Recommendation method and device based on decoupled learning and goal behavior-oriented learning

Abstract

The application discloses a recommendation method, in particular to a recommendation method and device based on decoupling learning and target behavior-oriented learning. The method comprises the following steps: constructing a user-item interaction graph under a behavior based on the interaction conditions of users and items under multiple behaviors; performing decoupling attribute domain representation processing on the user-item interaction graph to obtain item decomposition embedding corresponding to item nodes and user decomposition embedding corresponding to user nodes in multiple decoupled attribute domains; calculating the attention score between a user node and its adjacent item node in an attribute domain based on the item decomposition embedding and the user decomposition embedding for multiple attribute domains corresponding to the user-item interaction graph; performing aggregation processing on the user node and the item node based on the attention score to obtain user target embedding and item target embedding under the behavior; and performing analysis and processing on the user target embedding and the item target embedding under multiple behaviors to obtain a recommendation result. The method can obtain more accurate recommendation results.

Description

TECHNICAL FIELD

[0001] The application belongs to a recommendation method, in particular to a recommendation method and device based on decoupled learning and goal behavior oriented learning. BACKGROUND

[0002] In recommendation systems, utilizing various user behaviors such as clicks, favorites, and purchases is crucial to alleviate the inherent sparsity problem in single behavior data. Multi-behavior recommendation refers to a method of considering multiple behaviors of users (such as clicks, purchases, comments, etc.) in a recommendation system to make recommendations. This method can more comprehensively understand the interests and preferences of users, thereby improving the accuracy of recommendations and user satisfaction.

[0003] Existing multi-behavior recommendation methods mainly utilize graph network methods to model the semantic interaction between users and items under different behaviors. These graph network-based methods model the interaction relationship between users and items in different behaviors, aggregate the learned embeddings from different user-item behaviors to obtain the final user (item) embedding representation. Some methods utilize contrastive learning to further improve model performance by adjusting the user preferences between target behaviors and auxiliary behaviors. Although combining contrastive learning with graph neural networks significantly improves the performance of multi-behavior recommendation, multi-behavior recommendation still faces two main problems:

[0004] Firstly, these methods do not consider the fine-grained relationship between user preferences for specific item attributes between target behaviors and auxiliary behaviors. Item attributes reflect user preferences, and there is a correlation between multiple behaviors for specific item attributes.

[0005] Secondly, when learning user preferences for target behaviors, existing methods ignore noise data in auxiliary behaviors. Noise data in auxiliary behaviors mainly reflects certain item attributes that are not related to target user behaviors. Even some items that do not seem to be related to target behaviors still provide useful information for predicting target behaviors for specific item attributes.

[0006] In actual application scenarios, auxiliary behaviors such as clicks and adds that are inconsistent with target purchase behaviors have a negative impact on the accuracy of recommendations for target user behavior preferences. However, the commonly used method in existing technology only learns coarse-grained user preferences and fails to consider the subtle attributes of items that users focus on in different behaviors, resulting in poor accuracy of recommendation results. SUMMARY

[0007] The technical problem solved by the present application is that the above-mentioned method commonly used in the prior art usually only learns coarse-grained user preferences and fails to consider subtle attributes of items focused on by users in different behaviors, resulting in poor accuracy of recommendation results.

[0008] The content of the present application includes:

[0009] In a first aspect, the embodiments of the present application provide a recommendation method based on decoupling learning and target behavior-oriented learning, comprising:

[0010] Based on the interaction between users and items under multiple behaviors, a user-item interaction graph under each behavior is constructed, the user-item interaction graph comprising a user node for representing the user and an item node for representing the item, and the multiple behaviors comprising a target behavior and an auxiliary behavior;

[0011] The user-item interaction graph is subjected to decoupled attribute domain representation processing to obtain item decomposition embeddings corresponding to the item nodes and user decomposition embeddings corresponding to the user nodes in multiple decoupled attribute domains;

[0012] For multiple attribute domains corresponding to each user-item interaction graph, based on the item decomposition embeddings and the user decomposition embeddings, the attention scores between the user nodes and their adjacent item nodes under each attribute domain are calculated;

[0013] The user nodes and the item nodes are subjected to aggregation processing based on the attention scores to obtain user target embeddings and item target embeddings under each behavior;

[0014] The user target embeddings and the item target embeddings under multiple behaviors are subjected to analysis processing to obtain a recommendation result.

[0015] Optionally, the decoupled attribute domain representation processing of the user-item interaction graph to obtain the item decomposition embeddings corresponding to the item nodes and the user decomposition embeddings corresponding to the user nodes in multiple decoupled attribute domains comprises:

[0016] Multiple decoupled attribute domains are constructed, and user initial embeddings corresponding to the user nodes and item initial embeddings corresponding to the item nodes are obtained;

[0017] The item initial embeddings are decomposed in each attribute domain to obtain item decomposition embeddings in the attribute domain, and user decomposition embeddings in each attribute domain are determined based on the user initial embeddings.

[0018] Optionally, the decomposing the item initial embedding in each of the attribute domains to obtain an item decomposed embedding in the attribute domain comprises:

[0019] In each of the attribute domains, the item initial embedding is projected into an embedding of the attribute domain by using a pre-learned projection matrix to obtain an item decomposed embedding in the attribute domain, the item decomposed embedding in the attribute domain being:

[0020]

[0021] wherein, an item decomposed embedding e i an item initial embedding for representing the item node i, a projection matrix, the projection matrix being a learnable parameter matrix,

[0022] Optionally, the calculating an attention score between the user node and its adjacent item node in each of the attribute domains based on the item decomposed embedding and the user decomposed embedding comprises:

[0023] determining an average value of the item decomposed embeddings of all adjacent item nodes corresponding to the user node in the attribute domain as a query vector corresponding to the user node;

[0024] using the query vector corresponding to the user node to calculate the attention score between the user node and its adjacent item node in the attribute domain.

[0025] Optionally, the attention score between the user node and its adjacent item node is:

[0026]

[0027] wherein, a query vector corresponding to the user node u in the attribute domain a under the target behavior, a pre-learned parameter, an item decomposed embedding of the item node i under the attribute domain a.

[0028] Optionally, the aggregating the user node and the item node based on the attention score to obtain a user target embedding and an item target embedding under each of the behaviors comprises:

[0029] aggregate the user nodes and their adjacent item nodes to obtain a user aggregated embedding, and aggregate the item nodes and their adjacent user nodes to obtain an item aggregated embedding;

[0030] add the user aggregated embeddings in all the attribute domains under the same user-item interaction graph to obtain a user target embedding under the behavior, and add the item aggregated embeddings in all the attribute domains under the same user-item interaction graph to obtain an item target embedding under the behavior.

[0031] Optionally, the item is a commodity, the target behavior is purchase, and the auxiliary behaviors include at least one of clicking, collecting, and commenting.

[0032] The interaction conditions of the user and the item under the multiple behaviors are used to represent the conditions of the user performing the target behavior and the auxiliary behaviors on the commodity, and the attribute domains are used to represent the attributes of the commodity.

[0033] In a second aspect, an embodiment of the present application provides a recommendation device based on decoupling learning and target behavior oriented learning, which comprises:

[0034] A construction module is configured to construct a user-item interaction graph under each of the behaviors based on the interaction conditions of the user and the item under the multiple behaviors, wherein the user-item interaction graph comprises a user node representing the user and an item node representing the item, and the multiple behaviors include a target behavior and an auxiliary behavior.

[0035] A decoupling representation module is configured to perform decoupling attribute domain representation processing on the user-item interaction graph to obtain item decomposition embeddings corresponding to the item nodes and user decomposition embeddings corresponding to the user nodes in multiple decoupled attribute domains.

[0036] An attention calculation module is configured to calculate, for each of the attribute domains, an attention score between the user nodes and their adjacent item nodes under the attribute domain based on the item decomposition embeddings and the user decomposition embeddings.

[0037] An aggregation processing module is configured to aggregate the user nodes and the item nodes based on the attention scores to obtain a user target embedding and an item target embedding under each of the behaviors.

[0038] An analysis processing module is configured to perform analysis processing on the user target embeddings and the item target embeddings under the multiple behaviors to obtain a recommendation result.

[0039] In a third aspect, an embodiment of the present invention provides an electronic device comprising: a memory, a processor, and a program stored in the memory and executable on the processor; the processor is configured to read the program in the memory to implement the steps of the recommendation method based on decoupled learning and target behavior-oriented learning as described in the first aspect.

[0040] In a fourth aspect, an embodiment of the present invention provides a readable storage medium for storing a program, which, when executed by a processor, implements the steps of the recommendation method based on decoupled learning and target behavior-oriented learning as described in the first aspect.

[0041] The beneficial effect of the present invention is that, in this embodiment, the user-item interaction under different behaviors is characterized in multiple decoupled attribute domains, and the user's preferences are reflected through fine-grained item attributes. The user's preferences for different item attributes are learned based on a decoupled graph convolutional network to capture the fine-grained semantic interactions between users and items in different attribute spaces. The method provided by this embodiment can capture the subtle attributes of the items that users pay attention to in different behaviors, improve the understanding of user preferences and improve the quality of recommendations in complex behavioral environments, thereby obtaining more accurate recommendation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Attachment Figure 1 A flowchart of a recommendation method based on decoupled learning and goal-behavior-oriented learning provided by an embodiment of the present invention;

[0043] Attachment Figure 2 A diagram illustrating the architecture of a recommendation method based on decoupled learning and goal-behavior-oriented learning provided by an embodiment of the present invention;

[0044] Attachment Figure 3 A schematic diagram of the process of obtaining a user-project interaction diagram provided by an embodiment of the present invention;

[0045] Attachment Figure 4a A schematic diagram of the decoupled graph convolution part provided by an embodiment of the present invention;

[0046] Attachment Figure 4b for Figure 4a A partial enlarged schematic diagram of the corresponding part of the attribute domain 1;

[0047] Attachment Figure 4c for Figure 4a A partial enlarged schematic diagram of the corresponding part of the attribute domain K;

[0048] Attachment Figure 5 A schematic diagram of the decoupled contrastive learning part provided by an embodiment of the present invention;

[0049] Attachment Figure 6 Schematic diagram of a recommendation device based on decoupled learning and goal-behavior-oriented learning provided by an embodiment of the present invention;

[0050] attached Figure 7 The structural schematic diagram of an electronic device provided in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0051] The term "and / or" in the embodiments of the present application describes the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B can mean that A exists alone, A and B exist together, and B exists alone. The character " / " generally represents an "or" relationship between the associated objects. The term "multiple" in the embodiments of the present application means two or more, and other quantifiers are similar. The terms "first", "second", and the like in the specification and claims of the present application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the terms used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than that illustrated or described here, and the objects distinguished by "first", "second" are usually of the same type and do not limit the number of objects, for example, the first object can be one or more.

[0052] The technical solutions in the embodiments of the present application will be described clearly and completely in the embodiments of the present application in combination with the drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, and not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0053] Please see Figure 1 , Figure 1 The schematic diagram of a recommendation method based on decoupling learning and target behavior-oriented learning provided in an embodiment of the present application, the method specifically includes the following steps:

[0054] Step 101, based on the interaction between users and items under multiple behaviors, a user-item interaction graph under each behavior is constructed, the user-item interaction graph includes a user node for representing the user and an item node for representing the item, and the multiple behaviors include a target behavior and an auxiliary behavior;

[0055] Step 102, performing decoupling attribute domain representation processing on the user-item interaction graph to obtain item decomposition embedding corresponding to the item node and user decomposition embedding corresponding to the user node in multiple decoupled attribute domains;

[0056] Step 103, for each of the multiple attribute domains corresponding to the user-item interaction graph, based on the item decomposition embedding and the user decomposition embedding, calculating the attention score between the user node and its adjacent item node under each attribute domain.

[0057] At step 104, the user nodes and the item nodes are aggregated based on the attention scores to obtain user target embeddings and item target embeddings under each behavior;

[0058] At step 105, the user target embeddings and the item target embeddings under multiple behaviors are combined for analysis to obtain a recommendation result.

[0059] Referring to Figure 2 The embodiments of the present application also provide a recommendation method framework based on decoupling learning and target behavior guided learning, which includes four modules, namely, a decoupling attribute domain representation model for implementing step 102, a target behavior guided attention mechanism module for implementing step 103, a decoupling graph attention network module for implementing step 103, and a denoising contrastive learning module for training.

[0060] Different steps or modules will be specifically introduced below. In the specific implementation, the user embedding can also be referred to as a user feature or a user representation, and the item embedding can also be referred to as an item feature or an item representation. As a specific embodiment, the item can be an object targeted by the user's behavior, for example, when the behavior is purchase, the item is the purchased object.

[0061] Referring to Figure 3 In step 101, the interaction between users and items under multiple behaviors is first obtained, and in specific embodiments, the number of users and items can both be multiple. The multiple behaviors include a target behavior and auxiliary behaviors, and in some embodiments, the number of auxiliary behaviors is one or more. For example, in the specific embodiment shown in Figure 3 , the behaviors include purchase, add to cart, and click, wherein the purchase is the target behavior, and the add to cart and click are auxiliary behaviors.

[0062] In step 101, user-item interaction graphs under different behaviors are established respectively. In the user-item interaction graph under the behavior, if a user and an item have interacted under the behavior (for example, user A adds item A to the cart, that is, user A and item A have interacted under the purchase behavior), the user node representing the user and the item node representing the item are connected. In the specific embodiment shown in Figure 3 , the user-item interaction graphs corresponding to the three behaviors of purchase, add to cart, and click are obtained respectively.

[0063] In step 102, the item nodes and user nodes in the user-item interaction graph are projected into a plurality of decoupled attribute domains to represent the attraction of different item attributes to users under different behaviors. The correlation between user preferences in different behaviors is mainly reflected on different item attributes, and the noise data in auxiliary behaviors is also closely related to item attributes. Through this step, the user's preference for item attributes under different behaviors can be focused on, and therefore more accurate embedding can be obtained by focusing on specific item attributes.

[0064] Optionally, in some embodiments, the step 102 comprises:

[0065] A plurality of decoupled attribute domains are constructed, and user initial embedding corresponding to the user nodes and item initial embedding corresponding to the item nodes are obtained;

[0066] The item initial embedding is decomposed in each of the attribute domains to obtain item decomposition embedding in the attribute domain, and user decomposition embedding in each of the attribute domains is determined based on the user initial embedding.

[0067] In this embodiment, several decoupled attribute domains are constructed to represent different item attribute domains. The attribute domain is an implicit space, which is different from the explicit item category and user attribute, and is a potential semantic space representing implicit potential item attributes. For example, when the item is a clothing item, the attribute domain can represent the version, color, price, material, etc. of the clothes. In specific implementation, the number of attribute domains is not limited here, and each user-item interaction graph corresponds to a plurality of attribute domains.

[0068] It should be understood that, without special description, the attribute domain referred to in each embodiment includes each attribute domain corresponding to each user-item interaction graph (i.e. under all behaviors), and the steps performed in all attribute domains are the same, which will not be described in detail.

[0069] In some embodiments, each user and item is associated with an ID embedding. Let is an embedding matrix of user embedding initialization, and Q∈R N×D is an embedding matrix of item embedding initialization, where M represents the number of users, N represents the number of items, and D represents the embedding size. Formally, given a one-hot encoding embedding matrix and a one-hot encoding embedding matrix Specifically as follows:

[0070]

[0071]

[0072] where e uis a user initial embedding of a user u, is a one-hot encoding vector corresponding to a user u, e i is a project initial embedding of a project i, is a one-hot encoding vector of a project i.

[0073] In each attribute domain, a project decomposition embedding is obtained by decomposing the project initial embedding. Assuming A is A = 1, 2,..., |A|, A is a hyperparameter representing the number of decoupled attribute domains, the aforementioned e i is projected into the embedding of attribute a e A by using a projection matrix. Specifically, in some embodiments, the decomposing, in each of the attribute domains, the project initial embedding to obtain a project decomposition embedding in the attribute domain, comprises:

[0074] In each of the attribute domains, the project initial embedding is projected into the embedding of the attribute domain by using a pre-learned projection matrix to obtain a project decomposition embedding in the attribute domain, the project decomposition embedding in the attribute domain being:

[0075]

[0076] wherein, is a project decomposition embedding used to represent a project node i in an attribute domain a, i is a project initial embedding used to represent a project node i, is used to represent the projection matrix, the projection matrix being a learnable parameter matrix, || ||2 is used to represent the L2 norm. In some embodiments, is a projection matrix of an attribute domain shared by all projects.

[0077] It should be understood that in the present embodiment, the initial embedding of the input graph attention network is the same as the project decomposition embedding, i.e.,

[0078]

[0079] Optionally, in some embodiments, the specific manner of determining the user decomposition embedding in each of the attribute domains based on the user initial embedding is as follows:

[0080] The user decomposition embedding is obtained by using a transformation matrix and is taken as the initial embedding of the input graph attention network as follows:

[0081]

[0082] Since different attribute domains should contain different information about item attributes, otherwise multiple item embeddings degenerate into equivalent to one item embedding, it is necessary to separate item embeddings in different attribute domains. In order to better model capacity and interpretability, and avoid redundant information, an independence loss is introduced in the embodiment to constrain to ensure the independence of different attribute domains, and prevent them from degenerating into a single attribute domain.

[0083] See Figures 4a-4c , specifically, in the embodiment, mutual information can be used to encourage the separation of item decomposition embeddings in different attribute domains. For each item i, the mutual score is as follows:

[0084]

[0085] Where s(·) is a function that measures the similarity of two identical item decomposition embeddings in different attribute domains, τ represents the temperature in the softmax function as a hyperparameter, and A is the set of attribute domains.

[0086] In some embodiments, s(·) is set to a cosine similarity function, which is specifically as follows:

[0087]

[0088] The final dependency loss is composed of the mutual scores of all items, which is specifically calculated as follows:

[0089]

[0090] As Figure 2 shown, in step 103, an attention mechanism based on goal behavior guided learning is designed for each attribute domain, so as to realize the modification of the message passing of the graph neural network according to the user's preference for the target behavior. For all adjacent item nodes of a user in a specific attribute domain, if the item embedding in the attribute domain is closer to the user's target behavior preference, the information passed by the item node to the user node in the attribute domain should have a higher importance weight compared with other adjacent item nodes.

[0091] Optionally, in some embodiments, the step 103 comprises:

[0092] determining the average value of the item decomposition embeddings of all adjacent item nodes corresponding to the user node in the attribute domain as the query vector corresponding to the user node;

[0093] using the query vector corresponding to the user node to calculate the attention score between the user node and its adjacent item nodes in the attribute domain.

[0094] In the present embodiment, in order to learn the attention score of the user node u and its adjacent item node i in the attribute domain a, first, a query vector q that can reflect the target behavior preference of the user node u is learned. Specifically, in some embodiments, a query vector is calculated in each attribute domain in each user behavior, and different behaviors need to calculate different query vectors. The query vector is the average value of all item representations that the user has interacted with, and thus can represent the user preference. The average value of the item decomposition embedding of all item nodes that interact with the user node u in the attribute domain a is determined as the query vector, specifically as follows:

[0095]

[0096] wherein, is used to represent the query vector corresponding to the user node u in the attribute domain a under the target behavior, represents the nth item node adjacent to the user node u in the attribute domain a. e in the formula and Figure 4b correspond to the target historical behavior.

[0097] Then, the learned query vector can be used to calculate the attention score of the user node and its adjacent item node. Optionally, in some embodiments, the attention score between the user node and its adjacent item node is:

[0098]

[0099] wherein, is used to represent the query vector corresponding to the user node u in the attribute domain a under the target behavior, is a pre-learned parameter, [;] is used to represent the concatenation of vectors, is used to represent the item decomposition embedding of the item node i in the attribute domain a.

[0100] It should be understood that in each attribute domain, the attention score of each user node and its adjacent item node can be calculated. After obtaining the attention score of each user node and its adjacent item node, the attention score of each item node and its adjacent user node is also obtained.

[0101] Optionally, in some embodiments, the step 104 comprises:

[0102] based on the attention score, aggregating the user node and its adjacent item node to obtain a user aggregation embedding, and aggregating the item node and its adjacent user node to obtain an item aggregation embedding;

[0103] The user target embedding under the behavior is obtained by adding the user aggregated embeddings in all the attribute domains under the same user-item interaction graph, and the item target embedding under the behavior is obtained by adding the item aggregated embeddings in all the attribute domains under the same user-item interaction graph.

[0104] As shown in the following formula (1), after obtaining the attention score between the user node and the adjacent item node, a graph attention network is used for aggregation processing. Figure 2 As shown in the following formula (1), after obtaining the attention score between the user node and the adjacent item node, a graph attention network is used for aggregation processing.

[0105] In some embodiments, based on the LightGCN method, an aggregation function under a specific auxiliary behavior is designed, as shown in the following formula (2):

[0106]

[0107]

[0108] wherein, is used to represent the neighbor nodes of the user node u, is used to represent the neighbor nodes of the item node i; l is used to represent the number of layers of the graph attention network, is used to represent the user decomposition embedding of the user node u in the attribute domain a in the (l+1) th layer, is used to represent the item decomposition embedding of the item node i in the attribute domain a in the (l+1) th layer, is used to represent the attention score of the user u and the adjacent item node i in the attribute domain a in the l layer.

[0109] The value of the corresponding of each attribute domain under the target behavior is set to 1. For the attribute domains {1, 2, …, A} under the auxiliary behavior k, the embedding of the user node u in the attribute domain a under the auxiliary behavior k obtained by the L-layer graph attention network can be obtained as follows:

[0110]

[0111]

[0112] wherein, is used to represent the embedding of the user node u in the attribute domain a under the auxiliary behavior k obtained by the L-layer graph attention network, is used to represent the embedding of the item node i in the attribute domain a under the auxiliary behavior k obtained by the L-layer graph attention network.

[0113] By the above manner, the user target embedding of the user node under the behavior is obtained by summing the embeddings corresponding to the same user node in all attribute domains under the same behavior. Similarly, the item target embedding of the item node under the behavior is obtained by summing the embeddings corresponding to the same item node in all attribute domains under the same behavior.

[0114] In the application embodiment, a target behavior guided graph attention mechanism is designed in the graph convolution network to obtain a graph attention network. The target user behavior preference can guide the learning of user-item interaction learning in auxiliary behaviors. The attention mechanism is incorporated into the graph convolution network, which can effectively extract the item attributes that the user is interested in for the auxiliary behavior, while reducing the influence of irrelevant item attributes.

[0115] Optionally, in some embodiments, the analysis and processing based on the user target embedding and the item target embedding under the plurality of behaviors is performed to obtain a recommendation result, including:

[0116] determining a user final embedding based on the user target embedding under all behaviors, and determining an item final embedding based on the item target embedding under all behaviors;

[0117] performing analysis and processing based on the user final embedding and the item final embedding to obtain the recommendation result.

[0118] After obtaining the user target embedding and the target item embedding under each of the K behaviors {1, 2, …, K}, the user final embedding and the item final embedding under all behaviors can be further determined. In some embodiments, the user final embedding and the item final embedding are:

[0119]

[0120]

[0121] wherein, and denote a weight matrix, and denote a bias parameter. In a specific implementation, the analysis and processing based on the user final embedding and the item final embedding can obtain a final recommendation result, and the specific process can be referred to the description in the related art, which will not be repeated here.

[0122] The embodiment also provides a recommendation model, which comprises the decoupled attribute domain representation module, the target behavior guided attention mechanism module and the decoupled graph attention network module. In the process of training the recommendation model, in order to learn the correlation between the embedding representing the target behavior of the user and the auxiliary behavior, a contrast learning method is used to make the embedding of the target behavior and the auxiliary behavior of the user more consistent. Since the embedding of the auxiliary behavior contains a large amount of noise data, such as click behavior caused by the popularity of the item, the noise information in the embedding of the auxiliary behavior will be taken as a positive sample, which reduces the ability to distinguish between positive samples and negative samples in contrast learning.

[0123] In the embodiment, a denoising contrast learning method is provided to reduce the noise information in the positive sample, so as to maximize the effectiveness of contrast learning in the recommendation model. Through the embedding learning module based on the graph attention network, the user embedding of the auxiliary behavior can be obtained by effectively capturing the target behavior preference of the user-item interaction, which can reduce more noise data in the auxiliary behavior compared with the method based on the graph convolution network commonly used in the prior art.

[0124] As shown in Figure 5 Before contrast learning, a multi-attribute linearization attention mechanism is designed, the user target embedding of the target behavior is regarded as Q, and the user target embedding of the auxiliary behavior is regarded as K and V. First, the user target embedding is projected as follows:

[0125]

[0126]

[0127]

[0128] wherein, represents the user embedding matrix of the target behavior, represents the user embedding matrix of the kth auxiliary behavior of attribute a, represents a parameter matrix. The number of behaviors is K, and the first K-1 behaviors represent auxiliary behaviors, and the Kth behavior represents a target behavior.

[0129] Then, the linearized attention mechanism is used for the user target embedding of the kth auxiliary behavior of attribute a, and the attention weight is calculated using the following function:

[0130]

[0131] wherein, the feature map φ(.) is applied to the matrix Q and K by row. Specifically, φ(x) = elu(x) + 1, elu(.) represents an exponential linear unit activation function, is obtained by Composition. The linearized attention mechanism used in this embodiment can effectively reduce the complexity of the model.

[0132] In the above formula, when x is negative, the user target embedding matrix of the auxiliary behavior k is It can be calculated by adding a slice of the matrix of the attribute dimension as follows:

[0133]

[0134] Then, the user target embedding matrix of the auxiliary behavior can be obtained as And the user target embedding matrix of the target behavior is

[0135] By using contrastive learning to capture the fine-grained consistency between the target behavior and the auxiliary behavior of the user, a more in-depth and accurate understanding of the user's target behavior preference is achieved, and the following is obtained:

[0136]

[0137] In the above formula, τ is the temperature hyperparameter in softmax, and φ(.) represents the inner product of two vectors. The final contrastive loss function is as follows:

[0138]

[0139] In this embodiment, a denoising contrastive learning method is designed to adjust the user preference between the target behavior and the auxiliary behavior. By denoising learning the user's preference for different item attributes in the auxiliary behavior, higher weights are assigned to the user features in the attribute domain more similar to the target behavior, thereby improving the purity of the positive samples and achieving more effective contrastive learning.

[0140] To optimize the current module, in some embodiments, a pair-wise Bayesian personalized ranking (BPR) loss is used for calculation, which makes the similarity between the associated nodes. The BPR loss function is as follows:

[0141]

[0142] Where O represents the training sample set, σ represents the sigmoid function, the superscript T represents the transpose operation, e u represents the embedding of the user node u, e i represents the embedding of the item node i, e j represents the embedding of the item node j.

[0143] In some embodiments, the loss function for training the recommendation model is as follows:

[0144]

[0145] wherein a, b and m are hyperparameters that control the contrast loss, dependency loss and the recommendation task training loss and L2 regularization ratio respectively.

[0146] In the embodiment, the consistency between the target user behavior and the auxiliary user behavior can be better enhanced by the contrast learning through the setting of the contrast loss. In the embodiment, the importance weight of the user's preference for different attribute domains under a specific auxiliary behavior relative to the target user behavior is also learned, and by emphasizing the embedding of the user behavior in the important attribute domain, the noise data in the auxiliary behavior embedding can be effectively reduced, thereby realizing fine-grained contrast learning to reduce the influence of noise data on positive samples.

[0147] Optionally, in some embodiments, the item is a commodity, the target behavior is purchase, and the auxiliary behavior includes at least one of clicking, collecting and commenting.

[0148] The interaction of the user and the item under the plurality of behaviors is used to represent the case that the user performs the target behavior and the auxiliary behavior on the commodity, and the attribute domain is used to represent the attribute of the commodity.

[0149] The recommendation method based on decoupled learning and target behavior oriented learning provided by the embodiment can be used in the scenario of personalized commodity recommendation for users. The item can be a commodity targeted by the user's behavior, and the behavior performed by the user can be purchase, clicking, collecting and commenting, etc. In the embodiment, purchase is determined as the target behavior, and other behaviors except purchase are determined as auxiliary behaviors. Through historical data, data of various operations performed by the user on a plurality of commodities (i.e. the interaction of the user and the item under the plurality of behaviors) can be obtained.

[0150] In the scenario of personalized commodity recommendation, it is usually necessary to make personalized commodity recommendation to the user according to the user's historical behavior, however, the clicking, commenting and collecting behaviors performed by the user on the commodity do not necessarily mean that the user has a purchase preference for the commodity. The recommendation method commonly used in the prior art only learns coarse-grained user preference, and fails to consider the subtle attributes of the commodity focused on by the user in different behaviors. Therefore, when the purchase is taken as the target behavior to make personalized commodity recommendation to the user, the auxiliary behaviors such as clicking, commenting and collecting in the historical data of the user can have a negative impact on the recommendation result.

[0151] By the method provided in the embodiment, the user preference between the target behavior and the auxiliary behavior can be adjusted, the commodity attributes in which the user is interested in the auxiliary behavior can be effectively extracted, the influence of irrelevant commodity attributes is reduced, and then the fine-grained semantic interaction between the user and the commodity in different attribute spaces is captured, and the accuracy of the obtained recommendation result is improved.

[0152] See Figure 6 The embodiment of the application further provides a recommendation device 600 based on decoupling learning and target behavior oriented learning, which comprises:

[0153] A construction module 601 is configured to construct a user-item interaction graph under each behavior based on the interaction between the user and the item under a plurality of behaviors, wherein the user-item interaction graph comprises a user node for representing the user and an item node for representing the item, and the plurality of behaviors comprises a target behavior and an auxiliary behavior.

[0154] A decoupling representation module 602 is configured to perform decoupling attribute domain representation processing on the user-item interaction graph to obtain item decomposition embedding corresponding to the item node and user decomposition embedding corresponding to the user node in a plurality of decoupled attribute domains.

[0155] An attention calculation module 603 is configured to calculate, for each attribute domain corresponding to a plurality of attribute domains of each user-item interaction graph, an attention score between the user node and its adjacent item node in the attribute domain based on the item decomposition embedding and the user decomposition embedding.

[0156] An aggregation processing module 604 is configured to perform aggregation processing on the user node and the item node based on the attention score to obtain user target embedding and item target embedding under each behavior.

[0157] An analysis processing module 605 is configured to perform analysis processing on the user target embedding and the item target embedding under a plurality of behaviors to obtain a recommendation result.

[0158] Optionally, the decoupling representation module 602 comprises:

[0159] A first processing unit is configured to construct a plurality of decoupled attribute domains, and obtain user initial embedding corresponding to the user node and item initial embedding corresponding to the item node.

[0160] A decomposition unit is configured to decompose the item initial embedding in each attribute domain to obtain item decomposition embedding in the attribute domain, and determine user decomposition embedding in each attribute domain based on the user initial embedding.

[0161] Optionally, the decomposition unit is specifically configured to:

[0162] In each of the attribute domains, the project initial embedding is projected into the embedding of the attribute domain using the pre-learned projection matrix to obtain the project decomposition embedding in the attribute domain. The project decomposition embedding in the attribute domain is:

[0163]

[0164] in, The item decomposition embedding e is used to represent the item node i in the attribute domain a. i The initial embedding of the item used to characterize item node i, is used to represent the projection matrix, which is a learnable parameter matrix, and ||||2 is used to represent the L2 norm.

[0165] Optionally, the attention calculation module 603 includes:

[0166] A first determining unit is configured to determine an average value of item decomposition embeddings of all adjacent item nodes corresponding to the user node in the attribute domain as a query vector corresponding to the user node;

[0167] A calculation unit is used to use the query vector corresponding to the user node to calculate the attention score between the user node and its adjacent item nodes in the attribute domain.

[0168] Optionally, the attention score between the user node and its adjacent item node is:

[0169]

[0170] in, It is used to represent the query vector corresponding to the user node u in the attribute domain a under the target behavior. are pre-learned parameters, [;] are used to represent the concatenation of vectors, Item decomposition embedding used to represent item node i under attribute domain a.

[0171] Optionally, the aggregation processing module 604 includes:

[0172] an aggregation unit, configured to aggregate the user node and its adjacent item nodes based on the attention score to obtain a user aggregate embedding, and to aggregate the item node and its adjacent user nodes to obtain an item aggregate embedding;

[0173] An adding unit is configured to add the user aggregated embeddings in all the attribute domains under the same user-item interaction graph to obtain a user target embedding under the behavior, and add the item aggregated embeddings in all the attribute domains under the same user-item interaction graph to obtain an item target embedding under the behavior.

[0174] Optionally, the analysis processing module 605 comprises:

[0175] A second determining unit is configured to determine a user final embedding based on the user target embeddings under all the behaviors, and determine an item final embedding based on the item target embeddings under all the behaviors;

[0176] An analysis processing unit is configured to perform analysis processing based on the user final embedding and the item final embedding to obtain the recommendation result.

[0177] Optionally, in some embodiments, the item is a commodity, the target behavior is purchase, and the auxiliary behaviors include at least one of clicking, collecting and commenting.

[0178] The interaction conditions of the user and the item under the multiple behaviors are used to represent the conditions of the user performing the target behavior and the auxiliary behaviors on the commodity, and the attribute domains are used to represent the attributes of the commodity.

[0179] The recommendation apparatus 600 based on decoupled learning and target behavior oriented learning provided by the embodiments of the present application can execute the above-mentioned method embodiments, and has similar implementation principles and technical effects, which will not be described here again.

[0180] It should be noted that the division of units in the embodiments of the present application is illustrative, and is merely a logical function division. In actual implementation, another division manner can be used. In addition, each functional unit in each embodiment of the present application can be integrated in one processing unit, or each unit can be physically present alone, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0181] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a processor-readable storage medium. Based on such understanding, the technical solutions of the present application, essentially or in other words, the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to perform all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0182] As shown in Figure 7 The embodiments of the present application provide an electronic device 700, which comprises a memory 702, a processor 701, and a program stored in the memory 702 and capable of running on the processor 701; the processor 701 is used for reading the program in the memory 702 to realize the steps in the recommendation method based on decoupled learning and target behavior-oriented learning as described above.

[0183] The embodiment of the present application further provides a readable storage medium, and a program is stored on the readable storage medium. The program is executed by a processor to implement each process of the embodiment of the recommendation method based on decoupling learning and target behavior-oriented learning, and the same technical effects can be achieved. To avoid repetition, details are not described herein. The readable storage medium can be any available medium or data storage device accessible by the processor, including but not limited to a magnetic memory (for example, a floppy disk, a hard disk, a magnetic tape, a magneto-optical disk (MO), etc.), an optical memory (for example, a compact disk (CD), a digital versatile disk (DVD), a Blu-ray disk (BD), a high-definition versatile disk (HVD), etc.), and a semiconductor memory (for example, a read-only memory (ROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a non-volatile memory (NAND FLASH), a solid state disk (SSD), etc.).

[0184] It should be noted that in this document, the terms "comprising", "including", or any other variant thereof are intended to cover non-exclusive inclusions, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or further includes elements inherent to such a process, method, article or device. Without more limitations, the element defined by the statement "including a" does not exclude the presence of other identical elements in the process, method, article or device including the element.

[0185] Those skilled in the art can clearly understand the above-mentioned embodiment method can be realized by means of software and necessary general hardware platform, of course, also can be realized by hardware, but in many cases, the former is a better embodiment. According to such understanding, the technical scheme of the present application can be embodied in the form of software product, and the computer software product is stored in a storage medium (such as ROM / RAM, disk, optical disk), including a plurality of instructions to make a terminal (which can be a mobile phone, computer, server, air conditioner, or network equipment, etc.) execute the method described in various embodiments of the present application.

[0186] The embodiments of the present application are described above in combination with the drawings, but the present application is not limited to the above-mentioned specific embodiments, and the above-mentioned specific embodiments are only illustrative, not limiting, and those skilled in the art can make many forms without departing from the purpose of the present application and the scope protected by the claims under the inspiration of the present application, which all belong to the protection of the present application.

Claims

1. A recommendation method based on decoupled learning and goal behavior oriented learning, characterized in that, The method comprises the following steps: Based on the interaction between the user and the project under multiple behaviors, a user-project interaction graph under each behavior is constructed, the user-project interaction graph comprises a user node for representing the user and a project node for representing the project, and the multiple behaviors comprise a target behavior and an auxiliary behavior; Decoupling attribute domain representation processing is performed on the user-project interaction graph to obtain project decomposition embeddings corresponding to the project nodes and user decomposition embeddings corresponding to the user nodes in multiple decoupled attribute domains; For multiple attribute domains corresponding to each user-project interaction graph, attention scores between the user nodes and their adjacent project nodes under each attribute domain are calculated based on the project decomposition embeddings and the user decomposition embeddings; Based on the attention scores, the user nodes and the project nodes are aggregated to obtain user target embeddings and project target embeddings under each behavior; The user target embeddings and the project target embeddings under multiple behaviors are combined for analysis and processing to obtain a recommendation result; Wherein, the attention scores between the user nodes and their adjacent project nodes under each attribute domain are calculated based on the project decomposition embeddings and the user decomposition embeddings, comprising: The average value of the project decomposition embeddings of all adjacent project nodes corresponding to the user node in the attribute domain is determined as the query vector corresponding to the user node; The attention scores between the user nodes and their adjacent project nodes in the attribute domain are calculated using the query vector corresponding to the user node; Wherein, based on the attention scores, the user nodes and the project nodes are aggregated to obtain user target embeddings and project target embeddings under each behavior, comprising: Based on the attention scores, the user nodes and their adjacent project nodes are aggregated to obtain user aggregated embeddings, and the project nodes and their adjacent user nodes are aggregated to obtain project aggregated embeddings; The user target embeddings under the behavior are obtained by adding the user aggregated embeddings in all attribute domains under the same user-project interaction graph, and the project target embeddings under the behavior are obtained by adding the project aggregated embeddings in all attribute domains under the same user-project interaction graph; Wherein, the decoupling attribute domain representation processing is performed on the user-project interaction graph to obtain the project decomposition embeddings corresponding to the project nodes and the user decomposition embeddings corresponding to the user nodes in multiple decoupled attribute domains, comprising: Multiple decoupled attribute domains are constructed, and user initial embeddings corresponding to the user nodes and project initial embeddings corresponding to the project nodes are obtained; The project initial embeddings are decomposed in each attribute domain to obtain project decomposition embeddings in the attribute domain, and user decomposition embeddings in each attribute domain are determined based on the user initial embeddings; Wherein, the project initial embeddings are decomposed in each attribute domain to obtain project decomposition embeddings in the attribute domain, comprising: In each of the attribute domains, the item initial embedding is projected into an embedding of the attribute domain by using a pre-learned projection matrix, to obtain an item decomposition embedding in the attribute domain.

2. The method of claim 1 wherein, The item decomposition embedding in the attribute domain is: wherein, a project decomposition embedding e for characterizing an item node i in an attribute domain a, i an item initial embedding for characterizing an item node i, for characterizing the projection matrix, the projection matrix being a learnable parameter matrix, || ||2 for characterizing the L2 norm.

3. The method of claim 1 wherein, The attention score between the user node and its adjacent item node is: wherein, for characterizing a query vector corresponding to the user node u in the attribute domain a under the target behavior, is a pre-learned parameter, [;] for characterizing the concatenation of vectors, for characterizing an item decomposition embedding of the item node i under the attribute domain a.

4. The method of claim 1 wherein, The item is a commodity, the target behavior is purchase, and the auxiliary behavior includes at least one of clicking, collecting, and commenting. The interaction of the user and the item under the plurality of behaviors is used to represent the case that the user performs the target behavior and the auxiliary behavior on the commodity, and the attribute domain is used to represent the attribute of the commodity.

5. A recommendation apparatus based on decoupled learning and goal behavior oriented learning, characterized by, Comprise: A construction module is configured to construct a user-item interaction graph under each of the behaviors based on the interaction of the user and the item under the plurality of behaviors, the user-item interaction graph comprising a user node for representing the user and an item node for representing the item, the plurality of behaviors comprising a target behavior and an auxiliary behavior; A decoupling representation module is configured to perform decoupling attribute domain representation processing on the user-item interaction graph to obtain item decomposition embeddings of the item node and user decomposition embeddings of the user node in a plurality of decoupled attribute domains; An attention calculation module is configured to calculate, for each of a plurality of attribute domains corresponding to the user-item interaction graph, an attention score between the user node and its adjacent item node in the attribute domain based on the item decomposition embedding and the user decomposition embedding; An aggregation processing module is configured to perform aggregation processing on the user node and the item node based on the attention score to obtain a user target embedding and an item target embedding under each of the behaviors; An analysis processing module is configured to perform analysis processing in combination with the user target embedding and the item target embedding under the plurality of behaviors to obtain a recommendation result. The attention calculation module comprises: A first determination unit is configured to determine an average value of item decomposition embeddings of all adjacent item nodes corresponding to the user node in the attribute domain as a query vector corresponding to the user node; A calculation unit is configured to calculate the attention score between the user node and its adjacent item node in the attribute domain by using the query vector corresponding to the user node; The aggregation processing module comprises: An aggregation unit is configured to aggregate the user node and its adjacent item node based on the attention score to obtain a user aggregated embedding, and aggregate the item node and its adjacent user node to obtain an item aggregated embedding; An addition unit is configured to add the user aggregated embeddings in all the attribute domains under the same user-item interaction graph to obtain a user target embedding under the behavior, and add the item aggregated embeddings in all the attribute domains under the same user-item interaction graph to obtain an item target embedding under the behavior. The decoupling representation module comprises: A first processing unit is configured to construct a plurality of decoupled attribute domains, and obtain a user initial embedding corresponding to the user node and an item initial embedding corresponding to the item node; a decomposition unit configured to decompose the item initial embedding in each of the attribute domains to obtain an item decomposition embedding in the attribute domain, and determine a user decomposition embedding in each of the attribute domains based on the user initial embedding; wherein the decomposition unit is specifically configured to: in each of the attribute domains, project the item initial embedding into an embedding of the attribute domain by using a pre-learned projection matrix to obtain the item decomposition embedding in the attribute domain.

6. An electronic device comprising: A memory, a processor, and a program stored in the memory and capable of running on the processor; the processor is configured to read the program in the memory to implement the steps in the recommendation method based on decoupled learning and goal behavior oriented learning according to any one of claims 1 to 4.

7. A readable storage medium for storing a program, characterized by, The program is executed by the processor to implement the steps in the recommendation method based on decoupled learning and goal behavior oriented learning according to any one of claims 1 to 4.

Images