Recommendation method based on latent graph structure mining and user long and short term interest fusion

Abstract

The application discloses a recommendation method based on latent graph structure mining and user long-term and short-term interest fusion, predicts the probability of a user clicking a target item based on a user historical behavior sequence, and mainly comprises seven parts: a first part is to divide the user historical behavior sequence into a recent historical behavior sequence and a long-term historical behavior sequence; a second part is to obtain the short-term interest of the user by using a recurrent neural network based on the recent historical behavior sequence of the user; a third part is to mine a latent graph structure according to the similarity of the items in the user historical behavior sequence; a fourth part is to update the item vector by using a graph neural network based on the mined graph structure, and to obtain the long-term interest of the user by using a recurrent neural network based on the updated item vector; a fifth part is to fuse the short-term interest and the long-term interest of the user according to the user interest diversity preference; and a sixth part is to predict the click rate of the user on the item.

Description

Technical Field

[0001] This invention belongs to the field of Internet service technology, and in particular relates to a recommendation method based on latent graph structure mining and the fusion of users' long-term and short-term interests. Background Technology

[0002] In recent years, Chinese mobile internet users have been spending increasingly more time on short video platforms. Due to the short duration of short videos, users can watch many videos throughout the day. Typically, users browse different types of short videos on these platforms, indicating a strong diversity of user preferences. Research has shown that inputting users' historical behavior sequences into short-term interest modeling models actually improves the model's performance when the most recent behaviors are removed, further demonstrating the strong diversity of user preferences.

[0003] Traditional recommendation methods typically input a user's recent behavior sequence into the model to obtain the user's short-term interests, and then recommend short videos that the user is interested in. However, user behavior on short video platforms is extremely rich and dense. If the user's long-term historical behavior sequence is ignored and only the recent historical behavior sequence is considered, the rich information contained in the user's long-term historical behavior sequence will be lost, and it may also overfit to the user's recent interests, resulting in relatively one-dimensional recommendation results. On the other hand, if all the user's historical behavior sequence is directly input into the interest modeling model, the large amount of noise contained in the user's long-term historical behavior sequence will impair the model's recommendation performance, and the model complexity will be too high, with long training and running times. Summary of the Invention

[0004] The problem defined in this method is to predict the probability of a user clicking on a target item based on a user's behavior sequence. The mathematical notation involved includes: the user set in the platform is represented as... The set of items is represented as .user The historical behavioral sequence is any of these items subscript Indicates user Behavioral sequence The length of the recommendation process is a key factor. Existing recommendation methods typically input a user's recent behavior sequence into the model to obtain the user's short-term interests, and then recommend short videos that the user is interested in. However, user behavior on short video platforms is extremely rich and dense. If the user's long-term historical behavior sequence is ignored and only the user's recent historical behavior sequence is considered, the rich information contained in the user's long-term historical behavior sequence will be lost, and it may overfit to the user's recent interests, resulting in a relatively one-dimensional recommendation result. On the other hand, if all of the user's historical behavior sequences are directly input into the interest modeling model, the large amount of noise contained in the user's long-term historical behavior sequence will impair the model's recommendation performance, and the model complexity will be too high, with long training and running times.

[0005] Therefore, the present invention adopts the following technical solution:

[0006] A recommendation method based on latent graph structure mining and the fusion of users' long-term and short-term interests is presented below:

[0007] S1. Obtain the user's historical behavior sequence and divide the user's historical behavior sequence into recent historical behavior sequence and long-term historical behavior sequence according to the order in which the user's behavior towards items is executed.

[0008] S2. Based on the user's recent historical behavior sequence, a recurrent neural network is used to model and obtain the user's short-term interests;

[0009] S3. Based on the similarity of items in the user's historical behavior sequence, a filtering operation consisting of multiple loops is used to mine potential graph structures.

[0010] S4. Based on the mined graph structure, update the item vector using a graph neural network, and then obtain the user's long-term interests using a long short-term memory neural network based on the updated item vector;

[0011] S5. Based on the diversity of user interests and preferences, calculate the richness of user behavior, and combine the weights of short-term and long-term interests obtained through learning to fuse the user's short-term and long-term interests to obtain the final user interest representation.

[0012] S6. By calculating the similarity between the user's interest representation and the vector representation of the target item, the click-through rate of the user on the item is predicted, thus realizing item recommendation.

[0013] As a preferred embodiment, the specific implementation method of S1 is as follows:

[0014] Get the user set in the platform and item collection Acquire users Historical Behavioral Sequence any of these items subscript Indicates user Behavioral sequence Length; The vector representation is ; Sequence of user's historical behavior The middle Each item is represented as a sequence of a user's recent historical behaviors. ; The previous user behavior sequence Each item is represented as a sequence of a user's long-term historical behavior. .

[0015] As a preferred embodiment, the specific implementation method of S2 is as follows:

[0016] For users Long Short-Term Memory Neural Network is used to analyze the user's recent historical behavior sequences. Modeling is performed to obtain users' short-term interests. :

[0017]

[0018] in, It is a sequence of recent user behavior. The vector representation of , This represents a Long Short-Term Memory (LSTM) neural network. The output of the last hidden state of the LSM neural network represents the user's short-term interest. .

[0019] As a preferred embodiment, the specific implementation method of S3 is as follows:

[0020] Calculate the cosine similarity between preceding and subsequent items in a user's historical behavior sequence to obtain the initial adjacency matrix of the graph structure. ;matrix It is a lower triangular matrix, when the i-th row and j-th column of the matrix... Subscript satisfies hour, ; Represents items in the user's historical behavior sequence and The similarity, specifically the value of the items and vector representation and The cosine similarity; further, the initial adjacency matrix... Performing a filtering operation yields a sparse graph structure. The filtering operation is a multi-cycle process:

[0021] First, let the final graph structure contain the set of nodes as follows: The set to be expanded is , and Initialize to ;

[0022] The second step is to use the initial adjacency matrix. ,get The set of most similar items for each item , will set Put the items in the collection In, that is and update for ,Right now ;

[0023] The third step is to repeat step two K times, ultimately obtaining a sparse graph structure. And the adjacency matrix of the graph In each of the multiple iterations described above, the adjacency matrix corresponds to the second step. The change can be expressed by the formula:

[0024]

[0025] in, Represents a set All items , Indicates finding and items Most similar items Then give the adjacency matrix Value assigned .

[0026] As a preferred embodiment, the specific implementation method of S4 is as follows:

[0027] S41. Based on the mined graph structure, update the item vector using a graph neural network; the graph neural network is a multi-level structure, through a total... The information delivery layer incorporates user behavior information from longer periods; among which Nodes in the diagram The Layer item vector representation is Its update formula is:

[0028]

[0029]

[0030] in, Represents the adjacency matrix Perform regularization; It is a graph structure The degree matrix is ​​calculated using the following formula: The degree matrix is ​​a diagonal matrix; It is an adjacency matrix The regularized matrix, where the matrix element in the i-th row and j-th column is ; express The neighborhood group, express The Layer item vector representation; nodes in the graph No. Layer vector Initialize to ;go through After the information is passed between layers, the node is finally obtained. The updated item vector is ;

[0031] S42, Based on the user's recent historical behavior sequence The updated item vector representation is as follows Then, long short-term memory neural networks are used to obtain users' long-term interests. ,in .

[0032] As a preferred embodiment, the specific implementation method of S5 is as follows:

[0033] S51, The number of categories in the statistical data set is: The number of label types is ; The number of user actions under a certain category or tag is greater than As a criterion for determining whether a user is interested in a particular category or tag, the number of categories that the user is interested in is counted in the user's historical behavior sequence. and number of tags To obtain the richness of user behavior for:

[0034]

[0035] in, and These are hyperparameters that control the importance of category and tag information, respectively. ;

[0036] S52. Use a two-layer MLP model to learn the weights of short-term and long-term interests, and combine them with the richness of user behavior. The final long-term and short-term interest weights are obtained as follows:

[0037]

[0038]

[0039] in, and These are the model parameters of a two-layer MLP model, which are updated during training. (Superscript) It is the transpose symbol; express Activation function express Activation function; It is the weight of users' long-term interests learned by the model itself, through... The function normalizes it to Within the range, then multiplied by the richness of user behavior. Obtain the final user long-term interest weight ;

[0040] S53, Based on user long-term interest weights and user short-term interest weight long-term user interest and short-term interests By merging these elements, the final representation of user interests can be obtained. ;

[0041] As a preferred embodiment, the specific implementation method of S6 is as follows:

[0042] According to user User interest representation Representing user interests and target items vector representation Perform inner product calculations to predict user preference for the target item. Click-through rate :

[0043] .

[0044] Preferably, the recommendation model framework consisting of S1 to S6 needs to be pre-trained before actual reasoning; during the training process, the user's behavior towards the target item is analyzed. Click-through rate prediction Calculate the predicted click-through rate. and the true value of click-through rate The cross-entropy loss function is used to guide the update process of model parameters, and the Adam optimizer is used to update the model parameters.

[0045] Preferably, the cross-entropy loss function is calculated as follows:

[0046]

[0047] in, This is the actual value, representing whether the user clicked on the target item; It is the sigmoid function.

[0048] Preferably, the platform is a short video platform, and the item is a short video. The vector representation of the main image of the short video is initialized.

[0049] The beneficial technical effects of the present invention are as follows:

[0050] (1) In order to make full use of the information in the user's long-term behavior sequence, this method proposes to divide the user's historical behavior sequence into recent historical behavior sequence and long-term historical behavior sequence, and extract information about the user's long-term interests from the long-term historical behavior sequence and integrate it into the item vector of the user's recent historical behavior sequence.

[0051] (2) This method proposes to construct a one-way graph structure based on historical behavior sequence. Specifically, a lower triangular adjacency matrix is ​​first calculated based on the similarity of items, and then a multi-loop process is used to filter out valuable edge information from the lower triangular adjacency matrix.

[0052] (3) When integrating users’ long-term and short-term interests, this method takes into account the diversity of users’ preferences and simulates the degree of diversity of users’ preferences by calculating the richness of item categories and tags involved in user behavior. Attached Figure Description

[0053] Figure 1 This is a flowchart illustrating the recommendation method based on latent graph structure mining and the fusion of users' long-term and short-term interests according to the present invention.

[0054] Figure 2 This is a model diagram of the recommendation method based on latent graph structure mining and fusion of users' long-term and short-term interests in this invention. Detailed Implementation

[0055] To further understand the present invention, the recommendation method based on latent graph structure mining and fusion of user long-term and short-term interests provided by the present invention will be described in detail below with reference to specific embodiments. However, the present invention is not limited thereto. Non-essential improvements and adjustments made by those skilled in the art under the core guiding principles of the present invention are still within the protection scope of the present invention.

[0056] The problem defined in this method is to predict the probability of a user clicking on a target item based on the user's historical behavior sequence. In the short video field, the item is the short video. The mathematical notation involved includes: the user set in the platform is represented as... The set of items is represented as .user The historical behavioral sequence is any of these items subscript Indicates user Behavioral sequence The recommendation method based on latent graph structure mining and the fusion of short-term and long-term user interests consists of seven parts. The first part divides the user's historical behavior sequence into recent historical behavior sequence and long-term historical behavior sequence; the second part uses a recurrent neural network to obtain the user's short-term interests based on the user's recent historical behavior sequence; the third part mines the latent graph structure based on the similarity of items in the user's historical behavior sequence; the fourth part updates the item vectors using a graph neural network based on the mined graph structure; and uses a recurrent neural network to obtain the user's long-term interests based on the updated item vectors; the fifth part fuses the user's short-term and long-term interests based on the diversity of user interests; the sixth part predicts the user's click-through rate for items; and the seventh part designs a loss function based on the model characteristics.

[0057] like Figure 1 As shown, according to one embodiment of the present invention, the method includes the following steps:

[0058] S100: Obtain the user's historical behavior sequence and divide the user's historical behavior sequence into recent historical behavior sequence and long-term historical behavior sequence according to the order in which the user's actions on items are executed.

[0059] The user set in the platform is represented as The set of items is represented as Targeting the target users To obtain its historical behavior sequence any of these items subscript Indicates user Behavioral sequence The length. The vector representation is In this embodiment, the recommendation method can be applied to the short video field; therefore, the corresponding platform is a short video platform, and the item is a short video. Initialized as a vector representation of the main image of a short video. This represents the user's historical behavior sequence. The middle Each item is represented as a sequence of a user's recent historical behaviors. . The previous... Each item is represented as a sequence of a user's long-term historical behavior. It should be noted that the key parameter used here to distinguish between recent and long-term historical behavior sequences is... and The specific value needs to be optimized and determined based on the actual data, with the best final recommendation effect as the standard.

[0060] S200 uses a recurrent neural network to model and obtain users' short-term interests based on their recent historical behavior sequences.

[0061] The recurrent neural network used in this invention is a common method for sequence modeling, typically including long short-term memory neural networks and gated recurrent unit neural networks. In this embodiment, the recurrent neural network uses a long short-term memory (LSTM) neural network to process the user's recent historical behavior sequences. Modeling to obtain users' short-term interests :

[0062]

[0063] in, It is a sequence of recent user behavior. The vector representation of , This represents a Long Short-Term Memory (LSTM) neural network. The output of the last hidden state of the LSM neural network represents the user's short-term interest. .

[0064] S300 mines potential graph structures through a filtering operation consisting of multiple iterations, based on the similarity of items in the user's historical behavior sequence.

[0065] In this embodiment, the specific mining method for the above graph structure is as follows:

[0066] Calculate the cosine similarity between preceding and subsequent items in a user's historical behavior sequence to obtain the initial adjacency matrix of the graph structure. .matrix It is a lower triangular matrix, that is, when the i-th row and j-th column of the matrix is... Subscript satisfies hour, .and Represents items in the user's historical behavior sequence and The similarity, specifically the value of items and vector representation and The cosine similarity. This initial adjacency matrix. The graph is too dense and contains a lot of noise, requiring further filtering to obtain a sparse graph structure. The above filtering operation is a multi-loop process, and its specific loop process is as follows:

[0067] First, let the final graph structure contain the set of nodes as follows: The set to be expanded is , and Initialize to This means adding all items from the user's recent historical behavior sequence to the node set. and middle.

[0068] The second step is to use the initial adjacency matrix. ,get The set of most similar items for each item , will set Put the items in the collection In, that is and update for ,Right now .

[0069] The third step is to repeat step two K times, ultimately obtaining a sparse graph structure and its adjacency matrix. The value of the number of iterations, K, needs to be different depending on the scenario; a larger K allows for the incorporation of more long-term user interest information, but also increases noise. In this method... The graph structure is represented as follows: The corresponding adjacency matrix is ​​represented as The formula represents the adjacency matrix corresponding to each repetition of the second step in the above multiple iterations. The changes are as follows:

[0070]

[0071] in, Represents a set All items , Indicates finding and items Most similar items Then give the adjacency matrix Value assigned .

[0072] S400, based on the mined graph structure, uses a graph neural network to update the item vector, and then uses a long short-term memory neural network to obtain the user's long-term interests based on the updated item vector.

[0073] The essence of a graph neural network is to pass and integrate the information of neighboring nodes into the node in the graph. In this embodiment, the graph neural network is a multi-layered structure with a total of [number missing]. ,pass The information transmission across layers can incorporate user behavior information from longer periods. Here The value is related to K. Nodes in the diagram The Layer item vector representation is Its update formula is:

[0074]

[0075]

[0076] in, Represents the adjacency matrix Regularization is applied to prevent gradient explosion or vanishing during model training. It is a graph structure The degree matrix is ​​calculated using the following formula: The degree matrix is ​​a diagonal matrix. It is an adjacency matrix The regularized matrix, where the matrix element in the i-th row and j-th column is... . express The neighborhood group, express The Layered item vector representation. Nodes in the graph. No. Layer vector Initialize to .go through After the information is passed between layers, the node is finally obtained. The updated item vector is This vector incorporates long-term user preference information. At this point, it includes the user's recent historical behavior sequence. The vector representation is Then, using a long short-term memory neural network, we obtain a result based on... Gaining long-term user interest ,in .

[0077] S500 calculates the richness of user behavior based on the diversity of user interests and preferences, and combines the weights of short-term and long-term interests obtained through learning to fuse the user's short-term and long-term interests to obtain the final user interest representation.

[0078] The extracted long-term user interests incorporate more user behavioral information, resulting in richer user preferences. This method combines the richness of user behavior with the differences between long-term and short-term interests to fuse user interests. The richness of user historical behavior is determined by the richness of information such as categories and tags involved in the user's behavior. The number of categories in the statistical data set is [number missing]. The number of label types is The criterion for defining a user's interest in a category or tag is: the number of actions a user performs under a certain category or tag is greater than [a certain number]. This means that the user is considered interested in the category or tag. In this embodiment, Count the number of categories that users are interested in from their historical behavior sequences. and number of tags The richness of user behavior can be calculated using the following formula. for:

[0079]

[0080] in, and These are hyperparameters that control the importance of category and tag information, respectively. In this embodiment .

[0081] The greater the difference between long-term and short-term interests, the more singular the user's short-term interests are. This embodiment uses a two-layer MLP model to learn the weights of long-term and short-term interests, and combines this with the richness of user behavior. The final long-term and short-term interest weights are obtained as follows:

[0082]

[0083]

[0084] in, and These are the model parameters of a two-layer MLP model, which are updated during training. (Superscript) It is the transpose symbol. yes Activation function yes Activation function. It is the weight of users' long-term interests learned by the model itself, through... The function normalizes it to Within the range, then multiplied by the richness of user behavior. Obtain the final user long-term interest weight .

[0085] Finally, the user's interest is represented as .

[0086] The S600 predicts the click-through rate of items by calculating the similarity between the user's interest representation and the vector representation of the target item, thereby enabling item recommendation.

[0087] According to user User vectors Predicting user preferences for target items Click-through rate ,thing The vector is The method for calculating click-through rate is to take the inner product of the user vector and the item vector to obtain the similarity:

[0088]

[0089] It should be noted that steps S100 to S600 above actually constitute the model framework of the recommended method of the present invention, as follows: Figure 2 As shown. This model framework needs to be trained before it can be used for inference. The training process can be implemented through the following S700.

[0090] S700, based on the model characteristics, a loss function is designed to train the aforementioned model framework. In this embodiment, it can be obtained through the user's click-through rate prediction value for the target item. Calculate the predicted click-through rate. and the true value of click-through rate The error between the predicted click-through rate and the actual click-through rate is used to update the model parameters. and the true value of click-through rate The error can be expressed as cross-entropy, and the cross-entropy loss function guides the update process of the model parameters. The cross-entropy loss function is a current technology, and its formula is as follows:

[0091]

[0092] in, This is the actual value, representing whether the user clicked on the target item. It's the sigmoid function. It should be noted that constructing the training dataset is also a current technique, and will not be described in detail here. During actual training, the model parameters can be updated using the Adam optimizer.

[0093] In summary, this invention proposes a recommendation method for short video recommendation in the short video field based on latent graph structure mining and the fusion of users' long-term and short-term interests. First, the user's historical behavior sequence is divided into a long-term historical behavior sequence and a recent historical behavior sequence according to the time of the user's behavior. Then, a recurrent neural network is used to extract the user's short-term interests from the recent historical behavior sequence. Next, the graph structure of the user's historical behavior sequence is mined. Here, the graph structure is a unidirectional graph structure, meaning there are only edges between short videos viewed earlier and those viewed later. Based on the learned graph structure, a graph neural network is used to update the short video vector representation. The updated short video vectors in the user's recent historical behavior sequence have fused with the user's long-term interest information. Then, a recurrent neural network is used to model the user's recent historical behavior sequence to obtain the user's long-term interests. Finally, considering the user's diverse preference information, the user's long-term and short-term interests are fused, and short videos that the user may be interested in are recommended. This invention can effectively achieve the task of recommending short videos that match user interests.

[0094] The above description of the embodiments is provided to enable those skilled in the art to understand and apply the present invention. It will be apparent to those skilled in the art that various modifications can be made to the above embodiments, and the general principles described herein can be applied to other embodiments without inventive effort. Therefore, the present invention is not limited to the above embodiments, and any improvements and modifications made to the present invention by those skilled in the art based on the disclosure thereof should be within the scope of protection of the present invention.

Claims

1. A recommendation method based on latent graph structure mining and fusion of users' long-term and short-term interests, characterized in that: S100. Obtain the user's historical behavior sequence and divide the user's historical behavior sequence into recent historical behavior sequence and long-term historical behavior sequence according to the order in which the user performs actions on items. S200: Based on the user's recent historical behavior sequence, a recurrent neural network is used to model and obtain the user's short-term interests; S300. Based on the similarity of items in the user's historical behavior sequence, a filtering operation consisting of multiple loops is used to mine potential graph structures. S400. Based on the mined graph structure, update the item vector using a graph neural network, and then obtain the user's long-term interests using a long short-term memory neural network based on the updated item vector; the specific implementation method is as follows: S401. Based on the mined graph structure, update the item vector using a graph neural network; the graph neural network is a multi-level structure, through a total... The information delivery layer incorporates user behavior information from longer periods; among which ; Nodes in the diagram The Layer item vector representation is Its update formula is: ； ； in, Represents the adjacency matrix Perform regularization; It is a graph structure The degree matrix is ​​calculated using the following formula: The degree matrix is ​​a diagonal matrix; It is an adjacency matrix The regularized matrix, where the matrix element in the i-th row and j-th column is... ; express The neighborhood group, express The Layer item vector representation; This represents the item nodes in a graph structure. (The graph nodes are...) No. Layer vector Initialize to ;go through After the information is passed between layers, the node is finally obtained. The updated item vector is ; Indicates the number of times the loop is repeated; S402, Based on the user's recent historical behavior sequence The updated item vector representation is as follows Then, long short-term memory neural networks are used to obtain users' long-term interests. ,in ; S500: Based on the diversity of user interests and preferences, calculate the richness of user behavior, and combine the weights of short-term and long-term interests obtained through learning to fuse the user's short-term and long-term interests to obtain the final user interest representation. S600: By calculating the similarity between the user's interest representation and the vector representation of the target item, the click-through rate of the user on the item is predicted, thereby realizing item recommendation.

2. The recommendation method based on latent graph structure mining and fusion of user long-term and short-term interests as described in claim 1, characterized in that, The specific implementation method of S100 is as follows: Get the user set in the platform and item collection Acquire users Historical Behavioral Sequence any of these items subscript Indicates user Behavioral sequence Length; The vector representation is ; User historical behavior sequence The middle Each item is represented as a sequence of a user's recent historical behaviors. ; The previous user historical behavior sequence Each item is represented as a sequence of a user's long-term historical behavior. .

3. The recommendation method based on latent graph structure mining and fusion of user long-term and short-term interests as described in claim 2, characterized in that, The specific implementation method of S200 is as follows: For users Long Short-Term Memory Neural Network is used to analyze the user's recent historical behavior sequences. Modeling is performed to obtain users' short-term interests. : ； in, It is a sequence of recent user behavior. The vector representation of , This represents a Long Short-Term Memory (LSTM) neural network. The output of the last hidden state of the LSM neural network represents the user's short-term interest. .

4. The recommendation method based on latent graph structure mining and fusion of user long-term and short-term interests as described in claim 3, characterized in that, The specific implementation method of S300 is as follows: Calculate the cosine similarity between preceding and subsequent items in a user's historical behavior sequence to obtain the initial adjacency matrix of the graph structure. ;matrix It is a lower triangular matrix, when the i-th row and j-th column of the matrix... Subscript satisfies hour, ; Represents items in the user's historical behavior sequence and The similarity, specifically the value of the items and vector representation and The cosine similarity; further, the initial adjacency matrix... Performing a filtering operation yields a sparse graph structure. The filtering operation is a multi-cycle process: First, let the final graph structure contain the set of nodes as follows: The set to be expanded is , and Initialize to ; The second step is to use the initial adjacency matrix. ,get The set of most similar items for each item , will set Put the items in the collection In, that is and update for ,Right now ; The third step is to repeat step two K times, ultimately obtaining a sparse graph structure. And the adjacency matrix of the graph ; In each of the multiple iterations described above, the adjacency matrix corresponding to the second step is repeated. The change can be expressed by the formula: ； in, Represents a set All items , Indicates finding and items Most similar items Then give the adjacency matrix Value assigned .

5. The recommendation method based on latent graph structure mining and fusion of user long-term and short-term interests as described in claim 1, characterized in that, The specific implementation method of S500 is as follows: S501, The number of categories in the statistical data centralization system is: The number of label types is ; The number of user actions under a certain category or tag is greater than As a criterion for determining whether a user is interested in a particular category or tag, the number of categories that the user is interested in is counted in the user's historical behavior sequence. and number of tags To obtain the richness of user behavior for: ； in, and These are hyperparameters that control the importance of category and tag information, respectively. ; S502 uses a two-layer MLP model to learn the weights of short-term and long-term interests, and combines this with the richness of user behavior. The final long-term and short-term interest weights are obtained as follows: ； ； in, and These are the model parameters of a two-layer MLP model, which are updated during training. (Superscript) It is the transpose symbol; express Activation function express Activation function; It is the weight of users' long-term interests learned by the model itself, through... The function normalizes it to Within the range, then multiplied by the richness of user behavior. Obtain the final user long-term interest weight ; S503, Based on User Long-Term Interest Weighting and user short-term interest weight long-term user interest and short-term interests By merging these elements, the final representation of user interests can be obtained. .

6. The recommendation method based on latent graph structure mining and fusion of user long-term and short-term interests as described in claim 5, characterized in that, The specific implementation method of S600 is as follows: According to user User interest representation Representing user interests and target items vector representation Perform inner product calculations to predict user preference for the target item. Click-through rate : 。 7. The recommendation method based on latent graph structure mining and fusion of user long-term and short-term interests as described in claim 1, characterized in that, The recommendation model framework consisting of S100~S600 needs to be trained in advance before actual inference. During the training process, through the user's interaction with the target item Click-through rate prediction Calculate the predicted click-through rate. and the true value of click-through rate The cross-entropy loss function is used to guide the update process of model parameters, and the Adam optimizer is used to update the model parameters.

8. The recommendation method based on latent graph structure mining and fusion of user long-term and short-term interests as described in claim 7, characterized in that, The formula for calculating the cross-entropy loss function is: ； in, This is the actual value, representing whether the user clicked on the target item; It is the sigmoid function.

9. The recommendation method based on latent graph structure mining and fusion of user long-term and short-term interests as described in claim 2, characterized in that, The platform is a short video platform, and the items are short videos. The vector representation of the main image of the short video is initialized.

Images