Multi-source data fusion analysis system for e-commerce digital insights

Abstract

This invention provides a multi-source data fusion and analysis system for e-commerce digital insights, relating to the field of data processing technology. The system includes: selecting benchmark analysis units within a spatial relationship structure; calculating the fractal dimension features of feature vectors; performing multi-level recursive partitioning of the analysis units based on the fractal dimension values, with each level maintaining self-similarity to obtain feature partitions; mapping the feature vector distribution to the corresponding feature partitions; calculating the spatial distribution weights of each feature partition based on its vector density distribution and spatial positional relationship; adjusting the parameters of a deep learning network using these partition spatial distribution weights to obtain value assessment results; and processing the value assessment results through a multimodal fusion algorithm, generating personalized recommendation strategies based on a reinforcement learning framework, and calculating resource and intelligent routing algorithms to complete the digital insight analysis process. This invention can improve the efficiency of e-commerce data utilization and decision-making effectiveness.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a multi-source data fusion and analysis system for e-commerce digital insights. Background Technology

[0002] In the operation of e-commerce platforms, digital insights based on multi-source data such as user behavior, product information, and supply chain status are the core of achieving precise marketing and personalized recommendations. Currently, mainstream analysis systems typically rely on processing single-dimensional interactive behavior data such as user clicks and browsing, combined with structured information such as product tags, and using machine learning models for value assessment and recommendations. However, these existing technologies still have significant limitations when dealing with complex e-commerce scenarios. Specifically, e-commerce platforms generate massive amounts of multi-source heterogeneous data during operation, such as unstructured texts like product reviews and Q&A containing user subjective opinions, as well as dynamic supply chain data reflecting inventory and logistics timeliness. There are deep, non-linear intrinsic connections between these data and user behavior data. Existing methods mostly use simple data splicing or static weighted fusion to process them, making it difficult to fully capture and quantify the complex dynamic connections between these heterogeneous data sources. For example, the appearance of a negative review may instantly change the purchasing decision of a high-intent user, and supply chain fluctuations (such as logistics delays caused by a sudden epidemic in the shipping location) will also affect users' acceptance of products in real time. These dynamic and implicit connections are easily ignored or underestimated under the existing technological framework. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a multi-source data fusion and analysis system for e-commerce digital insights, which can improve the efficiency of e-commerce data utilization and decision-making effectiveness.

[0004] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0005] The first aspect is a multi-source data fusion and analysis system for e-commerce digital insights, including:

[0006] The data acquisition module is used to collect interaction behavior data across the entire chain and integrate unstructured text data and supply chain dynamic data to construct a multi-source heterogeneous raw dataset.

[0007] The processing module is used to preprocess and feature process multi-source heterogeneous raw datasets, including semantic parsing based on deep natural language processing and feature extraction based on time-series analysis, to generate a standardized feature matrix.

[0008] The mapping module is used to map the standardized feature matrix to a high-dimensional feature space to form a feature vector distribution. By constructing a convex polygon representation of each feature vector, it calculates the overlapping area between different feature vector distribution regions, quantifies the degree of overlap of feature distributions, and generates a spatial relationship structure.

[0009] The segmentation module is used to select the benchmark analysis unit in the spatial relationship structure. By calculating the fractal dimension features of the feature vector, the analysis unit is recursively divided into multiple levels based on the fractal dimension value. Each level of division maintains self-similarity characteristics, resulting in feature partitions.

[0010] The calculation module is used to map the feature vector distribution to the corresponding feature partitions, and calculate the spatial distribution weight of each feature partition based on the vector density distribution and spatial position relationship of each feature partition.

[0011] The adjustment module is used to adjust the parameters of the deep learning network using the weights distributed in the partition space to obtain the value evaluation results.

[0012] The output module processes the value assessment results through a multimodal fusion algorithm, generates personalized recommendation strategies based on a reinforcement learning framework, and computes resource and intelligent routing algorithms to complete the digital insight analysis process.

[0013] The above-described solution of the present invention has at least the following beneficial effects:

[0014] By leveraging semantic parsing from deep natural language processing, feature extraction from temporal analysis, and nonlinear dimensionality reduction and spatial relationship construction techniques, the system can uncover deep nonlinear relationships between multi-source data, effectively avoiding the problem of insufficient relationship capture caused by static weighting. Combining feature partitioning of recursive segmentation with spatial distribution weight calculation, the dynamic changes in data relationships are further quantified, enabling the effective identification of implicit relationships (such as the relationship between comment sentiment and purchase decision, and the relationship between supply chain fluctuations and user acceptance). By optimizing the parameters of the deep learning network through partitioned spatial distribution weights, the model can focus on high-value data partitions, reduce irrelevant interference, and make the value assessment results more consistent with the actual e-commerce operation scenario. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of a multi-source data fusion and analysis system for e-commerce digital insights provided in an embodiment of the present invention.

[0016] Figure 2 This is a flowchart illustrating the process of adjusting the parameters of a deep learning network using partitioned spatial distribution weights to obtain value evaluation results, provided by an embodiment of the present invention. Detailed Implementation

[0017] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0018] like Figure 1 As shown, embodiments of the present invention propose a multi-source data fusion and analysis system for e-commerce digital insights, comprising:

[0019] The data acquisition module is used to collect interaction behavior data across the entire chain and integrate unstructured text data and supply chain dynamic data to construct a multi-source heterogeneous raw dataset.

[0020] The processing module is used to preprocess and feature process multi-source heterogeneous raw datasets, including semantic parsing based on deep natural language processing and feature extraction based on time-series analysis, to generate a standardized feature matrix.

[0021] The mapping module is used to map the standardized feature matrix to a high-dimensional feature space to form a feature vector distribution. By constructing a convex polygon representation of each feature vector, it calculates the overlapping area between different feature vector distribution regions, quantifies the degree of overlap of feature distributions, and generates a spatial relationship structure.

[0022] The segmentation module is used to select the benchmark analysis unit in the spatial relationship structure. By calculating the fractal dimension features of the feature vector, the analysis unit is recursively divided into multiple levels based on the fractal dimension value. Each level of division maintains self-similarity characteristics, resulting in feature partitions.

[0023] The calculation module is used to map the feature vector distribution to the corresponding feature partitions, and calculate the spatial distribution weight of each feature partition based on the vector density distribution and spatial position relationship of each feature partition.

[0024] The adjustment module is used to adjust the parameters of the deep learning network using the weights distributed in the partition space to obtain the value evaluation results.

[0025] The output module processes the value assessment results through a multimodal fusion algorithm, generates personalized recommendation strategies based on a reinforcement learning framework, and computes resource and intelligent routing algorithms to complete the digital insight analysis process.

[0026] In this embodiment of the invention, by leveraging semantic parsing of deep natural language processing, feature extraction of temporal analysis, and nonlinear dimensionality reduction and spatial relationship construction techniques, the system can uncover deep nonlinear relationships between multi-source data, effectively avoiding the problem of insufficient relationship capture caused by static weighting. By combining feature partitioning of recursive segmentation with spatial distribution weight calculation, the dynamic changes of data relationships are further quantified, allowing implicit relationships (such as the relationship between comment sentiment and purchase decision, and the relationship between supply chain fluctuations and user acceptance) to be effectively identified. By optimizing the parameters of the deep learning network through partitioned spatial distribution weights, the model can focus on high-value data partitions, reduce irrelevant interference, and make the value assessment results more consistent with the actual e-commerce operation scenario.

[0027] In a preferred embodiment of the present invention, collecting end-to-end interaction behavior data and integrating unstructured text data and supply chain dynamic data to construct a multi-source heterogeneous raw dataset may include:

[0028] The system collects end-to-end interactive behavior data to obtain a raw behavioral data stream containing page browsing trajectories, product click sequences, and transaction operation logs. Specifically, it captures users' page browsing trajectory data in real time through seamless tracking scripts embedded in the e-commerce app and webpage frontend. This data covers the URLs of product detail pages visited by users, the duration of each page stay, the percentage position scrolled to, and operation records such as clicks on the "favorite" button and the "add to cart" button. Simultaneously, it captures users' product click operations through the e-commerce platform's backend interface listening service, recording the unique identifier of each click, the precise time of the click, the source page redirected before the click, and the target page path after the click, forming a continuous product click sequence. In addition, it extracts users' transaction operation logs by connecting to the core database of the e-commerce platform's transaction system, including the order time, unique order number, specific list of purchased products, selected payment channel, actual payment amount, and all status changes of the order from creation to delivery completion. The above page browsing trajectories, product click sequences, and transaction operation logs are integrated into a raw behavioral data stream in chronological order of data generation.

[0029] Based on the raw behavioral data stream, device identification and session reconstruction algorithms are used for data aggregation to obtain complete interactive session sequence data. Based on the interactive session sequence data, dynamic monitoring data from each link of the supply chain is collected to obtain a supply chain data set. Specifically, based on the raw behavioral data stream, unique feature information of user terminal devices is extracted through device identification technology, including device hardware model, operating system version number, browser kernel type, IP address hashed feature value, and hardware parameter combination of device sensors. The IP address hash encryption is to perform a hash operation on each field of the IP address and take the first 16 bits of the feature value combination. At the same time, the session reconstruction algorithm is used to aggregate the interactive data corresponding to the same device identifier. The session validity period is set to automatically terminate if there is no operation for 30 consecutive minutes after the user first accesses the platform. The timer starts from the time when the user first initiates a page request. If no interactive operation request is detected from the device within 30 consecutive minutes, the session is determined to be terminated. All page views, product clicks, and transaction operations within the validity period are concatenated in chronological order to form complete interactive session sequence data.

[0030] Based on the unique product identifiers and order numbers contained in the interactive session sequence data, the system connects to the supply chain management module of the e-commerce platform to collect dynamic monitoring data from each link of the supply chain. Specifically, it collects real-time data on the current inventory quantity, shelf location, inbound inspection time, and outbound packaging time of the products through IoT sensors deployed in various warehousing centers. It also obtains the logistics and transportation status data of the products by calling the supply chain data set API provided by logistics partners, including the merchant's shipping time, logistics transit node name, current real-time location of the package, estimated delivery time, and abnormal records such as delays and damages during transportation. At the same time, it collects information on the supplier's production plan progress, raw material inventory balance, and replenishment cycle. These inventory data, logistics data, and supplier data are then categorized and integrated into a supply chain data set.

[0031] Based on interactive conversation sequence data and supply chain data sets, unstructured text data is collected to obtain a text corpus. Data association processing is performed on the interactive conversation sequence data, supply chain data sets, and text corpus to obtain multi-source associated data. This data is then standardized to construct a multi-source heterogeneous raw dataset. Specifically, this includes: based on user unique identifiers, product unique identifiers, and order numbers from the interactive conversation sequence data, connecting to the e-commerce platform's comment management module, Q&A community database, and customer service chat log storage module to collect unstructured text data. This includes user reviews after purchasing products, questions raised by users in Q&A communities regarding product functions, usage methods, and after-sales policies, as well as replies from other users or platform officials, and real-time chat logs between users and the platform's online customer service. This text data undergoes preliminary organization, removing meaningless garbled characters and duplicate text content to form a structured text corpus. The final structured text corpus is based on user unique identifiers, product unique identifiers, and order numbers. As a core key field for association, a mapping relationship is established between interactive session sequence data, supply chain data sets, and text corpora. For example, the interactive session sequence data corresponding to the same user's unique identifier is associated with the product review text data published by that user; the supply chain data corresponding to the same product's unique identifier is associated with the user's click, add-to-cart, and purchase behavior data for that product; and the transaction operation logs corresponding to the same order number are associated with the logistics and transportation data of that order, thus obtaining multi-source associated data. The multi-source associated data is then standardized, unifying the time recording format of all data to UTC standard time, the character encoding format of all data to UTF8, and the units of all numerical data are unified. For example, the inventory quantity is unified to the number of pieces, the logistics and transportation time in minutes is divided by 60 to convert to hours, and the product price is unified to yuan. The text data undergoes a unified character encoding conversion and whitespace characters and invalid symbols are removed. Finally, a multi-source heterogeneous original dataset containing multiple types of data is constructed.

[0032] This embodiment achieves comprehensive collection and deep correlation of full-link behavioral data, unstructured text data, and supply chain dynamic data. By using specific identification association and unit unification methods, it breaks down information barriers between different types of data sources, fully integrates the core value of various types of data, and avoids the analytical limitations caused by a single data source.

[0033] In a preferred embodiment of the present invention, preprocessing and feature processing are performed on the multi-source heterogeneous original dataset, including semantic parsing based on deep natural language processing and feature extraction based on temporal analysis, to generate a standardized feature matrix, which may include:

[0034] The process involves preprocessing the multi-source heterogeneous original dataset to obtain a purified standard dataset. Based on this purified standard dataset, text corpus data is extracted, and text features are extracted to obtain semantic feature vectors. Specifically, this includes preprocessing the multi-source heterogeneous original dataset, encompassing three stages: data cleaning, data deduplication, and data completion. During data cleaning, data records lacking core fields such as unique user identifiers, unique product identifiers, or timestamps are directly removed. For numerical data, statistical analysis methods are used to identify outliers. First, the mean of all valid data for the numerical field is calculated. Then, the average of the squares of the differences between each data point and the mean is calculated to obtain the variance. The square root of the variance is then taken to obtain the standard deviation. If a value is greater than the mean plus three times the standard deviation, or less than the mean minus three times the standard deviation, it is considered an outlier. For constant values, outliers are removed. During data deduplication, the core field combination of each data entry is compared. If two data entries have identical user unique identifier, product unique identifier, timestamp, and operation type, they are considered duplicate data. One of them is retained and the remaining duplicate records are deleted. During data completion, missing non-core numerical fields are filled with the mean of all valid data for that field. First, the sum of all valid data for that field is calculated, and then divided by the number of valid data entries to obtain the mean. This mean is used to fill in the missing values. For missing non-core categorical fields, the mode with the highest frequency of occurrence in that field is used to fill in the missing values. The frequency of each categorical value in that field is counted, and the categorical value with the highest frequency is the mode. This mode is used to fill in the missing values. After the above processing, the purified standard dataset is obtained.

[0035] Based on the purified standard dataset, extract the text corpus data for text feature extraction. First, perform word segmentation. Adopt a word segmentation strategy that combines dictionary matching and statistical language models to split continuous text sequences into independent word units. Subsequently, remove common function words that have no actual semantic contribution through a preset stop word list, such as "de", "le", "he", "yu", etc.; construct a word embedding model for semantic feature transformation. Use all non-repeated words in the text corpus as the training samples of the model. Set the dimension of the word vectors output by the model to 256 dimensions. Adopt the SkipGram model architecture to build the network structure. The goal of model training is to maximize the co-occurrence probability of each target word and its adjacent context words. Iteratively optimize the model parameters through the stochastic gradient descent algorithm. Each time an iteration is performed, first select a batch of training samples, calculate the error between the predicted co-occurrence probability and the actual co-occurrence probability of the model, multiply the error by the learning rate to obtain the parameter adjustment amount, and subtract the adjustment amount from the word vector parameters to complete one update. Continuously iterate until the change in the loss function value of the model is less than 0.001 for 5 consecutive iterations. At this time, obtain the low-dimensional dense word vector representation corresponding to each word; for each text data, perform a weighted sum calculation on the word vectors corresponding to all the words it contains, introduce an attention mechanism to strengthen the weights of key semantic information. Key semantic words include evaluation words such as "good quality", "slow logistics", "high price" that directly reflect user attitudes or product attributes. First, count the number of occurrences of each word in each text. Combine with a preset key semantic word library. If a word is in the library, its weight is 1.5 times the weight of ordinary words. The weight of ordinary words is set to 1. Multiply the word vector of each word by the corresponding weight, add all the weighted word vectors, and then divide by the total number of words to obtain the semantic feature vector of this text.

[0036] Based on the purified standard dataset, interactive session sequence data is extracted and time-series feature mining is performed to obtain temporal feature vectors. Based on semantic and temporal feature vectors, a feature fusion algorithm is used to integrate multi-source features, resulting in a multi-source integrated feature set. Specifically, this includes: extracting interactive session sequence data from the purified standard dataset for time-series feature mining; firstly, each interactive session sequence is divided into fixed 10-minute time windows, and basic statistical features such as page views, product clicks, average page dwell time, and number of items added to cart are statistically analyzed within each time window. The average page dwell time is calculated by summing the page dwell times within the window and dividing by the number of page views. Then, the temporal correlation characteristics of the interactive behavior are analyzed. The time interval between two adjacent product clicks is calculated, i.e., the time of the second click minus the time of the first click. The click frequency change value of the same product in different consecutive time windows is calculated, i.e., the click frequency of the second window minus the click frequency of the first window. The click frequency is the number of clicks in the window divided by 10. The time span from the first browsing of a product to the final order placement is calculated, i.e., the order placement time minus the first browsing time. At the same time, a sequence pattern recognition method is used to mine the regular patterns of user behavior, such as the order of product categories that users browse continuously, and the complete behavioral path pattern from clicking a product to adding it to the shopping cart and then placing an order. After these basic statistical features, temporal derived features and sequence pattern features are quantized and encoded respectively, they are combined to form the temporal feature vector corresponding to each interactive session sequence.

[0037] Based on the obtained semantic feature vector and temporal feature vector, a weighted summation feature fusion algorithm is used to integrate multi-source features. The weight allocation is based on the actual contribution of the two types of features to e-commerce digital insights. By analyzing the accuracy of the two types of features in predicting user purchase decisions in historical data, the weight of semantic features is determined to be 0.4 and the weight of temporal features is determined to be 0.6. The first dimension value of the semantic feature vector is multiplied by 0.4, and the first dimension value of the temporal feature vector is multiplied by 0.6. The two results are added together to obtain the first dimension value of the integrated feature vector. All dimension values ​​are calculated in this way to obtain the multi-source integrated feature set.

[0038] Based on a multi-source integrated feature set, normalized feature data is obtained through feature normalization. Based on the normalized feature data, dimensional unification and format standardization are performed to generate a standardized feature matrix. Specifically, this includes: First, based on the multi-source integrated feature set, feature normalization eliminates dimensional differences between different feature dimensions. Using the min-max normalization method, the maximum and minimum values ​​of all feature values ​​under a given feature dimension are first identified. For each feature value, the minimum value is subtracted from the current feature value to obtain the difference. Then, the difference is divided by the result of the maximum minus the minimum value, mapping all feature values ​​to the interval between 0 and 1, thus obtaining normalized feature data. Second, based on the normalized feature data, dimensional unification is performed. The number of dimensions of all feature vectors is counted, and the maximum number of dimensions is identified. For feature vectors with a number of dimensions less than the maximum number of dimensions, zeros are added sequentially to the end of the vector until the number of dimensions equals the maximum number of dimensions, ensuring that all feature vectors have consistent dimensions. Simultaneously, all feature data is format standardized, with the data type uniformly set to floating-point. All feature vectors are arranged in rows and organized into a matrix form, ultimately generating a standardized feature matrix.

[0039] This embodiment achieves accurate semantic parsing of unstructured text data through specific word vector training and weighted calculation in deep natural language processing technology, mines the temporal correlation features of user behavior through detailed window division and feature calculation in time series analysis, and completes feature fusion and standardization processing by combining clear weight allocation and normalization calculation, effectively capturing the deep nonlinear correlation between multi-source heterogeneous data.

[0040] In a preferred embodiment of the present invention, the standardized feature matrix is ​​mapped to a high-dimensional feature space to form a feature vector distribution. By constructing a convex polygon representation of each feature vector, the overlapping area between different feature vector distribution regions is calculated to quantify the degree of overlap of the feature distributions and generate a spatial relationship structure, which may include:

[0041] Processing is based on a standardized feature matrix, mapping high-dimensional features to a low-dimensional manifold space to obtain the distribution of feature vectors in the low-dimensional space. Specifically, when processing each feature vector, the importance coefficient of each dimension is calculated. The calculation process is as follows: First, calculate the variance of each dimension by taking the difference between all sample values ​​in that dimension and the mean of that dimension, squared each difference, summing them, and then dividing by the total number of samples to obtain the variance of that dimension. Second, calculate the sum of the variances of all dimensions, divide the variance of each dimension by this sum, and obtain the importance coefficient of that dimension. In e-commerce scenarios, this applies to the sentiment dimension and behavioral temporal dimension of text semantics. The importance coefficients of key dimensions such as click frequency and supply chain inventory are typically between 0.1 and 0.2. After multiplying the value of each dimension of each feature vector by the corresponding importance coefficient, it is mapped to a two-dimensional or three-dimensional low-dimensional manifold space according to the neighborhood retention rule. The neighborhood retention rule is that if the Euclidean distance between any two feature vectors in the high-dimensional space is less than 0.2, then the Euclidean distance between the two in the low-dimensional space must be kept between 0.18 and 0.22 to ensure that feature vectors with similar user behavior or similar product attributes are still in similar positions in the low-dimensional space, and finally the distribution of feature vectors in the low-dimensional space is obtained.

[0042] Based on the distribution of feature vectors in low-dimensional space, the outer boundary points of each feature vector cluster are identified, resulting in a set of boundary points for each cluster. Based on this set, all boundary points are sorted clockwise to obtain an ordered sequence. Specifically, this involves: clustering all feature vectors in low-dimensional space. The first step calculates the Euclidean distance between each feature vector and all other feature vectors. Taking a two-dimensional vector as an example, with coordinates x1y1 for vector A and x2y2 for vector B, the difference between x1 and x2 is calculated and squared, and the difference between y1 and y2 is also squared. The square root of the sum of these two squared values ​​is then used to obtain the Euclidean distance. The second step calculates the average Euclidean distance between all pairs of vectors. This average is used as a set distance, and the number of other vectors within this set distance range around each vector is counted. The third step calculates the total distance between all vectors... The average number of surrounding vectors is used to group vectors with a number greater than this average into the same cluster, forming multiple feature vector clusters, such as product feature clusters with high click-through rates and high conversion rates, and user feature clusters with high-frequency browsing and low conversion rates. The geometric center of each cluster is calculated. In a two-dimensional scene, the sum of the x-coordinates of all vectors in the cluster is calculated first, and then divided by the number of vectors in the cluster to obtain the center x-coordinate. The center y-coordinate is calculated in the same way. The distance from each vector to the geometric center is calculated, and the average of all distances is calculated. Vectors with a distance greater than this average multiplied by 1.2 are marked as boundary point candidates. Points with a mutual distance less than a set threshold are removed from the candidate points. This threshold is the average of the pairwise distances of all vectors in the cluster divided by 2. The calculation method is to sum the pairwise distances of all vectors in the cluster, divide by the logarithm of the vectors to obtain the average, and then divide by 2. The remaining candidate points are the outer boundary point set.

[0043] Using the geometric center of the cluster as the reference point, calculate the angle between the line connecting each boundary point to the reference point and the positive direction of the horizontal coordinate axis. The calculation first calculates the longitudinal and lateral differences between the boundary points and the reference point. The longitudinal difference is the boundary point's y-coordinate minus the reference point's y-coordinate, and the lateral difference is the boundary point's x-coordinate minus the reference point's x-coordinate. If the lateral difference is not zero, divide the longitudinal difference by the lateral difference to obtain the tangent value. Determine the angle based on the tangent value; for example, a tangent of 1 results in an angle of 45 degrees. If the lateral difference is zero and the longitudinal difference is positive, the angle is 90 degrees; if the longitudinal difference is negative, the angle is 270 degrees. Sort the boundary points in ascending order of angle, from 0 degrees to 360 degrees, forming an ordered sequence of boundary points.

[0044] Based on an ordered sequence of boundary points, adjacent boundary points are connected sequentially to form an initial polygon boundary. Based on this initial polygon boundary, concave points within the polygon are identified, and boundary adjustment methods are used to eliminate these concave points, ensuring the polygon satisfies the convexity condition, thus obtaining an optimized polygon boundary. Specifically, this includes:

[0045] Based on the optimized polygon boundary, the first and last boundary points are connected to form a geometric boundary representation. Specifically, this involves connecting adjacent boundary points sequentially according to the ordered boundary point sequence, i.e., connecting the first point to the second point, the second point to the third point, and so on until the last point, forming a closed initial polygon boundary. The interior angles between each adjacent side of the initial polygon are calculated, and the angle formed by extending the next side forward and connecting it to the previous side is taken as the interior angle. For example, if side AB connects to side BC, and CB is extended to point D, then angle ABD is the interior angle. Points with interior angles less than 135 degrees are identified as concave points. For each concave point, its preceding and following points are found, and these two points are directly connected by a straight line, replacing the original concave point connecting edge. For example, if the concave point is B, the preceding point is A, and the following point is C, AC is directly connected to replace AB and BC. After adjustment, all interior angles are recalculated, and the interior angles of each adjacent edge are checked one by one to ensure that all interior angles are greater than or equal to 135 degrees, satisfying the convexity condition, thus obtaining the optimized polygon boundary. The optimized polygon boundary has connected adjacent points in sequence. At this time, the last boundary point of the sequence is connected to the first boundary point with a straight line to form a complete closed convex polygon. This convex polygon completely encloses all feature vectors of the corresponding cluster, and its boundary point coordinates and shape accurately reflect the spatial distribution range of the cluster, serving as the geometric boundary representation of the cluster.

[0046] Based on geometric boundary representation, the area of ​​overlapping regions between the boundaries of different feature vector clusters is calculated to obtain a feature distribution overlap metric. Based on the feature distribution overlap metric, spatial associations between feature vector clusters are established to generate a spatial relationship structure. Specifically, for any two clusters of convex polygons, their respective bounding rectangles are calculated. The left boundary of the bounding rectangle is the minimum x-coordinate of all vertices of the polygon, the right boundary is the maximum x-coordinate, the upper boundary is the maximum y-coordinate, the lower boundary is the minimum y-coordinate, the length is the right boundary minus the left boundary, and the width is the upper boundary minus the lower boundary. To determine if the circumscribed rectangles overlap, if the left boundary of rectangle A is greater than the right boundary of rectangle B, or the top boundary of rectangle A is less than the bottom boundary of rectangle B, then there is no overlap, and the overlap area is 0; otherwise, there is overlap. The overlap area is calculated using the scan line method. The first step is to determine the scan range as the x-coordinate range of the overlapping portion of the two circumscribed rectangles, i.e., the left boundary is the maximum value of the left boundaries of the two rectangles, and the right boundary is the minimum value of the right boundaries of the two rectangles. The second step is to set the scan line spacing to the minimum difference of the x-coordinates of all feature vectors, typically between 0.01 and 0.05, with the scan lines arranged sequentially from the left boundary to the right boundary. The third step is to generate two intersection points between each scan line and the boundaries of the two convex polygons, and record the y-coordinates of these intersection points. The first step is to calculate the area of ​​the interval by multiplying the distance between the two scan lines by the absolute value of the difference in the y-coordinates of the two intersection points. The fourth step is to add the areas of all intervals to obtain the area of ​​the overlapping region. The area of ​​the two convex polygons is calculated by using the shoelace formula, arranging the vertices of the polygons in order, and calculating the product of the x-coordinate of the adjacent vertex and the y-coordinate of the next vertex in turn, and summing them to obtain the first sum. The second sum is obtained by calculating the product of the y-coordinate of the adjacent vertex and the x-coordinate of the next vertex in turn. The absolute value of the first sum minus the second sum is divided by 2 to obtain the area of ​​the polygon. The overlapping region area is divided by the smaller area of ​​the two polygon areas to obtain the characteristic distribution overlap quantification index, which ranges from 0 to 1.

[0047] The association strength levels are classified according to the overlap metric: a metric greater than 0.3 indicates a strong association, such as a cluster of high-click products and a cluster of high-purchase products; a metric between 0.1 and 0.3 indicates a medium association, such as a cluster of users browsing a certain category of products and a cluster of users browsing related categories of products; and a metric less than 0.1 indicates a weak association, such as a cluster of unpopular products and a cluster of popular products. The geometric information of each cluster is recorded, including the x and y coordinates of the geometric center, the area of ​​the convex polygon, and the x and y coordinates of all boundary points. The geometric information of all clusters is integrated with the association strength levels between clusters to construct a spatial relationship structure that includes the spatial location, size, and degree of association of the clusters.

[0048] In this embodiment, convex polygons are used to delineate the spatial range of feature clusters, and the cluster association strength is quantified by calculating the specific overlapping area. The resulting spatial relationship structure can accurately reflect the distribution characteristics and association patterns of e-commerce multi-source data.

[0049] In a preferred embodiment of the present invention, a baseline analysis unit is selected in the spatial relationship structure. The fractal dimension features of the feature vectors are calculated, and the analysis unit is recursively partitioned at multiple levels based on the fractal dimension values. Each level of partitioning maintains self-similarity, resulting in feature partitioning. This partitioning may include:

[0050] Based on the spatial relationship structure, high-density regions in the feature vector distribution are identified to obtain the benchmark analysis unit. Based on the benchmark analysis unit, the fractal dimension value of the feature vector distribution is calculated to obtain the fractal dimension feature. Specifically, this includes: first, calculating the Euclidean distance between all pairs of feature vectors within the unit, and selecting the minimum value from all distances as the minimum spacing, with the minimum spacing ranging from 0.01 to 0.1; using twice the minimum spacing as the starting side length of the measurement box, with the starting side length ranging from 0.02 to 0.2; then calculating the side length of the circumscribed square of the benchmark analysis unit, and taking the side length of all feature vectors within the unit... The larger of the differences between the maximum and minimum x-coordinate values ​​and the maximum and minimum y-coordinate values ​​is the unit side length. The unit side length ranges from 0.5 to 5. Half of the unit side length is used as the maximum side length of the measurement box, which ranges from 0.25 to 2.5. Five different box side lengths are set with equal step sizes. The step size is the maximum side length minus the initial side length, divided by 4. The initial side length is then added to the step size to obtain the five side lengths. For example, if the initial side length is 0.02 and the maximum side length is 0.25, the step size is 0.0575, and the five side lengths are 0.02, 0.0775 ... 135, 0.1925, 0.25, with an overall value range of 0.02 to 2.5; Starting from the top left corner of the unit, the square measuring boxes of each side length are moved horizontally according to the side length of the boxes to cover the entire unit. After covering the horizontal range, they are moved vertically according to the side length, layer by layer, covering the entire unit without omission or repetition; The number of boxes containing at least one feature vector under each side length is counted, with a larger number for smaller side lengths, ranging from 10 to 500. For example, the number is 300 to 500 when the side length is 0.02, and 10 to 30 when the side length is 2.5; The natural logarithm of the side length of each box is calculated. And the natural logarithm of the number of boxes containing vectors, for example, the logarithm of a side length of 0.02 is approximately -3.912, and the logarithm of a quantity of 300 is approximately 5.704; the logarithm of a side length of 0.0775 is approximately -2.554, and the logarithm of a quantity of 120 is approximately 4.787; in ascending order of side length, calculate the logarithmic difference between two adjacent side lengths (the latter minus the former) and the logarithmic difference of the corresponding quantity (the latter minus the former), for example, the logarithmic difference of side lengths -2.554 minus -3.912 equals 1.358, and the logarithmic difference of quantity 4.787 minus 5.704 equals -0.917.

[0051] Divide the logarithmic difference of each quantity by the logarithmic difference of the corresponding side length to obtain five slope values, ranging from 0.8 to 3.2. For example, the above calculation result is -0.917 divided by 1.358, approximately 0.675. If it exceeds the range, check the side length setting and statistical results. Add the five slope values ​​together and divide by 5 to obtain the average value. This average value is the fractal dimension value, ranging from 1 to 3.

[0052] Fractal dimension features are extracted based on fractal dimension values: When the fractal dimension value is close to 1, the feature vector exhibits a linear or banded distribution within the cell, reflecting the feature aggregation dominated by a single attribute in e-commerce scenarios, such as product features divided only by price range; when the value is between 1.5 and 2.5, the feature vector exhibits a planar distribution of moderate complexity, reflecting the feature aggregation of multiple attributes collaboratively associated, such as product features combining price, sales volume, and positive review rate; when the value is close to 3, the feature vector exhibits a dense and complex three-dimensional distribution, reflecting the feature aggregation of multiple dimensions cross-associated, such as comprehensive features integrating user behavior, product attributes, and supply chain status. This quantitative value combined with the distribution pattern forms a complete fractal dimension feature.

[0053] Based on fractal dimension features, the method determines the number of recursive partitioning levels to obtain partitioning parameters. Specifically, it uses fractal dimension values ​​of 1.5 and 2.5 as partitioning thresholds. When the fractal dimension value is less than 1.5, the number of recursive partitioning levels is set to two; when the fractal dimension value is between 1.5 and 2.5, the number of partitioning levels is set to three; and when the fractal dimension value is greater than 2.5, the number of partitioning levels is set to four. This number of partitioning levels is the partitioning parameter, and the partitioning level ranges from 2 to 4. In e-commerce scenarios, the fractal dimension values ​​of feature clusters such as high-frequency user interaction behaviors and core product attributes are mostly between 1.5 and 2.5, corresponding to a partitioning level of three.

[0054] Based on the partitioning parameters, the granularity of each partition level is calculated to obtain the granularity parameters for each level. Based on the granularity parameters for each level, the benchmark analysis unit is partitioned into first-level feature partitions, specifically including: the partitioning granularity refers to the total number of sub-units obtained after each level of partitioning, with a value ranging from 2 to 6; the calculation method for the first-level partitioning granularity is to multiply the fractal dimension value by 2 and then round down to the nearest integer, ensuring that the result falls between 2 and 6. For example, a fractal dimension value of 1.2 multiplied by 2 equals 2.4, rounded down to the nearest integer 2; a fractal dimension value of 1.8 multiplied by 2 equals 3.6, rounded down to the nearest integer 4; a fractal dimension value of 2.8 multiplied by 2 equals 5.6, rounded down to the nearest integer 6; the horizontal and vertical directions are determined according to the granularity during partitioning. The number of vertical divisions is determined such that the product of the number of horizontal divisions and the number of vertical divisions equals the granularity, and the two are as close as possible. For example, when the granularity is 2, it is 1 horizontal division and 2 vertical divisions or 2 horizontal divisions and 1 vertical division; when the granularity is 3, it is 1 horizontal division and 3 vertical divisions or 3 horizontal divisions and 1 vertical division; when the granularity is 4, it is 2 horizontal divisions and 2 vertical divisions; when the granularity is 5, it is 1 horizontal division and 5 vertical divisions or 5 horizontal divisions and 1 vertical division; when the granularity is 6, it is 2 horizontal divisions and 3 vertical divisions or 3 horizontal divisions and 2 vertical divisions. According to the determined number of divisions, the baseline analysis unit is evenly divided along the horizontal and vertical directions to obtain the corresponding number of sub-regions. Each sub-region is the first-level feature partition.

[0055] Based on the first-level feature partitions, the self-similarity characteristics of each partition are verified to obtain the self-similarity verification results. Specifically, this includes: calculating the fractal dimension value of each first-level feature partition; subtracting the fractal dimension value of each first-level partition from the fractal dimension value of the corresponding benchmark analysis unit, and taking the absolute value of the difference. Using 0.2 as the self-similarity judgment threshold, if the absolute value of the difference is less than 0.2, the first-level partition is determined to satisfy the self-similarity characteristic; if the absolute value of the difference is greater than or equal to 0.2, it is determined not to satisfy the self-similarity characteristic. The absolute value of the difference ranges from 0 to 1.0. The self-similarity verification results of each first-level partition are recorded. For example, if the fractal dimension value of the benchmark analysis unit is 1.8, and the dimension value of a certain first-level partition is 1.9, the absolute value of the difference is 0.1, which is less than 0.2, thus satisfying self-similarity; if the dimension value of a certain first-level partition is 1.5, the absolute value of the difference is 0.3, which is greater than 0.2, thus not satisfying self-similarity.

[0056] Based on the self-similarity verification results, subsequent partitioning is performed on each level of partition. The granularity of each partitioning level is dynamically adjusted based on the fractal dimension value to ensure that the self-similarity characteristic is maintained, and the preliminary feature partitions are finally obtained. Specifically, for the first-level partition that satisfies the self-similarity characteristic, the granularity of the next level of partitioning is dynamically adjusted according to its fractal dimension value. If the fractal dimension value of the partition is higher than the fractal dimension value of the benchmark analysis unit, the granularity of the next level of partitioning is increased by 1 based on the granularity of the current level. If the fractal dimension value of the partition is lower than the fractal dimension value of the benchmark analysis unit, the granularity of the next level of partitioning is decreased by 1 based on the granularity of the current level. The adjusted granularity should be kept between 2 and 6. If the adjusted granularity is less than 2, it is fixed at 2. If it is greater than 6, it is fixed at 6. According to the adjusted granularity, the first-level partition is partitioned into a second level using the segmentation method in step 1014 to obtain the second-level feature partitions. Then repeat the self-similarity verification process in step 1015 to verify the second-level partitions. For the second-level partitions that meet the self-similarity characteristics, continue to adjust the granularity and perform the next level of partitioning in the above way until the number of partitioning levels is completed. For partitions that do not meet the self-similarity characteristics in each level of partitioning, no further partitioning will be performed, and they will be directly retained as the feature partitions of the current level. Finally, all the partitions obtained by the partitioning together constitute the preliminary feature partitions.

[0057] Based on the initial feature partitions, the partition structure is optimized to obtain feature partitions. Specifically, this includes: counting the number of feature vectors in each initial feature partition; using 1% of the total number of feature vectors as a merging threshold, with this threshold ranging from 10 to 400 (determined based on the total number of vectors from 1000 to 40000); if a preliminary partition contains fewer vectors than this threshold, it is merged with its adjacent feature partitions; during merging, the difference in fractal dimension values ​​between the merged partition and all its adjacent partitions is calculated, and the adjacent partition with the smallest difference is selected for merging, ensuring that the merged partitions still maintain similarity. The structural characteristics of the adjacent partitions are analyzed. After merging, the boundaries between adjacent partitions are checked. The angle between the two lines formed by three adjacent points on the boundary and the horizontal direction is calculated, i.e., the angle between the first and second lines. The angle change value is obtained by subtracting the angle of the first line from the angle of the second line. 90 degrees is used as the smoothing threshold. If the angle change value is greater than 90 degrees, the boundary segment is smoothed. The processing method is to directly connect the two ends of the boundary segment and delete the inflection points in the middle to make the boundary lines smoother. After merging and smoothing, the final feature partition is obtained.

[0058] This embodiment deeply mines the spatial distribution and intrinsic structural characteristics of features throughout the entire process, improving the rationality and accuracy of feature partitioning and enhancing the relevance and reliability of the analysis results.

[0059] In a preferred embodiment of the present invention, mapping the feature vector distribution to the corresponding feature partitions and calculating the partition spatial distribution weights based on the vector density distribution and spatial positional relationship of each feature partition may include:

[0060] Based on the feature vector distribution and feature partitioning, feature vectors are assigned to corresponding feature partitions to obtain partitioning results. Based on the partitioning results, the vector distribution density within each feature partition is calculated to obtain partition density distribution data. Specifically, this includes: calculating the geometric center (coordinate range 0 to 1) and boundary range (each dimension boundary range 0 to 1) of each feature partition; for each feature vector (coordinate range 0 to 1), calculating its Euclidean distance (range 0 to √√d, where d is the dimension after dimensionality reduction, 2 or 3) to the geometric centers of all feature partitions; and simultaneously determining whether the vector is within the boundary range of a certain partition. If a feature vector is within a given partition, it is directly assigned to that partition; if it is not within the boundary of any partition, it is assigned to the nearest partition, ensuring that all feature vectors are contained in a unique partition, thus forming the partition assignment result. The total number of feature vectors contained in each feature partition is counted (range 50 to 1000), and the spatial volume of the partition is measured (area in 2D space, volume in 3D space, range 0.5 to 10). The total number of feature vectors in the partition is divided by the spatial volume of the partition to obtain the vector distribution density of the partition, with a density range of 0.1 to 200. The vector distribution densities of all partitions together constitute the partition density distribution data.

[0061] Based on the partition density distribution data, the relative positional relationship of each feature partition in the feature space is calculated to obtain partition spatial relationship data; based on the partition density distribution data and partition spatial relationship data, each feature partition is evaluated to obtain the partition spatial distribution weight, specifically including:

[0062] Calculate the geometric center of the entire feature space (the average of all feature vector coordinates, ranging from 0 to 1). Then, calculate the Euclidean distance between the geometric center of each feature partition and the geometric center of the feature space, ranging from 0.1 to 5, reflecting the degree of deviation of the partition's center position. Simultaneously, calculate the Euclidean distance (ranging from 0.1 to 5) between each partition and the geometric centers of all other partitions. Count the number of adjacent high-density partitions for each partition (the number of high-density partitions with a distance less than τ, ranging from 1 to 8). The degree of deviation and the number of adjacent partitions together constitute the spatial relationship data of the partitions. Compare the partition vector distribution density with the maximum density of all partitions to obtain a density score. The density score is calculated as follows: The range is 0 to 1. The calculation method is as follows: divide the density of the partition by the maximum density of all partitions; multiply the inverse of the distance between the partition and the center of the feature space by 0.5 to obtain the position deviation score (range 0 to 0.5); then divide the number of adjacent partitions of the partition and the high-density partitions by the maximum number of adjacent partitions of all partitions to obtain the adjacent density score (range 0 to 0.5); add the two to obtain the position score (range 0 to 1); add the density score and the position score to obtain the partition comprehensive score (range 0 to 2); divide the comprehensive score of each partition by the sum of the comprehensive scores of all partitions to obtain the partition spatial distribution weight, the weight value ranges from 0 to 1, and the sum of the weights of all partitions is 1.

[0063] In this embodiment, the spatial distribution weights of the partitions combine density and location as dual dimensions to quantify the importance of the partitions, providing a precise basis for model parameter initialization and avoiding blind initialization.

[0064] like Figure 2 As shown, in another preferred embodiment of the present invention, adjusting the parameters of the deep learning network using partitioned spatial distribution weights to obtain a value evaluation result may include:

[0065] Based on the spatial distribution weights of the partitioned network, the connection parameters of the deep neural network are initialized to obtain a parameter-optimized deep neural network. Specifically, this includes constructing a deep neural network containing an input layer, hidden layers, and an output layer. The number of nodes in the input layer is consistent with the dimension of the standardized feature matrix (range 256 to 1000). Three hidden layers are set, with the number of nodes in each layer being twice (512 to 2000), 1.5 times (384 to 1500), and once (256 to 1000) the number of nodes in the input layer, respectively. The output layer has one node (corresponding to the value assessment score). The hidden layers use... The ReLU activation function is used, and the output layer uses the Sigmoid activation function. When initializing the network connection parameters, the connection parameters from the input layer to the first hidden layer are set to the corresponding partition spatial distribution weights multiplied by a random value between 0.001 and 0.01. That is, each connection parameter is equal to the partition weight corresponding to that feature multiplied by a random initial value. The random initial value ranges from 0.001 to 0.01, and the final value of the connection parameter ranges from 0.0001 to 0.01. This ensures that features related to high-weight partitions have a stronger influence in the initial stage, resulting in a deep neural network with optimized parameters.

[0066] Based on the parameter-optimized deep neural network, the standardized feature matrix is ​​processed through forward propagation to obtain a preliminary value assessment output. Specifically, the process includes: feeding the standardized feature matrix (with elements ranging from 0 to 1) as input data into the parameter-optimized deep neural network; after receiving the feature data, each input layer node multiplies the value of each node with its corresponding connection parameter, and then sums all the product results to obtain the weighted sum of that node, with the weighted sum ranging from -10 to 10; feeding the weighted sum into the first hidden layer, where it is calculated using the ReLU activation function, i.e., outputting the weighted sum when it is greater than 0 and outputting 0 when it is less than or equal to 0, thus obtaining the output result of the first hidden layer (with values ​​ranging from 0 to 10); feeding this result into subsequent hidden layers in the same way, performing weighted summation and activation function calculations sequentially, and finally feeding it into the output layer, where it is calculated using the Sigmoid activation function, i.e., outputting 1 divided by 1 plus the negative weighted sum of the natural exponent, thus obtaining the preliminary value assessment output corresponding to each feature vector, with the assessment output ranging from 0 to 1.

[0067] Based on the preliminary value assessment output and the spatial distribution weights of the partition, the parameters of the deep neural network are iteratively optimized to obtain the optimized network parameter configuration. The value assessment result is then calculated, specifically including: using the mean square error between the preliminary value assessment output and the actual value label (the value level of the product or user determined based on historical transaction data and user feedback, with a value range of 0 to 1) as the loss function. The loss function is calculated by summing the squares of the differences between the assessment output and the actual label of all samples and dividing by the total number of samples. The initial value range of the loss function is 0.1 to 0.5, and the target value range is 0 to 0.01. Stochastic gradient descent is used for iterative parameter optimization with a learning rate of 0.001 (ranging from 0.0005 to 0.002). Each iteration selects a batch of samples, representing 10% of the total samples (ranging from 5% to 15%). The gradient of the loss function with respect to each network connection parameter is calculated, ranging from -1 to 1, reflecting the rate of change of the loss function with respect to the parameters. The gradient is multiplied by the learning rate to obtain the parameter adjustment (ranging from -0.002 to 0.002). The current connection parameter is subtracted from the adjustment to complete one update. Simultaneously, the adjustment is corrected based on the spatial distribution weights of the partitions: the parameter adjustment for high-weight partitions (weights greater than 0.1) is multiplied by 1.2, and the parameter adjustment for low-weight partitions (weights less than 0.05) is multiplied by 0.8. This iterative process is repeated 50 to 200 times until the loss function value changes by less than 0.0001 for 10 consecutive iterations. The iteration stops at this point, and the network output evaluation result, ranging from 0 to 1, is the final value evaluation result.

[0068] In this embodiment, the deep neural network enhances the influence of key features through weight adaptation initialization and iterative optimization, thereby improving the accuracy and stability of value assessment and making the assessment results more in line with the actual needs of e-commerce scenarios.

[0069] In a preferred embodiment of the present invention, the value assessment results are processed through a multimodal fusion algorithm, a personalized recommendation strategy is generated based on a reinforcement learning framework, and resource and intelligent routing algorithms are calculated to complete the digital insight analysis process, which may include:

[0070] Based on the value assessment results, a multimodal fusion algorithm is used to integrate textual features, behavioral features, and supply chain features to obtain multimodal feature integration results. Specifically, this includes extracting textual semantic feature vectors (256 dimensions, value range 0 to 1), user behavior temporal feature vectors (256 dimensions, value range 0 to 1), and supply chain dynamic feature vectors (256 dimensions, value range 0 to 1, including quantitative features corresponding to inventory, logistics, and production progress). Based on the previously obtained value assessment results (value range 0 to 1), the fusion weights of each modality feature are determined. The weight of textual semantic features is the value assessment result multiplied by 0.3, with a weight range of 0 to 0.3; the weight of behavioral temporal features is the value assessment result multiplied by 0.4, with a weight range of 0 to 0.4; and the weight of supply chain dynamic features is the value assessment result multiplied by 0.3, with a weight range of 0 to 0.3. Multiply the value of each dimension of each modal feature vector by the corresponding fusion weight, and then add the values ​​of the corresponding dimensions of the three modal feature vectors to obtain the multimodal feature integration result. The integrated feature vector has 256 dimensions, and the value of each dimension ranges from 0 to 1.

[0071] Based on the results of multimodal feature integration, an e-commerce environment state analysis is performed using a reinforcement learning framework to obtain e-commerce environment state features. Based on these features, a policy generation algorithm is used to construct a personalized recommendation strategy, resulting in a preliminary recommendation strategy. Specifically, this includes: constructing a reinforcement learning framework, using the multimodal feature integration results as the environment state input. The environment state includes three dimensions: user preference features, product value features, and supply chain security features, with each dimension ranging from 0 to 1; the agent is the recommendation decision module, with the action space consisting of recommended product combinations within the e-commerce platform. Each action corresponds to a recommended order of products, with the number of product combinations ranging from 10 to 50; the reward function is set as the user click-through rate multiplied by the reward function after the recommendation strategy is implemented. The sum of 0.4, conversion rate multiplied by 0.5, and repurchase rate multiplied by 0.1 is used, with click-through rate, conversion rate, and repurchase rate ranging from 0 to 1, and reward value ranging from 0 to 1. The policy network of the reinforcement learning model is initialized, with a network structure consistent with the deep neural network, and parameters initialized to random values ​​between 0.001 and 0.01. Experience data is generated through interaction with the environment, with 100 to 1000 experience samples generated per interaction. The samples include state, action, reward, and new state, and the samples are stored in the experience replay pool. The policy network is trained based on the experience samples, with 100 to 500 iterations to optimize the action selection probability, so that the agent selects the action that can obtain the maximum cumulative reward in the current state, thus obtaining the preliminary recommendation policy.

[0072] Based on the initial recommendation strategy, iterative optimization is performed to obtain an optimized personalized recommendation strategy. Based on the optimized personalized recommendation strategy, a resource allocation algorithm is used to calculate the required computational resource allocation scheme, obtaining the resource allocation result. Specifically, this includes: applying the initial recommendation strategy to a subset of users, representing 10% of the total users (ranging from 5% to 20%), collecting actual user behavior data, including click-through rate, conversion rate, and repurchase rate (all ranging from 0 to 1), and calculating the difference between the actual reward value and the model's predicted reward value (ranging from -1 to 1). The product combination and recommendation order of the recommendation strategy are adjusted based on this difference. For example, if the actual click-through rate of a product is lower than the predicted value (negative difference), the weight of that product in the recommendation list is reduced by 0.1 times the absolute value of the difference (ranging from 0 to 0.1). Simultaneously, the strategy is adjusted in conjunction with dynamic supply chain data. If the inventory gap of a product accounts for more than 30% of the total inventory, its recommendation frequency is reduced by a percentage equal to the inventory gap percentage (ranging from 0 to 1). Repeat the above process for 3 to 5 iterations. After each round of optimization, apply the strategy to a larger user group, collect data and adjust it until the actual reward value increases by less than 0.01 for three consecutive rounds, thus obtaining the optimized personalized recommendation strategy.

[0073] The computational requirements of the optimized personalized recommendation strategies are analyzed. The CPU resources required for each recommendation strategy are calculated as the number of products included in the strategy multiplied by the computational resources per product, with each product occupying 0.1 CPU cores (ranging from 0.05 to 0.15 CPU cores). The total CPU resource requirement ranges from 0.5 to 10 CPU cores. Memory resources are calculated as the amount of feature data corresponding to the strategy multiplied by the storage coefficient, with each 1000-dimensional feature occupying 1MB of memory (ranging from 0.5 to 1.5MB). The total memory resource requirement ranges from 100 to 1000MB. Based on the value assessment results of the recommendation strategies, resource allocation priorities are determined: strategies with high value assessment results (greater than 0.8) have a priority of 1, medium value (0.3 to 0.8) has a priority of 2, and low value (less than 0.3) has a priority of 3. Computational resources are allocated according to priority order: strategies with priority 1 receive 50% of the total CPU resources and 50% of the total memory resources, strategies with priority 2 receive 30%, and strategies with priority 3 receive 20%, ensuring that high-value recommendation strategies receive sufficient computational resources.

[0074] Based on the resource allocation results, the data transmission path is determined through intelligent routing algorithms to obtain a routing configuration scheme. Based on the personalized recommendation strategy, resource allocation results, and routing configuration scheme, the digital insight analysis process is completed, which specifically includes: collecting real-time status data of each server node of the e-commerce platform, with transmission latency ranging from 1 to 100 milliseconds, bandwidth utilization ranging from 30% to 90%, and packet loss rate ranging from 0.1% to 5%. The evaluation metric for data transmission paths is set as the sum of transmission delay multiplied by 0.5, bandwidth utilization multiplied by 0.3, and packet loss rate multiplied by 0.2, with the metric value ranging from 0.1 to 0.9. The smaller the value, the better the path. For each personalized recommendation strategy's corresponding data stream, the evaluation metric for all possible transmission paths from the data processing node to the user access node is calculated, with the number of paths ranging from 3 to 10. The path with the smallest metric value is selected as the optimal transmission path. If multiple paths have the same metric value with a difference of less than 0.05, the path passing through the high-priority resource allocation node is selected first. Finally, the transmission path for each data stream is determined, forming a routing configuration scheme. Combining the personalized recommendation strategy, resource allocation results, and routing configuration scheme, the e-commerce digital insight analysis process is completed.

[0075] This embodiment integrates various core features through multimodal fusion, and the reinforcement learning framework dynamically adapts to changes in the e-commerce environment. The generated personalized recommendation strategy is more in line with user preferences and dynamic scenarios. With reasonable resource allocation and intelligent routing scheme, the recommendation strategy is ensured to be implemented efficiently. The whole process captures the deep dynamic correlation of multi-source heterogeneous data, improving the accuracy and practicality of digital insights.

[0076] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A multi-source data fusion and analysis system for e-commerce digital insights, characterized in that, include: The data acquisition module is used to collect interaction behavior data across the entire chain and integrate unstructured text data and supply chain dynamic data to construct a multi-source heterogeneous raw dataset. The processing module is used to preprocess and feature process multi-source heterogeneous raw datasets, including semantic parsing based on deep natural language processing and feature extraction based on time-series analysis, to generate a standardized feature matrix. The mapping module is used to map the standardized feature matrix to a high-dimensional feature space to form a feature vector distribution. By constructing a convex polygon representation of each feature vector, it calculates the overlapping area between different feature vector distribution regions, quantifies the degree of overlap of feature distributions, and generates a spatial relationship structure. The segmentation module is used to select the benchmark analysis unit in the spatial relationship structure. By calculating the fractal dimension features of the feature vector, the analysis unit is recursively divided into multiple levels based on the fractal dimension value. Each level of division maintains self-similarity characteristics, resulting in feature partitions. The calculation module is used to map the feature vector distribution to the corresponding feature partitions and calculate the spatial distribution weight of each feature partition based on the vector density distribution and spatial position relationship of each feature partition. The adjustment module is used to adjust the parameters of the deep learning network using the weights distributed in the partition space to obtain the value evaluation results. The output module processes the value assessment results through a multimodal fusion algorithm, generates personalized recommendation strategies based on a reinforcement learning framework, and computes resource and intelligent routing algorithms to complete the digital insight analysis process.

2. The multi-source data fusion and analysis system for e-commerce digital insights according to claim 1, characterized in that, Collect end-to-end interaction behavior data and integrate unstructured text data and supply chain dynamic data to construct a multi-source heterogeneous raw dataset, including: Collect end-to-end interactive behavior data to obtain a raw behavioral data stream containing page browsing trajectory, product click sequence, and transaction operation log; Based on the raw behavioral data stream, data aggregation processing is performed using device identification and session reconstruction algorithms to obtain complete interactive session sequence data; based on the interactive session sequence data, dynamic monitoring data of each link in the supply chain is collected to obtain a supply chain data set; Based on interactive conversation sequence data and supply chain data sets, unstructured text data is collected to obtain a text corpus. Data association processing is performed on the interactive conversation sequence data, supply chain data sets, and text corpus to obtain multi-source associated data, which is then standardized to construct a multi-source heterogeneous original dataset.

3. The multi-source data fusion and analysis system for e-commerce digital insights according to claim 2, characterized in that, Preprocessing and feature processing are performed on the multi-source heterogeneous original dataset, including semantic parsing based on deep natural language processing and feature extraction based on temporal analysis, to generate a standardized feature matrix, including: The multi-source heterogeneous original dataset is preprocessed to obtain a purified standard dataset; based on the purified standard dataset, text corpus data is extracted, text features are extracted, and semantic feature vectors are obtained. Based on the purified standard dataset, interactive session sequence data is extracted, and time series feature mining is performed to obtain time series feature vectors. Based on semantic feature vectors and time series feature vectors, a feature fusion algorithm is used to integrate multi-source features to obtain a multi-source integrated feature set. Based on the multi-source integrated feature set, normalized feature data is obtained through feature normalization; based on the normalized feature data, dimensional unification and format standardization are performed to generate a standardized feature matrix.

4. The multi-source data fusion and analysis system for e-commerce digital insights according to claim 3, characterized in that, The standardized feature matrix is ​​mapped to a high-dimensional feature space to form a feature vector distribution. By constructing a convex polygon representation of each feature vector, the overlapping area between different feature vector distribution regions is calculated to quantify the degree of overlap of the feature distributions and generate a spatial relationship structure, including: Based on the standardized feature matrix, the high-dimensional features are mapped to the low-dimensional manifold space to obtain the distribution of feature vectors in the low-dimensional space. Based on the distribution of feature vectors in low-dimensional space, a boundary detection algorithm is used to identify the outer boundary points of the feature vector cluster, and a geometric boundary representation of the feature vector cluster is constructed based on the boundary points. Based on geometric boundary representation, the area of ​​overlapping regions between the boundaries of different feature vector clusters is calculated to obtain a feature distribution overlap metric; based on the feature distribution overlap metric, spatial correlation between feature vector clusters is established to generate a spatial relationship structure.

5. The multi-source data fusion and analysis system for e-commerce digital insights according to claim 4, characterized in that, Based on the distribution of feature vectors in low-dimensional space, a boundary detection algorithm is used to identify the outer boundary points of the feature vector cluster. A geometric boundary representation of the feature vector cluster is constructed based on these boundary points, including: Based on the distribution of feature vectors in low-dimensional space, the outer boundary points of each feature vector cluster are identified, resulting in a set of boundary points for the feature vector clusters. Based on the set of boundary points for the feature vector clusters, all boundary points are sorted clockwise to obtain an ordered sequence of boundary points. Based on the ordered boundary point sequence, adjacent boundary points are connected sequentially to form the initial polygon boundary; based on the initial polygon boundary, concave points in the polygon are identified, and boundary adjustment methods are used to eliminate the concavities to ensure that the polygon satisfies the convexity condition, thus obtaining the optimized polygon boundary. Based on the optimized polygon boundary, the first and last boundary points are connected to form a geometric boundary representation.

6. The multi-source data fusion and analysis system for e-commerce digital insights according to claim 5, characterized in that, A baseline analysis unit is selected within the spatial relationship structure. The fractal dimension features of the feature vectors are calculated, and the analysis unit is recursively partitioned at multiple levels based on these fractal dimension values. Each level of partitioning maintains self-similarity, resulting in feature partitions, including: Based on the spatial relationship structure, high-density regions in the feature vector distribution are identified to obtain the benchmark analysis unit; based on the benchmark analysis unit, the fractal dimension value of the feature vector distribution is calculated to obtain the fractal dimension feature. Based on the fractal dimension features, the method determines the number of levels of recursive partitioning to obtain the partitioning parameters; Based on the partitioning parameters, the benchmark analysis unit is divided into multiple levels. The granularity of each level is adjusted based on the fractal dimension value to ensure self-similarity and obtain preliminary feature partitioning. Based on the initial feature partitioning, the partitioning structure is optimized to obtain the feature partitions.

7. The multi-source data fusion and analysis system for e-commerce digital insights according to claim 6, characterized in that, Based on the partitioning parameters, the baseline analysis unit is divided into multiple levels. Each level of partitioning adjusts the granularity based on the fractal dimension value to ensure self-similarity, resulting in preliminary feature partitions, including: Based on the partitioning parameters, the granularity of each partitioning level is calculated to obtain the partitioning granularity parameters at each level; based on the partitioning granularity parameters at each level, the benchmark analysis unit is partitioned at the first level to obtain the first-level feature partition; Based on the first-level feature partitioning, the self-similarity characteristics of each partition are verified, and the self-similarity verification results are obtained. Based on the self-similarity verification results, subsequent partitioning is carried out at each level. The granularity of each partitioning is dynamically adjusted based on the fractal dimension value to ensure that the self-similarity property is maintained, and finally the preliminary feature partitioning is obtained.

8. The multi-source data fusion and analysis system for e-commerce digital insights according to claim 7, characterized in that, The feature vector distribution is mapped to the corresponding feature partitions. Based on the vector density distribution and spatial positional relationship of each feature partition, the spatial distribution weights of the partitions are calculated, including: Based on the feature vector distribution and feature partitioning, the feature vectors are assigned to the corresponding feature partitions to obtain the partitioning assignment results; based on the partitioning assignment results, the vector distribution density within each feature partition is calculated to obtain the partition density distribution data. Based on the partition density distribution data, the relative positional relationship of each feature partition in the feature space is calculated to obtain the partition spatial relationship data; based on the partition density distribution data and the partition spatial relationship data, each feature partition is evaluated to obtain the partition spatial distribution weight.

9. The multi-source data fusion and analysis system for e-commerce digital insights according to claim 8, characterized in that, By utilizing the spatially distributed weights of the partitioned region to adjust the parameters of the deep learning network, the value evaluation results are obtained, including: Based on the partitioned spatial distribution weights, the connection parameters of the deep neural network are initialized to obtain a parameter-optimized deep neural network. Based on the parameter-optimized deep neural network, the standardized feature matrix is ​​processed by the forward propagation calculation method to obtain the preliminary value assessment output; Based on the preliminary value assessment output and the weights distributed in the partition space, the parameters of the deep neural network are iteratively optimized to obtain the optimized network parameter configuration, and the value assessment result is calculated.

10. The multi-source data fusion and analysis system for e-commerce digital insights according to claim 9, characterized in that, The value assessment results are processed using a multimodal fusion algorithm, personalized recommendation strategies are generated based on a reinforcement learning framework, and resource and intelligent routing algorithms are computed to complete the digital insight analysis process, including: Based on the value assessment results, a multimodal fusion algorithm is used to integrate text features, behavioral features, and supply chain features to obtain the multimodal feature integration result; Based on the results of multimodal feature integration, an e-commerce environment state analysis is performed using a reinforcement learning framework to obtain e-commerce environment state features; based on the e-commerce environment state features, a policy generation algorithm is used to construct a personalized recommendation strategy to obtain a preliminary recommendation strategy. Based on the initial recommendation strategy, iterative optimization is performed to obtain the optimized personalized recommendation strategy; based on the optimized personalized recommendation strategy, a resource allocation algorithm is used to calculate the required computing resource allocation scheme and obtain the resource allocation result; Based on the resource allocation results, a data transmission path is determined through an intelligent routing algorithm to obtain a routing configuration scheme; based on the personalized recommendation strategy, resource allocation results, and routing configuration scheme, a digital insight analysis process is completed.

Images