A multi-dimensional data-driven self-brand growth value prediction method

Abstract

The present application relates to the technical field of data processing and business intelligence, and discloses a multi-dimensional data driven self-brand growth value prediction method, which extracts a main feature vector of a previous time step, calculates a dynamic weight for a same-product white label control group to generate a dynamic benchmark, and accordingly peels off a pure brand-driven incremental index, and combines a repeat purchase decay rate and a search index to construct a three-dimensional state vector; through eigenvalue decomposition on a state covariance matrix, a current main feature vector and a maximum eigenvalue are obtained, and the main feature vector is fed back to a weight calculation link to form a closed loop; finally, the decomposition result and a historical state sequence are respectively input into a pre-trained machine learning model to output a brand life cycle stage and a future brand asset valuation. The present application accurately anchors a reference system through a dynamic feedback mechanism, and realizes comprehensive and accurate diagnosis and prediction of brand value.

Description

Technical Field

[0001] This invention relates to the fields of data processing and business intelligence technology, specifically a multi-dimensional data-driven method for predicting the growth value of proprietary brands. Background Technology

[0002] With the development of the retail industry, private label brands have become an important asset for businesses. Accurately assessing the growth value of private label brands is crucial for companies to formulate market strategies.

[0003] Existing brand valuation methods typically measure brand premium by comparing the brand with similar products in the market. However, these methods often use static or fixed reference points, making it difficult to effectively distinguish between the brand's own growth and the impact of fluctuations in the overall market environment. When the industry as a whole is in an upward or downward cycle, this fixed comparison method can easily misinterpret macroeconomic noise as changes in the brand's own value, leading to biased evaluation results and affecting the accuracy of the conclusions.

[0004] Furthermore, traditional evaluation systems have limitations in their selection of metrics. Most evaluation models rely excessively on immediate financial indicators such as price and sales volume, while neglecting deeper factors that reflect long-term brand health. For example, key information such as brand loyalty reflected in repeat purchase behavior and consumer mindshare represented by proactive search behavior is often simplified or omitted in existing models. This one-sided reliance on metric dimensions makes the evaluation results unable to fully reflect the true state of brand equity, potentially obscuring potential risks or missing development opportunities.

[0005] Meanwhile, most existing technologies remain at the level of descriptive analysis of historical data, lacking the ability to dynamically diagnose the brand's stage in its life cycle and predict future value growth. Evaluation results are often retrospective, failing to provide forward-looking guidance to decision-makers. Consequently, brand managers are unable to adjust operational strategies in a timely manner to respond to market changes, limiting the initiative and scientific nature of brand management. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a multi-dimensional data-driven method for predicting the growth value of proprietary brands. This method solves the problems of existing technologies, such as the static evaluation benchmark making the results susceptible to market noise interference, the single evaluation dimension making it difficult to fully reflect brand health, and the lagging evaluation conclusions lacking forward-looking decision support.

[0007] To achieve the above objectives, this invention provides a multi-dimensional data-driven method for predicting the growth value of proprietary brands, comprising the following steps:

[0008] At a set discrete time step, acquire the business data stream of the target own brand and the data stream of the control group sample set of white-label products in the same category;

[0009] Extract the principal feature vector of the previous time step, construct a three-dimensional fluctuation vector for each reference product in the sample set of the white-label control group of the same category, calculate the absolute value projection of the three-dimensional fluctuation vector of each reference product on the principal feature vector of the previous time step, calculate the dynamic weight for the sample set of the white-label control group of the same category based on the absolute value projection, and generate the benchmark price and benchmark sales data of the white-label of the same category in the current time step.

[0010] Based on the business data flow of the target proprietary brand, the benchmark price of the same category of white-label products, and the benchmark sales data, pure brand-driven incremental indicators are extracted.

[0011] By combining the pure brand-driven incremental indicators, the repurchase time interval decay rate, and the normalized pure brand keyword active search index, a three-dimensional state vector representing the state of the current time step node is constructed.

[0012] Within the set sliding time window, a state covariance matrix is ​​constructed based on the three-dimensional state vector, and eigenvalue decomposition is performed on the state covariance matrix to generate the principal eigenvector and the maximum eigenvalue of the current time step. At the same time, the principal eigenvector output at the current time step is used as the principal eigenvector of the previous time step when calculating the next time step.

[0013] Based on the maximum eigenvalue and the principal eigenvector, a classification feature vector is constructed and input into a pre-trained machine learning classification model to output the diagnostic results of the brand life cycle transition stage. At the same time, the historical three-dimensional state vector sequence is input into a pre-trained time series prediction model to output the brand asset quantitative valuation at the predicted time node.

[0014] Preferably, the step of constructing a three-dimensional state vector representing the state of the current time step node specifically includes:

[0015] Obtain the pure brand-driven incremental metrics;

[0016] The repurchase time interval decay rate is calculated based on the set of repurchase order timestamps contained in the business data stream of the target proprietary brand.

[0017] Based on the brand keyword search frequency index contained in the business data stream of the target proprietary brand, the normalized pure brand keyword active search index is obtained after normalization processing by sliding window maximum and minimum values.

[0018] The three-dimensional state vector is constructed by concatenating the pure brand-driven incremental indicators, the repurchase time interval decay rate, and the normalized pure brand keyword active search index into a matrix.

[0019] Preferably, the steps of generating the principal feature vector and the largest eigenvalue at the current time step specifically include:

[0020] Multiple consecutive three-dimensional state vectors are extracted within the sliding time window to construct the state covariance matrix;

[0021] Perform eigenvalue decomposition on the state covariance matrix to obtain the maximum eigenvalue and the principal eigenvector.

[0022] Preferably, the step of extracting pure brand-driven incremental indicators specifically includes:

[0023] Determine whether the price parameter value of the target private label is higher than the benchmark price value of the same category of white-label products and whether the benchmark sales data is non-zero;

[0024] When the determination result is yes, the pure brand-driven incremental indicator is extracted from the total transaction data; and when the determination result is no, the pure brand-driven incremental indicator is assigned a value of zero.

[0025] Preferably, the step of calculating dynamic weights for the control group sample set of the same product category specifically includes:

[0026] A negative natural exponent function is introduced as the kernel function, and the absolute value projection is used as the input of the negative natural exponent function to generate the dynamic weights.

[0027] Preferably, the steps for constructing the three-dimensional wave vector specifically include:

[0028] The first-order time difference data of the reference product at the current discrete time step is extracted to obtain the price parameter difference value. At the same time, the missing repurchase and search-related parameters are padded with constant zeros, and then combined to form the three-dimensional fluctuation vector.

[0029] Preferably, the specific steps for outputting diagnostic results of the brand lifecycle transition stage include:

[0030] The classification feature vector is input into the pre-trained machine learning classification model, which serves as a support vector machine classifier, to output diagnostic results for the brand lifecycle transition stage.

[0031] Preferably, the specific steps for outputting the quantitative valuation of brand assets at the predicted time point include:

[0032] The historical three-dimensional state vector sequence is input into the pre-trained time series prediction model, which is a long short-term memory network model, to output the brand asset quantitative valuation at the predicted time node.

[0033] The historical three-dimensional state vector sequence is obtained by arranging multiple consecutive three-dimensional state vectors in chronological order.

[0034] Preferably, the business data stream includes the target proprietary brand's price parameters, sales parameters, repeat purchase order timestamp set, and brand keyword search frequency index.

[0035] Preferably, the method further includes initialization, the initialization step comprising:

[0036] Assign equal initial weights to each reference product in the same category of white-label control group sample set, and initialize the principal feature vector of the previous time step as a unit direction vector.

[0037] This invention provides a multi-dimensional, data-driven method for predicting the growth value of proprietary brands. It offers the following advantages:

[0038] 1. This invention constructs a dynamic white-label benchmark of the same category and uses the principal feature vector to adjust the weight of the reference product. This effectively filters out the common interference of macroeconomic factors and industry cycles, avoids the evaluation bias caused by using a fixed reference system, and ensures the objectivity and adaptability of the comparison benchmark. This allows for a more accurate extraction of the value increment driven purely by the brand's own effect, thus improving the accuracy and credibility of the prediction results.

[0039] 2. This invention constructs a three-dimensional state vector that includes pure brand-driven incremental growth, repurchase decay rate, and active search index. It organically combines immediate financial monetization capability, long-term user loyalty, and potential consumer mindshare, overcoming the one-sidedness of traditional methods that rely solely on a single financial indicator to evaluate brand value. This makes the portrayal of brand growth status more comprehensive and three-dimensional, and can capture subtle changes in brand health.

[0040] 3. This invention combines machine learning classification models with time series prediction models, which can not only diagnose the current life cycle stage of a brand, but also quantitatively predict its future asset value. It provides brand managers with decision-making basis that combines current state awareness with future trend prediction, which helps to realize the strategic shift from passive response to proactive planning and improves the scientific and forward-looking nature of brand management. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of the system module architecture of the present invention;

[0042] Figure 2 This is a schematic diagram of the method steps of the present invention;

[0043] Figure 3 This is a comparison chart of the predicted results and actual values ​​of the present invention;

[0044] Figure 4 This is a dynamic diagnostic probability diagram for brand lifecycle stages according to the present invention;

[0045] Figure 5 This is a comparison chart of the errors of different prediction methods of the present invention. Detailed Implementation

[0046] The technical solutions in 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.

[0047] See attached document Figure 1 This invention provides a multi-dimensional data-driven system for predicting the growth value of proprietary brands. This system runs on a computing device equipped with a processor, storage medium, and a relational database, and may include:

[0048] The data acquisition initialization module is used to extract the business data stream of the target proprietary brand and the data stream of the control group sample set of the same category of white-label products from the relational database at a set discrete time step, and to perform the assignment operation of the initial state of the system.

[0049] The dynamic benchmark anchoring module is used to calculate dynamic weights based on the absolute value projection of the fluctuation vectors of each reference product onto the principal feature vector, and to generate benchmark price and benchmark sales data of the same category of white-label products at the current time step using the dynamic weights.

[0050] The incremental extraction module is used to combine the price and sales parameters of the target private label brand, and extract pure brand-driven incremental indicators from the total transaction data when the private label price is higher than the benchmark price of white-label products in the same category and the benchmark sales are not zero.

[0051] The state space modeling module is used to combine pure brand-driven incremental indicators, the repurchase time interval decay rate of the target private label brand, and the normalized pure brand keyword active search index to construct a three-dimensional state vector representing the state of the current time step node.

[0052] The eigenvalue decomposition feedback module is used to construct the state covariance matrix within a set sliding time window, perform eigenvalue decomposition on the state covariance matrix to obtain the maximum eigenvalue and the principal eigenvector, and write the principal eigenvector in reverse to the system cache as the benchmark anchoring input parameter for the next time step calculation.

[0053] The stage prediction module is used to construct classification feature vectors based on the maximum eigenvalue and the principal eigenvector, and input them into a pre-trained machine learning classification model to output diagnostic results of the brand life cycle transition stages. At the same time, the historical three-dimensional state vector sequence is input into a pre-trained time series prediction model to output the brand asset quantitative valuation at the predicted time point.

[0054] See attached document Figure 2 This invention provides a multi-dimensional data-driven method for predicting the growth value of proprietary brands, comprising the following steps:

[0055] S10, at the set discrete time step, extract the business data stream of the target own brand and the data stream of the control group sample set of the same category of white brand from the relational database. The business data stream includes price parameters, sales parameters, repurchase order timestamp set and brand keyword search frequency index. In the initial state of the system, assign equal initial weight values ​​to each reference product in the control group sample set of the same category of white brand and initialize the main feature vector of the system state.

[0056] S20: Extract the main feature vector of the previous time step cached in the relational database, calculate the dynamic weight based on the absolute value projection of the fluctuation vector of each reference product onto the main feature vector, and use the dynamic weight to perform weighted aggregation on the data of the control group sample set of the same category of white-label products to generate the benchmark price and benchmark sales data of the same category of white-label products at the current time step.

[0057] S30: Read the price and sales parameters of the target private label brand, and perform numerical comparison calculations by combining the benchmark price and benchmark sales data of the same category of white-label brands. Under the condition that the private label brand price is higher than the benchmark price of the same category of white-label brands and the benchmark sales are not zero, extract the pure brand-driven incremental indicators from the total transaction data.

[0058] S40: Obtain pure brand-driven incremental indicators, combine the repurchase time interval decay rate of the target private label brand and the normalized pure brand keyword active search index, and construct a three-dimensional state vector representing the current time step node state.

[0059] S50: Extract multiple consecutive three-dimensional state vectors within the set sliding time window to construct a state covariance matrix. Perform eigenvalue decomposition on the state covariance matrix to obtain the maximum eigenvalue and the corresponding principal eigenvector. Write the principal eigenvector output at the current time step in reverse into the system cache as the benchmark anchoring input parameter for the next time step calculation.

[0060] S60 constructs a classification feature vector based on the maximum eigenvalue and the principal eigenvector, and inputs it into a pre-trained machine learning classification model for identification, outputting diagnostic results for the brand lifecycle transition stage; at the same time, it inputs the historical three-dimensional state vector sequence within the set sliding time window into a pre-trained time series prediction model, outputting the brand asset quantitative valuation at the predicted time node.

[0061] The specific implementation of step S10 is described in detail here. Regarding the data acquisition and system initialization process, the system sequentially performs data extraction and assignment operations according to a predefined logical order.

[0062] In this embodiment, to ensure that multi-source heterogeneous data can support subsequent matrix operations and benchmark decoupling, the system first constructs a unified spatiotemporal observation system. Specifically, the system sets the discrete time step parameters of the computing device and the initial calculation time. The system continuously and periodically performs data processing according to the preset discrete time step. For the setting of the discrete time step, those skilled in the art can configure it to a specific time span such as hours, days, or natural weeks based on the update frequency of the business data. The conversion of time units is a well-known technique in the art and will not be elaborated here. Since the underlying raw data in the business system usually has different sampling frequencies, to prevent data misalignment caused by time asynchrony, the system establishes time window alignment logic. This logic divides the time axis of the computing device into independent time intervals that are closed on the left and open on the right, so that continuous transaction events falling into the same time interval are forcibly aggregated into slices of data corresponding to the discrete time step. Based on the above alignment benchmark, the system defines the first node for starting data processing as the initial time moment and maps subsequent iterative processing processes to discrete time steps.

[0063] After unifying the observation benchmarks, the computing device periodically reads various parameters of the target proprietary brand from the relational database storing business metrics by executing structured query statements. To meet the requirement of full disclosure, the aforementioned relational database can be implemented using MySQL or PostgreSQL data storage systems. The read business data stream includes the target proprietary brand's price parameters, sales parameters, repeat purchase order timestamp set, and brand keyword search frequency metric at the current time step. Regarding the rationale for selecting specific input parameters, the linkage between price and sales parameters directly maps the product's basic supply and demand elasticity. The repeat purchase order timestamp set is used to extract the time decay momentum representing user retention trends, while the brand keyword search frequency metric independently reflects the active mindshare after deviating from platform algorithm recommendations. These three independent dimensions together constitute an orthogonal information source for evaluating implicit brand assets. The repeat purchase order timestamp set records the specific absolute time points of multiple transactions under the same user identifier, and the brand keyword search frequency metric records the number of times the user actively inputs the corresponding brand name in the system within that time step.

[0064] Considering that the parameter fluctuations of a single business entity are often mixed with common interferences from the macroeconomy and industry cycles, relying solely on its own time-series data would lead to serious evaluation biases. Therefore, based on the established physical attribute matching rules, the system filters a control group sample set of similar white-label products in a relational database and extracts its data stream. To establish an objective reference system with benchmark significance, the system searches the database for reference products without registered brand trademarks. The search criteria stipulate that the core physical attributes of the reference products must be consistent with the target's own brand. The matching rules for physical attributes specifically stipulate that the products must have the same production materials, specifications, or core functional parameters. The system includes multiple selected reference products into the control group sample set of similar white-label products and extracts the price and sales parameters of each reference product at the current time step.

[0065] As a preferred approach, to ensure the unbiasedness of subsequent feedback calculations, it is necessary to perform the allocation of initial weight values ​​for reference products and the initialization of the system's main feature vector in the initial system state. At the initial time, the system has not yet accumulated state sequence data for feedback adjustment; therefore, each reference product in the same category of white-label control group sample set is assigned an absolutely equal benchmark reference weight. The physical meaning of this operation is that, under initial conditions without prior knowledge, the system assumes, based on the maximum entropy principle, that the contribution rate of all unbranded products conforming to physical attributes to the industry baseline drift follows a uniform distribution. The initial weight assignment formula executed by the computing device is:

[0066] ;

[0067] in, Indicates the first The initial weight values ​​of each reference product at the initial time; This represents the total number of reference products included in the control group sample set of the same category of white-label products. In order to ensure the statistical significance of subsequent covariance and to prevent hardware error anomalies caused by the denominator approaching 0 in the division operation, the value range of this total number parameter is set to a positive integer greater than or equal to 10. The integer index number for the reference product.

[0068] To trigger the feedback loop of the subsequent state-space matrix, the system initializes the principal eigenvector at the initial moment with a unit direction vector, which serves as the reference parameter for calculating the projection in the next loop step. The formula for initializing the principal eigenvector is as follows:

[0069] ;

[0070] in, This represents the principal eigenvector set by the system at the initial moment; The transpose operator represents the matrix. The three-dimensional unit vector described above corresponds to the coordinate axis directions of the three state variables—brand-driven increment, repeat purchase decay, and search index—that will be constructed subsequently. Through the above parameter configuration and spatiotemporal alignment, the computing device has completed the memory loading of the basic business data stream and the state configuration operations of the computing start node, providing a basic environment for the subsequent generation of dynamic benchmark data.

[0071] This section details the process of extracting the principal eigenvectors from the previous time step and calculating the dynamic benchmark. To filter out common noise in the market and accurately anchor the reference system, the system executes the following sub-steps based on state feedback control logic.

[0072] S201, In this embodiment, constructing a same-dimensional fluctuation vector is a prerequisite for aligning spatial projection with indicators. Directly comparing the absolute values ​​of parameters is often susceptible to interference from inherent differences in the manufacturing costs of goods. To eliminate such static basis noise, the system is based on state dynamics, focusing on the relative changes in the time series. The system constructs a three-dimensional fluctuation vector for each reference product in the same category of white-label control group sample set, with a dimension strictly consistent with the state vector. In specific implementation, the system extracts the first-order time difference data of the reference product at the current discrete time step, converting the static data slices into dynamic vectors representing the trend of business changes. The formula for constructing the three-dimensional fluctuation vector is:

[0073] ;

[0074] in, Indicates the first A reference product at discrete time step The three-dimensional wave vector below; Indicates the first A reference product at discrete time step The price parameter difference between it and its previous time step; Indicates the first The repurchase parameter difference value corresponding to each reference product; Indicates the first The difference in search parameters corresponding to each reference product; This represents the matrix transpose operator. Because some white-label products lack complete repeat purchase and search tracking records in the underlying business system, the system performs constant zero-padding when extracting these missing fields. This operation is not only an algebraic means to maintain the consistency of vector dimensions required for subsequent spatial inner product operations, but also reflects the objective physical fact: white-label products themselves do not possess brand mental attributes, and the theoretical change kinetic energy of the corresponding dimensions should be zero. Therefore, forced zero-padding is entirely consistent with business logic and does not introduce distortion.

[0075] In S202, in linear algebra and spatial analytic geometry, the dot product of two high-dimensional vectors can intuitively reflect the degree of directional convergence in their spatial distribution. The computing device extracts the principal feature vector from the previous time step cached in the relational database, thereby quantifying the spatial overlap between the fluctuation trends of each reference product and the core development direction of the proprietary brand. The system resolves the algebraic relationship between the two through inner product operations. Because the underlying linear algebra algorithm library used for feature decomposition often produces feature vectors with randomly flipped signs, directly relying on the original vector inner product result can easily induce disordered oscillations in the feedback control link. To address this engineering defect, the computing device nests an absolute value envelope operation outside the dot product layer. The specific absolute value projection calculation formula is as follows:

[0076] ;

[0077] in, Indicates the first A reference product at discrete time step The absolute value of the projected value; This indicates the previous time step cached by the system in the relational database. The main eigenvectors; This represents the vector dot product operator. This represents the absolute value operator. The absolute value projection quantifies the degree to which the reference product is assimilated by the private label trend.

[0078] S203, the system performs adaptive mapping of dynamic weights based on the calculated absolute value projection values ​​to reduce the weight of affected reference products. As a preferred approach, the system introduces a negative natural exponential function as the kernel function. The technical purpose is that when the absolute value projection value of a reference product tends to increase, the system determines that its price and sales fluctuations have deviated from the objective macro-market trend and are instead constrained by the assimilation interference of its own brand's traffic spillover; therefore, it must be assigned a lower benchmark reference weight. The dynamic weight calculation formula executed by the system is configured as follows:

[0079] ;

[0080] in, Indicates the first A reference product at discrete time step Dynamic weights under; Represented by natural constant An exponential function with base 0; This represents a preset smoothing constant used to adjust the magnitude of the exponentially decaying gradient. Regarding the basis for determining the value of this smoothing constant, those skilled in the art typically define it within the interval (0, 1] based on the statistical variance of historical control group data, to prevent individual extreme values ​​from causing an excessively steep weight distribution. This represents the chain addition operator; This indicates the total number of reference products included in the control group sample set of the same product category; This is a loop iteration variable used in continuous addition operations, specifically for iterating over reference products; Indicates the first The absolute value projected value of a reference product.

[0081] To prevent system crashes caused by excessively large negative exponents leading to underflow and the denominator returning to zero during floating-point operations, the computing device incorporates minimum value truncation protection logic before executing division instructions. If the accumulated value of the denominator is detected to be less than the processor's preset minimum machine precision parameter, the system will immediately suspend the weight update operation of the current time step and maintain the weight allocation state of the previous time step, thereby ensuring the robustness of the overall computing architecture.

[0082] S204, the system uses the updated dynamic weights to perform weighted aggregation on the data of the control group sample set of the same category of white-label products, generating a high-purity benchmark reference line for subsequent indicator decoupling. The computing device performs scalar multiplication and summation on the dynamic weights of each reference product and their original parameters at the current time step, outputting the benchmark price and benchmark sales data of the same category of white-label products at the current time step. The corresponding weighted aggregation formulas are as follows:

[0083] ;

[0084] ;

[0085] in, Indicates at discrete time step The benchmark price of similar white-label products generated by the aggregation process; Indicates the first A reference product at discrete time step Price parameters; Indicates at discrete time step Benchmark sales volume of similar white-label products generated by aggregation; Indicates the first A reference product at discrete time step The system outputs benchmark parameters based on the aforementioned weight rebalancing process. These parameters effectively eliminate local disturbances parasitic on the proprietary brand, thus providing a robust and interference-resistant comparative coordinate system for quantifying pure brand premium.

[0086] This section details the decoupling and extraction process of purely brand-driven incremental indicators. In economics and commodity pricing theory, the final transaction value of a commodity is typically composed of the manufacturing cost of basic materials and the premium of intangible assets. However, in complex market environments, this premium signal is often obscured by price disturbances such as promotions and inventory clearance. The system extracts the quantitative incremental value purely generated by brand effects from the mixed business data stream through the following sub-steps.

[0087] S301, the computing device reads the price and sales parameters of the target brand at the current discrete time step via the internal data bus. In this embodiment, the price parameter directly reflects the terminal selling price at the current time slice, while the sales parameter records the actual number of units sold within the corresponding time interval. These two parameters constitute a representation of the final transaction result of the target brand under the market supply and demand game.

[0088] S302, the system extracts benchmark prices and sales volume data for white-label products in the same category, and performs numerical comparison calculations and screening for extreme boundary conditions. To ensure the effectiveness of decoupling, strict trigger conditions are set within the computing device. First, the system determines whether the price of its own brand is significantly higher than the benchmark price of white-label products in the same category. The physical basis for this determination is that price is the most direct monetization of brand value. Only when the selling price of its own brand breaks through the average price baseline of basic manufacturing and distribution in the same category does the system recognize that it has substantial brand premium capability; if the price is lower than or equal to the benchmark, it indicates that the current sales behavior is dominated by non-brand factors such as price reductions or inventory clearance.

[0089] Simultaneously, the system checks in parallel whether the baseline sales volume is greater than a preset zero-point threshold. As a preferred approach, this zero-point threshold can be configured to 1 based on the minimum transaction unit. The technical purpose of this conditional branching mechanism is to prevent the loss of baseline data due to extreme situations such as stockouts or network fluctuations at a specific time step. If the baseline sales volume is zero, it means that the comparison reference system itself has become invalid. At this time, any comparison calculation will produce distorted results that are meaningless to the business, so logical blocking is necessary.

[0090] S303, under the specific condition that the price of the proprietary brand is higher than the benchmark price of white-label products in the same category and the benchmark sales volume is not zero, the system extracts a pure brand-driven incremental indicator from the total transaction data. This indicator is obtained by multiplying the brand premium per unit product by the transaction volume, and its physical meaning is to quantify the total excess cash flow contribution of brand awareness to the enterprise at the current time step. The specific extraction formula executed by the system is as follows:

[0091] ;

[0092] in, This indicates the target private label brand in discrete time steps. The pure brand-driven incremental indicators are expressed in international standard currency units (such as USD, CNY, etc.). This indicates the target private label brand in discrete time steps. The following are the price parameters; This indicates the target private label brand in discrete time steps. The sales figures below are in the internationally recognized unit of pcs; This represents the scalar multiplication operator.

[0093] If any of the trigger conditions in S302 above is not met, the system will execute the engineering protection logic and directly assign a zero value to the pure brand-driven incremental indicator. Through the nested calculation method of algebraic difference and volume integral, the system successfully eliminates the basic commodity value components and interference noise introduced by price reduction and promotion activities, and outputs a pure brand-driven incremental indicator with high independence and accuracy, providing core data support for the subsequent construction of a high-dimensional state space model.

[0094] In modern control theory and system dynamics, state-space representation provides a standardized mathematical framework that unifies multiple observational indicators with different physical dimensions into a single vector describing the intrinsic characteristics of a system at a given moment. To achieve multi-dimensional dynamic tracking of brand assets, the system merges the instantaneous incremental indicators decoupled from previous steps with retention and search indicators representing long-term user mindset through the following sub-steps, constructing a three-dimensional state vector that fully represents the state of the current time step.

[0095] S401, to quantify the strength of user loyalty and consumption habits, the system needs to extract a decay rate indicator with continuously changing characteristics from the discrete set of repeat purchase order timestamps. In this embodiment, the computing device first traverses the set of all users who have made two or more purchases within the current time step and calculates the time interval between each pair of adjacent orders. Subsequently, the final decay rate is obtained by performing a nonlinear mapping on the average time interval of all users. As a preferred method, the system uses a negative exponential function for mapping, the physical meaning of which is that the longer the average repeat purchase interval of users, the faster the brand stickiness decays, and the smaller the corresponding indicator value. The specific calculation formula is as follows:

[0096] ;

[0097] ;

[0098] in, This indicates the target private label brand in discrete time steps. The repurchase time interval decay rate is a dimensionless scalar value within the interval (0, 1]. Represented by natural constant An exponential function with base 0; The preset time decay coefficient is used to adjust the sensitivity of the repurchase interval to the final indicator. Its value depends on the characteristics of the product's repurchase cycle. A larger value can be used for fast-moving consumer goods, while a smaller value can be used for durable goods. The average repurchase interval for all users within the current time step is expressed in seconds (s). In time step The total number of users who made at least two purchases within the specified period; For users Total number of purchases; For users No. The absolute timestamp of the purchase is usually a Unix timestamp, in seconds (s). Integer index label representing the user; This represents the loop iteration variable in a chain addition operation, specifically used to iterate through the number of purchases. To ensure robustness of the calculation, if at the current time step... There are no repeat customers (i.e.) The system will directly provide... Assign a preset maximum value or The value is assigned to zero to characterize the unmeasurable user stickiness during this period.

[0099] S402, to eliminate the dimensional impact of fluctuations in the absolute value of search frequency across different time scales and to enable comparison with other indicators within a unified mathematical space, the system needs to perform normalization on the original pure brand term active search index. In this embodiment, the system employs the sliding window minimum-maximum normalization (Min-Max Scaling) method. This method can linearly map data to the [0, 1] interval while dynamically adapting to long-term data drift. Its calculation formula is:

[0100] ;

[0101] in, This represents the normalized search index for pure brand keywords. The raw search frequency is extracted directly from the database, and the unit is counts. and In the past The maximum and minimum search frequencies within a sliding history window consisting of several time steps. (Regarding the length of the sliding window...) The value of is determined based on the natural cycle of the business. For example, for a time step measured in days, it can be . The value is set to 90 to cover business fluctuations over a quarter, thus striking a balance between smoothing out short-term noise and responding to medium-term trends. To prevent calculation errors caused by a zero denominator in the special case of constant search frequency, if the system detects... and If they are equal, then directly... The value is assigned to the median, 0.5.

[0102] In step S403, after independently quantifying the three mutually orthogonal core dimensions brought by the brand—(1) immediate financial gains, (2) user behavior retention, and (3) potential mindshare—the system performs matrix concatenation of these indicators with the pure brand-driven incremental indicators output from the previous step S30 to construct a three-dimensional state vector representing the current time step node state. The formula for constructing this vector is:

[0103] ;

[0104] in, Indicates at discrete time step The three-dimensional state vector below. This state vector constitutes a complete, multi-dimensional snapshot of brand health, providing a standardized mathematical input for subsequent analysis of covariance to reveal its inherent dynamic structure and evolutionary trends.

[0105] In multivariate statistical analysis and principal component analysis (PCA), the covariance matrix reveals the linear correlation structure among variables within multidimensional data, while its eigenvalue decomposition further extracts the most significant directions of data variation. To extract the core evolutionary dynamics of brand assets from continuous state vector time series and construct an adaptive feedback mechanism, the system performs calculations through the following sub-steps.

[0106] S501, to capture recent dynamics of brand status without being influenced by outdated historical data, the system sets a fixed-length sliding time window and constructs a state covariance matrix based on the continuous three-dimensional state vectors within the window. In this embodiment, the length of the sliding time window can be configured according to the requirements of business response sensitivity, for example, set to 30 time steps to cover approximately one month of business data fluctuations. To obtain an unbiased estimate of the sample covariance, the system uses the Bessel correction method, the specific calculation formula of which is:

[0107] ;

[0108] ;

[0109] in, Indicates at discrete time step The 3×3 state covariance matrix constructed below; The length of the sliding window; In time step The three-dimensional state vector below; It is the mean vector of all state vectors within the window; This represents the loop iteration variable in a continuous addition operation, specifically used to iterate through time steps.

[0110] The diagonal elements of this matrix represent the variances of pure brand-driven incremental growth, repurchase decay rate, and search index, respectively, while the off-diagonal elements quantify the degree of synergistic change among these three pairs. To ensure the statistical significance and non-singularity of the covariance matrix, the sliding window length must be greater than the dimension of the state vector (i.e., 3). As a preferred approach, it is generally recommended to use a length of 10 or more to obtain a more robust statistical estimate.

[0111] S502, the computing device performs eigenvalue decomposition on the generated state covariance matrix to extract its largest eigenvalue and the corresponding normalized principal eigenvector. This eigenvalue decomposition problem is algebraically represented by solving the following equation:

[0112] ;

[0113] in, The eigenvalues ​​of the matrix are, in physical terms, the variance of the state vector projected onto the direction of the corresponding eigenvector. This corresponds to an eigenvector, whose physical meaning is a specific direction in the state space. After solving the problem, the system extracts the largest eigenvalue. With the corresponding principal feature vector Main eigenvectors This points to the direction of the most dramatic change in brand asset status within the current time window, i.e., the main axis of brand development. For example, if this vector has the largest component in the pure brand-driven incremental dimension, it indicates that recent brand value growth is mainly driven by financial monetization. For this eigenvalue decomposition problem, those skilled in the art can use standard numerical computation libraries (such as LAPACK) to implement it. To prevent the risk of singular or near-singular covariance matrices due to high data collinearity, the system calculates the condition number of the matrix before performing the decomposition. If it exceeds a preset matrix condition number tolerance threshold (e.g., 10), the system will proceed accordingly. 6 If the matrix is ​​ill-conditioned, the eigenvalue decomposition result from the previous time step will be used.

[0114] In addition, to assess the dominance of the extracted principal eigenvectors, the system also calculates the contribution rate of the largest eigenvalue. :

[0115] ;

[0116] in, It is the sum of all eigenvalues; This represents the loop iteration variable in a continuous addition operation, specifically used to iterate over feature values. If this contribution rate... If the value is below a preset decision threshold (e.g., 0.5), it indicates that there is no clear dominant direction at present. In this case, the system can selectively pause feedback updates to avoid tracking a vague and unstable signal.

[0117] S503, assuming the principal feature vector is sufficiently dominant, the computing device will move to the current time step. The principal eigenvectors obtained by solving the following steps As a key feedback signal, it is stored in a designated system cache field via a database write operation. This vector will be used as the basis for step S20 in the next discrete time step. Input parameters required for performing dynamic reference anchoring Through this feedback loop design, the benchmark used by the system to identify market noise is no longer fixed, but can be dynamically adjusted in real time to track the main direction of the brand's own development, thereby achieving a more accurate and intelligent anchoring of the reference system.

[0118] Based on the brand status time series data generated in the previous steps, the system further performs diagnostic identification and predictive analysis, aiming to determine the current macro life cycle stage of the brand (diagnosis) and predict its future value trend (prediction) in parallel.

[0119] S601, To achieve accurate classification of brand lifecycle stages, the system first constructs a classification feature vector that can characterize the core dynamics of the system. In this embodiment, this feature vector is the maximum eigenvalue calculated in the previous time step (step S502). With the main feature vector It is pieced together. Its physical significance lies in the largest eigenvalue. It quantifies the drasticness or momentum of recent brand status changes, while the main feature vector... This indicates the main direction of change. The formula for constructing the classification feature vector is as follows:

[0120] ;

[0121] in, It is a four-dimensional eigenvector. Due to the eigenvalues... The numerical range and the principal feature vector The component ranges of a (unit vector) vary greatly. To avoid model training being dominated by dimensions, the system performs Z-score standardization on the feature vector before inputting it into the model.

[0122] S602, the system inputs the standardized classification feature vector into a pre-trained machine learning classification model to identify the life cycle stage of the brand. As a preferred approach, this model employs a Support Vector Machine (SVM) classifier and is configured with a Radial Basis Function (RBF) kernel.

[0123] Model Output: The direct output of the model is one of four predefined, independent lifecycle stage labels: {Introduction, Growth, Maturity, Decline}. To quantify the uncertainty of the classification results, in this embodiment, the system further converts the SVM output into posterior probabilities for each stage using techniques such as Platt scaling, for example, {Introduction: 0.05, Growth: 0.85, Maturity: 0.1, Decline: 0.0}, thereby providing decision-makers with richer confidence information.

[0124] Model Training: This SVM classifier requires offline training. The training sample set consists of standardized classification feature vectors from historical time series data. The corresponding ground truth labels are manually labeled by business experts based on market reports, sales performance, and key operational events for each historical period. During training, the regularization parameter of the SVM and the width parameter of the RBF kernel are optimized using k-fold cross-validation.

[0125] S603, To quantitatively predict the future asset value of a brand, the system employs a deep learning model specifically designed for time series forecasting. In this embodiment, the model is a Long Short-Term Memory (LSTM) network.

[0126] Model Input: The input to the LSTM model is a sequence of past data. Three-dimensional state vectors at consecutive time steps The sequence formed, i.e. Among them, the length of the playback window. The value of depends on the length of the business cycle; for example, it can be set to 90 time steps. Before constructing the input sequence, the system standardizes all state vectors using the mean and standard deviation of the training set to ensure the numerical stability of the network input.

[0127] Model structure: This LSTM network consists of an input layer, two stacked LSTM layers (each containing, for example, 64 hidden units, followed by a Dropout layer to prevent overfitting), and a fully connected output layer. The output layer contains only one neuron and uses a linear activation function.

[0128] Model output: The model output is the result of a time step in the future. Pure brand-driven incremental growth Predicted value Among them, predictive horizon It can be configured according to business needs.

[0129] Model Training: This LSTM network also requires offline training. Training data is generated by sliding a window across a sequence of historical state vectors, producing a large number of (input sequence, target value) sample pairs. The training objective is to minimize the difference between the predicted and true values, and the loss function uses Mean Squared Error (MSE).

[0130] ;

[0131] in, Indicates mean square error; This represents the number of samples in the training batch. For the first Predicted values ​​for each sample; For the first The true value of each sample; This represents the loop iteration variable in the addition operation, specifically used to iterate through samples within a batch. The training process uses the Adam optimizer for gradient descent and parameter updates.

[0132] S604, the computing device integrates the classification results (including confidence levels) of the lifecycle stages with the prediction results of future asset value, and outputs them in the form of a graphical user interface or API interface. For example, the system can display the current stage: growth stage (confidence level 85%), with a projected 5% increase in brand incremental growth in the next cycle, providing brand managers with complete decision support information from current status diagnosis to future trend warning.

[0133] Specific application examples:

[0134] This embodiment takes a target proprietary brand (XXX smart noise-canceling headphones) as the target object and aims to use the system and method of the present invention to predict its brand growth value.

[0135] Background setting:

[0136] Target own brand: XXX smart noise-canceling headphones;

[0137] Discrete time step: Data acquisition and calculation are performed on a daily basis. (representing the number of days);

[0138] The sample set for the white-label control group of the same product category: Based on the set physical attribute matching rules, 15 reference products sold on the same e-commerce platform, using the same noise reduction chip solution, but without registered brand trademarks were selected from the relational database to form the sample set. ).

[0139] Computing equipment: running Linux, configured with an Intel Xeon processor, 128GB of memory, and deploying a PostgreSQL relational database.

[0140] Initial moment: That is, the first day that the target's own brand is launched for sale.

[0141] Step S10: Data Acquisition and System Initialization

[0142] Data extraction: The system in At the initial time, the business data stream of the target private label brand is extracted from the relational database for the first time, including price parameters. Sales parameters The system began recording timestamps of repeat purchase orders and brand keyword search frequency metrics. Simultaneously, it extracted data streams (price and sales parameters) from 15 reference products within the same category's white-label control group sample set.

[0143] Initial weight allocation: Based on the principle of maximum entropy, the system assigns absolutely equal initial weight values ​​to each reference product in the white-label control group sample set of the same category:

[0144] ;

[0145] Principal eigenvector initialization: To trigger subsequent feedback loop closure, the system initializes the initial principal eigenvector with a unit direction vector.

[0146] ;

[0147] The vector is written in reverse to the system cache as the baseline anchor input parameter for the next time step calculation.

[0148] Step S20: Dynamic reference anchoring (with) (For example)

[0149] Assume the system has reached the discrete time step. .

[0150] Extracting the main feature vector: The system caches the previous time step from the relational database ( Extracting the main feature vector Assume the vector is This indicates that the most dramatic changes in brand status recently have been primarily driven by pure brand-driven incremental growth.

[0151] Constructing fluctuation vectors: The system constructs a three-dimensional fluctuation vector for each reference product in the control group sample set of the same category of white-label products. Because white-label products lack repeat purchase and search tracking records, the system performs constant zero-padding when extracting missing fields.

[0152] Calculate the absolute value projection: For each reference commodity, the system calculates its fluctuation vector in... absolute projection on For example, the price of the third reference product rose sharply on that day, and its fluctuation trend highly overlapped with the core development direction of the private label brand, leading to... The value is relatively large.

[0153] ;

[0154] Calculate dynamic weights: The system performs adaptive mapping of dynamic weights based on the absolute value projection values. The larger third reference product was deemed to be affected by assimilation from the spillover of its own-brand traffic and was therefore assigned a lower benchmark reference weight. .

[0155] ;

[0156] Generate benchmark data: The system uses updated dynamic weights to perform weighted aggregation on the data of the control group sample set of the same category of white-label products, generating benchmark price and benchmark sales data of the same category of white-label products at the current time step:

[0157] ;

[0158] ;

[0159] Step S30: Incremental extraction ( )

[0160] Obtain Own Brand Parameters: The system reads the price parameters of the target own brand at the current discrete time step. With sales parameters .

[0161] Condition determination: The system performs numerical comparison calculations.

[0162] ;

[0163] ;

[0164] Extracting incremental metrics: Under the condition that all judgment conditions are met, the system extracts pure brand-driven incremental metrics from the total transaction data.

[0165] ;

[0166] Step S40: State-space modeling ( )

[0167] Calculate the repurchase attenuation rate: The system calculates the average repurchase time interval from the set of repurchase order timestamps. The decay rate of the repurchase interval is obtained by mapping through a negative exponential function. .

[0168] Calculate the normalized search index: the original brand keyword search frequency index for the day is... In the past 90 days ( Within the sliding history window, the system performs maximum and minimum value normalization processing to obtain... .

[0169] Constructing a 3D state vector: The system concatenates the above three independent dimensions of indicators into a matrix to construct a 3D state vector representing the state of the current time step node.

[0170] ;

[0171] Step S50: Feature decomposition and feedback ( )

[0172] Constructing the covariance matrix: The system sets the sliding time window length to 30 ( Extract the continuous three-dimensional state vector from day 71 to day 100, and construct a 3x3 state covariance matrix. .

[0173] Execution feature decomposition: computing device pair Perform eigenvalue decomposition to obtain the largest eigenvalue. and the corresponding principal feature vector .

[0174] Assumption This vector points in the direction of the most dramatic change in brand asset status within the current time window.

[0175] Reverse write cache: The computing device outputs the principal feature vector at the current time step. Write the reverse data to the system cache as the next time step. The baseline anchoring input parameters are used during calculation to complete the feedback closed-loop design.

[0176] Step S60: Stage Prediction ( )

[0177] Phase Diagnosis: System Construction of Classification Feature Vectors The model outputs the posterior probabilities for each stage: {Introduction: 2%, Growth: 92%, Maturity: 6%, Decline: 0%}.

[0178] Value prediction: The system uses the three-dimensional state vector sequence of the past 90 time steps. Input a pre-trained Long Short-Term Memory (LSTM) network model. The model outputs a predicted time point (e.g., the 30th day in the future). The quantitative valuation of brand assets, namely .

[0179] Output Results: The system integrates diagnostic and predictive results, for example: The target private label brand is currently in its growth stage (92% confidence level), with strong core brand momentum. It is estimated that in one month, the daily incremental growth driven by the brand alone will reach 52,000 yuan.

[0180] Experimental verification and effect comparison

[0181] To verify the effectiveness of the method of this invention, we collected daily business data from three new consumer electronics private label brands on a leading domestic e-commerce platform for 500 days from their launch, forming an experimental dataset. We then compared the method of this invention with two benchmark methods.

[0182] It contains business data from 3 brands across 1500 time steps. The first 400 days of data were used as the training set, and the last 100 days of data were used as the test set.

[0183] Comparison method:

[0184] Method 1 (ARIMA Model): A classic time series forecasting model that uses only the pure brand-driven incremental (BDI) univariate series of the target brand's history for forecasting, without considering multidimensional information and external market benchmarks.

[0185] Method 2 (VAR Model): Vector Autoregression (VAR) model, a commonly used multivariate time series model. This model uses three variables simultaneously: BDI, repurchase rate, and search index, but employs a static, equally weighted average price of white-label products as the market benchmark and lacks a feedback mechanism.

[0186] The method of this invention: adopts the complete process of this invention, including dynamic benchmark anchoring, state space modeling, feature decomposition feedback, and SVM-based stage diagnosis and LSTM-based value prediction.

[0187] Evaluation indicators:

[0188] Value prediction tasks: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE).

[0189] Phase diagnostic tasks: accuracy and macro average F1 score.

[0190] Experimental results:

[0191] Table 1. Performance Comparison of Value Prediction (BDI)

[0192] method MAE (RMB) RMSE (RMB) MAPE (%) Method 1 (ARIMA) 7，852.6 9，941.3 18.5% Method 2 (VAR) 4，120.1 5，388.7 9.7% Method of the present invention 1，986.4 2，605.9 4.6%

[0193] Table 2. Comparison of diagnostic performance at different lifecycle stages

[0194] method Accuracy Macro F1-Score Method 2 (VAR+SVM) 78.3% 0.76 Method of the present invention 94.1% 0.93

[0195] Conclusions and analysis of figures:

[0196] Combined with appendix Figure 3As shown in the figure, this graph compares the predictive performance of Pure Brand-Driven Increment (BDI) over a test period of 100 time steps. The horizontal axis represents time steps (days), and the vertical axis represents BDI (yuan). In the graph, the actual BDI value is represented by a solid black line, while the **predicted BDI value (using the method of this invention)** is represented by a dashed gray line. It is clear from the graph that throughout the test period, the predicted value represented by the dashed gray line and the actual value represented by the solid black line exhibit a high degree of fit. Specifically, during the rapid growth phase of brand value from 0 to approximately 35 time steps, and the subsequent stabilization phase, the overall trend, growth rate, and details of local fluctuations of the two curves are largely consistent. The BDI value ranges from nearly zero yuan initially to over 80,000 yuan, and the method of this invention consistently provides closely following predictions. This intuitively demonstrates the superior predictive accuracy and robustness of the method of this invention in the quantitative valuation of brand assets.

[0197] Combined with appendix Figure 4 As can be seen, this chart, presented as a stacked area diagram, displays the dynamic diagnostic probabilities of brand lifecycle stages. The horizontal axis represents time steps (days), and the vertical axis represents stage probability, ranging from 0 to 1. According to the legend, the four lifecycle stages—introduction, growth, maturity, and decline—are represented by four different shades of gray, from lightest to darkest. In the initial testing phase (approximately the first 20 time steps), the lightest gray area, representing the introduction stage, dominates, with its probability gradually decreasing from over 0.7. Simultaneously, the area of ​​the next lightest gray area, representing the growth stage, rapidly expands, and its probability correspondingly rises from below 0.2 to become dominant. After approximately the 25th time step, the probability of the growth stage stabilizes at an absolutely high level above 0.8, while the darker areas representing the maturity and decline stages maintain extremely low probabilities. This smooth transition from the introduction to the growth stage accurately reflects the dynamic shifts in the brand lifecycle, validating the effectiveness and accuracy of the system's diagnostic function.

[0198] Combined with appendix Figure 5As shown in the figure, the bar chart quantitatively compares the errors of different prediction methods. The horizontal axis represents the prediction method, including ARIMA, VAR, and the method of this invention; the vertical axis represents the mean absolute percentage error (%). The numerical label at the top of each bar clearly indicates its error magnitude. Specifically, the ARIMA method has the highest error at 18.5%; the VAR method has a moderate error at 9.7%; and the method of this invention has the lowest error at only 4.6%. Visually, the bar representing the method of this invention (represented by the darkest gray) is shorter than the corresponding bars for the ARIMA method (represented by light gray) and the VAR method (represented by medium gray). This result clearly and quantitatively demonstrates that the method of this invention, by introducing dynamic benchmark anchoring and feature decomposition feedback mechanisms, has a decisive advantage over traditional time series models in filtering out market noise and improving prediction accuracy.

Claims

1. A multi-dimensional data-driven method for predicting the growth value of proprietary brands, characterized in that, Includes the following steps: At a set discrete time step, acquire the business data stream of the target own brand and the data stream of the control group sample set of white-label products in the same category; Extract the principal feature vector of the previous time step, construct a three-dimensional fluctuation vector for each reference product in the sample set of the white-label control group of the same category, calculate the absolute value projection of the three-dimensional fluctuation vector of each reference product on the principal feature vector of the previous time step, calculate the dynamic weight for the sample set of the white-label control group of the same category based on the absolute value projection, and generate the benchmark price and benchmark sales data of the white-label of the same category in the current time step. Based on the business data flow of the target proprietary brand, the benchmark price of the same category of white-label products, and the benchmark sales data, pure brand-driven incremental indicators are extracted. By combining the pure brand-driven incremental indicators, the repurchase time interval decay rate, and the normalized pure brand keyword active search index, a three-dimensional state vector representing the state of the current time step node is constructed. Within the set sliding time window, a state covariance matrix is ​​constructed based on the three-dimensional state vector, and eigenvalue decomposition is performed on the state covariance matrix to generate the principal eigenvector and the maximum eigenvalue of the current time step. At the same time, the principal eigenvector output at the current time step is used as the principal eigenvector of the previous time step when calculating the next time step. Based on the maximum eigenvalue and the principal eigenvector, a classification feature vector is constructed and input into a pre-trained machine learning classification model to output the diagnostic results of the brand life cycle transition stage. At the same time, the historical three-dimensional state vector sequence is input into a pre-trained time series prediction model to output the brand asset quantitative valuation at the predicted time node.

2. The method for predicting the growth value of a proprietary brand driven by multi-dimensional data according to claim 1, characterized in that, The specific steps for constructing a three-dimensional state vector representing the state of the current time step node include: Obtain the pure brand-driven incremental metrics; The repurchase time interval decay rate is calculated based on the set of repurchase order timestamps contained in the business data stream of the target proprietary brand. Based on the brand keyword search frequency index contained in the business data stream of the target proprietary brand, the normalized pure brand keyword active search index is obtained after normalization processing by sliding window maximum and minimum values. The three-dimensional state vector is constructed by concatenating the pure brand-driven incremental indicators, the repurchase time interval decay rate, and the normalized pure brand keyword active search index into a matrix.

3. The method for predicting the growth value of a proprietary brand driven by multi-dimensional data according to claim 1, characterized in that, The specific steps for generating the principal feature vector and the largest eigenvalue at the current time step include: Multiple consecutive three-dimensional state vectors are extracted within the sliding time window to construct the state covariance matrix; Perform eigenvalue decomposition on the state covariance matrix to obtain the maximum eigenvalue and the principal eigenvector.

4. The method for predicting the growth value of a proprietary brand driven by multi-dimensional data according to claim 1, characterized in that, The specific steps to extract pure brand-driven incremental metrics include: Determine whether the price parameter value of the target private label is higher than the benchmark price value of the same category of white-label products and whether the benchmark sales data is non-zero; When the determination result is yes, the pure brand-driven incremental indicator is extracted from the total transaction data; and when the determination result is no, the pure brand-driven incremental indicator is assigned a value of zero.

5. The method for predicting the growth value of a proprietary brand driven by multi-dimensional data according to claim 1, characterized in that, The specific steps for calculating dynamic weights for the control group sample set of the same product category include: A negative natural exponent function is introduced as the kernel function, and the absolute value projection is used as the input of the negative natural exponent function to generate the dynamic weights.

6. The method for predicting the growth value of a proprietary brand driven by multi-dimensional data according to claim 1, characterized in that, The specific steps for constructing a three-dimensional wave vector include: The first-order time difference data of the reference product at the current discrete time step is extracted to obtain the price parameter difference value. At the same time, the missing repurchase and search-related parameters are padded with constant zeros, and then combined to form the three-dimensional fluctuation vector.

7. The method for predicting the growth value of a proprietary brand driven by multi-dimensional data according to claim 1, characterized in that, The specific steps for outputting diagnostic results for the brand lifecycle transition stages include: The classification feature vector is input into the pre-trained machine learning classification model, which serves as a support vector machine classifier, to output diagnostic results for the brand lifecycle transition stage.

8. The method for predicting the growth value of a proprietary brand driven by multi-dimensional data according to claim 1, characterized in that, The specific steps for outputting a quantitative valuation of brand assets at the predicted time point include: The historical three-dimensional state vector sequence is input into the pre-trained time series prediction model, which is a long short-term memory network model, to output the brand asset quantitative valuation at the predicted time node. The historical three-dimensional state vector sequence is obtained by arranging multiple consecutive three-dimensional state vectors in chronological order.

9. The method for predicting the growth value of a proprietary brand driven by multi-dimensional data according to claim 1, characterized in that, The business data stream includes the target proprietary brand's price parameters, sales parameters, repeat purchase order timestamp set, and brand keyword search frequency metrics.

10. The method for predicting the growth value of a proprietary brand driven by multi-dimensional data according to claim 1, characterized in that, It also includes initialization, the initialization steps of which include: Assign equal initial weights to each reference product in the same category of white-label control group sample set, and initialize the principal feature vector of the previous time step as a unit direction vector.

Images