A recommendation method, apparatus and electronic device

Abstract

A recommendation method comprises: determining a user event graph (UEG) based on training data, the UEG comprising: a user node, an item node, a context node, and an intent node, wherein the intent node is connected with the user node, the item node, and the context node respectively; performing graph convolution on user features, item features, context features, and intent features based on the UEG to obtain a target feature set, the target feature set comprising target user features, target item features, and target context features; and recommending a target item to a target user based on the target feature set and the context features of the target user in a target environment, the target item being an item matching the context features in the target environment. In this way, the user node, the item node, and the context node are associated through the intent node, so that the context features can be integrated into the graph convolution, and better user features and item features can be obtained, thereby improving the accuracy of subsequent recommendation services.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence (AI) technology, and more particularly to a recommendation method, apparatus, and electronic device. Background Technology

[0002] Recommendation services are crucial tools for alleviating information overload and improving user experience, and also a significant means of monetizing AI. The key data source for building recommendation services is the interaction between users and items (e.g., clicks). Utilizing this interaction data allows for the prediction of user preferences and the recommendation of items that users might be interested in. The accuracy of the recommended items significantly impacts the user experience. Therefore, improving the accuracy of recommendation services is a pressing technical challenge that needs to be addressed. Summary of the Invention

[0003] This application provides a recommendation method, apparatus, electronic device, computer storage medium, computer product, and chip that can incorporate contextual features into graph convolution, thereby improving the accuracy of recommendation services.

[0004] Firstly, this application provides a recommendation method, which includes: determining a User Event Graph (UEG) based on training data, wherein the UEG includes user nodes, item nodes, context nodes, and intent nodes, wherein the intent nodes are connected to the user nodes, item nodes, and context nodes respectively; the user nodes are the nodes containing the user's user features; the item nodes are the nodes containing the item features of the items selected by the user; the context nodes are the nodes containing the context features that influence the user's selection of items; and the intent nodes are the nodes containing the intent features of the user's intent to select items; based on the UEG, performing graph convolution on the user features, item features, context features, and intent features to obtain a target feature set, which includes target user features, target item features, and target context features; and recommending target items to the target user based on the target feature set and the context features of the target user in the target environment, wherein the target items are items that match the context features in the target environment.

[0005] In this way, by associating user nodes, item nodes, and context nodes through intent nodes, contextual features can be incorporated into graph convolution, resulting in better user and item features and improving the accuracy of subsequent recommendation services.

[0006] In one possible implementation, intent features are derived based on user features and context features.

[0007] In one possible implementation, there are multiple context features, and the method further includes: determining the weight value of each context feature based on the user features and each context feature; and determining the intent feature based on each context feature and its weight value.

[0008] In one possible implementation, based on UEG, graph convolution is performed on user features, item features, context features, and intent features to obtain target user features. Specifically, this includes: determining the first user feature associated with the first intent feature in the h-th layer of the convolution graph based on the first intent feature in the (h-1)-th layer and the item features associated with the first intent feature in the (h-1)-th layer; and determining the target user feature based on the first user features obtained from each layer of the convolution graph when the number of convolutions in the convolution graph reaches a preset number.

[0009] Furthermore, the method also includes: when there are multiple item features associated with the first intent feature in the (h-1)th layer, for each item feature associated with the first intent feature, a second user feature associated with the first intent feature in the hth layer is determined based on the first intent feature in the (h-1)th layer and the item feature, so as to obtain multiple second user features; based on the multiple second user features, a first user feature associated with the first intent feature in the hth layer is determined; when the number of convolutions in the convolutional graph reaches a preset number, a target user feature is determined based on the first user features obtained in each layer of the convolutional graph.

[0010] In one possible implementation, based on UEG, graph convolution is performed on user features, item features, context features, and intent features to obtain target item features. Specifically, this includes: determining the first item feature in the h-th layer of the convolution graph based on the user features associated with the first item feature in the (h-1)-th layer and the intent features associated with the first item feature in the (h-1)-th layer; and determining the target item feature based on the first item features obtained in each layer of the convolution graph when the number of convolutions in the convolution graph reaches a preset number.

[0011] Furthermore, the method also includes: when there are multiple user features associated with the first item feature in the (h-1)th layer, for each user feature associated with the first intent feature, a second item feature in the hth layer is determined based on the user feature and the intent feature associated with the first item feature in the (h-1)th layer, so as to obtain multiple second item features; based on the multiple second item features, the first item feature is determined in the hth layer; when the number of convolutions in the convolutional graph reaches a preset number, the target item feature is determined based on the first item features obtained in each layer of the convolutional graph.

[0012] In one possible implementation, based on UEG, graph convolution is performed on user features, item features, context features, and intent features to obtain target context features. Specifically, this includes: determining the first intent feature in layer h of the convolution graph based on the user features and item features associated with the first intent feature in layer (h-1) of the convolution graph; determining the first context feature in layer h based on the first intent feature in layer h, the first context feature associated with the first intent feature in layer (h-1), and the first user feature associated with the first intent feature in layer h; and determining the target context feature based on the first context features obtained from each layer of the convolution graph when the number of convolutions in the convolution graph reaches a preset number.

[0013] Furthermore, the method also includes: when there are multiple user features associated with the first context feature in layer h, for each user feature associated with the first context feature, based on the first intent feature in layer h and the first context feature associated with the first intent feature in layer (h-1), a second context feature in layer h is determined to obtain multiple second context features; based on the multiple second context features, the first context feature associated with the first context feature in layer h is determined; when the number of convolutions in the convolutional graph reaches a preset number, the target context feature is determined based on the first context features obtained in each layer of the convolutional graph.

[0014] In one possible implementation, the method further includes: when a first context feature is associated with multiple first intent features, filtering the multiple first intent features during at least one convolution process to obtain a desired first intent feature, wherein the similarity value between the desired first intent feature and the first context feature is higher than a preset similarity value.

[0015] Secondly, this application provides a recommendation apparatus, comprising: a graph construction module, used to determine a User Event Graph (UEG) based on training data, wherein the UEG includes: user nodes, item nodes, context nodes, and intent nodes, wherein the intent nodes are connected to the user nodes, item nodes, and context nodes respectively; the user nodes are the nodes containing the user's user features; the item nodes are the nodes containing the item features of the item selected by the user; the context nodes are the nodes containing the context features that influence the user's selection of the item; and the intent nodes are the nodes containing the intent features of the user's intent to select the item; a graph convolution module, used to perform graph convolution on the user features, item features, context features, and intent features based on the UEG to obtain a target feature set, wherein the target feature set includes target user features, target item features, and target context features; and a recommendation module, used to recommend target items to the target user based on the target feature set and the context features of the target user in the target environment, wherein the target items are items that match the context features in the target environment.

[0016] In one possible implementation, intent features are derived based on user features and context features.

[0017] In one possible implementation, there are multiple context features. The graph convolution module is also used to determine the weight values ​​of each context feature based on the user features and each context feature; and to determine the intent features based on each context feature and its weight values.

[0018] In one possible implementation, the graph convolution module is specifically used to: determine the first user feature associated with the first intention feature in the h-th layer of the convolution graph based on the first intention feature in the (h-1)-th layer of the convolution graph and the item feature associated with the first intention feature in the (h-1)-th layer of the convolution graph; and determine the target user feature based on the first user feature obtained in each layer of the convolution graph when the number of convolutions in the convolution graph reaches a preset number.

[0019] Furthermore, the graph convolution module is also used to: when there are multiple item features associated with the first intent feature in the (h-1)th layer, for each item feature associated with the first intent feature, determine the second user feature associated with the first intent feature in the hth layer based on the first intent feature in the (h-1)th layer and the item feature, so as to obtain multiple second user features; determine the first user feature associated with the first intent feature in the hth layer based on the multiple second user features; when the number of convolutions in the convolution graph reaches a preset number, determine the target user feature based on the first user features obtained in each layer of the convolution graph.

[0020] In one possible implementation, the graph convolution module is specifically used to: determine the first item feature in the h-th layer of the convolution graph based on the user feature associated with the first item feature in the (h-1)-th layer of the convolution graph, and the intent feature associated with the first item feature in the (h-1)-th layer; and determine the target item feature based on the first item feature obtained in each layer of the convolution graph when the number of convolutions in the convolution graph reaches a preset number.

[0021] Furthermore, the graph convolution module is also used to: when there are multiple user features associated with the first item feature in the (h-1)th layer, for each user feature associated with the first intent feature, determine the second item feature in the hth layer based on the user feature and the intent feature associated with the first item feature in the (h-1)th layer, so as to obtain multiple second item features; based on the multiple second item features, determine the first item feature in the hth layer; when the number of convolutions in the convolution graph reaches a preset number, determine the target item feature based on the first item features obtained in each layer of the convolution graph.

[0022] In one possible implementation, the graph convolution module is specifically used to: determine the first intent feature in the h-th layer of the convolution graph based on the user feature and item feature associated with the first intent feature in the (h-1)-th layer of the convolution graph; determine the first context feature in the h-th layer based on the first intent feature in the h-th layer, the first context feature associated with the first intent feature in the (h-1)-th layer, and the first user feature associated with the first intent feature in the h-th layer; and determine the target context feature based on the first context features obtained in each layer of the convolution graph when the number of convolutions in the convolution graph reaches a preset number.

[0023] Furthermore, the graph convolution module is also used to: when there are multiple user features associated with the first context feature in the h-th layer, for each user feature associated with the first context feature, determine the second context feature in the h-th layer based on the first intent feature in the h-th layer and the first context feature associated with the first intent feature in the (h-1)-th layer, so as to obtain multiple second context features; determine the first context feature in the h-th layer based on the multiple second context features; when the number of convolutions in the convolution graph reaches a preset number, determine the target context feature based on the first context features obtained in each layer of the convolution graph.

[0024] In one possible implementation, the graph convolution module is further configured to: when a first context feature is associated with multiple first intent features, filter the multiple first intent features during at least one convolution process to obtain the desired first intent feature, wherein the similarity value between the desired first intent feature and the first context feature is higher than a preset similarity value.

[0025] Thirdly, this application provides an electronic device, comprising: at least one memory for storing a program; and at least one processor for executing the program stored in the memory; wherein, when the program stored in the memory is executed, the processor is configured to execute the method described in the first aspect or any possible implementation thereof.

[0026] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when run on a processor, causes the processor to perform the method described in the first aspect or any possible implementation thereof.

[0027] Fifthly, this application provides a computer program product that, when run on a processor, causes the processor to perform the method described in the first aspect or any possible implementation thereof.

[0028] In a sixth aspect, this application provides a chip including at least one processor and an interface; the at least one processor obtains program instructions or data through the interface; the at least one processor is used to execute program line instructions to implement the method described in the first aspect or any possible implementation of the first aspect.

[0029] It is understood that the beneficial effects of the second to sixth aspects mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of a graph neural network structure provided in an embodiment of this application;

[0031] Figure 2 This is a schematic diagram of the structure of a service recommendation model provided in an embodiment of this application;

[0032] Figure 3 This is a schematic diagram of a user event graph provided in an embodiment of this application;

[0033] Figure 4 This is a schematic diagram of another user event diagram provided in an embodiment of this application;

[0034] Figure 5 This is a flowchart illustrating a recommended method provided in an embodiment of this application;

[0035] Figure 6 This is a schematic diagram of the structure of a recommended device provided in an embodiment of this application;

[0036] Figure 7 This is a schematic diagram of the structure of a chip provided in an embodiment of this application. Detailed Implementation

[0037] In this article, the term "and / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. The symbol " / " in this article indicates that the related objects are in an "or" relationship; for example, A / B means A or B.

[0038] The terms "first" and "second," etc., used in the specification and claims herein are used to distinguish different objects, not to describe a specific order of objects. For example, "first response message" and "second response message," etc., are used to distinguish different response messages, not to describe a specific order of response messages.

[0039] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0040] In the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more, for example, multiple processing units means two or more processing units, multiple elements means two or more elements, etc.

[0041] First, the terms used in this application example will be explained.

[0042] 1. User characteristics

[0043] User characteristics refer to the features that identify different users. Examples include user identifiers, gender, age, region, occupation, and registration information on a webpage.

[0044] 2. Item characteristics

[0045] Item features refer to the characteristics that identify the item selected by the user. For example, the attributes of the selected item. For instance, when an item is an e-commerce product, item features might include: the item's title, category, and size; when an item is a music product, item features might include: the song's category, name, singer, composer, and lyricist; when an item is a blog post, item features might include: the article's title, author, and category.

[0046] 3. Contextual features

[0047] Contextual features refer to external factors that influence a user's decision-making during interaction with an item. For example, when a user browses an article, contextual features may include one or more of the following: the article's keywords, author, category, the user's location, and time.

[0048] Next, the technical solutions involved in the embodiments of this application will be described.

[0049] Generally, contextual features can significantly influence a user's item selection. For example, in restaurant recommendations, current time and location can effectively filter out unsuitable candidates; in e-commerce, clicks in recent sessions provide a strong signal for a user's next purchase. Therefore, effectively integrating contextual features into a context-aware recommendation system (CARS) for user preference prediction can significantly improve the accuracy of the recommendations output by CARS.

[0050] In CARS systems, matrix factorization (MF), factorization machines (FM), or neural networks (NN) can be used to model the interactions between user features, item features, and contextual features to provide recommendation services. However, MF and FM methods rely on shallow models, making them ill-suited for handling complex user behaviors or data inputs. NN methods, on the other hand, often neglect higher-order structural information in the data, i.e., the relationships between data points, thus making it difficult to learn higher-order structural information. Additionally, graph neural networks (GNNs) can integrate neighbor information using message passing mechanisms, allowing nodes to access information from higher-order neighbors through multi-layer stacking. Therefore, using graph neural networks can improve the accuracy of recommendation services.

[0051] For example, Figure 1The structure of a graph neural network is shown. For example... Figure 1 As shown, the graph neural network may include an encoder, graph convolution layers (GCLs), and a decoder.

[0052] The encoder's input includes three features: user features, item features, and context features. The encoder encodes both user and item features into multi-hot encoding, and then performs average pooling on the encoded features to obtain the initial representation: p u and q i As for contextual features, since they are not refined in the graph convolutional layer, they are not average pooled; only the set of contextual features v is obtained. c .

[0053] In the graph convolutional layer, the output p of the encoder can be processed. u v c and q i Convolutional operations are performed. However, standard graph convolutions do not consider edge features. Edges often carry contextual features, resulting in poor output from graph convolutional layers and impacting the accuracy of subsequent recommendation services. A decoder can decode the output of graph convolutional layers to obtain the desired result.

[0054] To improve the accuracy of recommendation services, this application provides a service recommendation model that incorporates contextual features into graph convolution to obtain better user and item features, thereby improving the accuracy of subsequent recommendation services.

[0055] For example, Figure 2 An architecture for a service recommendation model is shown. For example... Figure 2 As shown, the service recommendation model 200 may include: a graph construction module 210, a graph convolution module 220, and a recommendation module 230.

[0056] The graph construction module 210 is primarily used to construct a user-event graph (UEG) based on training data, connecting user features, item features, and context features. The training data can include data associated with user features, data associated with item features, and data associated with context features. User features, item features, and context features can all be obtained by encoding the corresponding data.

[0057] The graphing module 210 can link "user characteristics, item characteristics, and contextual characteristics" by constructing "intent nodes." For example, taking a user characteristic as an example, the UEG constructed by the graphing module 210 can be as follows: Figure 3 As shown, in Figure 3 Middle,U i Q represents user characteristics. i V represents the characteristics of an item. i Representing contextual features, T i Indicative features; Figure 3 The node containing the intended feature is the intended node. Figure 3 As can be seen, user features, item features, and contextual features are linked together through the node where the intent feature is located. The intent feature can be understood as the feature corresponding to a user's intent to select a particular item.

[0058] It is understandable that the UEG graph constructed by the graphing module 210 can include: intent nodes, user nodes, item nodes, and context nodes. Intent nodes are the nodes containing intent features, user nodes are the nodes containing user features, item nodes are the nodes containing item features, and context nodes are the nodes containing context features. In the constructed UEG graph, one intent feature can correspond to one item feature, and one item feature can correspond to one intent feature, or multiple intent features. For example, as... Figure 4 As shown, in this UEG, item feature Q2 corresponds to intent features T2 and T3, respectively, and item feature Q4 corresponds to intent feature T5. When the constructed UEG contains only one user feature, the item feature corresponding to that user feature can also correspond to one intent feature, or multiple intent features; conversely, the intent feature corresponding to that user can correspond to one item feature.

[0059] Intent features can be obtained from user features and context features. In some embodiments, intent features can be determined using the following "Formula 1".

[0060]

[0061] Among them, U e For the e-th user feature, T c For U e The corresponding c-th intent feature, V i For U e The corresponding i-th context feature, α i Let W0 be the weight value of the i-th context feature, where W0 ∈ R. d×1 W1 and W2∈R d×d y∈R d×1For example, see [link to relevant documentation]. Figure 3 From Formula 1 above, we can determine that: T1 = α 11 *V1+α 12 *V2,T2=α 21 *V1+α 22 *V2+α 23 *V3.

[0062] Graph convolution module 220 is mainly used to perform convolution operations on the UEG constructed by graph construction module 210 to obtain the required user features, item features, and context features. Since intent nodes are added to the constructed UEG, traditional graph convolution methods cannot be used. Therefore, this application proposes a graph convolution method that can fully utilize the key role of intent nodes to explore the relationships between user features, item features, and context features. The convolution methods for each feature are described below.

[0063] (1) User characteristics

[0064] For user characteristics, intent nodes can be used as bridges to pass information about item characteristics and contextual characteristics to user characteristics, as shown in "Formula 2" below.

[0065]

[0066] in, This refers to the user features of the i-th user in the h-th layer of the convolutional graph; It is an intent feature associated with the i-th user in the (h-1)-th layer; It is the intention feature in the (h-1)th layer. Related item characteristics.

[0067] For each user feature, when there are multiple associated item features, a user feature can be calculated from each item feature, resulting in multiple user features. To reduce subsequent computational load and improve computational speed, these user features can be aggregated into one and used as the user feature in the h-th layer. This allows for a clear understanding of which items a user has interacted with. For example, multiple user features can be aggregated into one by averaging, which can be done using "Formula 3" below.

[0068]

[0069] in, For the user features obtained by aggregation in the h-th layer, This refers to the user features obtained in the h-th layer based on the x-th item feature.

[0070] Finally, after the number of convolutions reaches a preset number, the user features can be fine-tuned based on the user features obtained from each convolutional layer to obtain the final user features. For example, the user features obtained from each convolutional layer can be averaged, and the average value can be used as the final desired user features. This can be calculated using the following "Formula 4".

[0071]

[0072] in, The final calculated user features are represented by H, where H is the number of convolutional layers. These are the user features obtained in the h-th layer.

[0073] It should be understood that besides using averaging for fine-tuning, other methods (such as variance) can also be used, and this is not limited here. Furthermore, in addition to fine-tuning the user features after convolution, the last obtained user features can also be directly used as the final desired user features. The specific choice depends on the actual situation and is not limited here.

[0074] (2) Item characteristics

[0075] For item features, intent nodes can be used as a bridge to pass information about user features and context features to item features, as shown in "Formula 5" below.

[0076]

[0077] in, This refers to an item feature in the h-th layer of the convolutional graph that is associated with the i-th user; This refers to the user features of the i-th user in the (h-1)-th layer of the convolutional graph; It is in the (h-1)th layer that the characteristics of the item are... An intentional feature associated with the relationship.

[0078] For each item feature, when there are multiple associated user features, an item feature can be calculated from each user feature, resulting in multiple item features. To reduce subsequent computational load and improve computational speed, these item features can be aggregated into one and used as the item feature in the h-th layer. This allows for a clear understanding of which users have interacted with an item. For example, multiple item features can be aggregated into one by averaging, which can be done using "Formula 6" below.

[0079]

[0080] in, The item features obtained by aggregation in the h-th layer, The item features are obtained in the h-th layer based on the x-th user feature.

[0081] Finally, after the number of convolutions reaches the preset number, for any item feature at any node, the item features can be fine-tuned based on the item features obtained from each convolutional layer to obtain the final item features. For example, the item features obtained from each convolutional layer can be averaged, and the average value can be used as the final required item features. This can be calculated using the following "Formula 7".

[0082]

[0083] in, The final calculated item features are represented by H, where H is the number of convolutional layers. The features of the items obtained in the h-th layer.

[0084] It should be understood that, besides using the averaging method for fine-tuning, other methods (such as variance) can also be used, and this is not limited here. Furthermore, after convolution, in addition to fine-tuning the item features, the last obtained item features can also be directly used as the final desired item features. The specific choice depends on the actual situation and is not limited here.

[0085] (3) Contextual features

[0086] For contextual features, information about user features and item features can be passed to intent features associated with contextual features, as shown in Formula 8 below.

[0087]

[0088] in, It refers to an intent feature associated with the i-th user in the h-th layer of the convolutional graph; This refers to the user features of the i-th user in the (h-1)-th layer of the convolutional graph; It is the intention feature in the (h-1)th layer. Related item characteristics.

[0089] Additionally, using the aforementioned "Formula 1", the user features in the h-th layer of the convolutional graph related to the i-th user and the intent features in the (h-1)-th layer of the convolutional graph can be analyzed. A context feature is associated with the data and processed to obtain the weight value corresponding to that context feature. Next, the corresponding feature of the context feature in the h-th layer of the convolutional graph can be determined using the following "Formula 9".

[0090]

[0091] in, This refers to the features in the h-th layer of the convolutional graph that correspond to the intent features. An associated contextual feature; This is achieved by using the aforementioned "Formula 1" to combine the user features of the i-th user in the h-th layer of the convolutional graph and the contextual features in the (h-1)-th layer of the convolutional graph. The context features at the corresponding nodes are processed.

[0092] For each context feature, when there are multiple associated user features, a context feature can be calculated from each user feature, resulting in multiple context features. To reduce subsequent computational load and improve computational speed, these context features can be aggregated into one and used as the context feature in the h-th layer. This allows for a clear understanding of which items a user interacted with in which contexts. For example, multiple context features can be aggregated into one by averaging, which can be done using "Formula 10" as described below.

[0093]

[0094] in, The context features obtained by aggregating in the h-th layer, This refers to the contextual features obtained in the h-th layer based on the x-th user feature.

[0095] Finally, after the number of convolutions reaches the preset number, for any node's context features, the context features can be fine-tuned based on the context features obtained from each convolutional layer to obtain the final context features. For example, the context features obtained from each convolutional layer can be averaged, and the average value can be used as the final required context features, which can be calculated using "Formula 11" below.

[0096]

[0097] in, The final calculated context features are represented by H, where H is the number of convolutional layers. The context features obtained in the h-th layer.

[0098] It should be understood that besides using averaging for fine-tuning, other methods (such as variance) can also be used, and this is not limited here. Furthermore, in addition to fine-tuning the context features after convolution, the last obtained context features can also be directly used as the final required context features. The specific choice depends on the actual situation and is not limited here.

[0099] Additionally, when the contextual features in the h-th layer of the convolutional graph are obtained... Then, it can be passed back to the intent node. In this process, the context features and user features in the h-th layer can be processed first using the aforementioned "Formula 1" to obtain the intent features in the h-th layer.

[0100] After performing the convolution operation, we can obtain finely tuned user features / item features / contextual features:

[0101]

[0102] in, For user characteristics, For the characteristics of the item, For contextual features.

[0103] In some embodiments, when a context feature is associated with multiple intent features, the convolution process on the context feature may be affected by noise. Therefore, to reduce the impact of noise, intent features closely related to the context feature can be selected first, and then the convolution operation can be performed. Simultaneously, since some intent features are discarded for the context feature, the computational load can be reduced subsequently, thus improving computational speed and model training efficiency.

[0104] One possible implementation is to first determine the first similarity value between the context feature and each intent feature using algorithms such as Euclidean distance, absolute value, or cosine similarity. Then, based on these first similarity values, an average similarity value is determined. Next, a second similarity value between each intent feature and the average similarity value can be determined. Finally, when the second similarity value is less than a preset similarity, it indicates a high similarity between the context feature and the intent feature corresponding to the second similarity value, and therefore the intent feature corresponding to the second similarity value can be retained; otherwise, it indicates a low similarity between the context feature and the intent feature corresponding to the second similarity value, and therefore the intent feature may be noise, so it can be discarded.

[0105] The recommendation module 230 is mainly used to recommend services to the target user based on the features output by the graph convolution module 220. For example, the recommendation module 230 can match the contextual features of the target user in the current environment with the features output by the graph convolution module 220 to recommend items to the target user that match the contextual features of their current environment.

[0106] Next, based on the content described above, a recommended method provided by an embodiment of this application will be introduced. It is understood that this method is proposed based on the content described above, and some or all of the content of this method can be found in the relevant descriptions above.

[0107] Please see Figure 5 , Figure 5 This is a flowchart illustrating a recommended method provided in an embodiment of this application. It is understood that this method can be executed by any device, equipment, platform, or cluster of devices with computing and processing capabilities. Figure 5 As shown, the recommended method includes the following steps:

[0108] S501. Based on the training data, determine the User Event Graph (UEG). The UEG includes: user nodes, item nodes, context nodes, and intent nodes. The intent nodes are connected to the user nodes, item nodes, and context nodes respectively. The user node is the node where the user's user features are located. The item node is the node where the item features of the item selected by the user are located. The context node is the node where the context features that affect the user's selection of the item are located. The intent node is the node where the intent features of the user's intent to select the item are located.

[0109] In this embodiment, the training data can be, but is not limited to, pre-configured data for the user. This data may include data related to user characteristics, data related to item characteristics, and data related to contextual characteristics. After the user configures the training data, a User Event Graph (UEG) can be determined based on the training data. The UEG includes: 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 the nodes containing the user's user characteristics, item nodes are the nodes containing the item characteristics of the item selected by the user, context nodes are the nodes containing the contextual characteristics that influence the user's item selection, and intent nodes are the nodes containing the intent characteristics of the user's intent to select the item. For example, the UEG can be... Figure 3 or Figure 4 As shown. For example, this step can be, but is not limited to, the aforementioned... Figure 2 The composition module 210 in the middle is executed.

[0110] In some embodiments, the intent features in the UEG can be obtained based on user features and context features. Where there are multiple context features, the weight value of each context feature can be determined first based on the user features and each context feature. Then, the intent features are determined based on each context feature and its weight value. See "Formula 1" above for details.

[0111] S502. Based on UEG, perform graph convolution on user features, item features, context features and intent features to obtain a target feature set, which includes target user features, target item features and target context features.

[0112] In this embodiment, after obtaining the UEG, graph convolution can be performed on user features, item features, context features, and intent features based on the UEG to obtain a target feature set. The target feature set includes target user features, target item features, and target context features. For example, this step can be, but is not limited to, the aforementioned... Figure 2 The graph convolution module 220 in the middle is executed.

[0113] In some embodiments, for target user features, the first user feature associated with the first intent feature in the h-th layer of the convolutional graph can be determined based on the first intent feature in the (h-1)-th layer of the convolutional graph and the item feature associated with the first intent feature in the (h-1)-th layer (see the description in "Formula 2" above). When the number of convolutions in the convolutional graph reaches a preset number, the target user feature is determined based on the first user features obtained in each layer of the convolutional graph. For example, the user feature obtained from the last convolution can be used as the target user feature, or the target user feature can be obtained through "Formula 4" above.

[0114] Furthermore, when there are multiple item features associated with the first intent feature in layer (h-1), for each item feature associated with the first intent feature, a second user feature associated with the first intent feature in layer h can be determined based on the first intent feature in layer (h-1) and any item feature, thus obtaining multiple second user features. For example, the aforementioned "Formula 2" can be used to process the first intent feature and any item feature in layer (h-1) to obtain the second user feature associated with the first intent feature in layer h. Then, based on the multiple second user features, the first user feature associated with the first intent feature in layer h can be determined; that is, a first user feature can be obtained by aggregating multiple second user features. For example, the aforementioned "Formula 3" can be used to process multiple second user features. Finally, when the number of convolutions in the convolutional graph reaches a preset number, the target user feature is determined based on the first user features obtained in each layer of the convolutional graph.

[0115] For the target item feature, the first item feature in layer h of the convolutional graph can be determined firstly based on the user feature associated with the first item feature in layer (h-1) of the convolutional graph, and the intent feature associated with the first item feature in layer (h-1) of the convolutional graph (see the description in "Formula 5" above). Then, when the number of convolutions in the convolutional graph reaches a preset number, the target item feature is determined based on the first item features obtained in each layer of the convolutional graph. For example, the item feature obtained from the last convolution can be used as the target item feature, or the target item feature can be obtained through "Formula 7" above.

[0116] Furthermore, when there are multiple user features associated with the first item feature in layer (h-1), for each user feature associated with the first intent feature, a second item feature in layer h is determined based on that user feature and the intent feature associated with the first item feature in layer (h-1), thus obtaining multiple second item features. For example, the aforementioned "Formula 5" can be used to process any user feature in layer (h-1) and the intent feature associated with the first item feature in layer (h-1) to obtain the second item feature in layer h. Then, based on the multiple second item features, a first item feature can be determined in layer h, i.e., multiple second item features can be aggregated to obtain a single first item feature. For example, the aforementioned "Formula 6" can be used to process multiple second item features. Finally, when the number of convolutions in the convolutional graph reaches a preset number, the target item feature is determined based on the first item features obtained in each layer of the convolutional graph.

[0117] For the target contextual features, the first intent feature in the h-th layer of the convolutional graph can be determined based on the user features and item features associated with the first intent feature in the (h-1)-th layer of the convolutional graph. For example, the first intent feature in the h-th layer of the convolutional graph can be determined using the aforementioned "Equation 8".

[0118] Next, based on the first intent feature in layer h, the first context feature associated with the first intent feature in layer (h-1), and the first user feature associated with the first intent feature in layer h, the first context feature in layer h can be determined. For example, the first user feature associated with the first intent feature in layer h and the first context feature associated with the first intent feature in layer (h-1) can be processed using the aforementioned "Formula 1" to determine the weight value corresponding to the first context feature. Then, the first intent feature in layer h and the determined weight can be processed using the aforementioned "Formula 9" to obtain the first context feature in layer h.

[0119] Finally, when the number of convolutions in the convolutional graph reaches a preset number, the target context feature is determined based on the first context features obtained from each layer in the convolutional graph. For example, the context feature obtained from the last convolution can be used as the target context feature, or the target context feature can be obtained through the aforementioned "Formula 11".

[0120] Furthermore, when there are multiple user features associated with the first context feature in layer h, for each user feature associated with the first context feature, a second context feature in layer h can be determined based on the first intent feature in layer h and the first context feature associated with the first intent feature in layer (h-1), thus obtaining multiple second context features. For example, the aforementioned "Formula 9" can be used to process any user feature in layer h, the first intent feature in layer h, and the first context feature associated with the first intent feature in layer (h-1), to obtain a second context feature in layer h. Then, based on the multiple second context features, a first context feature can be determined in layer h, i.e., multiple second context features can be aggregated to obtain a single first context feature. For example, the aforementioned "Formula 10" can be used to process multiple second context features. Finally, when the number of convolutions in the convolutional graph reaches a preset number, the target context feature is determined based on the first context features obtained from each layer in the convolutional graph.

[0121] In some embodiments, when a first context feature is associated with multiple first intent features, the multiple first intent features are filtered during at least one convolution process to obtain a desired first intent feature, wherein the similarity value between the desired first intent feature and the first context feature is higher than a preset similarity value. This removes the influence of noise and improves computational efficiency and accuracy.

[0122] After determining the target user characteristics, target item characteristics, and target context characteristics, S503 can be executed.

[0123] S503. Based on the target feature set and the contextual features of the target user in the target environment, recommend target items to the target user. The target items are items that match the contextual features in the target environment.

[0124] In this embodiment, after obtaining the contextual features of the target user in the target environment, the target user and its corresponding contextual features can be matched with the target feature set to obtain target items that are suitable for the target user and match the contextual features in the target environment, and to recommend target items to the target user.

[0125] In this way, by associating user nodes, item nodes, and context nodes through intent nodes, contextual features can be incorporated into graph convolution, resulting in better user and item features and improving the accuracy of subsequent recommendation services.

[0126] Based on the methods in the above embodiments, this application provides a recommended apparatus.

[0127] like Figure 6 As shown, the recommendation device 600 includes: a graph construction module 601, a graph convolution module 602, and a recommendation module 603. The graph construction module 601 is used to determine a User Event Graph (UEG) based on training data. The UEG includes 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, item nodes are nodes containing item features of items selected by the user, context nodes are nodes containing context features influencing the user's item selection, and intent nodes are nodes containing intent features of the user's intent to select an item. The graph convolution module 602 is used to perform graph convolution on user features, item features, context features, and intent features based on the UEG to obtain a target feature set, which includes target user features, target item features, and target context features. The recommendation module 603 is used to recommend target items to the target user based on the target feature set and the target user's context features in the target environment. Target items are items that match the context features in the target environment. For example, the graph construction module 601 can be, but is not limited to, the aforementioned... Figure 2 The graph construction module 210 is the same as that described above, and the graph convolution module 602 may be, but is not limited to, the same as that described above. Figure 2 The graph convolution module 220 is the same as that described above, and the recommended module 603 may be, but is not limited to, the same as that described above. Figure 2 The recommended module 230 is the same.

[0128] In some embodiments, intent features are derived based on user features and context features.

[0129] In some embodiments, there are multiple context features. The graph convolution module 602 is further configured to determine the weight value of each context feature based on the user features and each context feature; and to determine the intent feature based on each context feature and the weight value of each context feature.

[0130] In some embodiments, the graph convolution module 602 is specifically used to: determine a first user feature associated with the first intention feature in the h-th layer of the convolution graph based on the first intention feature in the (h-1)-th layer of the convolution graph and the item feature associated with the first intention feature in the (h-1)-th layer; and determine a target user feature based on the first user feature obtained in each layer of the convolution graph when the number of convolutions in the convolution graph reaches a preset number.

[0131] Furthermore, the graph convolution module 602 is also used to: when there are multiple item features associated with the first intent feature in the (h-1)th layer, for each item feature associated with the first intent feature, determine the second user feature associated with the first intent feature in the hth layer based on the first intent feature in the (h-1)th layer and the item feature, so as to obtain multiple second user features; determine the first user feature associated with the first intent feature in the hth layer based on the multiple second user features; when the number of convolutions in the convolution graph reaches a preset number, determine the target user feature based on the first user features obtained in each layer of the convolution graph.

[0132] In some embodiments, the graph convolution module 602 is specifically used to: determine the first item feature in the h-th layer of the convolution graph based on the user feature associated with the first item feature in the (h-1)-th layer of the convolution graph and the intent feature associated with the first item feature in the (h-1)-th layer; and determine the target item feature based on the first item feature obtained in each layer of the convolution graph when the number of convolutions in the convolution graph reaches a preset number.

[0133] Furthermore, the graph convolution module 602 is also used to: when there are multiple user features associated with the first item feature in the (h-1)th layer, for each user feature associated with the first intent feature, determine the second item feature in the hth layer based on the user feature and the intent feature associated with the first item feature in the (h-1)th layer, so as to obtain multiple second item features; based on the multiple second item features, determine the first item feature in the hth layer; when the number of convolutions in the convolution graph reaches a preset number, determine the target item feature based on the first item features obtained in each layer of the convolution graph.

[0134] In some embodiments, the graph convolution module 602 is specifically configured to: determine a first intent feature in layer h of the convolution graph based on user features and item features associated with the first intent feature in layer (h-1) of the convolution graph; determine a first context feature in layer h based on the first intent feature in layer h, a first context feature associated with the first intent feature in layer (h-1), and a first user feature associated with the first intent feature in layer h; and determine a target context feature based on the first context features obtained in each layer of the convolution graph when the number of convolutions in the convolution graph reaches a preset number.

[0135] Furthermore, the graph convolution module 602 is also used to: when there are multiple user features associated with the first context feature in the h-th layer, for each user feature associated with the first context feature, determine the second context feature in the h-th layer based on the first intent feature in the h-th layer and the first context feature associated with the first intent feature in the (h-1)-th layer, so as to obtain multiple second context features; determine the first context feature in the h-th layer based on the multiple second context features; when the number of convolutions in the convolution graph reaches a preset number, determine the target context feature based on the first context features obtained in each layer of the convolution graph.

[0136] In some embodiments, the graph convolution module 602 is further configured to: when a first context feature is associated with multiple first intent features, filter the multiple first intent features in at least one convolution process to obtain a desired first intent feature, wherein the similarity value between the desired first intent feature and the first context feature is higher than a preset similarity value.

[0137] It should be understood that the above-described device is used to execute the methods in the above embodiments. The implementation principle and technical effect of the corresponding program modules in the device are similar to those described in the above methods. The working process of the device can be referred to the corresponding process in the above methods, and will not be repeated here.

[0138] Based on the methods in the above embodiments, this application provides an electronic device. The electronic device may include: at least one memory for storing a program; and at least one processor for executing the program stored in the memory; wherein, when the program stored in the memory is executed, the processor executes the methods in the above embodiments.

[0139] Based on the methods in the above embodiments, this application provides a computer-readable storage medium storing a computer program that, when run on a processor, causes the processor to execute the methods in the above embodiments.

[0140] Based on the methods in the above embodiments, this application provides a computer program product, characterized in that, when the computer program product is run on a processor, the processor executes the methods in the above embodiments.

[0141] Based on the methods described in the above embodiments, this application also provides a chip. Please refer to... Figure 7 , Figure 7 This is a schematic diagram of a chip structure provided in an embodiment of this application. Figure 7 As shown, chip 700 includes one or more processors 701 and interface circuitry 702. Optionally, chip 700 may also include a bus 703. Wherein:

[0142] The processor 701 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed through integrated logic circuits in the hardware of the processor 701 or through software instructions. The processor 701 may be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods and steps disclosed in the embodiments of this application. The general-purpose processor may be a microprocessor or any conventional processor.

[0143] The interface circuit 702 can be used to send or receive data, instructions or information. The processor 701 can use the data, instructions or other information received by the interface circuit 702 to process the data, instructions or other information, and can send the processed information out through the interface circuit 702.

[0144] Optionally, chip 700 also includes memory, which may include read-only memory and random access memory, and provides operation instructions and data to the processor. A portion of the memory may also include non-volatile random access memory (NVRAM).

[0145] Optionally, the memory stores executable software modules or data structures, and the processor can execute corresponding operations by calling the operation instructions stored in the memory (which may be stored in the operating system).

[0146] Optionally, the interface circuit 702 can be used to output the execution results of the processor 701.

[0147] It should be noted that the functions of the processor 701 and the interface circuit 702 can be implemented through hardware design, software design, or a combination of hardware and software; no restrictions are imposed here.

[0148] It should be understood that each step of the above method embodiments can be completed by hardware logic circuits or software instructions in a processor.

[0149] It is understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application. In addition, in some possible implementations, each step in the above embodiments may be selectively executed according to the actual situation, and may be partially or fully executed, which is not limited here.

[0150] It is understood that the processor in the embodiments of this application can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. A general-purpose processor can be a microprocessor or any conventional processor.

[0151] The method steps in the embodiments of this application can be implemented in hardware or by a processor executing software instructions. The software instructions can consist of corresponding software modules, which can be stored in random access memory (RAM), flash memory, read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, portable hard disks, CD-ROMs, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can reside in an ASIC.

[0152] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted through the computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0153] It is understood that the various numerical designations used in the embodiments of this application are merely for descriptive convenience and are not intended to limit the scope of the embodiments of this application.

Claims

1. A recommendation method, characterized in that, The method includes: Based on training data, a User Event Graph (UEG) is determined. The UEG includes: user nodes, item nodes, context nodes, and intent nodes. The intent nodes are connected to the user nodes, item nodes, and context nodes, respectively. The user nodes are the nodes containing the user's user features, the item nodes are the nodes containing the item features of the item selected by the user, the context nodes are the nodes containing the context features that influence the user's item selection, and the intent nodes are the nodes containing the intent features of the user's intent to select the item. The intent features are obtained based on the user features and the context features. Based on the UEG, graph convolution is performed on the user features, the item features, the context features and the intent features to obtain a target feature set, which includes target user features, target item features and target context features; Based on the target feature set and the contextual features of the target user in the target environment, target items are recommended to the target user, wherein the target items are items that match the contextual features in the target environment.

2. The method according to claim 1, characterized in that, The context features are multiple, and the method further includes: Based on the user characteristics and each of the context features, the weight values ​​of each of the context features are determined respectively; The intent feature is determined based on each of the context features and the weight value of each of the context features.

3. The method according to any one of claims 1-2, characterized in that, Based on the UEG, graph convolution is performed on the user features, the item features, the context features, and the intent features to obtain the target user features, specifically including: Based on the first intent feature in the (h-1)th layer within the convolutional graph, and the item features associated with the first intent feature in the (h-1)th layer, a first user feature associated with the first intent feature in the h-th layer within the convolutional graph is determined; When the number of convolutions in the convolutional graph reaches a preset number, the target user features are determined based on the first user features obtained from each layer of the convolutional graph.

4. The method according to claim 3, characterized in that, The method further includes: When there are multiple item features associated with the first intent feature in the (h-1)th layer, for each item feature associated with the first intent feature, a second user feature associated with the first intent feature in the hth layer is determined based on the first intent feature in the (h-1)th layer and the item feature, so as to obtain multiple second user features. Based on the plurality of second user features, a first user feature associated with the first intent feature is determined in the h-th layer; When the number of convolutions in the convolutional graph reaches a preset number, the target user features are determined based on the first user features obtained from each layer of the convolutional graph.

5. The method according to any one of claims 1-2, characterized in that, Based on the UEG, graph convolution is performed on the user features, the item features, the context features, and the intent features to obtain the target item features, specifically including: Based on the user features associated with the first item feature in the (h-1)th layer within the convolutional graph, and the intent features associated with the first item feature in the (h-1)th layer, the first item feature in the h-th layer within the convolutional graph is determined; When the number of convolutions in the convolutional graph reaches a preset number, the target item features are determined based on the first item features obtained from each layer of the convolutional graph.

6. The method according to claim 3, characterized in that, Based on the UEG, graph convolution is performed on the user features, the item features, the context features, and the intent features to obtain the target item features, specifically including: Based on the user features associated with the first item feature in the (h-1)th layer within the convolutional graph, and the intent features associated with the first item feature in the (h-1)th layer, the first item feature in the h-th layer within the convolutional graph is determined; When the number of convolutions in the convolutional graph reaches a preset number, the target item features are determined based on the first item features obtained from each layer of the convolutional graph. The method further includes: When there are multiple user features associated with the first item feature in the (h-1)th layer, for each user feature associated with the first intent feature, a second item feature in the hth layer is determined based on the user feature and the intent feature associated with the first item feature in the (h-1)th layer, so as to obtain multiple second item features. Based on the plurality of second item features, determine the features that are related to the first item features in the h-th layer; When the number of convolutions in the convolutional graph reaches a preset number, the target item features are determined based on the first item features obtained from each layer of the convolutional graph.

7. The method according to any one of claims 1-2, characterized in that, Based on the UEG, graph convolution is performed on the user features, the item features, the context features, and the intent features to obtain the target context features, specifically including: Based on the user features and item features associated with the first intent feature in the (h-1)th layer within the convolutional graph, the first intent feature in the hth layer within the convolutional graph is determined; Based on the first intent feature in the h-th layer, the first context feature associated with the first intent feature in the (h-1)-th layer, and the first user feature associated with the first intent feature in the h-th layer, the first context feature in the h-th layer is determined; When the number of convolutions in the convolutional graph reaches a preset number, the target context feature is determined based on the first context feature obtained from each layer in the convolutional graph.

8. The method according to claim 7, characterized in that, The method further includes: When there are multiple user features associated with the first context feature in the h-th layer, for each user feature associated with the first context feature, a second context feature in the h-th layer is determined based on the first intent feature in the h-th layer and the first context feature associated with the first intent feature in the (h-1)-th layer, so as to obtain multiple second context features. Based on the plurality of second context features, determine the relationship with the first context feature in the h-th layer; When the number of convolutions in the convolutional graph reaches a preset number, the target context feature is determined based on the first context feature obtained from each layer in the convolutional graph.

9. The method according to any one of claims 1-2, characterized in that, The method further includes: When a first context feature is associated with multiple first intent features, the multiple first intent features are filtered during at least one convolution process to obtain a desired first intent feature, wherein the similarity value between the desired first intent feature and the first context feature is higher than a preset similarity value.

10. A recommended device, characterized in that, The device includes: A graph construction module is used to determine a User Event Graph (UEG) based on training data. The UEG includes user nodes, item nodes, context nodes, and intent nodes. The intent nodes are connected to the user nodes, item nodes, and context nodes, respectively. The user nodes are the nodes containing user features, the item nodes are the nodes containing item features of the items selected by the user, the context nodes are the nodes containing context features that influence the user's item selection, and the intent nodes are the nodes containing intent features of the user's intent to select the items. The intent features are obtained based on the user features and the context features. The graph convolution module is used to perform graph convolution on the user features, the item features, the context features and the intent features based on the UEG to obtain a target feature set, which includes target user features, target item features and target context features; The recommendation module is used to recommend target items to the target user based on the target feature set and the contextual features of the target user in the target environment. The target items are items that match the contextual features in the target environment.

11. An electronic device, characterized in that, include: At least one memory for storing programs; At least one processor for executing the program stored in the memory; When the program stored in the memory is executed, the processor is used to execute the method as described in any one of claims 1-9.

12. A computer-readable storage medium storing a computer program that, when run on a processor, causes the processor to perform the method as described in any one of claims 1-9.

13. A computer program product, characterized in that, When the computer program product is run on a processor, the processor causes the processor to perform the method as described in any one of claims 1-9.

14. A chip, characterized in that, Includes at least one processor and interface; The at least one processor obtains program instructions or data through the interface; The at least one processor is used to execute the program instructions to implement the method as described in any one of claims 1-9.

Images