User demand mining method based on feature fusion density perception, terminal and medium

Abstract

The invention relates to the technical field of data mining and product demand analysis, and discloses a user demand mining method based on feature fusion density perception, a terminal and a medium. The method comprises the following steps: collecting product rejection data, extracting a semantic feature vector from a preprocessed data text, obtaining a user behavior sequence by using a session identifier in meta-information to generate a behavior feature vector, and generating a situation feature vector based on time and environment attributes in the meta-information; calculating the attention weight of each dimension feature, and performing weighted fusion on the semantic feature vector, the behavior feature vector and the situation feature vector by using the attention weight to generate a multi-dimensional feature representation vector; clustering the multi-dimensional feature representation vectors to obtain a plurality of clusters; and feature portrait construction and value evaluation are carried out on the cluster, a demand description document is generated, and optimization algorithm parameters are fed back according to a verification result. According to the method, the efficiency and the quality of mining potential user requirements from neglected rejection data can be remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data mining and product demand analysis, in particular to a user demand mining method based on feature fusion density perception, a terminal and a medium. BACKGROUND

[0002] In the product development process, due to inadequate design considerations and insufficient understanding of user needs, valuable data is often rejected by product recognition models. These rejected data contain important user demand information and product improvement directions. Effective mining of these data can provide valuable decision-making basis for product iteration. Currently, the processing of rejected data mainly relies on manual expert analysis, simple rule filtering and traditional clustering analysis, but these methods have obvious defects such as low efficiency, high cost, poor adaptability, etc., and are difficult to meet the demand of large-scale data analysis, which has become a bottleneck restricting the speed of product optimization.

[0003] In terms of feature analysis, existing technologies mainly focus on single-dimensional feature extraction, some pay attention to text content analysis, some focus on user behavior data, and some try simple feature combination. Although in recent years some enterprises have begun to use machine learning and deep learning technologies for semi-automatic analysis, but still need a lot of manual intervention, and the model performance is highly dependent on the quality of training data, and the generalization ability is limited. Traditional clustering algorithms such as K-means and DBSCAN are easily disturbed by noise when dealing with complex multi-dimensional data, and it is difficult to find small-scale but high-consistency data clusters, and these small data clusters often represent important emerging user needs.

[0004] In summary, the existing technology has the problems of insufficient automation, insufficient multi-dimensional feature integration, weak recognition ability of small demand, poor dynamic adaptability, and imperfect demand value evaluation mechanism in processing rejected data to mine user needs. These limitations restrict the ability of enterprises to efficiently discover and utilize valuable user needs from rejected data. Therefore, there is an urgent need for a technology that can automatically and efficiently identify and mine high-value user needs from complex multi-dimensional rejected data to support enterprises to better understand user needs and accelerate product iteration and optimization. SUMMARY

[0005] To solve the technical problems in the prior art, the present application provides a user demand mining method based on feature fusion density perception, a terminal and a medium, which can significantly improve the efficiency and quality of mining potential user needs from neglected rejected data.

[0006] To achieve the above purpose, the present application provides the following technical solutions: The present application discloses a user demand mining method based on feature fusion density perception, comprising: S1. Collect product refusal data and pre-process the data to extract meta-information of the data; S2. Extract semantic feature vectors from the pre-processed data text, use the session identifier in the meta-information to obtain user behavior sequences to generate behavior feature vectors, and generate context feature vectors based on the time and environment attributes in the meta-information; S3. Calculate the attention weight of each dimension feature, and use the attention weight to weight and fuse the semantic feature vector, the behavior feature vector and the context feature vector to generate a multi-dimensional feature representation vector; S4. Cluster the multi-dimensional feature representation vector using a density difference perception clustering algorithm, which constructs a local density field through an adaptive bandwidth kernel density estimation, defines a density difference coefficient to quantify the local variation characteristics of the density field, identifies the clustering center based on the density gradient through a climbing iteration, and divides the clustering boundary through a density saddle point detection to obtain multiple clustering clusters; S5. Construct a feature portrait and evaluate the value of the clustering cluster to generate a demand description document.

[0007] As a further improvement of the above scheme, the specific process of step S2 includes: At the semantic feature extraction level, input the text description of the refusal data into a pre-trained language model for word embedding representation to capture the basic semantic features of the text; through a deep bidirectional Transformer network, the basic semantic features are enhanced in context semantics to extract general semantic features; and then through a domain adaptation layer, the general semantic features are processed for domain specialization to obtain a semantic feature vector; At the behavior feature extraction level, the behavior sequence in the user session is obtained, each behavior event in the sequence is mapped to a vector representation, and the vector representation of the behavior sequence is processed through a time series attention network to extract a behavior feature vector; At the context feature extraction level, the environmental context information including device technical parameters, geographic location information, network environment state and use time features is obtained, and a multi-layer perception network is used for nonlinear fusion to obtain a context feature vector.

[0008] As a further improvement of the above scheme, step S4 specifically includes: S41. Construct a local density estimator, for each data point in the feature space , use an adaptive bandwidth Gaussian kernel function to calculate its local density value : ; In the formula, is a bandwidth parameter that is adaptively adjusted according to the local data distribution characteristics of point , so that the density estimation has different resolutions in different regions. is an arbitrary data point in the feature space except for the point ; is a data set consisting of all data points in the feature space; is the data set ; is the total number of data points in the data set ; is the Euclidean norm; is the adaptive bandwidth parameter, so that the density estimation has different resolutions in different regions, and the calculation formula is: ; wherein is the global reference bandwidth parameter; is the initial density value of the point obtained by the fixed bandwidth pre-estimation; is the geometric mean of the initial density values of all data points; is the sensitivity parameter, with a value range of [0, 1], which controls the adaptive degree of the bandwidth to the local density; S42. Calculate the density gradient vector and the density Hessian matrix, and the expression formulas are as follows: ; ; wherein is the gradient vector of the density function at the point , indicating the direction in which the density grows fastest; is the Hessian matrix of the density function at the point , describing the second-order curvature characteristics of the density field; is the partial derivative operator; is the local density function; are the first to the th coordinate components in the feature space, respectively; is the dimension of the multi-dimensional feature representation vector; the superscript represents vector transposition; S43. Define the density difference coefficient, which is used to quantify the local variation characteristics of the density field at the data point: ; wherein is the density difference coefficient of the point , which is used to quantify the local variation characteristics of the density field at the point; is the minimum eigenvalue of the density Hessian matrix at the point , representing the second-order variation rate of the density field in the direction of minimum curvature; S44. Perform density climbing iteration based on the density gradient vector to identify the initial clustering center; wherein, for each data point , the iterative update is performed along the direction of the density gradient: ; wherein, is the position of the point at the tth iteration; is the position of the point after the (t+1)th iteration update; is an adaptive step size parameter, which controls the moving distance of each iteration; is the density gradient vector at the point ; is the Euclidean norm of the gradient vector; The calculation formula of , is the base step size; is the global average gradient norm; The iteration termination condition is wherein is the gradient norm threshold; the maximum eigenvalue; the points in the convergence point that satisfy the local density maximum value condition are identified as the initial cluster centers, and the local density maximum value condition is that all eigenvalues of the Hessian matrix are negative; S45. Introduce a density saddle point detection mechanism to divide the cluster boundary, and the specific process is as follows: Define the saddle point judgment condition: the point is a density saddle point if and only if the following conditions are met: I. The modulus of the density gradient vector tends to zero: ; wherein, is the gradient threshold of the saddle point judgment; II. The Hessian matrix has both positive and negative eigenvalues; For the adjacent cluster centers and identified initially, search for a point that satisfies the saddle point judgment condition along the direction of the line connecting the two points; when the density saddle point s between the two density maximum value points and satisfies the following conditions, it is determined that and belong to different cluster clusters: ; wherein, is the local density value of the point , and are the local density values of the point and the point respectively; is the minimum value; Density ratio threshold parameter, controls the maximum ratio of density at saddle point to density at cluster center, value range is (0, 1); DDC threshold parameter, controls the minimum intensity of density difference at boundary.

[0009] As a further improvement of the above scheme, in step S4, hierarchical multi-resolution analysis is also performed, specifically including: Perform density estimation and clustering under multiple different bandwidth parameter settings respectively; perform stability analysis on the clustering results under each resolution, calculate the overlap degree of the clustering clusters under different resolutions; retain the clustering structure that stably exists under multiple resolutions, filter out noise clusters and pseudo clusters that only appear at a single resolution.

[0010] As a further improvement of the above scheme, in step S4, the following objective function is optimized Adjust the boundary of the clustering cluster: ; In the formula, Total number of clustering clusters; is the data point set contained in the th clustering cluster; is the center point of the th clustering cluster, i.e. the local density maximum point, is the local density value of the center point of the th clustering cluster; is the density saddle point between the th clustering cluster and the th clustering cluster; Regularization parameter, used to balance the consistency of intra-cluster density and the clarity of inter-cluster boundary; Expected lower threshold of density difference coefficient; Take the maximum value.

[0011] As a further improvement of the above scheme, in step S5, the specific process of feature image construction and value evaluation includes: S51. Extract the feature image of each clustering cluster, including: extract the high-frequency keywords in the cluster as semantic labels through the TF-IDF algorithm, count the behavior sequence pattern distribution of the data points in the cluster, and summarize the context attribute distribution features of the data in the cluster; S52. Calculate the multi-factor value score of each clustering cluster, the calculation formula is as follows: ; In the formula, is the multi-factor value score of the th clustering cluster; is the size factor, wherein For clusters The number of data points is determined, and the effect of extreme values ​​is smoothed using a logarithmic transformation. It is the cohesion factor. ,in For the first The average distance from data points within each cluster to the cluster center. Global average distance As a unique factor, , For the first Data points within each cluster; The difference factor is calculated by taking the cosine similarity between the cluster feature vector and each feature vector in the existing product feature library, and then taking the complement of the maximum similarity. ,in For the first The average eigenvector of each cluster; For the first A feature representation of an existing product function; This refers to the number of existing functions; Cosine similarity; As a business value factor, ,in For data points The value score for each user is calculated based on their historical spending, activity level, and membership level. to Let be the weight coefficients of each factor, satisfying ; S53. Based on value rating The clusters are sorted, and those with scores higher than a preset threshold are selected. Clusters are used as high-value requirement candidates to generate structured requirement description documents that include requirement background, user group characteristics, core functional points and implementation priorities.

[0012] As a further improvement to the above solution, the user demand mining method also includes: S6. Based on the verification results of the aforementioned requirements description document, optimize the algorithm parameters. The specific process includes: S61. Collect requirement verification result data, including the set of requirements that have passed verification. and the set of requirements that failed validation ; S62. Adjust the feature fusion weights based on the verification results, enhancing the weights of feature dimensions relevant to the verification pass requirement. The updated formula is as follows: ; In the formula, For the updated number dimensional feature weights, The weights before the update. For learning rate, For the first The demand corresponds to the cluster in the th... Attention score on dimensional features For the first The demand corresponds to the cluster in the th... Attention score on dimensional features and Sets and The number of elements; S63. Adjust the clustering threshold parameter based on the validation results, and dynamically update the parameter according to the difference between the validation pass rate and the target pass rate. The update formula is as follows: ; ; In the formula, and These are the density ratio threshold parameters before and after the update, respectively; and These are the DDC threshold parameters before and after the update, respectively. and These are the adjustment step sizes for β and γ, respectively; This represents the actual pass rate. To verify the pass rate of the target; S64. Based on the validation results, optimize the weight coefficients of each factor in step S52, and minimize the following loss function using gradient descent: ; The weight update formula is as follows: ; In the formula, Optimize the loss function for the weights; The learning rate; and These are the first and second pages after the update, respectively. Factor weights; For the loss function on the th Partial derivatives of each weight; S65. Add the feature vectors corresponding to the verified requirements to the historical high-value pattern library for subsequent calculation of historical similarity factors: ; In the formula, No. Dimensional historical pattern library; For the first an average feature vector of the verification passed requirements in the dimension; a verification passed requirement set newly added in this round.

[0013] As a further improvement of the above-mentioned scheme, step S1 specifically comprises: establishing an interface connection with the product rejection model, collecting product rejection data, the product rejection data including user input content, operation record, submission time and device information; performing cleaning processing on the collected product rejection data to remove invalid data, outliers and duplicate records; performing format standardization processing on the cleaned product rejection data to convert data of different sources and types into a unified structured format; extracting and labeling meta-information of the product rejection data, the meta-information including rejection timestamp, rejection reason code, user session identifier, device type identifier and user unique identifier; The specific process of step S3 comprises: aligning the semantic feature vector, the behavior feature vector and the context feature vector through a projection matrix respectively, projecting the feature vectors of different dimensions to a unified feature space; calculating attention scores for the projected feature vectors respectively, the calculation of the attention scores comprehensively considering information entropy of the feature vectors, similarity with historical high-value patterns and feature confidence; normalizing the attention scores to obtain attention weights of the features of each dimension; performing weighted summation on the three types of feature vectors using the attention weights to obtain a fused feature vector; performing layer normalization on the fused feature vector to obtain a final multi-dimensional feature representation vector.

[0014] The application further discloses a computer terminal, including a memory, a processor and a computer program stored in the memory and executable on the processor, when the processor executes the computer program, the steps of the feature fusion density perception based user demand mining method are realized.

[0015] The application further discloses a computer readable storage medium, which stores a computer program, when the processor executes the program, the steps of the feature fusion density perception based user demand mining method are realized.

[0016] Compared with the prior art, the application has the following beneficial effects: 1. The user demand mining method based on feature fusion density perception disclosed in the application breaks through the limitation of traditional methods that only focus on single-dimensional analysis by feature extraction and integration of three key dimensions of semantics, behavior and context, realizes comprehensive capture and understanding of user demand, reduces human resource consumption and analysis cost while improving the efficiency and quality of mining potential user demand from rejected data, provides reliable technical support for enterprises to find innovative opportunities from neglected data, and accelerates the product optimization iteration process.

[0017] 2. The density difference perception clustering algorithm designed in the application can solve the problem that traditional clustering methods easily misjudge small high-value demand as noise by density gradient calculation and local density peak detection, thereby improving the recognition ability of small but highly cohesive demand clusters.

[0018] 3. The application realizes full-process automation from rejected data collection to demand value evaluation, reduces manual intervention to the minimum, greatly improves the efficiency and processing scale of demand mining, and enables enterprises to respond to market changes more quickly.

[0019] 4. The application constructs a demand value evaluation system and feedback loop mechanism based on multi-dimensional indicators, so that the system can learn and improve from actual application results, maintain long-term performance advantages and value output, thereby improving product resource allocation efficiency and the implementation ratio of high-return demand. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 The flowchart of the user demand mining method based on feature fusion density perception in embodiment 1 of the application.

[0021] Figure 2 The architecture diagram of the user demand mining method based on feature fusion density perception in embodiment 1 of the application.

[0022] Figure 3 The clustering result visualization (t-SNE dimension reduction) diagram in embodiment 1 of the application.

[0023] Figure 4 The value score distribution histogram in embodiment 1 of the application.

[0024] Figure 5 The structural schematic diagram of the computer terminal in embodiment 2 of the application. DETAILED DESCRIPTION

[0025] The technical solutions in the embodiments of the present application will be clearly and completely described with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative effort belong to the scope of the present application.

[0026] Embodiment 1 In view of many challenges existing in current refusal data analysis, such as single feature extraction, insufficient clustering algorithm for identifying minority needs, low efficiency of artificial analysis and the like, the present application provides a user new demand automatic mining method based on feature fusion density perception. The method realizes accurate identification and evaluation of high-value user needs in refusal data by designing a multi-dimensional feature representation mechanism and an improved density clustering algorithm.

[0027] Referring to Figure 1 The embodiment provides a user demand mining method based on feature fusion density perception, which comprises steps S1-S5.

[0028] S1. Collect product refusal data, clean and format standardize the data, and extract the meta information of the data.

[0029] The step S1 of the present application relates to a comprehensive data collection and preprocessing process. First, an interface connection with a product refusal model is established, and product refusal data is collected in real time or periodically in batches. The product refusal data includes but is not limited to the following environmental context information data: (1) user input content: including user-submitted text queries, voice-transcribed text, image description text and other original input information; (2) operation record: including the interface operation sequence of the user before and after submitting the request, such as clicking, sliding, inputting, deleting, retrying and the like; (3) submission time: including the accurate timestamp of request initiation, user session start time, historical access time distribution and other time dimension information; (4) device information: including device type, operating system version, application version, screen resolution, network type, geographic location and the like.

[0030] After collection is completed, multi-level data preprocessing operations are performed: first, data cleaning is performed, specifically including: (1) identifying and removing invalid data: identifying and removing blank input, garbled input, test data and other records without actual meaning; (2) outlier processing: detecting and processing outliers in numerical fields, for fields such as operation duration and request frequency, using the interquartile range (IQR) method to identify outliers, and deciding to remove or correct according to business rules; (3) Duplicate record removal: Based on the combination key of user identification, request content, and timestamp, identify and merge duplicate submitted records, and retain the most complete data version.

[0031] Then, format standardization processing is performed to convert data of different sources and types into a unified structured format, specifically including: (1) Text standardization: Unified character encoding into UTF-8 format, performing full-angle half-angle conversion, case normalization, special character cleaning, and other text normalization operations; (2) Time format unification: Convert time stamps of different time zones and formats into the ISO8601 standard format of UTC time; (3) Numerical normalization: Standardize continuous numerical fields to eliminate dimensional differences; (4) Category coding: Adopt a unified coding scheme for discrete fields and establish a globally consistent category mapping table.

[0032] Finally, the meta-information of the data is extracted and labeled, including: (1) Recognition timestamp: Records the exact time when product recognition determination occurs; (2) Recognition reason code: Labels the specific reason category leading to recognition, such as pre-defined categories like "out of business scope", "intent recognition failure", "entity extraction failure", "knowledge base no match", etc.; (3) User session identifier: Associates all interaction records of the same user in the same session, supporting subsequent behavior sequence analysis; (4) Device type identifier: Labels the device type of the request source, such as mobile, PC, smart terminal, etc.; (5) User unique identifier: Assigns each user a desensitized unique identifier under the premise of meeting privacy protection requirements.

[0033] After the above preprocessing process, the original recognition data is converted into structured and standardized data records, each record containing original input content, complete context information, and rich meta-information labels, providing high-quality data input for subsequent multi-dimensional feature extraction.

[0034] S2. Extract semantic feature vectors from pre-processed data texts, use session identifiers in meta-information to obtain user behavior sequences to generate behavior feature vectors, and generate context feature vectors based on time and environmental attributes in meta-information.

[0035] The three-dimensional feature extraction framework is adopted to systematically capture the multi-element information of the refusal data, and the three-dimensional feature extraction mechanism ensures that the expression form of user demand can be comprehensively captured from three complementary angles of semantic expression, behavior mode and use scene. At the semantic feature extraction level, a bidirectional Transformer architecture optimized for refusal scene is used for deep semantic modeling. Specifically, the following three sub-steps are included: (1) Word embedding representation The text description of the input refusal data , wherein is the th word token in the text, is the length of the word token sequence. First, word embedding representation is performed by a pre-trained language model: ; In the formula, is the word embedding matrix, and the dimension is ; is the embedding vector of the th word token; is the word embedding dimension, which is in the embodiment. The word embedding layer captures the basic semantic features of the text and maps discrete word token symbols to continuous vector space.

[0036] (2) Contextual semantic enhancement

[0037] Subsequently, contextual semantic enhancement is performed by a deep bidirectional Transformer network: ; In the formula, represents a bidirectional Transformer encoder stacked by layers, and is the general semantic feature vector, is the hidden layer dimension. The bidirectional Transformer simultaneously models the forward and backward context dependencies of the text through the self-attention mechanism, effectively extracting the user intent and demand information implicitly contained in the text. Specifically, the [CLS] position vector of the last layer output of the Transformer or the average pooling of all position vectors is taken as the semantic representation of the entire text.

[0038] (3) Then, the application domain adaptation layer is applied to the domain-specific processing of the general semantic features: ; In the formula, is the semantic feature vector after domain adaptation, is the semantic feature dimension; is a dimension reduction projection matrix, is a bottleneck layer dimension, ; is an adaptation layer projection matrix; and is a bias vector; is a nonlinear activation function, and the present embodiment adopts a GELU activation function. The domain adaptation layer makes the model more sensitive to the language expression of a specific business domain by fine-tuning on data of the specific business domain, and can identify terms and expression patterns specific to the domain.

[0039] In the aspect of behavior feature extraction, the system retrieves and associates the complete behavior sequence of the same user in the same session from the behavior log database based on the user session identifier in the meta information. Let the user behavior sequence be , where is the length of the behavior sequence, and each behavior event is represented as an attribute tuple: ; In the formula, is the behavior type, and the value range includes predefined behavior categories such as click, swipe, input, delete, submit, return, retry, etc.; is the timestamp of the behavior trigger; is the behavior duration; is the target object identifier of the behavior action; is the additional parameter of the behavior, such as input content, swipe direction, etc.

[0040] Each behavior event is mapped to a vector representation by a special encoding function: ; In the formula, is the vector representation of the th behavior event, is the behavior embedding dimension; is a behavior type embedding function that maps discrete behavior types to dense vectors; is a time encoding function that encodes the periodicity of the timestamp using a sine position encoding method; is a duration encoding function that performs linear mapping after logarithmic transformation of the duration; is a target object embedding function; represents vector concatenation operation.

[0041] Subsequently, the behavior sequence is processed using a time sequence attention network: ; The specific calculation process of the time sequence attention network is as follows: ; ; wherein, is the behavior feature vector; is the attention weight of the th behavior event; is the time interval from the th behavior event to the current time; , is the learnable weight matrix, is the attention hidden layer dimension; is the attention score vector; is the bias term. The time sequence attention network pays particular attention to the time sequence relationship and conversion law of the behavior pattern, and gives higher weight to the recent behavior, extracts the time sequence features of the user operation habits and intention expression.

[0042] At the level of context feature extraction, the system comprehensively analyzes the user's environmental context information based on the time attribute and environmental attribute in the meta information. The context information contains four dimensions of data, namely device technical parameters, geographic location information, network environment state and use time characteristics, wherein: the device technical parameters cover device type, operating system, screen size, memory capacity, etc.; the geographic location information includes country, city, latitude and longitude, etc.; the network environment state includes network type, network delay, bandwidth, etc.; the use time characteristics include time of day, day of the week, whether it is a holiday, etc. For the context data of each dimension, the system respectively performs feature encoding processing: for category type features, an embedding encoding method is used to map discrete categories to dense vectors; for numerical type features, after normalization, linear mapping is used to convert them into vectors of the same dimension. Through the above encoding methods, the system respectively processes the context data of the four dimensions to obtain the device feature vector , the geographic feature vector , the network feature vector and the time feature vector . Then, the encoded vectors of each dimension are spliced, and a multi-layer perception network is used for non-linear fusion to learn the interaction between different context features, and finally generate the context feature vector representing the user's use scenario.

[0043] S3. The semantic feature vector, the behavior feature vector and the context feature vector are adaptively weighted and fused to generate a multi-dimensional feature representation vector.

[0044] First, to ensure the comparability of different dimensional features, the three types of feature vectors are dimensionally aligned and projected: ; ; ; where, , , is the projected feature vector, is the unified fusion feature dimension; , , is the projection matrix of each dimension.

[0045] Then, the attention score of each feature vector is calculated to evaluate its relative importance in the current sample. The calculation of the attention score considers three factors: (1) Information entropy factor: measures the information richness of the feature vector ; where, is the normalized feature component; is a small constant to prevent numerical underflow. The higher the information entropy, the more uniform the feature distribution, and the greater the information quantity. and are calculated in the same way.

[0046] (2) Historical similarity factor: measures the similarity between the current feature and the historical high-value pattern ; where, is the prototype vector of the th historical high-value semantic pattern; is the number of historical patterns; is the cosine similarity function. and are calculated in the same way.

[0047] (3) Confidence factor: measures the confidence of the feature extraction module on the current sample ; where, and are the parameters of the confidence evaluation network; is the Sigmoid activation function, which maps the confidence to the interval.

[0048] Integrating the above three factors, the original attention score of each dimension is calculated: ; ; ; where, , 、 is the weighing coefficient of information entropy, historical similarity and confidence, satisfying The original attention score is normalized to obtain the attention weight of each dimension: ; ; ; The attention weight satisfies the normalization condition: This calculation process ensures that the weight distribution reflects both the current sample characteristics and historical experience rules.

[0049] According to the calculated attention weight, a weighted fusion operation is performed: ; wherein, is the feature vector after weighted fusion.

[0050] To ensure the numerical stability and comparability of the fused features, the system further applies layer normalization processing: ; wherein, is the final multi-dimensional feature representation vector; is the vector mean; is the vector variance; and are learnable scaling and offset parameters; is a small constant to prevent division by zero. Layer normalization processing eliminates the problems of dimension difference and inconsistent numerical range.

[0051] The advantage of this adaptive fusion mechanism is that it can automatically adjust the feature weight for different types of rejection data: it gives higher weight to semantic features for data with clear semantic expression, strengthens behavior features for data with obvious behavior patterns, and highlights situational features for data with special usage scenarios, thereby ensuring that the fused features can retain the key information of each dimension to the maximum extent and provide high-quality feature representation for subsequent clustering analysis.

[0052] S4. Clustering the multi-dimensional feature representation vector using a density difference perception clustering algorithm, the algorithm constructs a local density field through an adaptive bandwidth kernel density estimation, defines a density difference coefficient to quantify the local variation characteristics of the density field, identifies the clustering center through a climbing iteration based on the density gradient, and divides the clustering boundary through density saddle point detection, to obtain multiple clustering clusters.

[0053] The application designs a density difference perception clustering algorithm (DD-SCAN), which is specially optimized for the identification of small but high value demand clusters in the rejection data. The traditional density clustering algorithm such as DBSCA has obvious limitations when dealing with uneven density distribution: when there are high-density and low-density regions in the data set, using a globally uniform density threshold will cause the effective clustering of the low-density region to be incorrectly determined as noise, or multiple clusters of the high-density region to be incorrectly merged. The algorithm effectively overcomes this limitation through adaptive density estimation and density difference perception mechanism.

[0054] The algorithm first constructs a high-precision local density estimator. For each data point in the feature space , the local density value of the data point is calculated by kernel density estimation (KDE), and an adaptive bandwidth Gaussian kernel function is used: ; In the formula, is the local density estimation value of the data point ; is the total number of data points in the data set ; is any data point in the data set, and the entire data set is traversed; is the Euclidean distance between the data point and ; is the adaptive bandwidth parameter at the data point ; is the Gaussian kernel function, defined as: where is the feature vector dimension; is the normalized distance.

[0055] The adaptive bandwidth parameter is dynamically adjusted according to the local data distribution characteristics of the data point , and the calculation formula is: ; In the formula, is the global reference bandwidth, which is determined by the Silverman rule: where is the average of the standard deviations of each dimension of the data; is the initial density value estimated using a fixed bandwidth ; is the geometric mean of the initial density values of all data points, ; is the sensitivity parameter, which controls the response degree of the bandwidth to the local density, and the value range is In this embodiment, we take .

[0056] This adaptive bandwidth strategy enables density estimation to achieve lower bandwidth and higher resolution in high-density regions, and higher bandwidth and stronger smoothing effect in low-density regions, thus obtaining accurate density estimates across different density areas. Kernel density estimation provides a smoother and more accurate density measure compared to simple neighborhood counting.

[0057] In obtaining the density field Based on this, a density gradient vector is introduced. and density Hessian matrix The concept of is used to calculate the first and second derivatives of the density field through numerical differentiation.

[0058] The density gradient vector is calculated using the central difference method: ; ; In the formula, For data points The density gradient vector at a given location indicates the direction of the fastest local density growth. The eigenvector of the eigenvector One portion, ; For the first One standard basis vector; The differential step size is taken in this embodiment. The density Hessian matrix is ​​the second-order partial derivative matrix of the density field: ; In the formula, For data points The density Hessian matrix at a given location is a symmetric matrix. For the density function with respect to the first The and the first The second-order mixed partial derivatives of each coordinate component are numerically calculated using the second-order central difference method: ;

[0059] Eigenvalue analysis of the Hessian matrix reveals the curvature characteristics of the density field: at density maxima, all eigenvalues ​​of the Hessian matrix are negative; at density minima, all eigenvalues ​​are positive; and at saddle points, the eigenvalues ​​have mixed signs.

[0060] Using the differential properties of the density gradient and the Hessian matrix, the density difference coefficient (DDC) is defined: ; In the formula, For data points The density difference coefficient of DDC is a non-negative scalar; The Euclidean norm of the density gradient vector, reflecting the intensity of density variation; The smallest eigenvalue of the Hessian matrix (algebraic minimum), reflecting the curvature of the density field in the flattest direction; The absolute value is taken.

[0061] The DDC index takes into account both the intensity and curvature of density variation, and has the following properties: (1) In the flat density region, the gradient norm is close to zero, and the DDC value is small; (2) Near the density extreme point, the gradient norm is close to zero, and the DDC value is small; (3) In the region boundary where the density changes sharply, both the gradient norm and the absolute value of the curvature are large, and the DDC value is large. Therefore, DDC can effectively identify the region boundary where the density changes significantly, and these boundaries are usually the natural dividing lines between different clustering clusters.

[0062] In the clustering formation process, an improved density climbing strategy is used to identify local density maximum points as cluster centers. For each data point , the following iterative process is used to climb along the density gradient direction: ; In the formula, is the position of the point at the th iteration, and the initial value ; is the position of the point after the th iteration; is the density gradient vector at the current position; is the norm of the gradient vector, used for normalization to obtain the unit direction vector; is an adaptive step size parameter, which is dynamically adjusted according to the local density gradient size: ; In the formula, is the basic step size; is the global average gradient norm. When the gradient is large, the step size is close to , and when the gradient is small, the step size is correspondingly reduced to ensure the stability of the iterative process.

[0063] The iteration termination condition is: , is the gradient norm threshold; is the maximum eigenvalue; when the gradient norm is small enough and all eigenvalues of the Hessian matrix are negative, it indicates that the current position is a local density maximum point, and the iteration converges. When the iteration converges, the converged point is identified as the cluster center. All data points that converge to the same density maximum point are grouped into the same initial clustering cluster.

[0064] To accurately identify small high-value clusters and divide cluster boundaries, a density saddle point detection mechanism is introduced. Density saddle points are special critical points in the density field, which increase in density in some directions and decrease in density in other directions, and usually mark the boundaries between different clusters. The identification of the saddle point is based on the eigenvalue structure of the Hessian matrix. For a critical point (satisfying ), if its Hessian matrix has both positive and negative eigenvalues, then s is a density saddle point.

[0065] For the initially identified adjacent cluster centers and , the algorithm searches for a density saddle point along the connecting line direction. Let be the saddle point between and , if the following conditions are met, it is determined that and belong to different cluster clusters: ; In the formula, is the density saddle point between the two cluster centers; and are two adjacent density maximum points (cluster centers); is the local density value at the saddle point ; and are the local density values at the cluster centers and , respectively; is the minimum value function; is a density ratio threshold parameter, which controls the maximum ratio of the density at the saddle point to the density of the cluster center, and the value range is , and in the embodiment ; is a DDC threshold parameter that controls the minimum intensity of the density difference at the boundary, and in the embodiment , where is the average value of all data point DDC values. The meaning of the above conditions is: (1) The density at the saddle point should be significantly lower than the density of the cluster center on both sides, indicating that there is a clear density trough between the two clusters; (2) The density difference coefficient at the saddle point should be large enough, indicating that this position is indeed a boundary region with sharp density changes. If the saddle point does not meet the above conditions, it is considered that the density difference between the two density maximum points is not enough to constitute a cluster boundary, and the corresponding initial cluster should be merged.

[0066] To cope with data distribution at different scales and ensure the robustness of clustering results, a hierarchical multi-resolution analysis strategy is adopted. The specific steps are as follows: (1) Multi-scale density estimation: Density estimation is performed at multiple different global reference bandwidths , respectively, where is the resolution level number, and is taken as in the embodiment. The bandwidth sequence is set as a geometric progression:

[0067] (2) Multi-scale clustering: The clustering process is independently performed at each resolution level to obtain sets of clustering results .

[0068] (3) Clustering stability analysis: For each pair of adjacent resolution level clustering results, the overlap degree between clustering clusters is calculated: , where is the th clustering cluster at the th resolution level; denotes the cardinality (number of elements) of the set.

[0069] (4) Stable clustering screening: The clustering structure that exists stably at multiple resolution levels is retained. The stability score of a clustering cluster is defined as: , where is an indicator function; is the overlap threshold, which is taken as in the embodiment. The higher the stability score, the more likely it is to represent the true data structure. Clustering clusters with stability scores below the threshold are filtered out, effectively eliminating noise clusters and pseudo-clusters that only appear at a single resolution.

[0070] Finally, the boundaries of the clustering clusters are fine-tuned by optimizing the following objective function: ; In the formula, is the objective function to be minimized; is the final number of clustering clusters; is the data point set contained in the th clustering cluster; is the data point in the clustering cluster ; is the center (density maximum point) of the th clustering cluster; and are data points and clustering center local density value of the i-th cluster; is the density saddle point between the j-th cluster and the k-th cluster; is the density saddle point between the j-th cluster and the k-th cluster; is the density difference coefficient at the saddle point; is the regularization parameter, which controls the balance between the two loss terms, and is taken as 0.1 in the embodiment; ; is the expected lower threshold of the density difference coefficient, and is taken as 0.1 in the embodiment; ; is the maximum value function.

[0071] The first term of the objective function is the intra-cluster density consistency loss, which encourages data points in the same cluster to have similar density values, thereby forming a density-uniform cluster; the second term is the inter-cluster boundary clarity regularization term, which encourages the saddle point between different clusters to have a higher DDC value, thereby forming a clear cluster boundary. By minimizing the objective function through the gradient descent method, the cluster membership of the boundary data points is fine-tuned until the objective function converges. This optimization process ensures the density consistency within the cluster and the clear boundary between the clusters, which is particularly suitable for identifying high-cohesion small data clusters reflecting specific user needs in the rejection data.

[0072] S5. Feature profiling and value evaluation of the cluster, generation of demand description documents, and feedback of optimization algorithm parameters according to the verification results.

[0073] The last step of the present application is the demand identification and verification process, which aims to convert the clustering analysis results into operational product demand suggestions and establish a closed-loop optimization mechanism to continuously improve the algorithm performance. First, the system performs in-depth feature analysis on each identified cluster, including: performing word frequency statistics and TF-IDF analysis on the original text of all data points in the cluster, extracting Top-N keywords as semantic tags, and at the same time, using a text summary model to generate a summary semantic description of the cluster; statistical analysis of the behavior sequence pattern distribution of the data points in the cluster, identification of high-frequency behavior sequences and typical behavior transition paths; aggregation of the context feature distribution of the data points in the cluster, including device type distribution, time period distribution, and regional distribution, etc.

[0074] Subsequently, the system applies a multi-factor scoring model to evaluate the value of each cluster, taking into account the following five factors: (1) Scale factor : measures the data size of the cluster , where is the number of data points in the cluster . Log transformation is used to smooth the impact of extreme values.

[0075] (2) Cohesion factor​ : measure the tightness of a cluster , where is the average distance of data points in the cluster to the center; is the global average distance. Higher cohesion factor means the cluster is tighter.

[0076] (3) distinctness factor : measure the difference of a cluster from other clusters , where is the center of another cluster. Higher distinctness factor means the cluster is more different from other clusters.

[0077] (4) disparity factor : measure the gap between the cluster and existing product features , where is the average feature vector of the cluster; is the feature representation of the th existing product feature; is the number of existing features; is the cosine similarity. Higher disparity factor means the demand is more likely to be a new demand not covered by existing products.

[0078] (5) business value factor : estimate the potential business value based on user attributes and behavior features , where is the value score of data point corresponding user, calculated based on user historical consumption, activity level, membership level, etc.

[0079] Combine the above five factors to calculate the value score of a cluster: ; where, is the comprehensive value score of the th cluster; is the weight coefficient of each factor, satisfying , which can be adjusted according to business priorities. In this embodiment, the initial setting is , , , , .

[0080] According to the value score, filter high-value demand candidates (C ), automatically generate a structured requirement description document. The document contains the following: (1) Requirement background: describe the user pain points and scenarios corresponding to this requirement, automatically generated based on semantic tags and context attributes. (2) User group characteristics: describe the typical user portrait corresponding to this requirement, including device preferences, use time, regional distribution, etc. (3) Core function points: extract the core functions expected by users based on behavior pattern analysis and semantic tags. (4) Implementation priority suggestion: give priority suggestions based on value scoring, scale factor and implementation difficulty estimation. (5) Typical case display: list the selected typical samples to facilitate product personnel to understand the requirements intuitively. For potential requirements with high scores, the system further designs targeted verification schemes: (1) A / B test design: for function improvements that can be quickly implemented, design division schemes, sample size calculations, core indicator definitions, etc. for experimental groups and control groups. (2) User survey questionnaire: automatically generate questionnaire topics based on requirement characteristics, including requirement confirmation questions, priority ranking questions, and function detail preference questions. (3) Prototype verification scheme: for complex new function requirements, design the function range and verification indicators of the minimum viable product prototype.

[0081] S6. After verification is complete, the system feeds back the result data to the algorithm framework, forming a closed-loop optimization mechanism: (1) Feature fusion weight update: collect requirement verification results, and distinguish between the set of verified requirements and the set of unverified requirements . Analyze the feature differences between the two types of requirements in the semantic, behavior, and context dimensions, and adjust the feature fusion weights in step S3: where is the feature dimension; and are the average attention weights of the verified and unverified requirements in that dimension, respectively; is the learning rate. This update mechanism enhances the feature dimension weights related to verified requirements.

[0082] (2) Cluster threshold parameter adjustment: according to the difference between the verification pass rate and the target pass rate, dynamically adjust the clustering threshold parameters and : ; ; where and are the updated density saddle point detection threshold parameters; and are the previous threshold values; and are the adjustment steps of β and γ, respectively; is the actual verification pass rate; The target verification pass rate is verified. If the verification pass rate is lower than the target, the threshold is increased to filter out clusters with higher confidence; if the verification pass rate is higher than the target, the threshold is lowered to find more potential needs.

[0083] (3) Value evaluation model optimization: based on the verification result, the factor weight of the value scoring model in step S5 is optimized, and the verification error is minimized by gradient descent: ; ; Wherein is the weight optimization loss function; is the learning rate. The optimization target makes the verified needs obtain higher value scores, and the unverified needs obtain lower scores.

[0084] (4) Historical pattern library update: the feature vector corresponding to the verified needs is added to the historical high-value pattern library for the calculation of the historical similarity factor: ; In the formula, The dimensional historical pattern library; is the average feature vector of the first verified needs in this dimension; is the newly added verified needs set in this round.

[0085] This complete demand identification and verification process ensures full-link support from data discovery to demand landing. Through continuous feedback optimization, the system can continuously learn and adapt to business characteristics, significantly improving the efficiency and accuracy of product demand mining.

[0086] This embodiment takes the search function identification data of an e-commerce platform as an example to show the complete process of the present application.

[0087] In the data collection and preprocessing phase, the system connects the search engine backend system of the e-commerce platform through the pre-configured data interface, collects all the rejected user search request data within a month, totaling about 2 million records. The original data collected includes user query text, session identifier, user ID (after desensitization processing), query timestamp, device information, geographic location code, and rejection reason code, etc. The system first performs data cleaning operations, filters obvious abnormal data through the set rules, including empty queries, pure number non-semantics queries, robot batch requests, etc., reduces the data set size to 1.5 million; then performs data standardization processing, including query text character standardization, case unification, special symbol processing, etc., to ensure the consistency of subsequent analysis; then extracts rejection related meta information, such as rejection reason distribution statistics, time distribution characteristics, etc., to provide background reference for subsequent analysis. To balance the calculation efficiency and analysis effect, the system uses stratified sampling method to sample the cleaned data. Taking the rejection reason category, device type, and time period as the stratification basis, ensure that the proportion of each layer sample is consistent with the distribution of the original data, and finally extract 30,000 records as the experimental data set, which ensures the representativeness of the data while significantly reducing the computational cost.

[0088] In the multi-dimensional feature extraction stage, the system first processes the semantic features, uses a bidirectional Transformer network (such as the BERT-base model) fine-tuned for the e-commerce field to encode each query text, and obtains a word embedding matrix containing context information. Then, through a 12-layer Transformer encoder, the context semantic is enhanced, and the output of the [CLS] position is extracted as a 768-dimensional general semantic representation. Considering the particularity of the e-commerce search scene, the system further applies a domain adaptation layer to specialize the semantic representation. The adaptation layer is fine-tuned on labeled data in the e-commerce search field, which can better identify brand names, product attribute words, and promotion terms specific to the field. Finally, a 512-dimensional semantic feature vector is output. For behavior features, the system extracts the user behavior sequence within 30 minutes before and after each search request from the user session log, including browsing product categories, clicking product details, adding to the shopping cart, and collecting operations, as well as operation interval time and dwell time. Each behavior event is encoded into a composite vector containing behavior type embedding, time position encoding, dwell time encoding, and target object embedding. Through a time attention network, these behavior sequences are encoded into a 192-dimensional behavior feature vector, with a time decay mechanism giving recent behaviors higher attention weights. In terms of context features, the system integrates user device types (such as PC, mobile, smart speaker, etc.), query time periods (weekday / weekend, morning / afternoon / night), geographic location (city level, region type), and network environment (WiFi / 4G / 5G) information, respectively, and encodes or embeds them into a 128-dimensional context feature vector after a three-layer perception machine network fusion.

[0089] In the adaptive feature fusion link, the system first projects the three types of feature vectors with different dimensions into a common feature space of 512 dimensions through a linear projection layer, eliminating the influence of dimension differences. In the attention weight calculation stage, the system dynamically determines the weight of each feature dimension by considering three factors: the information entropy factor evaluates the information richness of each dimension, with a high entropy value indicating that the dimension contains more effective information; the historical similarity factor identifies similar feature patterns to historical verification needs by comparing with a pre-constructed high-value pattern library; and the confidence factor is output by each feature extraction module, reflecting the reliability of feature extraction. After weighted combination, the three factors are normalized by the Softmax function to obtain the dynamic attention weight of each dimension. Correlation analysis shows that in the e-commerce search scenario, the semantic feature has a high correlation with the user's real demand, the behavior feature is helpful to judge the demand urgency, and the context feature plays an important role in distinguishing exclusive demands in specific scenarios. In this embodiment, the average weight of the semantic feature is about 0.45, the behavior feature is about 0.31, and the context feature is about 0.24. As the data analysis proceeds, the system will automatically adjust the weight distribution according to different types of rejection data, for example, for queries containing obvious professional terms (such as "ANC noise reduction earphones"), the semantic feature weight is increased to 0.68; for behavior sequences showing clear purchase intention, the behavior feature weight is correspondingly increased. The fused feature vector is subjected to layer normalization to eliminate dimension differences and stabilize the numerical distribution, forming a unified 512-dimensional feature representation for subsequent clustering analysis.

[0090] In the adaptive bandwidth kernel density estimation stage, the system first calculates the global reference bandwidth using the Silverman rule, obtaining an initial value of about 0.28. Considering the non-uniform distribution of the rejection data, the system calculates an adaptive bandwidth for each data point: a smaller bandwidth is used in dense areas to maintain resolution, and a larger bandwidth is used in sparse areas to ensure estimation stability. Based on the adaptive bandwidth Gaussian kernel function, the system calculates the local density value of all data points. Then the system calculates the density gradient and Hessian matrix. The central difference method is used to calculate the density gradient vector of each data point, representing the direction and rate of density change; at the same time, the Hessian matrix of the density function is calculated, and its eigenvalues are analyzed to determine the geometric properties of the point. Based on the density value, gradient modulus and Hessian matrix eigenvalue, the system calculates the density difference coefficient (DDC) of each data point, which presents a higher value in the cluster boundary area, providing a basis for subsequent boundary division. In the density climbing and cluster center identification stage, each data point iteratively climbs along the density gradient direction, and the step size is adaptively adjusted according to the DDC value. A larger step size is used in low DDC areas to accelerate convergence, and a smaller step size is used in high DDC areas to accurately locate. When the gradient modulus is less than a certain threshold, the iteration is stopped, and the data points converging to the same position are classified into the same cluster, and the convergence endpoint is the candidate cluster center. In this embodiment, 127 candidate cluster centers are identified. The system further performs density saddle point detection and cluster merging determination. For adjacent candidate cluster centers, search for the density minimum point (saddle point) on the line connecting them. If the saddle point density is lower than a certain proportion (50% in this embodiment) of the density of the two cluster centers, and the DDC value of the saddle point exceeds 10% of the average, then the two sides are determined as independent clusters; otherwise, the two candidate clusters are merged. After this processing, 127 candidate clusters are merged into 89 preliminary clusters. To improve the stability of the clusters, the system uses a hierarchical multi-resolution analysis method. Five different reference bandwidths (geometric progression from 0.14 to 0.57) are used to perform clustering, and the performance of each cluster at different resolutions is observed. If a cluster exists stably at least at three resolution levels (overlapping degree exceeds 70%), it is considered a stable cluster and retained; otherwise, it is filtered out. In this embodiment, 72 cluster clusters pass the stability test. Finally, the cluster boundary is optimized. The system constructs an optimization function targeting cluster compactness and separation, introduces DDC as a regularization constraint, and adjusts the attribution of data points in the boundary area through gradient descent iteration, so that the cluster boundary is distributed along the high DDC area as much as possible. After 100 rounds of iterative optimization, 72 clusters with clear boundaries are finally formed. The size distribution of these clusters presents a long-tail characteristic, with the largest cluster containing about 7600 data, and most clusters having a size between 50 and 500, showing the diversity and niche of user demand. The cluster quality evaluation shows that the silhouette coefficient reaches 0.412, indicating that the cluster structure has good cohesion and separation.

[0091] In the demand identification and verification phase, the system conducts multi-dimensional analysis and value assessment on the 72 clusters. The system first extracts the feature portrait of each cluster, including high-frequency keywords and TF-IDF weight distribution, shared behavior pattern sequences, main usage scenario characteristics, and user group portrait, forming a complete cluster feature description. In the multi-factor value assessment section, the system considers five value factors: the scale factor reflects the potential number of users, and the logarithmic normalization processing avoids the dominance of large clusters; the cohesion factor measures the consistency of data within the cluster, and high cohesion indicates that the demand is clear and concentrated; the uniqueness factor assesses the difference between the data within the cluster and the global distribution, and high uniqueness indicates that the demand has distinct characteristics; the difference factor measures the gap between this demand and existing functions, and high difference means that there is a large room for product improvement; the business value factor considers business indicators such as user spending capacity and purchase willingness. According to the business characteristics of the e-commerce platform, the system sets the difference factor weight to 0.25, the cohesion and uniqueness factor weights to 0.20 each, the business value factor weight to 0.20, and the scale factor weight to 0.15. The weighted sum of each factor gives the comprehensive value score. The scoring results show that there are three clustering clusters with particularly high business value: the first high-value cluster (comprehensive score 0.828) represents the demand for "combined product search" as users try to search for multiple related products at the same time (such as "red dress with silver high heels"), the cluster contains about 2800 data, and the user group is mainly composed of women aged 25-35, with a usage proportion of 43% on weekends; the second high-value cluster (comprehensive score 0.793) reflects the search failure caused by the use of dialect or heavy accent in voice input, indicating the necessity of the "dialect voice recognition" function, the cluster contains about 1500 data, with a proportion of 34% for smart speaker devices and 71% for users in third and fourth-tier cities; the third high-value cluster (comprehensive score 0.789) shows the pattern of users searching for seasonal products in advance when they are not yet available, reflecting the demand for "seasonal product pre-display", the cluster has a higher user spending capacity, with an average single price that is 42% higher than the platform average. The system automatically generates detailed demand documents for these three high-value demands, including demand background description, target user group characteristics, core function points, typical case listing, and implementation priority suggestions, providing complete decision-making references for product teams. At the same time, the system designs targeted A / B test verification schemes: for the "combined product search" function, 10% of platform users are selected as the test group and 10% as the control group, and a two-week function test is implemented, with core observation indicators including search success rate, click rate, conversion rate, single price, and dwell time. Based on the verification results, the system feeds back the verification data to the algorithm framework, optimizes the clustering parameters and feature weights, and improves the system's sensitivity to similar demand patterns.The whole demand mining process is completed from data collection to demand verification, which takes significantly less time than the traditional manual analysis method, and discovers small but high-value demands that may be ignored by expert methods, fully demonstrating the practical value and technical advancement of the present application.

[0092] To verify the effectiveness of the method proposed in the present application, the DD-SCAN algorithm is compared with three mainstream clustering algorithms in the comparative experiment, and the clustering results are visualized, as shown in Table 1: Table 1: Comparison of clustering results of algorithms

[0093] As can be seen from the table, the DD-SCAN proposed in the present application identifies 52 clustering clusters in terms of the number of clusters, which is much more than the 15 of K-means and the 28 of DBSCAN, indicating that the algorithm can discover more subdivided user demand categories. In terms of small cluster identification capability, DD-SCAN successfully identifies 23 small clusters (with less than 100 samples), which is nearly twice that of HDBSCAN, which is crucial for mining small but high-value long-tail demands. In terms of silhouette coefficient, DD-SCAN reaches 0.433, indicating that the clustering results have better cohesion and separation.

[0094] To visually display the clustering effect, the present embodiment uses the t-SNE algorithm to reduce the 512-dimensional feature vector to two dimensions for visualization, and the perplexity parameter is set to 30. Figure 3 The clustering result distribution of the DD-SCAN algorithm is shown, where different colors represent different clustering clusters, and the red five-point star mark indicates the three high-value demand clusters identified after value evaluation. As can be observed from the figure, the DD-SCAN algorithm successfully identifies multiple clustering clusters distributed in various regions of the feature space, including several small clusters with clear boundaries, verifying the adaptive ability of the algorithm to density differences. The three high-value demand clusters are located in different regions of the feature space, indicating that these demands have significant feature differences. Figure 4 The value score distribution histogram of the 52 clustering clusters is shown. The score distribution is approximately normally distributed, with a mean of 0.412 and a standard deviation of 0.186. The red dashed line in the figure marks the high-value threshold τ=0.7, and the clustering clusters exceeding this threshold are determined to be high-value potential demands. The statistical results show that there are 3 clustering clusters (accounting for 5.8%) whose value scores exceed the threshold, corresponding to the "combined commodity search" (V=0.828), "dialect voice recognition" (V=0.793) and "seasonal commodity advance display" (V=0.789) three types of demands, which will be preferentially pushed to the product team for manual review and A / B test verification.

[0095] The experimental results show that the clustering method based on the density difference perception can effectively identify the user demand mode hidden in the rejection data, and has a significant advantage in the minority demand mining and clustering quality. Through the multi-factor value evaluation model, the system can accurately screen out potential demands with high commercial value from a large number of clustering clusters, and provide data-driven decision support for product iteration.

[0096] The core innovation points of the present application mainly include two aspects: first, a multi-dimensional feature representation and fusion mechanism is proposed, which can comprehensively capture user demand features from three key dimensions of semantics, behavior and context, and realize effective integration of different dimensional features through adaptive weight calculation; second, a novel density difference perception clustering algorithm is designed, which realizes accurate identification of small high-value demand clusters easily ignored by traditional clustering methods through density gradient calculation and local density peak detection. These innovations enable the present application to improve the demand mining accuracy while significantly improving the analysis efficiency and coverage.

[0097] Embodiment 2 The embodiment provides a computer terminal, including a memory, a processor and a computer program stored in the memory and executable on the processor, when the processor executes the computer program, the steps of the user demand mining method based on feature fusion density perception are realized.

[0098] As shown in Figure 5 , the computer terminal provided in the embodiment includes at least one processor 101 and a memory 102 connected with the at least one processor 101, and the specific connection medium between the processor 101 and the memory 102 is not limited in the embodiment. Figure 5 In the embodiment, the connection between the processor 101 and the memory 102 is taken by way of example. Figure 5 The bus 100 is represented by a thick line, and the connection mode between other components is only schematically described, and is not limited. The bus 100 can be divided into an address bus, a data bus, a control bus and the like, and for the sake of representation, Figure 5 In the embodiment, only one thick line is represented, but it does not mean that there is only one bus or one type of bus. Alternatively, the processor 101 can also be called a controller, and the name is not limited.

[0099] In the embodiment, the memory 102 stores instructions executable by the at least one processor 101, and the at least one processor 101 can execute the foregoing method by executing the instructions stored in the memory 102.

[0100] The processor 101 is the control center of the apparatus, and can connect all parts of the apparatus through various interfaces and lines. The apparatus performs various functions and processes data by running or executing instructions stored in the memory 102 and calling data stored in the memory 102, thereby monitoring the apparatus as a whole.

[0101] In a possible design, the processor 101 can include one or more processing units, and the processor 101 can integrate an application processor and a modem processor. The application processor mainly processes an operating system, a user interface, and an application program, and the modem processor mainly processes wireless communication. It can be understood that the modem processor can also not be integrated into the processor 101. In some embodiments, the processor 101 and the memory 102 can be implemented on the same chip, and in some embodiments, they can also be implemented on separate chips respectively.

[0102] The processor 101 can be a general-purpose processor, for example, a central processing unit (CPU), a digital signal processor, an application-specific integrated circuit, a field programmable gate array, or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component, and can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the feature fusion density-aware user demand mining method disclosed in Embodiment 1 can be directly embodied by a hardware processor for execution, or be executed by a combination of hardware and software modules in the processor 101.

[0103] The memory 102, as a non-volatile computer readable storage medium, can be used to store non-volatile software programs, non-volatile computer executable programs and modules. The memory 102 can include at least one type of storage medium, for example, can include flash memory, hard disk, multimedia card, card type memory, random access memory (RAM), static random access memory (SRAM), programmable read-only memory (PROM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), magnetic memory, magnetic disk, optical disk, etc. The memory 102 is any other medium capable of carrying or storing desired program code in the form of instructions or data structures and capable of being accessed by a computer, but is not limited thereto. The memory 102 in the present embodiment can also be a circuit or any other device capable of realizing a storage function, used to store program instructions and / or data.

[0104] By designing and programming the processor 101, the code corresponding to the user demand mining method based on feature fusion density perception introduced in the foregoing embodiments can be fixed into the chip, so that the chip can execute the steps of the user demand mining method based on feature fusion density perception as shown in the foregoing embodiments at runtime. Figure 1 How to design and program the processor 101 is a technology known to those skilled in the art, and will not be described here.

[0105] Embodiment 3 The present embodiment provides a computer readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the steps of the user demand mining method based on feature fusion density perception as described in Embodiment 1.

[0106] The computer readable storage medium can include a flash memory, a hard disk, a multimedia card, a card-type memory (e.g., SD or DX memory, etc.), a random access memory (RAM), a static random access memory (SRAM), a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), a programmable read-only memory (PROM), a magnetic memory, a magnetic disk, an optical disk, etc. In some embodiments, the storage medium can be an internal storage unit of the computer device, such as a hard disk or a memory of the computer device. In other embodiments, the storage medium can also be an external storage device of the computer device, such as a plug-in hard disk, a smart media card, a secure digital card, a flash memory card, etc. equipped on the computer device. Of course, the storage medium can also include both the internal storage unit and the external storage device of the computer device. In the present embodiment, the memory is generally used to store an operating system and various application software installed in the computer device, etc. In addition, the memory can also be used to temporarily store various data that have been output or will be output.

[0107] The above description is merely preferred embodiments of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art, according to the technical range disclosed in the present application and the inventive concept, should be covered within the protection scope of the present application.

Claims

1. A user demand mining method based on feature fusion density awareness, characterized in that, include: S1. Collect product rejection data, preprocess the data, and extract the metadata. S2. Extract semantic feature vectors from the preprocessed data text, use the session identifier in the metadata to obtain user behavior sequences to generate behavior feature vectors, and generate context feature vectors based on the time and environment attributes in the metadata. S3. Calculate the attention weights of each dimension of features, and use the attention weights to perform weighted fusion of semantic feature vectors, behavioral feature vectors and contextual feature vectors to generate multidimensional feature representation vectors; S4. Cluster the multidimensional feature representation vector using a density difference-aware clustering algorithm. This algorithm constructs a local density field through adaptive bandwidth kernel density estimation, defines a density difference coefficient to quantify the local change characteristics of the density field, identifies cluster centers through climbing iteration based on density gradient, and divides the cluster boundaries through density saddle point detection to obtain multiple clusters. S5. Construct feature profiles and conduct value assessments for the clusters, and generate a requirement description document.

2. The user demand mining method based on feature fusion density awareness according to claim 1, characterized in that, The specific process of step S2 includes: At the semantic feature extraction level, the text description of the rejected data is input into a pre-trained language model for word embedding representation to capture the basic semantic features of the text; the basic semantic features are then enhanced with contextual semantics through a deep bidirectional Transformer network to extract general semantic features; and the general semantic features are then processed with a domain adaptation layer to obtain a semantic feature vector. At the behavioral feature extraction level, the behavioral sequence in the user session is obtained, each behavioral event in the sequence is mapped to a vector representation, and the vector representation of the behavioral sequence is processed by a temporal attention network to extract the behavioral feature vector. At the context feature extraction level, environmental context information including device technical parameters, geographical location information, network environment status and usage time features is obtained, and nonlinear fusion is performed through a multilayer perceptron network to obtain the context feature vector.

3. The user demand mining method based on feature fusion density awareness according to claim 1, characterized in that, Step S4 specifically includes: S41. Construct a local density estimator for each data point in the feature space. The local density value is calculated using an adaptive bandwidth Gaussian kernel function. : In the formula, As a base point The bandwidth parameter is adaptively adjusted according to the local data distribution characteristics, so that the density estimation has different resolutions in different regions; For the feature space, except for the point Any data point outside, The dataset consists of all data points in the feature space; For dataset The total number of data points in the middle; The Gaussian kernel function; It is the Euclidean norm; An adaptive bandwidth parameter is used to enable density estimation with different resolutions in different regions. The calculation formula is as follows: In the formula, This is the global baseline bandwidth parameter; Points obtained for fixed bandwidth estimation The initial density value; The geometric mean of the initial density values ​​for all data points; This is a sensitivity parameter with a value range of [0,1], which controls the degree to which the bandwidth adapts to local density. S42. Calculate the density gradient vector and the density Hessian matrix, expressed by the following formulas: In the formula, For point The gradient vector of the density function at a given point represents the direction of the fastest density growth; For point The Hessian matrix of the density function describes the second-order curvature property of the density field; For partial derivative operators; It is a local density function; The first to the second in the feature space, respectively. Each coordinate component The dimension of the multidimensional feature representation vector; superscript Indicates vector transpose; S43. Define the density difference coefficient to quantify the local variation characteristics of the density field at data points: In the formula, For point The density difference coefficient is used to quantify the local variation characteristics of the density field at that point; For point The smallest eigenvalue of the density Hessian matrix at a given point represents the second rate of change of the density field along the direction of minimum curvature. S44. Perform density ramping iterations based on the density gradient vector to identify initial cluster centers; where, for each data point Iterative updates are performed along the density gradient direction: In the formula, The position of the point at the t-th iteration; The position of the point after the (t+1)th iteration; The adaptive step size parameter controls the distance moved in each iteration; For point Density gradient vector at; Let be the Euclidean norm of the gradient vector; The calculation formula is: , Base step size; It is the global average gradient norm; The iteration termination condition is ,in The gradient norm threshold; The largest eigenvalue; the points in the convergence points that satisfy the local density maximum condition are identified as the initial cluster centers, where the local density maximum condition is that all eigenvalues ​​of the Hessian matrix are negative. S45. A density saddle point detection mechanism is introduced to delineate cluster boundaries. The specific process is as follows: Define the saddle point determination criteria: point A point is a density saddle point if and only if the following condition is satisfied:

1. The magnitude of the density gradient vector approaches zero: In the formula, The gradient threshold for determining saddle points; II. Hessian Matrix Both positive and negative eigenvalues ​​exist simultaneously; For the initially identified adjacent cluster centers and Search along the direction of the line connecting them for points that satisfy the saddle point criterion; when two density maxima are found... and When the density saddle point s between them satisfies the following condition, it is determined that... and Belonging to different clusters: In the formula, For point The local density value, and Points and points The local density value; To obtain the minimum value; The density ratio threshold parameter controls the maximum ratio of the density at the saddle point to the density at the cluster center, and its value ranges from (0,1). This is the DDC threshold parameter, which controls the minimum intensity of density difference at the boundary.

4. The user demand mining method based on feature fusion density awareness according to claim 3, characterized in that, In step S4, a hierarchical multi-resolution analysis is also performed, specifically including: Density estimation and clustering were performed under multiple different bandwidth parameter settings; stability analysis was conducted on the clustering results at each resolution, and the overlap of clusters at different resolutions was calculated; clustering structures that are stable at multiple resolutions were retained, while noisy clusters and pseudo-clusters that only appear at a single resolution were filtered out.

5. The user demand mining method based on feature fusion density awareness according to claim 3, characterized in that, In step S4, the following objective function is optimized. Adjust the boundaries of the clusters: In the formula, This represents the total number of clusters. For the first The set of data points contained in each cluster; For the first The center point of each cluster, i.e., the local density maxima, For the first Local density values ​​of the center points of each cluster; For the first The cluster and the first Density saddle points between clusters; This is a regularization parameter used to balance intra-cluster density consistency and inter-cluster boundary clarity; This represents the expected lower bound threshold for the density difference coefficient. To obtain the maximum value.

6. The user demand mining method based on feature fusion density awareness according to claim 5, characterized in that, In step S5, the specific process of constructing the feature profile and evaluating its value includes: S51. Extract feature profiles for each cluster, including: extracting high-frequency keywords within the cluster as semantic labels using the TF-IDF algorithm, statistically analyzing the behavioral sequence pattern distribution of data points within the cluster, and summarizing the contextual attribute distribution features of the data within the cluster; S52. Calculate the multi-factor value score for each cluster, using the following formula: In the formula, For the first Multifactor value scoring of individual clusters; As a scale factor, ,in For clusters The number of data points is determined, and the effect of extreme values ​​is smoothed using a logarithmic transformation. It is the cohesion factor. ,in For the first The average distance from data points within each cluster to the cluster center. Global average distance As a unique factor, , For the first Data points within each cluster; The difference factor is calculated by taking the cosine similarity between the cluster feature vector and each feature vector in the existing product feature library, and then taking the complement of the maximum similarity. ,in For the first The average eigenvector of each cluster; For the first A feature representation of an existing product function; This refers to the number of existing functions; Cosine similarity; As a business value factor, ,in For data points The value score for each user is calculated based on their historical spending, activity level, and membership level. to Let be the weight coefficients of each factor, satisfying ; S53. Based on value rating The clusters are sorted, and those with scores higher than a preset threshold are selected. Clusters are used as high-value requirement candidates to generate structured requirement description documents that include requirement background, user group characteristics, core functional points and implementation priorities.

7. The user demand mining method based on feature fusion density awareness according to claim 6, characterized in that, Also includes: S6. Based on the verification results of the aforementioned requirements description document, optimize the algorithm parameters. The specific process includes: S61. Collect requirement verification result data, including the set of requirements that have passed verification. and the set of requirements that failed validation ; S62. Adjust the feature fusion weights based on the verification results, enhancing the weights of feature dimensions relevant to the verification pass requirement. The updated formula is as follows: In the formula, For the updated number dimensional feature weights, The weights before the update. For learning rate, For the first The demand corresponds to the cluster in the th... Attention scores on dimensional features For the first The demand corresponds to the cluster in the th... Attention scores on dimensional features and Sets and The number of elements; S63. Adjust the clustering threshold parameter based on the validation results, and dynamically update the parameter according to the difference between the validation pass rate and the target pass rate. The update formula is as follows: In the formula, and These are the density ratio threshold parameters before and after the update, respectively; and These are the DDC threshold parameters before and after the update, respectively. and These are the adjustment step sizes for β and γ, respectively; This represents the actual pass rate. To verify the pass rate of the target; S64. Based on the validation results, optimize the weight coefficients of each factor in step S52, and minimize the following loss function using gradient descent: The weight update formula is as follows: In the formula, Optimize the loss function for the weights; The learning rate; and These are the first and second pages after the update, respectively. Factor weights; For the loss function on the th Partial derivatives of each weight; S65. Add the feature vectors corresponding to the verified requirements to the historical high-value pattern library for subsequent calculation of historical similarity factors: In the formula, No. Dimensional historical pattern library; For the first The verification passes the average feature vector of the requirement in this dimension; This is the set of newly added verification requirements in this round.

8. The user demand mining method based on feature fusion density awareness according to claim 1, characterized in that, Step S1 specifically includes: Establish an interface connection with the product rejection model to collect product rejection data, which includes user input content, operation records, submission time, and device information; clean the collected product rejection data to remove invalid data, outliers, and duplicate records; standardize the format of the cleaned product rejection data to convert data from different sources and types into a unified structured format; extract and label the metadata of the product rejection data, which includes rejection timestamp, rejection reason code, user session identifier, device type identifier, and user unique identifier; The specific process of step S3 includes: Semantic feature vectors, behavioral feature vectors, and contextual feature vectors are dimensionally aligned using projection matrices, projecting feature vectors of different dimensions onto a unified feature space. Attention scores are calculated for each projected feature vector, taking into account the information entropy of the feature vector, its similarity to historical high-value patterns, and feature confidence. The attention scores are then normalized to obtain attention weights for each feature dimension. These attention weights are then used to weight and sum the three types of feature vectors to obtain a fused feature vector. Finally, the fused feature vector undergoes layer normalization to obtain the final multidimensional feature representation vector.

9. A computer terminal, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the user demand mining method based on feature fusion density awareness as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the user demand mining method based on feature fusion density awareness as described in any one of claims 1 to 8.

Images