A cross-platform e-commerce user alignment method based on user commodity interest

Abstract

The application belongs to the field of network alignment, and particularly relates to a cross-platform e-commerce user alignment method based on user commodity interest; the method comprises the following steps: obtaining user data of different e-commerce platforms; constructing an original universal graph according to the user data; performing dual hypergraph conversion and edge pruning processing on the original universal graph to obtain a user commodity relationship graph; calculating a feature projection of commodities according to the user commodity relationship graph; calculating a user interest feature vector according to the feature projection of commodities; processing user basic attribute data according to the user commodity relationship graph to obtain a user attribute feature vector; fusing the user interest feature vector and the user attribute feature vector to obtain user information representation; performing information matching according to the user information representation to obtain a user alignment result of two e-commerce platforms; the application has high alignment accuracy, is helpful for commodity recommendation of e-commerce platforms, and has high practicability.

Description

Technical Field

[0001] This invention belongs to the field of network alignment, specifically relating to a cross-platform e-commerce user alignment method based on user product interests. Background Technology

[0002] With the rapid global proliferation of online platforms, various e-commerce service platforms have emerged, such as Taobao, JD.com, and Pinduoduo in China, and Amazon and eBay internationally. These platforms offer users a wide range of personalized services. To meet diverse needs, users often register accounts on multiple e-commerce platforms instead of limiting themselves to a single platform. These users act as bridges connecting different networks, linking and integrating multiple e-commerce platforms. Similarly, merchants also choose to open stores on multiple platforms in hopes of achieving better sales.

[0003] User alignment allows for the aggregation of isolated data from diverse sources, creating a new database richer in information. For example, in recommendation systems on online platforms, genuine e-commerce platform user data is often anonymized to protect user privacy, raising questions about its authenticity. Furthermore, highly unstructured data leads to inconsistencies in attributes and network structures during user alignment, hindering the full extraction of data value. Constructing a high-quality user alignment model through heterogeneous information network alignment will facilitate a comprehensive understanding of user characteristics, thereby providing more personalized recommendation services.

[0004] In recent years, many scholars have conducted extensive research on network alignment methods, which can be broadly categorized into two main types: rule-based methods and statistical learning-based methods. Early approaches analyzed the similarity of attributes between users in two different networks to determine whether users distributed across heterogeneous networks belonged to the same user. Other approaches, with a limited number of seed nodes, studied the one-to-one anchor relationships between users in heterogeneous networks, proposing user account matching algorithms based on social, spatial, temporal, and textual attributes. There are also alignment methods based on partitioned indexes, which prune entities in the network beforehand, eliminating unlikely similarity pairs, building an index, and finally assigning potentially matching entity pairs to one or more small partitions as candidate entity pairs. However, with increasing user privacy awareness and the unverifiable nature of information, the quality of user data used for alignment tasks has become low, leading to unsatisfactory results.

[0005] Current research largely focuses on social network alignment, with limited research on e-commerce network alignment. Social networks lack the crucial concept of termination behavior, which significantly impacts network alignment, and existing alignment models fall short in this task, exhibiting several problems:

[0006] 1. Heterogeneous network entity relationships are complex and diverse. E-commerce platforms contain multiple entities such as users and products, and the relationships between these entities are complex and diverse, which makes it difficult to represent the features of these entities.

[0007] 2. Insufficient representation of hidden information in network structure. The original features only consider the mapping of local single nodes in the network space, which is insufficient to accurately capture the link relationships between nodes, affecting the model's ability to represent user behavior features.

[0008] 3. E-commerce users' interests exhibit cross-platform continuity. E-commerce platforms generate massive amounts of data, and a user's interests and behaviors on a single platform are merely a selective projection of that user's behavior. It is difficult to effectively prevent cross-platform users who purchase the same products from being identified as the same person. Summary of the Invention

[0009] To address the shortcomings of existing technologies, this invention proposes a cross-platform e-commerce user alignment method based on user product interests. This method includes:

[0010] S1: Obtain user data from the first e-commerce platform and the second e-commerce platform respectively. The user data includes basic user attribute data, basic product attribute data, and historical user behavior data.

[0011] S2: Construct the original hypergraph based on user data; perform dual hypergraph transformation and edge pruning on the original hypergraph to obtain the user-product relationship graph;

[0012] S3: Calculate the feature projection of the product based on the user product relationship graph;

[0013] S4: Calculate the user interest feature vector based on the feature projection of the product;

[0014] S5: Process the user's basic attribute data based on the user-product relationship diagram to obtain the user attribute feature vector;

[0015] S6: Use a gating mechanism to fuse user interest feature vectors and user attribute feature vectors to obtain user information representation;

[0016] S7: Use the Euclidean distance between the user information representations of the two e-commerce platforms to perform user alignment and obtain the user alignment results of the two e-commerce platforms.

[0017] Preferably, the process of calculating the feature projection of a product includes:

[0018] S31: Construct a sequence of user-interested products based on the user-product relationship graph;

[0019] S32: Calculate the feature influence coefficients between products in the user's interest product sequence, normalize the feature influence coefficients, and obtain the first weight coefficient of the product;

[0020] S33: Define the user interest decay function, and calculate the feature projection of the product based on the product's first weight coefficient and the user interest decay function.

[0021] Furthermore, the formula for calculating the characteristic influence coefficient is as follows:

[0022]

[0023] in, The feature influence coefficient of the j-th item on the i-th item in the user's interest item sequence Φ, where σ represents the activation function. c represents the attention coefficient. i Let c represent the basic feature vector of the i-th item in the user's interest item sequence Φ, || denotes the concatenation operation. j This represents the basic feature vector of the j-th item in the user's interest item sequence Φ.

[0024] Furthermore, the formula for calculating the feature projection of a commodity is:

[0025]

[0026] in, Let represent the feature projection of the i-th item in the user's interest item sequence Φ, K represent the number of heads in the multi-head attention, and σ represent the activation function. c represents the first weight coefficient of the j-th item in the user's interest item sequence Φ relative to the i-th item. j Let represent the basic feature vector of the j-th item in the user's interest item sequence Φ. Let Φ represent the set of all items in interest group Φ except for the i-th item, and let IntDecay represent the user interest decay function.

[0027] Preferably, the process of calculating user interest feature vectors includes:

[0028] S41: Use a single-layer artificial neural network to process the feature projection of the product to obtain the basic feature vector of the product;

[0029] S42: Aggregate the basic feature vector and feature projection of the product to obtain the second weight coefficient of the product;

[0030] S43: Calculate the feature vector representation of the user's interest product sequence based on the product's basic feature vector and the product's second weight coefficient;

[0031] S44: Aggregate the feature vector representations of all user interest product sequences to obtain the user interest feature vector.

[0032] Furthermore, the formula for calculating the second weighting coefficient is as follows:

[0033]

[0034]

[0035] in, Z represents the intermediate parameter, where X represents the number of items in the user's interest item sequence, and Z represents the intermediate parameter. Φ Let E represent the basic feature vector of the product, and let E represent the basic interest and behavior representation of the user. Let 'b' represent the feature projection of the product, and 'b' represent the deviation vector. This represents the second weighting coefficient.

[0036] Furthermore, the formula for calculating the feature vector representation of a user's interest item sequence is as follows:

[0037]

[0038] Among them, Z′ Φ The feature vector representation of a user's interest product sequence Z represents the second weighting coefficient, X represents the number of items in the user's interest item sequence, and Z represents the second weighting coefficient. Φ This represents the feature projection of the product, and Q represents the location code.

[0039] Preferably, the process of fusing user interest feature vectors and user attribute feature vectors includes:

[0040] S61: Calculate the combined weights based on the user attribute feature vector and the user interest feature vector;

[0041] S62: Calculate the user information representation based on the combined weights, user attribute feature vectors, and user interest feature vectors.

[0042] Furthermore, the formula for calculating the portfolio weights is:

[0043] ρ = sigmoid(W) α Z u +W β A(u)′+γ)

[0044] Where ρ represents the combined weight, W α W represents the parameters of the first neural gating layer. β Z represents the parameters of the second neural gating layer, γ represents the parameters of the third neural gating layer, sigmoid() represents the activation function, and Z represents the parameter of the third neural gating layer. uLet A(u)' represent the user interest feature vector, and let A(u)' represent the user attribute feature vector.

[0045] Furthermore, the formula for calculating the user information representation is as follows:

[0046] Z = ρ⊙Z u +(1-ρ)⊙A(u)′

[0047] Where Z represents the user information representation, ρ represents the combined weight, and Z u Let A(u)' represent the user interest feature vector, A(u)' represent the user attribute feature vector, and ⊙ represent the product of corresponding elements of the two matrices.

[0048] The beneficial effects of this invention are as follows: This invention processes user data from different e-commerce platforms, obtaining user attribute feature vectors and user interest feature vectors from the perspectives of user attributes and user interest in products, respectively. By fusing these two feature vectors, a user information representation is obtained, and users from different e-commerce platforms are aligned based on this representation. This invention solves the problem of significant differences in the performance of the same user across different platforms due to the cross-platform continuity of user interests, which leads to difficulties in alignment. It further explores the influencing factors of e-commerce user alignment, achieving high alignment accuracy, helping to discover more hidden user information, and improving the effectiveness of marketing, recommendation, and other merchant activities. Attached Figure Description

[0049] Figure 1 This is a flowchart of the cross-platform e-commerce user alignment method based on user product interests in this invention;

[0050] Figure 2 This is a schematic diagram of user alignment on an e-commerce platform in this invention;

[0051] Figure 3 This is a schematic diagram of e-commerce network conversion in this invention. Detailed Implementation

[0052] 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.

[0053] This invention proposes a cross-platform e-commerce user alignment method based on user product interests, such as... Figure 1 Figure 2 As shown, the method includes the following:

[0054] S1: Obtain user data from the first e-commerce platform and the second e-commerce platform respectively. The user data includes basic user attribute data, basic product attribute data, and historical user behavior data.

[0055] User data from the first and second e-commerce platforms can be obtained by downloading from the public APIs of the e-commerce platforms or from existing databases provided by the e-commerce platforms. User data includes basic user attribute data, basic product attribute data, and user historical behavior data. Basic product attribute data refers to the basic characteristics of the product, such as product name, product category, and product price. Basic user attribute data includes user nickname, gender, and place of origin. User historical behavior data refers to records of user interest actions such as browsing, favorites, adding to cart, and placing orders for products.

[0056] Raw data is usually unstructured and cannot be directly used for data analysis. Simple data cleaning can structure most unstructured data. For example, this involves deleting duplicate data and cleaning up invalid information; data cleaning can be applied to acquired user data to obtain cleaned user data.

[0057] S2: Construct the original general graph based on user data; perform dual hypergraph transformation and edge pruning on the original general graph to obtain the user-product relationship graph.

[0058] For e-commerce platforms, users and products are crucial components of the entire network, and their complex behavioral relationships form a heterogeneous network. Most current research implicitly captures link information from entity representations, but in e-commerce network alignment tasks, more accurate discovery of link features is needed because the entity association information represented by links significantly impacts task performance. To represent this complex heterogeneous network relationship and accurately discover link features, this invention constructs an original graph based on user data, abstracting the network into a graph G, such as... Figure 3 As shown, the e-commerce network transformation method is used to construct dual hypergraphs based on the original graphs of the two e-commerce platforms, and edge pruning is performed on the original graphs.

[0059] S21: Perform dual hypergraph transformation on the original graph.

[0060] Because the links between entities are extremely sparse, the number of non-zero elements in the adjacency matrix is ​​much smaller than the number of zero elements. Therefore, a sparse representation of the link list L∈R is used. (2m) To represent the interaction relationships between entities:

[0061] L={v start v end}

[0062] Among them, v start It is the starting entity of the link, vend is the closing entity of the link, and m is the number of links; therefore, the e-commerce network is represented as:

[0063] G = (A, L, H)

[0064] Where A represents entity features, i.e., the basic attribute features of a specific user or product, and H represents specific link features, i.e., user interest and behavior features (browsing, favorites, adding to cart, and placing orders).

[0065] By using the concept of dual space, the roles of nodes and edges in the network are interchanged, resulting in a transformed list of hyperlinks. Represented as:

[0066]

[0067] Where e * Indicates a hyperlink, v * Indicates the hyperlink e * Related entities, where D represents the sum of the degrees of all entities in the hypergraph, v i Let n represent the entities in the hypergraph, and n represent the number of entities in the hypergraph; the specific transformation process is as follows:

[0068]

[0069]

[0070] Where 1≤i≤m, This represents the first column and the (2i-1)th row of the link list, L 1，i This represents the first column and the i-th row of the link list.

[0071] Unlike other hypergraphs, in the original hypergraph, each link connects exactly two entities. Therefore, the hypergraph after dual transformation is a 2-regular graph, i.e., D = 2m. By replacing entity A with link H, the transformed graph G is obtained. * It can be represented as:

[0072] G = (A, L, H) → G * =(H, L) * A)

[0073] Similarly, the transformed dual hypergraph can be restored to the original ordinary graph, as shown below:

[0074] G * =(H, L) * (A) → G = (A, L, H)

[0075] This indicates that the transformation is reversible and does not lose any features or topological information of the graph. Therefore, the final e-commerce network graph G and the transformed hypergraph G are... *Represented as:

[0076] G = (A, L, H)

[0077] G * =(H, L) * A)

[0078] While a standard graph divides a single entity vertically, a supergraph clusters entities horizontally based on the links between them. For alignment tasks, a supergraph provides a more suitable data dimension than a standard graph.

[0079] S22: Perform edge pruning on the original map.

[0080] Define the user action influencing factor: A user's interest in a product is reflected through their interactive actions with that product. Users can browse, add to cart, favorite, or purchase products, and these actions, to varying degrees, reflect the user's level of interest in the product. Therefore, the user interest action influencing factor is defined as follows:

[0081]

[0082] Among them, browse, addCart, collect, and buy represent the user's actions of browsing, adding to cart, favorite, and purchasing products, respectively.

[0083] E-commerce platforms contain a vast amount of user interest and behavior data, and the same individual's interest and behavior can change across different platforms due to various factors. Therefore, rather than constantly adding data, appropriately generalizing user marginal behaviors allows the model to focus more on learning the core features of users. Since e-commerce networks typically have non-Euclidean data structures, traditional discrete convolution cannot handle them appropriately. Considering that Graph Convolutional Neural Networks (GCNs) implement graph convolution for graph data, this invention applies the GCN entity learning method to the hypergraph G after performing a dual hypergraph transformation on the ordinary graph G. * In entity learning, it is represented as:

[0084]

[0085] in, Represents the dual hypergraph G * Entity features at layer n+1 Represents the dual hypergraph G * In all neighboring entity characteristics at level n, N(e;L) * ) is a dual hypergraph G *The AGGREGATE function aggregates the neighboring entity set N of the link e, combining the features of all neighboring entities in the nth layer to obtain an entity feature in the n+1th layer. The UPDATE function then updates the information representation of the link e.

[0086] The initial representation of a hypergraph entity is obtained by aggregating the representations of the two entities associated with the link in the original hypergraph. Since different interest actions have varying impacts on the user, an interest action influence factor, IntAct, is applied to the initial hypergraph entity representation to make the associated features more realistic. Thus, the hypergraph G... * The initial representation of the entity is:

[0087] A′=IntAct·(a1+a2)

[0088] Where A′ is the hypergraph entity feature representation, and a1 and a2 are the two entity feature representations associated with this hypergraph entity in the original general graph. Thus, the link representation of the original general graph G, i.e., the user interest behavior representation learned by the dual hypergraph, is:

[0089] E = GCN(H, L) * ,A′)

[0090] Then, by using the tanh activation function, its link score can be obtained as follows:

[0091] P = tanh(E) = tanh(GCN(H, L)) * ,A′))

[0092] in Retaining the k entities with the highest scores yields the final pruned graph, i.e., the user-product relationship graph, represented as:

[0093] G′=(A,L,H idx ); idx = top k (P)

[0094] S3: Calculate the feature projection of the product based on the user product relationship graph.

[0095] S31: Construct a sequence of user-interested products based on the user-product relationship graph.

[0096] A user's series of interest actions regarding a certain type of product can be grouped into a sequence. This set can also be viewed as a flowchart with skippable steps. For example, for a product, a user's interest actions might be browsing, adding to favorites, adding to cart, and placing an order, or a user could skip intermediate actions and directly browse and place an order; browsing is a necessary condition for the subsequent actions. This set can largely represent a user's interest in this type of product. Based on the user-product relationship graph, the user interest product sequence Φ can be constructed as follows:

[0097] Φ = (c1, c2, ..., c X )

[0098] Where X is the number of items in the user's interest item sequence, and c i , i∈[1,k] represents the basic feature vector of the i-th item in the user interest item sequence of category c, referring to the item. This item is the item for which the user has taken an interest action. For example, if the user browses items c1, c2, and c3 in sequence, and then adds items c4 and c5 to their cart in sequence, then Φ=(c1, c2, c3, c4, c5). c4 and c5 can essentially be items c1 and c2 respectively, meaning that the same item can appear repeatedly in the user interest item sequence. For a user, there can be multiple user interest item sequences.

[0099] S32: Calculate the feature influence coefficients among products in the user's interest product sequence, normalize the feature influence coefficients, and obtain the first weight coefficient of the product.

[0100] User interest and behavioral characteristics are revealed through the products they interact with. Then, the original generalized graph is used to learn user generalized interests on the pruned user-product relationship graph G′.

[0101] The learning of user generalized interests is conducted from two dimensions: the product layer and the interest sequence layer. The learning of user generalized interests at the product layer includes the following:

[0102] Without precise context, the learning outcomes for entities are ambiguous. However, for user interests, using the user's interactions with a particular type of goods as context can provide a sufficiently intuitive understanding of the user's interest structure. Therefore, given a sequence of interest items for user i, Φ = (c1, c2, ..., c... X Here, 'c' represents products of the same category, which can be used to limit the context for feature learning. It's obvious that even for the same product, the impact on the user should differ depending on the type of association (browsing, favorites, adding to cart, purchasing). Furthermore, implicit associations should exist between products. For example, in a product pair (i, j), the influence coefficients of product i and product j should be different. Using a self-attention mechanism, the asymmetric weight coefficients between entities can be obtained, thus yielding the feature influence coefficients between products in the user's interest product sequence. The calculation formula can be expressed as:

[0103]

[0104] in, The feature influence coefficient of the j-th product on the ith product in the user interest product sequence Φ; σ represents the activation function, preferably the tanh function; c represents the attention coefficient. i Let c represent the basic feature vector of the i-th item in the user's interest item sequence Φ, || denotes the concatenation operation. j Let represent the basic feature vector of the j-th item in the user's interest item sequence Φ; similarly, the formula for calculating the feature influence coefficient of the i-th item on the j-th item in the user's interest item sequence Φ is:

[0105]

[0106] The feature influence coefficients are normalized using the softmax function to obtain the first weight coefficient of the product, expressed as:

[0107]

[0108] in, This represents the first weight coefficient of the j-th item in the user's interest item sequence Φ relative to the i-th item. Let Φ represent the set of all items in interest group Φ except for the i-th item.

[0109] S33: Define the user interest decay function, and calculate the feature projection of the product based on the product's first weight coefficient and the user interest decay function.

[0110] User interest in products changes over time, on an individual basis, and thus persists across different platforms. Only by reducing the engagement of historical interest actions can newer interest actions achieve better engagement, because the more recent a user's interest action in a product, the stronger their current interest in that product. This invention defines a user interest decay function as:

[0111] IntDecay=exp(-α×lastHours)

[0112] Where lastHours represents the number of hours since the last interest action, and α∈[0,1] represents the decay coefficient, which can be selected based on practical experience. Preferably, α is 0.5.

[0113] The feature projection of a product is calculated based on its first weight coefficient and the user interest decay function: The first weight coefficients of the products are aggregated, and after time decay processing, the feature projections of products in the user interest product sequence Φ based on the user interest product sequence Φ are obtained. The calculation formula is:

[0114]

[0115] in, σ represents the feature projection of the i-th item in the user's interest item sequence Φ; σ represents the activation function, preferably the Elu function; and IntDecay represents the user interest decay function.

[0116] Due to the scale-free nature of heterogeneous graphs, graph data exhibits significant variance. Furthermore, self-attention mechanisms can lead to weights excessively focusing on their own positions. Therefore, this problem is mitigated by extending the process to multi-head attention, processing low-dimensional projections separately before concatenation. Thus, for the feature projection of item i... Represented as:

[0117]

[0118] Where K represents the number of heads receiving multi-head attention, preferably K is 4.

[0119] S4: Calculate the user interest feature vector based on the feature projection of the product.

[0120] For user generalization interest learning at the interest sequence level, the following are included:

[0121] Focusing solely on learning about the products themselves is insufficient to encompass the user's interest profile; therefore, it is also necessary to learn the entire sequence of products that the user is interested in.

[0122] S41: Use a single layer of artificial neural network to process the feature projection of the product to obtain the basic feature vector of the product.

[0123] Projecting all product features of the product sequence Φ The basic feature vector Z of this sequence is obtained through a single layer of MLP (Artificial Neural Network). Φ , is represented as:

[0124]

[0125] S42: Aggregate the basic feature vector and feature projection of the product to obtain the second weight coefficient of the product.

[0126] The average value of the feature vector of the interest product sequence and the aggregated product feature projections is taken, and the weight coefficient of the interest product sequence is obtained by normalization. The calculation formula is as follows:

[0127]

[0128]

[0129] in, Z represents the intermediate parameter, where X represents the number of items in the user's interest item sequence, and Z represents the intermediate parameter. Φ Let E represent the basic feature vector of the product, and let E represent the basic interest and behavior representation of the user. Let 'b' represent the feature projection of the product, and 'b' represent the deviation vector. This represents the second weighting coefficient.

[0130] S43: Calculate the feature vector representation of the user's interest product sequence based on the product's basic feature vector and the product's second weight coefficient.

[0131] The order in which users interact with products based on their interests should influence their interest representation. Furthermore, the distance from each entity to other entities varies; that is, for a given sequence of interest products Φ, the compact centrality of each product is different. Therefore, when calculating the feature vector representation of the user's interest product sequence, positional encoding is added to the product feature projections to allow the vectors to represent temporal and positional information. The feature vector representation Z of the user's interest product sequence is calculated based on the product's basic feature vector and the product's second weight coefficient. Φ The formula is:

[0132]

[0133] Among them, Z′ Φ The feature vector representation of a user's interest product sequence Z represents the second weighting coefficient, X represents the number of items in the user's interest item sequence, and Z represents the second weighting coefficient. Φ This represents the feature projection of the product, and Q represents the location code.

[0134] S44: Aggregate the feature vector representations of all user interest product sequences in the user interest product sequence to obtain the user interest feature vector.

[0135] Considering that each user has multiple sequences of items of interest, the final user interest feature vector Z can be obtained by aggregating the feature vector representations of all user interest item sequences. u , is represented as:

[0136]

[0137] Where Y represents the number of user interest sequences.

[0138] S5: Process the user's basic attribute data based on the user-product relationship diagram to obtain the user attribute feature vector.

[0139] Based on the user-product relationship graph G′ obtained by pruning the original graph, the user's basic attribute data is processed, and a transformation matrix M is introduced. φ This allows user attributes to be projected into the same feature space as the interest vector, as shown below:

[0140] A(u)′=A(u)·M φ

[0141] Where A(u) is the user's basic feature vector, and A(u)′ represents the user's attribute feature vector.

[0142] S6: Use a gating mechanism to fuse user interest feature vectors and user attribute feature vectors to obtain user information representation.

[0143] The rapid development of the internet has led to the rise of numerous e-commerce platforms. Differentiated platform decisions are key to their success in the competition, and this also influences users' shopping choices. User interests, more accurately described as user attributes, are constantly changing over time. The interests a user displays on a platform can only be considered a selective projection of those interests onto that platform. A significant difference between e-commerce network alignment and general alignment tasks is that cross-platform user behavior should be highly similar, but not identical. A user's interest trends should be an integration of the interest trends displayed across their accounts on different platforms—that is, complementary interests, which naturally include many overlapping interests. Therefore, after obtaining user interest representations, it is necessary to modify these cross-platform interest representations and integrate user attributes.

[0144] A gating-based alignment algorithm is used to fuse user interest feature vectors and user attribute feature vectors, including the following:

[0145] S61: Calculate the combined weights based on the user attribute feature vector and the user interest feature vector.

[0146] A neural gating layer is introduced to adaptively merge users' product interest features and user attribute features. For users with dispersed interests, their interest performance across different platforms should be reduced, while their attribute-based features should be favored. First, the combined weight ρ should be calculated, expressed as:

[0147] ρ = sigmoid(W) α Z u +W β A(u)′+γ)

[0148] Where ρ represents the combined weight; W α W represents the parameters of the first neural gating layer. β Z represents the parameters of the second neural gating layer, γ represents the parameters of the third neural gating layer, used to extract the salient parts of the two feature representations; sigmoid() represents the activation function, Z... u Let A(u)' represent the user interest feature vector, and let A(u)' represent the user attribute feature vector.

[0149] S62: Calculate the user information representation based on the combined weights, user attribute feature vectors, and user interest feature vectors.

[0150] The final user information representation Z is obtained by smoothly combining the weights. The calculation formula is as follows:

[0151] Z = ρ⊙Z u +(1-ρ)⊙A(u)′

[0152] Where Z represents the user information representation, ρ represents the combined weight, and Z u Let A(u)' represent the user interest feature vector, A(u)' represent the user attribute feature vector, and ⊙ represent the product of corresponding elements of the two matrices.

[0153] S7: Use the Euclidean distance between the user information representations of the two e-commerce platforms to perform user alignment and obtain the user alignment results of the two e-commerce platforms.

[0154] For cross-platform e-commerce network alignment tasks, the goal of representation learning is to group users from two e-commerce networks together, while keeping them away from other users. Therefore, when finally matching information, even if they are in different networks, only the regularized distance between the two user nodes needs to be considered. Taking the first e-commerce network G1 as the support network and the second e-commerce network G2 as the query network, the correspondence is determined by calculating the Euclidean distance between the final feature representations (i.e., user information representations) of users in G1 and G2, thus determining whether the alignment is successful. The Euclidean distance of the user information representations is expressed as:

[0155]

[0156] For each user node in query network G2 Traversing all user nodes in the support network G1 can be viewed as an n-classification problem. When a certain user node in the support network G1... Query user nodes in network G2 When the Euclidean distance is minimized, the two user nodes are classified into the same category, i.e., they are aligned. The loss function is then iteratively calculated, and is expressed as follows:

[0157]

[0158] in, Let C represent the loss function in the nth iteration, and C represent the number of users.

[0159] For each user node in the support network G1, the user with the highest similarity can be found in the query network G2 for alignment. When the loss function is minimized, the alignment of the two e-commerce platforms is completed, and the user alignment result of the two e-commerce platforms is obtained.

[0160] User alignment aggregates and consolidates isolated data from different sources, creating a new database with richer information. For example, recommendation systems on online platforms can gain a comprehensive understanding of user characteristics based on alignment results, enabling them to provide more personalized product recommendations. Furthermore, alignment can increase the frequency of product recommendations, stimulating orders and increasing transaction success rates, thereby boosting revenue for e-commerce platforms and merchants. Simultaneously, it enhances user satisfaction and loyalty to the platform.

[0161] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A cross-platform e-commerce user alignment method based on user product interests, characterized in that, include: S1: Obtain user data from the first e-commerce platform and the second e-commerce platform respectively. The user data includes basic user attribute data, basic product attribute data, and historical user behavior data. S2: Construct the original hypergraph based on user data; perform dual hypergraph transformation and edge pruning on the original hypergraph to obtain the user-product relationship graph; S3: Calculate the feature projection of the product based on the user-product relationship graph; the process of calculating the feature projection of the product includes: S31: Construct a sequence of user-interested products based on the user-product relationship graph; S32: Calculate the feature influence coefficients among items in the user's interest item sequence, normalize the feature influence coefficients to obtain the first weight coefficients of the items; the formula for calculating the feature influence coefficients is: ； in, User Interest Product Series The characteristic influence coefficient of the j-th product on the i-th product. This represents the activation function. Indicates the attention coefficient. Represents a sequence of products that represent user interests. The basic feature vector of the i-th item in the dataset. This indicates a splicing operation. Represents a sequence of products that represent user interests. The basic feature vector of the j-th item in the dataset; S33: Define the user interest decay function, and calculate the feature projection of the product based on the product's first weight coefficient and the user interest decay function; the formula for calculating the product's feature projection is: ； in, Represents a sequence of products that represent user interests. The feature projection of the i-th item in the middle, The number of heads representing multi-head attention. Represents a sequence of products that represent user interests. The first weighting coefficient of the j-th item to the i-th item. Indicating in interest groups The set of all other items except the i-th item. This represents the user interest decay function; S4: Calculate the user interest feature vector based on the feature projection of the product; the process of calculating the user interest feature vector includes: S41: Use a single-layer artificial neural network to process the feature projection of the product to obtain the basic feature vector of the product; S42: Aggregate the basic feature vector and feature projection of the product to obtain the second weight coefficient of the product; S43: Calculate the feature vector representation of the user's interest product sequence based on the product's basic feature vector and the product's second weight coefficient; S44: Aggregate the feature vector representations of all user interest product sequences to obtain the user interest feature vector; S5: Process the user's basic attribute data based on the user-product relationship diagram to obtain the user attribute feature vector; S6: Use a gating mechanism to fuse user interest feature vectors and user attribute feature vectors to obtain user information representation; S7: Use the Euclidean distance between the user information representations of the two e-commerce platforms to perform user alignment and obtain the user alignment results of the two e-commerce platforms.

2. The method for cross-platform e-commerce user alignment based on user product interests according to claim 1, characterized in that, The formula for calculating the second weighting coefficient is: ； ； in, Indicates intermediate parameters. This indicates the number of items in the user's interest item sequence. Represents the basic feature vector of a product. This represents the user's basic interest and behavior. Represents the feature projection of the product. Represents the deviation vector. This represents the second weighting coefficient.

3. The method for cross-platform e-commerce user alignment based on user product interests according to claim 1, characterized in that, The formula for calculating the feature vector representation of a user's interest item sequence is: ； in, The feature vector representation of a user's interest product sequence This represents the second weighting coefficient. This indicates the number of items in the user's interest item sequence. Represents the feature projection of the product. This indicates the position code.

4. The method for cross-platform e-commerce user alignment based on user product interests according to claim 1, characterized in that, The process of fusing user interest feature vectors and user attribute feature vectors includes: S61: Calculate the combined weights based on the user attribute feature vector and the user interest feature vector; S62: Calculate the user information representation based on the combined weights, user attribute feature vectors, and user interest feature vectors.

5. A cross-platform e-commerce user alignment method based on user product interests according to claim 4, characterized in that, The formula for calculating the portfolio weights is: ； in, Indicates the portfolio weight. Indicates the parameters of the first neural gating layer. This represents the parameters of the second neural gating layer. This represents the parameters of the third neural gating layer. This represents the activation function. Represents the user interest feature vector. This represents the user attribute feature vector.

6. A cross-platform e-commerce user alignment method based on user product interests according to claim 4, characterized in that, The formula for calculating user information representation is: ； in, This represents user information. Indicates the portfolio weight. Represents the user interest feature vector. Represents the user attribute feature vector. This indicates that the corresponding elements of two matrices are multiplied.

Images