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

Abstract

The application 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 collects product rejection data, extracts a semantic feature vector from a preprocessed data text, acquires a user behavior sequence by using a session identifier in meta information to generate a behavior feature vector, and generates a context feature vector based on time and environment attributes in the meta information; the attention weight of each dimension feature is calculated, the semantic feature vector, the behavior feature vector and the context feature vector are weighted and fused by using the attention weight, and a multi-dimensional feature representation vector is generated; the multi-dimensional feature representation vector is clustered to obtain a plurality of clustering clusters; feature portrait construction and value evaluation are performed on the clustering clusters to generate a demand description document, and algorithm parameter optimization is fed back according to a verification result. The application can significantly improve the efficiency and quality of mining potential user demand from neglected rejection data.

Description

Technical Field

[0001] This invention relates to the fields of data mining and product demand analysis technology, specifically to a user demand mining method, terminal, and medium based on feature fusion density perception. Background Technology

[0002] During product development, high-value data is often incorrectly rejected by product rejection models due to inadequate design considerations and insufficient understanding of user needs. This rejected data contains important user needs information and product improvement directions; effectively mining this data can provide valuable decision-making basis for product iteration. Currently, processing rejected data mainly relies on human expert analysis, simple rule filtering, and traditional cluster analysis. However, these methods have obvious drawbacks such as low efficiency, high cost, and poor adaptability, making them difficult to cope with large-scale data analysis needs and becoming a bottleneck restricting the speed of product optimization.

[0003] In feature analysis, existing technologies mostly focus on single-dimensional feature extraction. Some focus on text content analysis, some on user behavior data, and some attempt simple feature combinations. Although some companies have begun to adopt machine learning and deep learning technologies for semi-automated analysis in recent years, significant manual intervention is still required, and model performance is highly dependent on the quality of training data, with limited generalization ability. Traditional clustering algorithms such as K-means and DBSCAN are easily affected by noise when processing complex multidimensional data, making it difficult to discover small but internally consistent data clusters, which often represent important emerging user needs.

[0004] In summary, existing technologies for mining user needs from rejection data suffer from several limitations, including insufficient automation, inadequate integration of multi-dimensional features, weak identification of niche needs, poor dynamic adaptability, and an imperfect demand value assessment mechanism. These limitations restrict enterprises' ability to efficiently discover and utilize valuable user needs from rejection data. Therefore, there is an urgent need for a technological method that can automatically and efficiently identify and mine high-value user needs from complex, multi-dimensional rejection data to support enterprises in better understanding user needs and accelerating product iteration and optimization. Summary of the Invention

[0005] To address the technical problems existing in the prior art, this invention provides a user demand mining method, terminal, and medium based on feature fusion density perception, which can significantly improve the efficiency and quality of mining potential user demands from ignored rejection data.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] This invention discloses a user demand mining method based on feature fusion density awareness, comprising:

[0008] S1. Collect product rejection data, preprocess the data, and extract the metadata.

[0009] 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.

[0010] 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;

[0011] 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.

[0012] S5. Construct feature profiles and conduct value assessments for the clusters, and generate a requirement description document.

[0013] As a further improvement to the above scheme, the specific process of step S2 includes:

[0014] 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.

[0015] 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.

[0016] 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.

[0017] As a further improvement to the above scheme, step S4 specifically includes:

[0018] 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. :

[0019] ;

[0020] 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:

[0021] ;

[0022] 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.

[0023] S42. Calculate the density gradient vector and the density Hessian matrix, expressed by the following formulas:

[0024] ;

[0025] ;

[0026] 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;

[0027] S43. Define the density difference coefficient to quantify the local variation characteristics of the density field at data points:

[0028] ;

[0029] 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.

[0030] 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:

[0031] ;

[0032] In the formula, This represents 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;

[0033] 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.

[0034] S45. A density saddle point detection mechanism is introduced to delineate cluster boundaries. The specific process is as follows:

[0035] Define the saddle point determination criteria: point A point is a density saddle point if and only if the following condition is satisfied:

[0036] 1. The magnitude of the density gradient vector approaches zero: In the formula, The gradient threshold for saddle point determination;

[0037] II. Hessian Matrix Both positive and negative eigenvalues ​​exist simultaneously;

[0038] 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:

[0039] ;

[0040] 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.

[0041] As a further improvement to the above scheme, step S4 also includes hierarchical multi-resolution analysis, specifically including:

[0042] 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.

[0043] As a further improvement to the above scheme, in step S4, the following objective function is optimized. Adjust the boundaries of the clusters:

[0044] ;

[0045] 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.

[0046] As a further improvement to the above scheme, step S5 includes the following specific process for feature profile construction and value assessment:

[0047] 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;

[0048] S52. Calculate the multi-factor value score for each cluster, using the following formula:

[0049] ;

[0050] 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 ;

[0051] 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.

[0052] As a further improvement to the above solution, the user demand mining method also includes:

[0053] S6. Based on the verification results of the aforementioned requirements description document, optimize the algorithm parameters. The specific process includes:

[0054] S61. Collect requirement verification result data, including the set of requirements that have passed verification. and the set of requirements that failed validation ;

[0055] 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:

[0056] ;

[0057] 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;

[0058] 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:

[0059] ;

[0060] ;

[0061] 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;

[0062] 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:

[0063] ;

[0064] The weight update formula is as follows:

[0065] ;

[0066] 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;

[0067] S65. Add the feature vectors corresponding to the verified requirements to the historical high-value pattern library for subsequent calculation of historical similarity factors:

[0068] ;

[0069] 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.

[0070] As a further improvement to the above scheme, step S1 specifically includes:

[0071] 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;

[0072] The specific process of step S3 includes:

[0073] 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.

[0074] The present invention also 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, it implements the steps of the user demand mining method based on feature fusion density perception as described above.

[0075] The present invention also discloses a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the user demand mining method based on feature fusion density perception as described above.

[0076] Compared with the prior art, the beneficial effects of the present invention are:

[0077] 1. The user demand mining method based on feature fusion density perception disclosed in this invention breaks through the limitations of traditional methods that only focus on single-dimensional analysis by extracting and integrating features from three key dimensions: semantics, behavior, and context. It achieves a comprehensive capture and understanding of user needs, thereby improving the efficiency and quality of mining potential user needs from rejected data while reducing human resource consumption and analysis costs. It provides reliable technical support for enterprises to discover innovation opportunities from overlooked data and accelerates the product optimization and iteration process.

[0078] 2. This invention designs a density difference-aware clustering algorithm. By calculating density gradients and detecting local density peaks, it can solve the problem that traditional clustering methods easily misjudge niche high-value needs as noise, thereby improving the ability to identify small but cohesive demand clusters.

[0079] 3. This invention automates the entire process from collecting rejection data to assessing demand value, minimizing human intervention and significantly improving the efficiency and scale of demand mining, enabling enterprises to respond to market changes more quickly.

[0080] 4. This invention constructs a demand value assessment system and feedback closed-loop mechanism based on multi-dimensional indicators, enabling the system to 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 demands. Attached Figure Description

[0081] Figure 1 This is a flowchart of the user demand mining method based on feature fusion density perception in Embodiment 1 of the present invention.

[0082] Figure 2 This is an architecture diagram of the user demand mining method based on feature fusion density awareness in Embodiment 1 of the present invention.

[0083] Figure 3 This is a visualization of the clustering results (t-SNE dimensionality reduction) in Embodiment 1 of the present invention.

[0084] Figure 4 This is a histogram of value score distribution in Embodiment 1 of the present invention.

[0085] Figure 5 This is a schematic diagram of the structure of the computer terminal in Embodiment 2 of the present invention. Detailed Implementation

[0086] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0087] Example 1

[0088] To address the numerous challenges in current rejection data analysis, such as limited feature extraction, insufficient identification of niche needs by clustering algorithms, and low efficiency of manual analysis, this invention provides an automatic method for mining new user needs based on feature fusion and density awareness. This method achieves accurate identification and evaluation of high-value user needs in rejection data through the design of a multi-dimensional feature representation mechanism and an improved density clustering algorithm.

[0089] Please see Figure 1 This embodiment provides a user demand mining method based on feature fusion density perception, including steps S1 to S5.

[0090] S1. Collect product rejection data, clean and standardize the data, and extract the metadata.

[0091] Step S1 of this invention involves a comprehensive data collection and preprocessing process. First, an interface connection with the product rejection model is established, and product rejection data is collected in batches in real time or periodically. The product rejection data includes, but is not limited to, the following environmental context information data: (1) User input content: including original input information such as text queries, speech-to-text, and image description text submitted by the user; (2) Operation records: including the user's interface operation sequence before and after submitting the request, such as click, swipe, input, delete, retry, and other behavioral events; (3) Submission time: including time dimension information such as the precise timestamp of the request initiation, the user session start time, and the distribution of historical access time; (4) Device information: including device type, operating system version, application version, screen resolution, network type, and geographical location.

[0092] After data collection is complete, multi-level data preprocessing operations are performed: First, data cleaning is carried out, specifically including:

[0093] (1) Identify and remove invalid data: Identify and remove records that have no practical meaning, such as blank input, garbled input, and test data;

[0094] (2) Outlier handling: Detect and handle outliers in numerical fields. For fields such as operation duration and request frequency, use the interquartile range (IQR) method to identify outliers and decide whether to remove or correct them according to business rules.

[0095] (3) Duplicate record removal: Based on the combination key of user ID, request content and timestamp, identify and merge duplicate submitted records, and retain the most complete data version.

[0096] Then, format standardization is performed to convert data from different sources and types into a unified structured format, specifically including:

[0097] (1) Text standardization: Unify the character encoding to UTF-8 format, and perform text standardization operations such as full-width and half-width conversion, case normalization, and special character cleanup;

[0098] (2) Unified time format: Convert timestamps from different time zones and formats into the ISO8601 standard format of UTC time;

[0099] (3) Numerical normalization: Standardize the continuous numerical fields to eliminate the difference in dimensions; (4) Category coding: Use a unified coding scheme for discrete fields to establish a globally consistent category mapping table.

[0100] Finally, the metadata of the data is extracted and labeled, and the metadata includes:

[0101] (1) Rejection timestamp: Records the precise time when the product rejection determination occurred;

[0102] (2) Rejection reason code: Mark the specific reason category that caused the rejection, such as "out of business scope", "intent recognition failure", "entity extraction failure", "no matching in knowledge base" and other predefined categories;

[0103] (3) User session identifier: Associated with all interaction records of the same user in the same session, supporting subsequent behavior sequence analysis;

[0104] (4) Device type identifier: Marks the type of device from which the request originates, such as mobile terminal, PC terminal, smart terminal, etc.;

[0105] (5) User unique identifier: Under the premise of meeting privacy protection requirements, each user is assigned a unique identifier after desensitization.

[0106] After the above preprocessing steps, the raw rejection data is transformed into structured and standardized data records. Each record contains the original input content, complete contextual information, and rich meta-information labels, providing high-quality data input for subsequent multi-dimensional feature extraction.

[0107] 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.

[0108] This invention employs a three-dimensional feature extraction framework to systematically capture multi-dimensional information from rejection data. This three-dimensional feature extraction mechanism ensures a comprehensive capture of user needs from three complementary perspectives: semantic expression, behavioral patterns, and usage scenarios. At the semantic feature extraction level, this invention uses a bidirectional Transformer architecture optimized for rejection scenarios for deep semantic modeling. Specifically, it includes the following three sub-steps:

[0109] (1) Word embedding representation

[0110] Text description of input rejection data ,in For the first in the text Each token The length of the word sequence is given. First, word embeddings are represented using a pre-trained language model:

[0111] ;

[0112] In the formula, The word embedding matrix has a dimension of . ; For the first Embedding vectors of 1 word element; For word embedding dimension, this embodiment takes Word embedding layers capture the fundamental semantic features of text, mapping discrete lexical symbols to a continuous vector space.

[0113] (2) Contextual semantic enhancement

[0114] Subsequently, contextual semantic enhancement is performed using a deep bidirectional Transformer network:

[0115] ;

[0116] In the formula, express A layer-stacked bidirectional Transformer encoder, in this embodiment, is taken as... ; For general semantic feature vectors, This refers to the hidden layer dimension. The bidirectional Transformer simultaneously models the forward and backward contextual dependencies of the text through a self-attention mechanism, effectively extracting the implicit user intent and demand information within the text. Specifically, the [CLS] position vector output from the last layer of the Transformer, or average pooling of all position vectors, is taken as the semantic representation of the entire text.

[0117] (3) Next, the domain adaptation layer performs domain-specific processing on the general semantic features:

[0118] ;

[0119] In the formula, This is the semantic feature vector after domain adaptation. Semantic feature dimension; For the dimension-reduced projection matrix, As a bottleneck layer dimension, ; To adapt the layer projection matrix; and It is the bias vector; For non-linear activation functions, this embodiment uses the GELU activation function. The domain adaptation layer fine-tunes the model on data specific to a particular business domain, making the model more sensitive to the language expressions of that domain and enabling it to recognize domain-specific terms and expression patterns.

[0120] In terms of behavioral feature extraction, the system retrieves and associates complete behavioral sequences of the same user within the same session from the behavioral log database based on the user session identifier in the metadata. Let the user behavioral sequence be... ,in The length of the behavior sequence, for each behavior event. Represented as an attribute tuple:

[0121] ;

[0122] In the formula, This refers to the behavior type, with values ​​including predefined behavior categories such as click, swipe, input, delete, submit, return, and retry. The timestamp that triggered the action; Duration of the behavior; Identify the target object for the action; Additional parameters for the behavior, such as input content and sliding direction.

[0123] Each behavioral event Mapped to a vector representation using a dedicated encoding function:

[0124] ;

[0125] In the formula, For the first A vector representation of a behavioral event. Embed dimensions for behavior; The behavior type embedding function maps discrete behavior types to dense vectors; The time encoding function uses a sinusoidal positional encoding method to encode the periodic characteristics of the timestamp; This is a duration encoding function that performs a logarithmic transformation on the duration followed by a linear mapping. Embed a function into the target object; This indicates a vector concatenation operation.

[0126] Subsequently, a temporal attention network is used to process the behavioral sequences:

[0127] ;

[0128] The specific computation process of the temporal attention network is as follows:

[0129] ;

[0130] ;

[0131] In the formula, This is a behavioral feature vector; For the first Attention weights for each behavioral event; For the first The time interval between each behavioral event and the current moment; , The weight matrix is ​​a learnable matrix. For attention hidden layer dimensions; This is the attention score vector; This is the bias term. This temporal attention network pays special attention to the temporal relationships and transition patterns of behavioral patterns, assigns higher weight to recent behaviors, and extracts the temporal features of user operating habits and intention expressions.

[0132] At the contextual feature extraction level, the system comprehensively analyzes the user's environmental context information based on the time and environmental attributes in the metadata. Contextual information includes data in four dimensions: device technical parameters, geographic location information, network environment status, and usage time characteristics. Specifically: device technical parameters cover device type, operating system, screen size, memory capacity, etc.; geographic location information includes country, city, latitude and longitude, etc.; network environment status includes network type, network latency, bandwidth, etc.; and usage time characteristics include time of day, day of the week, and whether it is a holiday, etc. For each dimension of contextual data, the system performs feature encoding processing: for categorical features, an embedding encoding method is used to map discrete categories into dense vectors; for numerical features, normalization is performed followed by a linear mapping to convert them into vector representations of the same dimension. Through these encoding methods, the system processes the four dimensions of contextual data to obtain device feature vectors. Geographic feature vectors Network feature vectors and time feature vector Subsequently, the encoded vectors from each dimension are concatenated and nonlinearly fused through a multi-layer perceptron network to learn the interaction relationships between different contextual features, ultimately generating contextual feature vectors that represent the user's usage scenario. .

[0133] S3. Adaptively weighted and fused semantic feature vectors, behavioral feature vectors, and contextual feature vectors to generate multidimensional feature representation vectors.

[0134] First, to ensure the comparability of features across different dimensions, dimension-aligned projection is performed on the three types of feature vectors:

[0135] ;

[0136] ;

[0137] ;

[0138] In the formula, , , The projected feature vectors For a unified fusion feature dimension; , , represents the projection matrix for each dimension.

[0139] Then, attention scores are calculated for each of the three feature vectors to assess their relative importance in the current sample. The calculation of the attention score considers three factors:

[0140] (1) Information entropy factor: measures the information richness of feature vectors

[0141] ;

[0142] In the formula, These are the normalized feature components; A small constant to prevent numerical underflow. Higher information entropy indicates a more uniform feature distribution and a greater amount of information. and The calculation method is the same.

[0143] (2) Historical similarity factor: measures the degree of similarity between current features and historical high-value patterns.

[0144] ;

[0145] In the formula, For the first Prototype vectors of historically high-value semantic patterns; Number of historical patterns; This is the cosine similarity function. and The calculation method is the same.

[0146] (3) Confidence factor: measures the confidence level of the feature extraction module in the current sample.

[0147] ;

[0148] In the formula, and The parameters of the network are used to evaluate confidence levels; For the Sigmoid activation function, the confidence score is mapped to... Interval.

[0149] Taking into account the above three factors, calculate the original attention score for each dimension:

[0150] ;

[0151] ;

[0152] ;

[0153] In the formula, , , The coefficients for the weighting of information entropy, historical similarity, and confidence are given, satisfying the following conditions: The original attention scores are normalized to obtain the attention weights for each dimension:

[0154] ;

[0155] ;

[0156] ;

[0157] Attention weights satisfy the normalization condition: This calculation process ensures that the weight allocation reflects both the characteristics of the current samples and historical patterns.

[0158] Based on the calculated attention weights, perform a weighted fusion operation:

[0159] ;

[0160] In the formula, This is the feature vector after weighted fusion.

[0161] To ensure the numerical stability and comparability of the fused features, the system further applies layer normalization processing:

[0162] ;

[0163] In the formula, This is the final multidimensional feature representation vector; It is the vector mean; The variance is the vector variance. and These are learnable scaling and offset parameters; To prevent division by zero for small constants, layer normalization eliminates dimensional differences and inconsistencies in numerical ranges.

[0164] The advantage of this adaptive fusion mechanism is that it can automatically adjust feature weights for different types of rejected data: give higher weights to semantic features for data with clear semantic expression, strengthen the contribution of behavioral features for data with obvious behavioral patterns, and highlight the influence of contextual features for data with special use scenarios, thereby ensuring that the fused features can retain key information in each dimension to the maximum extent and provide high-quality feature representations for subsequent clustering analysis.

[0165] S4. Cluster the multidimensional feature representation vector using a density difference-aware clustering algorithm. The algorithm constructs a local density field through adaptive bandwidth kernel density estimation, defines a density difference coefficient to quantify the local variation 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.

[0166] This invention designs a density difference-aware clustering algorithm (DD-SCAN), which is specifically optimized for identifying niche but high-value clusters in rejected data. Traditional density clustering algorithms, such as DBSCA, have significant limitations when handling non-uniform density distributions: when high-density and low-density regions coexist in the dataset, using a globally uniform density threshold can lead to valid clusters in low-density regions being incorrectly identified as noise, or multiple clusters in high-density regions being incorrectly merged. This algorithm effectively overcomes this limitation through adaptive density estimation and a density difference-aware mechanism.

[0167] This algorithm first constructs a high-precision local density estimator. For each data point in the feature space... Its local density value is calculated using kernel density estimation (KDE). An adaptive bandwidth Gaussian kernel function is used:

[0168] ;

[0169] In the formula, For data points The local density estimate; For dataset The total number of data points in the data; For any data point in the dataset, sum the results and iterate through the entire dataset. For data points and The Euclidean distance between them; For data points Adaptive bandwidth parameters at the location; Let Gaussian kernel function be defined as: ,in The dimension of the feature vector; This is the normalized distance.

[0170] Adaptive bandwidth parameters Based on data points The local data distribution characteristics are dynamically adjusted, and the calculation formula is as follows:

[0171] ;

[0172] In the formula, The global baseline bandwidth is determined using the Silverman rule: ,in This represents the average of the standard deviations of each dimension of the data. To use fixed bandwidth The estimated initial density value; The geometric mean of the initial density values ​​of all data points. ; This is a sensitivity parameter that controls the bandwidth's response to local density; its value range is... In this embodiment, we take .

[0173] 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.

[0174] 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.

[0175] The density gradient vector is calculated using the central difference method:

[0176] ;

[0177] ;

[0178] 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 1 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:

[0179] ;

[0180] 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:

[0181] ;

[0182] 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.

[0183] Using the differential properties of the density gradient and the Hessian matrix, the density difference coefficient (DDC) is defined:

[0184] ;

[0185] In the formula, For data points The density difference coefficient is a non-negative scalar. It is the Euclidean norm of the density gradient vector, reflecting the intensity of density change; The smallest eigenvalue (algebraically smallest) of a matrix reflects the curvature of the density field in the gentlest direction; This indicates taking the absolute value.

[0186] The DDC index considers both the intensity and curvature of density changes and has the following properties: (1) In regions with flat density, the gradient norm is close to zero, and the DDC value is small; (2) Near density extrema, the gradient norm is close to zero, and the DDC value is small; (3) At the boundaries of regions with rapid density changes, both the absolute values ​​of the gradient norm and curvature are large, and the DDC value is large. Therefore, DDC can effectively identify the boundaries of regions with significant density changes, which are usually natural boundaries between different clusters.

[0187] During cluster formation, an improved density ramping strategy is employed to identify local density maxima as cluster centers. For each data point... The hill-climbing process is carried out along the density gradient direction through the following iterative process:

[0188] ;

[0189] In the formula, For the first The point position and initial value at the next iteration. ; For the first The point position after the next iteration; This is the density gradient vector at the current position; The norm of the gradient vector is used for normalization to obtain the unit direction vector; The adaptive step size parameter is dynamically adjusted based on the magnitude of the local density gradient.

[0190] ;

[0191] In the formula, Base step size; This is the global average gradient norm. When the gradient is large, the step size is close to... When the gradient is small, the step size is reduced accordingly to ensure the stability of the iteration process.

[0192] The iteration termination condition is: , The gradient norm threshold; The largest eigenvalue; when the gradient norm is sufficiently small and all eigenvalues ​​of the Hessian matrix are negative, it indicates that the current position is a local density maximum, and the iteration converges. When the iteration converges, the convergence point is... The data points that converge to the same density maxima are identified as cluster centers. All data points that converge to the same density maxima are grouped into the same initial cluster.

[0193] To accurately identify small, high-value clusters and delineate cluster boundaries, a density saddle point detection mechanism is introduced. A density saddle point is a special critical point in a density field, where density increases in some directions and decreases in others, typically marking the boundary between different clusters. Saddle point identification is based on the eigenvalue structure of the Hessian matrix. For critical points in a density field... (satisfy If its Hessian matrix If both positive and negative eigenvalues ​​exist, then s is a density saddle point.

[0194] For the initially identified adjacent cluster centers and The algorithm searches for density saddle points along the direction of the connecting line. Let... for and A saddle point between two points is determined if the following conditions are met. and Belonging to different clusters:

[0195] ;

[0196] In the formula, This is the density saddle point between the two cluster centers; and These are two adjacent density maxima (cluster centers). Saddle point Local density value at; and Cluster centers and Local density value at; The function is for finding the minimum value; The density ratio threshold parameter controls the maximum ratio of density at saddle points to density at cluster centers, with a value range of [value missing]. In this embodiment, we take ; The DDC threshold parameter controls the minimum intensity of density difference at the boundary; in this embodiment, it is taken as... ,in The value is the average of the DDC values ​​of all data points. The above conditions mean that: (1) the density at the saddle point should be significantly lower than the density of the cluster centers on both sides, indicating that there is a significant density trough between the two clusters; (2) the density difference coefficient at the saddle point should be large enough to indicate that the location is indeed a boundary region where the density changes drastically. If the saddle point does not meet the above conditions, it is considered that the density difference between the two density maxima is insufficient to constitute a cluster boundary, and the corresponding initial clusters should be merged.

[0197] To address data distributions at different scales and ensure the robustness of clustering results, a hierarchical multi-resolution analysis strategy is adopted. The specific steps are as follows:

[0198] (1) Multi-scale density estimation: on multiple different global baseline bandwidths Perform density estimation separately under the following settings: The number of resolution levels is taken in this embodiment. The bandwidth sequence is set according to a geometric progression: .

[0199] (2) Multi-scale clustering: The clustering process is performed independently at each resolution level to obtain... Clustering results .

[0200] (3) Cluster stability analysis: For each pair of clustering results at adjacent resolution levels, the overlap between clusters is calculated: ,in For the first The first layer One cluster; The cardinality (number of elements) of a set.

[0201] (4) Stable cluster selection: Retain cluster structures that are stable across multiple resolution levels. Define clusters. The stability score is: ,in For indicator functions; The overlap threshold is taken as [value] in this embodiment. A higher stability score indicates that the cluster can be identified at multiple scales and is more likely to represent the true data structure. Clusters with stability scores below a threshold are filtered out. The clustering effectively eliminates noisy clusters and pseudo-clusters that only appear at a single resolution.

[0202] Finally, the boundaries of the clusters are finely adjusted by optimizing the following objective function:

[0203] ;

[0204] In the formula, Let be the objective function to be minimized; The final number of clusters; For the first The set of data points contained in each cluster; For clusters Data points in; For the first The center of each cluster (density maxima); and Data points and cluster center The local density value; For the first The cluster and the first Density saddle points between clusters; is the density difference coefficient at the saddle point; The regularization parameter controls the balance between the two losses; in this embodiment, it is taken as... ; The expected lower bound threshold of the density difference coefficient is taken in this embodiment. ; This is the function for finding the maximum value.

[0205] The first term of the objective function is the intra-cluster density consistency loss, which encourages data points within the same cluster to have similar density values, thus forming clusters with uniform density. The second term is the inter-cluster boundary clarity regularization term, which encourages saddle points between different clusters to have higher DDC values, thus forming clear cluster boundaries. The objective function is minimized using the gradient descent method. The cluster assignment of boundary data points is fine-tuned until the objective function converges. This optimization process ensures density consistency within clusters and clear boundaries between clusters, making it particularly suitable for identifying highly cohesive small data clusters that reflect specific user needs in rejected data.

[0206] S5. Construct feature profiles and evaluate the value of the clusters, generate requirement description documents, and optimize algorithm parameters based on the verification results.

[0207] The final step of this invention is the requirement identification and verification process, which aims to transform clustering analysis results into actionable product requirement suggestions and establish a closed-loop optimization mechanism to continuously improve algorithm performance. First, the system performs deep feature analysis on each identified cluster, including: performing word frequency statistics and TF-IDF analysis on the original text of all data points within the cluster to extract Top-N keywords as semantic tags; simultaneously, using a text summarization model to generate a general semantic description of the cluster; statistically analyzing the distribution of behavioral sequence patterns of data points within the cluster to identify high-frequency behavioral sequences and typical behavioral transition paths; and summarizing the contextual feature distribution of data points within the cluster, including equipment type distribution, time period distribution, and geographical distribution.

[0208] Subsequently, the system applies a multi-factor scoring model to evaluate the value of each cluster, taking into account the following five factors:

[0209] (1) Scale factor : Measuring the size of the data in a cluster ,in For clusters The number of data points. The effect of extreme values ​​is smoothed using a logarithmic transformation.

[0210] (2) Cohesion factor : Measures the tightness of clusters ,in This represents the average distance from the data points within a cluster to the center of the cluster. This represents the global average distance. A higher cohesion factor indicates a more tightly clustered group.

[0211] (3) Uniqueness factor : Measures the degree of difference between a cluster and other clusters ,in The uniqueness factor is the center of the cluster. A higher uniqueness factor indicates that the cluster is more distinguishable from other clusters.

[0212] (4) Difference factor : Measuring the gap between the needs represented by clusters and the existing product functionality ,in This represents the average feature vector of the cluster. For the first A feature representation of an existing product function; This refers to the number of existing functions; Cosine similarity. A higher similarity indicates that the demand is more likely to be a new demand not covered by existing products.

[0213] (5) Business value factor Estimating potential business value based on user attributes and behavioral characteristics ,in For data points The value score for each user is calculated based on information such as the user's historical spending, activity level, and membership level.

[0214] Based on the above five factors, the value score of the cluster is calculated:

[0215] ;

[0216] In the formula, For the first The overall value score of each cluster; Let be the weight coefficients of each factor, satisfying This can be adjusted according to business priority; in this embodiment, it is initially set to... , , , , .

[0217] High-value demand candidates are selected based on value scores. ), automatically generate structured requirement description documents. The document contains the following: (1) Requirement background: describe the user pain points and scenarios corresponding to the requirement, automatically generated based on semantic tags and contextual attributes. (2) User group characteristics: describe the typical user profile corresponding to the requirement, including device preferences, usage time periods, geographical distribution, etc. (3) Core functional points: extract the core functions expected by users based on behavioral pattern analysis and semantic tags. (4) Implementation priority suggestions: give priority suggestions based on value scores, scale factors and implementation difficulty estimates. (5) Typical case display: list the selected typical samples to facilitate product personnel to intuitively understand the requirements. For potential requirements with high scores, the system further designs targeted verification schemes: (1) A / B test design: for functional improvements that can be quickly implemented, design the division scheme of experimental and control groups, sample size calculation, core indicator definition, etc. (2) User survey questionnaire: automatically generate questionnaire questions based on requirement characteristics, including requirement confirmation questions, priority ranking questions, functional detail preference questions, etc. (3) Prototype verification scheme: for more complex new functional requirements, design the functional scope and verification indicators of the minimum feasible product prototype.

[0218] S6. After verification, the system feeds back the results to the algorithm framework, forming a closed-loop optimization mechanism:

[0219] (1) Feature fusion weight update: Collect the requirements verification results and distinguish the sets of requirements that have passed the verification. With the set of requirements that failed validation Analyze the feature differences between the two types of needs in three dimensions: semantics, behavior, and context, and adjust the feature fusion weights in step S3 accordingly: ,in For feature dimensions; and These represent the average attention weights for requirements that pass and fail verification in this dimension; The learning rate is used. This update mechanism enhances and validates the feature dimensions and weights relevant to the requirements.

[0220] (2) Clustering threshold parameter adjustment: The clustering threshold parameter is dynamically adjusted according to the difference between the validation pass rate and the target pass rate. and :

[0221] ;

[0222] ;

[0223] In the formula, and The updated density saddle point detection threshold parameters; and The threshold before the update; and These are the adjustment step sizes for β and γ, respectively; This represents the actual pass rate. The threshold is set to the target pass rate. If the pass rate is lower than the target, the threshold is increased to filter out clusters with higher confidence; if the pass rate is higher than the target, the threshold is decreased to discover more potential needs.

[0224] (3) Value assessment model optimization: Based on the validation results, optimize the factor weights of the value scoring model in step S5, and minimize the validation error through gradient descent:

[0225] ;

[0226] ;

[0227] in Optimize the loss function for the weights; The learning rate is the learning rate. This optimization objective assigns higher value scores to requirements that pass validation and lower scores to requirements that fail validation.

[0228] (4) Historical pattern library update: The feature vectors corresponding to the verified requirements are added to the historical high-value pattern library for the calculation of historical similarity factors.

[0229] ;

[0230] 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.

[0231] This comprehensive requirement identification and verification process ensures end-to-end support from data discovery to requirement implementation. Through continuous feedback and optimization, the system can learn and adapt to business characteristics, significantly improving the efficiency and accuracy of product requirement mining.

[0232] This embodiment uses the data rejection function of an e-commerce platform as an example to demonstrate the complete process of the present invention.

[0233] During the data collection and preprocessing phase, the system connects to the e-commerce platform's search engine backend system through a pre-configured data interface to collect search request data from all rejected users within one month, totaling approximately 2 million records. The collected raw data includes key fields such as user query text, session identifiers, user IDs (after anonymization), query timestamps, device information, geolocation codes, and rejection reason codes. The system first performs data cleaning, filtering out obvious anomalous data through set rules, including empty queries, purely numeric queries without semantic meaning, and bot batch requests, reducing the dataset size to 1.5 million records. Subsequently, data standardization is performed, including character normalization of query text, case consistency, and handling of special symbols to ensure consistency in subsequent analysis. Next, rejection-related metadata is extracted, such as rejection reason distribution statistics and time distribution characteristics, providing background reference for subsequent analysis. To balance computational efficiency and analytical effectiveness, the system uses stratified sampling to sample the cleaned data. By using the categories of rejection reasons, device types, and time periods as the basis for stratification, we ensured that the sample ratio of each stratum was consistent with the distribution of the original data. Finally, we extracted 30,000 records as the experimental dataset, which significantly reduced computational costs while ensuring data representativeness.

[0234] In the multidimensional feature extraction stage, the system first processes semantic features, using a bidirectional Transformer network (such as the BERT-base model) fine-tuned for the e-commerce domain to encode each query text, obtaining a word embedding matrix containing contextual information. Subsequently, a 12-layer Transformer encoder is used for contextual semantic enhancement, extracting the output at the [CLS] position as a 768-dimensional general semantic representation. Considering the specificity of e-commerce search scenarios, the system further applies a domain adaptation layer to specialize the semantic representation. This adaptation layer is fine-tuned on labeled data in the e-commerce search domain, enabling better recognition of domain-specific expressions such as brand names, product attribute words, and promotional terms, ultimately outputting a 512-dimensional semantic feature vector. For behavioral features, the system extracts user behavior sequences from user session logs within 30 minutes before and after each search request, including actions such as browsing product categories, clicking product details, adding to cart, and favorites, as well as the intervals and durations of each action. Each behavioral event is encoded as a composite vector containing behavioral type embedding, time location encoding, dwell time encoding, and target object embedding. These behavioral sequences are then encoded into a 192-dimensional behavioral feature vector using a temporal attention network, where a time decay mechanism gives higher attention weight to recent behaviors. For contextual features, the system integrates information such as user device type (e.g., PC, mobile device, smart speaker), query time period (weekday / weekend, morning / afternoon / evening), geographic location (city level, region type), and network environment (WiFi / 4G / 5G). After one-hot encoding or embedding representation, these are fused and encoded into a 128-dimensional contextual feature vector via a three-layer perceptron network.

[0235] In the adaptive feature fusion stage, the system first projects the feature vectors of three different dimensions onto a 512-dimensional common feature space through a linear projection layer, eliminating the impact of dimensional differences. In the attention weight calculation stage, the system dynamically determines the weight of each feature dimension by comprehensively considering three factors: the information entropy factor assesses the information richness of each dimension, with high entropy values ​​indicating that the dimension contains more effective information; the historical similarity factor identifies feature patterns similar to historical verification needs by comparing with a pre-built high-value pattern library; and the confidence factor, output by each feature extraction module, reflects the reliability of feature extraction. These three factors are weighted and normalized using the Softmax function to obtain the dynamic attention weights for each dimension. Correlation analysis reveals that in e-commerce search scenarios, semantic features are highly correlated with users' actual needs, while behavioral features are helpful in judging the urgency of needs, and contextual features play an important role in distinguishing specific needs in particular scenarios. In this embodiment, the average weight of semantic features is approximately 0.45, behavioral features approximately 0.31, and contextual features approximately 0.24. As data analysis progresses, the system automatically adjusts weight allocation based on different types of rejected data. For example, for queries containing explicit technical terms (such as "ANC noise-canceling headphones"), the semantic feature weight is increased to 0.68; for behavioral sequences showing clear purchase intent, the behavioral feature weight is correspondingly increased. The fused feature vector undergoes layer normalization to eliminate dimensional differences and stabilize the numerical distribution, forming a unified 512-dimensional feature representation for subsequent cluster analysis.

[0236] In the adaptive bandwidth kernel density estimation stage, the system first calculates the global baseline bandwidth using the Silverman rule, obtaining an initial value of approximately 0.28. Considering the non-uniformity of the rejected data distribution, the system calculates an adaptive bandwidth for each data point: a smaller bandwidth is used in densely populated areas to maintain resolution, and a larger bandwidth is used in sparsely populated areas to ensure estimation stability. Based on the Gaussian kernel function of the adaptive bandwidth, the system calculates the local density values ​​of all data points. Subsequently, the system calculates the density gradient and Hessian matrix. The central difference method is used to calculate the density gradient vector for each data point, representing the direction and rate of density change; simultaneously, the Hessian matrix of the density function is calculated, and its eigenvalues ​​are analyzed to determine the geometric characteristics of the point. Based on the density value, gradient magnitude, and Hessian matrix eigenvalues, the system calculates the density difference coefficient (DDC) for each data point. This coefficient exhibits a higher value in the cluster boundary region, 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, with the step size adaptively adjusted according to the DDC value. A larger step size is used in low DDC regions to accelerate convergence, while a smaller step size is used in high DDC regions for precise positioning. Iteration stops when the gradient magnitude is less than a set threshold. Data points converging to the same position are grouped into the same cluster, and the convergence endpoint is the candidate cluster center. In this embodiment, a total of 127 candidate cluster centers were identified. The system further performs density saddle point detection and cluster merging determination. For adjacent candidate cluster centers, a density minimum point (saddle point) is searched on the line connecting them. If the density of the saddle point is lower than a certain proportion of the density of the cluster centers on both sides (set to 50% in this embodiment), and the DDC value of the saddle point exceeds 10% of the mean, then the two sides are determined to be independent clusters; otherwise, the two candidate clusters are merged. After this processing, the 127 candidate clusters are merged into 89 preliminary clusters. To improve clustering stability, the system adopts a hierarchical multi-resolution analysis method. Clustering is performed with 5 different baseline bandwidths (increasing geometrically from 0.14 to 0.57), and the performance of each cluster at different resolutions is observed. If a cluster is stable across at least three resolution levels (overlap exceeding 70%), it is considered stable and retained; otherwise, it is filtered out. In this embodiment, 72 clusters passed the stability test. Finally, cluster boundary optimization is performed. The system constructs an optimization function targeting intra-cluster compactness and inter-cluster segregation, introducing DDC as a regularization constraint. Gradient descent iteratively adjusts the assignment of data points in boundary regions, ensuring that cluster boundaries are distributed along high DDC regions as much as possible. After 100 rounds of iterative optimization, 72 clusters with clear boundaries are finally formed. The size distribution of these clusters exhibits a long-tail characteristic, with the largest cluster containing approximately 7600 data points, while most clusters range in size from 50 to 500 data points, reflecting the diversity and niche nature of user needs. Cluster quality assessment shows a silhouette coefficient of 0.412, indicating good cohesion and segregation in the cluster structure.

[0237] In the demand identification and verification phase, the system conducts multi-dimensional analysis and value assessment of 72 clusters. The system first extracts a feature profile for each cluster, including high-frequency keywords and TF-IDF weight distribution, shared behavioral pattern sequences, main usage scenario characteristics, and user group profiles, forming a complete cluster feature description. In the multi-factor value assessment stage, the system comprehensively considers five value factors: the scale factor reflects the number of potentially influential users, and logarithmic normalization is used to avoid excessive dominance by large clusters; the cohesion factor measures the consistency of data within the cluster, with high cohesion indicating a clear and concentrated demand; the uniqueness factor assesses the degree of difference between the data within the cluster and the global distribution, with high uniqueness indicating a distinctive demand; the difference factor measures the gap between the demand and existing functions, with high difference implying significant room for product improvement; and the commercial value factor comprehensively considers commercial indicators such as user spending power and purchase intention. Based on the business characteristics of e-commerce platforms, the system sets the weight of the difference factor to 0.25, the weights of cohesion and uniqueness factors to 0.20 each, the weight of the commercial value factor to 0.20, and the weight of the scale factor to 0.15. The weighted sum of all factors yields the comprehensive value score. The scoring results show that three clusters have particularly high business value: The first high-value cluster (overall score 0.828) is characterized by users attempting to search for multiple related products simultaneously (such as "red dress with silver high heels"), reflecting the demand for "combined product search." This cluster contains approximately 2,800 data points, with users primarily being women aged 25-35, and 43% using it on weekend afternoons. The second high-value cluster (overall score 0.793) mainly reflects search failures caused by users using dialects or heavily accented voice input, indicating the necessity of "dialect voice recognition" functionality. This cluster contains approximately 1,500 data points, with 34% using smart speaker devices and 71% being users from third- and fourth-tier cities. The third high-value cluster (overall score 0.789) shows a pattern of users searching for seasonal products before they are available, reflecting the demand for "early display of seasonal products." Users in this cluster have higher spending power, with an average order value exceeding the platform average by 42%. The system automatically generates detailed requirement documents for these three high-value requirements, including a description of the requirement background, characteristics of the target user group, core functionalities, typical case studies, and implementation priority suggestions, providing the product team with a complete decision-making reference. Simultaneously, the system designs a targeted A / B testing verification plan: for the "combined product search" function, 10% of platform users are selected as the test group, and 10% as the control group, conducting a two-week functional test. Key observation indicators include search success rate, click-through rate, conversion rate, average order value, and dwell time. Based on the verification results, the system feeds the verification data back to the algorithm framework to optimize clustering parameters and feature weights, improving the system's sensitivity to identifying similar requirement patterns.The entire demand mining process, from data collection to demand verification, takes significantly less time than traditional manual analysis methods. It also uncovers niche but high-value demands that might be overlooked by expert methods, fully demonstrating the practical value and technological advancement of this invention.

[0238] To verify the effectiveness of the method proposed in this invention, this embodiment compares the DD-SCAN algorithm with three mainstream clustering algorithms, and performs a visual analysis of the clustering results, as shown in Table 1:

[0239] Table 1: Comparison of Clustering Results by Algorithm

[0240]

[0241] As shown in the table, the proposed DD-SCAN identifies 52 clusters, significantly more than K-means' 15 and DBSCAN's 28, indicating that this algorithm can discover more subdivided categories of user needs. Regarding small cluster identification, DD-SCAN successfully identified 23 small clusters (less than 100 samples), nearly twice that of HDBSCAN, which is crucial for uncovering niche but high-value long-tail needs. In terms of silhouette coefficient, DD-SCAN reaches 0.433, indicating better cohesion and separability in the clustering results.

[0242] To visually demonstrate the clustering effect, this embodiment uses the t-SNE algorithm to reduce the 512-dimensional feature vector to two dimensions for visualization, with the perplexity parameter set to 30. Figure 3 The distribution of clustering results from the DD-SCAN algorithm is shown, with different colors representing different clusters. A red pentagram marks the three high-value demand clusters identified after value assessment. The figure shows that the DD-SCAN algorithm successfully identified multiple clusters distributed across various regions of the feature space, including several small but clearly defined clusters, validating the algorithm's adaptability to density differences. The three high-value demand clusters are located in different regions of the feature space, indicating significant characteristic differences among these demands. Figure 4 The histogram shows the value score distribution of 52 clusters. The score distribution approximates a normal distribution, 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; clusters exceeding this threshold are judged as high-value potential needs. Statistical results show that three clusters (accounting for 5.8%) have value scores exceeding the threshold, corresponding to three types of needs: "combined product search" (V=0.828), "dialect speech recognition" (V=0.793), and "pre-display of seasonal products" (V=0.789). These needs will be prioritized and pushed to the product team for manual review and A / B testing verification.

[0243] The experimental results above demonstrate that the density difference-aware clustering method proposed in this invention can effectively identify hidden user demand patterns in rejected data, particularly showing significant advantages in niche demand mining and cluster quality. Through a multi-factor value assessment model, the system can accurately screen potential demands with high commercial value from a large number of clusters, providing data-driven decision support for product iteration.

[0244] The core innovations of this invention are mainly reflected in two aspects: First, a multi-dimensional feature representation and fusion mechanism is proposed, which can comprehensively capture user demand features from three key dimensions: semantics, behavior, and context, and effectively integrate features from different dimensions through adaptive weight calculation. Second, a novel density difference-aware clustering algorithm is designed, which accurately identifies small, high-value demand clusters that are easily overlooked by traditional clustering methods through density gradient calculation and local density peak detection. These innovations enable this invention to significantly improve analysis efficiency and coverage while enhancing the accuracy of demand mining.

[0245] Example 2

[0246] This 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, it implements the steps of the user demand mining method based on feature fusion density perception as described in Embodiment 1.

[0247] like Figure 5 As shown, the computer terminal provided in this embodiment includes: at least one processor 101, and a memory 102 connected to at least one processor 101. This embodiment does not limit the specific connection medium between the processor 101 and the memory 102. Figure 5 The example shown is the connection between processor 101 and memory 102 via bus 100. Bus 100 is... Figure 5 The connections between other components are shown in bold lines and are for illustrative purposes only, not as limiting information. Bus 100 can be divided into address bus, data bus, control bus, etc., for ease of representation. Figure 5 The bus is represented by a single thick line, but this does not indicate that there is only one bus or one type of bus. Alternatively, the processor 101 may also be called a controller; there is no restriction on the name.

[0248] In this embodiment, the memory 102 stores instructions that can be executed by at least one processor 101. The at least one processor 101 can execute the aforementioned method by executing the instructions stored in the memory 102.

[0249] The processor 101 is the control center of the device. It can connect to various parts of the control device through various interfaces and lines. By running or executing instructions stored in memory 102 and calling data stored in memory 102, the processor can perform various functions and process data, thereby monitoring the device as a whole.

[0250] In one possible design, processor 101 may include one or more processing units. Processor 101 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user interface, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may also not be integrated into processor 101. In some embodiments, processor 101 and memory 102 may be implemented on the same chip; in some embodiments, they may also be implemented on separate chips.

[0251] Processor 101 can be a general-purpose processor, such as a central processing unit (CPU), digital signal processor, application-specific integrated circuit, field-programmable gate array or other programmable logic device, discrete gate or transistor logic device, or discrete hardware component, capable of implementing or executing 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 user demand mining method based on feature fusion density awareness disclosed in Embodiment 1 can be directly manifested as execution by a hardware processor, or execution by a combination of hardware and software modules in processor 101.

[0252] 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. Memory 102 may include at least one type of storage medium, such as 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 storage, magnetic disk, optical disk, etc. Memory 102 can be any other medium capable of carrying or storing desired program code in the form of instructions or data structures that can be accessed by a computer, but is not limited thereto. In this embodiment, memory 102 can also be a circuit or any other device capable of implementing storage functions for storing program instructions and / or data.

[0253] By designing and programming the processor 101, the code corresponding to the user demand mining method based on feature fusion density awareness described in the foregoing embodiments can be embedded into the chip, thereby enabling the chip to execute the code during runtime. Figure 1 The steps of the user demand mining method based on feature fusion density awareness are shown. How to design and program the processor 101 is a technique well-known to those skilled in the art and will not be described further here.

[0254] Example 3

[0255] This embodiment provides a computer-readable storage medium storing a computer program thereon. When the program is executed by a processor, it implements the steps of the user demand mining method based on feature fusion density perception as described in Embodiment 1.

[0256] The computer-readable storage medium may include flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the storage medium may be an internal storage unit of a computer device, such as the hard disk or memory of the computer device. In other embodiments, the storage medium may also be an external storage device of the computer device, such as a plug-in hard disk, smart memory card, secure digital card, flash memory card, etc., provided on the computer device. Of course, the storage medium may include both internal storage units and external storage devices of the computer device. In this embodiment, the memory is typically used to store the operating system and various application software installed on the computer device. In addition, the memory can also be used to temporarily store various types of data that have been output or will be output.

[0257] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

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 saddle point determination; 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, 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