An Automatic Interaction Method and System Based on a Retail Large Model

Abstract

The present invention relates to the field of computer technology, and particularly to an automatic interaction method and system based on a retail large model. Among them, an automatic interaction method based on a retail large model includes the following steps: S1: Real-time record user behavior; S2: Obtain a matching asset matrix in a preset asset library; S3: Sequentially associate and push several assets in the asset matrix, and use the retail large model to judge the assets preferred by the user according to the user behavior record, and generate a preferred asset list in order according to the degree of preference; S4: Establish a retrieval matrix according to the sorting rules preset for the selected asset characteristics, and retrieve in the asset matrix corresponding to the asset characteristics in the asset library through the retrieval matrix to obtain a matching asset matrix. The present invention can automatically optimize and push highly compatible assets for users according to the user profile and user behavior through the retail large model, can reduce the occupation of computing power resources, and improve the recommendation effect.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and particularly to an automatic interaction method and system based on a retail large model. Background Art

[0002] The retail industry is the part of commercial activities that directly faces consumers, covering various sales channels from physical stores to e-commerce. With the progress of technology and the change of consumer preferences, the retail industry is undergoing a profound transformation. More and more transaction activities are transferred to online platforms. At the same time, with the development of AI technology and big data analysis, in order to optimize product recommendations and improve transaction efficiency, AI technology is used to provide unique products and services for each customer according to user preferences and historical behaviors. For example, the retail recommendation method, system and medium based on data processing of a graph database disclosed in a Chinese patent (Publication No.: CN115797020B). In this patented technology, the method includes: collecting product interaction information and consumption characteristic information, inputting them into a retail event graph database to generate a retail relationship model organization tree, generating a target consumer group characteristic portrait, then extracting consumption tendency data and comparing it with the product data information in the graph database to obtain product retail intention data, processing the consumption tendency data and the product retail intention data to obtain a supply and sales fit correlation index, and then comparing it with a preset retail relationship fitting threshold to judge the effectiveness of retail product recommendations; thus, using the graph database to obtain consumption tendency data, performing a fit degree process with the product retail intention data obtained by comparing with the product data information, and judging the retail product recommendation situation, realizing the technology of recommending and judging consumer groups and retail products through graph database technology and big data. When it works, it not only needs to analyze the information and behaviors of users, but also needs to preprocess the stored products. The general processing method is to set data mapping relationships, and through a large number of data mapping relationships to process the association degree between products. By manual, specification library division or AI technology automatic processing and other methods, several associated products are added to the products, and finally the storage of associated products is realized. Although it can reduce the computing power resources occupied by the AI model to a certain extent, the large-scale mapping relationship table will also bring a large data storage pressure, and the operation and maintenance of adding or deleting data later is more complex, and the operating cost is higher. Summary of the Invention

[0003] The technical problem to be solved by the present invention: The existing product recommendation methods rely on AI technology. When facing the product processing work with a large amount of data, it will occupy more computing power resources, the recommendation effect is poor, the storage of products is also more difficult, the data storage pressure is large, the operation and maintenance of adding or deleting data later is more complex, and the operating cost is higher.

[0004] To solve the above technical problems, the first aspect of the present invention adopts the following technical solution: An automatic interaction method based on a retail large model, comprising the following steps:

[0005] S1: Access the user, obtain the historical user portrait, and record the user behavior in real time;

[0006] S2: Call the preset retail large model to obtain the matching asset matrix in the preset asset library according to the user portrait and user behavior;

[0007] S3: According to the fit degree between the user and the assets, sequentially associate and push several assets in the asset matrix, use the retail large model to judge the assets preferred by the user according to the user behavior record, and generate a preferred asset list in order according to the preference degree;

[0008] S4: Obtain several assets in the preferred asset list, divide and establish a retrieval matrix according to the preset sorting rules of the selected asset characteristics, retrieve in the asset matrix corresponding to the asset characteristics in the asset library through the retrieval matrix, obtain the matching asset matrix, and transfer to step S3.

[0009] When the present invention works, it can automatically optimize and push assets with high fit degree for the user according to the user portrait and user behavior through the retail large model, and can cyclically obtain the assets preferred by the user during the pushing process. At the same time, a retrieval matrix is established according to the asset characteristics of the preferred assets to retrieve in the asset library, so that the retail large model only needs to analyze the user, reducing the process of analyzing commodities. Only through feature matching can the scale of data to be called when pushing commodities be greatly reduced. While greatly reducing the occupied computing power resources, it can also improve the data security coefficient of the asset library.

[0010] Preferably, in the step S1, when accessing the user, obtaining the historical user portrait, and recording the user behavior in real time, the following steps are adopted: Access the user, obtain the historical user portrait, and record the user behavior in real time when the user browses. The recorded user behavior includes at least one of user search information, user browsing information, user page stay time, and user click record.

[0011] Preferably, in the step S2, when calling the preset retail large model to obtain the matching asset matrix in the preset asset library according to the user portrait and user behavior, the following steps are adopted:

[0012] A1: Obtain the user portrait and user behavior of the user, call the preset retail large model to split several asset characteristics related to the user according to the user portrait and user behavior of the user, dynamically weight the several asset characteristics according to the current scenario, and then use the method of aggregation statistics to obtain the statistical results of the several asset characteristics and arrange and store them to obtain the set of asset characteristics preferred by the user;

[0013] A2: The asset feature set is used to locate matching assets in the asset matrix corresponding to the selected asset feature type in the preset asset library, and the asset matrix of suitable data size is extracted from the asset matrix according to the matching assets using preset interception specifications.

[0014] Preferably, the asset library is established by the following steps:

[0015] B1: Determine the associated hierarchical distribution of several asset features, establish several source nodes, each source node is filled with asset features of the corresponding level as several first child nodes, and the several first child nodes are arranged in sequence into row vectors according to a preset sorting rule;

[0016] B2: Each first sub-node is filled with asset features of the corresponding level as a number of second sub-nodes, and the number of second sub-nodes are arranged in sequence into a column vector according to a preset sorting rule;

[0017] B3: Call the retail model to split each asset into several asset features, add the index of the asset to each asset feature, and configure the several asset features as matching second child nodes;

[0018] B4: Traverse several row vectors of several source nodes to form the elements of the first row of several asset matrices respectively, and fill the corresponding column vectors to form elements of several columns respectively, so as to generate several asset matrices and store them as asset libraries.

[0019] When the present invention is working, it can realize the storage of assets. By splitting asset features and storing assets through category classification, the association operations between several assets are reduced, and the data mapping relationship table can be saved when the asset scale is large. The pressure of data storage can be reduced to a certain extent. At the same time, it is also convenient to update the asset library when adding or reducing assets, and the workflow is optimized. By establishing several asset matrices, the hidden association information between different assets can be excavated, and the conversion rate of asset push can be improved while ensuring the compatibility of asset push, and no additional computing power resources are required, with high economic benefits.

[0020] Preferably, in the step S3, according to the matching degree between the user and the assets, several assets in the asset matrix are associated and pushed in sequence. When using the retail large model to judge the assets preferred by the user based on the user behavior records and generating a preferred asset list in the order of preference degree, the following steps are adopted: Using the retail large model to locate the matching assets in the matching asset matrix according to the user's preference feature set, and at the same time predicting at least one predicted preferred asset according to the user's preference feature set by the retail large model. Determine the associated push direction in the asset matrix through a preset push algorithm, and push several assets in the asset matrix in sequence. Use the retail large model to judge the assets preferred by the user based on the user behavior records and generate a preferred asset list in the order of preference degree.

[0021] Preferably, in the step S3, when determining the associated push direction in the asset matrix through a preset push algorithm and pushing several assets in the asset matrix in sequence, the following steps are adopted:

[0022] C1: Obtain the asset characteristics corresponding to the matching assets and the positions of the predicted preferred assets in the current asset matrix, calculate the initial vector from the matching assets to the predicted preferred assets, and calculate the unit vector. Calculate the position of the next preferred asset according to the unit vector and the preset step size;

[0023] C2: Intercept the matching asset matrix according to the position of the next preferred asset through a preset intercept specification. When all the matching assets in the current asset matrix have been pushed, go to step C3;

[0024] C3: When there are new preferred assets in the current asset matrix, go to step C4; otherwise, go to step C5;

[0025] C4: Calculate the unit vector from the current preferred asset to the preset preferred asset and the unit vector from the current preferred asset to the new preferred asset. Obtain the corrected unit vector through vector addition. Calculate the position of the next preferred asset according to the unit vector and the preset step size. Judge whether the position of the next preferred asset is the predicted position or whether the number of loops has reached the preset threshold. If so, end and record all the preferred assets to generate a preferred asset list, and go to step S4; otherwise, go to step C2;

[0026] C5: Dynamically adjust the step size according to the number of loops. Calculate the position of the next preferred asset according to the unit vector and the accumulated step size. Judge whether the position of the next preferred asset is the predicted position or whether the number of loops has reached the preset threshold. If so, end and record all the preferred assets to generate a preferred asset list, and go to step S4; otherwise, go to step C2.

[0027] When the present invention is working, the push algorithm is used to determine the associated push direction in the asset matrix, and the sequential push of several assets in the asset matrix can be realized. The push path can be corrected according to the real-time changes of user preferences, and the effect of asset push can be further improved, and the conversion rate can be improved. At the same time, the final push target can be corrected, and the user portrait can be further improved. Through multiple cycles of push, the user portrait can be accurately portrayed, thereby assisting the retail big model to push assets with higher fit to users.

[0028] Preferably, step C2 further includes the following steps: when a matching asset matrix is intercepted according to the position of the next preferred asset through a preset interception specification, the interception path and the asset matrix intercepted each time are synchronously recorded to obtain a number of optimized push matrices, and the several optimized push matrices are processed through the retail big model to obtain a number of asset push paths for reducing the consumption of computing resources.

[0029] Preferably, in step S4, a number of assets in the preferred asset list are obtained, and a search matrix is established according to the preset sorting rules of the selected asset features. The asset matrix corresponding to the asset features in the asset library is searched through the search matrix to obtain a matching asset matrix. When entering step S3, the following steps are adopted:

[0030] D1: Obtain several assets in the preferred asset list, cluster their asset features and determine at least one common asset feature and at least two similar asset features;

[0031] D2: Determine the matrix position of the selected asset according to the preset sorting rules based on the similar asset characteristics, obtain several groups of matrix positions, calculate the distance of each matrix position in each group of matrix positions and verify them, and eliminate the related asset characteristics whose distance exceeds the threshold;

[0032] D3: Obtain the preset interception specifications to establish a matrix, and fill in the relevant asset features that have passed the distance verification in several group matrix positions to obtain a search matrix;

[0033] D4: Search the asset matrix corresponding to the asset features in the asset library by template matching through the search matrix, obtain the matching asset matrix, and go to step S3.

[0034] When the present invention is working, a search matrix obtained by filling in a plurality of groups of matrix positions is used to search in the asset matrix. This can effectively mine the user's hidden preferences while ensuring that the intercepted asset matrix is highly consistent with the current user, without occupying additional computing resources, with high calling efficiency and good push effect.

[0035] Preferably, in the step D3, when obtaining a preset intercept specification to establish a matrix and filling relevant asset features that pass the distance verification in several group matrix positions to obtain a retrieval matrix, the following steps are adopted: obtaining a preset intercept specification to establish a matrix, traversing all elements according to several groups of matrix positions, filling appropriate asset features in sequence and ensuring that the corresponding asset features meet the distance constraints of each group of matrix positions, and repeatedly obtaining the retrieval matrix in sequence.

[0036] To solve the above technical problems, the second aspect of the present invention adopts the following technical solution: An automatic interaction system based on a retail large model, which applies an automatic interaction method based on a retail large model as described above. The automatic interaction system includes a processor and a memory. At least one instruction, at least one program, a code set or an instruction set is stored in the memory. The instruction, program, code set or instruction set is loaded and executed by the processor to perform an automatic interaction method based on a retail large model as described above.

[0037] The beneficial technical effects of the present invention include:

[0038] 1. The present invention can automatically optimize and push highly compatible assets for users according to user portraits and user behaviors through a retail large model, can cyclically obtain the assets preferred by users during the pushing process, and establish a retrieval matrix based on the asset features of the preferred assets to retrieve in the asset library, so that the retail large model only needs to analyze users, reducing the process of analyzing commodities. Only through feature matching can the scale of data to be called during commodity pushing be greatly reduced. While greatly reducing the occupied computing resources, it can also improve the data security coefficient of the asset library.

[0039] 2. The present invention can realize the storage of assets. By splitting asset features and storing assets through category division, the association operations between several assets are reduced. When the scale of assets is large, the data mapping relationship table can be saved, the pressure of data storage can be reduced to a certain extent, and it is also convenient to update the asset library when adding or deleting assets, optimizing the work process. By establishing several asset matrices, the hidden association information between different assets can also be mined, which can improve the conversion rate of asset pushing while ensuring the compatibility of asset pushing, and does not require additional computing resources, with high economic benefits.

[0040] 3. The present invention determines the associated pushing direction in the asset matrix through a pushing algorithm, realizes the sequential pushing of several assets in the asset matrix, can correct the pushing path in real time according to the changing user preferences, can further improve the effect of asset pushing and the conversion rate, and can also realize the correction of the final pushing target, can realize the further improvement of the user portrait, and complete the accurate portrayal of the user portrait through multiple cyclic pushings, thereby assisting the retail large model to push more compatible assets to users.

[0041] 4. The present invention uses a retrieval matrix obtained by filling several groups of matrix positions to retrieve in the asset matrix, which can effectively mine the hidden preferences of users while ensuring that the intercepted asset matrix highly matches the current user, without occupying additional computing power resources, with high call efficiency and good push effect.

[0042] Other features and advantages of the present invention will be disclosed in detail in the following specific embodiments and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The present invention will be further described below with reference to the accompanying drawings:

[0044] Figure 1 is a working flowchart of an automatic interaction method based on a retail large model;

[0045] Figure 2 is a working flowchart of establishing an asset library;

[0046] Figure 3 is a working flowchart of step S3 in an automatic interaction method based on a retail large model;

[0047] Figure 4 is a working flowchart of step S4 in an automatic interaction method based on a retail large model. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0048] The technical solutions of the embodiments of the present invention will be explained and described below with reference to the accompanying drawings of the embodiments of the present invention. However, the following embodiments are only the preferred embodiments of the present invention and not all of them. Based on the embodiments in the embodiments, other embodiments obtained by those skilled in the art without creative efforts fall within the protection scope of the present invention.

[0049] In the following description, terms such as "inner", "outer", "upper", "lower", "left", "right", etc. indicating directions or positional relationships are only for convenient description of the embodiments and simplification of the description, rather than indicating or implying that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore should not be construed as a limitation of the present invention. Embodiment 1:

[0050] Please refer to Figure 1 , this embodiment discloses an automatic interaction method based on a retail large model, including the following steps:

[0051] S1: Access the user, obtain the historical user profile, and record the user behavior in real time;

[0052] S2: Call the preset retail large model to obtain the matching asset matrix in the preset asset library according to the user profile and user behavior;

[0053] S3: According to the fit degree between the user and the assets, several assets in the asset matrix are sequentially associated and pushed. Use the retail large model to judge the assets preferred by the user based on the user behavior records, and generate a list of preferred assets in order according to the preference degree;

[0054] S4: Obtain several assets in the list of preferred assets, divide and establish a retrieval matrix according to the preset sorting rules of the selected asset characteristics, retrieve in the asset matrix corresponding to the asset characteristics in the asset library through the retrieval matrix, obtain the matching asset matrix, and transfer to step S3.

[0055] When this example works, it can automatically optimize and push assets with high fit degree for the user according to the user portrait and user behavior through the retail large model, and can cyclically obtain the assets preferred by the user during the pushing process. At the same time, a retrieval matrix is established according to the asset characteristics of the preferred assets to retrieve in the asset library, so that the retail large model only needs to analyze the user, reducing the process of analyzing commodities. Only through feature matching can the scale of data to be called when pushing commodities be greatly reduced. While greatly reducing the occupied computing power resources, it can also improve the data security coefficient of the asset library.

[0056] In this embodiment, in the step S1, when accessing the user, obtaining the historical user portrait, and recording the user behavior in real time, the following steps are adopted: access the user, obtain the historical user portrait, and record the user behavior in real time when the user browses. The recorded user behavior includes at least one of user search information, user browsing information, user page stay time, and user click record. For example, when the user browses the material interface, process interface, effect interface, etc. of the assets, the preferred asset characteristics and their corresponding types of the user can be captured by recording this part of the user behavior.

[0057] In specific implementation, in the step S2, when calling the preset retail large model to obtain the matching asset matrix in the preset asset library according to the user portrait and user behavior, the following steps are adopted:

[0058] A1: Obtain the user portrait and user behavior of the user, call the preset retail large model to split several asset characteristics related to the user according to the user portrait and user behavior of the user, dynamically weight the several asset characteristics according to the current scenario, and then use the method of aggregation statistics to obtain the statistical results of the several asset characteristics and arrange and store them to obtain the set of asset characteristics preferred by the user;

[0059] A2: Locate the matching assets in the asset matrix corresponding to the selected asset characteristic type in the preset asset library through the set of asset characteristics, and intercept the asset matrix with an appropriate data scale in the asset matrix according to the preset interception specifications for the matching assets.

[0060] Preferably, when dynamically assigning weights, the weights of each asset feature can be dynamically planned based on user behavior. At this time, the hidden preferences of each user can be mined through the retail big model based on the common preferences of similar customers, which can further improve the asset push effect.

[0061] See also Figure 2 As a further improvement of this embodiment, the asset library is established by the following steps:

[0062] B1: Determine the associated hierarchical distribution of several asset features, establish several source nodes, each source node is filled with asset features of the corresponding level as several first child nodes, and the several first child nodes are arranged in sequence into row vectors according to a preset sorting rule;

[0063] B2: Each first sub-node is filled with a number of second sub-nodes whose asset features are of a corresponding level, and the second sub-nodes are sequentially arranged into a column vector according to a preset sorting rule;

[0064] B3: Call the retail model to split each asset into several asset features, add the index of the asset to each asset feature, and configure the several asset features as matching second child nodes;

[0065] B4: Traverse several row vectors of several source nodes to form the elements of the first row of several asset matrices respectively, and fill the corresponding column vectors to form elements of several columns respectively, so as to generate several asset matrices and store them as asset libraries.

[0066] When this instance is working, it can realize the storage of assets. By splitting asset features and storing assets through category classification, it reduces the association operations between several assets, can save data mapping relationship tables when the asset scale is large, can reduce the pressure of data storage to a certain extent, and also facilitates updating the asset library when adding or reducing assets, optimizes the workflow, and can also mine hidden association information between different assets by establishing several asset matrices, which can improve the conversion rate of asset push while ensuring the fit of asset push, and does not require additional computing power resources, with high economic benefits.

[0067] In specific implementation, the sorting rule can be appropriately divided according to the type of each asset feature. For non-numerical features, manual or big data analysis is required to determine the relevance and then sort them sequentially. For numerical features, they can be sorted sequentially through numerical sorting. When sorting, interval sorting needs to be performed according to the preset distribution gradient, and the positions without data are set to zero, which can ensure the correlation of similar asset features in the asset matrix. Moreover, compared with manually adding mapping relationships or automatically adding them by an AI model, it not only improves the correlation efficiency but also facilitates the discovery of hidden correlation information between assets, further improving the fit of asset push.

[0068] Please refer to Figure 4 Preferably, in step S4, several assets in the preferred asset list are obtained, and a retrieval matrix is established by dividing according to the preset sorting rule of the selected asset features. The retrieval matrix is used to retrieve in the asset matrix corresponding to the asset features in the asset library to obtain a matching asset matrix. When transferring to step S3, the following steps are adopted:

[0069] D1: Obtain several assets in the preferred asset list, cluster their asset features, and determine at least one identical asset feature and at least two similar asset features;

[0070] D2: According to the similar asset features of the selected assets, determine their matrix positions respectively according to the preset sorting rule, obtain several groups of matrix positions, calculate the distances of each matrix position in each group of matrix positions and verify them, and eliminate the relevant asset features whose distances exceed the threshold;

[0071] D3: Obtain the preset truncation specification to establish a matrix, and fill in the relevant asset features that pass the distance verification in several groups of matrix positions to obtain a retrieval matrix;

[0072] D4: Use the retrieval matrix to retrieve in the asset matrix corresponding to the asset features in the asset library through template matching to obtain a matching asset matrix, and transfer to step S3.

[0073] When this example works, a retrieval matrix filled by several groups of matrix positions is used to retrieve in the asset matrix, which can ensure that the intercepted asset matrix highly fits the current user while effectively discovering the hidden preferences of the user, and does not require additional computing resources, has high call efficiency, and good push effect.

[0074] As a further improvement of this embodiment, in step D2, when calculating the distances of each matrix position in each group of matrix positions, the Chebyshev distance calculation operation can be used to calculate the maximum distance of the corresponding matrix in the asset matrix under different sorting rules. In actual work, the row coordinates of two matrix positions generally have a small deviation, and their column coordinates have a large deviation. Using the Chebyshev distance calculation operation can reduce the calculation pressure. Of course, any other suitable calculation method can also be used to meet the distance calculation requirements of this embodiment.

[0075] Preferably, in step D3, when obtaining a preset truncation specification to establish a matrix and filling the relevant asset features with qualified distance verification in several group matrix positions to obtain a retrieval matrix, the following steps are adopted: obtaining a preset truncation specification to establish a matrix, traversing all elements according to several groups of matrix positions, filling in appropriate asset features in sequence and ensuring that the corresponding asset features meet the distance constraints of each group of matrix positions, and repeatedly obtaining the retrieval matrix in sequence. Preferably, several groups of matrix positions can also be screened to control the filling quantity, which can improve the template matching efficiency while ensuring the matching efficiency and reduce the occupation of computing power resources. Embodiment 2:

[0076] Please refer to Figure 3 , this embodiment provides an automatic interaction method based on a retail large model. The same parts as other embodiments will not be elaborated here, and the differences will be described in detail below.

[0077] In this embodiment, in step S3, according to the degree of fit between the user and the assets, several assets in the asset matrix are sequentially associated and pushed. When using the retail large model to judge the assets preferred by the user according to the user behavior records and generating a preferred asset list in order of preference degree, the following steps are adopted: positioning the matching assets in the matching asset matrix through the retail large model according to the user's preference feature set, and at the same time predicting at least one predicted preferred asset according to the user's preference feature set through the retail large model. Determining the associated push direction in the asset matrix through a preset push algorithm, and sequentially pushing several assets in the asset matrix. Using the retail large model to judge the assets preferred by the user according to the user behavior records and generating a preferred asset list in order of preference degree.

[0078] In specific implementation, in step S3, when determining the associated push direction in the asset matrix through a preset push algorithm and sequentially pushing several assets in the asset matrix, the following steps are adopted:

[0079] C1: Obtain the asset features corresponding to the matching assets, predict the position of the preferred asset in the current asset matrix, calculate the initial vector from the matching assets to the predicted preferred asset, calculate the unit vector, and calculate the position of the next preferred asset according to the unit vector and the preset step size;

[0080] C2: Extract the matching asset matrix according to the position of the next preferred asset through the preset extraction specification. When the matching assets in the current asset matrix are all pushed, go to step C3;

[0081] C3: When there is a new preferred asset in the current asset matrix, go to step C4; otherwise, go to step C5;

[0082] C4: Calculate the unit vector from the current preferred asset to the preset preferred asset and the unit vector from the current preferred asset to the new preferred asset. Obtain the corrected unit vector through vector addition. Calculate the position of the next preferred asset according to the unit vector and the preset step size. Determine whether the position of the next preferred asset is the predicted position or whether the number of loops reaches the preset threshold. If so, end and record all preferred assets to generate a preferred asset list, and go to step S4; otherwise, go to step C2;

[0083] C5: Dynamically adjust the step size according to the number of loops. Calculate the position of the next preferred asset according to the unit vector and the accumulated step size. Determine whether the position of the next preferred asset is the predicted position or whether the number of loops reaches the preset threshold. If so, end and record all preferred assets to generate a preferred asset list, and go to step S4; otherwise, go to step C2.

[0084] When this example works, the associated push direction is determined in the asset matrix through the push algorithm to realize the sequential push of several assets in the asset matrix. It can correct the push path in real time according to the user's preferences, further improve the effect of asset push, increase the conversion rate, and at the same time can also correct the final push target, realize the further improvement of the user portrait, and complete the accurate portrayal of the user portrait through multiple loop pushes, so as to assist the retail large model to push assets with higher fit to users.

[0085] As a further improvement of this embodiment, in step C1, when obtaining the asset features corresponding to the matching assets, predicting the position of the preferred asset in the current asset matrix, calculating the initial vector from the matching assets to the predicted preferred asset, calculating the unit vector, and calculating the position of the next preferred asset according to the unit vector and the preset step size, the following steps are adopted. The position of the asset features corresponding to the matching assets is , the position of the predicted preferred asset is , when calculating the initial vector from the matching assets to the predicted preferred asset, the following formula is used:

[0086] ;

[0087] Wherein: is the initial vector from the matching asset to the predicted preference asset;

[0088] When calculating the unit vector from the matching asset to the predicted preference asset, the following formula is used:

[0089] ;

[0090] Wherein: is the unit vector from the matching asset to the predicted preference asset;

[0091] When calculating the position of the next preference asset, the following formula is used:

[0092] ;

[0093] ;

[0094] Wherein: is the actual step length from the current asset to the next preference asset;

[0095] is the rounding operation after step length calculation;

[0096] s is a preset step length, and the step length is set to be adjustable;

[0097] is the position of the next preference asset;

[0098] In the step C4, when calculating the unit vector from the current preference asset to the preset preference asset and the unit vector from the current preference asset to the new preference asset, the following formula is used:

[0099] ;

[0100] ;

[0101] Wherein: is the unit vector from the current preference asset to the preset preference asset;

[0102] is the unit vector from the current preference asset to the new preference asset;

[0103] When obtaining the corrected unit vector through vector addition, the following formula is used:

[0104] ;

[0105] ;

[0106] Wherein: is the correction vector obtained by performing vector addition on the unit vector from the current preferred asset to the preset preferred asset and the unit vector from the current preferred asset to the new preferred asset;

[0107] is the unit vector of the correction vector;

[0108] is to calculate the vector length through the Euclidean norm;

[0109] When calculating the position of the next preferred asset, the following formula is used:

[0110] ;

[0111] ;

[0112] In the step C5, when dynamically adjusting the step size according to the number of loops and calculating the position of the next preferred asset according to the unit vector and the accumulated step size, the following formula is used:

[0113] ;

[0114] ;

[0115] Wherein: is the accumulated step size.

[0116] Preferably, in the step C2, the following steps are further included. When intercepting a matching asset matrix according to the position of the next preferred asset through a preset interception specification, the interception path and each intercepted asset matrix are synchronously recorded to obtain a plurality of optimized push matrices. The plurality of optimized push matrices are processed by the retail large model to obtain a plurality of asset push paths for reducing the consumption of computing resources. In specific implementation, the predicted preferred asset can be set as an asset with high economic benefits. At this time, the user's preference can be non-invasively influenced, so that in multiple asset browsing, it is considered that the various asset characteristics of the asset with high economic benefits are in the best combination state, thereby increasing the preference degree for the asset with high economic benefits, and further increasing its conversion rate and economic benefits. Embodiment 3:

[0117] This embodiment provides an automatic interaction system based on a retail large model, which applies an automatic interaction method based on a retail large model as described in the above embodiment. The automatic interaction system includes a processor and a memory. At least one instruction, at least one program, a code set or an instruction set is stored in the memory. The instruction, program, code set or instruction set is loaded and executed by the processor to perform an automatic interaction method based on a retail large model as described in the above embodiment.

[0118] The beneficial technical effects of this embodiment include: The present invention can automatically optimize and push highly compatible assets for users according to user portraits and user behaviors through the retail large model, and can cyclically obtain the assets preferred by users during the pushing process. At the same time, a retrieval matrix is established based on the asset characteristics of the preferred assets to conduct retrieval in the asset library, enabling the retail large model to only analyze users, reducing the process of analyzing commodities, and only by feature matching, the scale of data to be called when pushing commodities can be greatly reduced. While greatly reducing the occupied computing power resources, the data security factor of the asset library can also be improved.

[0119] The above is only the specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Those skilled in the art should understand that the present invention includes but is not limited to the content described in the drawings and the above specific embodiments. Any modification that does not deviate from the functional and structural principles of the present invention will be included in the scope of the claims.

Claims

1. An automatic interaction method based on a retail large model, characterized in that It includes the following steps: S1: Access the user, obtain the historical user profile, and record the user behavior in real time; S2: Call the preset retail large model to obtain the matching asset matrix in the preset asset library according to the user profile and user behavior; S3: According to the fit degree between the user and the assets, sequentially associate and push several assets in the asset matrix, use the retail large model to judge the assets preferred by the user according to the user behavior record, and generate a preferred asset list in order according to the preference degree; S4: Obtain several assets in the preferred asset list, establish a retrieval matrix according to the preset sorting rules for the selected asset characteristics, retrieve in the asset matrix corresponding to the asset characteristics in the asset library through the retrieval matrix, obtain the matching asset matrix, and transfer to step S3; In step S4, when obtaining several assets in the preferred asset list, establishing a retrieval matrix according to the preset sorting rules for the selected asset characteristics, retrieving in the asset matrix corresponding to the asset characteristics in the asset library through the retrieval matrix, and obtaining the matching asset matrix and transferring to step S3, the following steps are adopted: D1: Obtain several assets in the preferred asset list, cluster their asset characteristics, and determine at least one identical asset characteristic and at least two similar asset characteristics; D2: Determine their matrix positions respectively according to the similar asset characteristics of the selected assets according to the preset sorting rules, obtain several groups of matrix positions, calculate the distances of each matrix position in each group of matrix positions respectively and verify, and eliminate the relevant asset characteristics whose distances exceed the threshold; D3: Obtain the preset truncation specification to establish a matrix, and fill in the relevant asset characteristics that pass the distance verification in several groups of matrix positions to obtain a retrieval matrix; In step D3, when obtaining the preset truncation specification to establish a matrix and filling in the relevant asset characteristics that pass the distance verification in several groups of matrix positions to obtain a retrieval matrix, the following steps are adopted: Obtain the preset truncation specification to establish a matrix, traverse all elements according to several groups of matrix positions, fill in the appropriate asset characteristics in turn and ensure that the corresponding asset characteristics meet the distance constraints of each group of matrix positions, and repeatedly obtain the retrieval matrix in turn; D4: Retrieve through the retrieval matrix in the asset matrix corresponding to the asset characteristics in the asset library by means of template matching, obtain the matching asset matrix, and transfer to step S3.

2. The automatic interaction method based on a retail large model according to claim 1, wherein: In step S1, when accessing the user, obtaining the historical user profile, and recording the user behavior in real time, the following steps are adopted: Access the user, obtain the historical user profile, and record the user behavior in real time when the user browses. The recorded user behavior includes at least one of user search information, user browsing information, user page stay time, and user click record.

3. The automatic interaction method based on a retail large model according to claim 1, characterized in that: In step S2, when calling the preset retail large model to obtain the matching asset matrix in the preset asset library according to the user profile and user behavior, the following steps are adopted: A1: Obtain the user profile and user behavior of the user, call a preset retail large model to split a number of asset features related to the user according to the user profile and user behavior of the user, after dynamically weighting a number of asset features according to the current scenario, use the method of aggregation statistics to obtain the statistical results of a number of asset features and arrange and store them to obtain the set of asset features preferred by the user; A2: Locate the matching assets in the asset matrix corresponding to the selected asset feature type in the preset asset library through the set of asset features, and intercept the asset matrix with an appropriate data scale in the asset matrix according to the preset intercept specification through the matching assets.

4. The automatic interaction method based on a retail large model according to claim 1, characterized in that: The establishment of the asset library adopts the following steps: B1: Determine the associated hierarchical distribution of a number of asset features, establish a number of source nodes, and fill each source node with the asset features of the corresponding level as a number of first child nodes, and a number of first child nodes are arranged in a row vector in sequence according to a preset sorting rule; B2: Fill each first child node with the asset features of the corresponding level as a number of second child nodes, and a number of second child nodes are arranged in a column vector in sequence according to a preset sorting rule; B3: Call the retail large model to split a number of asset features of each asset respectively, add the index of the asset to each asset feature, and then configure a number of asset features as matching second child nodes; B4: Traverse the row vectors of a number of source nodes to form the elements of the first row of a number of asset matrices respectively, and fill the corresponding column vectors to form the elements of several columns respectively, so as to generate a number of asset matrices and store them as an asset library.

5. The automatic interaction method based on a retail large model according to claim 4, characterized in that: In step S3, according to the fit degree between the user and the assets, a number of assets in the asset matrix are associated and pushed in sequence. When using the retail large model to judge the assets preferred by the user according to the user behavior record and generating a list of preferred assets in order according to the preference degree, the following steps are adopted. Locate the matching assets in the matching asset matrix through the set of preference features of the user by the retail large model, and at the same time predict at least one predicted preferred asset according to the set of preference features of the user by the retail large model. Determine the associated push direction in the asset matrix through a preset push algorithm, and push a number of assets in the asset matrix in sequence. Use the retail large model to judge the assets preferred by the user according to the user behavior record and generate a list of preferred assets in order according to the preference degree.

6. The automatic interaction method based on a retail large model according to claim 5, wherein: In step S3, when determining the associated push direction in the asset matrix through a preset push algorithm and pushing a number of assets in the asset matrix in sequence, the following steps are adopted: C1: Obtain the asset features corresponding to the matching assets and the position of the predicted preferred asset in the current asset matrix, calculate the initial vector from the matching assets to the predicted preferred asset, and calculate the unit vector, and calculate the position of the next preferred asset according to the unit vector and the preset step size; C2: Intercept the matching asset matrix according to the position of the next preferred asset through the preset intercept specification. When all the matching assets in the current asset matrix have been pushed, go to step C3; C3: When there is a new preferred asset in the current asset matrix, go to step C4; otherwise, go to step C5. C4: Calculate the unit vector from the current preferred asset to the preset preferred asset and the unit vector from the current preferred asset to the new preferred asset. Obtain the corrected unit vector through vector addition. Calculate the position of the next preferred asset based on the unit vector and the preset step size. Determine whether the position of the next preferred asset is the predicted position or whether the number of loops has reached the preset threshold. If so, end and record all preferred assets to generate a preferred asset list, and go to step S4; otherwise, go to step C2. C5: Dynamically adjust the step size according to the number of loops. Calculate the position of the next preferred asset based on the unit vector and the accumulated step size. Determine whether the position of the next preferred asset is the predicted position or whether the number of loops has reached the preset threshold. If so, end and record all preferred assets to generate a preferred asset list, and go to step S4; otherwise, go to step C2.

7. The automatic interaction method based on a retail large model according to claim 6, characterized in that: In step C2, the following steps are further included. When intercepting a matching asset matrix according to the position of the next preferred asset through a preset intercept specification, synchronously record the intercept path and each intercepted asset matrix to obtain a number of optimized push matrices. Process the number of optimized push matrices through the retail large model to obtain a number of asset push paths for reducing the consumption of computing resources.

8. An automatic interaction system based on a retail large model, applying an automatic interaction method based on a retail large model according to any one of claims 1 to 7, characterized in that: The automatic interaction system includes a processor and a memory. At least one instruction, at least one segment of program, code set or instruction set is stored in the memory. The instruction, program, code set or instruction set is loaded and executed by the processor to perform an automatic interaction method based on a retail large model according to any one of claims 1 to 7.

Images