations.

[0032] In this embodiment of the application, in order to identify the current stage of the cross-module consumption chain of the user and to adjust the chain mode of different stages according to the progress of school age transition, the following steps are also included in step 3: Step 5: Extract cross-module consumption sequences and generate a consumption link pattern library with stage annotations; From the historical order data of all users, extract cross-module service consumption sequences by user dimension and preset time window. Replace the service item in each sequence with a combined code of its module identifier and service category label.

[0033] Step 5 may specifically include: Step 501: Apply the prefix tree generation algorithm to the combined coded consumption sequences of all users. The input of the prefix tree generation algorithm is the set of combined coded consumption sequences of all users, and the output is a prefix tree structure containing the prefix paths of all sequences and the frequency of each path.

[0034] Step 502: Traverse all paths in the prefix tree, filter subsequences whose frequency exceeds the preset support threshold, and generate a high-frequency cross-module consumption link pattern library.

[0035] Step 503: Based on the service category tags contained in each link pattern, query the association mapping table between service category tags and school age stages, and label each link pattern with a school age stage affiliation tag. The school age stage affiliation tags include three types: source school age stage tag, target school age stage tag, and cross-school age stage tag, generating a consumption link pattern library with stage annotations.

[0036] It should be noted that the labeling rules for the above-mentioned school age stage attribution labels are as follows: when all service category labels in the link mode belong to the service category corresponding to the source school stage, they are labeled as source school stage labels; when all belong to the service category corresponding to the target school stage, they are labeled as target school stage labels; when they include service categories of both the source and target school stages, they are labeled as cross-school stage labels.

[0037] Step 6: Match the current household consumption link prefix and obtain the set of matching link patterns; Retrieve records of completed state migration events for the current household within a preset backtracking time window. Concatenate the module identifiers and service category labels corresponding to each event into a consumption link prefix sequence, ordered chronologically. Perform prefix matching between this consumption link prefix sequence and the stage-labeled consumption link pattern library to obtain a set of successfully matched link patterns, along with the original link confidence score and school-age stage affiliation label for each pattern.

[0038] It should be noted that the above original link confidence is a value calculated based on the frequency ratio of the matching path in the prefix tree during the prefix matching process, reflecting the statistical probability that the current consumer link prefix sequence will continue to the complete link mode.

[0039] Step 7: Perform staged modulation of the confidence level of the matched link mode based on the transition progress coefficient; Based on transition progress coefficient The original link confidence scores for each matching link mode are subjected to stage modulation, and the calculation method is as follows: For link patterns labeled with the target learning segment, their original link confidence scores are calculated. Multiply by the enhanced function value: ; For link patterns labeled with source learning segment tags, determine their original link confidence. Multiply by the decay function value: ; For link patterns labeled as cross-segment, the original link confidence level remains unchanged: ; in, For the first The original link confidence of the matching link pattern. To match the link mode number, This is the transition progress coefficient. The preset enhancement adjustment coefficient, The preset attenuation adjustment coefficient, The link confidence level after modulation of the target learning segment pattern. The link confidence level after modulation of the source segment pattern. This represents the link confidence level after modulation for cross-study-stage mode. Enhancement adjustment coefficient. and attenuation adjustment coefficient The range of values ​​is .

[0040] It should be noted that when Values and Values hour, This means that the link confidence drops to zero after modulation of the source segment mode, which is the limiting case for attenuation adjustment; in practical applications, The value can be set according to business needs to control the attenuation rate. When Approaching At that time, the link confidence of the target learning segment mode is enhanced to the greatest extent, while the link confidence of the source learning segment mode is reduced to the greatest extent, thus tilting the recommendation results toward the target learning segment related modules in the later stage of the transition period.

[0041] Step 8: Combine the post-modulated link confidence and the transition period-aware family demand fusion vector similarity to generate a stage-aware recommendation score and output the ranking result; For each unfinished link in a matching link pattern, retrieve a set of service items that match the service category tag of that link from the candidate service items of the corresponding module.

[0042] Step 8 may specifically include: Step 801: Using the cosine similarity algorithm, with the feature vectors of each candidate service item and the transition-period perceived family demand fusion vector as input, calculate the similarity score between each candidate service item and the family demand. Multiply the similarity score by the modulated link confidence of the corresponding link mode to generate a phase-period perceived multi-module dynamic association recommendation score.

[0043] Step 802: Sort candidate service items in descending order based on the multi-module dynamic association recommendation scores perceived at each stage. Group by link mode, and then sort by link sequence within each group. Select the service item combination corresponding to the link mode with the highest confidence after modulation as the primary package, and the service items of the second-best link mode as the alternative packages.

[0044] Step 803: Apply a transitional enhancement factor to candidate service items belonging to the target learning stage's associated module. It should be noted that the transitional enhancement factor is calculated as follows: As input, multiply the multi-module dynamic association recommendation score of stage perception by This ensures that the final score of candidate service items belonging to the target learning stage's associated module increases positively with the increase of the transition progress coefficient. Module diversity interval insertion processing is performed, which avoids the appearance of service items from the same module in adjacent positions in the ranking results, and alternates service items from different modules. The output is a multi-module dynamic association recommendation result with transition period awareness.

[0045] It should be noted that the specific method of the above module diversity interval insertion processing is as follows: In the candidate service item sequence arranged in descending order, if two adjacent service items belong to the same module, the different module service items with the highest multi-module dynamic association recommendation score in the subsequent sequence are retrieved and inserted between them, until the module identifiers of adjacent positions are all different or the sequence has been traversed.

[0046] This implementation calculates the remaining time between a child's role node and the next transition node in the learning stage. When the remaining time is less than the transition period threshold, a transition progress coefficient is generated. This coefficient is then used to linearly interpolate and fuse the current role feature vector with the behavioral migration pattern vector of the peer group. This allows the fused transition period hybrid role feature vector to gradually transition from the features of the current learning stage to those of the target learning stage over time. In the attention-weighted aggregation stage, the transition progress coefficient is encoded and concatenated to the query vector of the transition period role node. This enables the attention weight allocation during the aggregation process to distinguish between the gradual needs of the transition period and the fixed needs of the stable period. Consequently, the generated transition period-perceived family needs fusion vector accurately represents the mixed state of the coexistence of new and old module needs during the transition period, overcoming the shortcomings of static role identification where feature semantic drift causes recommendations to lag behind actual needs.

[0047] Meanwhile, this implementation method mines high-frequency cross-module consumption link patterns from the full user consumption sequence using a prefix tree generation algorithm and labels each pattern with its corresponding school age stage. This allows the system to perform prefix matching between the current family's consumption link prefix sequence and the stage-labeled consumption link pattern library, thereby identifying the user's current multi-step consumption link stage. Based on this, a transition progress coefficient is used to enhance or attenuate the confidence of link patterns belonging to different school age stages. This ensures that link patterns associated with the target school age stage gradually receive higher recommendation weights in the early stages of the transition period, while link patterns associated with the source school age stage gradually receive lower recommendation weights in the later stages. This achieves dynamic switching of recommendation results according to the school age transition progress. Therefore, this implementation method overcomes the shortcomings of recommendations triggered independently by a single state transition event, which cannot perceive the consumption link stage and cannot adjust cross-module association preferences according to the school age transition progress. It ensures that the recommendation results have both link coherence and reflect the gradual change trend of cross-module consumption patterns in families due to children's school age transition.

[0048] like Figures 2-9 As shown, based on the steps of the above-mentioned multi-module dynamic association recommendation method for the cultural circle mini-program, it is applied to the following example, where the application scenario is as follows: In the spring of 20XX, a cultural circle mini-program platform launched a multi-module dynamic association recommendation process for a family user (family ID FAM-0372). This family includes a child role node (role ID CHILD-01), currently in the last year of kindergarten and about to enter the first grade of primary school. The platform recorded the family's historical consumption behavior in multiple modules such as picture book reading, interest courses, stationery purchases, and parent-child activities, and executed the following steps based on this.

[0049] The application example includes: the system reads the birth date of CHILD-01, queries the school age transition schedule, calculates the remaining time value before the primary school enrollment node, and compares it with the preset transition period threshold.

[0050] Table 1 CHILD-01 Transitional Status Information The remaining time value of 5.5 months is less than the threshold of 12 months, triggering the transition period flag. The transition progress coefficient is calculated as follows, where... This represents the remaining time value. Transition period threshold: ; Transition progress coefficient Attached to the CHILD-01 node to form an enhanced character graph.

[0051] The system retrieves users of the same age group by the target grade level tag "first grade of primary school" and the corresponding age range, counts the changes in the frequency of use of service categories of each module before and after the transition period, extracts the typical behavior migration pattern vector of the transition period, and performs linear interpolation fusion with the current role feature vector.

[0052] Table 2 Current Role Feature Vector and Migration Pattern Vector (Frequency Components of Service Categories in Each Module) by As interpolation weights, a fusion calculation is performed on the "primary school supplementary materials" component, where For the current feature components, For migration mode components: ; The remaining components are calculated similarly to obtain the transitional period hybrid role feature vector. The weights of this vector in the target learning segment components such as primary school teaching materials and primary school courses are significantly higher than those in the current feature vector.

[0053] Replace the original feature vector of CHILD-01 in the enhanced character map with the hybrid character feature vector, and simultaneously... The input fully connected layer generates transition progress coefficient encoding, which is concatenated to the end of the attention query vector of CHILD-01 to form an extended query vector. Attention-weighted aggregation is performed on all family role nodes (including parent role nodes ADULT-01 and ADULT-02), outputting a transition-perceived family needs fusion vector. Because CHILD-01 carries the extended query vector, its attention weight for service categories related to the target educational stage is increased during the aggregation process compared to the non-transition state. The final fusion vector accurately reflects the coexistence of old and new module needs for the family.

[0054] The system retrieves service items that are simultaneously associated with the tags "kindergarten senior class" and "primary school first grade" from the candidate service items of each module, calculates the matching score between each candidate service item and the family's demand fusion vector using the cosine similarity algorithm, and outputs the recommendation results after sorting them in descending order.

[0055] Table 3. Matching Scores of Candidate Services (Partial) The system extracts cross-module consumption sequences from all users' historical order data within a preset time window, replacing service items with a combination of module identifier and service category label. A prefix tree generation algorithm is applied to all sequences, traversing all paths and filtering subsequences whose frequency exceeds a support threshold, generating a high-frequency cross-module consumption link pattern library. Then, the system queries the association mapping table based on the service category label in each pattern, marking the appropriate educational stage label.

[0056] Table 4. Partial List of Consumption Link Patterns Marked by Stage The system obtains the state transition event records that FAM-0372 has completed within the preset backtracking time window, concatenates them into the consumption link prefix sequence "Course·Primary School Transition", performs prefix matching with the pattern library, and obtains the successfully matched link pattern and its original link confidence.

[0057] Table 5 Prefix matching results and original link confidence by Enhancement coefficient of the target learning stage Source segment attenuation coefficient The raw link confidence scores for each matching link mode are subjected to staged modulation, whereby... This represents the confidence level of the original link.

[0058] Taking PAT-02 (target grade level) and PAT-01 (source grade level) as examples: ; ; Table 6 Link Confidence After Stage Modulation For candidate service items in each matching link mode's recommended stage, calculate their cosine similarity score with the transition-perceived family demand fusion vector, multiply it by the modulated link confidence of the corresponding mode, and generate a stage-perceived multi-module dynamic association recommendation score. Apply a transition-perceived enhancement factor to service items belonging to the target learning stage's association module. Then, perform module diversity interval insertion processing to output the final recommendation result.

[0059] Table 7 Final Recommendation Score and Output Ranking The "Elementary School Teaching Aid Set → Stationery Gift Box" combination corresponding to PAT-02 is the main package, while the "Entrance Experience Activity" corresponding to PAT-04 is the alternative. The reading module, stationery module, and parent-child module appear alternately to meet the requirements of module diversity and interval insertion.

[0060] Throughout the process, the data starts from CHILD-01's birth date and is used to calculate the transition progress coefficient based on the remaining time value. This coefficient is used in all subsequent steps: in the feature fusion step, it controls the interpolation ratio between the current feature and the transfer pattern; in the attention aggregation step, it influences the weight allocation in the form of encoding; in the stage modulation step, it enhances the confidence of the target learning segment link pattern and decays the confidence of the source learning segment; and in the final ranking step, it serves as the basis for the transition period enhancement factor.

[0061] In the final recommendation results, the scores of the target grade level related service items (primary school teaching aids, stationery gift boxes, and school experience activities) were significantly higher than those of the source grade level related service items. Moreover, the service items of different modules were arranged alternately, reflecting the cross-module related recommendation ability that gradually changes with the progress of school age transition.

[0062] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for dynamic association and recommendation of multiple modules in a cultural circle mini-program, characterized in that, Includes the following steps: Obtain the birth date and current school stage label of each child role node in the family role graph. Calculate the remaining time value of each child role node from the next school stage transition node based on the school age transition timetable. When the remaining time value is less than a preset transition period threshold, calculate the difference between the transition period threshold and the remaining time value and divide it by the transition period threshold to obtain the transition progress coefficient. Attach the transition period marker, target school stage label, and the transition progress coefficient to the corresponding child role node to generate an enhanced role graph. Using the transition progress coefficient as the interpolation weight, the current role feature vector and the typical behavior migration pattern vector of the peer group during the transition period are linearly interpolated and fused to generate a hybrid role feature vector during the transition period. The typical behavior migration pattern vector during the transition period is generated as follows: retrieve the set of users of the same age group according to the target learning stage label and age range, obtain the access records of each module service category of the user set of the same age group within the preset time window before and after the transition period, count the usage frequency before and after the transition period according to the module identifier and service category label, calculate the frequency change value, and concatenate the frequency change values ​​of all modules to form the typical behavior migration pattern vector during the transition period. The original feature vector of the corresponding role node in the enhanced role graph is replaced by the transitional hybrid role feature vector. The transition progress coefficient is encoded and concatenated to the query vector of the transitional role node. The transition progress coefficient is input into a fully connected layer. The fully connected layer performs a linear transformation on the transition progress coefficient and maps it through an activation function, outputting a fixed-dimensional vector with the same dimension as the query vector. The fixed-dimensional vector is concatenated to the end of the query vector of the corresponding role node to form an extended query vector. The extended query vector participates in attention weighted aggregation. The extended query vector is used as the query input of the corresponding role node in attention weighted aggregation. Attention weighted aggregation is performed on all role nodes. The input of attention weighted aggregation is the feature vector and extended query vector of each role node. The output is a family demand fusion vector, generating a transitional perceived family demand fusion vector. After generating the transition-perceived family demand fusion vector, the process further includes: extracting cross-module service consumption sequences from the historical order data of all users according to user dimensions and preset time windows; replacing the service items in each sequence with a combination code of their module identifier and service category label; applying a prefix tree generation algorithm to the combined coded consumption sequences of all users to generate a prefix tree structure; traversing all paths in the prefix tree structure; filtering subsequences whose frequency exceeds a preset support threshold; generating a high-frequency cross-module consumption link pattern library; querying the association mapping table between service category labels and school age stages based on the service category labels contained in each link pattern; labeling each link pattern with a school age stage affiliation label; and generating a stage-labeled consumption link pattern library. It also includes: obtaining the status migration event records that the current family has completed within the preset backtracking time window, combining and encoding the module identifier and service category label corresponding to each event into a consumption link prefix sequence in chronological order, performing prefix matching between the consumption link prefix sequence and the consumption link pattern library marked by the stage, and obtaining the set of successfully matched link patterns and the original link confidence and school age stage affiliation label of each pattern; For link patterns labeled with the target learning segment, their original link confidence scores are calculated. Multiply by the enhanced function value: ； For link patterns labeled with source learning segment tags, determine their original link confidence. Multiply by the decay function value: ； For link patterns labeled as cross-segment, the original link confidence level remains unchanged: ； in, For the first The original link confidence of the matching link pattern. To match the link mode number, This is the transition progress coefficient. The preset enhancement adjustment coefficient, The preset attenuation adjustment coefficient, The link confidence level after modulation of the target learning segment pattern. The link confidence level after modulation of the source segment pattern. To enhance the link confidence of the cross-segment mode after modulation, the adjustment coefficient is increased. and attenuation adjustment coefficient The range of values ​​is ; It also includes: for each unfinished link in a matching link pattern, retrieving a set of service items matching the service category label of that link from the candidate service items of the corresponding module; calculating the similarity score between the feature vector of each candidate service item and the fusion vector of the family demand using the cosine similarity algorithm, multiplying the similarity score by the modulated link confidence of the link pattern to generate a stage-aware multi-module dynamic association recommendation score; for candidate service items belonging to the target learning stage association module, multiplying their stage-aware multi-module dynamic association recommendation score by... As the final score, the multi-module dynamic association recommendation results with transition period awareness are sorted in descending order and output. Calculate the matching score between the feature vector of each candidate service item and the fusion vector of the family demand, sort them in descending order of the matching score, and output the multi-module dynamic association recommendation results.

2. The multi-module dynamic association recommendation method for cultural circle mini-programs according to claim 1, characterized in that, The linear interpolation fusion method is as follows: Transition progress coefficient As interpolation weights, for the current character feature vector Transitional period typical behavior transfer pattern vector The feature vectors of the transitional hybrid roles are fused together. The calculation method is as follows: ； in, For the first The current role feature vector of each transitional role node. This represents the vector of typical transitional behavior patterns. The transition progress coefficient calculated in step 1, For the generated transitional hybrid role feature vector, when near During the transition period, the hybrid role feature vector is dominated by the current role's features; when near During this period, the feature vector of the mixed role in the transition period is mainly composed of migration pattern features.

3. The multi-module dynamic association recommendation method for cultural circle mini-programs according to claim 1, characterized in that, The candidate service item feature vector is the encoded representation of each service item in terms of service category, applicable age group, and module affiliation dimension. The classification attributes of service category, applicable age group, and module affiliation are converted into numerical vectors through one-hot encoding. The matching score is calculated using the cosine similarity algorithm.

4. The multi-module dynamic association recommendation method for cultural circle mini-programs according to claim 1, characterized in that, The school age stage attribution labels include three types: source school stage label, target school stage label, and cross-school stage label. Among them, when all service category labels in the link mode belong to the service category corresponding to the source school stage, they are labeled as source school stage labels; when all belong to the service category corresponding to the target school stage, they are labeled as target school stage labels; and when they contain service categories of both the source school stage and the target school stage, they are labeled as cross-school stage labels.

5. The multi-module dynamic association recommendation method for cultural circle mini-programs according to claim 1, characterized in that, Also includes: Grouped by link mode, within each group arranged by link time sequence, the service item combination corresponding to the link mode with the highest confidence after modulation is selected as the main package, and the service items of the second best link mode are selected as the alternative packages; in the descending order of the candidate service item sequence, module diversity interval insertion processing is performed. When two adjacent service items belong to the same module, the service item of the different module with the highest final score is retrieved from the subsequent sequence and inserted between the two, until the module identifiers of adjacent positions are all different or the sequence is traversed.

6. A multi-module dynamic association recommendation system for cultural circle mini-programs, used to execute the multi-module dynamic association recommendation method for cultural circle mini-programs as described in any one of claims 1 to 5, characterized in that, include: The transition period status acquisition module is used to acquire the birth date and current school stage label of each child role node in the family role graph, calculate the remaining time value of each child role node from the next school stage transition node based on the school age transition timetable, and calculate the transition progress coefficient when the remaining time value is less than the preset transition period threshold to generate an enhanced role graph. The feature fusion module is used to linearly interpolate and fuse the current role feature vector with the typical behavior migration pattern vector of the peer group during the transition period using the transition progress coefficient as the interpolation weight, so as to generate a hybrid role feature vector during the transition period. The demand aggregation module is used to replace the original feature vector of the corresponding role node in the enhanced role graph with the feature vector of the transitional hybrid role, encode and concatenate the transition progress coefficient to the query vector of the transitional role node, perform attention weighted aggregation on all role nodes, and generate a transitional perceived family demand fusion vector. The recommendation output module is used to calculate the matching score between the feature vector of each candidate service item and the fusion vector of the family demand, sort them in descending order of the matching score, and output the dynamic association recommendation results of multiple modules.

Images