An e-commerce platform commodity recommendation method based on interest preferences

Abstract

The application belongs to the technical field of commodity recommendation management, and particularly relates to an e-commerce platform commodity recommendation method based on interest preferences; the method comprises the following steps: obtaining user data, and pre-processing the user data; extracting the pre-processed user data and inputting the same into an e-commerce platform commodity recommendation model based on interest preferences to obtain a commodity recommendation result for the user, and the e-commerce platform recommends commodities to the user according to the commodity recommendation result; the present application compensates for explicit feedback information, extracts hidden information, and fuses the explicit feedback information and the hidden information, so that the present application can more accurately capture the result of the user's interest preferences for commodities, stimulate the user's consumption desire, and has high practicability.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of commodity recommendation management, and particularly relates to an e-commerce platform commodity recommendation method based on interest preferences. BACKGROUND

[0002] With the popularity of mobile terminal devices and the vigorous development of the Internet, people's life gradually enters the information age, and the e-commerce platform is inseparable from modern life. People benefit from Internet information, and with the explosive growth of e-commerce commodity data, the information overload problem is becoming increasingly serious, but these redundant data that no one cares about often have hidden value. If the merchant can actively recommend the goods that meet the user's fuzzy needs to the user from the massive commodity information, it is expected to convert potential demand into actual demand, which not only can improve the sales volume of the e-commerce website, but also helps to meet the actual needs of users and improve the loyalty of users to the platform. Therefore, how to extract the information interested by the user from the massive data of the platform has become the focus of the e-commerce platform.

[0003] The recommendation algorithm is a tool to help users quickly find useful information. Unlike search engines, the recommendation algorithm does not require the user to provide explicit requirements, but analyzes the user's historical behavior as the user's interest, and thus actively recommends information that meets the user's interest and demand. Nowadays, the recommendation algorithm has become more and more important. On the one hand, the user's demand has uncertainty and fuzziness. According to Amazon statistics, only 16% of customers who shop on its website have clear purchase intentions. On the other hand, the recommendation algorithm plays an important role in the market competition of enterprises, and it has been widely used in various fields such as e-commerce (such as Amazon, Alibaba, etc.), information retrieval (such as iGoogle, Baidu, etc.), social networks, location services, news push (such as Google News, Toutiao, etc.). Therefore, the recommendation algorithm has attracted the attention of many researchers.

[0004] With the continuous expansion of e-commerce platform data and the gradual maturity of deep learning technology, nonlinear models such as deep learning have gradually become the mainstream of e-commerce platform recommendation system models. For example, Liu Y et al. (Liu Y, Yang S, Xu Y, et al. Contextualized graph attention network for recommendation with item knowledge graph[J]. IEEE Transactions on Knowledge and Data Engineering, 2021.) proposed a new graph neural network based on user interest preference, which explicitly uses the local and non-local graph context information of entities in the user-item graph. This graph attention mechanism captures user interest preferences for goods based on local information. The paper shows that the analysis method of capturing user interest preferences using graph neural networks can effectively analyze explicit feedback information, but it lacks compensation for explicit feedback behavior and the fusion of hidden information, making the result of capturing user interest preferences for goods not accurate enough, the recommendation result not accurate enough, and the user's desire for consumption not strong enough. SUMMARY

[0005] Current e-commerce platform recommendation model research still has some problems in improving the accuracy and interpretability of the recommendation model:

[0006] 1. Insufficient explicit feedback information on e-commerce platforms. Compared with the large amount of implicit feedback data generated by users, the amount of original explicit feedback information from users to goods is insignificant, making it difficult to accurately describe user interest preferences.

[0007] 2. Hidden relationships are complex and diverse in heterogeneous information space. The relationship between users and goods in the heterogeneous information network under the e-commerce platform is complex and diverse, and relying solely on explicit feedback data is not enough to accurately describe the user's true characteristics, resulting in a rough expression of the user's hidden information by the model.

[0008] 3. Differences in interest preference degree (difficult to accurately represent). User interest preferences are influenced by factors such as gender, age, and price of goods, so it is obviously a problem to quantify the degree of interest preference from the characteristics of the user and the neighboring nodes

[0009] To address the deficiencies in the prior art, the present application proposes an e-commerce platform goods recommendation method based on interest preference, which includes: obtaining user data, preprocessing the user data; inputting the preprocessed user data into an e-commerce platform goods recommendation model based on interest preference to obtain a goods recommendation result for the user, and the e-commerce platform recommends goods to the user according to the goods recommendation result;

[0010] The process of processing data using an interest-based e-commerce platform product recommendation model includes:

[0011] S1: Obtain explicit feedback information and extract relevant attributes based on user data;

[0012] S2: The explicit feedback information is compensated by a differential graph autoencoder to obtain the compensated feedback information adjacency matrix;

[0013] S3: Use a deep factorization machine to extract hidden information from user data and obtain user feature representations;

[0014] S4: A graph attention neural network is used to process the relevant attributes, the adjacency matrix of the feedback information, and the user feature representation to obtain the user's product recommendation results.

[0015] Preferably, the relevant attributes include user attributes, user activity level, interest preferences, basic product attributes, and interest incentive factors.

[0016] Furthermore, the formula for calculating user activity is:

[0017] ActDeg(u i )=μ×Cartcount(u i )+Orders(u i )

[0018] Among them, ActDeg(u i ) represents the i-th user u i User activity level, μ represents the penalty coefficient, Cartcount(u i ) represents the i-th user u i The number of items added to the cart, Orders(u i ) represents the i-th user u i The number of orders placed.

[0019] Furthermore, the formula for calculating interest preference is:

[0020]

[0021]

[0022] Among them, PerferDeg(l i→k ,u i ,c k ) represents the i-th user u i For the k-th item c k interest preference, u pi Represents the i-th user u i Interest preferences are represented by meg(u i ,ck represents embedding the effective information in the kth commodity c k in the embedding representation of the user u i , m represents the number of commodities.

[0023] Preferably, the process of compensating the explicit feedback information by using the differential graph autoencoder includes: the differential graph autoencoder includes an encoder and a decoder; obtaining an explicit feedback information matrix according to the explicit feedback information; mapping the explicit feedback information matrix to a latent variable feature space by using the encoder to obtain a latent feature representation; and mapping the latent feature representation to the original sample space by using the decoder to obtain a compensated feedback information adjacency matrix.

[0024] Further, the formula for calculating the compensated feedback information adjacency matrix is:

[0025]

[0026] wherein, represents the compensated feedback information adjacency matrix, ActDeg represents the user activity, σ represents the variance of the latent feature representation, and Z represents the latent feature representation.

[0027] Preferably, the process of extracting hidden information of user data by using the extremely deep factorization machine includes: the extremely deep factorization machine includes a CIN and a DNN;

[0028] The user's own attributes, interest incentive factors and commodity basic attributes are one-hot encoded and embedded to obtain original features of the user data;

[0029] The original features of the user data are processed by using the DNN to obtain implicit feature interactions;

[0030] The original features of the user data are processed by using the CIN to obtain explicit feature interactions;

[0031] The implicit feature interactions and the explicit feature interactions are weighted and aggregated to obtain user feature representations.

[0032] Preferably, the process of processing the related attributes, the feedback information adjacency matrix and the user feature representations by using the graph attention neural network includes:

[0033] The internal factor value is calculated according to the user activity and the user's own attributes; the external factor value is calculated according to the commodity basic attributes and the interest incentive factors, and the external factor is taken as the commodity node feature;

[0034] The internal factor value and the user feature representation are aggregated to obtain a user node feature;

[0035] An attention graph neural network is adopted to calculate the attention degree according to the feedback information adjacency matrix, user node features and commodity node features;

[0036] The attention degree is normalized to obtain an attention coefficient;

[0037] The recommendation index of commodities to users is calculated according to the attention coefficient and the interest preference degree of the users; the recommendation indexes are sorted, and the first N commodities are taken as the commodity recommendation results of the users.

[0038] Further, the formula for calculating the internal factor value is:

[0039] FactorsIn(u i )=ActDeg(u i )×User(u i )

[0040] Wherein, FactorsIn(u j ) represents the internal factor value of the i-th user u i , ActDeg(u i ) represents the user activity of the i-th user u i , and User(u i ) represents the user attribute of the i-th user u i .

[0041] The formula for calculating the external factor value is:

[0042] FactorsOut(u i )=Enc(u i ,c k )×Iattri(c k )

[0043] Wherein, FactorsOut(u i ) represents the external factor value of the i-th user u i , Enc(u i ,c k ) represents the interest incentive factor of the i-th user u i and the k-th commodity c k , and Iattri(c k ) represents the basic attribute of the k-th commodity c k .

[0044] Further, the formula for calculating the recommendation index of commodities to users is:

[0045]

[0046] Wherein, represents the recommendation index, This represents the attention coefficient of the i-th user to the k-th item, PferDeg(l) i→k ,u i ,c k ) represents the i-th user u i For the k-th item c k Interest preferences.

[0047] The beneficial effects of this invention are as follows: This invention analyzes user data from e-commerce platforms, uses a differential graph autoencoder to compensate explicit feedback information to obtain compensated explicit feedback information, employs a deep factorization machine to extract hidden information from user data, and finally obtains product recommendation results for users by fusing explicit feedback information and hidden information. Compared with existing technologies, this invention introduces rich explicit feedback data and hidden information from e-commerce platforms, which can predict the top N products that users are most interested in on the e-commerce platform and recommend products to users based on the prediction results. At the same time, it can also grasp individual consumption habits and specific needs, which helps e-commerce platforms achieve better profitability and improve user satisfaction with the e-commerce platform, making it highly practical. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the product recommendation method for e-commerce platforms based on interest preferences in this invention;

[0049] Figure 2 This is a schematic diagram of the explicit feedback information compensation process in this invention;

[0050] Figure 3 This is a schematic diagram illustrating the extraction of hidden information using the deep factorization machine in this invention. Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] This invention proposes a product recommendation method for e-commerce platforms based on interest preferences, such as... Figure 1 As shown, the method includes the following:

[0053] Acquire user data and preprocess it; extract the preprocessed user data and input it into the product recommendation model of the e-commerce platform based on interest preferences to obtain product recommendation results for users; the e-commerce platform recommends products to users based on the product recommendation results.

[0054] The user data can be obtained from a public e-commerce data website or directly querying the real-time user behavior data in the database provided by the e-commerce enterprise. Here, what needs to be obtained is the explicit feedback information of the platform users in a certain life cycle (such as clicking, ordering, rating, returning, collecting, etc.), the basic attribute data of the goods with which the user has interactive behavior (including category, price, release time, etc.), and the user's own basic attribute data (age, gender, occupation and region) and other data.

[0055] The user data is preprocessed, which includes simple cleaning of the obtained data. Generally, the original data obtained is unstructured and cannot be directly used for data analysis. Through simple data cleaning, most of the unstructured data can be structured. For example, deleting data items with abnormal values or control, cleaning invalid information, etc.

[0056] The preprocessed user data is input into the e-commerce platform product recommendation model based on interest preference, and the process of the e-commerce platform product recommendation model based on interest preference includes the following contents:

[0057] S1: Obtain explicit feedback information and extract related attributes according to user data; the related attributes include user's own attributes, user's activity, interest preference, product basic attributes, and interest incentive factors.

[0058] The explicit feedback information is obtained according to the explicit feedback information of the user in a certain life cycle:

[0059] The explicit feedback information is represented as:

[0060] The specific meaning is the interaction relationship between the user and the product, s nx represents the explicit feedback information value of the xth user and the nth product, i.e. the value of different interaction types (click, order, rating, return, collection). Preferably, the explicit feedback information values of clicking, ordering, rating, returning and collecting are 1, 2, 3, 4 and 5 respectively.

[0061] According to the user data, the related attributes are extracted:

[0062] The user's own attributes User(u i ):

[0063] The user's own attributes include age, gender, occupation and region, which have certain correlation with the generation of user interactive behavior. Therefore, the user's own attributes are defined as:

[0064] User(u i )=[age(u i ),gender(u i ),occup(ui ), zipCode(u i )

[0065] where age(u i ) is the age of user u i , gender(u i ) is the gender of user u i , and occup(u i ) and zipCode(u i ) are the occupation and the zip code of user u i , respectively.

[0066] User activity degree ActDeg(u i ):

[0067] In the e-commerce platform, the activity degree of a user can reflect the trust and satisfaction of user u i to the platform to some extent. The more active the user is, the more likely it is to generate certain consumption behavior. Therefore, the user activity degree is defined as:

[0068] ActDeg(u i ) = μ × Cartcount(u i ) + Orders(u i )

[0069] where Cartcount(u i ) represents the number of added-to-cart of user u i in the e-commerce platform, and Orders(u i ) represents the number of orders of user u i in the e-commerce platform. Since users with a large number of added-to-cart in the shopping cart may not generate consumption behavior, a penalty coefficient μ is added to balance the influence of the added-to-cart number, μ ∈ [0, 1], and the penalty coefficient is valued according to the average experience of the historical behavior of users on the e-commerce platform, preferably μ = 0.3.

[0070] Interest preference degree PerferDeg(l i→k , u i , c k ):

[0071] For user u i , the effective information of product c k in the explicit feedback information of this user, such as the category and price of c k , is embedded into the interest preference representation u i of user u pi , and the similarity between product c k and u pi is calculated to approximate the expression of user ui For product c k Interest preference refers to the potential motivation for users to make purchases based on the higher the similarity between products. Therefore, interest preference is defined as:

[0072]

[0073]

[0074] Among them, PerferDeg(l i→k ,u i ,c k ) represents the i-th user u i For the k-th item c k interest preference, l i→k Indicates user u i And product c k There is an interest relationship, u pi Represents the i-th user u i Interest preferences are represented by meg(u i c k ) indicates that the k-th item c k Effective information (product category, price, etc.) is embedded into the user's... i In the embedded representation, m represents the number of products for which the user has performed actions of interest such as clicking, placing an order, rating, returning, or adding to favorites.

[0075] Product basic attributes Iattri(c k ):

[0076] On e-commerce platforms, product attributes such as type, price, and release time not only indirectly reflect a person's preferences or needs but are also indispensable for product embedding. Therefore, the basic attributes of a product are defined as follows:

[0077] Iattri(c k ) = [typeAtt(c k ), priceAtt(c k ), timeAtt(c k )]

[0078] Among them, typeAtt(c k ) indicates product c k The type, priceAtt(c k ) indicates product c k The price, timeAtt(c k ) indicates product c k Release date.

[0079] Interest motivator Enc(ui , c k ):

[0080] Different feedback information of users reflects the income value of different goods in the e-commerce platform on the user, which invisibly affects the psychological trade-off of the user when facing the goods, and the platform will benefit from it when the scale tilts to the consumption behavior side. This income value is called interest incentive factor:

[0081]

[0082] Among them, isBuy represents whether the user u i buys the goods c k , and the value of isBuy is 5 if the goods are purchased; isCollect represents whether the user u i collects the goods c k , and the value is 3 if the goods are collected; isBrowse represents whether the user u i browses the goods c k , and the value is 2 if the goods are browsed; otherwise, the three are 1; and w is the normalization factor of the interest incentive factor, preferably w is 10.

[0083] S2: Adopting a differential graph autoencoder to compensate the explicit feedback information to obtain the compensated feedback information adjacency matrix.

[0084] In the recommendation system of the e-commerce platform, compared with a large number of goods, the explicit feedback data of the user often has the characteristics of data sparseness, which is easy to cause the problem of inaccurate prediction behavior. Therefore, it is very necessary to compensate for hidden, missing and other data to generate some meaningful data to make up for the lack of sample space before predicting the user behavior. In the user-goods space, the expression of user interest preference has a profound influence on the feedback behavior of the user, but in the recommendation scene of the e-commerce platform, the potential behavior information of the user behavior is not considered, and part of the potential behavior information plays a key role in the quantification of the interest preference. Therefore, the present application refers to the idea of evolutionary game and proposes a compensation method for compensating the explicit feedback information of the user based on the differential graph autoencoder. The probability distribution of the co-occurrence matrix in the collaborative filtering of the differential graph autoencoder VGAE is used to complete the self-encoding of the hidden vector of the goods node or the user node, and then the link prediction is made according to the hidden vector representation, so as to achieve the effect of information completion. The specific process is shown in Figure 2 .

[0085] The VGAE saves important features of the input vector in the reconstruction function, and obtains the latent feature representation of the sample by minimizing the reconstruction error. Meanwhile, the parameter amount of the reconstruction function is much smaller than that of the input vector, and the autoencoder can also play a role in dimension reduction and data compression. Since the autoencoder has a certain degree of freedom, it can achieve the effect of accurate dimension prediction after the "generalization" process of the autoencoder, so as to obtain good user explicit feedback information sample data and alleviate the problem of insufficient effective explicit feedback data.

[0086] The VGAE is similar to the autoencoder (AE) and is also composed of an encoder E and a decoder D. The explicit feedback information matrix is obtained according to the explicit feedback information; the encoding process maps the input sample, i.e., the explicit feedback information matrix of the user, to the latent variable feature space through E to obtain the latent feature representation Z; the decoding process maps the latent feature representation to the original sample space through D to obtain the compensated sample.

[0087] Suppose the reconstruction function including the encoding and decoding stages in the VGAE is r(e; s), then the total optimization objective function is as follows:

[0088]

[0089] Wherein, S represents the set of all edges in the user-goods bipartite graph, s represents the element in the explicit feedback information matrix, i.e., the explicit feedback information value, and e represents the output after encoding and decoding, i.e., the compensated explicit feedback information value.

[0090] Solving the objective function can obtain the compensated feedback information adjacency matrix, and the element in the compensated feedback information adjacency matrix is the compensated explicit feedback information value e. The solving process is as follows:

[0091] In the VGAE, the encoder q(Z|X, A) is calculated as follows:

[0092]

[0093] Wherein, A represents the explicit feedback information matrix, and the element is the explicit feedback information value of the user and the commodity interaction; X represents the node feature value, Z represents the latent feature representation, Q represents the number of input samples, N() represents the Gaussian distribution, z i represents the ith latent feature representation obtained according to the mean and variance, μ i represents the mean of the ith sample input, represents the variance of the ith sample input.

[0094] As shown in Figure 2 In the VGAE, the mean μ and the variance σ of the Gaussian distribution of the latent feature vector are calculated by using the advantage of the graph convolutional neural network GCN to aggregate the neighbor node information:

[0095]

[0096] μ = GCN μ (X, A)

[0097] logσ = GCN σ (X, A)

[0098] where GCN * (X, A) denotes the graph convolutional computation formula of mean or variance, D' denotes a degree matrix, W0 denotes an optimization parameter of μ, W1 denotes an optimization parameter of σ, ReLU denotes an activation function, GCN μ (X, A) denotes mean computation using a convolutional network, GCN σ (X, A) denotes variance computation using a graph convolutional network.

[0099] Although the autoencoder has certain "generalization ability", there is still overfitting problem in the encoding and decoding game process, so that the explicit feedback information is invalid compensation. Analogous to the idea of game theory, the description process of user interest preference is inseparable from the activity of the user on the platform, and the interest preference of the user stimulates the interaction behavior of the user and the goods. The higher the user activity is, the more explicit feedback information is, and the more explicit the interest preference is; on the contrary, the user with low activity is easy to cause interest preference expression deviation and recommend invalid goods due to lack of explicit feedback data, further leading to the decline of user activity. The present application regards the description of user interest preference as the selection of interested goods by the user, and proposes an EI-VGAE (explicit feedback information variational graph autoencoder) method, which integrates the user activity ActDeg into the VGAE, and the normalized ActDeg is used as a decoding factor to reduce the influence of insufficient explicit feedback information on the description deviation of user interest preference.

[0100] According to the user activity and the implicit feature representation, the compensated feedback information adjacency matrix is calculated, and is represented as:

[0101]

[0102] wherein, is the compensated feedback information adjacency matrix, and ActDeg is the user activity.

[0103] Cross entropy and KL divergence are used as loss functions:

[0104]

[0105]

[0106]

[0107]

[0108] S3: Extracting hidden information of user data by using an eXtreme deep factorization machine to obtain a user feature representation.

[0109] In a recommendation system, understanding the implicit feature interaction behind the user interaction behavior is of great significance to the prediction of recommended commodities, but the basic features extracted by hand often have the phenomenon of lack of expression ability and high cost. Since the original features rarely can achieve the optimal effect, in order to clearly mine the deeper feature interaction in the sparse interaction information network and accurately express the hidden driving factors of the user interaction behavior, the present application introduces an eXtreme deep factorization machine (xDeepFM) to learn the mutual relationship between user and commodity multi-features, and optimizes the method by considering the influence of time on the user interaction behavior.

[0110] As shown in Figure 3 , the xDeepFM mainly includes two parts: CIN (compressed interaction network) and DNN (deep neural network), which jointly learn multi-order interaction features in an explicit and implicit manner to extract hidden information of user data. After one-hot encoding, embedding representation processing of user's own attributes, interest incentive factors and commodity basic attributes, the original features of user data are obtained.

[0111] The original features of user data are processed by using the DNN to obtain implicit feature interaction. The DNN implicitly generates bit-level feature interaction from data, and the formula is as follows:

[0112] x 1 =δ(w 1 x 0 +b 1 )

[0113] x k =δ(w k x k-1 +b k )

[0114] Wherein, x 0 is the original feature embedding representation of user data, x k-1 is the output vector of the k-1 layer in the network, w k and b k are the first training parameter and the second training parameter of the k layer respectively, and δ is an activation function.

[0115] The final output vector of the DNN is linearly integrated into the original features, as shown below:

[0116]

[0117] where a is a linear weight, ω k+1 is the parameter of the k+1th hidden layer interaction output, represents the output of the DNN, i.e., the implicit feature interaction.

[0118] The CIN is used to process the original features of the user data to obtain explicit feature interaction. The CIN is responsible for constructing a limited order feature combination explicitly. Each layer of the multiple hidden layers of the CIN needs to be calculated according to the previous hidden layer and the original features, and the output calculation formula of the kth layer is as follows:

[0119]

[0120] where, represents the output matrix of the kth layer of the interaction neural network, and the input of the k-1th layer is m represents the number of original features, H k represents the number of neurons in the kth hidden layer, and the number of features in each layer is H i are not the same; X 0 ∈R m×D represents the original feature input, obtains H k *m D-dimensional features, i.e., a three-dimensional feature tensor, represents the Hadamard product, which calculates the product of the corresponding elements of two vectors, is a convolution kernel. The two-dimensional tensor is compressed to a one-dimensional tensor layer by layer through the convolution kernel, and the compression formula is as follows:

[0121]

[0122] where D represents the dimension of the two-dimensional tensor.

[0123] Finally, multiple one-dimensional feature vectors are spliced to obtain the output q of the CIN part, i.e., the explicit feature interaction, and the calculation formula is as follows:

[0124]

[0125] The implicit feature interaction and the explicit feature interaction are weighted and aggregated to obtain the user feature representation: the weighted output of the output representation of all hidden layers in the CIN and the output of the DNN is taken as the output result, i.e., the explicit hidden information is mined, which is regarded as the user feature representation F ui :

[0126]

[0127] where w0 and w1 represent the output weight of the hidden layer in the CIN and the output weight of the DNN, respectively.

[0128] S4: The related attributes, the feedback information adjacency matrix and the user feature representation are processed by using the graph attention neural network to obtain a commodity recommendation result of the user.

[0129] Since the user-commodity network of the e-commerce platform is non-Euclidean structure data, the traditional discrete convolution cannot properly aggregate the potential information in the neighborhood, and the graph attention neural network (HAN) realizes the convolution operation of the graph from the graph theory, therefore, the HAT is selected to locally aggregate the neighborhood information, the relationship between the potential behavior and hidden information of the user and the user interest preference is comprehensively considered, the attention coefficient of the user to the commodity is calculated according to the user node feature and the commodity node feature, and finally the commodities are recommended to the target user according to the attention coefficient of the commodity and the user interest preference degree.

[0130] The interest preference of the user is obtained by fusing the internal factor FactorsIn(u j ) and the external factor FactorsOut(u j ) of the user-commodity; the internal factor value is calculated according to the user activity and the user attribute:

[0131] FactorsIn(u i )=ActDeg(u i )×User(u i )

[0132] The external factor value is calculated according to the commodity basic attribute and the interest incentive factor, and the external factor is taken as the commodity node feature:

[0133] FactorsOut(u i )=Enc(u i ,c k )×Iattri(c k )

[0134] The internal factor value FactorsIn(u i ) and the user feature representation F ui are spliced to obtain the user node feature:

[0135] The attention degree is calculated according to the feedback information adjacency matrix , the user node feature and the commodity node feature by using the graph attention neural network:

[0136] The HAN introduces the node level attention degree based on different meta-paths in the propagation process. The node level attention degree is the attention degree to the neighbor node, for example, the attention degree of the user u to the commodity c. The node level attention degree is represented as:

[0137]

[0138] attention(U, C; e) = δ(w) e ·[U||W i C])

[0139] Where U represents the user node feature, C represents the product node feature, and e is the explicit feedback information value, which takes the value of the feedback information adjacency matrix. The corresponding value in the adjacency matrix represents the value in the adjacency matrix. Middle user node u i With a certain type of product node c k There exists an interaction relationship between them, namely, an edge (u, i). || represents a concatenation operation, w e This represents a trainable node attention vector used to calculate attention levels. W i It is a transformation matrix responsible for transforming different types of product nodes into the same feature space. It also represents the attention level of all adjacent nodes of a user. The attention coefficients of users towards products are obtained by normalizing the attention scores of different meta-paths using the softmax function:

[0140]

[0141] in, Let I represent the attention coefficient of the i-th user for the k-th item, and let u represent the user node. i The neighboring nodes.

[0142] Finally, multiple different attention coefficient matrices are obtained based on different meta-paths.

[0143] The recommendation index of a product to a user is calculated based on the attention coefficient and the user's interest preference. Taking into account multiple types of hidden information about the product, the user's preference for different types of products is expressed through the attention coefficient of multiple paths, thus characterizing interest preferences. Then, based on the attention coefficient matrix and the interest preference score (PerferDeg(l)... i→k u i c k ), calculate the recommendation index of users for products under different meta paths. To ensure the rationality of user interests and preferences, the softmax value of different meta-path preference degrees is used to represent the user's final recommendation index for the product:

[0144]

[0145] in, This indicates the recommendation index. represents the attention coefficient of the i-th user to the k-th commodity, PerferDeg(i, k) represents the interest preference degree of the i-th user u i→k , k) represents the interest preference degree of the i-th user u i , k) represents the interest preference degree of the i-th user u k , k) represents the interest preference degree of the i-th user u i , k) represents the interest preference degree of the i-th user u k , k) represents the interest preference degree of the i-th user u

[0146] The recommendation indexes are sorted, and the top N commodities with the highest recommendation indexes are taken as the commodity recommendation result of the user; N is an integer, and preferably, 10 is taken. The e-commerce platform recommends commodities to the user according to the commodity recommendation result.

[0147] The above examples further illustrate the purpose, technical solutions and advantages of the present application. It should be understood that the above examples are only preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made to the present application within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. An e-commerce platform commodity recommendation method based on interest preferences, characterized in that, The method comprises the following steps: obtaining user data and preprocessing the user data; extracting the preprocessed user data into an e-commerce platform product recommendation model based on interest preferences to obtain product recommendation results for the user, and the e-commerce platform recommending products to the user according to the product recommendation results; the process of processing data by the e-commerce platform product recommendation model based on interest preferences comprises: S1: obtaining explicit feedback information and extracting relevant attributes according to the user data; the relevant attributes include user attributes, user activity, interest preference, product basic attributes and interest incentive factors; S2: compensating the explicit feedback information by using a differential graph autoencoder to obtain a compensated feedback information adjacency matrix; the process of compensating the explicit feedback information by using the differential graph autoencoder comprises: the differential graph autoencoder comprises an encoder and a decoder; obtaining an explicit feedback information matrix according to the explicit feedback information; mapping the explicit feedback information matrix to a hidden variable feature space by using the encoder to obtain a hidden feature representation; mapping the hidden feature representation to the original sample space by using the decoder to obtain the compensated feedback information adjacency matrix; S3: extracting hidden information of the user data by using an ultra-deep factor decomposition machine to obtain a user feature representation; the process of extracting the hidden information of the user data by using the ultra-deep factor decomposition machine comprises: the ultra-deep factor decomposition machine comprises a CIN and a DNN; performing one-hot encoding and embedding representation processing on the user attributes, interest incentive factors and product basic attributes to obtain original features of the user data; processing the original features of the user data by using a DNN to obtain implicit feature interactions; processing the original features of the user data by using a CIN to obtain explicit feature interactions; performing weighted aggregation on the implicit feature interactions and the explicit feature interactions to obtain the user feature representation; S4: processing the relevant attributes, the feedback information adjacency matrix and the user feature representation by using a graph attention neural network to obtain product recommendation results for the user; the process of processing the relevant attributes, the feedback information adjacency matrix and the user feature representation by using the graph attention neural network comprises: calculating an internal factor value according to the user activity and the user attributes; calculating an external factor value according to the product basic attributes and the interest incentive factors and taking the external factor as a product node feature; aggregating the internal factor value and the user feature representation to obtain a user node feature; calculating an attention degree according to the feedback information adjacency matrix, the user node feature and the product node feature by using the graph attention neural network; performing normalization processing on the attention degree to obtain an attention coefficient; calculating a recommendation index of the product to the user according to the attention coefficient and the interest preference of the user; sorting the recommendation index and taking the first N products as the product recommendation results for the user. 2.The method of claim 1, wherein, The formula for calculating the user activity is: ActDeg(u i ) = μ × Cartcount(u i ) + Orders(u i ) wherein ActDeg(u i ) denotes the user activity of the i-th user u i , μ denotes a penalty coefficient, Cartcount(u i ) denotes the number of added-to-cart of the i-th user u i , and Orders(u i ) denotes the number of orders of the i-th user u i . 3.The method of claim 1, wherein, The formula for calculating the interest preference is: Among them, PerferDeg(l i→k ,u i ,c k ) represents the i-th user u i For the k-th item c k interest preference, u pi Represents the i-th user u i Interest preferences are represented by meg(u i ,c k ) indicates that the k-th item c k The effective information is embedded into the user's... i In the embedded representation, m represents the quantity of goods. 4.The method of claim 1, wherein, The formula for calculating the compensated feedback information adjacency matrix is: wherein, denotes the compensated feedback information adjacency matrix, ActDeg denotes the user activity, σ denotes the variance of the latent feature representation, and Z denotes the latent feature representation. 5.The method of recommending goods on an e-commerce platform based on interest preferences of claim 1, characterized in that, The formula for calculating the internal factor value is: FactorsIn(u i ) = ActDeg(u i ) x User(u i ) wherein, FactorsIn(u j ) represents the internal factor value of the i-th user u i , ActDeg(u i ) represents the user activity of the i-th user u i , and User(u i ) represents the user self attribute of the i-th user u i . The formula for calculating the external factor value is: FactorsOut(u i ) i ,c k ) × Iattri(c k ) wherein FactorsOut(u i ) denotes the external factor value of the i-th user u i , Enc(u i , c k ) denotes the interest incentive factor of the i-th user u i and the k-th commodity c k , and Iattri(c k ) denotes the commodity basic attribute of the k-th commodity c k .

6. The interest preference-based e-commerce platform commodity recommendation method of claim 1, wherein, The formula for calculating the recommendation index of the product to the user is: wherein, represents a recommendation index, represents the attention coefficient of the i-th user u i→k to the k-th commodity c i . k represents the interest preference degree of the i-th user u i to the k-th commodity c k .

Images