Hierarchical distribution and path optimization method and system for sales coaching tasks

Abstract

The application provides a hierarchical distribution and path optimization method and system for selling accompanying training tasks, relates to the technical field of data processing, and is based on the multidimensional characteristics of tasks. All accompanying training tasks are divided into difficulty level clusters with clear structures through a clustering algorithm. When recommending, a real-time vector is matched with a task level center and a task characteristic, so that a most suitable challenge and most relevant target personalized next task is dynamically generated, and dynamic optimization of a learning path is realized.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a method and system for hierarchical allocation and path optimization of sales coaching tasks. Background Technology

[0002] In the context of increasingly fierce industry competition, sales teams face the urgent challenge of quickly mastering strategic product knowledge (9+7+5 model) and adapting to diversified sales scenarios and complex customer communication. To help sales teams rapidly improve their sales capabilities in strategic products and services, innovative training methods and the introduction of intelligent technologies have become crucial.

[0003] However, existing technologies have rigidity in their assessment and recommendation mechanisms: First, the assessment system relies on a static, manually preset, and outdated knowledge framework, which leads to a disconnect between assessment results and actual effectiveness; second, the task recommendation adopts a one-size-fits-all linear model, which cannot capture the multidimensional dynamic changes in trainees' abilities and makes it difficult to provide truly suitable challenges.

[0004] Specifically, at the evaluation level, the system's judging criteria are fixed. For example, for ice-breaking skills in retail stores, the knowledge graph may have long contained fixed keywords such as weather and renovation. When, in real-world scenarios, due to digital transformation, top salespeople have switched to using new, pain-point-directing sales techniques, the static system, lacking new key intent nodes such as online reputation, Meituan reviews, and digital tools in its knowledge base, will mechanically determine that the trainee's response has a low match with the standard keywords, thus giving a negative evaluation of intent misunderstanding and failing to accurately measure the trainee's true business level.

[0005] Furthermore, at the recommendation level, the system's coaching strategy often relies solely on task tags (such as beginner, objection handling) or the learner's simple historical scores for linear recommendations, neglecting the fine-tuning between the task's overall difficulty (such as semantic complexity, required skill range, and historical average completion time) and the learner's real-time ability profile, such as the fluency of their current responses, semantic flexibility, and stability under pressure. For example, a learner with solid product knowledge but slow on-the-spot adaptability might be repeatedly recommended high-intensity simulated customer challenges by the system due to their high knowledge mastery score, leading to a decline in confidence due to continuous frustration. Conversely, a learner with advanced skills might still be assigned basic product feature repetition exercises due to the fixed system path, resulting in wasted training time and diminished motivation.

[0006] Therefore, the existing evaluation and recommendation systems are generally inefficient. Summary of the Invention

[0007] In view of the shortcomings of the existing technology, the purpose of this invention is to provide a method and system for hierarchical allocation and path optimization of sales coaching tasks, so as to solve the problems mentioned in the background technology.

[0008] To achieve the above objectives, this invention provides the following technical solution: a method for hierarchical allocation and path optimization of sales coaching tasks, comprising the following steps: First, based on a knowledge graph G and historical data of coaching tasks, extract multidimensional features of each coaching task and calculate the comprehensive similarity between tasks; second, use a clustering algorithm to divide all coaching tasks into multiple task level clusters corresponding to different difficulty levels, and generate a hierarchical data sequence containing the task level and its features; third, collect and process student interaction data to generate a student ability vector V. le Based on historically accumulated student ability vectors and practice path sequences, asynchronously and dynamically adjust the learning conversion cost weights between skill nodes in the knowledge graph G; based on the current student ability vector V... le The system matches the central features of each level cluster in the hierarchical data sequence to select preliminary candidate levels, and then matches these candidate levels with the specific training task features to calculate the optimal next training task. This optimal next training task is then assigned to the student, and after the student completes the task, the student's ability vector V is updated based on the new interaction data. le .

[0009] On the other hand, the present invention also includes a hierarchical allocation and path optimization system for sales coaching tasks, comprising an offline hierarchical module that pre-extracts multi-dimensional features of each coaching task based on a knowledge graph G and historical data of coaching tasks, calculates the comprehensive similarity between tasks, divides all coaching tasks into multiple task level clusters corresponding to different difficulty levels through a clustering algorithm, and generates a hierarchical data sequence containing the level to which the task belongs and its features; and a feature extraction module that collects and processes student interaction data to generate a student ability vector V. le The knowledge graph optimization module asynchronously and dynamically adjusts the learning conversion cost weights between skill nodes in the knowledge graph G based on historically accumulated student ability vectors and practice path sequences; the task matching module, based on the current student ability vector V... le The system matches the central features of each level cluster in the hierarchical data sequence to select preliminary candidate levels, and then matches these candidate levels with the specific training task features to calculate the optimal next training task. The allocation and update module assigns the optimal next training task to the student, and updates the student's ability vector V based on new interaction data after the student completes the task. le .

[0010] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention automates the distribution of training tasks for large-scale sales teams, optimizes the allocation efficiency of training resources, and enables refined scheduling of learning paths. By dynamically adjusting the learning conversion cost weights between skill nodes based on the actual ability vectors of all trainees, the knowledge graph iterates with business evolution and collective learning feedback, ensuring the timeliness of assessments. Secondly, based on the multi-dimensional features of tasks, the system uses a clustering algorithm to divide all training tasks into clearly structured difficulty level clusters. During recommendation, it matches the real-time trainee ability vector with the task level center and the task's own characteristics, thereby dynamically generating the most suitable and relevant personalized next task for each trainee. Furthermore, by extracting and calculating features including semantic matching degree, intent consistency, response efficiency, and operational accuracy from trainees' original interactive behaviors, such as voice, clicks, and navigation, and standardizing them into a unified trainee ability vector, the system transforms the trainee's comprehensive performance during training—including script relevance, depth of intent understanding, fluency of response, and operational accuracy—into a set of quantifiable and comparable indicators. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of the overall processing flow of the hierarchical allocation and path optimization method for sales coaching tasks proposed in one embodiment of the present invention.

[0012] Figure 2 This is a flowchart of task hierarchical clustering in one embodiment of the present invention.

[0013] Figure 3 This is a flowchart of the weighted attribution update process in one embodiment of the present invention. Detailed Implementation

[0014] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0015] The present invention will be further described in detail below with reference to the accompanying drawings, but this is not intended to limit the scope of the invention.

[0016] As one embodiment of the present invention, such as Figures 1-3 As shown, the proposed method for hierarchical allocation and path optimization of sales coaching tasks includes the following steps: Step S1: Based on the knowledge graph G and the historical data of the training tasks, extract the multidimensional features of each training task and calculate the comprehensive similarity between tasks. Then, use a clustering algorithm to divide all training tasks into multiple task level clusters corresponding to different difficulty levels, and generate a hierarchical data sequence containing the level to which the task belongs and its features.

[0017] Step S1-1: Construct a comprehensive feature vector F representing the semantics of the training task and the student's historical performance. task The specific process is as follows: Steps S1-11: In this embodiment, based on the pre-set script database in the system, a constructed sentence segmentation strategy is used to perform structured parsing on each training task, so as to parse the structured dialogue script of each training task into a series of semantically independent and independently evaluable minimum dialogue units, defined as atomic training units. The sentence segmentation strategy used is as follows: Based on the predefined standardized dialogue process framework in the sales knowledge graph, such as opening remarks, needs assessment, product recommendation, objection handling, and closing, logical segments are segmented to ensure that each logical segment corresponds to a complete sales skill execution step. Based on punctuation marks, such as periods, question marks, exclamation marks, and turn-taking markers, including speaker identifiers and preset pause duration thresholds, the text structure of each logical segment is segmented to obtain the natural turn units in the dialogue.

[0018] To facilitate understanding, taking the gigabit broadband promotion scenario training task as an example, it is broken down into multiple atomic training units, such as a customer stating that the current network speed cannot meet business needs, a salesperson inquiring about the store's daily network usage scenarios and peak hours, a salesperson introducing the advantages and stability of gigabit broadband uplink speed, or a customer raising concerns about the high monthly fee. After each atomic training unit is extracted, standardized preprocessing is performed: redundant data, including HTML tags and special characters, is removed, a unique unit identifier ID is generated for each atomic training unit, and a coding rule using the training task ID as the sequence number is adopted. A mapping relationship is established between this ID and the text content of the atomic training unit, the skill tags mapped from the knowledge graph G, and the temporal position in the original training task, and stored in the database.

[0019] Based on the above technical concept, the process of establishing the mapping relationship is as follows: After completing the text extraction and preprocessing of each atomic training unit, a corresponding data object is created. This object uses the unique ID of the atomic training unit as the primary key and contains the following fields: raw_text, which stores the cleaned raw dialogue text; skill_tag, which stores the skill node ID mapped from the sales knowledge graph G; and sequence_index, which records the absolute sequential position of the unit in its parent training task as an integer, starting from 0 and incrementing. Subsequently, these data objects are inserted into a relational database, such as MySQL, in the form of records. It should be noted that during this process, the skill_tag field is associated with the skill node table in the knowledge graph G through foreign key constraints or reference queries to ensure tag consistency. The sequence_index is automatically generated by incrementing the count when inserting all atomic training units in the same training task. Finally, through this ID primary key, all the association information of any atomic training unit is obtained, thereby establishing the mapping relationship.

[0020] Steps S1-12: Encode each raw_text using the pre-trained language model BERT. After average pooling, obtain the unit semantic vector. Concatenate all unit semantic vectors in order to obtain the original semantic vector E representing the overall semantics of the training task. raw Simultaneously, the set of skill nodes trained for this coaching task is extracted from the knowledge graph G, and its assessment feature vector F is obtained. node And for all the associated assessment feature vectors F node Averaging the vectors yields the standard skill vector F, representing the ideal skill requirement for the training task. skill The dimension is 5.

[0021] Steps S1-13: Retrieve all historical V data generated by trainees after practicing on this training task. le Vectors, calculated in each dimension (S sem ,S in E avg E std The mean of (ac) is used to form a vector H representing the historical average performance and difficulty of the training task. mean The dimension is 5; the original semantic vector E is transformed by constructing a fully connected layer. raw Standard skill vector F skill Compared with historical performance vector H mean By concatenating the data and setting the output dimension of the fully connected layer to a fixed 5, a comprehensive feature vector F representing the characteristics of the training task is generated. task The calculation formula is: F task =ReLU(W f ×[E raw;F skill ;H mean ]+b f ); In the formula, W f Let b be a trainable weight matrix. f The term is the bias term, [;] represents the vector concatenation operation, and ReLU is the activation function.

[0022] Furthermore, in steps S1-13, the training method or deterministic calculation rule for the comprehensive feature vector generation model (fully connected layer) is as follows: The weight matrix W of the fully connected layer is randomly initialized using He. f Bias term b f Typically initialized to a zero vector; the previous layer of neurons (here, the concatenated [E]) raw ;F skill ;H mean The output of the vector is passed as input to each neuron in the fully connected layer. The input is multiplied by the weight matrix, and the bias is added to obtain the output. Simultaneously, an activation function (ReLU) can be used to process the output, introducing non-linearity. Further, an appropriate loss function is selected based on the task type (e.g., classification or regression) to measure the difference between the model output and the true label. Based on the loss value calculated using the loss function, the loss on the weight matrix and bias term b is calculated using the backpropagation algorithm. f The gradient is calculated; an optimizer (such as stochastic gradient descent, Adam, etc.) updates the weight matrix and bias terms based on the calculated gradient, and the process of forward propagation, loss calculation, backpropagation, and parameter update is repeated continuously to make the model output better meet the task requirements. This series of processes is a current technique in neural network training.

[0023] Step S1-2: Based on the similarity of the characteristics of the training tasks, divide the training tasks into different difficulty levels, so that the system can adjust the difficulty level according to the student's real-time ability V. le The system recommends the most suitable next training task from the corresponding level to ensure the continuity of the learning path and the appropriateness of the challenge. The specific implementation method is as follows: Steps S1-21: Construct a multi-source similarity matrix W. For any two training tasks i and j, calculate the comprehensive similarity W. ij The formula is expressed as: ; In the formula, the first term is the comprehensive feature vector F of the training task. task The cosine similarity is used to measure the semantic similarity of the overall features of the training task. The second term is based on the historical performance vector H. meanThe Gaussian kernel similarity is used, where σ is the bandwidth parameter, which measures the similarity of the training tasks in terms of historical completion difficulty. The third term is the Jaccard similarity of the set of skill nodes associated with the training task, which measures the correlation at the skill level. α, β, and γ are adjustable fusion weights, satisfying α + β + γ = 1, and the optimal values ​​of parameters α, β, γ, and σ are determined through grid search and contour coefficient method. and These are the sets of skill nodes associated with training tasks i and j, respectively.

[0024] Steps S1-22: Based on the similarity matrix W, a clustering algorithm is used to achieve the optimal partitioning of the training task set, obtaining training task difficulty levels with clear structure and well-defined boundaries. The process includes: First, by analyzing the distribution of the W matrix values, the median or 70th percentile is taken as the association strength threshold θ. link It is used to determine whether two training tasks are similar enough to be associated; the cluster merging determination ratio ρ m The parameter is used to determine whether the correlation strength ratio between training task pairs within two training task clusters meets the merging criteria, such as setting it to 80%; and the maximum optimization rounds MaxR is used to control the maximum number of iterations for cluster optimization to prevent infinite loops.

[0025] Simultaneously, two logical storage structures are created in system memory, including an abnormal training task set D, used to store tasks whose similarity to all other training tasks in the entire training task set is below the association strength threshold θ. link The isolated training tasks; and the candidate level cluster set C, which is initially empty, is used to store the preliminary training task clusters obtained by scanning the association relationship, and to store the final training task clusters formed after merging operations in the subsequent iterative optimization process.

[0026] Secondly, based on the similarity matrix W, a binary association strength matrix A for training tasks is generated for anomaly screening. The judgment rule is as follows: if W ij ≥θ link Then A ij =1 indicates that training task i and training task j are strongly associated, and the latter is retained for subsequent cluster partitioning; otherwise, A ij =0. Then, anomaly screening for training tasks is performed: for each row i of matrix A, the sum of correlation Sum is calculated. i =Σ j A ij If Sum i =0 indicates that the similarity between training task i and all other training tasks in the set is less than θ. link If it is determined to be an abnormal training task, it is removed from the main set and stored in the abnormal training task set D.

[0027] Next, iterate through the association strength matrix A after removing abnormal training tasks to construct a preliminary candidate level cluster. Initialize an empty temporary cluster T. C Following the index order of the training tasks, perform an adjacency traversal on matrix A: when a training task p is encountered, if it is not assigned to any existing cluster, it is placed as a seed training task into a temporary cluster T. C Then search for all expressions that satisfy A' {pq} The training task q with a value of 1 is added to the temporary cluster T. C Next, for the temporary cluster T C For each newly added coaching task, recursively search for all its strongly related coaching tasks and add them, until no new coaching tasks are added. In the formula, p is the coaching task currently being traversed, serving as the starting point for cluster construction; q represents other coaching tasks in the same column as coaching task p in the matrix; and A' {pq} is a matrix element, representing the element in the p-th row and q-th column of matrix A', with a value of 0 or 1.

[0028] At this time, temporary cluster T C All the training tasks constitute a preliminary candidate cluster, which is stored as a whole in the candidate level cluster set C, and the temporary cluster T is cleared. C This will begin building the next cluster. It should be noted that this process repeats until all training tasks have been divided.

[0029] Based on the above technical concept, it can be understood that finding all expressions satisfying A' {pq} The purpose of the training task q with a value of 1 is to expand the adjacency relationship starting from the seed training task p, thereby efficiently aggregating strongly associated training tasks and forming a preliminary candidate cluster.

[0030] For ease of understanding, this embodiment uses a specific sales coaching scenario as an example. Assume that the remaining coaching task set after anomaly screening contains six coaching tasks: T1 (basic icebreaking), T2 (FABE introduction), T3 (handling price objections), T4 (competitive comparison), T5 (closing techniques), and T6 (online icebreaking). Let the association strength threshold be θ. link =0.75. The following table shows some rows of the binary correlation matrix A' obtained after removing abnormal training tasks, where the row and column indices correspond to training tasks T1 to T6. An example table is shown below:

[0031] Table 1 Training Assistant Task List

[0032]

[0033] Based on Table 1 above, a new cluster is constructed starting with training task T2 as the seed training task p. The algorithm traverses all columns q in matrix A' where p=2 (the second row), corresponding to training task T2. When q=1 (training task T1), A' {21} =1, which satisfies the condition, so T1 is added to the temporary cluster T. C When q=3 (training task T3), A' {23} =1, which satisfies the condition, so T3 is added to the temporary cluster T. C When q=4 (training task T4), A' {24} =1, which satisfies the condition, so T4 is added to the temporary cluster T. C When q=5 (training task T5) or q=6 (training task T6), A' {25} =0, A' {26} =0, the condition is not met, so it is not added.

[0034] Thus, after this round of adjacency expansion, the temporary cluster T... C The initial set of members {T2, T1, T3, T4} is obtained by gathering the seed training task T2 and all its directly and strongly related training tasks. This set forms an initial candidate cluster around product introduction and objection handling skills. Then, the related training tasks of newly added members (T1, T3, T4) are recursively checked until no further expansion is possible, thereby ensuring that all closely related training tasks within this local scope are obtained.

[0035] Finally, the candidate-level cluster set C is iteratively optimized by evaluating the cluster center similarity in each iteration to achieve dynamic merging. In practice: a. In each round of optimization, calculate the cluster center: for each candidate cluster c in set C. k Calculate its cluster center vector C' k The center vector is formed by the F of all training tasks within the cluster. task The arithmetic mean of the vectors is obtained as follows: ; In the formula, |c k |For set c k The cardinality represents the total number of training tasks belonging to this candidate cluster. The comprehensive feature vector of training task i includes the original semantics of the training task, standard skill requirements, and historical average performance.

[0036] b. Through the cluster center vector C' k The cosine similarity formula is used to calculate the similarity (i.e., the cosine similarity of the center vectors) between any two clusters cm and cn: ; c. Perform refined merging checks, and set the similarity interval [θ] that needs to be refined. low ,θ high For example, if set to [0.80, 0.92], the value is determined based on experts' experience in classifying the difficulty levels of training tasks, while meeting the actual business needs of the scenario. For Csim... mn If Csim mn ≥θ high If it is, then mark it as mergeable; if Csim mn ≤θ low If θ low <Csim mn <θ high Then, perform a refined check: calculate the similarity value between cluster cm and all training task pairs (a total of |cm||cm| pairs) in the original similarity matrix W, and count the similarity values ​​of W. {ij} ≥θ link Calculate the number of sparring partners and their proportion (P) to the total number of sparring partners. mn If P mn ≥ρ m If the two clusters are closely related, they are considered mergingable. Otherwise, they are not merged.

[0037] d. Merge all cluster pairs marked as mergeable in this round to form a new, larger cluster, and update the candidate level cluster set C. Simultaneously, repeat the iterative evaluation of cluster center similarity steps described above until no cluster merging occurs in two consecutive optimization rounds, or the maximum number of optimization rounds (MaxR) is reached. At this point, each cluster in C represents a final determined difficulty level for the training task.

[0038] S1-3: In specific implementation, based on the obtained final candidate level cluster set C, a two-dimensional data sequence is generated through the following steps: First, calculate the comprehensive feature vector F of all training tasks within each cluster of set C. task In the preset difficulty-related dimensions (historical average performance vector H) mean The overall mean of the mean of the clusters is used to sort all clusters in ascending order based on this mean. Then, ordered level labels L1 to Lk are assigned sequentially, where k is the total number of clusters. Secondly, establish a mapping relationship between the level labels and the cluster center vectors to ensure that each label Lx corresponds to the center vector C' of its cluster. x This serves as a benchmark for assessing the match between the student's ability and the level of the training task in subsequent training task recommendations.

[0039] Finally, traverse all training tasks and construct a two-dimensional data sequence. The specific operation is as follows: take a single training task in each cluster of set C as the data row, and take the four attributes of each training task as the data column, and organize them according to the following rules: the first column is the unique identifier of the training task, Task. ID The second column is the level label L. evel The third column is the comprehensive feature vector F of this training task. task The fourth column is the center vector C of the cluster to which the training task belongs. 'k Finally, the system iterates through every training task in all clusters and extracts its Task from the database. ID and F task And determine L based on its cluster. evel and the corresponding C 'k The data is then filled row by row into the sequence. The final output is a two-dimensional data sequence. It can be understood that the row indices of the two-dimensional data sequence correspond to different training tasks, and the column indices correspond to the four attribute fields mentioned above, fully including the training task classification information.

[0040] For newly added training tasks between two periodic hierarchical updates, the system employs an incremental hierarchical mechanism: It extracts the comprehensive feature vector of the new task, calculates its cosine similarity to the center vector of each existing level cluster, assigns it to the level cluster with the highest similarity, and temporarily assigns the corresponding level label to that cluster. Simultaneously, the task's temporary hierarchical information is appended to the hierarchical data sequence (the center vector of its cluster temporarily retains the original center vector of that cluster). When the next periodic regeneration occurs, the task will participate in the complete clustering process, and its temporary hierarchical information will be overwritten by the formal hierarchical results.

[0041] Step S2: Collect and process student interaction data to generate student ability vector V le The specific process is as follows: Step S2-1: Collect student interaction logs in real time to form a dataset Braw representing the student's original interaction behavior, including: timestamp data, used to record the time of each interaction behavior; operation type data, used to identify the category of interaction behavior, including four preset types of data: voice input, option click, scene switching, and interface navigation; interface coordinate data, which records the student's operation position in the graphical user interface in the form of pixel coordinate pairs; and content data, which records the specific content corresponding to the operation type, such as ASR transcribed text, clicked option text, and switched scene ID.

[0042] Step S2-2: Generate student interaction feature set F in The specific steps are as follows: Step S2-21: Perform continuity verification on the timestamp data and correct any detected abnormal timestamps. The preset interval threshold is preferably set to 5 seconds, and can be dynamically adjusted using sliding window statistics (e.g., 3 times the standard deviation). When correcting abnormal timestamps, it's not enough to simply remove them; relevant information about these abnormal records, such as timestamp and operation type, can also be recorded for subsequent analysis of the causes of the anomalies.

[0043] Furthermore, the breakpoint sequence can be completed to ensure that the timestamp data is monotonically increasing. Operation type data is categorized into four preset categories—voice input, option clicks, scene switching, and interface navigation—using a log processing program. Records that cannot be categorized are marked as unknown operations.

[0044] When completing breakpoint sequences, more factors can be considered. For example, if the data exhibits periodicity or a trend, these characteristics can be incorporated into the interpolation process, rather than simply using linear interpolation. Furthermore, different interpolation strategies can be implemented, selecting the appropriate method based on the characteristics of the data. For instance, linear interpolation may be sufficient for relatively stable data, while more advanced interpolation methods such as cubic spline interpolation can be considered for data with drastic changes.

[0045] After completing the timestamp data, it is necessary to strictly verify again whether the timestamp data is monotonically increasing. If it is found that the data does not meet the monotonically increasing condition, adjustments should be made or an error should be reported in a timely manner.

[0046] The interface coordinate data is categorized into predefined functional areas, including at least the script prompt area and the product information display area, using a coordinate mapping algorithm. The cumulative dwell time and operation frequency of each area are calculated to analyze the distribution of student attention. Content data is generated based on the operation type: if it is voice input, the text converted by the Automatic Speech Recognition (ASR) engine and the confidence score (range [0,1]) are stored, and the text is segmented, stop words are removed, and a word sequence is generated; if it is an option click, the identifier of the clicked option and the displayed text are recorded; if it is a scene switch, the name or ID of the target scene is recorded.

[0047] In this embodiment, the proposed coordinate mapping algorithm operates as follows: First, the collected raw pixel coordinates are normalized to convert absolute coordinates into relative coordinates relative to the current device screen or application window size, thereby eliminating deviations caused by different display resolutions. Second, based on the layout position and boundary coordinates of predefined functional areas, such as the prompt area, product information display area, and progress control area, a matching rule for coordinates to functional areas is established. By determining which functional area's boundary polygon the normalized coordinate point falls within, the operation position is classified. Finally, a time window analysis is performed on the sequence of coordinate points that continuously fall into the same functional area within the same session to calculate the student's cumulative dwell time and active interactions, such as click and swipe frequency, in that area.

[0048] In this embodiment, for ease of understanding, the four preset categories are: voice input, which corresponds to the student's real-time voice interaction, such as oral responses in training tasks in the script database preset in the system; option click, which corresponds to the student selecting preset structured dialogue options in the graphical interface; scene switching, which corresponds to the student switching between different training scenarios, such as clothing stores and restaurants along the street; and interface navigation, which corresponds to page jumps, tab switching, and navigation behaviors within the system.

[0049] Step S2-22: Based on the cleaned data, generate a set of student interaction features F. in ,include: Based on the system, the standard response text corresponding to the current training task is extracted from the pre-set script database of the current training task, and then encoded into a standard semantic vector V using the pre-trained language model BERT. std The student's speech recognition text is also encoded into a student semantic vector V. user The semantic matching degree is calculated using the following cosine similarity formula, and the result is normalized to the [0,1] interval to serve as an indicator for evaluating the semantic fit between the student's response and the standard script: ; Extract a set of keywords K from the pre-built sales knowledge graph G that are related to the current training skill points, such as icebreakers. Encode each keyword into a keyword vector k using the same BERT model described above. j Calculate the student semantic vector V. user With each keyword vector k j Euclidean distance d j Take the minimum distance d min =min(d1,d2,…,d n ), where n is the number of distances; and based on a preset threshold th, the minimum distance is converted into an intent relevance score to assess whether the trainee's response touches upon the core business intent: The preset threshold th is set by analyzing the distribution of historical data, and is generally taken as the 75th percentile of all historical minimum distances.

[0050] Response efficiency characteristics, calculating the time interval Δt between adjacent operations. i =t i -t i-1 Statistical average response time Δt avg and the standard deviation of response time Δt std To assess the fluency and hesitation of trainees' responses.

[0051] Operational accuracy features: For option click operations, compare the student's selected option with the standard recommended option for the current step, and calculate the matching accuracy ac = number of matches / total number of clicks.

[0052] In this embodiment, the proposed current training skill point represents a specific sales ability node extracted from the sales knowledge graph G that the trainee is currently training, used to define the expected goals and assessment standards for this round of coaching. In practical applications, the system uses the assessment requirements associated with the current training skill point, such as keyword sets and standard scripts, to analyze, match, and quantify the trainee's original interactive behavior, thereby determining the extent to which the trainee's behavior conforms to or achieves the training goal of that skill point.

[0053] Step S2-3: Standardize the generated feature values ​​to generate the trainee's ability vector V. le The specific steps are as follows: For semantic matching degree S sem The value calculated and normalized using cosine similarity is used as the standardized feature value: for the intention matching degree S in The minimum Euclidean distance, which has been converted into a fraction based on a preset threshold th, and normalized to a value in [0,1], is used as the standardized feature value; for response efficiency features (including average response time Δt) avg and the standard deviation of response time Δt std Normalization is performed using a preset threshold: The maximum permissible average response time threshold Ta is set. max and the maximum permissible standard deviation threshold T s max Calculate the standardized average response time efficiency E avg and standardized response stability E std After ensuring that both fall within the [0,1] interval, this value is used as the standardized feature value; for the operation accuracy feature ac, it is directly used as the standardized feature value. The aforementioned preset thresholds are generally set based on the 95th percentile of the historical response times of all trainees.

[0054] Finally, the trainees' ability vector V is formed by combining them in a fixed order. le =[Ssem ,S in E avg E std At this point, the student's original interaction data is mapped into a unified capability state vector.

[0055] In one embodiment of the present invention, it should be noted that the evaluation model of the sales training system in the prior art often relies on a statically constructed and fixed knowledge system, i.e., a static knowledge graph, to analyze the performance of trainees. This leads to a disconnect between the evaluation standards and the rapidly changing sales practice and individualized learning patterns.

[0056] Step S3: Based on the historically accumulated student ability vectors and practice path sequences, asynchronously and dynamically adjust the learning conversion cost weights between skill nodes in the knowledge graph G.

[0057] In this embodiment, a dynamic learning mechanism is constructed to optimize the pre-built sales knowledge graph G, thereby improving the accuracy of the evaluation. The specific steps are as follows: S3-1: Instantiate a topological graph G based on a predefined script database, including nodes in three layers: knowledge points, skills, and scenes, and perform static quantization on the connection edges between nodes: Use a graph attention network (GAT) to calculate semantic association weights, and the attention coefficients are calculated using the following formula: ; In the formula, , Let be the feature vectors of nodes i and j, W be the weight matrix, and a be the learnable vector. This represents vector concatenation, where N(i) is the set of neighboring nodes of node i.

[0058] The attention coefficient α obtained through calculation ij The features of neighboring nodes are weighted and aggregated to generate the final representation vector of node i containing contextual semantics. Based on this, the semantic association degree between any two nodes i and j in the graph is further calculated. This reflects the semantic closeness between nodes within the topological structure of the knowledge graph G. ; In the formula, Let be the final representation vector of node i, containing the node's own features and its local topological context information. Let be the final representation vector of node j. Let v be the representation vector of any node k, and let v represent the set of all nodes in the knowledge graph G, including all knowledge points, skills and scene nodes.

[0059] At the same time, based on expert experience, an initial learning transformation cost weight w is assigned to each edge in the graph. cin As a baseline for dynamic adjustment, the preset learning difficulty from the source skill node to the target skill node is quantified, and each skill node is labeled with an assessment feature vector F extracted from the script database and formalized. node =[w1,w2,...,w m ], where m is 5, and the dimension is set to V. le Consistency corresponds to the ideal requirements in five aspects: semantics, intent, efficiency, stability, and accuracy.

[0060] S3-2: The system periodically (e.g., daily or after accumulating a certain number of student interaction records) extracts historically accumulated V data from the database. le And its corresponding practice path sequence. This historical data does not include the current interaction that the student is currently engaged in, but rather records of all completed coaching tasks up to the end of the previous cycle.

[0061] S3-3: Extract sub-vectors of specific skill nodes based on weighted attribution and update edge weights.

[0062] Since a student's performance in any training task is often the result of the combined effect of multiple skills, directly attributing the student's overall ability vector for that task to a single skill node B would lead to evaluation distortion. Therefore, this invention constructs a weighted attribution mechanism based on task-skill association strength and graph semantic association, enabling the separation of sub-vectors corresponding to specific skill nodes B from the overall ability vector. Specifically: S3-3-1: Determine the strength of the association between tasks and skill nodes.

[0063] Through large-scale expert evaluation or in-depth analysis of massive task data, the degree of correlation between each training task and various skill nodes in the knowledge graph is clarified. This correlation strength reflects the task's dependence on the corresponding skill node. For example, a training task about product sales may have a high correlation strength with skill nodes such as product knowledge explanation and customer needs mining. The correlation strength between the task and skill node B is denoted as... (All correlation strengths have been normalized to the [0,1] interval).

[0064] S3-3-2: Obtain the overall capability vector V le .

[0065] After the student completes the training task, the system will generate an overall ability vector V with fixed dimensions. le Each dimension of this vector corresponds to a specific capability dimension, such as semantic understanding, intent recognition, efficiency, stability, and accuracy, comprehensively reflecting the trainee's overall performance in the training task.

[0066] S3-3-3: As in S3-1, calculate the semantic correlation between skill nodes. .

[0067] S3-3-4: Extracting the subvector V of a specific skill node le (B)

[0068] 1. When extracting sub-vectors of specific skill nodes, firstly, the correlation strength between the current training task and each skill node, calculated in step S3-3-1, is used. Based on a preset activation threshold θ, a set of skill node contexts activated by the current task is constructed. To ensure that the target node always participates in the calculation, the target node is forcibly included in this set, while irrelevant nodes with a correlation strength lower than the threshold are filtered out. The preset activation threshold is set based on business experience or historical data distribution.

[0069] 2. Based on the global semantic association matrix of the knowledge graph calculated in step S3-3-3 Extract the semantic association between the target node B and each other node j except itself from the above activation set.

[0070] 3. To avoid the problem that the traditional averaging strategy is easily diluted by weakly correlated nodes, a max pooling strategy is used here to calculate the contextual semantic aggregation degree. That is, only the maximum semantic correlation between the target node and other nodes in the set is taken as the metric.

[0071] If the activation set contains only the target node itself and no other nodes, then the context semantic aggregation degree is directly applied. Assigning a value of 1, taking the maximum value, means that the semantic support of the target node depends entirely on itself.

[0072] 4. Set the pre-defined correlation strength for this task. The dynamic calculation of contextual semantic aggregation degree mentioned above (Representing a graph structure environment) a weighted fusion is performed to obtain a comprehensive weight coefficient. The fusion formula is: ; in, and It is a weighting parameter used to balance the influence of task relevance strength and semantic relevance, and + =1, which can be adjusted according to the actual situation, and is generally set based on historical data.

[0073] Furthermore, based on the comprehensive weighting coefficient From the overall capability vector Vle(V le =[Ssem ,S in E avg E std Extract the subvector V corresponding to skill node B from [ac]). le (B), that is: .

[0074] 5. Employ a mapping function, such as the weighted average method, to represent the trainee's actual ability subvector V at the target node B. le (B) Mapped to a scalar evaluation score Score(V) le (B)). Assume the weights of each capability dimension are ω1, ω2, ..., ω5, and ,but: Score(V le (B))= V le (B)i; Among them, V le (B)i represents the subvector V le The value of the i-th dimension of (B).

[0075] S3-4: Based on the update rules, adjust the learning transformation cost weights of the edges from skill node A to skill node B: ; In the formula, η is the learning rate, R is the expected performance benchmark, which can be set to 0.75, and S... core (V le (B) represents the trainee's actual ability vector V at the target node B. le (B) Mapping to a function of scalar evaluation scores, such as a weighted average. It should be noted that, to prevent over-adjustment of weights, w... cnew Truncation is performed, limiting the range to [0, 2w]. cin Within the range.

[0076] Through the above steps, sub-vectors of specific skill nodes can be extracted from the overall ability vector, and these sub-vectors can be used to dynamically adjust the knowledge graph, enabling the knowledge graph to be continuously optimized based on the trainees' actual performance and improving the accuracy of the assessment.

[0077] Understandably, if a student smoothly transitions from practice at node A to node B and performs well, specifically demonstrated by a high score, then... A negative value reduces the weight of the path, while a positive value increases it, thus continuously optimizing the graph using feedback data from all students' practice.

[0078] It is important to note that this knowledge graph update process is independent of the real-time task recommendation in step S4. When performing optimal task matching in step S4, the version of the knowledge graph after the previous asynchronous update is used, not the version containing the current student's interaction data.

[0079] Step S4: Based on the current student's ability vector V le The system matches the central features of each level cluster in the hierarchical data sequence to select preliminary candidate levels, and then matches them with the specific training task features within the candidate levels to calculate the next optimal training task.

[0080] Step S4-1: Based on the system, calculate the student's current ability vector V. le It is viewed as a 1×5 (1 row, 5 columns) row vector, where each column represents a specific capability index. Simultaneously, the cluster center vector C' of all tasks is extracted from the two-dimensional data sequence. k This forms an N×5 matrix, where N is the total number of training tasks, and each column corresponds to the same semantic dimension of the training task level center feature vector. Perform similarity matching: For the j-th capability metric, such as S sem Semantic matching score is calculated by comparing its value with the set of N values ​​in the j-th column of the fourth (last) subsequence.

[0081] In this embodiment, the comparison method involves analyzing the overlap of numerical distributions between two datasets. For example, for each ability dimension, the system first calculates the minimum and maximum values ​​of all task level centers within that dimension, forming a numerical interval. Then, the student's ability value for that dimension is placed within this interval, and its relative position is calculated. If the student's ability value is exactly equal to the minimum value, the weight is set to zero; if it is equal to the maximum value, the weight is set to one; if it is in the middle, the weight is determined according to the ratio of the ability value to the interval length—that is, the closer the ability value is to the maximum value, the closer the weight is to one, and vice versa. This mapping relationship is linear, ensuring monotonically increasing and a uniquely determined result. Through this operation, a dimension-matching weight ω is calculated for each ability indicator j. j This allows for the quantification of the correlation between the trainee's ability indicators and the corresponding feature dimensions of the training task level center.

[0082] Finally, a set of weights [ω1,ω2,...,ω5] is output to determine the contribution ratio of each capability index to the final similarity calculation in subsequent row matching.

[0083] Step S4-2: Using the student's ability vector V le As a benchmark, it is compared and matched one by one with each row of training task data in the two-dimensional data sequence: First, calculate Vle C' of each row of the two-dimensional data sequence k The weighted cosine similarity between vectors is calculated by matching the obtained dimensions with the weights ω. j Substituting these values, we weight each dimension when calculating the vector dot product and magnitude, thus obtaining a level matching score that reflects the degree of matching between the student's ability and the central level of the training task. Based on this score, we select the top k training task rows with the highest scores as candidates. At this point, the top k training task rows represent the training task levels that are closest to the student's current ability profile.

[0084] In this embodiment, k is dynamically set according to the size of the candidate pool, and is taken as 10% of the total number of tasks.

[0085] Secondly, among these k candidate training task rows, V is further calculated. le The combined feature vector F of each training task itself task The cosine similarity between them is used to obtain the matching score of the training task characteristics.

[0086] Finally, based on the similarity score of the level matching and the matching score of the tutoring task characteristics (e.g., the comprehensive score = 0.7 × level matching score + 0.3 × task characteristic matching score), the k candidate tutoring tasks are sorted, and the task with the highest comprehensive score is determined as the optimal matching task. The complete data of the optimal matching tutoring task row is output, including its Task. ID , belonging to L evel And the corresponding C 'k Vectors are used to match and locate students' abilities with specific training tasks.

[0087] Step S5: Assign the next optimally matched coaching task to the student, and update the student's ability vector V based on the new interaction data after the student completes the task. le The specific implementation steps include: The task that determines the optimal matching coaching task. ID Assign the AI ​​tutor to the current student, start the corresponding AI tutoring session, and record the complete matching logic as a log entry, including the student's ability vector V used. le The matched training partner task level L evel The similarity scores and final ranking results are linked to the student's learning profile. After the student completes the recommended tutoring task, the system collects the new interaction data generated during the task and inputs it into step S2 for processing, thereby generating an updated student ability vector V. le' This indicates possible changes in the trainee's ability status.

[0088] Subsequently, based on V le'The system uses the latest two-dimensional data sequence, which is updated as new training tasks are added to the database or through periodic clustering, to achieve a new round of matching and recommendation.

[0089] In this embodiment, the matching recommendation results include at least the determined next optimal coaching task. ID and its corresponding level label; the historical progress trajectory includes at least the trainee's performance on each ability indicator, based on a historical sequence of V. le Trend data that changes over time, obtained through vector calculation.

[0090] The system will visualize the matching recommendation results, the student's current ability vector, and historical progress trajectory in the form of a dynamic dashboard.

[0091] As a second aspect of the present invention, a hierarchical allocation and path optimization system for sales coaching tasks is proposed, comprising: an offline hierarchical module, which pre-extracts multi-dimensional features of each coaching task based on a knowledge graph G and historical data of coaching tasks, calculates the comprehensive similarity between tasks, divides all coaching tasks into multiple task level clusters corresponding to different difficulty levels through a clustering algorithm, and generates a hierarchical data sequence containing the level to which the task belongs and its features; and a feature extraction module, which collects and processes student interaction data to generate a student ability vector V. le The knowledge graph optimization module asynchronously and dynamically adjusts the learning conversion cost weights between skill nodes in the knowledge graph G based on historically accumulated student ability vectors and practice path sequences; the task matching module, based on the current student ability vector V... le The system matches the central features of each level cluster in the hierarchical data sequence to select preliminary candidate levels, and then matches these candidate levels with the specific training task features to calculate the optimal next training task. The allocation and update module assigns the optimal next training task to the student, and updates the student's ability vector V based on new interaction data after the student completes the task. le .

[0092] The sales coaching task hierarchical allocation and path optimization device provided in this embodiment of the invention can execute the sales coaching task hierarchical allocation and path optimization method provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0093] The system also includes one or more processors and memory.

[0094] The memory is used to store operable instructions that, when executed by the one or more processors, cause the one or more processors to perform operations, including the flow of the sales coaching task tiered allocation and path optimization method of the foregoing embodiments, especially... Figure 1 The flowchart of the method is shown.

[0095] Other aspects disclosed in this invention also propose a computer-readable medium for storing software including instructions executable by one or more computers, which, upon execution, cause the one or more computers to perform operations, including the hierarchical allocation and path optimization process for sales coaching tasks described in the foregoing embodiments, particularly... Figure 1 The flowchart of the method is shown.

[0096] It should be recognized that embodiments of the present invention may be implemented or carried out by computer hardware, a combination of hardware and software, or by computer instructions stored in a non-transitory computer-readable storage medium.

[0097] The method can be implemented using standard programming techniques, including a non-transitory computer-readable storage medium configured with a computer program in the computer program, wherein the storage medium is configured such that the computer operates in a specific and predefined manner.

[0098] Each program can be implemented in a high-level procedural or object-oriented programming language to communicate with the computer system; however, if required, the program can be implemented in assembly or machine language.

[0099] In any case, the language can be either compiled or interpreted.

[0100] Furthermore, for this purpose, the program can run on programmed application-specific integrated circuits.

[0101] The processes described herein (or variations and / or combinations thereof) can be executed under the control of one or more computer systems configured with executable instructions, and can be implemented by hardware or a combination thereof as code (e.g., executable instructions, one or more computer programs, or one or more applications) that commonly executes on one or more processors. The computer program includes a plurality of instructions executable by one or more processors.

[0102] Furthermore, the method can be implemented in any suitable type of computing platform, including but not limited to personal computers, minicomputers, mainframes, workstations, networked or distributed computing environments, standalone or integrated computer platforms, or in communication with charged particle tools or other devices.

[0103] Various aspects of the present invention can be implemented in machine-readable code stored on a non-transitory storage medium or device, whether portable or integrated into a computing platform, such as a hard disk, optical read and / or write storage medium, RAM, ROM, etc., such that it can be read by a programmable computer, and when the storage medium or device is read by the computer, it can be used to configure and operate the computer to perform the processes described herein.

[0104] Furthermore, machine-readable code, or parts thereof, can be transmitted via wired or wireless networks.

[0105] When such media includes instructions or programs that combine with a microprocessor or other data processor to implement the steps described above, the invention described herein includes these and other different types of non-transitory computer-readable storage media.

[0106] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for hierarchical allocation and path optimization of sales coaching tasks, characterized by: Includes the following steps: Based on the knowledge graph G and historical data of the training tasks, multi-dimensional features of each training task are extracted in advance and the comprehensive similarity between tasks is calculated. All training tasks are divided into multiple task level clusters corresponding to different difficulty levels through clustering algorithms, and a hierarchical data sequence containing the level to which the task belongs and its features is generated. Collect and process student interaction data to generate student ability vector V le ; Based on the historically accumulated student ability vectors and practice path sequences, the learning conversion cost weights between skill nodes in the knowledge graph G are asynchronously and dynamically adjusted. Based on the current student ability vector V le The system matches the central features of each level cluster in the hierarchical data sequence to select preliminary candidate levels, and matches them with the specific training task features within the candidate levels to calculate the next training task that best fits the system. The next optimally matched training task is assigned to the student, and the student's ability vector V is updated based on the new interaction data after the student completes the task. le； The dynamic adjustment of the knowledge graph G includes the following steps: Instantiating a knowledge graph G containing three layers of nodes—knowledge points, skills, and scenarios—based on the system script database; calculating the semantic association weights between nodes using a graph attention network; setting an initial learning conversion cost weight for each edge; and labeling skill nodes with assessment feature vectors extracted from the script; extracting skill node identifiers corresponding to consecutively completed coaching tasks from the student's interaction log in chronological order to form a practice path sequence; periodically extracting historically accumulated student ability vectors and corresponding practice path sequences from the database; dynamically adjusting the learning conversion cost weights on corresponding edges according to update rules, and truncating the calculated new weights to limit them within a preset reasonable range; dividing all coaching tasks into multiple task level clusters using a clustering algorithm, including: for each coaching task, parsing the dialogue script into atomic coaching units and encoding the original semantic vector using a pre-trained language model; simultaneously, extracting assessment feature vectors of associated skill nodes from the knowledge graph G and calculating the historical average performance vector; and fusing the original semantic vector, assessment feature vector, and historical average performance vector to generate a comprehensive feature vector F for the task. task Construct a similarity matrix W, and for any two training tasks i and j, calculate the comprehensive similarity W. ij Set a threshold for association strength, generate a binary association matrix, and screen and remove isolated tasks that are not associated with other tasks; traverse the association matrix, recursively expand and construct preliminary candidate clusters through adjacency relationships, iteratively calculate the cluster center similarity and perform fine-grained merging checks until the cluster structure is stable, and each cluster in the final candidate level cluster set C represents a specific training task difficulty level.

2. The optimization method according to claim 1, characterized in that: Generate the student ability vector V le The steps include: Real-time collection of student interaction logs in the graphical user interface forms a raw dataset including timestamp data, operation type data, interface coordinate data, and content data; the raw dataset is cleaned and structured; and a set of student interaction features is generated based on the cleaned data. The set of student interaction features includes at least: the semantic matching degree between student responses and standard dialogue, calculated based on a pre-trained language model; The degree of intent alignment is calculated based on the distance between trainees' responses and keyword vectors in the sales knowledge graph; Response efficiency features calculated based on timestamp data, and operation accuracy features calculated based on operation type data matching. The semantic matching degree, intent consistency degree, standardized response efficiency feature, and operation accuracy feature are combined in a fixed order to form the student ability vector V. le .

3. The optimization method according to claim 2, characterized in that: The specific steps for generating hierarchical data sequences include: Calculate the overall mean of the comprehensive feature vector of all training tasks in each cluster of the final candidate level cluster set C on the preset difficulty-related dimensions, and sort all clusters in ascending order according to the overall mean, and assign ordered level labels in sequence. Establish a mapping relationship between each level label and the center vector of the corresponding cluster; traverse all training tasks, constructing a two-dimensional data sequence with each task as a row, where each row of data includes: the unique identifier of the training task, its level label, and its own comprehensive feature vector F. task and the center vector of the cluster to which it belongs; The hierarchical data sequence is periodically regenerated in response to updates to the knowledge graph G.

4. The optimization method according to claim 3, characterized in that: The specific matching calculation process for calculating the next optimally matched training task includes: Based on the current student ability vector V le For each dimension, the distribution relationship between the numerical values ​​and the set of values ​​of all cluster center vectors in the hierarchical data sequence in the corresponding dimension is analyzed, and a dimension matching weight is calculated for each ability dimension; using the dimension matching weight, the trainee's ability vector V is calculated. le The weighted cosine similarity between the data and the cluster center vector recorded in each row of the hierarchical data sequence is used as the hierarchical matching score, and the top k tasks with the highest scores are selected as primary candidates; for the primary candidate tasks, the student ability vector V is calculated. le Combined with the feature vector F of each task itself task The cosine similarity between the two is used as the task feature matching score; for each candidate task, the level matching score and the task feature matching score are fused according to a preset weight to obtain a comprehensive score, and the task with the highest comprehensive score is determined as the next best-fit training task.

5. The optimization method according to claim 1, characterized in that: The specific steps for screening and removing isolated tasks that are not related to other tasks during clustering are as follows: based on the association strength threshold θ link Transform the comprehensive similarity matrix W into a binary correlation matrix A, where if W ij ≥θ link Then A ij =1, otherwise A ij =0; For each row i of the binary correlation matrix, calculate the sum of correlation degrees Sum. i If the sum of the correlation degrees of a certain row i is Sum i If the value is 0, the corresponding training task in the corresponding row is determined to be an abnormal training task, and it is removed from the main task set and stored in the abnormal training task set.

6. The optimization method according to claim 1, characterized in that: The method further includes: matching recommendation results and the student's current ability vector V. le The historical progress trajectory across various capability dimensions is visualized in the form of a dynamic dashboard.

7. The optimization method according to claim 2, characterized in that: The semantic matching degree calculation process is as follows: extract standard response text from the preset script database corresponding to the current training task, and encode it into a standard semantic vector through a pre-trained language model; encode the student's speech recognition text into a student semantic vector through the same pre-trained language model; calculate the similarity between the standard semantic vector and the student semantic vector through the cosine similarity formula, and normalize the result to the [0,1] interval. The obtained value is the semantic matching degree. The steps for processing the original dataset are as follows: verify the continuity of timestamps and fill in the breakpoints, classify the operation types into preset categories, and classify the interface coordinates into predefined functional areas through a coordinate mapping algorithm to analyze the attention distribution.

8. A hierarchical allocation and path optimization system for sales coaching tasks, based on the hierarchical allocation and path optimization method for sales coaching tasks according to any one of claims 1 to 7, characterized in that: include: The offline grading module extracts multi-dimensional features of each training task based on the knowledge graph G and historical data of the training tasks, calculates the comprehensive similarity between tasks, divides all training tasks into multiple task level clusters corresponding to different difficulty levels through a clustering algorithm, and generates a grading data sequence containing the level to which the task belongs and its features. The feature extraction module collects and processes student interaction data to generate a student ability vector V. le ; The knowledge graph optimization module dynamically adjusts the learning conversion cost weights between skill nodes in the knowledge graph G asynchronously based on the historically accumulated student ability vectors and practice path sequences. The task matching module is based on the current student's ability vector V. le The system matches the central features of each level cluster in the hierarchical data sequence to select preliminary candidate levels, and matches them with the specific training task features within the candidate levels to calculate the next training task that best fits the system. The assignment update module assigns the next best-fit training task to the student and updates the student's ability vector V based on the new interaction data after the student completes the task. le .

Images