Recommendation method and device

Abstract

The invention provides a recommendation method and device. The recommendation method comprises the steps of obtaining recommendation request information; a cross-domain enhanced refined embedded feature set corresponding to a target user is searched from a refined embedded feature database, and cross-domain enhanced user refined embedded features are determined based on user refined embedded features of the target user in a target domain and preference features of a paired user in a source domain. The paired users are determined based on user refined embedding features and user intention embedding features corresponding to the target users in the target domain and user refined embedding features and user intention embedding features corresponding to the users in the source domain; and based on the context information and the refined embedded feature set, determining a recommendation result of the target user in the target domain. Cross-domain links are established through cross-domain user intention pairing, and preference knowledge of cross-domain users with similar intentions is migrated, so that a more accurate recommendation result is obtained.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence (AI) technology, and in particular to a recommendation method and device. Background Art

[0002] The rapid development of computer technology and the internet has brought about tremendous changes to society. While improving people's lives, it has also introduced new challenges, such as information overload. As data volumes continue to increase, recommendation systems are playing an increasingly important role. They not only provide users with effective information from massive amounts of data, reducing the difficulty of obtaining information, but also enhance the usability and value of data.

[0003] Due to the closed nature of the recommendation field, individual recommendation models may tend to recommend certain categories of items to users, leading to the phenomenon of information cocoons. Therefore, achieving diversity in user recommendation results has become a key issue. Furthermore, recommendation systems often require a large amount of historical interaction records to achieve effective recommendations. How to implement recommendations for a system with less historical data and achieve good results is also a current research focus in this field. Summary of the Invention

[0004] The embodiments of the present application provide a recommendation method and device, which establish cross-domain links by pairing cross-domain user intentions and transfer preference knowledge (also referred to as preference information) of cross-domain users with similar intentions to make the recommendation results more diverse and more accurate.

[0005] In a first aspect, the present application provides a recommendation method, comprising obtaining recommendation request information, the recommendation request information carrying target user information and target user context information in a target domain; searching from a refined embedding feature database based on the target user information to obtain a refined embedding feature set corresponding to the target user, the refined embedding feature set including cross-domain (CD) enhanced user refined embedding features, item refined embedding features, and context refined embedding features, wherein the cross-domain enhanced user refined embedding features are determined based on the first user refined embedding features and the preference features of the paired user in the source domain, the first user refined embedding features being the user refined embedding features corresponding to the target user in the target domain (targetdomain), and the paired user in the source domain (source domain), indicating preference information determined based on historical interaction information corresponding to the paired user in the source domain; the paired user is determined based on the refined user embedding features and user intention embedding features corresponding to the target user in the target domain, and the refined user embedding features and user intention embedding features corresponding to each user in the source domain; the first user refined embedding features, the item refined embedding features, and the context refined embedding features are obtained based on feature extraction of the target user's historical interaction information in the target domain, and the target domain and the source domain are different recommendation domains; based on the context information and the refined embedding feature set, determining the recommendation result of the target user in the target domain.

[0006] Optionally, the target user information carried in the recommendation request may be the target user's identity information. The target user's identity information may be any information that can identify the user, such as the target user's identity document (ID). For example, user Li Si opens a food ordering app, triggering the generation of a recommendation request that carries user Li Si's ID.

[0007] The recommendation method provided in this application constructs cross-domain links between users in the source domain and users in the target domain by embedding features of user intentions, then transfers the preference knowledge of users paired with target users from the source domain through cross-domain links, and then uses the preference knowledge of users in the source domain to enhance the preference knowledge of users in the target domain, and then uses the enhanced preference knowledge to predict recommendation results, breaking through the closed target recommendation domain, achieving diversity in recommendation results, and improving the accuracy of recommendation results.

[0008] In one possible implementation, a specific implementation of determining paired users based on the refined user embedding features and user intention embedding features corresponding to the target user in the target domain, and the refined user embedding features and user intention embedding features corresponding to each user in the source domain is as follows: determining the position distribution of each user in the target domain based on the refined user embedding features and user intention embedding features corresponding to each user in the target domain; determining the position distribution of each user in the source domain based on the refined user embedding features and user intention embedding features corresponding to each user in the source domain; determining a transformation matrix that transforms the position distribution of each user in the target domain into the position distribution of each user in the source domain with the goal of minimizing transmission cost; determining a paired user to be paired with the target user from each user in the source domain based on the refined user embedding features and user intention embedding features corresponding to the target user in the target domain, and the transformation matrix.

[0009] By modeling the link construction between cross-domain users as an optimal transport (OT) problem, the user refined embedding features and user intention embedding features of each user in the source domain, as well as the user refined embedding features and user intention embedding features of each user in the target domain, are used to determine the transformation matrix T. Then, the transformation matrix T is used to find paired users with the target user from the source domain. In other words, the obtained transformation matrix T can be used to establish links between users in the two domains (source domain and target domain), thereby achieving accurate pairing (also known as link establishment) between users in the source domain and users in the target domain. Preference knowledge is then transferred between the paired users, assisting in recommending target users in the target domain, making the recommendation results more accurate.

[0010] In another possible implementation, a specific implementation of determining the cross-domain enhanced user refined embedding feature based on the first user refined embedding feature and the preference feature of the paired user in the source domain is: based on the transformation matrix, the preference feature of the paired user in the source domain is converted to the target domain for representation to obtain the cross-domain preference feature; the cross-domain preference feature and the first user refined embedding feature are fused to obtain the cross-domain enhanced user refined embedding feature.

[0011] Using the transformation matrix T, the preference features of the paired users in the source domain who are paired with the target user are transferred to the target user in the target domain, and the cross-domain transferred knowledge embedding of the target user is obtained. This knowledge embedding is then fused with the refined embedding features of the target user in the target domain to obtain the cross-domain enhanced user refined embedding features. The cross-domain enhanced user refined embedding features include more interactive preference information and contextual preference information. Using the cross-domain enhanced user refined embedding features for recommendation prediction can obtain more accurate recommendation results and improve user experience.

[0012] In another possible implementation, the paired user's preference features in the source domain include the second user's refined embedding features and / or the paired user's user intent embedding features in the source domain, where the second user's refined embedding features are the paired user's corresponding refined embedding features in the source domain. In other words, the paired user's preference features in the source domain include the paired user's refined embedding features and user intent embedding features in the source domain.

[0013] In one example, the user intention embedding feature corresponding to the target user in the target domain is determined based on multiple user event intention embedding features corresponding to the target user in the target domain, and the multiple user event intention embedding features corresponding to the target user in the target domain are obtained based on the intention feature extraction of multiple historical interaction events corresponding to the target user in the target domain; the user intention embedding feature corresponding to each user in the source domain is determined based on multiple user event intention embedding features corresponding to each user in the source domain, and the multiple user event intention embedding features corresponding to each user in the source domain are obtained based on the intention feature extraction of multiple historical interaction events corresponding to each user in the source domain.

[0014] In another possible implementation, a specific implementation of the first user refined embedding feature, item refined embedding feature and context refined embedding feature is obtained based on feature extraction of the target user's historical interaction information in the target domain: based on the target user's historical interaction information in the target domain, a user event graph is constructed, wherein the user event graph includes user nodes, item nodes, context nodes and intention nodes, and the intention nodes are connected to the user nodes, item nodes and context nodes respectively. The user node is the node where the user feature corresponding to the target user is located, the item node is the node where the item feature corresponding to the item having an interactive relationship with the target user is located, the context node is the node where the context feature corresponding to the context of the target user's interaction with the item is located, and the intention node is the node where the intention feature corresponding to the intention of the target user's interaction with the item is located; the user event graph is used as the input of the graph convolutional neural network, and after the graph convolution operation of each node in the user event graph is performed by multiple layers of graph convolutional neural network layers, the first user refined embedding feature, item refined embedding feature and context refined embedding feature are output.

[0015] In this possible implementation, a user event graph is constructed based on historical interaction information in the recommendation field, and user nodes, item nodes, and context nodes are associated through intent nodes, so that context features can be incorporated into graph convolution, considering context information during interaction, and incorporating context information into user intent nodes, thereby increasing the accuracy of subsequent use of user intent to construct links between the source domain and the target domain.

[0016] In another possible implementation, the intent feature of the intent node is determined based on the user feature of the user node connected to the intent node and the context feature of the context node. That is, the user's interaction preference and context preference can be captured through the intent node.

[0017] In another possible implementation, the user event graph is used as the input of the graph convolutional neural network, and a specific implementation of outputting the first user refined embedding feature is as follows: performing a graph convolution operation on the intent feature of each intent node connected to the user node in the previous layer, and the item feature of the item node connected to each intent node, to obtain each sub-user feature of the user node in the current layer; performing an aggregation operation on each sub-user feature of the user node to obtain the embedding feature of the user node in the current layer; and obtaining the first user refined embedding feature based on the embedded features of the user node after being updated by each graph convolutional neural network layer in the multi-layer graph convolutional neural network layer.

[0018] By using the intention node as the center, information about items and context features is propagated to the user, making it easier for the user node to capture the interactive item features and context features. The refined user embedding features finally obtained by the graph convolution capture the type of items that the user will interact with under those context features.

[0019] In another possible implementation, the user event graph is used as the input of the graph convolutional neural network. After the multi-layer graph convolutional neural network layer performs graph convolution operations on each node in the user event graph, a specific implementation of outputting the refined embedding features of the item is as follows: performing graph convolution operations on the intent features of each intent node connected to the item node in the previous layer and the user features of the user node connected to each intent node to obtain each sub-item feature of the item node in the current layer; performing aggregation operations on each sub-item feature of the item node to obtain the embedding features of the item node in the current layer; and obtaining the refined embedding features of the item node based on the embedded features of the updated item nodes in each layer of the multi-layer graph convolutional neural network layer.

[0020] By using the intent node as the central node, information about users and context features is propagated to items, making it easier for item nodes to capture the user features and context features connected to them. This allows the refined item embedding features ultimately obtained by graph convolution to capture the type of user interacting with the item under that context feature.

[0021] In another possible implementation, the user event graph is used as the input of the graph convolutional neural network, and a specific implementation of outputting the context-refined embedding feature is as follows: a graph convolution operation is performed on the user features of the user node connected to the intention node in the previous layer and the item features of the item node to obtain the intention features of the intention node in the current layer; based on the intention features of the intention node in the current layer and the attention weights of the context nodes connected to the intention node, each sub-context feature of the context node in the current layer is determined; an aggregation operation is performed on each sub-context feature of the context node in the current layer to obtain the embedding features of the context node in the current layer; and the context-refined embedding feature is obtained based on the embedding features of the context nodes updated by each graph convolutional neural network layer in the multi-layer graph convolutional neural network layer.

[0022] In another possible implementation, a specific implementation of determining the recommendation results for the target user in the target domain based on the context information and the refined embedded feature set is as follows: using the context information and the refined embedded feature set as input to the recommendation network, outputting the target user's interaction probability for each item in the target domain; and determining the recommendation results for the target user in the target domain based on the target user's interaction probability for each item in the target domain.

[0023] By using the refined feature set after cross-domain enhancement to predict recommendation results, on the one hand, we can break through the closed target recommendation field and the information cocoon to achieve diversity in recommendation results. On the other hand, we can integrate the refined embedded feature set of user preference knowledge from other domains (source domains) to effectively capture user interests and improve the accuracy of recommendation results for users in the target domain.

[0024] Optionally, the recommendation network includes a decoder that uses refined embeddings enhanced with cross-domain information to predict the user's preference for items in context. Factorization machines (FMs) can be used as the decoder implementation. FMs are the core component of the recommendation network. FMs are linear models with better interpretability and lower complexity than multi-layer perceptrons (MLPs).

[0025] In another possible implementation, the recommendation method provided by the present application also includes: detecting that the target user performs a target operation on the recommendation result display page, the target operation is used to request a result analysis page for displaying the recommendation results; in response to the target operation, displaying the result analysis page, the result analysis page includes an explanatory diagram of the recommendation results, and the explanatory diagram at least indicates the process of migrating the preference information of the paired user in the source domain to the target user in the target domain.

[0026] Thanks to the method of pairing users in the source and target domains through user intent embedding features and the method of processing user event graphs through graph convolution, user intentions and contextual preferences are traceable, so a user interface (UI) can be implemented to demonstrate the cross-domain information transfer process and explain the recommendation results.

[0027] In another possible implementation, the preference information of the paired users in the source domain includes the interaction preference distribution information and context preference distribution information of the paired users in the source domain, the explanatory diagram indicates the interaction preference distribution and context preference distribution information of the paired users in the source domain, as well as the interaction preference distribution information and context preference distribution information of the target user in the target domain after aggregating the interaction preference distribution information and context preference distribution information of the paired users in the source domain, and the recommendation results determined based on the interaction preference distribution information and context preference distribution information of the target user in the target domain; wherein, the interaction preference distribution information of the paired users in the source domain indicates the preference degree of the paired users for each item feature in the source domain, the context preference distribution information of the paired users in the source domain indicates the preference degree of the paired users for each context feature in the source domain, the interaction preference distribution information of the target users in the target domain indicates the preference degree of the target users for each item feature in the target domain, and the context preference distribution information of the target users in the target domain indicates the preference degree of the target user for each context feature in the target domain.

[0028] Through the explanatory diagram, the reasons for the recommendation results are shown to users, such as the transfer process of the user's preference knowledge in the source domain to the target user in the target domain, the user's preference for each context, and the preference for each item feature. In this way, users can have a clear understanding of the reasons for the recommendation results.

[0029] For example, when a user clicks the "?" icon on the recommendation interface, a result analysis interface is displayed to the user in response to this click. This interface displays an explanation diagram, which shows the contextual preference distribution and interaction preference distribution of users in the source domain paired with the target user, an illustration of the transfer and fusion of the source domain user's preference distribution knowledge to the target user in the target domain, and the contextual preference distribution and interaction preference distribution of the target user in the target domain that has fused the source domain user knowledge. Optionally, the contextual preference distribution can be displayed as line weights, and the interaction preference distribution can also be displayed as line weights, which clearly reflects the specific information about items that users attach greater importance to in certain contextual scenarios. In other words, the recommendation results are highly interpretable, improving the user experience.

[0030] In another possible implementation, the historical interaction information in the target domain is located in the terminal device; the feature extraction step of obtaining the first user refined embedding features, the item refined embedding features, and the context refined embedding features based on the historical interaction information of the target user in the target domain is performed in the terminal device; and the step of determining the cross-domain enhanced user refined embedding features based on the first user refined embedding features and the preference features of the paired user in the source domain is performed in the cloud.

[0031] The recommended method of this application is implemented using an end-cloud collaborative architecture, and the feature extraction step of the user's historical interaction information is completed in the terminal device, and the user intention matching is completed in the cloud. This achieves accurate identification of user intentions and completion of cross-domain data transmission operations without uploading user data to the cloud, thereby protecting user privacy to the greatest extent possible.

[0032] In another possible implementation, the refined embedded feature database is updated according to a preset period. For example, the user event graph is updated every 48 hours using the historical interaction data generated within 48 hours, and then the refined embedded feature set corresponding to each user is updated to ensure the quality and freshness of the refined embedded feature set of each user in the refined embedded feature library.

[0033] In a second aspect, the present application provides a recommendation method, including obtaining recommendation request information, which carries information of a target user; determining refined user embedding features and user intention embedding features of the target user in a source domain, wherein the refined user embedding features and user intention embedding features of the target user in the source domain are obtained based on feature extraction of historical interaction information of the target user in the source domain; determining a paired user paired with the target user in the target domain based on the refined user embedding features and user intention embedding features corresponding to the target user in the source domain, and the refined user embedding features and user intention embedding features of each user in the target domain, wherein the refined user embedding features and user intention embedding features of each user in the target domain are obtained based on feature extraction of historical interaction information of each user in the target domain, the source domain and the target domain are different recommendation domains, and the number of interaction events of the target user in the target domain is less than the number of interaction events of the target user in the source domain; and determining a recommendation result for the target user in the target domain based on the historical interaction information of the paired users in the target domain.

[0034] By embedding user intent features, we can identify paired users with similar preferences to the target user in the target domain. We can then leverage the historical interaction information of paired users in the target domain to make recommendations to the target user in the target domain, addressing the issue of limited target user interaction data and poor recommendation accuracy. For example, for a new recommendation system, target user A has almost zero interaction data. The user event graph constructed based on this interaction data contains little user preference knowledge, making it impossible to predict useful recommendation results. In this case, we can embed user intent features (including interaction preference distribution and contextual preference distribution) based on target user A's source domain, which has richer interaction data. We can then identify user B in the target domain who has an intent pairing with the target user. We can then recommend items that user B interacted with in similar contextual scenarios in the target domain to target user A, enabling us to recommend items of interest to target user A in a cold start scenario.

[0035] In one possible implementation, based on the refined user embedding features and user intention embedding features of the target user in the source domain, as well as the refined user embedding features and user intention embedding features of each user in the target domain, a paired user to be paired with the target user in the target domain is determined. A specific implementation is: based on the refined user embedding features and user intention embedding features corresponding to each user in the target domain, the position distribution of each user in the target domain is determined; based on the refined user embedding features and user intention embedding features corresponding to each user in the source domain, the position distribution of the said each user in the source domain is determined; with the goal of minimizing the transmission cost, a transformation matrix is ​​determined to transform the position distribution of each user in the source domain into the position distribution of each user in the target domain; based on the refined user embedding features, user intention embedding features, and the transformation matrix corresponding to the target user in the source domain, the paired user to be paired with the target user is determined from each user in the target domain.

[0036] In another possible implementation, the user intention embedding feature corresponding to the target user in the source domain is determined based on the multiple user event intention embedding features corresponding to the target user in the source domain, and the multiple user event intention embedding features corresponding to the target user in the source domain are obtained based on the intention feature extraction of the multiple historical interaction events corresponding to the target user in the source domain; the user intention embedding feature corresponding to each user in the target domain is determined based on the multiple user event intention embedding features corresponding to each user in the target domain, and the multiple user event intention embedding features corresponding to each user in the target domain are obtained based on the intention feature extraction of the multiple historical interaction events corresponding to each user in the target domain.

[0037] In another possible implementation, a specific implementation of determining the refined user embedding features of the target user in the source domain is: based on the historical interaction information of the target user in the source domain, constructing a user event graph, wherein the user event graph includes user nodes, item nodes, context nodes and intention nodes, and the intention nodes are connected to the user nodes, item nodes and context nodes respectively. The user node is the node where the user features corresponding to the target user are located, the item node is the node where the item features corresponding to the item having an interactive relationship with the target user are located, the context node is the node where the context features corresponding to the context of the target user's interaction with the item are located, and the intention node is the node where the intention features corresponding to the intention of the target user's interaction with the item are located; the user event graph is used as the input of the graph convolutional neural network, and after the graph convolution operation of each node in the user event graph is performed by multiple layers of graph convolutional neural network layers, the refined user embedding features of the target user in the source domain are output.

[0038] In another possible implementation, the user event graph is used as the input of the graph convolutional neural network. After the graph convolution operation of each node in the user event graph is performed by multiple layers of graph convolutional neural network layers, the refined user embedding features of the target user in the source domain are output. A specific implementation is as follows: the intent features of each intent node connected to the user node in the previous layer and the item features of the item node connected to each intent node are performed to obtain each sub-user feature of the user node in the current layer; each sub-user feature of the user node is aggregated to obtain the embedding features of the user node in the current layer; based on the embedded features of the user nodes updated by each layer of the graph convolutional neural network layer in the multi-layer graph convolutional neural network layer, the refined user embedding features of the target user in the source domain are obtained.

[0039] In another possible implementation, the recommendation method provided by the present application also includes: detecting that the target user performs a target operation on the recommendation result display page, the target operation is used to request a result analysis page for displaying the recommendation result; in response to the target operation, the result analysis page is displayed, the result analysis page includes an explanatory diagram of the recommendation result, and the explanatory diagram at least indicates the pairing process of the target user in the source domain and the paired user in the target domain.

[0040] In another possible implementation, the explanatory diagram indicates the target user's interaction preference distribution information and contextual preference distribution information in the source domain, as well as the paired user's interaction preference distribution information and contextual preference distribution information in the target domain, and the recommendation results determined based on the paired user's historical interaction information in the target domain; wherein, the target user's interaction preference distribution information in the source domain indicates the target user's preference degree for each item feature in the source domain, the target user's contextual preference distribution information in the source domain indicates the target user's preference degree for each contextual feature in the source domain, the paired user's interaction preference distribution information in the target domain indicates the paired user's preference degree for each item feature in the target domain, and the paired user's contextual preference distribution information in the target domain indicates the paired user's preference degree for each contextual feature in the target domain.

[0041] On the third aspect, the present application provides a recommendation device, including an acquisition module, a search module and a recommendation module, wherein the acquisition module is used to obtain recommendation request information, and the recommendation request information carries the information of the target user and the context information of the target user in the target domain; the search module is used to search from the refined embedding feature database based on the information of the target user to obtain the refined embedding feature set corresponding to the target user, and the refined embedding feature set includes cross-domain enhanced user refined embedding features, item refined embedding features and context refined embedding features, wherein the cross-domain enhanced user refined embedding features are determined based on the first user refined embedding features and the preference features of the paired user in the source domain, and the first user refined embedding features are the target user's context information in the target domain. The corresponding user refined embedding features in the target domain, the preference features of the paired user in the source domain indicate the preference information determined based on the corresponding historical interaction information of the paired user in the source domain, the paired user is determined based on the user refined embedding features and user intention embedding features corresponding to the target user in the target domain, and the user refined embedding features and user intention embedding features corresponding to each user in the source domain, the first user refined embedding features, item refined embedding features and context refined embedding features are obtained based on feature extraction of the historical interaction information of the target user in the target domain, and the target domain and source domain are different recommendation domains; the recommendation module is used to determine the recommendation results of the target user in the target domain based on the context information and the refined embedding feature set.

[0042] In another possible implementation, a specific implementation of determining paired users based on the user refined embedding features and user intention embedding features corresponding to the target user in the target domain, and the user refined embedding features and user intention embedding features corresponding to each user in the source domain is as follows: determining the position distribution of each user in the target domain based on the user refined embedding features and user intention embedding features corresponding to each user in the target domain; determining the position distribution of each user in the source domain based on the user refined embedding features and user intention embedding features corresponding to each user in the source domain; determining a transformation matrix that transforms the position distribution of each user in the target domain into the position distribution of each user in the source domain with the goal of minimizing transmission cost; determining a paired user to be paired with the target user from each user in the source domain based on the user refined embedding features and user intention embedding features corresponding to the target user in the target domain, and the transformation matrix.

[0043] In another possible implementation, a specific implementation of determining the cross-domain enhanced user refined embedding feature based on the first user refined embedding feature and the preference feature of the paired user in the source domain is: based on the transformation matrix, the preference feature of the paired user in the source domain is converted to the target domain for representation to obtain the cross-domain preference feature; the cross-domain preference feature and the first user refined embedding feature are fused to obtain the cross-domain enhanced user refined embedding feature.

[0044] In another possible implementation, the preference feature of the paired user in the source domain includes a second user refined embedding feature and / or a user intention embedding feature of the paired user in the source domain, and the second user refined embedding feature is the corresponding user refined embedding feature of the paired user in the source domain.

[0045] In another possible implementation, the user intention embedding feature corresponding to the target user in the target domain is determined based on the multiple user event intention embedding features corresponding to the target user in the target domain, and the multiple user event intention embedding features corresponding to the target user in the target domain are obtained based on the intention feature extraction of the multiple historical interaction events corresponding to the target user in the target domain; the user intention embedding feature corresponding to each user in the source domain is determined based on the multiple user event intention embedding features corresponding to each user in the source domain, and the multiple user event intention embedding features corresponding to each user in the source domain are obtained based on the intention feature extraction of the multiple historical interaction events corresponding to each user in the source domain.

[0046] In another possible implementation, a specific implementation of the first user refined embedding feature, item refined embedding feature and context refined embedding feature is obtained based on feature extraction of the target user's historical interaction information in the target domain: based on the target user's historical interaction information in the target domain, a user event graph is constructed, wherein the user event graph includes user nodes, item nodes, context nodes and intention nodes, and the intention nodes are connected to the user nodes, item nodes and context nodes respectively. The user node is the node where the user feature corresponding to the target user is located, the item node is the node where the item feature corresponding to the item having an interactive relationship with the target user is located, the context node is the node where the context feature corresponding to the context of the target user's interaction with the item is located, and the intention node is the node where the intention feature corresponding to the intention of the target user's interaction with the item is located; the user event graph is used as the input of the graph convolutional neural network, and after the graph convolution operation of each node in the user event graph is performed by multiple layers of graph convolutional neural network layers, the first user refined embedding feature, item refined embedding feature and context refined embedding feature are output.

[0047] In another possible implementation, the intent feature of the intent node is determined based on the user feature on the user node connected to the intent node and the context feature on the context node.

[0048] In another possible implementation, the user event graph is used as the input of the graph convolutional neural network, and a specific implementation of outputting the first user refined embedding feature is as follows: performing a graph convolution operation on the intent feature of each intent node connected to the user node in the previous layer, and the item feature of the item node connected to each intent node, to obtain each sub-user feature of the user node in the current layer; performing an aggregation operation on each sub-user feature of the user node to obtain the embedding feature of the user node in the current layer; and obtaining the first user refined embedding feature based on the embedded features of the user node after being updated by each graph convolutional neural network layer in the multi-layer graph convolutional neural network layer.

[0049] In another possible implementation, the user event graph is used as the input of the graph convolutional neural network. After the multi-layer graph convolutional neural network layer performs graph convolution operations on each node in the user event graph, a specific implementation of outputting the refined embedding features of the item is as follows: performing graph convolution operations on the intent features of each intent node connected to the item node in the previous layer and the user features of the user node connected to each intent node to obtain each sub-item feature of the item node in the current layer; performing aggregation operations on each sub-item feature of the item node to obtain the embedding features of the item node in the current layer; and obtaining the refined embedding features of the item node based on the embedded features of the updated item nodes in each layer of the multi-layer graph convolutional neural network layer.

[0050] In another possible implementation, the user event graph is used as the input of the graph convolutional neural network, and a specific implementation of outputting the context-refined embedding feature is as follows: a graph convolution operation is performed on the user features of the user node connected to the intention node in the previous layer and the item features of the item node to obtain the intention features of the intention node in the current layer; based on the intention features of the intention node in the current layer and the attention weights of the context nodes connected to the intention node, each sub-context feature of the context node in the current layer is determined; an aggregation operation is performed on each sub-context feature of the context node in the current layer to obtain the embedding features of the context node in the current layer; and the context-refined embedding feature is obtained based on the embedding features of the context nodes updated by each graph convolutional neural network layer in the multi-layer graph convolutional neural network layer.

[0051] In another possible implementation, the recommendation module is specifically configured to: use the context information and the refined embedding feature set as inputs to the recommendation network, output the target user's interaction probability with each item in the target domain; and determine the recommendation result for the target user in the target domain based on the target user's interaction probability with each item in the target domain.

[0052] Optionally, the recommendation network includes a decoder that uses refined embeddings enhanced with cross-domain information to predict the user's preference for items in context. FM can be used as the decoder implementation, with FM serving as the core component of the recommendation network. FM is a linear model with better interpretability and lower complexity than MLP.

[0053] In another possible implementation, the recommendation device provided by the present application also includes: a recommendation explanation module, which is used to detect that the target user performs a target operation on the recommendation result display page, and the target operation is used to request a result analysis page for displaying the recommendation results; in response to the target operation, the result analysis page is displayed, and the result analysis page includes an explanation diagram of the recommendation results, and the explanation diagram at least indicates the process of migrating the preference information of the paired user in the source domain to the target user in the target domain.

[0054] In another possible implementation, the preference information of the paired users in the source domain includes the interaction preference distribution information and context preference distribution information of the paired users in the source domain, the explanatory diagram indicates the interaction preference distribution information and context preference distribution information of the paired users in the source domain, the interaction preference distribution information and context preference distribution information of the target user in the target domain after aggregating the interaction preference distribution information and context preference distribution information of the paired users in the source domain, and the recommendation results determined based on the interaction preference distribution information and context preference distribution information of the target user in the target domain; wherein, the interaction preference distribution information of the paired users in the source domain indicates the preference degree of the paired users for each item feature in the source domain, the context preference distribution information of the paired users in the source domain indicates the preference degree of the paired users for each context feature in the source domain, the interaction preference distribution information of the target user in the target domain indicates the preference degree of the target user for each item feature in the target domain, and the context preference distribution information of the target user in the target domain indicates the preference degree of the target user for each context feature in the target domain.

[0055] In another possible implementation, the historical interaction information in the target domain is located in the terminal device; the feature extraction step of obtaining the first user refined embedding features, the item refined embedding features, and the context refined embedding features based on the historical interaction information of the target user in the target domain is performed in the terminal device; and the step of determining the cross-domain enhanced user refined embedding features based on the first user refined embedding features and the preference features of the paired user in the source domain is performed in the cloud.

[0056] In another possible implementation, the refined embedded feature database is updated according to a preset period. For example, the user event graph is updated every 48 hours using the historical interaction data generated within 48 hours, and then the refined embedded feature set corresponding to each user is updated to ensure the quality and freshness of the refined embedded feature set of each user in the refined embedded feature library.

[0057] In a fourth aspect, the present application provides a recommendation device, including an acquisition module, a determination module, a pairing module and a recommendation module, wherein the acquisition module is used to obtain recommendation request information, and the recommendation request information carries information of the target user; the determination module is used to determine the user refined embedding features and user intention embedding features of the target user in the source domain, and the user refined embedding features and user intention embedding features of the target user in the source domain are obtained based on feature extraction of the target user's historical interaction information in the source domain; the pairing module is used to determine the paired user with the target user in the target domain based on the user refined embedding features and user intention embedding features corresponding to the target user in the source domain, and the user refined embedding features and user intention embedding features of each user in the target domain, and the user refined embedding features and user intention embedding features of each user in the target domain are obtained based on feature extraction of the historical interaction information of each user in the target domain, the source domain and the target domain are different recommendation domains, and the number of interaction events of the target user in the target domain is less than the number of interaction events of the target user in the source domain; the recommendation module is used to determine the recommendation result of the target user in the target domain based on the historical interaction information of the paired user in the target domain.

[0058] In another possible implementation, the pairing module is specifically used to: determine the position distribution of each user in the target domain based on the user refined embedding features and user intention embedding features corresponding to each user in the target domain; determine the position distribution of each user in the source domain based on the user refined embedding features and user intention embedding features corresponding to each user in the source domain; determine the transformation matrix that transforms the position distribution of each user in the source domain into the position distribution of each user in the target domain with the goal of minimizing the transmission cost; determine the paired user to be paired with the target user from each user in the target domain based on the user refined embedding features, user intention embedding features, and the transformation matrix corresponding to the target user in the source domain.

[0059] In another possible implementation, the user intention embedding feature corresponding to the target user in the source domain is determined based on the multiple user event intention embedding features corresponding to the target user in the source domain, and the multiple user event intention embedding features corresponding to the target user in the source domain are obtained based on the intention feature extraction of the multiple historical interaction events corresponding to the target user in the source domain; the user intention embedding feature corresponding to each user in the target domain is determined based on the multiple user event intention embedding features corresponding to each user in the target domain, and the multiple user event intention embedding features corresponding to each user in the target domain are obtained based on the intention feature extraction of the multiple historical interaction events corresponding to each user in the target domain.

[0060] In another possible implementation, the determination module is specifically used to: construct a user event graph based on the historical interaction information of the target user in the source domain, wherein the user event graph includes user nodes, item nodes, context nodes and intention nodes, and the intention nodes are connected to the user nodes, item nodes and context nodes respectively. The user node is the node where the user feature corresponding to the target user is located, the item node is the node where the item feature corresponding to the item having an interactive relationship with the target user is located, the context node is the node where the context feature corresponding to the context of the target user's interaction with the item is located, and the intention node is the node where the intention feature corresponding to the intention of the target user's interaction with the item is located; the user event graph is used as the input of the graph convolutional neural network, and after the graph convolution operation of each node in the user event graph is performed by multiple layers of graph convolutional neural network layers, the refined user embedding features of the target user in the source domain are output.

[0061] In another possible implementation, the user event graph is used as the input of the graph convolutional neural network. After the graph convolution operation of each node in the user event graph is performed by multiple layers of graph convolutional neural network layers, the refined user embedding features of the target user in the source domain are output. A specific implementation is as follows: the intent features of each intent node connected to the user node in the previous layer and the item features of the item node connected to each intent node are performed to obtain each sub-user feature of the user node in the current layer; each sub-user feature of the user node is aggregated to obtain the embedding features of the user node in the current layer; based on the embedded features of the user nodes updated by each layer of the graph convolutional neural network layer in the multi-layer graph convolutional neural network layer, the refined user embedding features of the target user in the source domain are obtained.

[0062] In another possible implementation, the recommendation device provided by the present application also includes a recommendation explanation module, which is used to: detect that the target user performs a target operation on the recommendation result display page, and the target operation is used to request a result analysis page for displaying the recommendation result; in response to the target operation, the result analysis page is displayed, and the result analysis page includes an explanation diagram of the recommendation result, and the explanation diagram at least indicates the pairing process between the target user in the source domain and the paired user in the target domain.

[0063] In another possible implementation, the explanatory diagram indicates the target user's interaction preference distribution information and contextual preference distribution information in the source domain, as well as the paired user's interaction preference distribution information and contextual preference distribution information in the target domain, and the recommendation results determined based on the paired user's historical interaction information in the target domain; wherein, the target user's interaction preference distribution information in the source domain indicates the target user's preference degree for each item feature in the source domain, the target user's contextual preference distribution information in the source domain indicates the target user's preference degree for each contextual feature in the source domain, the paired user's interaction preference distribution information in the target domain indicates the paired user's preference degree for each item feature in the target domain, and the paired user's contextual preference distribution information in the target domain indicates the paired user's preference degree for each contextual feature in the target domain.

[0064] In a fifth aspect, an embodiment of the present application provides a computing device comprising a memory and a processor, wherein the memory stores instructions, and when the instructions are executed by the processor, the method described in the first aspect and / or the second aspect is implemented.

[0065] In a sixth aspect, an embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method described in the first aspect and / or the second aspect is implemented.

[0066] In a seventh aspect, an embodiment of the present application further provides a computer program or a computer program product, which includes instructions that, when executed, cause a computer to execute the method described in the first aspect and / or the second aspect.

[0067] In an eighth aspect, an embodiment of the present application further provides a chip comprising at least one processor and a communication interface, wherein the processor is used to execute the method described in the first aspect and / or the second aspect. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 A system architecture diagram of a recommendation system that applies the recommendation method provided in an embodiment of the present application is shown;

[0069] Figure 2 A schematic diagram of a method flow chart of a recommended method provided in an embodiment of the present application;

[0070] Figure 3 The following is a schematic diagram of the architecture of the C-UEG-EL framework provided in an embodiment of the present application;

[0071] Figure 4 A schematic diagram of a process for constructing a user event graph is shown;

[0072] Figure 5Shows a schematic diagram of information propagation of user features in the recommendation field;

[0073] Figure 6 A schematic diagram showing the information propagation of item features in the recommendation field;

[0074] Figure 7 A schematic diagram of information propagation of contextual features in the recommendation field is shown;

[0075] Figure 8 A cross-domain link building diagram is shown;

[0076] Figure 9 A schematic diagram of cross-domain information dissemination is shown;

[0077] Figure 10 A schematic diagram showing an example of explaining a recommendation result is shown;

[0078] Figure 11 A schematic diagram of a method flow chart of another recommended method provided in an embodiment of the present application;

[0079] Figure 12 A schematic diagram showing an example of explaining a recommendation result is shown;

[0080] Figure 13 A schematic diagram of the structure of a recommended device provided in an embodiment of the present application;

[0081] Figure 14 A schematic diagram of the structure of a recommended device provided in an embodiment of the present application;

[0082] Figure 15 A schematic diagram of the structure of a computing device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0083] The term "and / or" as used herein describes an association relationship between related objects. Features can exist in three relationships. For example, A and / or B can be characterized by: A alone, A and B together, or B alone. The symbol " / " in this document indicates an OR relationship between related objects. For example, A / B indicates either A or B.

[0084] The terms "first" and "second" in this specification and claims are used to distinguish different objects rather than to describe a specific order of objects. For example, the terms "first user refined embedded feature" and "second user refined embedded feature" are used to distinguish different user refined embedded features rather than to describe a specific order of user refined embedded features.

[0085] In the embodiments of this application, words such as "exemplary" or "for example" are used to describe features as examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in the embodiments of this application should not be interpreted as being more preferred or advantageous over other embodiments or designs. Rather, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0086] In the description of the embodiments of the present application, unless otherwise specified, "multiple" means two or more than two, for example, multiple nodes refer to two or more nodes, etc.; multiple interaction events refer to two or more interaction events, etc.

[0087] The user refined embedding features / item refined embedding features / context refined embedding features in the embodiments of the present application are used to indicate a user event graph constructed based on historical interaction information (including nodes representing user nodes, item nodes, context nodes and intention nodes, and the intention nodes are connected to the user nodes, item nodes and context nodes respectively). After the user event graph is input into the graph convolutional neural network, the graph convolutional neural network processes each node in the user event graph and outputs a user refined embedding vector / item refined embedding vector / context refined embedding vector. The user refined embedding vector contains high-order semantic features of the user, the item refined embedding vector contains high-order semantic features of the item, and the context refined embedding vector contains high-order semantic features of the context.

[0088] It can be understood that when the user event graph of users in the target domain is input into the graph convolutional neural network, what is obtained are the user refined embedding features, item refined embedding features and context refined embedding features corresponding to the users in the target domain; when the user event graph of users in the source domain is input into the graph convolutional neural network, what is obtained are the user refined embedding features, item refined embedding features and context refined embedding features corresponding to the users in the source domain.

[0089] To mine and utilize potentially beneficial collaborative information between different recommendation scenarios, cross-domain recommendation typically treats any two recommendation scenarios as a source domain (with the corresponding user interactions serving as auxiliary data) and a target domain (with the corresponding user interactions serving as target data). Cross-domain knowledge transfer and collaboration are then achieved by building effective links between the two domains. Based on the similarity between two recommendation scenarios at different levels, they can be categorized into the following types:

[0090] 1. Cross-domain based on content level: In two recommendation scenarios, there is common content or features (such as keywords, tags) between users or items.

[0091] 2. Cross-domain based on user hierarchy: In two recommendation scenarios, there are common users, but the item hierarchies are different.

[0092] 3. Cross-domain based on item level: There are common items in two recommendation scenarios, but different user groups.

[0093] In related technologies, most cross-domain recommendation work relies on the proportion of overlapping information related to users or items, which is prone to performance bottlenecks when the proportion is insufficient; in addition, ignoring the contextual information during interaction may introduce noise in user or item matching when building links.

[0094] For example, related technology 1 proposes a cross-domain recommendation scheme (which can be called the ER4CDR scheme). This scheme is a typical example of cross-domain recommendation technology. They regard the interaction between the two domains as two user-item bipartite graphs, and then associate the nodes of the two graphs through the same users or items in the two domains to form a fusion graph. For the fusion graph, the ER4CDR scheme designs a heterogeneous network embedding method based on a hierarchical attention mechanism. It learns to select more representative neighbors or feature edges and assigns them higher weights. The first layer embeds the neighbor nodes and edges of the aggregated target node, and the second layer aggregates them together; for the two single-domain graphs, conventional graph neural networks are used for vertex representation learning. Overall, the ER4CDR scheme simultaneously applies graph representation learning of single-domain graphs and graph representation learning of fusion graphs to improve recommendation performance.

[0095] Related technology 2 proposes another cross-domain recommendation scheme (which can be called GWCDR). The GWCDR scheme is different from the ER4CDR scheme in that it does not use the same information in the two domains as the associated objects, but instead calculates the correlation between users and items in content or features in the embedding space and minimizes the difference between the two domains to achieve the link between the two domains, and then designs subsequent cross-domain recommendation schemes.

[0096] As can be seen, related cross-domain recommendation solutions rely on the same users or items in both domains to construct cross-domain links. However, this approach is limited by insufficient overlapping information. Furthermore, most work only utilizes user and item information, failing to fully exploit additional information, such as context. Ignoring contextual information during interactions can introduce noise into user or item matching when constructing links.

[0097] In order to get rid of the dependence on overlapping users or overlapping items and improve the accuracy of cross-domain links, the embodiment of the present application proposes a recommendation method and device that effectively matches user intentions between the source domain and the target domain to build beneficial cross-domain links, thereby screening reasonable source domain neighbors for users in the target domain for recognition transfer. The recommendation method and device proposed in the embodiment of the present application considers the interaction preferences and contextual preferences of the user group in the current domain (target domain), and further enhances the preference modeling and prediction of the user group by migrating user knowledge with similar interaction preferences or contextual preferences from the source domain through cross-domain links.

[0098] The specific implementation scheme of the recommended method and device provided in the embodiments of the present application is described in detail below.

[0099] Figure 1 FIG1 shows a system architecture diagram of a recommendation system that applies the recommendation method provided in the embodiment of the present application. Figure 1 As shown in Figure 3, the recommendation system mainly includes offline processing stage and online processing stage.

[0100] In the offline processing stage, the recommendation model is trained through historical interaction data from different recommendation fields (the recommendation model includes an embedding feature extraction network and a recommendation network), and the trained recommendation model (embedding feature extraction network) is used to extract features from the historical interaction data of each user to obtain the cross-domain enhanced refined embedding features corresponding to each user in different recommendation fields. The cross-domain enhanced refined embedding features corresponding to each user are saved in the refined embedding feature database, and the trained recommendation network is put online for processing and used as an online model to predict user preferences and recommend items that meet user preferences.

[0101] In the online recommendation processing phase, in response to a recommendation request, the refined cross-domain embedded features corresponding to the user in the target domain are searched from the refined embedded feature database based on the user ID. These refined cross-domain embedded features and the contextual information carried in the recommendation request are input into the online model, which then outputs the interaction probability between the user and each item. Items are then recommended to the user based on these interaction probabilities. For example, the item with the highest predicted interaction probability may be recommended, or multiple items may be recommended to the user, sorted in descending order of interaction probability.

[0102] Take the recommended domain A (hereinafter referred to as domain A for the convenience of description) and the recommended domain B (hereinafter referred to as domain B for the convenience of description) as examples.

[0103] During the offline processing phase, behavior tables are extracted from the A-domain log and the B-domain log. These logs include actions such as impressions and clicks, as well as the contextual information when these actions occurred. The behavior tables include user-item interaction information (e.g., user A purchased item a, and user B purchased item b), as well as the contextual information when these interactions occurred (e.g., user A purchased item a on Valentine's Day). Labels are generated based on the behavior tables, indicating the user's preference for the item. For example, if user A purchased item a, the label for this interaction event is item a, indicating that user A's interaction preference is item a, meaning that user A is likely to interact with item a.

[0104] Then, basic data tables are extracted from the user and item libraries of domains A and B, respectively. These tables contain features of each user and item, such as age, occupation, and gender, and features of items such as price, texture, and origin. Feature mining is performed on the basic data tables and behavior tables to generate an offline feature library. The offline feature library contains feature data of interaction events across different recommendation domains, such as user features, item features, interaction features, and context features. A user event graph is constructed based on the offline feature library. The user event graph consists of user nodes, item nodes, context nodes, and intent nodes. Intent nodes are connected to user nodes, item nodes, and context nodes, respectively. User nodes are nodes containing user features corresponding to users, item nodes are nodes containing item features corresponding to items with which the target user interacts, context nodes are nodes containing context features corresponding to the context of the user's interaction with the item, and intent nodes are nodes containing intent features corresponding to the target user's intention in interacting with the item. Labels and the corresponding user event graph (constructed from historical interaction events corresponding to the labels) constitute training samples, which are then used to train the recommendation model. The embedded feature extraction network in the trained recommendation model can be used to extract refined, cross-domain-enhanced embedded features for each user, resulting in a refined embedded feature database. These refined, cross-domain-enhanced embedded features include preference knowledge from other domains, enabling more accurate recommendations for users in the target domain.

[0105] During the online processing phase, when a user enters a recommendation domain system, for example, domain A (e.g., a hotel reservation interface), a recommendation request is triggered. Optionally, the recommendation request includes user information (e.g., user ID), current context, and information about items to be recommended, such as an item list. In this case, domain A is the target domain, and domain B is the source domain (e.g., a restaurant domain). The recommendation system searches the refined embedding feature database for the user ID in domain A, along with the cross-domain enhanced refined embedding features. The cross-domain enhanced refined embedding features and related information (e.g., current context and recommended item information) are then fed into the recommendation network. The system then predicts the user's interaction probability with items in domain A and recommends items based on this probability. For example, the system can sort items in descending order based on their predicted interaction probability. This means the system can sequentially display items in different locations as recommendations for the user. Users browse items in different locations and interact with them (with higher-ranked items being more likely to interact), such as browsing, selecting, adding to favorites, and downloading. At the same time, the actual interaction behaviors of users will be stored in the log as the training data for the next round. By training the model in the offline processing stage, the model parameters are updated and / or the refined embedding feature database is updated to improve the prediction effect of the model.

[0106] When the user needs it, for example, when the user clicks the "?" icon on the recommendation interface, it means that the user wants to know why the recommendation is made. The recommendation system will display the result analysis interface to the user based on the model inference data. The result analysis interface displays an explanation graph, through which the user can clearly understand the origin of the recommendation results, increase the explainability of the recommendation, and improve the user experience.

[0107] Figure 2 A flowchart of a method for a recommendation method provided in an embodiment of the present application. The method can be executed by any device, equipment, platform or device cluster with computing capabilities. The embodiment of the present application does not specifically limit the specific computing device for executing the method, and a suitable computing device can be selected for execution as needed. For example, it can be executed on a terminal device (such as a smart phone, personal computer, and tablet computer, etc.), or it can be executed on a terminal device and a cloud device (such as a cloud-side server), that is, it is implemented using an end-cloud collaborative architecture, or it can be completed on a cloud device. For the sake of convenience in the following description, the form of the execution subject will not be distinguished, and they are all described as recommendation systems. Figure 2 As shown, the recommendation method provided in the embodiment of the present application includes steps S201 to S203.

[0108] In step S201 , recommendation request information is obtained, where the recommendation request information carries information of a target user and context information of the target user in a target domain.

[0109] When a user enters a recommendation system, such as a hotel reservation interface, a recommendation request will be triggered. The generated recommendation request will carry the user's (i.e., target user) information and context information. The recommendation request is used to request a recommendation for the user, such as recommending a hotel that meets the user's preferences, allowing the user to quickly find the hotel they like, freeing the user from the tedious task of retrieving massive amounts of hotel information and improving the user experience.

[0110] In another example, the recommendation request also includes item information, such as an item list, which includes a certain number of items that have been preliminarily screened. For example, in the field of hotel recommendations, the item list includes 500 preliminarily screened hotels that the user may choose.

[0111] For example, when a user opens an application with a hotel reservation function on a smart terminal (such as a smartphone), a recommendation request will be triggered. For example, the recommendation system collects the user's user information (such as user ID), context information, and item information, and then carries the user information, context information, and item information into the recommendation request information, and sends the recommendation request to the recommendation server, requesting the recommendation server to recommend hotels that the user may like to the user based on the user's preferences.

[0112] It can be understood that contextual information refers to external information about a user interacting with an item or before the interaction, which influences the user's interaction decision. For example, in restaurant recommendations, the current time, location, and historical interaction information may all influence the user's decision to choose a restaurant. For example, if the current time is morning and the location is Hangzhou, and the user has previously eaten steamed buns for breakfast, a restaurant specializing in Hangzhou soup dumplings may be recommended to the user. Therefore, in the field of restaurant recommendations, contextual information may include one or more of the current time, location, and historical interaction information.

[0113] The types of items are different in different recommendation fields. For example, in the hotel recommendation field, the items are hotels; in the restaurant recommendation field, the items are restaurants; and in the video recommendation field, the items are videos. This application does not specifically limit the categories of items.

[0114] In step S202, a search is performed in the refined embedding feature database based on the target user's information to obtain a refined embedding feature set corresponding to the target user.

[0115] After receiving a recommendation request, the recommendation system responds by searching the refined embedding feature database, processed offline, to obtain the target user's corresponding refined embedding feature set. As described above, the refined embedding feature database includes the cross-domain enhanced refined embedding features corresponding to each user. Based on the user ID in the recommendation request, the refined embedding feature database can be used to search for the target user's cross-domain enhanced refined embedding features.

[0116] The refined embedding features after cross-domain enhancement learn the preference knowledge of two domains. Compared with the refined embedding features of a single domain that only learns the preference knowledge of the domain itself, it is richer and more complete, and is more conducive to predicting recommendation results that meet user preferences and more diverse recommendation results, thus realizing cross-domain recommendations.

[0117] The following describes in detail how the recommendation method provided in the embodiment of the present application obtains a refined embedded feature database during the offline processing stage, i.e., cross-domain enhanced refined embedded features corresponding to each user in each recommendation field.

[0118] Cross-domain recommendation differs from general recommendation systems in that it requires attention to the transfer and migration of information between different domains. When performing cross-domain recommendation, it is necessary to consider two domains: the source domain and the target domain.

[0119] The source domain can be understood as the domain from which knowledge transfer is performed, including rich interactive information; the target domain is the target recommendation domain. For example, if a user opens an app for restaurant reservations, the restaurant recommendation domain is the target domain. The hotel recommendation domain, which has rich interactive information and is similar to restaurant recommendations, is the source domain. By transferring preference knowledge from the hotel recommendation domain to the restaurant recommendation domain, the user's preference knowledge in the restaurant domain can be supplemented and enhanced, which helps to provide users with more accurate recommendations of restaurants that suit their preferences, breaking the inherent recommendation routine and making recommendations more diverse.

[0120] The purpose of cross-domain recommendation is to learn the features of the target domain based on the existing features of the source domain and the target domain, so as to recommend information to the target user in the target domain.

[0121] The key to cross-domain recommendation is how to accurately build effective links between two domains to achieve cross-domain knowledge transfer and collaboration. In other words, the key to cross-domain recommendation is how to identify beneficial source domain neighbors for users in the target domain to perform knowledge transfer. To this end, the present embodiment provides a cross-domain user-event collaborative graph embedding learning (C-UEG-EL) framework.

[0122] Figure 3 The schematic diagram of the C-UEG-EL framework provided in the embodiment of the present application is shown. Figure 3 As shown, the C-UEG-EL framework provided by the embodiment of the present application consists of two modules, including a cross-domain user-event collaborative graph convolution module and a recommendation module. The cross-domain user-event collaborative graph convolution module is used to use the useful knowledge of the source domain to learn refined embeddings (refined vectors) of users, items and contexts, and the recommendation module uses refined embeddings to improve the performance of downstream recommendation models.

[0123] Depend on Figure 3 It can be seen that the C-UEG-EL provided in the embodiment of the present application constructs a user event graph for users in the source domain and a user event graph for users in the target domain respectively. The user event graph includes user nodes, item nodes, intent nodes and context nodes, and the intent nodes are connected to the user nodes, item nodes and context nodes respectively. The intention features on the intent nodes can capture the user's interaction preferences and context preferences. The intent nodes can be used to achieve cross-domain user pairing, pairing users with similar user intentions in the two domains, and construct cross-domain links between paired users in the two domains based on the pairing, providing an accurate cross-domain information dissemination path for subsequent knowledge transfer.

[0124] Before performing cross-domain intent matching, a user event graph is first constructed for users in the source domain and users in the target domain. The user event graph is then input into a graph convolutional neural network to obtain single-domain user refined embedding features, item refined embedding features, and context refined embedding features.

[0125] Figure 4 A schematic diagram of the process of constructing a user event graph is shown in FIG. Figure 4 As shown in the figure, we first construct a personal graph of the user in the recommendation domain (source domain or target domain) based on historical interaction event information, then obtain the intent node attention graph based on the personal graph, and then derive the user event graph based on the intent node attention graph.

[0126] It can be understood that the user node is the node where the user features corresponding to the target user are located, the item node is the node where the item features corresponding to the item that has an interactive relationship with the target user are located, the context node is the node where the context features corresponding to the context of the target user's interaction with the item are located, and the intention node is the node where the intention features corresponding to the intention of the target user's interaction with the item are located.

[0127] In a recommendation domain, there are M user sets U={u1,u2,…,u M}、N items V={v1,v2,…,v N} and R contexts C={c1,c2,…,c R In addition, there are J features for the user A={A 1 ,A 2 ,…,A J} and the item has K features B={B 1 ,B 2 ,…,B K}.

[0128] like Figure 4 As shown in the personal graph, a user may have multiple item interaction histories, and each interaction is called an event. In the event set S, the i-th event s i Contains user information, item information and context information. The i-th event can be expressed as s i The corresponding embedding representation is written as The user embedding and its features are averaged to obtain Items are At this time i The embedding can be simplified as

[0129] This can be achieved by setting the encoder (Encoder) For example, user features, item features, and context features are respectively used as the input of the encoder, and the encoder encodes them to obtain For example, the encoder multi-hot encodes the input user features, item features, and context features, and then performs an average pooling operation on the encoded features to obtain the initialized embedding representation vectors.

[0130] In order to better identify user interests, the embodiment of the present application adopts the following calculation method to obtain the embedded representation of user event intention in user event i: The implementation formula is as follows:

[0131]

[0132]

[0133] in, W1,W2∈R L×L , is a trainable parameter, d is the dimension of the embedding, is the attention weight of the i-th context.

[0134] It should be noted that user features refer to user attribute information that can distinguish different users, such as user ID, gender, age, region, and occupation. Item features refer to item attribute information that can distinguish different items. For example, when the recommendation area is restaurants, restaurant features (i.e., item features) may include high-end, robot service, free parking, Chinese food, and popular.

[0135] Since the constructed user event graph has newly added intent nodes, the traditional graph convolution method is no longer applicable. This application provides a graph convolution method that uses intent nodes as hubs to associate user nodes, item nodes, and context nodes to explore the relationship between users, items, and context.

[0136] The following introduces the graph convolution method of each feature respectively.

[0137] 1. User Characteristics

[0138] For user characteristics, such as Figure 5 As shown in Figure 2, the intent node can be used as the center to propagate information about item features and context features to the user node. The specific implementation formula is as follows:

[0139]

[0140] in is the embedding representation of the user associated with the i-th event passed to the h-th layer of the graph convolution, and are the item and intent node embeddings at layer h-1, respectively, where For each user u, aggregate the information of all events associated with it to obtain the embedding of layer h. The specific formula is as follows:

[0141]

[0142] The above formula is to capture what type of items user u will interact with under what context features.

[0143] Finally, after each graph convolution layer is processed as described above, the embedded features of the user nodes processed by each graph convolution layer are obtained. The embedded features of the user nodes are refined to obtain the refined embedded features of the user in the single recommendation domain. For example, the embedded features of the user nodes processed by each graph convolution layer can be averaged to obtain the refined embedded features of the user in the single recommendation domain. That is, by averaging the embeddings of the user nodes in each layer, a refined user embedding (i.e., refined user embedding features) can be obtained. The specific formula is as follows:

[0144]

[0145] in, is the refined embedding feature of the user in the single recommendation domain graph convolution output, H is the number of graph convolution layers, is the embedded feature of the user node obtained after processing by the h-th graph convolution layer.

[0146] It should be noted that in addition to fine-tuning using averaging, other methods (such as variance) can also be used for fine-tuning, which are not limited here. Furthermore, after completing the graph convolution, in addition to fine-tuning the embedding features of the user nodes, the embedding features of the user nodes obtained after the last graph convolution layer can also be directly used as the final required refined embedding features of the user nodes. The specific method depends on the actual situation and is not limited here.

[0147] 2. Item Characteristics

[0148] For item features, such as Figure 6 As shown in Figure 2, the intent node can be used as the center to propagate information about user features and context features to item nodes. The specific implementation formula is as follows:

[0149]

[0150] in is the user embedding at layer h-1, The embedding of item v at layer h can be obtained by aggregating the information of all events associated with it. The specific formula is as follows:

[0151]

[0152] This formula can capture what type of users interact with item v under what contextual features.

[0153] Finally, after each graph convolution layer is processed as described above, the embedding features of the item nodes processed by each graph convolution layer are obtained. The embedding features of the item nodes are fine-tuned to obtain the refined embedding features of the items in the single recommendation domain. For example, the embedding features of the item nodes processed by each graph convolution layer can be averaged to obtain the refined embedding features of the items in the single recommendation domain. In other words, by averaging the embeddings of the item nodes in each layer, the refined item embedding (i.e., the refined item embedding feature) can be obtained. The specific formula is as follows:

[0154]

[0155] It should be noted that in addition to fine-tuning using averaging, other methods (such as variance) can also be used for fine-tuning, which are not limited here. Furthermore, after completing the graph convolution, in addition to fine-tuning the embedding features of the user nodes, the embedding features of the user nodes obtained after the last graph convolution layer can also be directly used as the final required refined embedding features of the user nodes. The specific method depends on the actual situation and is not limited here.

[0156] 3. Contextual Features

[0157] In order to propagate user and item information to context features, such as Figure 7 As shown, the information is first propagated to the intention node. The specific implementation formula is as follows:

[0158]

[0159] Then, the context feature receives different information from the intent node according to the attention distribution. The specific formula is as follows:

[0160]

[0161] It represents the attention weight between the intent node and the i-th context node. The attention weight is determined based on the i-th user feature and the i-th context feature. The specific implementation formula is as follows:

[0162]

[0163] For the i-th context node, there may be multiple intent nodes connected to it. At this time, the information transmitted by the i-th context node and the multiple intent nodes connected to it can be used to embed the context features of the h-th layer. The specific implementation formula is as follows:

[0164]

[0165] Finally, after each graph convolution layer is processed as described above, the embedded features of the context nodes processed by each graph convolution layer are obtained. The embedded features of the context nodes are fine-tuned to obtain the context-refined embedded features of the context in the single recommendation domain. For example, the embedded features of the context nodes processed by each graph convolution layer can be averaged to obtain the context-refined embedded features of the context in the single recommendation domain. In other words, by averaging the embeddings of the context nodes of each layer, a refined context embedding (i.e., user refined embedding features) can be obtained. The specific formula is as follows:

[0166]

[0167] It should be noted that Finally, we need to feed the information back to the intent node to obtain the embedding of the h+1th layer intent node.

[0168] After performing the graph convolution operation, a set of refined embedding features in a single recommendation domain can be obtained, namely, user refined embedding features Item refined embedding features and contextually refined embedding features

[0169] User intent is then used to pair users between different recommendation domains, that is, user intent is used to identify beneficial neighbors from the source domain for the target user in the target domain, that is, user intent is used to pair the target user with users in the source domain, so as to facilitate the subsequent transfer of knowledge from users in the source domain to the target domain.

[0170] For example, in order to better capture the interests of users in each domain, the intent node embeddings in the event set of each user in the domain are averaged to obtain the corresponding user average intent embedding through the following formula, and the embedding is used in subsequent cross-domain intent pairing.

[0171]

[0172] We expect that the information dissemination of users in the user event graph can not only contain knowledge in their own fields, but also contain useful knowledge in other fields. To achieve this goal, a key step is to establish reliable links between two users in different fields, and these links are more likely to carry knowledge that is beneficial to user preference prediction. In this embodiment of the application, we model the link construction between cross-domain users as an optimal transport (OT) problem. First, we regard the two domains as two discrete distributions:

[0173]

[0174]

[0175] in and represents the position of each user in the distribution. In this embodiment, we simply use a linear calculation with two control weights, namely and They can also be viewed as virtual enhanced user nodes further derived from user nodes and context nodes on the user event graph.

[0176] In the OT problem, the sum of μ and υ is constrained to be 1, that is, ∑ i μ i=∑ j υ j =1, for simplicity, we assume that each user has an average quality, i.e. and Next, based on the defined cost function The optimal solution to the OT problem aims to find an alignment matrix T (also called a transformation matrix T) that minimizes the total cost of transforming from μ to υ, where the constraints (∑ i T ij =υ j and ∑ j T ij =μ i )

[0177]

[0178] in is the set of all possible alignment matrices. In practice, the above equation can be efficiently solved using representative methods for OT problems, such as the Sinkhorn iterative algorithm. Since the Sinkhorn iterative algorithm is a mature algorithm, its specific algorithm details are omitted here for brevity. Formally, this process can be described as follows:

[0179]

[0180] Based on the optimized alignment matrix T, we can establish links between the user nodes of the two user event graphs to facilitate subsequent cross-domain information transfer. It should be noted that if a user appears in two domains, the two nodes representing the user in the user event graph will naturally be linked together. For ease of understanding, the cross-domain link construction diagram is as follows Figure 8 As shown in the figure, the dotted circle represents the virtual enhanced user node.

[0181] Through the C-UEG-EL provided by the embodiment of the present application, cross-domain links are explored and utilized during information dissemination. The embodiment of the present application uses the cross-domain links formed by pairing intentions as a bridge, and passes the preference knowledge (including contextual preferences and interaction preferences) of the source domain neighbors with cross-domain links to the corresponding users in the target domain during the graph convolution process, and enhances the preference modeling of the user in combination with the preference knowledge in its own domain. Thanks to the accurate pairing relationship and efficient convolution method, the model can capture user interests more effectively.

[0182] After a single domain undergoes intra-domain information propagation, it obtains refined embeddings of users, items, and contexts. Then calculate the user's average intention using the method introduced above And further calculate to get the alignment matrix T. Figure 9As shown, at this time we can identify a set of beneficial neighbors in the source domain for each target domain user and transfer their knowledge because they have some similar interaction preferences or context preferences. The implementation formula is as follows:

[0183]

[0184] in represents the knowledge embedding transferred from the source domain for user j based on cross-domain links. We combine this knowledge embedding with the refined embedding of user j and leverage the acquired cross-domain knowledge to enhance the preference learning of this user. The implementation formula is as follows:

[0185]

[0186] Finally, after performing cross-domain user event collaborative graph convolution, we can obtain a set of corresponding refined embeddings, namely, cross-domain enhanced refined embedding features. The implementation formula is as follows:

[0187]

[0188] Through the method described in C-UEG-EL provided in the embodiment of the present application, the historical interaction information of users in each domain is processed to obtain cross-domain enhanced refined embedding features of users in each domain.

[0189] In step S203 , based on the context information and the refined embedded feature set, the recommendation result for the target user in the target domain is determined.

[0190] The context information carried in the recommendation request and the refined embedding feature set are As input to downstream recommendation task models, such as the recommendation network in the online processing stage mentioned above, the output is the probability of the target user interacting with each item in the target domain. Items are then recommended to the user based on this interaction probability. For example, the item with the highest predicted interaction probability may be recommended, or multiple items may be recommended to the user in descending order of interaction probability.

[0191] Optionally, the recommendation network includes a decoder that uses refined embeddings enhanced with cross-domain information to predict the user's preference for items in context. FM can be used as the decoder implementation, with FM serving as the core component of the recommendation network. FM is a linear model with better interpretability and lower complexity than MLP.

[0192] Of course, in some other examples, the recommendation network can also be an MLP.

[0193] In another example, the recommendation method provided by the present application also includes: detecting that the target user performs a target operation on the recommendation result display page, the target operation is used to request a result analysis page for displaying the recommendation result; in response to the target operation, the result analysis page is displayed, the result analysis page includes an explanatory diagram of the recommendation result, and the explanatory diagram at least indicates the process of migrating the preference knowledge of the paired users in the source domain to the target user in the target domain.

[0194] Thanks to the method of pairing users in the source domain and the target domain through user intent embedding features, and the method of processing user event graphs using graph convolution, user intentions and contextual preferences are traceable, so a user interface can be implemented that demonstrates the cross-domain information transfer process and explains the recommendation results.

[0195] Optionally, the explanatory diagram indicates the interaction preference distribution and context preference distribution of the paired users in the source domain, the interaction preference distribution and context preference distribution of the target user in the target domain after aggregating the interaction preference distribution and context preference distribution of the paired users in the source domain, and the recommendation results determined based on the interaction preference distribution and context preference distribution of the target user in the target domain; wherein, the interaction preference distribution of the paired users in the source domain indicates the preference degree of the paired users for each item feature in the source domain, the context preference distribution indicates the preference degree of the paired users for each context feature in the source domain, the interaction preference distribution of the target user in the target domain indicates the preference degree of the target user for each item feature in the target domain, and the context preference distribution indicates the preference degree of the target user for each context feature in the target domain.

[0196] Through the explanatory diagram, the reasons for the recommendation results are shown to users, such as the transfer process of the user's preference knowledge in the source domain to the target user in the target domain, the user's preference for each context, and the preference for each item feature. In this way, users can have a clear understanding of the reasons for the recommendation results.

[0197] For example, when a user clicks the "?" icon on the recommendation interface, a result analysis interface is displayed to the user in response to this click. This interface displays an explanation diagram, which shows the contextual preference distribution and interaction preference distribution of users in the source domain paired with the target user, an illustration of the transfer and fusion of the source domain user's preference distribution knowledge to the target user in the target domain, and the contextual preference distribution and interaction preference distribution of the target user in the target domain that has fused the source domain user knowledge. Optionally, the contextual preference distribution can be displayed as line weights, and the interaction preference distribution can also be displayed as line weights, which clearly reflects the specific information about items that users attach greater importance to in certain contextual scenarios. In other words, the recommendation results are highly interpretable, improving the user experience.

[0198] Figure 10 A schematic diagram of an example of explaining a recommendation result is shown. Figure 10 As shown, the target user is "Li Si", the target domain is the hotel recommendation domain, Li Si clicks to enter the hotel reservation interface, triggering a recommendation request, and in response to the recommendation request, the recommendation system uses the recommendation method provided in the embodiment of the present application to recommend a hotel to user Li Si. When Li Si wants to know why it is recommended, he clicks the "question mark" icon on the recommendation interface. The recommendation system responds to the operation and switches to the result analysis interface. The result analysis interface displays an explanation diagram, which is drawn based on the original data of the model. The recommendation system will first coordinate the data of the same user in different domains, and modify Li Si's contextual preferences and interaction preferences in the hotel domain through knowledge migration from the restaurant domain. Using the migrated data, it is calculated that in similar context scenarios, hotel (id: 424) may be more in line with user interests, so hotel (id: 424) is recommended to user Li Si first.

[0199] In another example, the recommendation method provided in the embodiment of the present application can be implemented through the architecture of the end-cloud system, where the processing process involving user privacy data is placed on the terminal for processing, and other steps are placed on the cloud for execution. In this way, while ensuring efficiency, user privacy data is not uploaded to the cloud, thereby protecting user privacy.

[0200] For example, the initial encoding of the user's historical interaction data is completed at the terminal. That is, the encoder is deployed on the terminal device, and the graph convolution processing of the encoded data, as well as the intent matching and inference prediction based on contextual information and refined embedded feature sets are completed in the cloud. In this way, the encoded user's historical interaction data will hide private information and protect user privacy.

[0201] For another example, the process of directly processing the user's historical interaction information is executed on the terminal, such as extracting the target user's historical interaction information in the target domain to obtain the target user's refined user embedding features, item refined embedding features and context refined embedding features in the target domain and executing them in the terminal device; other processes that do not directly process the user's historical interaction information are executed in the cloud, such as determining the cross-domain enhanced user refined embedding features based on the target user's refined user embedding features in the target domain and the paired user's preference features in the source domain and executing them in the cloud.

[0202] The embodiment of the present application also provides another recommendation method that can solve the problem of sparse historical interaction data or cold start.

[0203] Figure 11A flowchart of another method for recommending a method according to an embodiment of the present application. The method can be executed by any device, equipment, platform or device cluster with computing capabilities. The embodiment of the present application does not specifically limit the specific computing device for executing the method, and a suitable computing device can be selected for execution as needed. For the sake of convenience in the following description, the form of the executing entity is not distinguished, and all are described as recommendation systems. Figure 11 As shown, the recommendation method provided in the embodiment of the present application includes steps S1101 to S1104.

[0204] In step S1101 , recommendation request information is obtained.

[0205] When a user enters a recommendation system, such as a hotel reservation interface, a recommendation request will be triggered. The generated recommendation request will carry the user's information (such as user ID). The recommendation request is used to request a recommendation for the user, such as recommending a hotel that meets the user's preferences, allowing the user to quickly find the hotel they like, freeing the user from the retrieval of massive hotel information and improving the user experience.

[0206] In another example, the recommendation request also includes contextual information.

[0207] For example, when a user opens an application with a hotel reservation function on a smart terminal (such as a smartphone), a recommendation request will be triggered. For example, the recommendation system collects the user's user information (such as user ID) and context information, and then carries the user information and context information into the recommendation request information, and sends the recommendation request to the recommendation server, requesting the recommendation server to recommend hotels that the user may like to the user based on the user's preferences.

[0208] In step S1102 , the user refined embedding features and user intention embedding features of the target user in the source domain are determined.

[0209] The target user's historical interaction data in the source domain is relatively rich, including rich preference knowledge, such as contextual preference knowledge and interaction preference knowledge. Therefore, feature extraction of the target user's historical interaction data in the source domain can obtain more accurate contextual preference knowledge and interaction preference knowledge.

[0210] Based on the target user's historical interaction data in the target domain, an event-user graph is constructed. This event-user graph is then fed into a graph convolutional network (GCN) to obtain a refined embedding of the target user in the source domain. The graph convolution operation performed by the GCN on the event-user graph is described in the previous section and will not be repeated here for brevity.

[0211] User intention embedding feature is user average intention embedding User average intent embedding The calculation method of is described above and will not be repeated here for the sake of brevity.

[0212] In step S1103, based on the refined user embedding features and user intention embedding features of the target user in the source domain, and the refined user embedding features and user intention embedding features of each user in the target domain, a paired user with the target user in the target domain is determined.

[0213] Adopt the refined embedding features of target users in the source domain and user intent embedding features And the user-refined embedding features of each user in the target domain and user intent embedding features Pairing users in the source domain with users in the target domain is modeled as an O / T problem. The alignment matrix T is then solved and used to establish links between users in the source domain and the target domain, i.e., pairing users in the source domain and the target domain. The specific method for solving the alignment matrix is ​​described above and will not be repeated here for brevity.

[0214] In step S1104 , based on the historical interaction information of the paired users in the target domain, a recommendation result for the target user in the target domain is determined.

[0215] Through the above steps, the user paired with the target user in the target domain is found, and then the historical interaction information of the user in the target domain is used to recommend information to the target user in the target domain.

[0216] For example, a user named Li Si who rarely books hotels or has never booked a hotel in a hotel booking application opens an application with a hotel booking function. In this case, the hotel recommendation domain is the target domain, and the user Li Si is the target user. Since Li Si has less historical interaction information in the hotel domain, it is impossible to extract effective preference knowledge, and therefore it is impossible to make effective recommendations to Li Si in the hotel domain. However, Li Si has more historical interaction information about restaurants in a certain application. In this case, the restaurant domain is the source domain. At this time, the recommendation algorithm provided by this application can be applied. By using the intent matching method, a user named Wang Wu whose preferences are close to Li Si’s can be found in the hotel domain. Then, Wang Wu’s historical interaction information can be used to recommend hotels to Li Si. For example, the hotels that Wang Wu frequently books can be directly recommended to Li Si; for another example, the hotels can be recommended to Li Si in descending order according to the number of times Wang Wu has booked them; for another example, the hotels can be recommended to Li Si in descending order according to the degree of Wang Wu’s favorable comments on the hotels.

[0217] In another example, the recommendation request also carries context information, and determining the recommendation result for the target user in the target domain is also related to the context information.

[0218] For example, after identifying a user paired with the target user in the target domain, a recommendation result for the target user in the target domain is determined based on the context information and the user's historical interaction information in the target domain. For example, the context information can be used to find items that the target user has interacted with in similar contexts from the user's historical interaction information in the target domain, and then the items are recommended to the target user.

[0219] In another example, the recommendation method provided by the present application also includes: detecting that the target user performs a target operation on the recommendation result display page, the target operation is used to request a result analysis page for displaying the recommendation result; in response to the target operation, the result analysis page is displayed, the result analysis page includes an explanatory diagram of the recommendation result, and the explanatory diagram at least indicates the pairing process of the target user in the source domain and the paired user in the target domain.

[0220] Thanks to the method of pairing users in the source domain and the target domain through user intent embedding features, and the method of processing user event graphs using graph convolution, user intentions and contextual preferences are traceable, so a user interface can be implemented that displays the cross-domain user pairing process and explains the recommendation results.

[0221] Optionally, the explanatory diagram indicates the target user's interaction preference distribution and context preference distribution in the source domain, as well as the paired user's interaction preference distribution and context preference distribution in the target domain, and the recommendation results determined based on the paired user's historical interaction information in the target domain; wherein, the target user's interaction preference distribution in the source domain indicates the target user's preference for each item feature in the source domain, the context preference distribution indicates the target user's preference for each context feature in the source domain, the paired user's interaction preference distribution in the target domain indicates the paired user's preference for each item feature in the target domain, and the context preference distribution indicates the paired user's preference for each context feature in the target domain.

[0222] Through the explanatory diagram, the reasons for the recommendation results are shown to the user, such as the users paired with the target user in the target domain, the preference distribution of the target user in the source domain, the preference distribution of the paired users in the target domain, and the information recommended to the target user based on the historical interaction information of the paired users. In this way, the user can have a clear understanding of the reasons for the recommendation results.

[0223] For example, when a user clicks the "?" icon on the recommendation interface, a result analysis interface is displayed to the user in response to this click. This interface displays an explanation diagram, which shows the contextual preference distribution and interaction preference distribution of the target user in the source domain, the users paired with the target user in the target domain, and the contextual preference distribution and interaction preference distribution of the user in the target domain. Optionally, the contextual preference distribution can be displayed as connection weights, and the interaction preference distribution can also be displayed as connection weights, which clearly reflects the user's preference for the item information in certain contextual scenarios. In other words, the recommendation results are highly interpretable, improving the user experience.

[0224] Figure 12 A schematic diagram of an example of explaining a recommendation result is shown. Figure 12 As shown, the target user is "Li Si" and the target domain is the hotel recommendation domain. Li Si clicks to enter the hotel reservation interface, triggering a recommendation request. In response to the recommendation request, the recommendation system uses the recommendation method provided in the embodiment of the present application to recommend a hotel to user Li Si. When Li Si wants to know why it is recommended, he clicks the "question mark" icon on the recommendation interface. The recommendation system responds to the operation and switches to the result analysis interface. The result analysis interface displays an explanation diagram, which is drawn based on the original data of the model. Since Li Si lacks interaction in the hotel domain, the system first reads Li Si's interaction preference distribution and context preference distribution in the restaurant domain, and then finds the user Wang Wu in the hotel domain who has an intention pairing with Li Si, and recommends the hotel (id 424) that Wang Wu has booked and is satisfied with in a similar context scenario to Li Si.

[0225] The recommendation method provided in the embodiment of the present application combines the user's interaction preference and average contextual intention preference to construct a more refined user intention, which is used to replace the coarse-grained user embedding in previous work for cross-domain intention matching. On the basis of pairing, a cross-domain link is constructed between users of the two domains to provide an accurate cross-domain information propagation path for subsequent graph convolution operations. Thanks to the accurate pairing relationship and efficient graph convolution method, the recommendation model can more effectively capture user interaction preferences and contextual preferences; at the same time, thanks to the method of pairing users in the source domain and the target domain through user intention embedding features, and the method of processing user event graphs by graph convolution, user intentions and contextual preferences are traceable, so it is possible to realize a user interface that displays the cross-domain information transfer process and then explains the recommendation results, so that the recommendation results are highly interpretable and the user experience is improved.

[0226] In order to verify the effectiveness of the recommendation method provided by the embodiment of this application, we conducted a large number of experiments on three data sets to prove it. Table 1 shows the information statistics of the three data sets. We grouped the three data sets into two groups and designed a total of 6 tasks to compare the effects of the cross-domain model (C-UEG-EL) and the single-domain model (UEG-EL). We report the task results in Tables 2-4. From the results in Tables 2-4, we can obtain the following observations:

[0227] 1) Whether building cross-domain links and knowledge transfer based on Yelp-OH or Amazon Books, both improve the performance of Yelp-NC. In particular, using Yelp-OH, which is more similar to Yelp-NC, produces relatively more significant results.

[0228] 2) For Yelp-OH, by migrating knowledge from Yelp-NC or Amazon-Book, the C-UEG-EL provided by the present embodiment has a clear advantage in HR metrics while maintaining a comparable trend in NDCG metrics. This also shows that the C-UEG-EL provided by the present embodiment helps to more comprehensively capture user preferences;

[0229] 3) Similarly, the C-UEL-EL provided in the embodiments of the present application can significantly improve the performance of UEG-EL on Amazon-Book. This is probably because the C-UEG-EL provided in the embodiments of the present application can better model and learn user preferences in Amazon-Book by collaborating with the contextual preferences of Yelp-NC or Yelp-OH.

[0230]

[0231] Table 1 Statistics of public datasets

[0232]

[0233] Table 2 Cross-domain experiment results 1

[0234]

[0235] Table 3 Cross-domain experiment results 2

[0236]

[0237] Table 4 Cross-domain experiment results 3

[0238] Based on the same concept as the aforementioned recommendation method embodiment, the present application embodiment also provides a recommendation device 1300, which can be deployed on any device, equipment, platform or equipment cluster with computing capabilities to implement the recommendation method provided in the present application embodiment, so as to establish cross-domain links through cross-domain user intention pairing, transfer the preference knowledge of cross-domain users with similar intentions, so as to make the recommendation results more diverse and more accurate. The recommendation device 1300 includes a method for implementing Figure 2-10 The units or modules of each step in the recommended method are shown.

[0239] Figure 13 This is a schematic diagram of a recommended device provided in an embodiment of the present application. Figure 13 As shown, the recommendation device 1300 includes at least an acquisition module 1301, a search module 1302 and a recommendation module 1303, wherein the acquisition module 1301 is used to obtain recommendation request information, and the recommendation request information carries the target user's information and the target user's context information in the target domain; the search module 1302 is used to search from the refined embedding feature database based on the target user's information to obtain a refined embedding feature set corresponding to the target user, and the refined embedding feature set includes cross-domain enhanced user refined embedding features, item refined embedding features and context refined embedding features, wherein the cross-domain enhanced user refined embedding features are determined based on the first user refined embedding features and the paired user's preference features in the source domain, and the first user refined embedding features are selected based on the target user's information. The feature is the user refined embedding feature corresponding to the target user in the target domain, the preference feature of the paired user in the source domain indicates the preference information determined based on the historical interaction information corresponding to the paired user in the source domain, the paired user is determined based on the user refined embedding feature and user intention embedding feature corresponding to the target user in the target domain, and the user refined embedding feature and user intention embedding feature corresponding to each user in the source domain, the first user refined embedding feature, the item refined embedding feature and the context refined embedding feature are obtained based on feature extraction of the historical interaction information of the target user in the target domain, and the target domain and the source domain are different recommendation domains; the recommendation module 1303 is used to determine the recommendation result of the target user in the target domain based on the context information and the refined embedding feature set.

[0240] In another possible implementation, a specific implementation of determining paired users based on the user refined embedding features and user intention embedding features corresponding to the target user in the target domain, and the user refined embedding features and user intention embedding features corresponding to each user in the source domain is as follows: determining the position distribution of each user in the target domain based on the user refined embedding features and user intention embedding features corresponding to each user in the target domain; determining the position distribution of each user in the source domain based on the user refined embedding features and user intention embedding features corresponding to each user in the source domain; determining a transformation matrix that transforms the position distribution of each user in the target domain into the position distribution of each user in the source domain with the goal of minimizing transmission cost; determining a paired user to be paired with the target user from each user in the source domain based on the user refined embedding features and user intention embedding features corresponding to the target user in the target domain, and the transformation matrix.

[0241] In another possible implementation, a specific implementation of determining the cross-domain enhanced user refined embedding feature based on the first user refined embedding feature and the preference feature of the paired user in the source domain is: based on the transformation matrix, the preference feature of the paired user in the source domain is converted to the target domain for representation to obtain the cross-domain preference feature; the cross-domain preference feature and the first user refined embedding feature are fused to obtain the cross-domain enhanced user refined embedding feature.

[0242] In another possible implementation, the preference feature of the paired user in the source domain includes a second user refined embedding feature and / or a user intention embedding feature of the paired user in the source domain, and the second user refined embedding feature is the corresponding user refined embedding feature of the paired user in the source domain.

[0243] In another possible implementation, the user intention embedding feature corresponding to the target user in the target domain is determined based on the multiple user event intention embedding features corresponding to the target user in the target domain, and the multiple user event intention embedding features corresponding to the target user in the target domain are obtained based on the intention feature extraction of the multiple historical interaction events corresponding to the target user in the target domain; the user intention embedding feature corresponding to each user in the source domain is determined based on the multiple user event intention embedding features corresponding to each user in the source domain, and the multiple user event intention embedding features corresponding to each user in the source domain are obtained based on the intention feature extraction of the multiple historical interaction events corresponding to each user in the source domain.

[0244] In another possible implementation, a specific implementation of the first user refined embedding feature, item refined embedding feature and context refined embedding feature is obtained based on feature extraction of the target user's historical interaction information in the target domain: based on the target user's historical interaction information in the target domain, a user event graph is constructed, wherein the user event graph includes user nodes, item nodes, context nodes and intention nodes, and the intention nodes are connected to the user nodes, item nodes and context nodes respectively. The user node is the node where the user feature corresponding to the target user is located, the item node is the node where the item feature corresponding to the item having an interactive relationship with the target user is located, the context node is the node where the context feature corresponding to the context of the target user's interaction with the item is located, and the intention node is the node where the intention feature corresponding to the intention of the target user's interaction with the item is located; the user event graph is used as the input of the graph convolutional neural network, and after the graph convolution operation of each node in the user event graph is performed by multiple layers of graph convolutional neural network layers, the first user refined embedding feature, item refined embedding feature and context refined embedding feature are output.

[0245] In another possible implementation, the intent feature of the intent node is determined based on the user feature on the user node connected to the intent node and the context feature on the context node.

[0246] In another possible implementation, the user event graph is used as the input of the graph convolutional neural network, and a specific implementation of outputting the first user refined embedding feature is as follows: performing a graph convolution operation on the intent feature of each intent node connected to the user node in the previous layer, and the item feature of the item node connected to each intent node, to obtain each sub-user feature of the user node in the current layer; performing an aggregation operation on each sub-user feature of the user node to obtain the embedding feature of the user node in the current layer; and obtaining the first user refined embedding feature based on the embedded features of the user node after being updated by each graph convolutional neural network layer in the multi-layer graph convolutional neural network layer.

[0247] In another possible implementation, the user event graph is used as the input of the graph convolutional neural network. After the multi-layer graph convolutional neural network layer performs graph convolution operations on each node in the user event graph, a specific implementation of outputting the refined embedding features of the item is as follows: performing graph convolution operations on the intent features of each intent node connected to the item node in the previous layer and the user features of the user node connected to each intent node to obtain each sub-item feature of the item node in the current layer; performing aggregation operations on each sub-item feature of the item node to obtain the embedding features of the item node in the current layer; and obtaining the refined embedding features of the item node based on the embedded features of the updated item nodes in each layer of the multi-layer graph convolutional neural network layer.

[0248] In another possible implementation, the user event graph is used as the input of the graph convolutional neural network, and a specific implementation of outputting the context-refined embedding feature is as follows: a graph convolution operation is performed on the user features of the user node connected to the intention node in the previous layer and the item features of the item node to obtain the intention features of the intention node in the current layer; based on the intention features of the intention node in the current layer and the attention weights of the context nodes connected to the intention node, each sub-context feature of the context node in the current layer is determined; an aggregation operation is performed on each sub-context feature of the context node in the current layer to obtain the embedding features of the context node in the current layer; and the context-refined embedding feature is obtained based on the embedding features of the context nodes updated by each graph convolutional neural network layer in the multi-layer graph convolutional neural network layer.

[0249] In another possible implementation, the recommendation module 1303 is specifically used to: use the context information and the refined embedded feature set as input to the recommendation network, output the target user's interaction probability with each item in the target domain; and determine the recommendation result for the target user in the target domain based on the target user's interaction probability with each item in the target domain.

[0250] Optionally, the recommendation network includes a decoder that uses refined embeddings enhanced with cross-domain information to predict the user's preference for items in context. FM can be used as the decoder implementation, with FM serving as the core component of the recommendation network. FM is a linear model with better interpretability and lower complexity than MLP.

[0251] In another possible implementation, the recommendation device 1300 provided in the present application also includes: a recommendation interpretation module 1304, which is used to detect that the target user performs a target operation on the recommendation result display page, and the target operation is used to request a result analysis page for displaying the recommendation results; in response to the target operation, the result analysis page is displayed, and the result analysis page includes an explanatory diagram of the recommendation results, and the explanatory diagram at least indicates the process of migrating the preference information of the paired user in the source domain to the target user in the target domain.

[0252] In another possible implementation, the preference information of the paired users in the source domain includes the interaction preference distribution information and context preference distribution information of the paired users in the source domain, the explanatory diagram indicates the interaction preference distribution information and context preference distribution information of the paired users in the source domain, the interaction preference distribution information and context preference distribution information of the target user in the target domain after aggregating the interaction preference distribution information and context preference distribution information of the paired users in the source domain, and the recommendation results determined based on the interaction preference distribution information and context preference distribution information of the target user in the target domain; wherein, the interaction preference distribution information of the paired users in the source domain indicates the preference degree of the paired users for each item feature in the source domain, the context preference distribution information of the paired users in the source domain indicates the preference degree of the paired users for each context feature in the source domain, the interaction preference distribution information of the target user in the target domain indicates the preference degree of the target user for each item feature in the target domain, and the context preference distribution information of the target user in the target domain indicates the preference degree of the target user for each context feature in the target domain.

[0253] In another possible implementation, the historical interaction information in the target domain is located in the terminal device; the feature extraction step of obtaining the first user refined embedding features, the item refined embedding features, and the context refined embedding features based on the historical interaction information of the target user in the target domain is performed in the terminal device; and the step of determining the cross-domain enhanced user refined embedding features based on the first user refined embedding features and the preference features of the paired user in the source domain is performed in the cloud.

[0254] In another possible implementation, the refined embedded feature database is updated according to a preset period. For example, the user event graph is updated every 48 hours using the historical interaction data generated within 48 hours, and then the refined embedded feature set corresponding to each user is updated to ensure the quality and freshness of the refined embedded feature set of each user in the refined embedded feature library.

[0255] The recommendation device 1300 according to the embodiment of the present application may be configured to execute the method described in the embodiment of the present application, and the above and other operations and / or functions of each module in the recommendation device 1300 are respectively to implement Figure 2-10 For the sake of brevity, the corresponding processes of each method in are not repeated here.

[0256] Based on the same concept as the aforementioned embodiment of another recommendation method, the present application embodiment also provides a recommendation device 1400, which can be deployed on any device, equipment, platform or equipment cluster with computing capabilities to implement the recommendation method provided in the present application embodiment, so as to achieve cross-domain user intention pairing and use the historical interaction information of the paired users in the target domain to recommend the target user, thereby solving the cold start problem. The recommendation device 1400 includes a method for implementing Figure 11 The units or modules of each step in the recommended method are shown.

[0257] Figure 14 This is a schematic diagram of a recommended device provided in an embodiment of the present application. Figure 14 As shown, the recommendation device 1400 includes at least an acquisition module 1401, a determination module 1402, a pairing module 1403 and a recommendation module 1404, wherein the acquisition module 1401 is used to obtain recommendation request information, and the recommendation request information carries the information of the target user; the determination module 1402 is used to determine the user refined embedding features and user intention embedding features of the target user in the source domain, and the user refined embedding features and user intention embedding features of the target user in the source domain are obtained based on feature extraction of the historical interaction information of the target user in the source domain; the pairing module 1403 is used to extract the user refined embedding features and user intention embedding features of the target user in the source domain based on the target user corresponding to the target user in the source domain. The embedded features and user intention embedded features, as well as the refined user embedded features and user intention embedded features of each user in the target domain, are used to determine the paired users paired with the target user in the target domain. The refined user embedded features and user intention embedded features of each user in the target domain are obtained based on feature extraction of historical interaction information of each user in the target domain. The source domain and the target domain are different recommendation domains. The number of interaction events of the target user in the target domain is less than the number of interaction events of the target user in the source domain. The recommendation module 1404 is used to determine the recommendation results of the target user in the target domain based on the historical interaction information of the paired users in the target domain.

[0258] In another possible implementation, the pairing module 1403 is specifically used to: determine the position distribution of each user in the target domain based on the user refined embedding features and user intention embedding features corresponding to each user in the target domain; determine the position distribution of each user in the source domain based on the user refined embedding features and user intention embedding features corresponding to each user in the source domain; determine the conversion matrix that converts the position distribution of each user in the source domain into the position distribution of each user in the target domain with the goal of minimizing the transmission cost; determine the paired user with the target user from each user in the target domain based on the user refined embedding features, user intention embedding features, and the conversion matrix corresponding to the target user in the source domain.

[0259] In another possible implementation, the user intention embedding feature corresponding to the target user in the source domain is determined based on the multiple user event intention embedding features corresponding to the target user in the source domain, and the multiple user event intention embedding features corresponding to the target user in the source domain are obtained based on the intention feature extraction of the multiple historical interaction events corresponding to the target user in the source domain; the user intention embedding feature corresponding to each user in the target domain is determined based on the multiple user event intention embedding features corresponding to each user in the target domain, and the multiple user event intention embedding features corresponding to each user in the target domain are obtained based on the intention feature extraction of the multiple historical interaction events corresponding to each user in the target domain.

[0260] In another possible implementation, the determination module 1402 is specifically used to: construct a user event graph based on the historical interaction information of the target user in the source domain, wherein the user event graph includes user nodes, item nodes, context nodes and intention nodes, and the intention nodes are connected to the user nodes, item nodes and context nodes respectively. The user node is the node where the user feature corresponding to the target user is located, the item node is the node where the item feature corresponding to the item having an interactive relationship with the target user is located, the context node is the node where the context feature corresponding to the context of the target user's interaction with the item is located, and the intention node is the node where the intention feature corresponding to the intention of the target user's interaction with the item is located; the user event graph is used as the input of the graph convolutional neural network, and after the graph convolution operation of each node in the user event graph is performed by multiple layers of graph convolutional neural network layers, the user refined embedding features of the target user in the source domain are output.

[0261] In another possible implementation, the user event graph is used as the input of the graph convolutional neural network. After the graph convolution operation of each node in the user event graph is performed by multiple layers of graph convolutional neural network layers, the refined user embedding features of the target user in the source domain are output. A specific implementation is as follows: the intent features of each intent node connected to the user node in the previous layer and the item features of the item node connected to each intent node are performed to obtain each sub-user feature of the user node in the current layer; each sub-user feature of the user node is aggregated to obtain the embedding features of the user node in the current layer; based on the embedded features of the user nodes updated by each layer of the graph convolutional neural network layer in the multi-layer graph convolutional neural network layer, the refined user embedding features of the target user in the source domain are obtained.

[0262] In another possible implementation, the recommendation device 1400 provided in the present application also includes a recommendation explanation module 1405, which is used to: detect that the target user performs a target operation on the recommendation result display page, and the target operation is used to request a result analysis page for displaying the recommendation results; in response to the target operation, display the result analysis page, and the result analysis page includes an explanation diagram of the recommendation results, and the explanation diagram at least indicates the pairing process between the target user in the source domain and the paired user in the target domain.

[0263] In another possible implementation, the explanatory diagram indicates the target user's interaction preference distribution information and contextual preference distribution information in the source domain, as well as the paired user's interaction preference distribution information and contextual preference distribution information in the target domain, and the recommendation results determined based on the paired user's historical interaction information in the target domain; wherein, the target user's interaction preference distribution information in the source domain indicates the target user's preference degree for each item feature in the source domain, the target user's contextual preference distribution information in the source domain indicates the target user's preference degree for each contextual feature in the source domain, the paired user's interaction preference distribution information in the target domain indicates the paired user's preference degree for each item feature in the target domain, and the paired user's contextual preference distribution information in the target domain indicates the paired user's preference degree for each contextual feature in the target domain.

[0264] The present application also provides a computing device comprising at least one processor, a memory, and a communication interface, wherein the processor is configured to execute Figure 2-12 The method described.

[0265] Figure 15 A schematic diagram of the structure of a computing device provided in an embodiment of the present application.

[0266] like Figure 15 As shown, the computing device 1500 includes at least one processor 1501, a memory 1502 and a communication interface 1503. The processor 1501, the memory 1502 and the communication interface 1503 are communicatively connected, and the communication connection can be achieved by a wired manner (such as a bus) or by a wireless manner. The communication interface 1503 is used to send and / or receive data sent by other devices; the memory 1502 stores computer instructions, and the processor 1501 executes the computer instructions to execute the recommendation method in the aforementioned method embodiment to establish a cross-domain link through cross-domain user intention pairing, migrate the preference knowledge of cross-domain users with similar intentions, so as to make the recommendation results more diverse, and make the recommendation results more accurate, and solve the problems of cold start and sparse interaction data of target users in the target domain.

[0267] It should be understood that in the embodiment of the present application, the processor 1501 may be a central processing unit (CPU), or may be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0268] The memory 1502 may include a read-only memory and a random access memory, and provides instructions and data to the processor 1501. The memory 1502 may also include a nonvolatile random access memory.

[0269] The memory 1502 may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. Among them, the non-volatile memory may be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory may be a random access memory (RAM), which is used as an external cache. By way of example and not limitation, many forms of RAM are available, such as static RAM (SRAM), dynamic random access memory (DRAM), synchronous DRAM (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link DRAM (SLDRAM), and direct rambus RAM (DR RAM).

[0270] It should be understood that the computing device 1500 according to the embodiment of the present application can execute the implementation of the embodiment of the present application. Figure 2-12 The method shown in the figure is described in detail above and will not be repeated here for the sake of brevity.

[0271] An embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the above-mentioned method is implemented.

[0272] An embodiment of the present application provides a chip, which includes at least one processor and an interface, wherein the at least one processor determines program instructions or data through the interface; the at least one processor is used to execute the program instructions to implement the method mentioned above.

[0273] An embodiment of the present application provides a computer program or a computer program product, which includes instructions. When the instructions are executed, the computer is caused to perform the above-mentioned method.

[0274] Those skilled in the art should further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the above description has generally described the composition and steps of each example according to function. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art may use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0275] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein may be implemented as software modules executed by hardware, a processor, or a combination of the two. The software modules may be placed in random access memory (RAM), internal memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0276] The specific implementation methods described above further illustrate the purpose, technical solutions and beneficial effects of this application. It should be understood that the above description is only the specific implementation methods of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of this application should be included in the scope of protection of this application.

Claims

1. A recommendation method, characterized in that: include: Obtaining recommendation request information, where the recommendation request information carries information about a target user and context information about the target user in a target domain; Based on the information of the target user, a search is performed in a refined embedding feature database to obtain a refined embedding feature set corresponding to the target user, wherein the refined embedding feature set includes cross-domain enhanced user refined embedding features, item refined embedding features, and context refined embedding features, wherein the cross-domain enhanced user refined embedding features are determined based on the first user refined embedding features and the preference features of the paired user in the source domain, the first user refined embedding features are the user refined embedding features corresponding to the target user in the target domain, the preference features of the paired user in the source domain indicate the preference information determined based on the historical interaction information corresponding to the paired user in the source domain, the paired user is determined based on the user refined embedding features and user intention embedding features corresponding to the target user in the target domain, and the user refined embedding features and user intention embedding features corresponding to each user in the source domain, the first user refined embedding features, the item refined embedding features, and the context refined embedding features are obtained based on feature extraction of the historical interaction information of the target user in the target domain, and the target domain and the source domain are different recommendation domains; Based on the context information and the refined embedding feature set, a recommendation result for the target user in the target domain is determined.

2. The method according to claim 1, characterized in that The method comprises: Determining a location distribution of each user in the target domain based on the refined user embedding features and the user intention embedding features corresponding to each user in the target domain; Determining the location distribution of each user in the source domain based on the refined user embedding features and the user intention embedding features corresponding to each user in the source domain; With the goal of minimizing transmission cost, determining a transformation matrix that transforms the position distribution of each user in the target domain into the position distribution of each user in the source domain; The paired user to be paired with the target user is determined from users in the source domain based on the user refined embedding feature, the user intention embedding feature, and the conversion matrix corresponding to the target user in the target domain.

3. The method according to claim 2, characterized in that The method comprises: Based on the conversion matrix, the preference features of the paired users in the source domain are converted into the target domain for representation to obtain cross-domain preference features; The cross-domain preference feature and the first user refined embedding feature are fused to obtain the cross-domain enhanced user refined embedding feature.

4. The method according to any one of claims 1 to 3, characterized in that The preference features of the paired user in the source domain include a second user refined embedding feature and / or a user intention embedding feature of the paired user in the source domain, where the second user refined embedding feature is a user refined embedding feature corresponding to the paired user in the source domain.

5. The method according to any one of claims 1 to 4, characterized in that The user intention embedding feature corresponding to the target user in the target domain is determined based on multiple user event intention embedding features corresponding to the target user in the target domain, and the multiple user event intention embedding features corresponding to the target user in the target domain are obtained based on intention feature extraction of multiple historical interaction events corresponding to the target user in the target domain; The user intention embedding features corresponding to each user in the source domain are determined based on multiple user event intention embedding features corresponding to each user in the source domain, and the multiple user event intention embedding features corresponding to each user in the source domain are obtained based on intention feature extraction of multiple historical interaction events corresponding to each user in the source domain.

6. The method according to any one of claims 1 to 5, characterized in that The method comprises: Based on the historical interaction information of the target user in the target domain, a user event graph is constructed, wherein the user event graph includes a user node, an item node, a context node, and an intention node, wherein the intention node is connected to the user node, the item node, and the context node, respectively; the user node is a node where the user feature corresponding to the target user is located; the item node is a node where the item feature corresponding to the item having an interactive relationship with the target user is located; the context node is a node where the context feature corresponding to the context of the target user's interaction with the item is located; and the intention node is a node where the intention feature corresponding to the intention of the target user's interaction with the item is located; The user event graph is used as the input of a graph convolutional neural network. After a multi-layer graph convolutional neural network layer performs a graph convolution operation on each node in the user event graph, the first user refined embedding feature, the item refined embedding feature, and the context refined embedding feature are output.

7. The method according to claim 6, characterized in that The intention feature of the intention node is determined based on the user feature on the user node connected to the intention node and the context feature on the context node.

8. The method according to claim 6 or 7, characterized in that The method of using the user event graph as the input of a graph convolutional neural network, performing a graph convolution operation on each node in the user event graph through multiple layers of graph convolutional neural network layers, and outputting the first user refined embedding feature includes: Performing a graph convolution operation on the intent features of each intent node connected to the user node in the previous layer and the item features of the item nodes connected to each intent node to obtain each sub-user feature of the user node in the current layer; Performing an aggregation operation on each sub-user feature of the user node to obtain an embedded feature of the user node of the current layer; Based on the updated embedding features of the user node in each graph convolutional neural network layer in the multi-layer graph convolutional neural network layer, the first user refined embedding features are obtained.

9. The method according to any one of claims 6 to 8, characterized in that: The user event graph is used as the input of the graph convolutional neural network, and a graph convolution operation is performed on each node in the user event graph through multiple layers of the graph convolutional neural network to output the refined embedding features of the item, including: Performing a graph convolution operation on the intent features of each intent node connected to the item node in the previous layer and the user features of the user node connected to each intent node to obtain each sub-item feature of the item node in the current layer; Performing an aggregation operation on each sub-item feature of the item node to obtain an embedded feature of the item node at the current layer; Based on the updated embedding features of the item nodes in each graph convolutional neural network layer in the multi-layer graph convolutional neural network layer, the refined embedding features of the item are obtained.

10. The method according to any one of claims 6 to 9, characterized in that: The user event graph is used as the input of the graph convolutional neural network, and a graph convolution operation is performed on each node in the user event graph through a multi-layer graph convolutional neural network layer to output the context-refined embedding feature, including: Performing graph convolution on the user features of the user node and the item features of the item node connected to the intent node in the previous layer to obtain the intent features of the intent node in the current layer; Determine each sub-context feature of the context node of the current layer based on the intention feature of the intention node of the current layer and the attention weight of the context node connected to the intention node; Performing an aggregation operation on each sub-context feature of the context node of the current layer to obtain an embedded feature of the context node of the current layer; The context-refined embedding features are obtained based on the embedded features of the context nodes updated in each graph convolutional neural network layer in the multi-layer graph convolutional neural network layer.

11. The method according to any one of claims 1 to 10, characterized in that The determining, based on the context information and the refined embedding feature set, a recommendation result for the target user in the target domain includes: Taking the context information and the refined embedding feature set as inputs of a recommendation network, and outputting the probability of the target user interacting with each item in the target domain; Based on the interaction probability of the target user to each item in the target domain, a recommendation result for the target user in the target domain is determined.

12. The method according to any one of claims 1 to 11, characterized in that Also includes: detecting that the target user performs a target operation on the recommendation result display page, wherein the target operation is used to request display of a result analysis page of the recommendation result; In response to the target operation, the result analysis page is displayed, the result analysis page including an explanatory diagram of the recommendation result, the explanatory diagram at least indicating a process of migrating the preference information of the paired user in the source domain to the target user in the target domain.

13. The method according to claim 12, characterized in that The preference information of the paired user in the source domain includes interaction preference distribution information and context preference distribution information of the paired user in the source domain, the explanatory diagram indicates the interaction preference distribution information and context preference distribution information of the paired user in the source domain, the interaction preference distribution information and context preference distribution information of the target user in the target domain after aggregating the interaction preference distribution information and context preference distribution information of the paired user in the source domain, and the recommendation result determined based on the interaction preference distribution information and context preference distribution information of the target user in the target domain; Among them, the interaction preference distribution information of the paired user in the source domain indicates the preference degree of the paired user for each item feature in the source domain, the context preference distribution information of the paired user in the source domain indicates the preference degree of the paired user for each context feature in the source domain, the interaction preference distribution information of the target user in the target domain indicates the preference degree of the target user for each item feature in the target domain, and the context preference distribution information of the target user in the target domain indicates the preference degree of the target user for each context feature in the target domain.

14. The method according to any one of claims 1 to 13, characterized in that The historical interaction information in the target domain is located in the terminal device; A feature extraction step of obtaining the first user refined embedding feature, the item refined embedding feature, and the context refined embedding feature based on historical interaction information of the target user in the target domain is performed in the terminal device; The step of determining the cross-domain enhanced user refined embedding feature based on the first user refined embedding feature and the preference feature of the paired user in the source domain is performed in the cloud.

15. The method according to any one of claims 1 to 14, characterized in that The refined embedding feature database is updated according to a preset period.

16. A recommendation method, characterized in that: include: Obtaining recommendation request information, wherein the recommendation request information carries information of a target user; Determining refined user embedding features and user intention embedding features of the target user in the source domain, where the refined user embedding features and user intention embedding features of the target user in the source domain are obtained based on feature extraction of historical interaction information of the target user in the source domain; Determining a paired user with the target user in the target domain based on the refined user embedding features and user intention embedding features corresponding to the target user in the source domain, as well as the refined user embedding features and user intention embedding features of each user in the target domain, wherein the refined user embedding features and user intention embedding features of each user in the target domain are obtained based on feature extraction of historical interaction information of each user in the target domain, the source domain and the target domain are different recommendation domains, and the number of interaction events of the target user in the target domain is less than the number of interaction events of the target user in the source domain; Based on the historical interaction information of the paired users in the target domain, a recommendation result for the target user in the target domain is determined.

17. The method according to claim 16, characterized in that The determining a paired user with the target user in the target domain based on the refined user embedding features and user intention embedding features of the target user in the source domain, and the refined user embedding features and user intention embedding features of each user in the target domain, includes: Determining a location distribution of each user in the target domain based on the refined user embedding features and the user intention embedding features corresponding to each user in the target domain; Determining the location distribution of each user in the source domain based on the refined user embedding features and the user intention embedding features corresponding to each user in the source domain; With the goal of minimizing transmission cost, determining a transformation matrix that transforms the position distribution of each user in the source domain into the position distribution of each user in the target domain; The paired user to be paired with the target user is determined from among users in the target domain based on the refined user embedding feature, the user intention embedding feature, and the conversion matrix corresponding to the target user in the source domain.

18. The method according to claim 17, characterized in that The user intention embedding feature corresponding to the target user in the source domain is determined based on multiple user event intention embedding features corresponding to the target user in the source domain, and the multiple user event intention embedding features corresponding to the target user in the source domain are obtained based on intention feature extraction of multiple historical interaction events corresponding to the target user in the source domain; The user intention embedding features corresponding to each user in the target domain are determined based on multiple user event intention embedding features corresponding to each user in the target domain, and the multiple user event intention embedding features corresponding to each user in the target domain are obtained based on intention feature extraction of multiple historical interaction events corresponding to each user in the target domain.

19. The method according to any one of claims 16 to 18, characterized in that: The determining of the refined user embedding features of the target user in the source domain includes: Based on the historical interaction information of the target user in the source domain, a user event graph is constructed, wherein the user event graph includes a user node, an item node, a context node, and an intention node, wherein the intention node is connected to the user node, the item node, and the context node, respectively; the user node is a node where the user feature corresponding to the target user is located; the item node is a node where the item feature corresponding to the item having an interactive relationship with the target user is located; the context node is a node where the context feature corresponding to the context of the target user's interaction with the item is located; and the intention node is a node where the intention feature corresponding to the intention of the target user's interaction with the item is located; The user event graph is used as the input of the graph convolutional neural network. After the graph convolution operation of each node in the user event graph is performed by multiple layers of graph convolutional neural network layers, the refined user embedding features of the target user in the source domain are output.

20. The method according to claim 19, characterized in that The user event graph is used as the input of the graph convolutional neural network, and a graph convolution operation is performed on each node in the user event graph through multiple layers of the graph convolutional neural network to output the user refined embedding features of the target user in the source domain, including: Performing a graph convolution operation on the intent features of each intent node connected to the user node in the previous layer and the item features of the item nodes connected to each intent node to obtain each sub-user feature of the user node in the current layer; Performing an aggregation operation on each sub-user feature of the user node to obtain an embedded feature of the user node of the current layer; Based on the updated embedding features of the user nodes in each graph convolutional neural network layer in the multi-layer graph convolutional neural network layer, the refined embedding features of the target user in the source domain are obtained.

21. The method according to any one of claims 16 to 20, characterized in that Also includes: detecting that the target user performs a target operation on the recommendation result display page, wherein the target operation is used to request display of a result analysis page of the recommendation result; In response to the target operation, the result analysis page is displayed, the result analysis page including an explanatory diagram of the recommendation result, the explanatory diagram at least indicating a pairing process between the target user in the source domain and the paired user in the target domain.

22. The method according to claim 21, characterized in that The explanation graph indicates the interaction preference distribution information and context preference distribution information of the target user in the source domain, the interaction preference distribution information and context preference distribution information of the paired user in the target domain, and the recommendation result determined based on the historical interaction information of the paired user in the target domain; Among them, the interaction preference distribution information of the target user in the source domain indicates the target user's preference degree for each item feature in the source domain, the context preference distribution information of the target user in the source domain indicates the target user's preference degree for each context feature in the source domain, the interaction preference distribution information of the paired user in the target domain indicates the paired user's preference degree for each item feature in the target domain, and the context preference distribution information of the paired user in the target domain indicates the paired user's preference degree for each context feature in the target domain.

23. A recommendation device, characterized in that: include: an acquisition module, configured to acquire recommendation request information, wherein the recommendation request information carries information of a target user and context information of the target user in a target domain; A search module is configured to search a refined embedding feature database based on the target user's information to obtain a refined embedding feature set corresponding to the target user, wherein the refined embedding feature set includes cross-domain enhanced user refined embedding features, item refined embedding features, and context refined embedding features, wherein the cross-domain enhanced user refined embedding features are determined based on first user refined embedding features and preference features of paired users in a source domain, the first user refined embedding features are user refined embedding features corresponding to the target user in the target domain, the preference features of paired users in the source domain indicate preference information determined based on historical interaction information corresponding to the paired users in the source domain, the paired users are determined based on user refined embedding features and user intention embedding features corresponding to the target user in the target domain, and user refined embedding features and user intention embedding features corresponding to each user in the source domain, the first user refined embedding features, the item refined embedding features, and the context refined embedding features are obtained based on feature extraction of historical interaction information of the target user in the target domain, and the target domain and the source domain are different recommendation domains; A recommendation module is configured to determine a recommendation result for the target user in the target domain based on the context information and the refined embedding feature set.

24. A recommendation device, characterized in that include: An acquisition module is used to acquire recommendation request information, wherein the recommendation request information carries information of a target user; a determination module, configured to determine refined user embedding features and user intention embedding features of the target user in the source domain, wherein the refined user embedding features and user intention embedding features of the target user in the source domain are obtained based on feature extraction of historical interaction information of the target user in the source domain; a pairing module, configured to determine a pairing user to be paired with the target user in the target domain based on the refined user embedding features and user intention embedding features corresponding to the target user in the source domain, and the refined user embedding features and user intention embedding features of each user in the target domain, wherein the refined user embedding features and user intention embedding features of each user in the target domain are obtained based on feature extraction of historical interaction information of each user in the target domain, the source domain and the target domain are different recommendation domains, and the number of interaction events of the target user in the target domain is less than the number of interaction events of the target user in the source domain; The recommendation module is configured to determine a recommendation result for the target user in the target domain based on historical interaction information of the paired users in the target domain.

25. A computing device comprising a memory and a processor, characterized in that: Instructions are stored in the memory, and when the instructions are executed by the processor, the method according to any one of claims 1 to 22 is implemented.

26. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 22 is implemented.

Images