CN116757763A

CN116757763A - Electronic commerce recommendation method and system based on knowledge graph deep learning

Info

Publication number: CN116757763A
Application number: CN202310733112.7A
Authority: CN
Inventors: 郭志勇
Original assignee: Guangzhou Ketuo Technology Co ltd
Current assignee: Guangzhou Ketuo Technology Co ltd
Priority date: 2022-11-09
Filing date: 2023-06-19
Publication date: 2023-09-15
Also published as: CN115439197A

Abstract

The embodiment of the invention relates to the technical field of electronic commerce, and discloses an electronic commerce recommendation method based on knowledge graph deep learning, which comprises the following steps: receiving browsing information transmitted by an application interface end; transmitting the browsing information to a pre-constructed commodity relation map model to extract entity relation so as to obtain a customer commodity interaction matrix; and transmitting the browsing information, the commodity interaction matrix of the customer and the commodity relation map model to a recommendation algorithm model to conduct recommendation identification so as to determine commodity recommendation results of corresponding customers. According to the recommendation method, through deep learning of the knowledge graph and the browsing log, a recommendation result reflecting browsing preference of a customer is output; the application interface module receives the retrieval of the recommendation result by the upper layer application, and realizes the integration of the recommendation application; the recommendation method of the embodiment can capture the preference characteristics of the customers on the commodities, support personalized commodity recommendation and improve commodity browsing efficiency of the customers on the electronic commerce platform.

Description

Electronic commerce recommendation method and system based on knowledge graph deep learning

Technical Field

The invention relates to the technical field of electronic commerce, in particular to an electronic commerce recommendation method and system based on knowledge graph deep learning.

Background

At present, a textile electronic commerce platform belongs to an industry electronic commerce platform, and a B2B electronic commerce mode combining online exhibition and offline transaction is adopted, so that the textile electronic commerce platform is various in commodity types and huge in customer number. In order to improve browsing efficiency and experience of customers, a commodity recommendation system needs to be developed, and commodities are actively recommended to customers according to the preference of the customers. Existing recommendation systems generally make combined recommendations based on registered users, but in textile electronic commerce platforms, customers are mainly unregistered tourists, evaluation and transaction records are scarce, and preference performance is limited to browsing records. The method brings the challenges of difficult identification of customers, difficult extraction of preference characteristics, difficult guarantee of recommendation performance and the like for the construction of a recommendation system. Therefore, designing a solution capable of improving the recommendation accuracy is a technical problem to be solved by those skilled in the art.

Disclosure of Invention

Aiming at the defects, the embodiment of the invention discloses an electronic commerce recommendation method based on knowledge graph deep learning, which can timely capture preference characteristics of customers on commodities, support personalized recommendation based on address information and promote commodity browsing efficiency of the customers on an electronic commerce platform.

The first aspect of the embodiment of the invention discloses an electronic commerce recommendation method based on knowledge graph deep learning, which comprises the following steps:

receiving browsing information transmitted by an application interface end, wherein the browsing information is operation content at an electronic commerce platform of a corresponding user, and comprises commodity information and a browsing log; the browse log comprises browser ip address information; determining customer information based on the browser ip address information;

sending the browsing information to a pre-constructed commodity relation map model to extract entity relation so as to obtain a customer commodity interaction matrix; describing commodity attribute relations and commodity attribution relations by adopting a resource description framework in the commodity relation map model, wherein the commodity attribute relations are expressed in a form of triples of commodity numbers, attribute relations and attribute value numbers, and the commodity attribution relations are expressed in a form of triples of commodity numbers, attribution and store numbers;

and transmitting the browsing information, the commodity interaction matrix of the customer and the commodity relation map model to a recommendation algorithm model for recommendation identification so as to determine commodity recommendation results of corresponding customers.

As an optional implementation manner, in the first aspect of the embodiment of the present invention, after receiving the browsing information transmitted by the application interface, the method further includes:

performing data cleaning operation on the browsing information to filter non-commodity browsing records and commodity field missing browsing records in the browsing log;

and judging the validity of the browsing information subjected to the filtering operation, when the browsing times or browsing time of the same commodity by a customer exceeds a first set value, determining that the commodity is valid for the customer, and otherwise, filtering the corresponding browsing information.

As an optional implementation manner, in the first aspect of the embodiment of the present invention, the commodity relational graph model is constructed by the following steps:

acquiring commodity attribute information at a corresponding electronic commerce platform;

extracting a plurality of commodity entities, commodity attributes and store entities from the commodity attribute information, and determining relations among the plurality of commodity entities, between the commodity entities and the store entities and between the commodity entities and the commodity attributes;

and constructing a commodity relation map model based on the relations among the commodity entities, the commodity entities and the store entities and the commodity attributes.

As an optional implementation manner, in the first aspect of the embodiment of the present invention, after extracting a plurality of commodity entities, commodity attributes, and store entities from the commodity attribute information, the method further includes:

determining browsing adjacency relations among commodities based on browsing relations of the commodities of the customers from the historical browsing information; the browsing adjacent relation is that if a plurality of commodities are browsed in one session period, the commodities are mutually browsed adjacent relation;

when the number of times of the mutual browsing adjacency relationship reaches a second set value, the existence of the browsing adjacency relationship between the commodities is determined.

As an optional implementation manner, in the first aspect of the embodiment of the present invention, the commodity relational graph model further includes the following steps:

if the current commodity recommendation graph construction process is not the first commodity relationship graph model construction process, the relationships among the commodity entities are fused into the previous commodity relationship graph model to obtain the commodity entities of the current commodity relationship graph model and the relationships among the commodity entities.

As an optional implementation manner, in the first aspect of the embodiment of the present invention, the recommended algorithm model includes a prediction function model, and the prediction function model is constructed by the following steps:

Dividing the set of customer-merchandise interactions (u, i) into a positive set of interactions based on the travel log LAnd interaction negative set->If there is an access record of customer u to item i in the travel log L, then a determination is made +.>Otherwise->The formalization is expressed as follows:

gathering the interaction positive setInteractive negative set->Inputting the commodity browsing adjacency graph NG and the commodity knowledge graph KG into a pre-constructed prediction function formula to predict so as to obtain a corresponding prediction result; the prediction function formula is as follows:

wherein ,representation->F represents a prediction function, U is customer information, U is a customer matrix, and u= { U1, U2,..mu.m }, I is merchandise information, I is a merchandise matrix, and i= { I1, I2,..in }.

In an optional implementation manner, in the first aspect of the embodiment of the present invention, after the sending the browsing information to a pre-built commodity relationship graph model to perform entity relationship extraction to obtain a customer commodity interaction matrix, the method further includes:

combining the browsing information with a commodity relation map KG, and extracting the intention of a customer-commodity interaction individual by adopting a KGIN algorithm to obtain corresponding individual intention characteristics;

combining the access log L with a commodity browsing adjacency graph NG, and extracting the customer-commodity interaction group intention of the group by adopting graph aggregation operation to obtain corresponding group intention characteristics;

And adopting an attention mechanism to perform feature fusion on the individual intention features and the group intention features to form fusion features, and evaluating interest scores of customers on commodities according to the fusion features.

In a first aspect of the present embodiment, the combining the browsing information with the commodity relationship graph model, and extracting the intent of the customer-commodity interaction individual by using a KGIN algorithm to obtain a corresponding individual intent set includes:

decomposing the acquired customer-commodity group (u, i) into an intention set c= { (u, P, i) |p e P }; reorganizing the original customer interaction record into an intention graph IG;

the method comprises the steps of associating an intention with a commodity relation map KG by adopting a KGIN algorithm, defining the intention p of a customer as nonlinear combination of different relations in a knowledge map, wherein the embedded expression of the intention is as follows:

wherein ,e_r Is an initial representation vector of the relation r, which is given to the attention score alpha (r, p) to quantify the importance thereof, w _rp Trainable weights, being specific relation r and specific intent p, are d-dimensional random initial vectors, w _r′p Representing other intent weights under the same relationship;

in the intent diagram IG, N is used _u = { (p, i) | (u, p, i) ∈c } to represent intent awareness history and first order connectivity around customer u, embedded representation of customer is represented with intent awareness information integration of history by embedded formulas:

in the formula ,is a feature representation of KGIN first-order path aggregation for customer u,as indicated by the dot product, β (u, p) is the weight for distinguishing the intention of the individual customer, e _p ^T Representation e _p Is the transposed vector of (c), exp represents the exponential operation, e _p′ ^T Representing a transpose of the other intent vector;

the KGIN algorithm adopts a characteristic aggregation mode based on the relation and the nodes to obtain the characteristic aggregation of the commodity i

wherein ,N_i Is a triplet set associated with commodity i as a header entity,then the tail entity v corresponding to commodity i is embedded with a representation,/->

The steps are repeated to obtain deeper information representation, and each layer of information is polymerized in multiple layers to obtain a wider receptive field, so that the final commodity and customer characteristics of KGIN are represented as follows:

wherein Q is the number of polymerization layers;

combining the access log L with a commodity browsing adjacency graph NG, and extracting the group intention of customer-commodity interaction of the group by adopting graph aggregation operation to obtain corresponding group intention characteristics

Acquiring browsing adjacent relation characteristics of customers and commodities by adopting first-order aggregation, wherein browsing adjacent relation characteristics e of commodities i _i ^Nei The expression is as follows:

wherein ,N_i ^Nei All commodity sets representing the browsing adjacency relationship with commodity i, j is N _i ^Nei Commodity e _j Is the initial representation vector for commodity j;

description of customer preference according to customer-commodity interaction relationship to obtain customer browsing adjacency feature, browsing adjacency feature e of customer u _u ^Nei The expression is as follows:

wherein ,N_u Is a browsing commodity collection for customer u.

In a first aspect of the embodiment of the present invention, the feature fusing the individual intention feature and the group intention feature to form a fused feature, and evaluating the interest score of the customer in the commodity according to the fused feature includes:

performing splicing operation on the browsing adjacent feature and the KGIN feature on the same dimension according to a splicing formula to obtain a splicing feature vector, wherein the splicing formula is as follows:

e _u ^concat ＝e _u ^Nei ||e _u ^KGIN

e _i ^concat ＝e _i ^Nei ||e _i ^KGIN

wherein ,e_u ^concat and e_i ^concat Respectively representing the splicing characteristics of a customer u and a commodity i, wherein I represents vector splicing operation;

the spliced feature vector and the attention weight w _att Multiplying and normalizing to select the weight key of key feature _weight ；

Employing the same attention weight w for customer and merchandise features _att According to key _weight Performing point multiplication on the spliced characteristic to obtain a fusion characteristic e _u ^fusion and e_i ^fusion ；

Inputting the fusion characteristics into a prediction formula to predict so as to obtain interest scores of customers on commodities, wherein the prediction formula is as follows:

The second aspect of the embodiment of the invention discloses an electronic commerce recommendation system based on knowledge graph deep learning, which comprises the following components:

and a receiving module: the method comprises the steps of receiving browsing information transmitted by an application interface end, wherein the browsing information is operation content of an electronic commerce platform of a corresponding user, and comprises commodity information and a browsing log; the browse log comprises browser ip address information; determining customer information based on the browser ip address information;

entity extraction module: the method comprises the steps of sending browsing information to a pre-constructed commodity relation map model to extract entity relations so as to obtain a customer commodity interaction matrix; describing commodity attribute relations and commodity attribution relations by adopting a resource description framework in the commodity relation map model, wherein the commodity attribute relations are expressed in a form of triples of commodity numbers, attribute relations and attribute value numbers, and the commodity attribution relations are expressed in a form of triples of commodity numbers, attribution and store numbers;

and a recommendation module: and the browsing information and the commodity interaction matrix of the customer are transmitted to a recommendation algorithm model to conduct recommendation identification so as to determine commodity recommendation results of corresponding customers.

A third aspect of an embodiment of the present invention discloses an electronic device, including: a memory storing executable program code; a processor coupled to the memory; the processor invokes the executable program code stored in the memory to execute the electronic commerce recommendation method based on the knowledge graph deep learning disclosed in the first aspect of the embodiment of the present invention.

A fourth aspect of the embodiment of the present invention discloses a computer-readable storage medium storing a computer program, where the computer program causes a computer to execute the electronic commerce recommendation method based on knowledge-graph deep learning disclosed in the first aspect of the embodiment of the present invention.

Compared with the prior art, the embodiment of the invention has the following beneficial effects:

according to the electronic commerce recommendation method based on the knowledge graph deep learning, the knowledge graph and the browsing log are subjected to deep learning, and recommendation results reflecting browsing preference of customers are output; the application interface module receives the retrieval of the recommendation result by the upper layer application, and realizes the integration of the recommendation application; the recommendation method of the embodiment can capture the preference characteristics of the customers on the commodities, support personalized commodity recommendation and improve commodity browsing efficiency of the customers on the electronic commerce platform.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings that are needed in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.

FIG. 1 is a schematic flow diagram of an electronic commerce recommendation method based on knowledge-graph deep learning according to an embodiment of the present invention;

FIG. 2 is a schematic flow chart of filtering browsing information according to an embodiment of the present invention;

FIG. 3 is a schematic flow chart of the commodity relational graph model construction disclosed in the embodiment of the invention;

FIG. 4 is a schematic flow diagram of a recommendation algorithm model construction disclosed in an embodiment of the present invention;

FIG. 5 is a technical framework diagram of a recommendation system disclosed in an embodiment of the present invention;

FIG. 6 is an ontology model diagram of a commodity knowledge graph according to an embodiment of the present invention;

figure 7 is a block diagram of a KGCN model disclosed in an embodiment of the present invention;

fig. 8 is a diagram showing the information transfer of the KGCN model disclosed in the embodiment of the present invention;

FIG. 9 is a schematic diagram of a first order polymerization process of the KGCN model according to the embodiment of the invention;

FIG. 10 is a schematic diagram of a recommendation principle disclosed in an embodiment of the present invention;

FIG. 11 is a deployment diagram of an e-commerce recommendation system disclosed in an embodiment of the present invention;

FIG. 12 is an exemplary diagram of an electronic knowledge map of a textile disclosed in an embodiment of the invention;

FIG. 13 is a graph showing the variation of algorithm recommendation accuracy with the number of learning iterations disclosed in the embodiments of the present invention;

FIG. 14 is a graph comparing AUC with view adjacency according to the present disclosure;

FIG. 15 is a graph showing a comparison of the presence or absence of browsing adjacencies of ACCs disclosed in embodiments of the present invention;

FIG. 16 is a block diagram of a KGIN-DF algorithm according to an embodiment of the invention

FIG. 17 is a schematic drawing of KGIN feature extraction according to an embodiment of the disclosure;

FIG. 18 is a schematic view of browsing adjacency feature extraction in accordance with an embodiment of the present invention;

FIG. 19 is a schematic diagram of attention feature fusion as disclosed in an embodiment of the present invention;

fig. 20 is a schematic structural diagram of an electronic commerce recommendation system based on knowledge-graph deep learning according to an embodiment of the present invention.

Detailed Description

The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

It should be noted that the terms "first," "second," "third," "fourth," and the like in the description and in the claims of the present invention are used for distinguishing between different objects and not necessarily for describing a particular sequential or chronological order. The terms "comprises," "comprising," and "having," and any variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, system, article, or apparatus that comprises a list of steps or elements is not necessarily limited to those steps or elements expressly listed or inherent to such process, method, article, or apparatus.

Classical recommendation algorithms can be divided into three categories, content filtering, collaborative filtering, and model filtering. All three types of classical algorithms need explicit customer transaction scoring records, and for textile e-commerce platforms lacking scoring records, the browsing logs need to be converted into scoring records, such as conversion according to browsing times or duration, but in fact, browsing reflects customer preference far weaker than transaction and scoring records, and conversion from browsing to scoring introduces a lot of noise to cause poor recommendation performance. For textile e-commerce recommendations, the following problems still remain to be solved: identification of non-registered guests; the knowledge graph construction problem suitable for deep learning; the deep learning framework design problem supporting multiple algorithms. The embodiment of the invention discloses an electronic commerce recommendation method, an electronic commerce recommendation system, electronic equipment and a storage medium based on knowledge graph deep learning, wherein a recommendation result reflecting browsing preference of a customer is output by performing deep learning on a knowledge graph and a browsing log; the application interface module receives the retrieval of the recommendation result by the upper layer application, and realizes the integration of the recommendation application; the recommendation method of the embodiment can capture the preference characteristics of the customers on the commodities, support personalized commodity recommendation and improve commodity browsing efficiency of the customers on the electronic commerce platform.

The recommendation method based on the knowledge graph can not only recommend according to the historical browsing log, but also combine commodity relations, and is a development trend of the current recommendation technology. Two analogies, path-based coding and depth-based learning, can be categorized according to the technology. The recommendation algorithm based on the path coding relies on experience of people, sets a recommendation path based on the relation between a customer and commodities, calculates commodity similarity along the recommendation path by taking the commodities accessed by the customer as a starting point, and recommends based on the similarity. The recommendation method based on the path coding is not generalized enough due to the fact that the recommendation method depends on human experience.

The recommendation method based on deep learning takes a knowledge graph and a customer interaction record as input, predicts the browsing probability of unknown customers on commodities through deep learning, enriches the recommendation path from the customers to the commodities, can relieve the influence of scoring record deletion, and can be regarded as a hybrid algorithm based on content and collaborative filtering. Currently, three classical models are mainly included: rippleNet, KGCN (Knowledge Graph Convolutional Networks) and KGAT (Knowledge Graph Attention Network) and derivative series thereof. RippleNet strengthens embedded representation of nodes by utilizing multi-hop relations of project connection, KGCN realizes more accurate recommendation by weighting map relations, KGAT introduces a customer-commodity interaction matrix into a knowledge map and weights the map relations to realize recommendation. The CKAN (Collaborative Knowledge-aware Attentive Network) model is a novel derivative knowledge graph recommendation algorithm, and compared with the three classical models, the model mainly improves the initial representation of customers, and expresses the customers by using an interactive commodity set, so that the recommendation can be performed only by the interaction record of the customers and the commodities, the application range is improved, and the learning frequency is reduced.

The textile electronic commerce platform belongs to an industry electronic commerce platform, adopts a B2B electronic commerce mode combining online exhibition and offline transaction, has various commodity types and huge customer quantity, and has strong recommendation requirements. However, customers are mainly unregistered tourists, evaluation and transaction records are lack, and preference feature extraction can only rely on browsing records. The scarcity of the related information of the customer preference causes poor effect of the traditional recommendation algorithm such as the customer collaborative filtering, the commodity collaborative filtering, the singular value decomposition and the like based on the customer-commodity evaluation matrix. The embodiment of the invention provides a textile e-commerce recommendation system based on knowledge graph deep learning, which is used for identifying customers by using browsing end IP addresses according to the characteristic of the stability of browsing end of B2B E-commerce customers, so as to solve the identification problem of unregistered customers;

the correlation between commodities is established through commodity attributes and browsed features, and the defect of the information of the preferential relationship of customers is overcome; the recommendation performance is ensured by establishing a system framework supporting a plurality of knowledge map deep learning recommendation algorithms. The system comprises: the knowledge graph construction module is used for constructing a knowledge graph reflecting entity relations of commodities, shops, customers and the like based on commodity attributes, shop attributes and browsing logs of the textile electronic commerce platform after data extraction, conversion and loading. In particular, a browsing adjacency between goods is defined based on the browsing log, and a direct association between goods is established. The deep learning module builds a unified deep learning data set and an evaluation index system on the basis of the knowledge-graph relation triples, provides a RippleNet, KGCN, KGAT, CKAN and other knowledge-graph deep learning recommendation algorithm, carries out deep learning based on the knowledge graphs and the browsing logs, and outputs recommendation results reflecting browsing preferences of customers. And the application interface module receives the retrieval of the recommended result by the upper application and realizes the integration of the recommended application. The recommendation system can capture the preference characteristics of the customers on the commodities, support personalized commodity recommendation and improve commodity browsing efficiency of the customers on the electronic commerce platform.

Example 1

Referring to fig. 1, fig. 1 is a flow chart of an electronic commerce recommendation method based on knowledge-graph deep learning according to an embodiment of the present invention. The execution main body of the method described in the embodiment of the invention is an execution main body composed of software or/and hardware, and the execution main body can receive related information in a wired or/and wireless mode and can send a certain instruction. Of course, it may also have certain processing and storage functions. The execution body may control a plurality of devices, such as a remote physical server or cloud server and related software, or may be a local host or server and related software that performs related operations on a device that is located somewhere, etc. In some scenarios, multiple storage devices may also be controlled, which may be located in the same location or in different locations than the devices. As shown in fig. 1, the electronic commerce recommendation method based on knowledge map deep learning includes the following steps:

s101: receiving browsing information transmitted by an application interface end, wherein the browsing information is operation content at an electronic commerce platform of a corresponding user, and comprises commodity information and a browsing log; the browse log comprises browser ip address information; determining customer information based on the browser ip address information;

This step is mainly for acquiring basic browsing data. Specifically, the application interface mainly provides a recommendation function call for an application program, the recommendation function call is provided in the form of http service, the service program searches offline recommendation documents obtained through deep learning according to an application request, and the search result is responded to the recommendation result in a Json format. The interface details are designed as follows: the interface protocol is http get; the request parameters are UserId (user identification), model (Model name), code (return code), msg (return message) and data (return data), wherein the Model is enumerated in { RippleNet, KGCN, KGAT, CKAN, … }; the return code: with http definition, success is indicated as 200; and (3) returning a message: query success, or failure reasons, such as customer absence; returning data: (commodity Id, customer browse probability) list, output as Json object.

More preferably, fig. 2 is a schematic flow chart of filtering operation on browsing information according to an embodiment of the present invention, as shown in fig. 2, after receiving browsing information transmitted by an application interface, the method further includes:

s1011: performing data cleaning operation on the browsing information to filter non-commodity browsing records and commodity field missing browsing records in the browsing log;

S1012: and judging the validity of the browsing information subjected to the filtering operation, when the browsing times or browsing time of the same commodity by a customer exceeds a first set value, determining that the commodity is valid for the customer, and otherwise, filtering the corresponding browsing information.

The client-commodity browsing relationship is mainly constructed based on a browsing log, and the browsing relationship is expressed as (client IP, browsing, commodity Id) by taking the browser IP address as a client entity identifier on the assumption that the browsing log contains information such as the browser IP address, browsing time, browsing commodity Id and the like, so that the relationship between the client and the commodity is constructed. The problem corresponding to the non-registered user information is solved through the mode.

To enhance the effectiveness of the browsing relationship, a determination of effective browsing is added, and only if the number of browsing the commodity exceeds 2, the effective browsing is considered to be effective for the commodity by the customer, otherwise, the effective browsing is considered to be unintentional browsing for the commodity by the customer, and the value is not great for the recommendation. Firstly, invalid browses are filtered in a browse log, and then triples (customer IP, browses and commodity Ids) are generated for the valid browses. Through the steps, more accurate pushing operation on customers is realized.

S102: sending the browsing information to a pre-constructed commodity relation map model to extract entity relation so as to obtain a customer commodity interaction matrix; describing commodity attribute relations and commodity attribution relations by adopting a resource description framework in the commodity relation map model, wherein the commodity attribute relations are expressed in a form of triples of commodity numbers, attribute relations and attribute value numbers, and the commodity attribution relations are expressed in a form of triples of commodity numbers, attribution and store numbers;

in the embodiment of the invention, the textile electronic commerce platform is displayed on line, the off-line transaction mode is adopted, registered customers are sparse, transaction data and scoring data are sparse, and based on the consideration of the availability of the data, the recommendation system is mainly built on the basis of data such as shops, commodities, browsing logs and the like, the commodities are in attribution relationship with the shops, and the commodities are related through attribute entities and have no direct relationship.

The module firstly extracts data such as shops, commodities, browsing logs and the like from a data interface of the textile electronic commerce platform, then washes the data, and filters non-commodity (exhibition, shops, notification and the like) browsing records and browsing records with missing commodity Id fields in the browsing logs. And loading objects such as shops, commodities, browsing and the like into the program for entity conversion.

The entities in the knowledge graph mainly comprise: stores, merchandise attributes (e.g., composition, weave, use, etc.), customers, etc. Store and commodity object are directly converted into store and commodity entity in knowledge graph, for commodity attribute, each attribute is a kind of entity, each discrete value in the value domain of this kind of attribute can be converted into commodity attribute entity, for example, component attribute, can be extracted: cotton, wool, mulberry silk, jean, and the like.

In particular, for the extraction of customer entities, since the guest is mainly unregistered, information cannot be extracted from the customer member registry like a general e-commerce system, where it is indirectly extracted from the travel log. Because of the transaction characteristic of the B2B, the customer of the textile e-commerce platform is mainly a buyer of a clothing processing enterprise, one clothing processing enterprise possibly has a plurality of buyers, but the aim of browsing commodities is to purchase materials for the enterprise, and the e-commerce platform is mainly accessed through an enterprise network, so that the browsing end IP recorded in the browsing log is stable. Here, a customer is defined as a group of purchasing agents of a clothing processing enterprise, rather than a single purchasing agent, and is identified by using a browsing IP in a browsing log, where an IP represents a customer entity.

More preferably, fig. 3 is a schematic flow chart of the construction of a commodity relational graph model according to the embodiment of the present invention, and as shown in fig. 3, the commodity relational graph model is constructed by the following steps:

s1021: acquiring commodity attribute information at a corresponding electronic commerce platform;

s1022: extracting a plurality of commodity entities, commodity attributes and store entities from the commodity attribute information, and determining relations among the plurality of commodity entities, between the commodity entities and the store entities and between the commodity entities and the commodity attributes;

s1023: and constructing a commodity relation map model based on the relations among the commodity entities, the commodity entities and the store entities and the commodity attributes.

The commodity attribute relationship construction process specifically includes taking commodity, store and commodity attribute value as entity objects, constructing commodity attribute relationship and commodity-store attribution relationship based on loaded commodity information, wherein the commodity and store are both described by adopting unique Id as entity identification, the relationship expression adopts RDF standard, RDF (Resource Description Framework ) is a resource description language, and the object relationship and attribute are described in the form of (head entity, relationship, tail entity) or (entity, attribute value).

For commodity attribute relationship, commodity is a head entity, and the relationship is specific attribute, for example, commodity attribute of textile mainly comprises: product classification, composition, weave, use, color, elasticity, application scope, manufacturing process, etc., each attribute is a type of relationship, the tail entity is an attribute value, and each attribute entity is assigned a unique Id identifier. After processing, the commodity attribute relationship will be expressed as a triplet of (commodity Id, attribute relationship, attribute value Id). The assignment relation between commodities and stores is a one-to-many relation, namely, one store has a plurality of commodities, and one commodity only belongs to one store and is converted into a triplet form (commodity Id, belonging to store Id).

In the process of constructing the customer-commodity browsing relationship, the customer-commodity browsing relationship is mainly constructed based on a browsing log, and the browsing relationship is expressed as (customer IP, browsing, commodity Id) by taking the browser IP address as a customer entity identifier on the assumption that the browsing log contains information such as the browser IP address, browsing time, browsing commodity Id and the like, so that the relationship between the customer and the commodity is constructed. To enhance the effectiveness of the browsing relationship, a determination of effective browsing is added, and only if the number of browsing the commodity exceeds 2, the effective browsing is considered to be effective for the commodity by the customer, otherwise, the effective browsing is considered to be unintentional browsing for the commodity by the customer, and the value is not great for the recommendation. Firstly, invalid browses are filtered in a browse log, and then triples (customer IP, browses and commodity Ids) are generated for the valid browses.

More preferably, after extracting the plurality of commodity entities, commodity attributes, and store entities from the commodity attribute information, the method further comprises:

s1021a: determining browsing adjacency relations among commodities based on browsing relations of the commodities of the customers from the historical browsing information; the browsing adjacent relation is that if a plurality of commodities are browsed in one session period, the commodities are mutually browsed adjacent relation;

s1021b: when the number of times of the mutual browsing adjacency relationship reaches a second set value, the existence of the browsing adjacency relationship between the commodities is determined.

In order to enhance the direct connection between commodities, the browsing adjacency relationship between commodities is derived based on the browsing relationship between the customers and the commodities, namely, the customers u browse three commodities of va, vb and vc in a session period (such as 60 minutes), and then va, vb and vc are browsing adjacency relationship with each other. In order to reduce the accidental browsing adjacency, a filtering rule is further established, namely that the number of times of browsing adjacency reaches a certain threshold (such as 2 times) before the existence of the browsing adjacency between the two commodities is determined. Browsing adjacency shows potential combination relations among commodities, such as a customer has an order of jeans, he can browse jean fabric, copper buttons, leather decorations and the like repeatedly in a mall, and the commodities browsed once are likely to be combined into a batch of clothes, but no combination relation among the commodities exists in available data, and the defect can be overcome through derivation of browsing adjacency. Based on the above relation construction, the ontology model of the textile e-commerce knowledge graph can be generalized, as shown in fig. 6.

The browsing adjacency relation is a relation between commodities developed according to the browsing log, the direct connection between the commodities is realized, and the recommended performance when the browsing adjacency relation exists is compared by adopting a KGCN algorithm. The AUC and ACC indices are shown in fig. 14 and 15, respectively. In the figure, the browsing adjacency is identified as having no-consequence relationship, and the non-browsing adjacency is identified as having no-consequence relationship. From the above description and fig. 14 and 15, it can be seen that AUC and ACC with browsing adjacency are significantly better than those without browsing adjacency, which can also verify the effectiveness of browsing adjacency proposed by the embodiment of the present invention.

More preferably, the commodity relational graph model further comprises the following steps:

Through the steps, continuous updating and optimization of the commodity relation graph model is realized, so that a more accurate graph relation model is constructed, and more accurate data recommendation is performed.

S103: and transmitting the browsing information, the commodity interaction matrix of the customer and the commodity relation map model to a recommendation algorithm model for recommendation identification so as to determine commodity recommendation results of corresponding customers.

As shown in fig. 5, the deep learning module includes functional sub-modules such as deep learning dataset construction, deep learning recommendation algorithm library configuration, deep learning task starting, recommendation result export, and the like. The proposal of the embodiment of the invention further processes the commodity relation document and the content in the browsing relation document based on the common characteristics of the deep learning recommendation algorithm of the knowledge graph such as RippleNet, KGCN, KGAT, CKAN and the like, so that the proposal can adapt to the training of the deep learning algorithm.

The deep learning dataset is first defined as follows:

(1) Customer-commodity interaction matrix Y: assuming that the textile e-commerce recommendation scene comprises m customers U= { U1, U2, & gt, um } and n commodities V= { V1, V2, & gt, vn }, and a customer-commodity interaction matrix Y epsilon R can be obtained according to the browsing relationship document derived by the knowledge graph construction module ^m×n For the followingWhen a browsing relation triplet exists between u and v, y _uv =1, otherwise y _uv =0. Note that y _uv =0, does not mean that customer u is not interested in commodity v, perhaps because u fails to find v.

(2) Knowledge graph G: the customers, shops, commodities and their attribute values (such as specific merchants, product components, uses, etc.) in the knowledge graph are collectively referred to as entity set E, and the commodity attribute relationship and the customer-commodity browsing relationship are collectively referred to as relationship set R, then the knowledge graph can be expressed as g= { (h, R, t) |h, t E, R E R }. (h, r, t) a relational triplet, such as { knitting jacquard thousand-gauge, merchant, tellong } the merchant representing the commodity "knitting jacquard thousand-gauge" is "tellong", because there is duplication of commodity and merchant names, in the actual triplet expression, the entity is identified by id. The path between commodities can be achieved through commodity knowledge maps, for example, the path between commodities such as "colored cloud yarn" of another class of commodities such as "Styrofoam" can be achieved through { colored cloud yarn, merchant, style }, and the path between the commodity knowledge maps and "knitting jacquard thousand bird lattice". Note that during the graph traversal, edges that are made up of relationships are considered bi-directional edges.

The deep learning data set construction module mainly converts the content of the browsing relationship document into a customer-commodity interaction matrix Y according to the definition of the data set, and converts the content of the browsing relationship document and the commodity relationship document into a commodity knowledge graph G.

Based on the data set, the customer-commodity recommendation problem can be abstracted into a prediction problem of the customer-commodity interaction relationship, namely given Y and G, U epsilon U, V epsilon V, and prediction Y _uv . The task of the recommendation algorithm is to construct a prediction function model as shown in the formula (1).

in the formula ,i.e. the predicted value of the interaction relation (i.e. browsing) of u and v,Θ is the parameter set of the function f.

The deep learning recommendation algorithm library configuration, rippleNet, KGCN, KGAT, CKAN and other knowledge map deep learning recommendation algorithms are the generation programs of the function f in the formula (1), and are composed of neural networks with parameter sets theta, each algorithm has a specific neural network structure, and the parameter sets theta are obtained through deep learning.

The module mainly performs unified configuration on context parameters required by the running of the algorithms. Comprising the following steps: training a data path, namely a commodity relation document and an access path for browsing the relation document; a training result output path; super parameters of each algorithm, including: learning rate, number of layers of neural network, number of nodes of each layer of neurons, etc., and personalized parameters of each algorithm. Starting the deep learning task and deriving the recommendation result, starting the deep learning task according to a certain scheduling strategy (such as once per week) as required, generating a new customer-commodity browsing probability prediction function f after the completion of the learning task, and calculating the browsing probability of any (u, v) pair in Y based on the function The larger indicates the greater potential interest of customer u in commodity v. For each customer u, according to +.>Sorting the goods v and filtering y _uv The K top products are considered as a recommended product set of u if the products are =1. />

Because the size of the customers and the commodity is large, the combined number of the customers and the commodity is more than 10 hundred million, and the online calculation has obvious delay, the recommendation is performed by adopting an offline calculation mode, namely, the background calculation is performed according to the function f, the offline recommendation commodity list of all the customers is generated at one time, the result is stored as a text file, and the content format is as follows:

customer 1: commodity 11, commodity 12, commodity 13, …, commodity 1K;

customer 2: commodity 21, commodity 22, commodity 23, …, commodity 2K;

customer 3: commodity 31, commodity 32, commodity 33, …, commodity 3K;

more preferably, the recommendation algorithm model includes a prediction function model, wherein the prediction function model is:

where U is customer information, U is customer matrix, and u= { U1, U2, …, um }, V is merchandise information, V is merchandise matrix, and v= { V1, V2, …, vn }, Θ is a parameter set of function f>Browsing predicted values for u and v;

fig. 4 is a schematic flow chart of a recommended algorithm model construction disclosed in the embodiment of the invention, and as shown in fig. 4, the recommended algorithm model is constructed by the following steps:

S1031: expressing the entity and the relation in the commodity relation map model as d-dimensional embedded vectors through a random initialization vector or a Trans-embedding expression algorithm;

s1032: modifying the embedded vectors of entities and relationships by a domain information dissemination and aggregator to obtain a final domain representation vu; wherein, the linear combination of the neighbor information is utilized to describe the field information of the commodity;

s1033: defining a browsing probability prediction function based on the customer information u and the final field representation vu, and performing model calculation through a cross entropy function until a set condition is met to complete construction of a recommendation algorithm model, wherein the cross entropy function is as follows:

wherein yi represents the actual browsing relationship of (u, v) in the customer commodity interaction matrix Y, +.>For the predicted (u, v) pair browsing probability, u is customer information, and v is merchandise information.

In the embodiment of the invention, in order to further illustrate the working mechanism of the deep learning algorithm in the system framework, a KGCN (Knowledge Graph Convolutional Networks) deep learning algorithm is taken as an example, and a recommendation algorithm calculation process based on knowledge graph deep learning is illustrated.

In the embodiment of the invention, a data set customer-commodity interaction matrix Y and a knowledge graph G required by KGCN deep learning are constructed, and the basic idea of KGCN is as follows: the entity (including all entities such as customers, commodities, shops, commodity attributes and the like) and the relation in G are expressed as d-dimensional embedded vectors (d is set manually) by adopting a random initialization vector or a Trans-series graph embedded expression algorithm, and then the embedded vectors of the entity and the relation are corrected through neighborhood information propagation and aggregation. The entities involved in the calculation and the relation symbol are embedded vectors thereof unless otherwise specified below. The algorithm has the following two main characteristics: the local neighborhood structure can be better captured and stored in each entity by aggregating neighbor information; the aggregation weight between the neighbor entities depends on the specific customer u and the relation r, so that the semantic information of G is represented, and the personalized interest of the customer in the relation is also represented.

The algorithm structure is shown in fig. 7. As shown in formula (1), the input of the KGCN forward computing process is a customer-commodity pair (u, v), and the customer-commodity interaction matrix Y, the knowledge graph G and the algorithm parameter set Θ are output as the browsing probability of the customer u to the commodity v

Specifically, firstly, defining a preference score function of a customer u and a relation r, wherein the preference score function is shown as a formula (2) and is used for describing the preference degree of the customer on the commodity relation, such as the preference materials and the preference uses of the customer in a textile e-commerce platform; the specific preference formula is as follows:

in the formula ,the preference score of u to r is represented, g is a preference function, and an inner product function is generally selected, namely, the inner product of the embedded vectors of u and r is calculated, and the more similar the inner product value is, the higher the preference score is.

Then, the neighborhood information of the commodity v is characterized by using the linear combination of the neighborhood information, and is defined as a formula (3):

in the formula r_ve The relationship between the commodity v and the customer entity e is represented, wherein the entity in the entity G includes the commodity itself, and also stores, customers, commodity attribute entities and the like, and N (v) represents the set of all the entities directly connected to v.Is->The calculation is shown as a formula (4);

in order to avoid overlarge neighborhood information calculation pressure caused by the existence of too many neighbors of a certain node v, an over-parameter K can be added in the input parameters, and if the number of the neighbors of v exceeds K, only K neighbors are selected for calculation instead of all the neighbors, so that the calculation mode of each batch is kept stable.

The neighborhood representation of v at this time is denoted asAnd satisfies the condition shown in the formula (5).

s(v)→{e|e∈N(p)}，|S(p)|≤K (5)

The last step of the KGCN layer is to aggregate the vector representation of commodity node v and its neighborhood representation into a single vector, where v is the vector generated in the previous iteration of updating, and one of the following three aggregators may be selected for calculation.

Summing (sum), adding the two representation vectors, then performing a nonlinear transformation, as shown in equation (6),

splicing aggregation (concat), namely splicing two expression vectors, namely two vectors with the dimension d, splicing the two expression vectors with the dimension d to be 2d, and then performing nonlinear transformation, namely the expression (7):

neighbor aggregation (neighbor), representing the neighborhood of commodity v asAnd outputting after nonlinear transformation, as shown in formula (8):

in the above formula, σ () is a nonlinear activation function, for example, relu, sigmoid, W is a linear transformation matrix, b is a bias term, and is a basic element of a full connection layer of GCN, and is also a component part of an algorithm parameter set Θ, which needs to be obtained through learning.

For the above calculation process, there is an example as shown in fig. 8.

In the above diagram, assuming that k=2, in the KGCN model, the representation of the relationship in the knowledge graph G is a non-directional edge, in the diagram, the blue node is a commodity entity v, h=1 is a first-order neighbor thereof, h=2 is a second-order neighbor thereof, the blue arrow is the direction of information transfer, and the training process is a process that the commodity v continuously aggregates information of h-order neighbors thereof, and the first-order aggregation is shown in fig. 9.

Let k=And 4, green circles are neighborhood representations, and blue circles are aggregated representations. In the iterative process, for example, for the h+1th iteration, the vector representation of the entity e obtained in the h iteration is used as an initial value, the neighborhood representation of the current commodity v is updated and calculated, and then the neighborhood representation of the h iteration is aggregated with the neighborhood representation of the h iteration to obtain the neighborhood representation of the h+1th v, and the final neighborhood representation of v is assumed to be the vector v _u . Finally, then based on u and v _u Defining a browse probability prediction function, wherein a sigmoid function is adopted as the prediction function, as shown in a formula (9):

in model learning, the loss function is defined as a cross entropy function, as in equation (10):

in the formula ,y_i The (u, v) in the interaction matrix Y is represented as a 1, not 0,log is a logarithmic function for the predicted (u, v) versus browsing probability.

More preferably, the recommendation algorithm model is any one of a RippleNet recommendation model, a KGCN recommendation model, a KGAT recommendation model and a CKAN recommendation model; each recommendation model stores corresponding neural network parameters;

and/or the aggregator is any one of sum aggregation, splice aggregation and neighbor aggregation;

and/or, the recommendation method adopts an offline computing mode to conduct recommendation operation, background computation is conducted according to a pre-built knowledge graph and a recommendation algorithm model, offline recommendation lists of all customers are generated at one time, and the results are stored as text files;

And/or, in the recommendation method, the commodity relation graph model and the recommendation algorithm model are both deployed in a GPU deep learning server, and the application interface is deployed in a Web server. The three modules of knowledge graph construction, deep learning, application interfaces and the like can be independently deployed on different servers, and according to the calculation characteristics, the first two modules are suggested to be deployed in a GPU deep learning server, and the application interfaces are deployed in a Web server, and the deployment is shown in FIG. 11.

As shown in FIG. 10, customer u1 browses three products, i1, i2, i3, u2 browses two products, i3, i4, and u3 browses only one product, i 5. If the collaborative filtering recommendation algorithm is used, then i4 is recommended to u1, while i5 is not. From the blue dotted box, it can be seen that the relation r3 between i3 browsed by u1 and i5 not browsed points to e2, and i5 should have a certain recommendation value from the recommendation perspective. However, the baseline algorithm ignores more auxiliary information because of no embedding of the knowledge graph, and the recommendation effect is unreasonable.

The KGCN model combines the auxiliary characteristics of the knowledge graph, and has better effect on personalized recommendation. The model assigns a weight to each relationship based on customer attention, and it performs neighbor aggregation with bias based on this weight. As shown in the above graph, since u1 browses three products i1, i2 and i3, the first two products have a common relation (r 2), and the third product has a single relation (r 3), then u1 can be considered to be more conscious of the main use of the products, and then some products with e1 as attributes can be completely recommended to customers, such as product i4 in the graph, and the recommendation probability is high. The customer may be relatively less careful in the two relationships of r3, r4 for the product, and then a product i5 as shown in the figure may be recommended, but less forceful than the recommendation. For the complete knowledge graph, the relation among products is rich, the calculation flow of the prediction score is more complex, and the final recommended products are more diversified.

Because stores, commodities and browsing logs in the textile electronic commerce platform are dynamically updated, in order to ensure the recommendation effectiveness, a data update strategy is established, customers, commodities and browsing logs are synchronized regularly, and a list of the recommended commodities of the customers is updated in a periodic learning mode, such as once a week.

Based on certain light-weight electronic commerce network data, a knowledge graph construction module is operated, commodity attribute relation construction and customer-commodity relation construction are achieved, and an entity relation RDF triplet file is derived. In order to realize visual presentation of the knowledge graph, the RDF triples are imported into the Neo4j graph database example by using a tool, and then the knowledge graph is browsed by using a Neo4j graph presentation tool, and the example is shown in fig. 12.

A deep learning server is deployed with a RippleNet, KGCN, KGAT, CKAN and other knowledge map deep learning algorithm, logic according to data set construction, recommendation algorithm library configuration, task starting and recommendation result export is operated, a recommendation result is finally obtained, the performance of the algorithm is calculated, and the final recommendation performance of each algorithm is shown in table 1. Fig. 13 shows the variation of the recommended Accuracy (ACC) of each algorithm with the number of learning iterations (epochs).

Algorithm	AUC	F1	ACC
				CKAN	0.8629	0.8090	0.7657
KGAT	0.8766	0.8159	0.8241
				KGCN	0.8575	0.8100	0.7782
Ripplenet	0.8464	0.8011	0.8004

Table 1 recommended performance of deep learning algorithm

As can be seen from table 1, there is a difference in overall performance between depth algorithms, such as ACC of KGAT algorithm is 0.0584 higher than CKAN algorithm, indicating that it is necessary to use various recommended algorithms. From the curves, it can be seen that the convergence rate of each algorithm, KGAT, although having a slower convergence rate, can obtain good results.

The User collaborative filtering (User-CF), commodity collaborative filtering (Item-CF) and singular matrix decomposition model (BiasSVD) are adopted, the customer-commodity interaction matrix Y is taken as input to learn, and the obtained ACC indexes are shown in the table 2.

Algorithm	User-CF	Item-CF	BiasSVD
				ACC	0.32％	0.07％	10.52％

TABLE 2 recommended performance of conventional algorithms

It can be seen from the table that the recommendation accuracy of the three traditional algorithms is very low, and the difference from the deep learning algorithm is great. The method shows that the browsing log which only depends on information sparsity cannot realize effective recommendation, and the knowledge graph is used for constructing commodity information association and performing deep learning according to the commodity information association, so that the method is an effective way for solving the problem of information sparsity recommendation.

Example two

Besides the mode adopted by the embodiment, the characteristic extraction modeling can be further carried out, and the KGIN algorithm has two improvements compared with the algorithm of KGCN, KGAT, CKAN and the like which directly extracts commodity characteristics based on the node relation of the knowledge graph: (1) The relationship combination based on the knowledge graph defines the interactive intention of the customer to the commodity so as to obtain better model capacity and interpretability. (2) Customer-commodity relationships are considered in the context of more fine-grained relationship paths long-term semantics. Because of the scarcity of customer-commodity interaction information, customers have strong relevance to browsing different commodities. The application further improves the KGIN algorithm and provides a KGIN (KGNI with dual feature fusion, KGIN-DF) deep learning model with double feature fusion, and the main idea is shown in figure 16; in the drawing, two basic models for extracting features are shown in a virtual frame, the KGIN model is arranged on the upper left, the browsing adjacency model is arranged on the lower left, and the obtained features are processed into final customers and commodities through a concentration feature fusion mechanismFeature vectors.Representing a splicing operation->Representing inner product manipulation, ++>Representing matrix alignment multiplication>Representing softmax manipulation,/->Representing a dot product operation.

The algorithm firstly randomly initializes the entity and relation in the commodity knowledge graph KG and the customer into a high-dimensional real value initial representation vector generated by using a xavier_uniform algorithm, and in the following calculation, the entity and relation in the commodity knowledge graph KG, the customer u, the commodity attribute entity v and the relation r are respectively represented by ei, eu and ev, and the vector dimension is d. Then taking as input the initial representation vector eu, ei of the (u, i) pair, the interaction positive set of the (u, i) pair The attribution score is taken as output, and the embedded feature vectors of u and i are extracted through deep learning. The main ideas of the proposed KGIN-DF algorithm can be described as three parts:

(1) Combining the access log L with the commodity knowledge graph KG, and extracting the intention of a customer-commodity traffic individual graph by adopting the existing KGIN algorithm;

(2) And combining the access log L with the commodity browsing adjacency graph NG, and extracting the customer-commodity interaction group intention of the group by adopting graph aggregation operation.

(3) And (3) fusing the two parts of intention characteristics obtained in the steps (1) and (2) by using an attention mechanism to form fused characteristics, and evaluating the interest score of the customer-commodity. The (1) part multiplexes the basic KGIN algorithm, and the (2) and (3) parts are the expansion parts proposed by the application.

The interaction intention between the customers and the commodities is captured based on the commodity knowledge graph and the commodity browsing adjacency graph respectively, so that more accurate commodity recommendation is provided. It is desirable to achieve multi-granularity intent extraction of (u, i) through dual feature fusion, thereby improving accuracy of recommendations.

KGIN is one of recommendation algorithms based on knowledge-graph sensing, and KGIN is composed of two key parts: (1) Customer intent modeling, utilizing a plurality of potential intents to describe a customer-commodity relationship, and defining each intent as a fine-grained combination of relationships, while encouraging disagreement graphs to be independent of each other; (2) Relationship path aware aggregation emphasizes relationship dependencies in remote connectivity, thereby preserving the overall semantics of relationship paths. The reason why KGIN is selected in the present application is that modeling of relationship intent accords with the behavior of customers concerning the natural attribute relationship of textiles, for example, one customer purchases a certain cloth more seriously concerning its component attribute, while another customer focuses on factors such as weave and color. 17 shows the general flow of KGIN.

Node p in the figure ₁ ，p ₂ ,. it represents customer intention preferences, node i ₁ ，i ₂ ,. the representative interaction positive setCommodity of (a), node v ₁ ，v ₂ ,. it is a node in the knowledge graph, i.e. the attribute entity of the commodity. The KGIN splits each customer-commodity interaction record into a plurality of intent relationships, so that customer characteristics are extracted according to the disagreement graph. The dashed boxes represent customer and merchandise feature extraction formulas, respectively.

KGIN defines intent as the reason for customer selection of merchandise, reflecting the commonality of all customer behaviors. Assuming that P is the set of intentions shared by all customers, KGIN divides a unified customer-commodity relationship into |p| (|| representing the number of elements contained in the set, the same applies hereinafter), and decomposes each (u, i) pair into c= { (u, P, i) |p e P }, and the original customer interaction record is reorganized into a new heterogram called an intent graph IG. But identifyThe semantics of the intent are difficult, so KGIN correlates the intent with knowledge-graph relationships, defines the intent p of the customer as a nonlinear combination of different relationships in the knowledge-graph, embedding the representation e _p The following formula is shown:

in the formula ,e_r Is an initial representation vector of the relationship r, which is given to the attention score α (r, p) to quantify its importance. w (w) _rp Trainable weights, being specific relation r and specific intent p, are d-dimensional random initial vectors, w _r′p Representing other intent weights in the same relationship. It should be noted that the attention score α (r, p) is a unified intention modeling for the whole customer, and of course, the intention of the whole customer is independent, so that an independent modeling module is introduced to distinguish the intention of each customer, and common point mutual information, cosine distance and the like are distinguished.

In the intent diagram IG, N is used _u = { (p, i) | (u, p, i) ∈c } to represent intent awareness history and first order connectivity around customer u, technically embedded representation of customer with intent awareness information integration of history

in the formula ,is a characteristic representation of KGIN first-order path aggregation for customer u, a-> As indicated by the following formula, β (u, p) is the weight for distinguishing the intention of the individual customer:

wherein e_p ^T Representation e _p Is the transposed vector of (c), exp represents the exponential operation, e _p， ^T Representing a transpose of the other intent vector. The KGIN adopts a characteristic aggregation mode based on the relationship and the node, and specifically, the characteristic aggregation of the commodity i is shown in the following formula:

wherein ,N_i Is a triplet set associated with commodity i as a header entity, Then i corresponds to the tail entity v embedded representation, < >>

After the customer intention representation and the commodity characteristic representation of the first-layer convolution are obtained, the above steps can be repeated to obtain a deeper information representation, and multiple layers of information are aggregated to obtain a wider receptive field, so that the final commodity and the customer characteristic representation of the KGIN are shown in the following formula.

Wherein Q is the number of polymerization layers. After the neighbor relation graph NG is obtained, the browsing adjacent relation features of the customer and the commodity are obtained by adopting simple first-order aggregation, namely commodity features connected with the direct edges of the commodity are added and summed, and the calculation is shown in fig. 18. Node i in the figure ₁ ，i ₂ ,. it is indicated that browsing the goods in the adjacency graph NG, node u ₁ ，u ₂ ,. it represents a customer. The edges are the browsing adjacency relationship between the commodities and the access relationship between the customers and the commodities. The dashed boxes represent customer and merchandise feature extraction formulas.

The first order convolution is adopted because the first order (direct edge) intuitively represents the browsing adjacency condition of the commodity, and thus, the browsing adjacency feature e of the commodity i _i ^Nei The expression is as follows:

in the formula N_i ^Nei All commodity sets representing the browsing adjacency relationship with commodity i, j is N _i ^Nei Commodity e _j Is the initial representation vector for commodity j. While the browsing adjacent feature of the customer describes the preference of the customer according to the customer-commodity interaction relation, and the browsing adjacent feature e of the customer u _u ^Nei The following formula is shown:

wherein N_u Is a browsing commodity collection for customer u. The KGIN features representing the intent of the customer-commodity interaction individual were obtained separately at the early stage andBrowsing abutment feature representing customer-commodity interaction group intention +.> andThe part fuses the two parts to form +.> andCompared with the adjacent feature browsing, the KGIN feature has larger receptive field and wider information content because it is subjected to multi-purpose multi-level aggregation operation, but when the information content is too large, noise is introduced to weaken the effect of graph aggregation, so that a attention feature fusion mechanism is introduced, as shown in fig. 19.

First define a learning attention weight w _att The parameter vector technically splices the browsing adjacent feature and the KGIN feature in the same dimension, and the parameter vector is shown in the following formula:

e _u ^concat ＝e _u ^Nei ||e _u ^KGIN

e _i ^concat ＝e _i ^Nei ||e _i ^KGIN

in the formula ,e_u ^concat and e_i ^concat The splicing characteristics of the customer u and the commodity i are respectively represented, and the I represents vector splicing operation. It should be noted that the same attention weight applies to different sources in this section, i.e. the same attention weight is used for both commodity and customer feature fusion. Spliced feature vector e ^concat And w is equal to _att Multiplying and normalizing to select the weight key of key feature _weight As shown in the formula:

in the formula ,Representing a bitwise multiplication of the matrix with the vector. Note that the same attention weight w is used for both customer and merchandise features _att . Finally according to the key _weight Performing point multiplication on the spliced characteristic to obtain a fusion characteristic, wherein the fusion characteristic is represented by the following formulaThe following is shown:

e _u ^fusion ＝key _weight，1 ⊙e _u ^Nei +key _weight，2 ⊙e _u ^KGIN

e _i ^fusion ＝key _weight，1 ⊙e _i ^Nei +key _weight，2 ⊙e _i ^KGIN

wherein key_weight，1 Representing keys _weight Vector and key composed of previous d-dimensional elements _weight，2 Then it belongs to a vector of post d-dimensional elements. After the attention fusion feature is acquired, the inner product similarity is adopted for prediction, and the prediction is shown in the following formula.

Where x represents the convolution operation of the vector. During training, a general BPR loss training model is adopted, and the following formula is shown:

wherein Is composed of interaction positive set->And interaction negative set->The composed training set, σ (·) is a sigmoid function, combined with regularization to prevent model overfitting and gradient disappearance, the final loss function is shown in the following equation.

Where Θ is the parameter set of the model, e.gw _rp and w_att Lambda is control L ₂ Regularized hyper-parameters to prevent overfitting.

The study mainly compares KGIN-R with the following algorithm, which is recommended by using a knowledge graph: and (3) performing matrix decomposition on the customer-commodity interaction matrix by using an MF matrix decomposition and classical collaborative filtering method, and performing dimension reduction treatment on the matrix to be suitable for a large matrix, so as to obtain characteristic representations of customers and commodities respectively. (2) basic KGIN recommendation algorithm. And constructing a customer intention by utilizing the commodity knowledge graph KG, and deep learning and extracting intention characteristics by utilizing a customer-commodity access log L, wherein a commodity browsing adjacency graph NG does not participate in the algorithm model.

A full ranking strategy is recommended, specifically, for each customer in the dataset, the goods that the customer has effectively browsed are grouped into a positive set, and the rest are grouped into a negative set. The recommendation algorithm will score each (u, i) pair, select the K products with the highest scores for one customer u to form a recommendation list, and then calculate the Top-K recommendation performance. The study employed widely used evaluation criteria including: recall, normalized impairment accumulation gain NDCG (Normalized Discounted Cumulative Gain), and Hit rate HR (Hit Ratio). Recall represents the hit rate of the recommended products in the test set, HR can be considered as the hit rate of the customer, and the hit condition of the recommended products is expressed from the whole representation, namely, the model is considered to have the capability of recommending the preferred products for the customer as long as the recommended products are in the positive set, NDCG considers the attraction of the commodity ranking position of the recommended list to the customer, and under the condition of the same Recall, the higher the commodity ranking of the test set is, the higher the NDCG score is. Here, K is set to 1, 2, 10, and 20, respectively, and then an average value of all customer evaluation indexes is taken as a final evaluation index.

The "boost" of the algorithm column in the table represents the index boost percentage of the KGIN-DF algorithm relative to the KGIN algorithm, the optimal results in each evaluation index are indicated in bold, and the suboptimal results are indicated by underlining. It can be seen that the KGIN-DF model is optimal except for the Recall index when k=1, and the rest indexes are all optimal, while the MF model is worst and significantly lower than the KGIN-DF and KGIN models.

The performance advantages of KGIN-DF over KGIN can be attributed to: (1) The method comprises the steps of (1) extracting intention characteristics of a customer group based on a browsing adjacency relationship and (2) fusing double characteristics based on an attention mechanism, wherein the browsing adjacency relationship utilizes a customer collaborative filtering idea to induce direct connection between commodities from a customer-commodity browsing log, so that individual intention of customers can be perceived in advance according to the intention of the customer group for commodity access, and a more effective recommendation list is provided for the customers. From the NDCG index, the KGIN-DF achieves a significant improvement of 14.03 to 23.56 percent compared with the KGIN, and the index aims at the ranking score of the recommended commodities and can more comprehensively reflect the effectiveness of the recommendation list. The KGIN model is not used for extracting the collaborative features of the adjacent graphs, and the recommendation is biased to the individual interaction intention of the commodity knowledge graph depiction, so that the sequencing result is poorer than KGIN-DF, and the effectiveness of KG and NG dual-feature collaborative extraction and fusion provided by the application on improving the recommendation performance is further verified by combining.

The performance of the MF in experiments is poor, because the matrix decomposition model cannot fully utilize the relation of the knowledge graph to optimize the characteristic representation of the commodity, the characteristic of the commodity is built only by means of the access log L, when the customer-commodity interaction information is sparse, the extracted customer characteristic is incomplete, and the recommendation list is difficult to meet the customer intention. This also verifies the validity of the commodity knowledge graph KG in the recommendation.

From the analysis, it can be concluded that: (1) The commercial knowledge graph KG and KGIN algorithm constructed by the application can effectively extract the customer visit intention characteristics of the B2B textile electronic commerce platform; (2) The commodity browsing adjacency graph NG constructed by the application can further improve the recommendation performance of the B2B textile electronic commerce platform by combining the proposed improved algorithm KGIN-DF.

Example III

Referring to fig. 20, fig. 20 is a schematic structural diagram of an electronic commerce recommendation system based on knowledge-graph deep learning according to an embodiment of the present application. As shown in fig. 20, the electronic commerce recommendation system based on knowledge-graph deep learning may include:

the receiving module 21: the method comprises the steps of receiving browsing information transmitted by an application interface end, wherein the browsing information is operation content of an electronic commerce platform of a corresponding user, and comprises commodity information and a browsing log; the browse log comprises browser ip address information; determining customer information based on the browser ip address information;

entity extraction module 22: the method comprises the steps of sending browsing information to a pre-constructed commodity relation map model to extract entity relations so as to obtain a customer commodity interaction matrix; describing commodity attribute relations and commodity attribution relations by adopting a resource description framework in the commodity relation map model, wherein the commodity attribute relations are expressed in a form of triples of commodity numbers, attribute relations and attribute value numbers, and the commodity attribution relations are expressed in a form of triples of commodity numbers, attribution and store numbers;

Recommendation module 23: and the browsing information and the commodity interaction matrix of the customer are transmitted to a recommendation algorithm model to conduct recommendation identification so as to determine commodity recommendation results of corresponding customers.

The electronic commerce recommendation method, system, electronic equipment and storage medium based on knowledge-graph deep learning disclosed by the embodiment of the invention are described in detail, and specific examples are applied to the description of the principle and the implementation mode of the invention, and the description of the above embodiments is only used for helping to understand the method and the core idea of the invention; meanwhile, as those skilled in the art will have variations in the specific embodiments and application scope in accordance with the ideas of the present invention, the present description should not be construed as limiting the present invention in view of the above.

Claims

1. The electronic commerce recommendation method based on the knowledge graph deep learning is characterized by comprising the following steps of:

2. The electronic commerce recommendation method based on knowledge-graph deep learning of claim 1, wherein the recommendation algorithm model comprises a prediction function model, and the prediction function model is constructed by the following steps:

wherein ,representation->F represents a predictive function, U is customer information, U is a customer matrix, and u= { U1, U2, …, um }, I is merchandise information, I is a merchandise matrix, and i= { I1, I2, …, in }.

3. The method for electronic commerce recommendation based on knowledge-graph deep learning of claim 1, further comprising, after said sending the browsing information to a pre-built commodity relationship graph model for entity relationship extraction to obtain a customer commodity interaction matrix:

4. The method for e-commerce recommendation based on knowledge-graph deep learning of claim 3, wherein combining the browsing information with a commodity relationship graph model and extracting the intent of the customer-commodity interaction individual by using a KGIN algorithm to obtain a corresponding individual intent set comprises:

wherein ,e_r Is an initial representation vector of the relation r, which is given to the attention score alpha (r, p) to quantify the importance thereof, w _rp Trainable weights, being specific relation r and specific intent p, are d-dimensional random initial vectors, w _r′p Other meanings in the same relationGraph weights;

in the intent diagram IG, N is used _u = { (, i) |(, p, i) ∈c } to represent the intent awareness history and first order connectivity around customer u, the embedded representation of the customer is represented by an embedded formula with intent awareness information integration of the history:

wherein Q is the number of polymerization layers;

wherein ,N_u Is a browsing commodity collection for customer u.

5. The method for e-commerce recommendation based on knowledge-graph deep learning of claim 4, wherein employing an attention mechanism to feature-fuse the individual intent features and the group intent features to form fused features, and evaluating a customer's interest score in a commodity based on the fused features, comprises:

e _u ^concat ＝e _u ^Nei ||e _u ^KGIN

e _i ^concat ＝e _i ^Nei ||e _i ^KGIN

6. the method for recommending e-commerce based on knowledge-graph deep learning as set forth in claim 1, further comprising, after receiving the browsing information transmitted from the application interface, the steps of:

7. The electronic commerce recommendation method based on knowledge-graph deep learning of claim 1, wherein the commodity relation graph model is constructed by the following steps:

8. The method for e-commerce recommendation based on knowledge-graph deep learning of claim 7, further comprising, after extracting a plurality of commodity entities, commodity attributes, and store entities from the commodity attribute information:

9. An electronic commerce recommendation system based on knowledge graph deep learning is characterized by comprising:

10. A computer-readable storage medium storing a computer program, wherein the computer program causes a computer to execute the electronic commerce recommendation method based on knowledge-graph deep learning of any one of claims 1 to 8.