CN115374186A

CN115374186A - Data processing method and AI system based on big data

Info

Publication number: CN115374186A
Application number: CN202211200521.2A
Authority: CN
Inventors: 李烜宇; 韩叶刚
Original assignee: Individual
Current assignee: SHANGHAI COMPASS INFORMATION SCIENCE CO Ltd
Priority date: 2022-09-29
Filing date: 2022-09-29
Publication date: 2022-11-22
Anticipated expiration: 2042-09-29
Also published as: CN115374186B

Abstract

The invention relates to the technical field of data processing, in particular to a data processing method and an AI system based on big data. According to the big data-based data processing method and the AI system, the thought of the interest focus decision vector corresponding to each conversation behavior knowledge field is determined according to the element contribution value, so that the occupation of processing overhead is reduced, and the timeliness of user interest mining is guaranteed; based on the method, the resource consumption is reduced when the complex electric business interaction knowledge chain is dealt with, the time consumption is reduced when the complex electric business interaction knowledge chain is dealt with, and therefore the user interest mining efficiency aiming at the complex electric business interaction knowledge chain is ensured.

Description

Data processing method and AI system based on big data

Technical Field

The invention relates to the technical field of data processing, in particular to a data processing method and an AI system based on big data.

Background

"panning" from big data, mining implicit, unknown, potential relationships, models, and trends to decision-making from large amounts of data information, and using these knowledge and rules to build models for decision-making support, methods, tools, and processes that provide predictive decision-making support, are big data mining. Taking e-commerce services as an example, big data mining comprises methods of utilizing association rules, cluster analysis, classification and prediction, time sequence patterns, deviation detection, intelligent recommendation and the like, so that an e-commerce service party is helped to extract data asset values contained in data, and the intelligence degree and pertinence of the e-commerce services are improved. Now, with the continuous development of e-commerce services, it is difficult for the traditional big data mining technology to efficiently implement data mining analysis of e-commerce users, and therefore a more optimized scheme is urgently needed to deal with the complex mining of e-commerce service big data.

Disclosure of Invention

In order to solve the technical problems in the related art, the invention provides a data processing method and an AI system based on big data.

In a first aspect, an embodiment of the present invention provides a data processing method based on big data, which is applied to an AI system, and the method includes: determining a first element contribution value of each original index element in a first element relationship network corresponding to an electric business interaction knowledge chain and a second element contribution value of each original pairing element in a second element relationship network corresponding to the electric business interaction knowledge chain according to the number of conversation behavior knowledge fields in the electric business interaction knowledge chain to be mined; determining a common value between each original index element and each original pairing element respectively by combining the first element contribution value and the second element contribution value; for each conversation behavior knowledge field in the electric business service interaction knowledge chain, determining an interest focusing decision vector corresponding to the conversation behavior knowledge field by combining a common value between an original index element corresponding to the conversation behavior knowledge field and each original pairing element and each target bias element in a third element relation network corresponding to the electric business service interaction knowledge chain; and determining user interest mining information corresponding to the electric business interaction knowledge chain by combining the interest focusing decision vectors corresponding to the conversation behavior knowledge fields in the electric business interaction knowledge chain.

According to the design, common values between each original index element and each original pairing element are determined by means of the determined first element contribution value corresponding to each original index element and the determined second element contribution value corresponding to each original pairing element, and then interest focusing decision vectors corresponding to each conversation behavior knowledge field are determined according to each common value and each target bias element in a third element relation network; compared with the method that a plurality of focusing decision relationship networks are directly integrated to obtain an element contribution relationship network, and then the element contribution relationship network is used for carrying out quantitative operation on the conversation behavior knowledge fields to obtain interest focusing decision vectors corresponding to the conversation behavior knowledge fields, the embodiment of the invention determines the thought of the interest focusing decision vectors corresponding to the conversation behavior knowledge fields according to the element contribution values, so that the occupation of processing cost is reduced, and the timeliness of user interest mining is guaranteed. Therefore, the resource consumption is reduced when the complex electric business interaction knowledge chain is dealt with, the time consumption for dealing with the complex electric business interaction knowledge chain is also reduced, and the user interest mining efficiency aiming at the complex electric business interaction knowledge chain is further ensured.

In some exemplary embodiments, said determining a commonality value between each of said original index elements and each of said original pair elements, in combination with said first element contribution value and said second element contribution value, comprises: adjusting each original index element into a target index element meeting the setting requirement and adjusting each original pairing element into a target pairing element meeting the setting requirement by any one target AI algorithm unit in a plurality of set AI algorithm units; determining a first element operation result corresponding to each original pairing element according to a second element contribution value corresponding to each original pairing element and a target pairing element corresponding to the original pairing element; for each original index element, determining a second element operation result corresponding to the original index element by combining a target index element corresponding to the original index element and a first element contribution value corresponding to the original index element; and determining a common value between the original index element and each original pairing element by combining the second element operation result and the first element operation result corresponding to each original pairing element.

By means of the design, the original index elements are adjusted into the target index elements meeting the setting requirements and the original pairing elements are adjusted into the target pairing elements meeting the setting requirements by means of the target AI algorithm unit, the problem that the precision of the obtained common value is interfered by directly processing the original index elements with the previous quantized value smaller than 0 and/or the original pairing elements with the previous quantized value smaller than 0 is avoided, and the precision and the reliability of the common value are ensured.

In some exemplary embodiments, the determining, by combining the second element operation result and the first element operation result corresponding to each original paired element, a commonality value between each original paired element and each original indexed element comprises: determining a first eccentricity coefficient corresponding to the original index elements by combining the first sequence priority of the original index elements in the first element relationship network and the number of the conversation behavior knowledge fields, and determining a second eccentricity coefficient corresponding to each original pairing element according to the second sequence priority of each original pairing element in the second element relationship network; combining the first eccentricity coefficient and a target index element corresponding to the original index element to determine a third element operation result corresponding to the original index element; for each original pairing element, determining a fourth element operation result corresponding to the original pairing element by combining a second eccentricity coefficient corresponding to the original pairing element and a target pairing element corresponding to the original pairing element; and determining a common value between the original index element and each original pairing element by combining the second element operation result, the third element operation result, and the first element operation result and the fourth element operation result corresponding to each original pairing element.

By means of the design, the common value between each original index element and each original pairing element is determined by means of the determined first eccentricity coefficient corresponding to each original index element and the determined second eccentricity coefficient corresponding to each original pairing element, and the accuracy of the determined common value can be further guaranteed.

In some exemplary embodiments, the determining, by combining the common value between the original index element corresponding to the knowledge field of session behavior and each original pairing element, and each target bias element in the third element relationship network corresponding to the chain of electrical business interaction knowledge, an interest focusing decision vector corresponding to the knowledge field of session behavior includes: for each original pairing element, determining first feature calculation information corresponding to the original pairing element by combining a first element operation result corresponding to the original pairing element and a target bias element corresponding to the third element relation network and having the same third sequence priority as a second sequence priority corresponding to the original pairing element; determining second feature calculation information corresponding to the original paired elements by combining a fourth element operation result corresponding to the original paired elements and a target offset element corresponding to the third element relation network and having the same third sequence priority as the second sequence priority corresponding to the original paired elements; multiplying a second element operation result corresponding to the original index element with first feature calculation information corresponding to each original pairing element to obtain a fifth element operation result; multiplying a third element operation result corresponding to the original index element with second feature calculation information corresponding to each original pairing element to obtain a sixth element operation result; and determining an interest focusing decision vector corresponding to the conversation behavior knowledge field by combining summation data of common values between the original index elements corresponding to the conversation behavior knowledge field and the original paired elements, summation data of the operation result of the fifth element and summation data of the operation result of the sixth element.

By means of the design, the target bias elements matched with the original pairing elements can be found in the third element relation network through the third sequence priority and the second sequence priority, and therefore the element characteristic binary groups are determined. According to the element operation result corresponding to the original pairing element in the element characteristic binary group and the target pairing element of the element characteristic binary group, the first characteristic calculation information and the second characteristic calculation information corresponding to each element characteristic binary group can be accurately obtained. Furthermore, according to the accurate first feature calculation information and the second feature calculation information, an accurate interest focusing decision vector can be determined.

In some exemplary embodiments, before determining, according to the number of the session behavior knowledge fields in the electric business interaction knowledge chain to be mined, a first element contribution value of each original index element in a first element relationship network corresponding to the electric business interaction knowledge chain, the method further includes: acquiring the electric business interaction knowledge chain to be mined, and performing knowledge transformation processing on the electric business interaction knowledge chain to obtain the transformed electric business interaction knowledge chain; performing chain level sorting on the electric business interaction knowledge chain and the electric business interaction knowledge chain which is transformed to obtain the electric business interaction knowledge chain which is finished with sorting; respectively carrying out knowledge projection operation on the electric business interaction knowledge chain which is finished with arrangement by means of a target vector relationship network to obtain a first element relationship network, a second element relationship network and a third element relationship network which correspond to the electric business interaction knowledge chain; the first element relation network, the second element relation network and the third element relation network correspond to different target vector relation networks respectively.

By means of the design, the electric business interaction knowledge chain to be mined can be transformed on the expert knowledge level through knowledge transformation processing, the electric business interaction knowledge chain which is located on different expert knowledge levels and completes transformation is obtained, then the electric business interaction knowledge chain and the electric business interaction knowledge chain which completes transformation are subjected to chain level sorting, and the electric business interaction knowledge chain which sorts more detailed knowledge and completes sorting can be obtained. And performing knowledge projection operation by means of the organized e-commerce business interaction knowledge chain, so that the diversity/comprehensiveness of the element characteristics covered by each of the first element relationship network, the second element relationship network and the third element relationship network is improved.

In some exemplary embodiments, the determining, in combination with the interest focus decision vector corresponding to each of the session behavior knowledge fields in the electronic commerce interaction knowledge chain, user interest mining information corresponding to the electronic commerce interaction knowledge chain includes: sorting all conversation behavior knowledge fields in the sorted e-commerce business interaction knowledge chain and interest focusing decision vectors corresponding to the conversation behavior knowledge fields to obtain all the sorted interest focusing decision vectors; respectively carrying out moving average processing on each interest focusing decision vector which is finished with moving average to obtain each interest focusing decision vector which is finished with moving average, and carrying out vector sorting on each interest focusing decision vector which is finished with moving average and the interest focusing decision vector which is finished with sorting and corresponds to the interest focusing decision vector to obtain an original interest focusing decision vector which corresponds to each interest focusing decision vector which is finished with moving average; performing knowledge transformation processing on each original interest focusing decision vector to obtain each transformed original interest focusing decision vector, and performing vector sorting on each original interest focusing decision vector and each transformed original interest focusing decision vector corresponding to the original interest focusing decision vector to obtain each sorted original interest focusing decision vector; and performing numerical value extrusion processing on each original interest focusing decision vector which is finished to be sorted to obtain to-be-processed interest focusing decision vectors corresponding to each conversation behavior knowledge field, and determining user interest mining information corresponding to the e-commerce service interaction knowledge chain according to each to-be-processed interest focusing decision vector.

By means of the design, the interest focus decision vectors to be processed, which have various input contents and can reflect the user interest mining information more accurately, can be obtained by performing sorting operation, sliding average processing, knowledge conversion processing and numerical value extrusion processing on the interest focus decision vectors, and then the user interest mining information can be determined accurately according to each accurate interest focus decision vector to be processed.

In some exemplary embodiments, the determining user interest mining information corresponding to the e-commerce business interaction knowledge chain according to each to-be-processed interest focus decision vector includes: determining a set formed by the to-be-processed interest focusing decision vectors as a derived to-be-mined electrical business interaction knowledge chain, and taking the to-be-processed interest focusing decision vectors as each session behavior knowledge field in the derived to-be-mined electrical business interaction knowledge chain; skipping to the step of carrying out knowledge transformation processing on the electricity business interaction knowledge chain to obtain the transformed electricity business interaction knowledge chain until the skip accumulated value reaches the set accumulated value to obtain each target interest focusing decision vector corresponding to the electricity business interaction knowledge chain to be mined; and determining the user interest mining information according to each target interest focusing decision vector.

By the design, based on iterative processing, the features reflected by the knowledge fields of the conversation behaviors can be deeply and completely mined, so that a target interest focusing decision vector containing diversified features is obtained. And mining the user interest by using each target interest focusing decision vector, so that the accuracy of the determined user interest mining information can be guaranteed.

In some exemplary embodiments, the determining the user interest mining information according to each of the target interest focusing decision vectors includes: performing decision vector translation on each target interest focusing decision vector by means of at least one set interest analysis algorithm to obtain the user interest mining information; the raw material of the downstream interest analysis algorithm in the two interest analysis algorithms with continuous translation priority is the analysis result of the upstream interest analysis algorithm, and the analysis result of the interest analysis algorithm with the translation priority at the tail is the user interest mining information.

By means of the design, the feature translation is performed by means of the interest analysis algorithms, and the sufficient feature translation of the features reflected by each target interest focusing decision vector can be achieved, so that accurate user interest mining information is obtained.

In some exemplary embodiments, the big data based data processing method is performed by means of an AI machine learning model that is debugged in advance; the AI machine learning model is debugged based on the following manner: obtaining an e-commerce business interaction knowledge chain example to be mined; loading the electric business interaction knowledge chain example to an AI machine learning model to be debugged, processing the electric business interaction knowledge chain example by means of the AI machine learning model to be debugged, and determining each target interest focusing regression vector corresponding to the electric business interaction knowledge chain example; determining user interest regression information corresponding to the electricity business interaction knowledge chain example according to each target interest focusing regression vector; and determining regression analysis cost corresponding to the AI machine learning model to be debugged by combining the user interest regression information and the user interest prior information corresponding to the electric business service interaction knowledge chain example, and performing cyclic debugging on the AI machine learning model to be debugged by means of the regression analysis cost until the debugging termination requirement is met to obtain the debugged AI machine learning model.

By the design, the adjusted AI machine learning model has better interest mining performance, and the adjusted AI machine learning model is used for mining the user interest, so that accurate user interest mining information can be obtained. By means of the regression analysis cost determined according to the user interest regression information and the user interest prior information, the to-be-debugged AI machine learning model is subjected to cyclic debugging, the regression analysis accuracy of the AI machine learning model can be effectively improved, and the high-precision and high-reliability AI machine learning model can be obtained through debugging.

In some exemplary embodiments, the user interest regression information comprises first user interest regression information, the regression analysis cost comprises a decision discrimination cost; determining user interest regression information corresponding to the e-commerce business interaction knowledge chain example according to each target interest focusing regression vector, wherein the determining comprises the following steps: determining first user interest regression information corresponding to the electric business interaction knowledge chain example by means of each target interest focusing regression vector; determining a regression analysis cost corresponding to the AI machine learning model to be debugged by combining the user interest regression information, wherein the determining comprises the following steps: and connecting a decision discrimination model by an AI machine learning model, and determining the decision discrimination cost by combining the first user interest regression information and the user interest prior information.

According to the design, the first user interest regression information is determined directly by means of the target interest focusing regression vectors, and the to-be-debugged AI machine learning model is subjected to cycle debugging by means of the decision discrimination cost determined according to the first user interest regression information and the user interest prior information, so that the model regression analysis accuracy can be improved, and the regression analysis efficiency of the model can be guaranteed.

In some exemplary embodiments, the user interest regression information comprises second user interest regression information, the regression analysis cost comprises a probabilistic estimation cost; determining user interest regression information corresponding to the e-commerce business interaction knowledge chain example according to each target interest focusing regression vector, wherein the determining comprises the following steps: performing decision vector translation on each target interest focusing regression vector by means of at least one interest analysis algorithm set in the AI machine learning model to be debugged to obtain second user interest regression information; determining a regression analysis cost corresponding to the AI machine learning model to be debugged by combining the user interest regression information, wherein the determining comprises the following steps: and determining the probability estimation cost by combining the second user interest regression information and the user interest prior information.

According to the design, decision vector translation is carried out on each target interest focusing regression vector by means of an interest analysis algorithm to obtain second user interest regression information, and then cyclic debugging is carried out on the AI machine learning model to be debugged in a combined mode by means of decision discrimination cost and probability estimation cost determined according to the second user interest regression information and user interest prior information, so that the interest mining accuracy of the AI machine learning model obtained by debugging can be further guaranteed.

In a second aspect, the present invention also provides an AI system comprising a processor and a memory; the processor is connected with the memory in communication, and the processor is used for reading the computer program from the memory and executing the computer program to realize the method.

In a third aspect, the invention also provides a computer-readable storage medium, on which a program is stored, which program, when executed by a processor, performs the method described above.

Drawings

Fig. 1 is a schematic flowchart of a data processing method based on big data according to an embodiment of the present invention.

Fig. 2 is a schematic communication architecture diagram of an application environment of a data processing method based on big data according to an embodiment of the present invention.

Detailed Description

The method embodiments provided by the embodiments of the present invention may be executed in an AI system, a computer device, or a similar computing device. Taking the example of operating on an AI system, the AI system 10 may include one or more processors 102 (the processors 102 may include, but are not limited to, a processing device such as a microprocessor MCU or a programmable logic device FPGA, etc.) and a memory 104 for storing data, and optionally a transmission device 106 for communication functions. It will be understood by those skilled in the art that the above-described structure is illustrative only, and is not intended to limit the structure of the AI system described above. For example, AI system 10 may also include more or fewer components than shown above, or have a different configuration than shown above.

The memory 104 can be used for storing computer programs, for example, software programs and modules of application software, such as a computer program corresponding to a big data based data processing method in the embodiment of the present invention, and the processor 102 executes various functional applications and data processing by running the computer programs stored in the memory 104, that is, implementing the above method. The memory 104 may include high speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, the memory 104 may further include memory remotely located from the processor 102, which may be connected to the AI system 10 via a network. Examples of such networks include, but are not limited to, the internet, intranets, local area networks, mobile communication networks, and combinations thereof.

The transmission device 106 is used for receiving or transmitting data via a network. The above-described specific example of the network may include a wireless network provided by a communication provider of the AI system 10. In one example, the transmission device 106 includes a Network adapter (NIC), which can be connected to other Network devices through a base station so as to communicate with the internet. In one example, the transmission device 106 may be a Radio Frequency (RF) module, which is used for communicating with the internet in a wireless manner.

Based on this, please refer to fig. 1, fig. 1 is a schematic flow chart of a data processing method based on big data according to an embodiment of the present invention, the method is applied to an AI system, and further includes the technical solutions described in steps 101 to 104.

Step 101: according to the number of conversation behavior knowledge fields in the electric business interaction knowledge chain to be mined, determining a first element contribution value of each original index element in a first element relationship network corresponding to the electric business interaction knowledge chain and a second element contribution value of each original pairing element in a second element relationship network corresponding to the electric business interaction knowledge chain.

In an embodiment of the present invention, the electric commerce interaction knowledge chain to be mined may include a plurality of session behavior knowledge fields, the electric commerce interaction knowledge chain may be determined according to the obtained electric commerce sessions, and the different electric commerce sessions correspond to different electric commerce interaction knowledge chains. The number of the session behavior knowledge fields included in the electric business interaction knowledge chain is determined by the information amount of the electric business session, and the larger the information amount of the electric business session is, the more the session behavior knowledge fields are included in the electric business interaction knowledge chain is.

For example, the e-commerce interaction knowledge chain to be mined = [ feature1, feature2, feature3,. ] and feature G ], where feature reflects the session behavior knowledge field, feature1 reflects the first session behavior knowledge field in the e-commerce interaction knowledge chain feature, feature2 reflects the second session behavior knowledge field in the e-commerce interaction knowledge chain feature, feature3 reflects the third session behavior knowledge field in the e-commerce interaction knowledge chain feature, and similarly, feature G reflects the G-th session behavior knowledge field in the e-commerce interaction knowledge chain feature, and G is determined by the information volume of the e-commerce session. The session behavior knowledge field can be understood as feature information or description vector of the user session behavior.

In the practical application process, the inventor finds out through creative work that, in order to improve the interest mining accuracy of mining the user interest of the electricity business interaction knowledge chain to be mined, the following technical scheme needs to be implemented before the step 101 is performed.

(1) And acquiring an electric business interaction knowledge chain to be mined, and performing knowledge transformation processing on the electric business interaction knowledge chain to obtain the transformed electric business interaction knowledge chain.

In the embodiment of the invention, for any electric business interaction knowledge chain needing to be mined, the electric business interaction knowledge chain to be mined needs to be obtained first. For example, the chain of e-commerce interaction knowledge to be mined may be existing and available directly when interest mining is needed. Or, the electric business conversation to be mined may be obtained first, knowledge refinement/feature mining may be performed on the electric business conversation to obtain a plurality of conversation behavior knowledge fields corresponding to the electric business conversation, and thus, an electric business interaction knowledge chain to be mined, which is composed of the plurality of conversation behavior knowledge fields, may be obtained. The sequence of each conversation behavior knowledge field in the electric business interaction knowledge chain to be mined is determined by the distribution label of the conversation behavior knowledge field in the electric business conversation.

The transformed electric business interaction knowledge chain comprises a plurality of transformed conversation behavior knowledge fields, and the number of the conversation behavior knowledge fields in the electric business interaction knowledge chain to be mined is consistent with the number of the transformed conversation behavior knowledge fields in the transformed electric business interaction knowledge chain and has a one-to-one matching relationship.

The data processing method based on big data provided by the embodiment of the invention can be carried out by an AI machine learning model which is debugged in advance, and the AI machine learning model can be deployed in an AI system.

After the electric business interaction knowledge chain to be mined is obtained, knowledge transformation processing can be performed on the electric business interaction knowledge chain to be mined by means of a first feature mapping unit (feedforward neural network unit) in the AI machine learning model, so that the transformed electric business interaction knowledge chain is obtained.

For example, by means of the first feature mapping unit, a knowledge projection operation (feature mapping) may be performed on each session behavior knowledge field in the electrical business interaction knowledge chain to be mined to obtain a transformed session behavior knowledge field corresponding to each session behavior knowledge field, and the transformed session behavior knowledge fields form the transformed electrical business interaction knowledge chain.

(2) And carrying out chain level arrangement (fusion processing based on the position of the knowledge chain) on the electric business interaction knowledge chain and the electric business interaction knowledge chain which is transformed to obtain the electric business interaction knowledge chain which is finished with arrangement.

In an actual implementation process, for each conversation behavior knowledge field in the electric business interaction knowledge chain, the conversation behavior knowledge field and the conversation behavior knowledge field which is in the converted electric business interaction knowledge chain and is corresponding to the conversation behavior knowledge field and is subjected to conversion are sorted, and the conversation behavior knowledge field which is corresponding to the conversation behavior knowledge field and is subjected to sorting is obtained. And then, the organized conversation behavior knowledge fields corresponding to the conversation behavior knowledge fields in the electric business interaction knowledge chain form the electric business interaction knowledge chain with the organized conversation behavior knowledge fields.

For example, the to-be-mined e-commerce interaction knowledge chain feature link = [ feature1, feature2, feature3,. And feature G ], the transformed e-commerce interaction knowledge chain feature link ' = [ feature '1, feature2, feature3,. And feature ' G ], where feature '1 is the transformed session behavior knowledge field corresponding to feature1, feature '2 is the transformed session behavior knowledge field corresponding to feature2, feature '3 is the transformed session behavior knowledge field corresponding to feature3, and similarly, feature ' G is the transformed session behavior knowledge field corresponding to feature G. After the sequence arrangement, feature1 and feature '1 may be arranged to obtain a finishing session behavior knowledge field feature "1 corresponding to the feature1, feature2 and feature'2 may be arranged to obtain a finishing session behavior knowledge field feature"2 corresponding to the feature2, feature3 and feature '3 may be arranged to obtain a finishing session behavior knowledge field feature "3 corresponding to the feature3, and similarly, feature G and feature' G may be arranged to obtain a finishing session behavior knowledge field feature" G corresponding to the feature G. Further, the obtained electric commerce interaction knowledge chain completing the arrangement is feature link "= [ feature"1, feature "2, feature"3,. -, feature "G ].

(3) Respectively carrying out knowledge projection operation on the electric business interaction knowledge chain which is finished with arrangement by means of a target vector relationship network to obtain a first element relationship network, a second element relationship network and a third element relationship network which correspond to the electric business interaction knowledge chain; the first element relation network, the second element relation network and the third element relation network correspond to different target vector relation networks respectively.

In the embodiment of the invention, the number of the element characteristics included in each of the first element relational network, the second element relational network and the third element relational network is consistent with the number of the organized conversation behavior knowledge fields in the organized e-commerce interaction knowledge chain. For example, the number of the arranged conversation behavior knowledge fields in the arranged e-commerce interaction knowledge chain is G, and the number of the element features included in each of the first element relationship network, the second element relationship network and the third element relationship network is also G. The first, second and third element relationship networks can also be understood as different feature matrices or feature maps.

For example, the first element relational network is obtained by performing knowledge projection operation on the organized e-commerce interaction knowledge chain by means of a set first target vector relational network corresponding to the relational network. The first element relation network can be reflected by MAP1, and the original index elements (original query features) in MAP1 are used for reflecting the element features of the conversation behavior knowledge fields themselves which are finished with sorting.

And the second element relation network is obtained by performing knowledge projection operation on the sorted electric business interaction knowledge chain by means of a set second target vector relation network corresponding to the relation network. The second element relation network can be reflected by MAP2, and original pairing elements in the MAP2 are used for reflecting element characteristics after knowledge refinement/characteristic mining is carried out on the organized conversation behavior knowledge fields on an element attention level. The original matching element (original key feature) in MAP2 is used to match with the original index element.

And the third element relational network is obtained by performing knowledge projection operation on the organized e-commerce business interaction knowledge chain by means of a set third target vector relational network corresponding to the relational network. The third element relation network can be reflected by MAP3, and the target bias element in the MAP3 is used for reflecting element characteristics after knowledge refinement/characteristic mining is carried out on the finished conversation behavior knowledge field on another element attention level. The target bias element (original attention feature) in MAP3 is used to determine interest focus decision vectors corresponding to the respective session behavior knowledge fields in the electrical business interaction knowledge chain to be mined.

In an actual implementation process, after the electric business interaction knowledge chain which is finished with arrangement is obtained, the electric business interaction knowledge chain which is finished with arrangement can be loaded to a basis focus analysis unit (attention unit) in an AI machine learning model, a step of performing knowledge projection operation on the electric business interaction knowledge chain which is finished with arrangement by means of a set target vector relationship network is performed through the attention unit, and finally a first element relationship network, a second element relationship network and a third element relationship network which correspond to the electric business interaction knowledge chain are obtained.

In some examples, after the electric business interaction knowledge chain to be mined is obtained, the electric business interaction knowledge chain to be mined may also be directly loaded to the attention unit, the attention unit performs a step of performing knowledge projection operation on the electric business interaction knowledge chain to be mined by using a set target vector relationship network, and finally, a first element relationship network, a second element relationship network, and a third element relationship network corresponding to the electric business interaction knowledge chain are obtained.

The above first element contribution value is used to perform quantization operation (weighting process) on a target index element corresponding to an original index element in the first element relationship network, the target index element is obtained by transforming the original index element by a set target AI algorithm unit, and the first element contribution value of each original index element is used to perform quantization operation on the target index element corresponding to the original index element. The second element contribution value is used for performing quantization operation on a target paired element corresponding to an original paired element in the second element relationship network, the target paired element is obtained by transforming the original paired element by means of a set target AI algorithm unit, and the second element contribution value of each original paired element is used for performing quantization operation on the target paired element corresponding to the original paired element.

In an actual implementation process, a first element contribution value corresponding to an original index element can be determined according to the number of session behavior knowledge fields in an e-commerce business interaction knowledge chain to be mined and the first order priority of the original index element in a first element relation network. For example, the first element contribution value corresponding to each original index element may be calculated according to the related cosine contribution value (cosine weight). Similarly, the second element contribution value corresponding to the original paired element can be determined according to the number of the session behavior knowledge fields and the second order priority of the original paired element in the second element relationship network. For example, the second element contribution value corresponding to each original paired element may be calculated according to the related sine contribution value (sine weight).

In addition, for the step of determining the first element contribution value of each original index element, after the first element relationship network is determined, the first element contribution value corresponding to each original index element can be directly determined according to the number of the conversation behavior knowledge fields in the e-commerce business interaction knowledge chain and the first order priority of each original index element in the first element relationship network, so that the first element contribution value corresponding to each original index element is determined after the first element relationship network is determined and is directly obtained in the subsequent use. Alternatively, before actually determining the common value between an original index element and each original paired element, the first element contribution value corresponding to the original index element may be determined according to the first order priority of the original index element in the first element relationship network and the number of the session behavior knowledge fields, so that the first element contribution value corresponding to each original index element may be determined only when used, and is not determined in advance.

For the step of determining the second element contribution value of each original paired element, after the second element relationship network is determined, the second element contribution value corresponding to each original paired element may be directly determined according to the number of the session behavior knowledge fields in the e-commerce business interaction knowledge chain and the second order priority of each original paired element in the second element relationship network, so that the second element contribution value corresponding to each original paired element is determined after the second element relationship network is determined and is directly obtained during subsequent use. Alternatively, when actually applying a second element contribution value corresponding to an original paired element, the second element contribution value corresponding to the original paired element may be determined and used according to the number of the session behavior knowledge fields and the second order priority of the original paired element in the second element relationship network, and may not be determined in advance.

Step 102: and determining the common value between each original index element and each original pairing element according to the first element contribution value and the second element contribution value.

In the embodiment of the present invention, the common value is used to reflect the similarity between the original index element and the original pairing element.

In an actual implementation process, when determining a commonality value between an original index element and an original pairing element, a target index element corresponding to the original index element may be weighted by a first element contribution value corresponding to the original index element to obtain an element operation result. Meanwhile, the target paired elements corresponding to the original paired elements can be weighted by the contribution values of the second elements corresponding to the original paired elements, so as to obtain the element operation results. And then, determining a common value between the original index element and the original pairing element according to the two element operation results.

In this way, according to the first element contribution value and the target index element corresponding to each original index element, and the second element contribution value and the target paired element corresponding to each original paired element, the commonality value between each original index element and each original paired element can be determined.

In some possible design concepts, step 102 can be implemented exemplarily by the following related schemes.

Step 1021: and adjusting each original index element into a target index element meeting the setting requirement and adjusting each original pairing element into a target pairing element meeting the setting requirement by any one target AI algorithm unit in the set plurality of AI algorithm units.

In the embodiment of the present invention, the set AI algorithm units may include three AI algorithm units, that is, three kernel functions. The mining timeliness corresponding to the AI algorithm unit 1 is the fastest, the mining accuracy corresponding to the AI algorithm unit 3 is the best, and the mining timeliness corresponding to the AI algorithm unit 2 is the slowest. Any one of the above three AI algorithm units may be used as the target AI algorithm unit. For example, with respect to the use of the target AI algorithm element, the settings may be based on actual user interest mining tasks. For example, if the mining timeliness needs to be improved, the AI algorithm unit 1 may be selected as a target AI algorithm unit, if it needs to ensure that the mining accuracy is optimal, the AI algorithm unit 3 may be selected as a target AI algorithm unit, and if the mining timeliness is not balanced, the AI algorithm unit 2 may be selected as a target AI algorithm unit.

It can be understood that, according to the algorithm performances corresponding to the three AI algorithm units, each AI algorithm unit does not have a quantization value less than 0, and therefore, when any one of the AI algorithm units is selected as a target AI algorithm unit and an original index element and an original pairing element are transformed by the target AI algorithm unit, both the obtained target index element and the obtained target pairing element meet the setting requirements. Further, the relational network composed of the respective target index elements obtained by the transformation will also be a relational network in which the quantization intervals are not less than 0, and the relational network composed of the respective target pair elements obtained by the transformation will also be a relational network in which the quantization intervals are not less than 0.

For example, when the target AI algorithm unit is the AI algorithm unit 1, each original index element in the first element relation network may be loaded to the AI algorithm unit 1, so as to obtain a target index element corresponding to each original index element. Meanwhile, each original paired element in the second element relationship network may be loaded to the AI algorithm unit 1, so as to obtain a target paired element corresponding to each original paired element.

Step 1022: and determining a first element operation result corresponding to each original pairing element according to the second element contribution value corresponding to each original pairing element and the target pairing element corresponding to the original pairing element.

For example, for each original paired element, a product of the contribution value of the second element corresponding to the original paired element and the target paired element corresponding to the original paired element may be obtained, and a result obtained by the product may be used as a result of the operation of the first element corresponding to the original paired element.

Step 1023: for each original index element, determining a second element operation result corresponding to the original index element according to a target index element corresponding to the original index element and a first element contribution value corresponding to the original index element; and determining the common value between the original index element and each original pairing element according to the second element operation result and the first element operation result corresponding to each original pairing element.

In an actual implementation process, for each original index element, a product may be made between a target index element corresponding to the original index element and a first element contribution value corresponding to the original index element, and a result obtained after the product may be used as a second element operation result corresponding to the original index element.

Then, for the original index element and any original paired element, the second element operation result corresponding to the original index element and the transpose of the first element operation result corresponding to the original paired element may be multiplied, and according to the result obtained by the multiplication, the common value between the original index element and the original paired element is determined.

In some possible design concepts, the step 1023 of determining the common value between the original index element and each original paired element according to the second element operation result and the first element operation result corresponding to each original paired element can be exemplarily implemented by the following correlation scheme.

Step 10231: according to the first sequence priority of the original index elements in the first element relation network and the number of the conversation behavior knowledge fields, determining first eccentricity coefficients corresponding to the original index elements, and according to the second sequence priority of each original pairing element in the second element relation network, determining second eccentricity coefficients corresponding to each original pairing element.

In the embodiment of the present invention, the first order priority is determined by the distribution tag of the original index element in the first element relation network. For example, if the original index element is the first feature member in the first element relationship network, the first order priority corresponding to the original index element is 1, and if the original index element is the ith feature member in the first element relationship network, the first order priority corresponding to the original index element is i. The second order priority is determined by the distribution label of the original paired elements in the second element relation network.

Further, the first eccentricity coefficient is used for performing quantization operation on a target index element corresponding to the original index element. The second eccentricity coefficient is used for carrying out quantization operation on the target pairing element corresponding to the original pairing element. Thus, the eccentricity coefficient may be understood as a weighted weight, or may be understood as a sinusoidal weight.

Step 10232: and determining a third element operation result corresponding to the original index element according to the first eccentricity coefficient and the target index element corresponding to the original index element.

For each original index element, a product of the first eccentricity coefficient corresponding to the original index element and the target index element corresponding to the original index element may be obtained to obtain a third element operation result corresponding to the original index element.

Step 10233: and for each original pairing element, determining a fourth element operation result corresponding to the original pairing element according to the second eccentricity coefficient corresponding to the original pairing element and the target pairing element corresponding to the original pairing element.

In the embodiment of the present invention, for an original paired element, a product of a second eccentricity coefficient corresponding to the original paired element and a target paired element corresponding to the original paired element may be obtained to obtain a fourth element operation result corresponding to the original paired element.

Step 10234: and determining a common value between the original index element and each original pairing element according to the second element operation result, the third element operation result, the first element operation result corresponding to each original pairing element and the fourth element operation result.

In the embodiment of the present invention, for each original index element and any one original paired element, a product may be formed by transposing a second element operation result corresponding to the original index element and a first element operation result corresponding to the original paired element, a product may be formed by transposing a third element operation result corresponding to the original index element and a fourth element operation result corresponding to the original paired element, and the two product results may be added to obtain a common value between the original index element and the original paired element.

In an actual implementation process, a commonality value between any one original index element and any one original pairing element can be determined by means of the operation result and the target AI algorithm unit.

In some possible examples, in order to avoid an excessive processing overhead caused by directly integrating the first element relationship network and the second element relationship network, the first element relationship network and the second element relationship network may be respectively processed by the target AI algorithm unit to avoid directly integrating the first element relationship network and the second element relationship network, for example, the target AI algorithm unit may be used to perform the splitting process.

Further, under the condition that the target AI algorithm unit is the AI algorithm unit 3, a corresponding normalization function may be set to perform numerical value compression processing (normalization processing).

Step 103: and for each conversation behavior knowledge field in the electric business interaction knowledge chain, determining an interest focusing decision vector corresponding to the conversation behavior knowledge field according to the common value between the original index element corresponding to the conversation behavior knowledge field and each original pairing element and each target bias element in a third element relation network corresponding to the electric business interaction knowledge chain.

For example, for each session behavior knowledge field in the e-commerce service interaction knowledge chain, an interest focus decision vector (based on the attention feature of the interest level, used for interest prediction and mining) corresponding to the session behavior knowledge field may be determined according to a commonality value between the original index element Fi and each original pairing element FFj corresponding to the session behavior knowledge field, the target index element corresponding to the session behavior knowledge field, and the transposed result and each target bias element FFFj of each target pairing element.

In some possible design concepts, step 103 can be implemented by the following technical solutions.

Step 1031: and for each original pairing element, determining first feature calculation information corresponding to the original pairing element according to a first element calculation result corresponding to the original pairing element and a target offset element corresponding to a third sequence priority in a third element relation network and having the same second sequence priority as the original pairing element.

Step 1032: and determining second feature calculation information corresponding to the original pairing element according to a fourth element operation result corresponding to the original pairing element and a target offset element corresponding to a third sequence priority in a third element relation network and having the same second sequence priority as the original pairing element.

Step 1033: and multiplying the second element operation result corresponding to the original index element and the first feature calculation information corresponding to each original pairing element to obtain a fifth element operation result.

Step 1034: and multiplying a third element operation result corresponding to the original index element and second feature calculation information corresponding to each original pairing element to obtain a sixth element operation result.

Step 1035: and determining an interest focusing decision vector corresponding to the conversation behavior knowledge field by combining summation data of common values between the original index elements corresponding to the conversation behavior knowledge field and the original paired elements, summation data of the operation result of the fifth element and summation data of the operation result of the sixth element.

In the embodiment of the present invention, the third order priority is determined by the distribution label of the target bias element in the third element relation network. For example, if the target bias element is the first feature member in the third element relationship network, the third order priority corresponding to the target bias element is 1. The feature calculation information may be an intermediate operation result, and the element operation result may be weighted data or a weighted result.

For example, a to-be-mined electricity and commerce interaction knowledge chain feature link = ([ feature1, feature2, feature3,. And feature G ], and the session behavior knowledge fields are respectively feature1 to feature G, so that the obtained interest focus decision vector corresponding to feature1 is VEC _1, the obtained interest focus decision vector corresponding to feature2 is VEC _2, and similarly, the obtained interest focus decision vector corresponding to feature G is VEC _ G.

For another example, the first element relation network and the second element relation network are both G × d relation networks, and the processing overhead required for directly performing the product on the transposed relation network of the first element relation network and the second element relation network is G × d, and the processing overhead required for calculating the interest focus decision vector is reduced to G × d by performing the processing according to the technical scheme. Therefore, mining of a complex e-commerce business interaction knowledge chain is carried out according to the quick record, resource consumption is reduced, mining time consumption is reduced, and mining timeliness is effectively improved.

In the actual implementation process, the step of determining the interest focusing decision vector corresponding to each session behavior knowledge field can be performed by means of the attention unit in the debugged AI machine learning model, and finally the attention unit outputs the interest focusing decision vector corresponding to each session behavior knowledge field.

Step 104: and determining user interest mining information corresponding to the electric business interaction knowledge chain according to the interest focusing decision vector corresponding to each conversation behavior knowledge field in the electric business interaction knowledge chain.

In the embodiment of the invention, the user interest mining information is visual information (including but not limited to a series of data assets with certain value, such as service requirements, interest items, business preferences, push tags and the like of the user for the electric business) generated after mining the electric business interaction knowledge chain to be mined. For example, an e-commerce interaction knowledge chain to be mined is feature link = [ feature1, feature2, feature3,.. And feature G ], and user interest mining information corresponding to the e-commerce interaction knowledge chain may be Y = [ Y1, feature2, feature 3.. And yP ], where G is much larger than P. Here, Y1 to yP respectively reflect the mined different pieces of visualized information, and one piece of visualized information may be determined by a plurality of session behavior knowledge fields, so Y is much larger than P, and thus the number of session behavior knowledge fields is much larger than that of the visualized information.

In the actual implementation process, after the interest focusing decision vectors corresponding to the conversation behavior knowledge fields are obtained, the AI machine learning model can perform further interest prediction on the interest focusing decision vectors, so as to determine user interest mining information corresponding to the electric business service interaction knowledge chain.

According to the embodiment of the invention, the interest focus decision vector rule corresponding to each conversation behavior knowledge field is directly determined according to the element contribution value, so that the step of directly integrating a plurality of focus decision relationship networks is simplified, the occupation of processing overhead is reduced, and the timeliness of user interest mining is ensured. Moreover, the processing overhead is reduced, the resource consumption is reduced when the complex electric business interaction knowledge chain is responded, the time consumption for responding the complex electric business interaction knowledge chain is also reduced, and the user interest mining efficiency aiming at the complex electric business interaction knowledge chain is further ensured.

In some possible design concepts, step 104 can be implemented by the following related schemes.

Step 1041: and sorting all conversation behavior knowledge fields in the sorted electric business interaction knowledge chain and interest focusing decision vectors corresponding to the conversation behavior knowledge fields to obtain all the sorted interest focusing decision vectors.

In the embodiment of the invention, the electric business interaction knowledge chain after finishing the arrangement is an electric business interaction knowledge chain obtained by performing chain-level arrangement on the electric business interaction knowledge chain to be mined and the electric business interaction knowledge chain which is generated by the first feature mapping unit and is finished with the transformation.

Further, after obtaining the interest focusing decision vector corresponding to each session behavior knowledge field, for the interest focusing decision vector corresponding to each session behavior knowledge field, the interest focusing decision vector corresponding to the session behavior knowledge field and the sorted session behavior knowledge field corresponding to the session behavior knowledge field in the sorted e-commerce interaction knowledge chain may be sorted to obtain each sorted interest focusing decision vector. The number of the interest focusing decision vectors which are finished to be sorted is consistent with the number of the conversation behavior knowledge fields in the electric business interaction knowledge chain to be mined, and the interest focusing decision vectors which are finished to be sorted and the conversation behavior knowledge fields in the electric business interaction knowledge chain to be mined have a one-to-one matching relationship.

Step 1042: and respectively carrying out moving average processing on each interest focusing decision vector which is finished with moving average to obtain each interest focusing decision vector which is finished with moving average, and carrying out vector sorting on each interest focusing decision vector which is finished with moving average and the interest focusing decision vector which is finished with sorting and corresponds to the interest focusing decision vector to obtain an original interest focusing decision vector which corresponds to each interest focusing decision vector which is finished with moving average.

In the actual implementation process, each interest focusing decision vector after finishing sorting can be loaded to a sliding filtering unit (convolution unit) in the AI machine learning model, and each interest focusing decision vector after finishing sorting is respectively subjected to moving average processing by virtue of the convolution unit, so as to obtain each interest focusing decision vector after finishing moving average. And then carrying out vector sorting on each interest focusing decision vector which finishes the moving average and the sorted interest focusing decision vector corresponding to the interest focusing decision vector to obtain an original interest focusing decision vector corresponding to each interest focusing decision vector which finishes the moving average.

It can be understood that the number of the original interest focusing decision vectors is consistent with the number of the conversation behavior knowledge fields in the electric business interaction knowledge chain to be mined, and the original interest focusing decision vectors and the conversation behavior knowledge fields in the electric business interaction knowledge chain to be mined have a one-to-one matching relationship.

Step 1043: and carrying out knowledge transformation processing on each original interest focusing decision vector to obtain each transformed original interest focusing decision vector, and carrying out vector sorting on each original interest focusing decision vector and the transformed original interest focusing decision vector corresponding to the original interest focusing decision vector to obtain each sorted original interest focusing decision vector.

In the actual implementation process, each original interest focusing decision vector can be loaded to a second feature mapping unit (feedforward neural network unit) in the AI machine learning model, and the second feature mapping unit is used to perform knowledge transformation processing on each original interest focusing decision vector, so as to obtain each transformed original interest focusing decision vector. And then, vector sorting can be carried out on each original interest focusing decision vector and the transformed original interest focusing decision vector corresponding to the original interest focusing decision vector, so as to obtain each sorted original interest focusing decision vector.

Step 1044: and performing numerical value extrusion processing on each original interest focusing decision vector which is finished to be sorted to obtain to-be-processed interest focusing decision vectors corresponding to each conversation behavior knowledge field, and determining user interest mining information corresponding to the electric business interaction knowledge chain according to each to-be-processed interest focusing decision vector.

In the actual implementation process, each original interest focusing decision vector which is finished with sorting can be loaded to a standardization unit (normalization unit) in an AI machine learning model, numerical value extrusion processing is respectively carried out on each original interest focusing decision vector which is finished with sorting by means of the normalization unit, to-be-processed interest focusing decision vectors corresponding to each original interest focusing decision vector which is finished with sorting are obtained, and as each original interest focusing decision vector which is finished with sorting and a conversation behavior knowledge field in an electric business interaction knowledge chain to be mined have a one-to-one matching relationship, the to-be-processed interest focusing decision vector corresponding to each conversation behavior knowledge field in the electric business interaction knowledge chain to be mined is obtained.

Further, the AI machine learning model can directly mine the interest focus decision vector to be processed to determine user interest mining information corresponding to the electricity and business interaction knowledge chain.

Under some possible design ideas, the first feature mapping unit, the attention unit, the sliding filtering unit, the second feature mapping unit and the normalization unit belong to the elements of a feature processing sub-model (encoder) in the AI machine learning model. Therefore, after the electric business interaction knowledge chain to be mined is obtained, the electric business interaction knowledge chain to be mined can be loaded to the encoder, and each unit in the encoder can output user interest mining information corresponding to the electric business interaction knowledge chain to be mined according to the electric business interaction knowledge chain to be mined.

Under other possible design considerations, after obtaining each to-be-processed interest focus decision vector, a set formed by each to-be-processed interest focus decision vector may be determined as a derived to-be-mined electrical business interaction knowledge chain, and each to-be-processed interest focus decision vector is used as each session behavior knowledge field in the derived to-be-mined electrical business interaction knowledge chain.

And further, skipping to the step of carrying out knowledge transformation processing on the electric business interaction knowledge chain to obtain the transformed electric business interaction knowledge chain until the skip accumulated value reaches the set accumulated value to obtain each target interest focusing decision vector corresponding to the electric business interaction knowledge chain to be mined.

In the embodiment of the present invention, the set accumulated value is determined according to the number of encoders included in the AI machine learning model, and the number of encoders may be set according to a design situation.

For example, after each to-be-processed interest focusing decision vector is output by the first encoder, the analysis result of the first encoder may be used as a raw material of the second encoder, the input is processed by the second encoder to obtain each to-be-processed interest focusing decision vector generated by the second encoder, and each to-be-processed interest focusing decision vector generated by the second encoder may be used as a raw material of the third encoder. Wherein one target interest focus decision vector corresponds to one session behavior knowledge field in the chain of electrical business interaction knowledge to be mined. And finally, determining user interest mining information according to the generated target interest focusing decision vectors.

For example, the AI machine learning model may directly predict each generated target interest focus decision vector, and determine user interest mining information according to the prediction result.

Therefore, by means of a plurality of feature translation operations performed by a plurality of feature translators (decoders), the features reflected by the knowledge fields of the conversation behaviors can be deeply and completely mined, so that a target interest focusing decision vector containing diversified features can be obtained. And mining the user interest by using each target interest focusing decision vector, so that the accuracy of the determined user interest mining information can be guaranteed.

For another example, after obtaining each target interest focus decision vector, the decision vector translation may be performed on each target interest focus decision vector by means of at least one set interest analysis algorithm to obtain user interest mining information; the raw material of the downstream interest analysis algorithm in the two interest analysis algorithms with continuous translation priorities is the analysis result of the upstream interest analysis algorithm, and the analysis result of the interest analysis algorithm with the translation priority at the tail is the user interest mining information.

In an embodiment of the present invention, the interest resolution algorithm may be a feature translation algorithm. The number of interest analysis algorithms can be set according to actual user interest mining needs.

For example, the AI machine learning model may be deployed with 6 interest resolution algorithms, each corresponding to a distinct translation priority. The raw material of the downstream interest analysis algorithm in the two interest analysis algorithms with continuous translation priorities is the analysis result of the upstream interest analysis algorithm, and the analysis result of the interest analysis algorithm with the translation priority at the tail is the user interest mining information.

For example, after the analysis result of the 12 th encoder is the target interest focus decision vector, the target interest focus decision vector may be loaded to the first interest analysis algorithm with the translation priority of 1, the first interest analysis algorithm is used to perform decision vector translation on the target interest focus decision vector, so as to obtain and output the feature translation result corresponding to the target interest focus decision vector, then the analysis result of the first interest analysis algorithm may be used as a raw material loaded to the second interest analysis algorithm with the translation priority of 2, the second interest analysis algorithm is used to perform further feature translation on the feature translation result, so as to obtain the analysis result of the second interest analysis algorithm, the analysis result of the second interest analysis algorithm is used as a raw material of the third interest analysis algorithm with the translation priority of 3, and similarly, the analysis result of the sixth interest analysis algorithm with the translation priority of 6 is used as the final user interest mining information.

Under the condition that the AI machine learning model comprises the encoder and the decoder, after the electric business interaction knowledge chain to be mined is loaded into the AI machine learning model, the electric business interaction knowledge chain to be mined can be processed by the encoder and the decoder in the AI machine learning model, and finally, the decoder outputs user interest mining information. Or, under the condition that the AI machine learning model only includes the encoder, the user interest mining information may be output directly according to the target interest focusing decision vector after the most downstream encoder outputs each target interest focusing decision vector corresponding to the e-commerce business interaction knowledge chain to be mined. Whether a transcoder is required to be included in the AI machine learning model can be determined based on actual user interest mining needs.

In some possible design ideas, since the big data-based data processing method provided by the embodiment of the present invention can be performed by using an AI machine learning model for completing debugging, the embodiment of the present invention further provides a method for debugging the AI machine learning model, which may include the following contents.

Step 301: an example of an e-commerce business interaction knowledge chain to be mined is obtained.

In an embodiment of the present invention, several examples of session behavior knowledge fields may be included in the example of the e-commerce interaction knowledge chain to be mined.

Step 302: and loading the electric business interaction knowledge chain example to an AI machine learning model to be debugged, processing the electric business interaction knowledge chain example by means of the AI machine learning model to be debugged, and determining each target interest focusing regression vector corresponding to the electric business interaction knowledge chain example.

Further, one target interest focusing regression vector corresponds to one conversation behavior knowledge field example, the number of the target interest focusing decision vectors is consistent with the number of the conversation behavior knowledge field examples, and the target interest focusing regression vector is a regression analysis result corresponding to the conversation behavior knowledge field example generated after the conversation behavior knowledge field example is processed by the AI machine learning model to be debugged.

For example, each target interest focus regression vector may be the analytic result of the most downstream encoder in the AI machine learning model to be debugged.

Step 303: and determining user interest regression information corresponding to the electric business interaction knowledge chain examples according to the target interest focusing regression vectors.

In the embodiment of the invention, the user interest regression information is a visual regression analysis result generated after mining the electric business interaction knowledge chain example.

In an actual implementation process, interest prediction can be performed on each target interest focusing regression vector, and user interest regression information corresponding to the e-commerce business interaction knowledge chain example is determined according to an analysis result.

Step 304: and determining regression analysis cost corresponding to the AI machine learning model to be debugged according to the user interest regression information and the user interest prior information corresponding to the electric business service interaction knowledge chain example, and performing cyclic debugging on the AI machine learning model to be debugged by means of the regression analysis cost until the debugging termination requirement is met to obtain the debugged AI machine learning model.

In the embodiment of the invention, the user interest prior information is the prior knowledge (true value) corresponding to the e-commerce interaction knowledge chain example. The debugging termination requirement can be that the number of times of cycle debugging reaches a preset number of times and/or the accuracy of the debugged AI machine learning model reaches a preset accuracy.

In the practical implementation process, the regression analysis cost of the AI machine learning model to be debugged during user interest mining can be determined according to the user interest regression information and the user interest prior information, then the AI machine learning model to be debugged is subjected to cycle debugging by virtue of the regression analysis cost, and the debugged AI machine learning model is used as the finally debugged AI machine learning model under the condition that the debugging termination requirement is met.

In some possible design considerations, the user interest regression information may include first user interest regression information, and the regression analysis cost (predictive loss) may include a decision discrimination cost (classification loss).

In the embodiment of the present invention, the first user interest regression information may be user interest regression information obtained by directly processing a target interest focusing regression vector for an AI machine learning model to be debugged.

In an actual implementation process, under the condition that the target AI machine learning model to be debugged does not include a decoder, after each target interest focus regression vector corresponding to the e-commerce interaction knowledge chain example is output by a most downstream encoder in the AI machine learning model to be debugged, the AI machine learning model to be debugged may directly perform regression analysis processing on each target interest focus regression vector, and output first user interest regression information corresponding to the e-commerce interaction knowledge chain example.

Further, the decision discrimination cost may be a regression analysis cost determined without including the interest analysis algorithm. For example, the AI machine learning model may be connected to the decision-making discrimination model, and the decision-making discrimination cost may be determined according to the first user interest regression information and the user interest prior information.

In an actual implementation process, the first user interest regression information and the user interest prior information can be processed by means of a time series regression (CTC) strategy to obtain a decision judgment cost. And the decision discrimination cost can be improved by means of a time sequence regression strategy to obtain the improved decision discrimination cost.

Further, the AI machine learning model to be debugged can be debugged in a circulating manner by means of the improved decision-making judgment cost until the debugging termination requirement is met, so that the debugged AI machine learning model is obtained.

Under other possible design considerations, the user interest regression information may further include second user interest regression information, and the regression analysis cost may further include a probability estimation cost. The probabilistic estimation cost may be a regression analysis cost determined under conditions including an interest resolution algorithm. The probability estimation cost can also be understood as cross entropy loss.

For example, under the condition that the target AI machine learning model to be debugged includes not less than one interest analysis algorithm, the decision vector translation may be performed on each target interest focusing regression vector by using not less than one interest analysis algorithm set in the AI machine learning model to be debugged, so as to obtain the second user interest regression information.

For example, after outputting each target interest focusing regression vector corresponding to the e-commerce interaction knowledge chain example by a most downstream encoder in the AI machine learning model to be debugged, each target interest focusing regression vector may be loaded to a first decoder, and processed by the first decoder to obtain an analysis result of the first decoder, and then the analysis result of the first decoder is used as a raw material of a second decoder, and similarly, the analysis result of the most downstream decoder is used as second user interest regression information.

And then, determining the probability estimation cost corresponding to the AI machine learning model to be debugged according to the second user interest regression information and the user interest prior information.

And finally, carrying out cyclic debugging on the AI machine learning model to be debugged by means of the probability estimation cost until the debugging termination requirement is met, and obtaining the debugged AI machine learning model. Or, the AI machine learning model to be debugged can be debugged in a circulating manner by means of the probability estimation cost and the decision judgment cost, and the debugged AI machine learning model is obtained under the condition that the debugging termination requirement is met.

In the embodiment of the invention, if the probability estimation cost and the decision discrimination cost are used, the AI machine learning model to be debugged is jointly debugged in a circulating way, and when the debugged AI machine learning model is specifically applied, after a most downstream encoder in the AI machine learning model outputs a target interest focusing decision vector of regression analysis, user interest mining information can be directly determined by means of the target interest focusing decision vector; and performing decision vector translation on the target interest focusing decision vector by means of a decoder to determine user interest mining information.

For the case that the AI machine learning model to be debugged includes an encoder, after the most downstream encoder outputs each target interest focusing regression vector corresponding to the e-commerce interaction knowledge chain example, the first user interest regression information corresponding to the e-commerce interaction knowledge chain example may be determined directly according to each target interest focusing regression vector. And then, determining a decision judgment cost corresponding to the AI machine learning model to be debugged according to the first user interest regression information and the user interest prior information corresponding to the electric business interaction knowledge chain example.

For the case that the AI machine learning model to be debugged includes the encoder and the decoder, after the electrical business interaction knowledge chain example is loaded into the AI machine learning model to be debugged, the electrical business interaction knowledge chain example may be processed by means of the encoder and the decoder in the AI machine learning model, and finally, the decoder outputs second user interest regression information corresponding to the electrical business interaction knowledge chain example. Further, the probability estimation cost corresponding to the AI machine learning model to be debugged may be determined according to the second user interest regression information and the user interest prior information corresponding to the electricity business interaction knowledge chain example.

And then, carrying out cyclic debugging on the AI machine learning model to be debugged by means of at least one of the decision judgment cost and the probability estimation cost until the debugging termination requirement is met, so as to obtain the debugged AI machine learning model.

Based on the same or similar inventive concepts, please refer to fig. 2 in combination, which further provides an architectural diagram of an application environment 30 of a data processing method based on big data, including an AI system 10 and an e-commerce service platform system 20 that communicate with each other, and the AI system 10 and the e-commerce service platform system 20 implement or partially implement the technical solution described in the above method embodiments when they run.

Further, a computer-readable storage medium is provided, on which a program is stored, which when executed by a processor implements the above-described method.

The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention, and various modifications and changes may be made by those skilled in the art. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A big data-based data processing method is applied to a big data AI system, and the method comprises the following steps:

determining a first element contribution value of each original index element in a first element relationship network corresponding to the electric business interaction knowledge chain and a second element contribution value of each original pairing element in a second element relationship network corresponding to the electric business interaction knowledge chain according to the number of conversation behavior knowledge fields in the electric business interaction knowledge chain to be mined;

determining a commonality value between each original index element and each original pairing element respectively by combining the first element contribution value and the second element contribution value;

for each conversation behavior knowledge field in the electric business service interaction knowledge chain, determining an interest focusing decision vector corresponding to the conversation behavior knowledge field by combining a common value between an original index element corresponding to the conversation behavior knowledge field and each original pairing element and each target bias element in a third element relation network corresponding to the electric business service interaction knowledge chain;

and determining user interest mining information corresponding to the electric business interaction knowledge chain by combining the interest focusing decision vectors corresponding to the conversation behavior knowledge fields in the electric business interaction knowledge chain.

2. A method according to claim 1, wherein said determining a commonality value between each of said original index elements and each of said original mating elements in combination with said first element contribution value and said second element contribution value comprises:

adjusting each original index element into a target index element meeting the setting requirement and adjusting each original pairing element into a target pairing element meeting the setting requirement by any one target AI algorithm unit in a plurality of set AI algorithm units;

determining a first element operation result corresponding to each original pairing element according to a second element contribution value corresponding to each original pairing element and a target pairing element corresponding to the original pairing element;

for each original index element, determining a second element operation result corresponding to the original index element by combining a target index element corresponding to the original index element and a first element contribution value corresponding to the original index element;

and determining the common value between the original index element and each original pairing element by combining the second element operation result and the first element operation result corresponding to each original pairing element.

3. The method according to claim 2, wherein the determining the common value between the original index element and each of the original pair elements by combining the second element operation result and the first element operation result corresponding to each of the original pair elements comprises:

determining a first eccentricity coefficient corresponding to the original index elements by combining the first sequence priority of the original index elements in the first element relationship network and the number of the conversation behavior knowledge fields, and determining a second eccentricity coefficient corresponding to each original pairing element according to the second sequence priority of each original pairing element in the second element relationship network;

combining the first eccentricity coefficient and a target index element corresponding to the original index element to determine a third element operation result corresponding to the original index element;

for each original pairing element, determining a fourth element operation result corresponding to the original pairing element by combining a second eccentricity coefficient corresponding to the original pairing element and a target pairing element corresponding to the original pairing element;

and determining a common value between the original index element and each original pairing element by combining the second element operation result, the third element operation result, and the first element operation result and the fourth element operation result corresponding to each original pairing element.

4. The method of claim 1, wherein the determining the interest focusing decision vector corresponding to the knowledge field of conversation behavior by combining the commonality value between the original index element corresponding to the knowledge field of conversation behavior and each of the original pairing elements and each of the target bias elements in the third element relationship network corresponding to the knowledge chain of electric business interaction comprises:

for each original pairing element, determining first feature calculation information corresponding to the original pairing element by combining a first element operation result corresponding to the original pairing element and a target bias element corresponding to the third element relation network and having the same third sequence priority as a second sequence priority corresponding to the original pairing element;

determining second feature calculation information corresponding to the original pairing element by combining a fourth element operation result corresponding to the original pairing element and a target bias element corresponding to the third element relation network and having the same third sequence priority as the second sequence priority corresponding to the original pairing element;

multiplying a second element operation result corresponding to the original index element with first feature calculation information corresponding to each original pairing element to obtain a fifth element operation result;

multiplying a third element operation result corresponding to the original index element with second feature calculation information corresponding to each original pairing element to obtain a sixth element operation result;

and determining an interest focusing decision vector corresponding to the conversation behavior knowledge field by combining summation data of common values between the original index elements corresponding to the conversation behavior knowledge field and the original paired elements, summation data of the operation result of the fifth element and summation data of the operation result of the sixth element.

5. The method of claim 1, wherein before determining the first element contribution value of each original index element in the first element relationship network corresponding to the electric business interaction knowledge chain according to the number of the session behavior knowledge fields in the electric business interaction knowledge chain to be mined, the method further comprises:

acquiring the electric business interaction knowledge chain to be mined, and performing knowledge transformation processing on the electric business interaction knowledge chain to obtain the transformed electric business interaction knowledge chain;

performing chain level sorting on the electric business interaction knowledge chain and the electric business interaction knowledge chain which is transformed to obtain the electric business interaction knowledge chain which is finished with sorting;

respectively carrying out knowledge projection operation on the electric business interaction knowledge chain which is finished with arrangement by means of a target vector relationship network to obtain a first element relationship network, a second element relationship network and a third element relationship network which correspond to the electric business interaction knowledge chain; the first element relation network, the second element relation network and the third element relation network correspond to different target vector relation networks respectively.

6. The method of claim 5, wherein the determining the user interest mining information corresponding to the e-commerce interaction knowledge chain in combination with the interest focus decision vector corresponding to each of the session behavior knowledge fields in the e-commerce interaction knowledge chain comprises:

sorting all conversation behavior knowledge fields in the sorted e-commerce business interaction knowledge chain and interest focusing decision vectors corresponding to the conversation behavior knowledge fields to obtain all the sorted interest focusing decision vectors;

respectively carrying out moving average processing on each interest focusing decision vector which is finished with moving average to obtain each interest focusing decision vector which is finished with moving average, and carrying out vector sorting on each interest focusing decision vector which is finished with moving average and the interest focusing decision vector which is finished with sorting and corresponds to the interest focusing decision vector to obtain an original interest focusing decision vector which corresponds to each interest focusing decision vector which is finished with moving average;

performing knowledge transformation processing on each original interest focusing decision vector to obtain each transformed original interest focusing decision vector, and performing vector sorting on each original interest focusing decision vector and each transformed original interest focusing decision vector corresponding to the original interest focusing decision vector to obtain each sorted original interest focusing decision vector;

and performing numerical value extrusion processing on each original interest focusing decision vector which is finished to be sorted to obtain to-be-processed interest focusing decision vectors corresponding to each conversation behavior knowledge field, and determining user interest mining information corresponding to the e-commerce service interaction knowledge chain according to each to-be-processed interest focusing decision vector.

7. The method of claim 6, wherein determining user interest mining information corresponding to the e-commerce business interaction knowledge chain based on each pending interest focus decision vector comprises: determining a set formed by the to-be-processed interest focusing decision vectors as a derived to-be-mined electrical business interaction knowledge chain, and taking the to-be-processed interest focusing decision vectors as each session behavior knowledge field in the derived to-be-mined electrical business interaction knowledge chain; skipping to the step of carrying out knowledge transformation processing on the electricity business interaction knowledge chain to obtain the transformed electricity business interaction knowledge chain until the skip accumulated value reaches the set accumulated value to obtain each target interest focusing decision vector corresponding to the electricity business interaction knowledge chain to be mined; determining the user interest mining information according to each target interest focusing decision vector;

wherein the determining the user interest mining information according to each of the target interest focusing decision vectors includes: performing decision vector translation on each target interest focusing decision vector by means of at least one set interest analysis algorithm to obtain the user interest mining information; the raw material of the downstream interest analysis algorithm in the two interest analysis algorithms with continuous translation priority is the analysis result of the upstream interest analysis algorithm, and the analysis result of the interest analysis algorithm with the translation priority at the tail is the user interest mining information.

8. The method according to claim 1, characterized in that the big data based data processing method is performed by means of an AI machine learning model that is debugged in advance; the AI machine learning model is debugged based on the following manner: obtaining an e-commerce business interaction knowledge chain example to be mined; loading the electric business interaction knowledge chain example to an AI machine learning model to be debugged, processing the electric business interaction knowledge chain example by means of the AI machine learning model to be debugged, and determining each target interest focusing regression vector corresponding to the electric business interaction knowledge chain example; determining user interest regression information corresponding to the electricity business interaction knowledge chain example according to each target interest focusing regression vector; determining regression analysis cost corresponding to the AI machine learning model to be debugged by combining the user interest regression information and the user interest prior information corresponding to the e-commerce business interaction knowledge chain example, and performing cyclic debugging on the AI machine learning model to be debugged by means of the regression analysis cost until the debugging termination requirement is met to obtain the debugged AI machine learning model; the user interest regression information comprises first user interest regression information, and the regression analysis cost comprises decision discrimination cost.

9. The method of claim 8, wherein the determining user interest regression information corresponding to the example of the e-commerce business interaction knowledge chain based on each of the target interest focus regression vectors comprises: determining first user interest regression information corresponding to the electric business interaction knowledge chain example by means of each target interest focusing regression vector; determining a regression analysis cost corresponding to the AI machine learning model to be debugged by combining the user interest regression information, wherein the determining comprises the following steps: and connecting a decision discrimination model by an AI machine learning model, and determining the decision discrimination cost by combining the first user interest regression information and the user interest prior information.

10. The method of claim 9, wherein the user interest regression information comprises second user interest regression information, the regression analysis cost comprises a probabilistic estimation cost; determining user interest regression information corresponding to the e-commerce business interaction knowledge chain example according to each target interest focusing regression vector, wherein the determining comprises the following steps: performing decision vector translation on each target interest focusing regression vector by means of at least one interest analysis algorithm set in the AI machine learning model to be debugged to obtain second user interest regression information; determining a regression analysis cost corresponding to the AI machine learning model to be debugged by combining the user interest regression information, wherein the determining comprises the following steps: and determining the probability estimation cost by combining the second user interest regression information and the user interest prior information.

11. An AI system comprising a processor and a memory; the processor is communicatively connected to the memory, and the processor is configured to read the computer program from the memory and execute the computer program to implement the method of any one of claims 1 to 8.