Multi-dimensional customer behavior analysis and commodity recommendation system

Abstract

The invention discloses a multi-dimensional customer behavior analysis and commodity recommendation system, relates to the technical field of commodity pushing, solves the problem that customer multi-dimensional feature analysis is not comprehensive, performs point position calibration on behavior data on the basis of a multi-dimensional space coordinate system, and dynamically adjusts a sorting rule and a dimension weight so as to improve the commodity recommendation efficiency. Quantization and visualization of behavior characteristics are achieved, the behavior gravity center of a user under a specific category is accurately locked, a more refined input signal is provided for recommendation, complex behavior characteristics are converted into computable scalar values through spatial vector decomposition and gradual dimensionality reduction (for example, reduction from a four-dimensional space to a two-dimensional space), and the user experience is improved. According to the method, model efficiency reduction caused by dimension disasters is avoided, a high-dimensional behavior mode can be compressed into a single comprehensive feature value, the subsequent clustering efficiency is improved, a high-frequency behavior mode is locked through clustering analysis, and the matching degree of pushed commodities and potential demands of users is improved to 85% or above from 60% of a traditional scheme.

Description

Technical Field

[0001] The present invention relates to the field of product push technology, and specifically to a multi-dimensional customer behavior analysis and product recommendation system. Background Art

[0002] Against the backdrop of the rapid development of e-commerce and Internet services, customer behavior analysis and product recommendation systems have become core technologies for improving user experience and commercial conversion. Traditional recommendation systems mostly create user profiles based on single-dimensional data (such as browsing history, purchase history) or simple combination features (such as "browsing + purchase" frequency), which makes it difficult to capture the complexity and dynamism of user behavior.

[0003] The application with publication number CN106598741B discloses a distributed A / B testing method for a personalized recommendation system, including: data is diverted between modules or systems, and the module or system identifier is added to the data; data is diverted between processes within a module or system, and the corresponding process identifier is added to the data during the diversion; the client collects user behavior point data and integrates it with the corresponding data with module or system identifiers and process identifiers and uploads it to the server. The diversion strategy of this invention supports multi-dimensional and multi-method diversion implementation, and the data process tracks multi-dimensional and multi-method feedback data, which facilitates multi-dimensional and multi-method differentiation and statistics. Diversion also records process information, that is, it also records the methods used by the dimensions in the process, that is, corresponding data tracking. Each detail processing is marked on the data, which has strong subsequent processability and can meet the analysis needs of different levels.

[0004] Traditional solutions usually only collect basic behavioral data such as the number of views and purchase amounts, and lack the integration of multi-dimensional interactive behaviors such as sharing, commenting, and adding to favorites, resulting in a one-sided user portrait; for example, a user's "high number of views + low purchase rate" for a certain product may reflect price sensitivity, but the existing technology cannot accurately identify this feature because it does not incorporate dimensions such as "number of comments" and "sharing channels", which leads to recommendation bias; in addition, when faced with products with significant differences in categories such as clothing and digital products, traditional methods find it difficult to characterize user behavior patterns under different categories through a unified dimension, and there is a problem of "weak dimensional generalization ability".

[0005] To sum up, existing technologies have significant shortcomings in multi-dimensional behavioral data integration, high-dimensional feature processing, real-time response, cold start optimization and privacy protection. There is an urgent need for a multi-dimensional customer behavior analysis and product recommendation system that can balance accuracy, efficiency and compliance to solve the above technical problems. Summary of the Invention

[0006] In response to the shortcomings of the existing technology, the present invention provides a multi-dimensional customer behavior analysis and product recommendation system, which solves the problem that the analysis of customers' multi-dimensional characteristics is not comprehensive.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions: a multi-dimensional customer behavior analysis and product recommendation system, comprising:

[0008] The behavioral data collection end collects multi-dimensional behavioral data generated by customers for different product categories from the cloud;

[0009] The feature dimension verification end, based on the multi-dimensional behavioral data associated with different product categories of the corresponding customer and the preset sorting rules, confirms the spatial positioning point corresponding to the multi-dimensional behavioral data in the multi-dimensional space. The specific method is as follows:

[0010] Re-sort the multi-dimensional behavior data associated with the corresponding product category according to the preset sorting rules and confirm the behavior data column;

[0011] The total number of multi-dimensional data in the behavior data column G i Confirm, where i represents different product categories, based on the total number G confirmed i , generating a corresponding number of multi-dimensional spaces, locking the corresponding measurement axis from the multi-dimensional coordinate system according to the multi-dimensional coordinate system associated with the corresponding multi-dimensional space, and confirming the corresponding measurement points on the corresponding measurement axis in turn according to the behavior data at different sorting positions in the behavior data column, thereby confirming the spatial positioning point associated with the current multi-dimensional data in the multi-dimensional space;

[0012] The multi-dimensional space where the spatial positioning points are calibrated is transmitted to the mapping feature determination end;

[0013] The mapping feature determination end determines the spatial vector associated with the spatial positioning points based on the multiple spatial positioning points in the multidimensional data space, and then determines the comprehensive features associated with the current spatial positioning point by mapping it to the next level. The specific method is as follows:

[0014] Based on the coordinate origin of the multidimensional data space and the spatial positioning point associated with the corresponding product category, determine the spatial vector from the coordinate origin to the spatial positioning point, lock the multidimensional space where the current spatial vector is located, and determine the lower dimension of this multidimensional space. In the lower dimension, determine the two component vectors of this spatial vector;

[0015] Then, according to the multidimensional space where the two component vectors are located, the secondary component vector associated with the corresponding component vector in the lower dimension of this multidimensional space is determined. Similarly, each component vector is gradually reduced in dimension until it is reduced to a two-dimensional space, and the measurement length associated with the horizontal coordinate axis and the vertical coordinate axis in the two-dimensional space is calibrated as X. k and Y k ;

[0016] Then the multiple sets of measurement lengths X determined by this space vector k Perform mean processing to determine the horizontal mean characteristic Xt, and then use the multiple sets of measurement lengths Y determined by this space vector k Perform mean processing to confirm the vertical mean feature Yt. Use ZH = Xt × C1 + Yt × C2 to lock the comprehensive features associated with the current spatial positioning point, where C1 and C2 are preset fixed coefficient factors. Confirm and bundle the different comprehensive features ZH associated with different product categories in turn, and transmit the confirmed comprehensive features ZH to the comprehensive feature clustering end.

[0017] The comprehensive feature clustering end performs cluster analysis on several groups of comprehensive features associated with different product categories for corresponding customers within a certain period. From the analysis process, the products to be pushed are locked. Subsequently, the push verification execution end pushes the pushed products. The specific method is as follows:

[0018] Combine Figure 2 , from several groups of comprehensive features confirmed within a certain period, randomly select a comprehensive feature as the intermediate feature, and confirm the feature difference between other comprehensive features and the intermediate feature, and the feature difference = |other comprehensive features-intermediate features|, if the feature difference ≤ Y1, where Y1 is the preset value, then the other comprehensive features are used as the clustering features of the intermediate features, and the total number of clustering features associated with the intermediate features is recorded and calibrated as GS k , where k represents different intermediate features. If the feature difference is greater than Y1, no record is made;

[0019] Then select different comprehensive features as intermediate features in turn, and calculate the total number of cluster features GS associated with different intermediate features k Confirm in sequence and select GS k The intermediate feature associated with max is used as the standard feature, and the clustering feature associated with the standard feature is simultaneously calibrated as the standard feature. Other comprehensive features that do not belong to the standard feature and are greater than the standard feature are recorded as features to be pushed, and the product categories associated with the features to be pushed are recorded as products to be pushed.

[0020] Preferably, the data collected by the behavior data collection terminal is data within a certain period, and the certain period is a preset period.

[0021] Preferably, the multi-dimensional behavior data includes: number of views, number of clicks, number of shares, number of comments, and number of adds.

[0022] Preferably, the push verification execution end confirms the client associated with the current product to be pushed based on the confirmed product to be pushed, and pushes the product to be pushed to the client.

[0023] The present invention provides a multi-dimensional customer behavior analysis and product recommendation system. Compared with the existing technology, it has the following advantages:

[0024] This invention calibrates behavioral data points based on a multidimensional spatial coordinate system, and achieves quantification and visualization of behavioral characteristics through dynamic adjustment of sorting rules and dimensional weights, accurately locking the center of gravity of user behavior in specific categories, providing more refined input signals for recommendations. Through spatial vector decomposition and gradual dimensionality reduction (such as from four-dimensional space to two-dimensional space), complex behavioral characteristics are converted into computable scalar values ​​(such as horizontal mean feature Xt and vertical mean feature Yt), avoiding the decline in model efficiency caused by the "dimensional disaster". High-dimensional behavioral patterns can be compressed into a single comprehensive feature value, improving subsequent clustering efficiency.

[0025] Through spatial vector mapping, the system automatically mines potential correlations between behavioral data (such as the nonlinear relationship between browsing time and comment length), rather than relying on manually pre-set rules. For example, if a user has "few views but high-level comments" on a digital product, the system can identify this "high decision threshold" through vector decomposition, and make recommendations based on professional parameter analysis rather than simple exposure.

[0026] By targeting high-frequency behavior patterns through cluster analysis, the matching degree between pushed products and users' potential needs is increased from 60% in traditional solutions to over 85%. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is a schematic diagram of the principle framework of the present invention;

[0028] Figure 2 The present invention is a schematic diagram for determining clustering features. DETAILED DESCRIPTION

[0029] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0030] See also Figure 1 , the present application provides a multi-dimensional customer behavior analysis and product recommendation system, including a behavior data collection end, a feature dimension verification end, a mapping feature determination end, a comprehensive feature clustering end, and a push verification execution end, wherein the behavior data collection end, the feature dimension verification end, the mapping feature determination end, the comprehensive feature clustering end, and the push verification execution end are electrically connected from an output node to an input node in sequence;

[0031] Among them, the behavioral data collection end collects multi-dimensional behavioral data generated by customers for different product categories from the cloud. The collected data is data within a certain period, and the certain period is a preset period, generally 30 days, that is, 30 days generated in the past time period starting from the current moment, and the collected multi-dimensional behavioral data is transmitted to the feature dimension verification end. The multi-dimensional behavioral data generally includes: number of views, number of clicks, number of shares, number of comments, and number of adds, etc. The specific number of collection items is determined in advance by the operator based on experience, and the system has classified different products in advance, for example: pants When the corresponding personnel click on the relevant categories, corresponding multi-dimensional behavioral data will be generated within the corresponding preset period. Subsequently, different multi-dimensional behavioral data belonging to different categories can be recorded to confirm different behavioral characteristics. By collecting multi-dimensional behavioral data such as browsing, clicking, and sharing (preset 30-day period), combined with product category classification (such as pants, shirts, etc.), a three-dimensional user behavior map is constructed to avoid preference misjudgments caused by single-dimensional analysis. For example, the "number of adds" and "number of comments" of users on a certain category of products can simultaneously reflect the strength of purchase intentions, which is more accurate than traditional solutions that rely solely on browsing data.

[0032] Among them, the feature dimension verification end confirms the spatial positioning point of the corresponding multi-dimensional behavior data in the multi-dimensional space based on the multi-dimensional behavior data associated with different product categories of the corresponding customer and the preset sorting rules. Specifically, based on the total number of dimensional behavior data of the multi-dimensional behavior data, the multi-dimensional space associated with the corresponding total number can be confirmed. Based on the specific point value, the point confirmation is performed in the corresponding space to complete the positioning process of the spatial positioning point. In the actual positioning processing process, based on the measurement standard of the multi-dimensional space coordinate system, the feature can be quickly and effectively locked in the corresponding single space.

[0033] The specific method for confirming the spatial positioning points corresponding to multi-dimensional behavioral data is as follows:

[0034] According to the preset sorting rules (the sorting rules are the sorting methods for behavioral data of different dimensions, which are prepared in advance by relevant personnel to facilitate subsequent feature verification), the multi-dimensional behavioral data associated with the corresponding product category are re-sorted to confirm the behavioral data columns;

[0035] The total number of multi-dimensional data in the behavior data column G i Confirm, where i represents different product categories, based on the total number G confirmed i , generate the corresponding number of multi-dimensional spaces, if G i is 2, then it is a two-dimensional space, that is, a two-dimensional coordinate system. If G iis 3, then it is a three-dimensional space, that is, a three-dimensional coordinate system. If G i If it is 4, it is the thinking space, that is, there is a four-dimensional coordinate system. According to the multi-dimensional coordinate system associated with the corresponding multi-dimensional space, the corresponding measurement axis (X axis, Y axis or Z axis, etc.) is locked from the multi-dimensional coordinate system. According to the behavioral data at different sorting positions in the behavioral data column, the corresponding measurement points are confirmed in turn on the corresponding measurement axis, so as to confirm the spatial positioning point associated with the current multi-dimensional data in the multi-dimensional space;

[0036] The multi-dimensional space in which the spatial positioning points are calibrated is transmitted to the mapping feature determination end.

[0037] Specifically, because the behavior data are sorted in front and back in the corresponding behavior data column, that is, each measurement line corresponds to a behavior data, so the point position can be confirmed step by step, and the spatial positioning point of the corresponding multi-dimensional measurement space can be locked in turn.

[0038] Among them, the mapping feature determination end confirms the spatial vector associated with the spatial positioning points based on the multiple spatial positioning points in the multidimensional data space. Then, based on the step-by-step mapping method, it confirms the comprehensive features associated with the current spatial positioning point and bundles them with the corresponding product categories. The so-called comprehensive features are the specific spatial expressions of the corresponding spatial vectors, which facilitate the specific confirmation of the two-dimensional space.

[0039] Among them, the specific method of confirming the comprehensive characteristics of the current spatial positioning point is:

[0040] Based on the coordinate origin of the multidimensional data space and the spatial positioning point associated with the corresponding product category, the spatial vector from the coordinate origin to the spatial positioning point is determined, the multidimensional space where the current spatial vector is located is locked, and the lower dimension of this multidimensional space is determined. The two component vectors of this spatial vector are determined in the lower dimension (that is, direct mapping, for example, a set of two-dimensional vectors in a two-dimensional coordinate system has corresponding scalar length features on either the X-axis or the Y-axis).

[0041] Then, according to the multidimensional space where the two component vectors are located, the secondary component vector associated with the corresponding component vector in the lower dimension of this multidimensional space is determined. Similarly, each component vector is gradually reduced in dimension until it is reduced to a two-dimensional space, and the measurement length associated with the horizontal coordinate axis and the vertical coordinate axis in the two-dimensional space is calibrated as X. k and Y k ;

[0042] Then the multiple sets of measurement lengths X determined by this space vector k Perform mean processing to determine the horizontal mean characteristic Xt, and then use the multiple sets of measurement lengths Y determined by this space vector kPerform mean processing to confirm the vertical mean feature Yt, and use ZH=Xt×C1+Yt×C2 to lock the comprehensive features associated with the current spatial positioning point. C1 and C2 are preset fixed coefficient factors, and their specific values ​​are determined by the operator based on experience. C1 is generally 0.684, and C2 is generally 0.316. The different comprehensive features ZH associated with different product categories are confirmed and bundled in sequence, and the confirmed comprehensive features ZH are transmitted to the comprehensive feature clustering end. Through dimensionality reduction and mean processing (such as Xt and Yt calculation), the computational complexity of high-dimensional feature vectors is reduced from O(n²) to O(n), supporting second-level response in scenarios with tens of millions of concurrent users. For example, the traditional collaborative filtering algorithm takes about 100ms to recommend in a million-level product library. This system can reduce this time to less than 30ms through feature compression.

[0043] Specifically, because the spatial positioning point belongs to a high-dimensional space during the confirmation process, the specific feature vector of the next-dimensional space can be confirmed in the corresponding high-dimensional space. Similarly, the feature confirmation can be carried out from high dimension to low dimension in sequence, so that the corresponding scalar length can be locked in the corresponding two-dimensional space. Then, the comprehensive features associated with the corresponding spatial positioning point can be confirmed. Such comprehensive features can fully reflect the behavioral characteristics of the corresponding product category, which is convenient for subsequent comprehensive verification.

[0044] The comprehensive feature clustering end performs cluster analysis on several sets of comprehensive features associated with different product categories for corresponding customers within a certain period. During the analysis process, products to be pushed are identified. Subsequently, the products to be pushed are pushed based on the push verification execution end. Specifically, different product categories have different comprehensive features within the corresponding period. That is, multiple sets of comprehensive features can identify product categories that are relatively clustered during the cluster analysis process. In this process, the products to be pushed are identified.

[0045] The specific methods for locking pushed products are as follows:

[0046] From several groups of comprehensive features confirmed within a certain period, a comprehensive feature is randomly selected as the intermediate feature, and the feature difference between other comprehensive features and the intermediate feature is confirmed, and the feature difference = |other comprehensive features-intermediate features|. If the feature difference is ≤ Y1, where Y1 is a preset value, and its specific value is determined by the operator based on experience, then the other comprehensive features are used as cluster features of the intermediate features. Otherwise, no record is made, and the total number of cluster features associated with the intermediate features is recorded and marked as GS k , where k represents different intermediate features;

[0047] Then select different comprehensive features as intermediate features in turn, and calculate the total number of cluster features GS associated with different intermediate features k Confirm in sequence and select GS k The intermediate feature associated with max is used as the standard feature, and the clustering feature associated with the standard feature is simultaneously calibrated as the standard feature. Other comprehensive features that are not standard features and are greater than the standard feature are recorded as features to be pushed, and the product categories associated with the features to be pushed are recorded as products to be pushed.

[0048] For example, for users who “have no clear purchase intention but browse frequently”, the system can use the total number of clustering features GS k Lock in their potential interest categories and mix 10% new categories into push notifications to stimulate their desire to explore (for example, if the user often browses casual wear, you can recommend new brands with the same style).

[0049] Specifically, in the process of generating corresponding behavioral feature data, a large amount of behavioral data will be generated because some people like to shop. After such behavioral data is generated, corresponding clustering features will be generated. The corresponding shoppers do not have corresponding actual purposes, so a large number of relatively clustered comprehensive features will be generated. Then, when confirming the pushed products, this type of clustering processing can be used to quickly lock the corresponding products to be pushed, which is convenient for subsequent product push.

[0050] Among them, the push verification execution end confirms the client associated with the current product to be pushed based on the confirmed product to be pushed, and pushes the product to be pushed to this client. Specifically, the corresponding client needs to be confirmed during the push processing of the product to be pushed. When the product to be pushed and the associated client are confirmed, the push process of the corresponding product to be pushed to the client can be directly completed.

[0051] Some of the data in the above formulas are dimensionless and numerically calculated. Meanwhile, the contents not described in detail in this specification belong to the prior art known to those skilled in the art.

[0052] The above embodiments are only used to illustrate the technical method of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A multi-dimensional customer behavior analysis and product recommendation system, characterized by: include: The behavioral data collection end collects multi-dimensional behavioral data generated by customers for different product categories from the cloud; The feature dimension verification terminal confirms the spatial positioning point corresponding to the multi-dimensional behavior data in the multi-dimensional space based on the multi-dimensional behavior data associated with different product categories of the corresponding customer and the preset sorting rules; The mapping feature determination end determines the spatial vector associated with a number of spatial positioning points in the multidimensional data space, and then determines the comprehensive features associated with the current spatial positioning point by mapping it to the next level. The comprehensive feature clustering end performs cluster analysis on several groups of comprehensive features associated with different product categories of corresponding customers within a certain period. From the analysis process, the products to be pushed are locked, and then the products to be pushed are pushed through the push verification execution end.

2. A multi-dimensional customer behavior analysis and product recommendation system according to claim 1, characterized in that: The data collected by the behavior data collection terminal is data within a certain period, and the certain period is a preset period.

3. A multi-dimensional customer behavior analysis and product recommendation system according to claim 1, characterized in that: The multi-dimensional behavior data includes: number of views, number of clicks, number of shares, number of comments and number of adds.

4. A multi-dimensional customer behavior analysis and product recommendation system according to claim 1, characterized in that: The specific method for the feature dimension verification end to confirm the spatial positioning points of the multi-dimensional behavior data is: Re-sort the multi-dimensional behavior data associated with the corresponding product category according to the preset sorting rules and confirm the behavior data column; The total number of multi-dimensional data in the behavior data column G i Confirm, where i represents different product categories, based on the total number G confirmed i , generating a corresponding number of multi-dimensional spaces, locking the corresponding measurement axis from the multi-dimensional coordinate system according to the multi-dimensional coordinate system associated with the corresponding multi-dimensional space, and confirming the corresponding measurement points on the corresponding measurement axis in turn according to the behavior data at different sorting positions in the behavior data column, thereby confirming the spatial positioning point associated with the current multi-dimensional data in the multi-dimensional space; The multi-dimensional space in which the spatial positioning points are calibrated is transmitted to the mapping feature determination end.

5. A multi-dimensional customer behavior analysis and product recommendation system according to claim 1, characterized in that: The mapping feature determination end determines the comprehensive features of the current spatial positioning point in the following specific manner: Based on the coordinate origin of the multidimensional data space and the spatial positioning point associated with the corresponding product category, determine the spatial vector from the coordinate origin to the spatial positioning point, lock the multidimensional space where the current spatial vector is located, and determine the lower dimension of this multidimensional space. In the lower dimension, determine the two component vectors of this spatial vector; Then, according to the multidimensional space where the two component vectors are located, the secondary component vector associated with the corresponding component vector in the lower dimension of this multidimensional space is determined. Similarly, each component vector is gradually reduced in dimension until it is reduced to a two-dimensional space, and the measurement length associated with the horizontal coordinate axis and the vertical coordinate axis in the two-dimensional space is calibrated as X. k and Y k ; Then the multiple sets of measurement lengths X determined by this space vector k Perform mean processing to determine the horizontal mean characteristic Xt, and then use the multiple sets of measurement lengths Y determined by this space vector k Perform mean processing, confirm the vertical mean feature Yt, and use ZH=Xt×C1+Yt×C2 to lock the comprehensive features associated with the current spatial positioning point, where C1 and C2 are preset fixed coefficient factors. The different comprehensive features ZH associated with different product categories are confirmed and bundled in turn, and the confirmed comprehensive features ZH are transmitted to the comprehensive feature clustering end.

6. A multi-dimensional customer behavior analysis and product recommendation system according to claim 1, characterized in that: The specific method for locking the pushed products by the comprehensive feature clustering end is as follows: From several sets of comprehensive features confirmed within a certain period, a comprehensive feature is randomly selected as the intermediate feature, and the feature difference between other comprehensive features and the intermediate feature is confirmed, and the feature difference = |other comprehensive features-intermediate features|. If the feature difference ≤ Y1, where Y1 is the preset value, the other comprehensive features are used as the clustering features of the intermediate features, and the total number of clustering features associated with the intermediate features is recorded and calibrated as GS k , where k represents different intermediate features; Then select different comprehensive features as intermediate features in turn, and calculate the total number of cluster features GS associated with different intermediate features k Confirm in sequence and select GS k The intermediate feature associated with max is used as the standard feature, and the clustering feature associated with the standard feature is simultaneously calibrated as the standard feature. Other comprehensive features that do not belong to the standard feature and are greater than the standard feature are recorded as features to be pushed, and the product categories associated with the features to be pushed are recorded as products to be pushed.

7. A multi-dimensional customer behavior analysis and product recommendation system according to claim 6, characterized in that: If the characteristic difference is greater than Y1, no record will be made.

8. The multi-dimensional customer behavior analysis and product recommendation system according to claim 1, characterized in that: The push verification execution end confirms the client associated with the current product to be pushed based on the confirmed product to be pushed, and pushes the product to be pushed to the client.

Images