Article recommendation method, model processing method, computer equipment and storage medium

Abstract

The invention discloses an article recommendation method, a model processing method, computer equipment and a storage medium. The article recommendation method comprises the steps of determining user features of a target user, article features of candidate articles and interaction behaviors between the target user and multiple articles; inputting the user features and the article features into a first network, and predicting the interestingness of the target user to the candidate articles; inputting the article features and the interaction behaviors into a second network, and predicting a first probability of interaction between the target user and the candidate articles; and selecting recommended items from the candidate items according to the interestingness and the first probability. According to the method, different recommendation information is obtained through analysis from different dimensions, the recommended item of the target user can be determined from the candidate items more accurately subsequently, and better robustness is achieved in the face of influences of data sparsity and information noise, so that the recommendation effect is improved.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to an item recommendation method and a model processing method, a computer device and a storage medium (computer-readable storage medium). Background Art

[0002] A recommendation system is an information filtering system that aims to automatically filter and recommend items that users may be interested in based on their personalized needs and interests.

[0003] At present, most recommendation systems build user portraits and item features through communication, and use collaborative filtering and other methods to find similarities between users or items, so as to recommend items to users. However, this method is affected by data sparsity and information noise, especially when new users or new items appear, and lacks relevant knowledge, resulting in low recommendation accuracy. Summary of the invention

[0004] In response to the above technical problems, the embodiments of the present application provide an item recommendation method and a model processing method, a computer device and a storage medium to solve the technical problem of low recommendation accuracy in related technologies.

[0005] In a first aspect, an embodiment of the present application provides an item recommendation method, comprising:

[0006] Determining user characteristics of a target user, item characteristics of candidate items, and interaction behaviors between the target user and multiple items on a network platform;

[0007] Inputting the user features and the item features into a first network of a first model to predict the interest of the target user in the candidate item;

[0008] Inputting the item features and the interaction behavior into the second network of the first model to predict a first probability of the target user interacting with the candidate item;

[0009] A recommended item is selected from the candidate items to be recommended to the target user according to the interest degree and the first probability.

[0010] In a second aspect, an embodiment of the present application provides a model processing method, the method comprising:

[0011] Determining sample user characteristics of sample users, sample item characteristics of sample items, and sample interaction behaviors between the sample target users and a plurality of items;

[0012] Inputting the sample user features and the sample item features into a first initial network of an initial model to obtain the predicted interest of the sample user in the sample item;

[0013] Inputting the sample item features and the sample interaction behavior into a second initial network of the initial model to obtain a predicted probability of the sample user interacting with the sample item;

[0014] The initial model is updated according to the predicted interest and the predicted probability.

[0015] In a third aspect, an embodiment of the present application further provides an item recommendation device, the device comprising:

[0016] A determination module, used to determine user characteristics of a target user, item characteristics of candidate items, and interaction behaviors between the target user and multiple items on a network platform;

[0017] A prediction module, configured to input the user features and the item features into a first network of a first model to predict the interest of the target user in the candidate item;

[0018] The prediction module is further used to input the item features and the interaction behavior into the second network of the first model to predict a first probability of the target user interacting with the candidate item;

[0019] A recommendation module is used to select a recommended item to be recommended to the target user from the candidate items according to the interest level and the first probability.

[0020] In a fourth aspect, an embodiment of the present application further provides a model processing device, the device comprising:

[0021] An acquisition module, used to determine sample user characteristics of sample users, sample item characteristics of sample items, and sample interaction behaviors between the sample target users and multiple items;

[0022] A processing module, used for inputting the sample user characteristics and the sample item characteristics into a first initial network of an initial model to obtain the predicted interest of the sample user in the sample item;

[0023] The processing module is further used to input the sample item features and the sample interaction behavior into a second initial network of the initial model to obtain a predicted probability of the sample user interacting with the sample item;

[0024] A training module is used to update the initial model according to the predicted interest and the predicted probability.

[0025] In a fifth aspect, an embodiment of the present application also provides a computer device, comprising a memory storing a plurality of instructions; a processor loads instructions from the memory to execute the steps of the item recommendation method as described above or execute the steps of the item recommendation method as described above.

[0026] In a fourth aspect, an embodiment of the present application further provides a computer-readable storage medium, which stores a plurality of instructions, and the instructions are suitable for a processor to load to execute the steps of the item recommendation method as described above or to execute the steps of the item recommendation method as described above.

[0027] The scheme of the application embodiment is adopted. After determining the user characteristics, item characteristics and behavior queue, the user characteristics and item characteristics are first analyzed based on the first network in the trained first model to determine the correlation between different candidate items for the target user from the feature perspective, and then the item characteristics and user behavior are analyzed by the second network in the first model to determine the possible interactive behavior of the target user when facing different candidate items from the interactive behavior perspective, so that the target user can more accurately determine the recommended items from the candidate items by subsequently using the probability of interest interaction predicted from different dimensions. At the same time, different input information is processed by different network layers integrated in the first model, and different network layers can also share information during the training process. When facing the influence of data sparsity and information noise, it has better robustness, thereby improving the recommendation effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0029] Figure 1 A schematic diagram of the steps of the item recommendation method provided in an embodiment of the present application;

[0030] Figure 2 A schematic diagram of a process flow for determining user characteristics and item characteristics provided in an embodiment of the present application;

[0031] Figure 3 A schematic diagram of a process flow of selecting recommended items from candidate items based on different weights provided in an embodiment of the present application;

[0032] Figure 4 A schematic diagram of a step flow chart for training a first model provided in an embodiment of the present application;

[0033] Figure 5A schematic diagram of a process flow for calculating a model loss value provided in an embodiment of the present application;

[0034] Figure 6 A schematic diagram of a process flow for calculating loss values ​​by assigning different weights provided in an embodiment of the present application;

[0035] Figure 7 A schematic diagram of the structure of an item recommendation device provided in an embodiment of the present application;

[0036] Figure 8 A schematic diagram of the structure of a model processing device provided in an embodiment of the present application;

[0037] Fig. 9 A schematic diagram of the structure of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0038] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present application. At the same time, in the description of the embodiments of the present application, the terms "first", "second", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance. Thus, the features defined as "first" and "second" may explicitly or implicitly include one or more features. In the description of the embodiments of the present application, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined.

[0039] In order to better understand the item recommendation method and model processing method, computer equipment and storage medium provided by the present application, the application environment applicable to the present application is described below. In item recommendation, most of the recommendation systems in the related art recommend items to users by constructing user portraits and item features and based on collaborative filtering. For example, the similarity between user portraits can be used to recommend items purchased by other users with similar portraits to the target user, or other items similar to the items purchased by the target user in the past can be recommended to the target user by utilizing the similarity between item attributes. However, this method is easily affected by data sparsity and information noise. For example, when a new user appears, it is impossible to recommend other users with similar portraits to the user because the user's portrait cannot be constructed in time, or when a new item appears, it is difficult to recommend the item to the appropriate user because the item's attributes cannot be obtained in time and the similarity relationship between the item and other items cannot be established.

[0040] In order to solve the above problems, the present application provides an item recommendation method, which uses different network layers integrated in the model to analyze the possible implicit relationship between users and items from different dimensions, thereby reducing the impact of data sparsity and information noise, and can recommend items to users more accurately and effectively. Specifically, the item recommendation method provided in the present application is usually run in the item recommendation device in the form of a computer program, and the item recommendation device is usually set in the computer device in the form of a processor. The item recommendation device in the computer device executes the item recommendation method provided in the embodiment of the present application by running the computer program corresponding to the item recommendation method.

[0041] For details, please refer to Figure 1 , Figure 1 This is a schematic diagram of the steps of the item recommendation method provided in the embodiment of the present application, which specifically includes steps S110 to S140:

[0042] S110, determining user characteristics of a target user, item characteristics of candidate items, and interaction behaviors between the target user and multiple items on a network platform.

[0043] In the embodiment of the present application, the target user can be understood as the user to whom an item needs to be recommended, and the candidate item can be understood as the commodity that needs to be recommended further. Specifically, in different application scenarios, the target user refers to different user groups, and the corresponding candidate items are usually different. For example, in an e-commerce shopping platform, the target user usually refers to a user group with potential shopping intentions, and the corresponding candidate items usually refer to the commodities sold by the e-commerce shopping platform, or, in a multimedia resource platform, the target user usually refers to a user group that wants to retrieve and browse related multimedia resources, such as image resources, text resources, or video resources, and the corresponding candidate items usually refer to the multimedia resources uploaded in the media resource platform. Of course, in actual applications, the candidate items can also be based on certain initial screening conditions and are selected from all the items owned by the platform and are more suitable for recommending to users. The embodiment of the present application does not limit the target user and the candidate items.

[0044] Furthermore, user characteristics can usually be simply understood as user portraits, that is, feature information commonly used to describe user preferences generated by relevant platforms using some historical interaction information of users on the platform after obtaining user authorization in order to better provide services to users. For example, in an e-commerce shopping platform, historical interaction information includes the frequency of users adding items to shopping carts, the frequency of purchasing items, the frequency of returning or exchanging items, or the evaluation of items, etc. In a multimedia resource platform, historical interaction information includes the frequency and duration of users viewing individual multimedia resources, as well as the comments they made on the multimedia resources, etc.

[0045] Correspondingly, item features can usually be understood as the type of labels created by the relevant platforms using some attribute information of items in order to better push items. For example, in e-commerce shopping platforms, item features usually include the type of item, applicable population, price range, etc., while in multimedia resource platforms, item features usually include the style of the item, emotional value, etc. Of course, it should be noted that different platforms have different ways of constructing user features and item features, but in any case, the user features read from the platform can usually simply describe user preferences, and item features can also simply describe the attributes of the item.

[0046] In addition, in addition to determining user characteristics and item characteristics, in order to ensure the item recommendation effect provided by the embodiment of the present application, in the embodiment of the present application, the interaction behavior between the target user and multiple items on the network platform is further determined, wherein the interaction behavior can generally be understood as a behavior sequence of the latest several interaction behaviors performed by the target user on various types of goods collected from the network platform based on network technology means, such as crawler technology, that is, the interaction behavior between the target user and multiple items can generally exist in the form of a queue. For example, the queue may include likes for item A, browsing of item B, comments on item C, purchases of item D, and returns of item E, etc., that is, the queue generally also includes several related items that have interactive behaviors with the user. Of course, these related items are generally other items that are different from the candidate items.

[0047] Of course, in order to further improve the effect of item recommendation, in another feasible embodiment of the present application, the user features and item features that simply describe user preferences or item attributes mentioned above can be further enhanced by knowledge fusion. For details, please refer to Figure 2 , Figure 2 A schematic flow chart of a step of determining user characteristics and item characteristics provided in an embodiment of the present application, specifically, comprising steps S210 to S250:

[0048] S210, obtaining initial user features of a target user, initial item features of candidate items, and interaction behaviors between the target user and multiple items on the network platform.

[0049] In the embodiment of the present application, the initial user characteristics of the target user are the characteristic information commonly used to describe the user's preferences, that is, the user portrait, generated by the aforementioned platform using some historical interaction information of the user on the platform, and the initial item characteristics are the label types created by the platform using some attribute information of the item. In addition, the interaction behavior between the target user and multiple items on the network platform can refer to the explanation of the aforementioned step S110, which will not be elaborated in the embodiment of the present application.

[0050] S220: Input the interaction behavior into a language model to obtain a user description text of the target user.

[0051] In order to enhance user features, the embodiment of the present application provides an implementation scheme for enriching user feature expressions in combination with a language model, wherein the language model here usually refers to a large language model that has been trained with certain text response capabilities, such as the common GPT (Generative Pre-Trained) language model or GLM (Glancing Language Model). Of course, it is also feasible to use other trained large language models to realize user interest reasoning. The embodiment of the present application does not go into details about the types of language models.

[0052] Specifically, inputting the interaction behavior into the language model usually refers to constructing a specific prompt question template, and inputting it into the language model in combination with the interaction behavior between the target user and different items, so that the trained language model can infer the user's interests and preferences. For example, for ease of understanding, the prompt question template can be as follows: "The following is the user's latest viewing record [behavior sequence], please analyze what type of movie the user likes?", "The following is the user's purchased goods [behavior sequence], please analyze what kind of goods the user needs?", "The following is the user's latest behavior activity [behavior sequence], please analyze what kind of activities the user likes, etc.?", and after the above information is input into the language model, the language model can analyze and output user description text related to the target user, such as: "The user likes comedy and science fiction movies, and is not very interested in action movies" and so on.

[0053] S230, generating a target question sentence associated with the candidate item according to a preset question template, and inputting the target question sentence into a language model to obtain an item description text of the candidate item.

[0054] In order to enhance the features of the item, corresponding to the above, the embodiment of the present application also provides an implementation scheme for enriching the expression of item features by combining a language model. For example, by constructing a specific prompt question template, it is possible to combine some prior knowledge outside the item. For example, the common external knowledge describing the attributes of the item in the form of a knowledge graph is input into the language model, so that the trained language model can infer the attributes of the item and output it in the form of a descriptive text to obtain the item description text of the candidate item, such as "The director of the movie is Zhang San, and it is a science fiction film released in XXXX."

[0055] S240: Enhance the initial user features according to the user description text to obtain user features of the target user.

[0056] In an embodiment of the present application, on the basis of the foregoing, after obtaining the user description text by using the language model to respond to the prompt template of the fused user behavior sequence, an additional text representation vector can be obtained by further text encoding the user description text, and this text representation vector is fused with the initial user features pre-constructed by the platform, such as the user portrait, for example, the common feature interactive fusion of the text representation vector and the initial user features, which can ultimately obtain more information-rich, that is, enhanced user features, which can further improve the effect of subsequent item recommendations.

[0057] S250: Enhance the initial item features according to the item description text to obtain item features of the candidate item.

[0058] In the embodiment of the present application, similarly, on the basis of the foregoing, after constructing a specific prompt question template and using a language model to fuse external knowledge information to generate an item description text, additional text representation vectors can be obtained by further text encoding these item description texts. This text representation vector is fused with the initial item features such as item labels pre-constructed by the platform, such as feature interaction fusion, and more information-rich, enhanced item features can also be obtained, thereby further improving the effect of subsequent item recommendations.

[0059] The embodiment of the present application, by utilizing a language model to fuse user behavior sequences and external knowledge of items to enrich the user's feature expression and the item's feature expression, can enhance the model's understanding of user preferences and item knowledge, thereby providing users with more personalized recommendations in the subsequent recommendation process, further improving the accuracy of item recommendations.

[0060] S120, inputting the user features and the item features into a first network of a first model to predict the interest of the target user in the candidate item.

[0061] In the embodiment of the present application, the interest level can be understood as an evaluation index of the target user's interest in the candidate item. For example, more commonly, the interest level can be in the form of a score. The higher the score, the higher the possibility that the target user is interested in the candidate item. For example, taking the score between 0 and 1 as an example, when the target user's interest level in the candidate item is closer to 1, it indicates that the user is more likely to be interested in the candidate item. On the contrary, when the target user's interest level in the candidate item is close to 0, it indicates that the user is not likely to be interested in the candidate item.

[0062] Of course, it is understandable that the implementation of the above process usually needs to rely on the training of the model, that is, under normal circumstances, it is necessary to use training sample data to train the first network in the first model that can analyze the interest between the target user and each candidate item from the dimension of feature information. Specifically, in the embodiment of the present application, the first network is usually integrated into the first model in an integrated manner, that is, in the training process, it is usually necessary to synchronize the training with other network layers in the first model. For the specific training process, please refer to the relevant description of the subsequent embodiments.

[0063] S130: Input the item features and the interaction behavior into the second network of the first model to predict a first probability of the target user interacting with the candidate item.

[0064] In the embodiment of the present application, corresponding to the aforementioned step S120, in the first model, in addition to the first network that can determine the interest of the target user in each candidate item from the dimension of feature information, a second network that can determine whether the target user is likely to select a candidate item for interaction from the dimension of user interaction is generally included. For example, the item features and interaction behaviors can be input into the second network of the first model so that the model outputs a first probability to describe the possibility of the user interacting with different candidate items, such as purchasing, returning a candidate item, or browsing a candidate video. The first probability can be a set of interaction probabilities with values ​​between 0 and 1, wherein the closer the interaction probability is to 1, the higher the possibility that the user will click (or interact) with the candidate item next, and the more suitable it is to recommend the item to the user.

[0065] It is understandable that there is a certain correlation between the user's interest in an item and the user's interaction with the candidate item, but it is not a direct positive correlation. In other words, the candidate item with which the user interacts is not necessarily a high-interest item, that is, the interest and the first probability refer to the analysis of the user and the item from different dimensions. Therefore, the present application integrates the second network and the first network in the first model, so that the first network and the second network can share some parameters with each other, so as to better cope with the influence of data sparsity and information noise, and improve the recommendation effect of the first model for new users or new items.

[0066] Of course, in addition to the first network and the second network provided by the aforementioned first model, in order to further improve the effect of recommending items, as another feasible embodiment of the present application, the first model further includes a third network that can analyze user characteristics, item characteristics, and behavior queues as a whole, thereby predicting the target user's next complete interaction behavior with the candidate items. That is, the item recommendation method also includes:

[0067] The user features, the item features and the interactive behavior queue are input into the third network preset model of the preset integrated first model to obtain a predicted behavior queue sequence corresponding to the target user.

[0068] In the embodiment of the present application, the predicted behavior sequence generally includes the predicted interactive behaviors between the target user and multiple items in the candidate items. For example, the predicted user's next behavior is to purchase candidate item A, purchase candidate item B, or return candidate item A.

[0069] Of course, it should be noted that, similar to the aforementioned first network and second network, the training of the third network in the third model also depends on the overall training of the first model, and the specific training process will be further explained later.

[0070] S140: Select a recommended item to be recommended to the target user from the candidate items according to the interest level and the first probability.

[0071] In the embodiment of the present application, after the first network and the second network of the first model respectively perform dimensional analysis and obtain the interest degree and the first probability of interaction of the target user corresponding to different candidate items, by further integrating the information of the two, the recommended items that are more suitable for recommendation to the target user can be further selected from the candidate items.

[0072] Specifically, selecting a recommended item that is more suitable for recommending to a target user from candidate items can be based on a single piece of information in the first probability of interest or interaction. Of course, considering that when training the first network and the second network, the two networks are integrated in the meaning model and trained uniformly, usually, during the training process, the importance of different networks can be assigned by a specified weighting coefficient. Therefore, similarly, in the embodiment of the present application, selecting a recommended item that is more suitable for recommending to a target user from candidate items based on the first probability of interest and interaction can also be based on the weighting coefficient that describes the importance of different networks to achieve weighted fusion of the first probability of interest and interaction. For example, refer to Figure 3 , Figure 3 A schematic flow chart of a step of screening recommended items from candidate items based on different weights provided in an embodiment of the present application, specifically, comprising steps S310 to S330:

[0073] S310, determining weight coefficients corresponding to the first network and the second network in the first model.

[0074] In the embodiment of the present application, in combination with the above-mentioned related description, it can be known that in the integrated training process of the first network and the second network, there is usually a given weighting coefficient to assign the importance of different networks. For example, as a common implementation scheme, the importance of the model can be determined by assigning different weights to the loss values ​​calculated by different initial networks during the training and updating of the first model. Therefore, in the embodiment of the present application, the first weight corresponding to the first network in the first model and the second weight corresponding to the second network are the weight coefficient set for the loss value calculated by the first network during training and the weight coefficient set for the loss value calculated by the second network during training during the training of the first model.

[0075] Of course, the first weight and the second weight may also be other preset parameters based on the first model that can describe the importance of the first network and the second network during training, or the first weight and the second weight may also be fixed preset values, which is not limited in the embodiments of the present application.

[0076] S320: Determine a recommendation score of the candidate item to the target user according to the weight coefficient, the interest level and the first probability.

[0077] In the embodiment of the present application, by using the first weight and the second weight that describe the importance of different networks as weighting coefficients, the interest and the first probability of interaction are weighted respectively, and different dimensions can be finally integrated to determine the recommendation scores of different candidate items for the target user. Of course, it should be noted that when the interest and the first probability are weighted by the weight coefficient, it is necessary to control the normalization of the interest and the first probability of interaction, that is, it is necessary to map the obtained interest and the first probability of interaction to the same value space of 0 to 1, so as to achieve weighted fusion of the interest and the first probability and obtain a more accurate recommendation score.

[0078] S330: Select a recommended item to be recommended to the target user from the candidate items based on the recommendation score.

[0079] In the embodiment of the present application, by further fusing the interest degree and the first probability to obtain a recommendation score, the recommended items recommended to the target user can be selected from the candidate items based on the recommendation score. Among them, the specific recommendation rules can be set based on actual needs. For example, as a common feasible implementation scheme, several items with higher recommendation scores can be selected from the candidate items based on the size relationship of the recommendation scores as the recommended items selected and recommended to the target user. Alternatively, as another feasible implementation scheme, it can also be based on the size relationship between the recommendation score and the preset score threshold, so that the items with scores higher than the preset score threshold are selected from the candidate items as the recommended items selected and recommended to the target user. Of course, in addition to the basis of the implementation scheme provided above, those skilled in the art can also set other implementation schemes based on the recommendation score to select recommended items from the candidate items and recommend them to the target user based on actual needs.

[0080] The solution provided in the embodiment of the present application can make the final recommended items more in line with user needs, thereby further effectively improving the recommendation effect of the items by weighting the interest and the first probability of interaction by using the weight coefficients describing the importance assigned to different networks when training the first model.

[0081] The scheme of the application embodiment is adopted. After determining the user characteristics, item characteristics and behavior queue, the user characteristics and item characteristics are first analyzed based on the first network in the trained first model to determine the correlation between different candidate items for the target user from the feature perspective, and then the item characteristics and user behavior are analyzed by the second network in the first model to determine the possible interactive behavior of the target user when facing different candidate items from the interactive behavior perspective, so that the target user can more accurately determine the recommended items from the candidate items by subsequently using the probability of interest interaction predicted from different dimensions. At the same time, different input information is processed by different network layers integrated in the first model, and different network layers can also share information during the training process. When facing the influence of data sparsity and information noise, it has better robustness, thereby improving the recommendation effect.

[0082] Of course, it can be understood that in the aforementioned process of implementing item recommendation, it depends on the training of the first model, wherein the first model includes multiple networks, such as the first network and the second network provided above. Therefore, as another feasible embodiment of the present application, a training process for the first model is further provided, wherein the first network and the second network in the first model can generally be based on the model idea of ​​multi-task learning, and are trained synchronously using the same batch of training data. For details, please refer to Figure 4 , Figure 4 A schematic flow chart of a step of training a first model provided in an embodiment of the present application, specifically, comprising steps S410 to S440:

[0083] S410, determining sample user characteristics of a sample user, sample item characteristics of a sample item, and sample interaction behaviors between the sample target user and a plurality of items.

[0084] In the embodiments of the present application, sample users and sample items generally refer to sample data collected from an open platform through network technologies such as crawlers or extracted from a dedicated model training database for model training, and the corresponding sample user characteristics and sample item characteristics are similar to the user characteristics of the target user and the item characteristics of the candidate items mentioned above, so the user portrait of the sample user and the sample label information of the sample item can be simply understood.

[0085] S420, inputting the sample user characteristics and the sample item characteristics into a first initial network of an initial model to obtain the predicted interest of the sample user in the sample item.

[0086] In order to realize the training of the first model, in the embodiments of the present application, a specific network model structure is usually adopted and the parameters are randomized to generate an initial model, wherein the initial model usually includes a first initial network and a second initial network, so that the first initial network and the second initial network can be trained separately through the loss value later, and they are respectively used as the first network and the second network in the first model obtained by the final training.

[0087] Of course, after inputting the sample user characteristics and sample item characteristics into the first initial network of the initial model, the predicted interest of the sample user corresponding to the sample item can also be predicted. Of course, since the parameters in the first initial network are randomly initialized, the first initial network often cannot accurately predict the predicted interest of the sample user corresponding to the sample item, that is, the first initial network needs to be trained through the interest labels of the sample user corresponding to the sample item.

[0088] S430, inputting the sample item features and the sample interaction behavior into a second initial network of the initial model to obtain a predicted probability of the sample user interacting with the sample item.

[0089] In the embodiment of the present application, similar to the above-mentioned prediction of the predicted interest of the sample user corresponding to the sample item through the first initial network, the sample item characteristics and the sample interaction behavior can be input into the second initial network of the initial model to predict the predicted probability of the sample user interacting with the sample item. Of course, since the parameters in the second initial network are randomly initialized, the second initial network often cannot accurately predict the predicted probability of the sample user interacting with the sample item, that is, it is necessary to train the second initial network through the interest labels of the sample user corresponding to the sample item.

[0090] S440: Update the initial model according to the predicted interest level and the predicted probability.

[0091] In an embodiment of the present application, after the predicted interest and predicted interaction probability of the sample user for the sample item are respectively predicted by the first initial network and the second initial network, by using the above information to construct the loss value with the true label, the initial model can be updated based on the idea of ​​back propagation, so that the first initial network and the second initial network in the initial model can improve the prediction accuracy of the sample user and the sample item, until the final first model is trained.

[0092] Among them, it can be understood that updating the initial model based on the predicted interest and predicted probability often requires multiple iterations of training. That is to say, as a feasible implementation scheme, after updating the initial model based on the predicted interest and predicted probability and obtaining the updated model, by inputting the sample users and sample items into the updated model again, and obtaining the updated predicted interest and the updated predicted probability through the updated first initial network and the second initial network, the loss value of the updated model can be evaluated again, and the updated model can be iteratively updated again based on the idea of ​​back propagation, until the predicted interest and predicted probability obtained by inputting the sample users and sample items into the updated model at a certain time meet the preset requirements, for example, after several consecutive update rounds and the labels between the sample users and the sample items meet the preset requirements, the currently trained model can be used as the final trained target model.

[0093] Of course, to facilitate understanding of the above scheme, the following will specifically illustrate the update process of the model using a round of update process during iterative training, as detailed below.

[0094] See also Figure 5 , Figure 5 A schematic flow chart of a step of constructing a loss value to train an initial model provided in an embodiment of the present application, specifically, comprising steps S510 to S540:

[0095] S510: Determine a recommended tag between the sample user and the sample item.

[0096] In the embodiment of the present application, the recommendation label between the sample user and the sample item can be understood as the interest label between the sample user and the sample item. Specifically, the recommendation label between the sample user and the sample item includes recommendation and non-recommendation, which correspond to the information of interest and non-interest, respectively.

[0097] S520: Determine a first loss value corresponding to the first initial network according to the predicted interest level and a first preset value corresponding to the recommended tag.

[0098] In the embodiment of the present application, combined with the above-mentioned related description, it can be known that the recommendation label usually includes recommendation and non-recommendation. Therefore, in order to facilitate the comparison with the predicted interest, the recommendation label is usually mapped to a corresponding numerical value. For example, as a common feasible implementation scheme, when the recommendation label is recommended, its corresponding first preset value can be set to 1, and when the recommendation label is not recommended, its corresponding first preset value can be set to 0, and the predicted interest is usually a value between 0 and 1. Therefore, the difference between the predicted interest and the first preset value corresponding to the recommendation label, for example, cross entropy, can be given to determine the first loss value corresponding to the first initial network. For example, when the recommendation label is recommendation, that is, its corresponding first preset value is 1, at this time, the predicted interest between the sample user and the sample item predicted by the first initial network is approximately close to 1, indicating that the first initial network has a more accurate prediction effect. At this time, the first loss value can be set to be smaller. On the contrary, the predicted interest between the sample user and the sample item predicted by the first initial network is approximately close to 0, which indicates that the prediction effect of the first initial network on the interest between the sample user and the sample item is not ideal. At this time, it is necessary to determine a larger first loss value to update the parameters in the first initial network.

[0099] S530: Determine a second loss value corresponding to the second initial network according to the predicted probability and a second preset value corresponding to the sample interaction behavior.

[0100] In the embodiment of the present application, similar to the above, the predicted probability can be understood as the possibility that the sample user and the sample item predicted by the second initial network will have interactive behavior, and obviously, the sample interactive behavior between the sample target user and multiple items can be used as the real interactive behavior label between the sample user and multiple items. For example, when the sample target user has interactive behavior with a candidate item, the corresponding second preset value can be set to 1, and conversely, when the sample target user does not have interactive behavior with a candidate item, the corresponding second preset value can be set to 0. At this time, the effect of the second initial network predicting the probability of interactive behavior between the sample user and the sample item can be determined based on the predicted probability and the second preset value corresponding to the sample interactive behavior. For example, when there is an interactive behavior between the sample user and the sample item, that is, when the second preset value is 1, if the probability of the interactive behavior between the sample user and the sample item predicted by the second initial network is closer to 1, it indicates that the second initial network has a more accurate prediction effect, and a smaller second loss value can be determined. Conversely, if the probability of the interactive behavior between the sample user and the sample item predicted by the second initial network is closer to 0, it indicates that the prediction effect of the second initial network on the interactive behavior between the sample user and the sample item is not ideal. At this time, a larger second loss value needs to be determined to update the parameters in the second initial network.

[0101] S540: Update the initial model according to the first loss value and the second loss value until a target model is obtained.

[0102] In an embodiment of the present application, based on the first loss value of the first initial network and the second loss value of the second initial network determined in the aforementioned manner, based on the idea of ​​back propagation, and using the first loss value and the second loss value to determine the model loss value of the initial model, the initial model can be updated until the conditions are met, and a target model that can accurately predict the interest and interaction behavior probability between the user and the item is obtained. The target model can be regarded as the first model mentioned above, which is used to realize item recommendations in subsequent item recommendation scenarios.

[0103] Specifically, in the process of updating the initial model according to the first loss value and the second loss value, the first initial network and the second initial network are considered to be integrated into the initial model to predict the association between users and items from different dimensions, so as to make recommendations. Different dimensions have different influences on item recommendations. Therefore, as another feasible embodiment of the present application, the first loss value and the second loss value can be further weighted by different weight coefficients, so that in the process of updating the initial model, one dimension can be given priority, that is, the prediction result of a certain initial network can be considered. For details, please refer to Figure 6 , Figure 6 A schematic flow chart of a step of training a model based on a loss value provided in an embodiment of the present application, specifically, comprising steps S610 to S640:

[0104] S610: Determine the update times of the initial model.

[0105] In the embodiment of the present application, combined with the above-mentioned related description, it can be known that in the process of updating the initial model, it is usually necessary to rely on the iterative update of the initial model. Therefore, the number of updates that have been performed on the initial model can be determined by a counter.

[0106] S620: Determine a first weight coefficient corresponding to the first initial network and a second weight coefficient corresponding to the second initial network according to the number of updates.

[0107] In the embodiment of the present application, at different update times, that is, at different stages of the update, different weight coefficients, such as the first weight coefficient and the second weight coefficient, can be set for the first initial network and the second initial network respectively, so that at different update stages of the model, the focus is on updating the first initial network or the second initial network from different dimensions.

[0108] S630: Determine a model loss value of the initial model according to the first weight coefficient, the second weight coefficient, the first loss value, and the second loss value.

[0109] In an embodiment of the present application, after determining the first weight coefficient and the second weight coefficient corresponding to the number of updates, the corresponding loss value is weighted according to the first weight coefficient and the second weight coefficient, and the model loss value calculated by integrating the emphasis of different dimensions can be obtained.

[0110] S640: Update the initial model according to the model loss value.

[0111] In the embodiment of the present application, based on the model loss value obtained by the above calculation, the initial model is iteratively updated using the idea of ​​back propagation, and training can be performed until the final target model is obtained.

[0112] In addition, it should be noted that the above only uses the first network and the second network to integrate and train the first model. In fact, combined with the above related description, it can be seen that the first model can also include a third network that can realize the overall analysis of user characteristics, item characteristics and interactive behaviors, so as to predict the next complete behavior queue of the target user. Therefore, as another feasible implementation scheme of the present application, the multi-task learning process of the first initial network and the second initial network also includes the joint training of the third initial network, that is, the initial model also includes the third initial network. At this time, the step of training the first initial model also includes:

[0113] The sample user features, the sample item features and the sample interaction behavior queue are input into the third initial network to obtain a sample predicted behavior sequence.

[0114] Among them, similar to the first initial network and the second initial network, the third initial network can realize the overall analysis of user characteristics, item characteristics and interactive behaviors. At this time, in the step of determining the model loss value of the first initial model according to the predicted interest and the predicted probability, the output result of the third initial network also needs to be considered, that is, the updating of the initial model according to the predicted interest and the predicted probability includes:

[0115] The initial model is updated according to the predicted interest, the predicted probability and a third loss value corresponding to the sample predicted behavior sequence.

[0116] Of course, in order to improve the training effect of the third initial network, the idea of ​​generative adversarial thinking can also be integrated in the process of training the third initial network, that is, a virtual predicted behavior sequence is generated by the generator and input into the discriminator for judgment. At this time, the model loss value of the final first initial model is the fusion of the loss value between the first preset value corresponding to the predicted interest and the recommended label mentioned above, the second loss value between the predicted probability and the second preset value of the sample interaction behavior, and the third loss value of the virtual predicted behavior sequence incorrectly identified by the discriminator. The training of the third initial network is realized by combining the idea of ​​generative adversarial thinking, which can enable the trained third network to more accurately predict more realistic user behavior sequences, thereby further improving the effect of subsequent models for item recommendation.

[0117] It should be understood that, although each step in the flowcharts involved in the above-mentioned embodiments is displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0118] Based on the same inventive concept, the embodiment of the present application also provides an item recommendation device for implementing the item recommendation method involved above. The implementation solution provided by the device to solve the problem is similar to the implementation solution recorded in the above method, so the specific limitations in the embodiment of the item recommendation device provided below can refer to the limitations of the item recommendation device method above, and the specific limitations will not be repeated here.

[0119] This embodiment also provides an item recommendation device, which can be integrated into a terminal device. Figure 7 As shown, the item recommendation device may include:

[0120] A determination module 710 is used to determine user characteristics of a target user, item characteristics of candidate items, and interaction behaviors between the target user and multiple items on a network platform;

[0121] A prediction module 720, configured to input the user features and the item features into a first network of a first model to predict the interest of the target user in the candidate item;

[0122] The prediction module 720 is further configured to input the item features and the interaction behavior into the second network of the first model to predict a first probability of the target user interacting with the candidate item;

[0123] The recommendation module 730 is configured to select a recommended item to be recommended to the target user from the candidate items according to the interest level and the first probability.

[0124] In some embodiments, the recommendation module is also used to determine the weight coefficients corresponding to the first network and the second network in the first model; determine the recommendation score of the candidate item to the target user based on the weight coefficients, the interest level and the first probability; and select a recommended item from the candidate items to recommend to the target user based on the recommendation score.

[0125] In some embodiments, the prediction module is also used to input the user characteristics, the item characteristics and the interaction behavior into the third network of the first model to obtain a predicted behavior sequence; the predicted behavior sequence includes predicted interaction behaviors between the target user and multiple items in the candidate items; the recommendation module is also used to determine the target item corresponding to the interaction behavior from the candidate items based on the predicted interaction behaviors contained in the predicted behavior sequence; and select recommended items from the target items to be recommended to the target user based on the interest level and the first probability.

[0126] In some embodiments, the determination module is also used to obtain the initial user characteristics of the target user, the initial item characteristics of the candidate items, and the interaction behavior between the target user and multiple items on the network platform; input the interaction behavior into the language model to obtain the user description text of the target user; generate a target question associated with the candidate item according to a preset question template, and input the target question into the language model to obtain the item description text of the candidate item; enhance the initial user characteristics according to the user description text to obtain the user characteristics of the target user; enhance the initial item characteristics according to the item description text to obtain the item characteristics of the candidate item.

[0127] The scheme of the application embodiment is adopted. After determining the user characteristics, item characteristics and behavior queue, the user characteristics and item characteristics are first analyzed based on the first network in the trained first model to determine the correlation between different candidate items for the target user from the feature perspective, and then the item characteristics and user behavior are analyzed by the second network in the first model to determine the possible interactive behavior of the target user when facing different candidate items from the interactive behavior perspective, so that the target user can more accurately determine the recommended items from the candidate items by subsequently using the probability of interest interaction predicted from different dimensions. At the same time, different input information is processed by different network layers integrated in the first model, and different network layers can also share information during the training process. When facing the influence of data sparsity and information noise, it has better robustness, thereby improving the recommendation effect.

[0128] Based on the same inventive concept, the embodiment of the present application also provides a model processing device for implementing the model processing method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in the embodiment of the model processing device provided below can refer to the limitations of the model processing device method above, and the specific limitations will not be repeated here.

[0129] This embodiment also provides a model processing device, which can be integrated into a terminal device. Figure 8 As shown, the model processing device may include:

[0130] An acquisition module 810 is used to determine sample user characteristics of sample users, sample item characteristics of sample items, and sample interaction behaviors between the sample target users and multiple items;

[0131] A processing module 820 is used to input the sample user characteristics and the sample item characteristics into a first initial network of an initial model to obtain the predicted interest of the sample user in the sample item;

[0132] The processing module 820 is further configured to input the sample item features and the sample interaction behavior into a second initial network of the initial model to obtain a predicted probability of the sample user interacting with the sample item;

[0133] The training module 830 is used to update the initial model according to the predicted interest and the predicted probability.

[0134] In some embodiments, the training module is also used to determine the recommended tags between the sample users and the sample items; determine a first loss value corresponding to the first initial network based on the predicted interest and a first preset value corresponding to the recommended tags; determine a second loss value corresponding to the second initial network based on the predicted probability and a second preset value corresponding to the sample interaction behavior; and update the initial model based on the first loss value and the second loss value until the target model is obtained.

[0135] In some embodiments, the training module is also used to determine the number of updates of the initial model; determine a first weight coefficient corresponding to the first initial network and a second weight coefficient corresponding to the second initial network according to the number of updates; determine a model loss value of the initial model according to the first weight coefficient, the second weight coefficient, the first loss value and the second loss value; and update the initial model according to the model loss value.

[0136] In some embodiments, the processing module is also used to input the sample user characteristics, the sample item characteristics and the sample interaction behavior queue into the third initial network to obtain a sample predicted behavior sequence; the training module is also used to update the initial model according to the predicted interest, the predicted probability and the third loss value corresponding to the sample predicted behavior sequence.

[0137] Based on the same inventive concept, the embodiment of the present application also provides a computer device, which may be a server or a terminal device, and includes a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps of the above-mentioned item recommendation method when executing the computer program. Thus, various functions are realized, such as:

[0138] Determining user characteristics of a target user, item characteristics of candidate items, and interaction behaviors between the target user and multiple items on a network platform;

[0139] Inputting the user features and the item features into a first network of a first model to predict the interest of the target user in the candidate item;

[0140] Inputting the item features and the interaction behavior into the second network of the first model to predict a first probability of the target user interacting with the candidate item;

[0141] According to the interest level and the first probability, a recommended item is selected from the candidate items to be recommended to the target user. Or the following functions are implemented:

[0142] Determining sample user characteristics of sample users, sample item characteristics of sample items, and sample interaction behaviors between the sample target users and a plurality of items;

[0143] Inputting the sample user features and the sample item features into a first initial network of an initial model to obtain the predicted interest of the sample user in the sample item;

[0144] Inputting the sample item features and the sample interaction behavior into a second initial network of the initial model to obtain a predicted probability of the sample user interacting with the sample item;

[0145] The initial model is updated according to the predicted interest and the predicted probability.

[0146] In one embodiment, the computer device is a terminal device, for example, its internal structure diagram can be as follows: Fig. 9 As shown. The computer device includes a processor, a memory, an input / output interface, a communication interface, a display unit and an input device. The processor, the memory and the input / output interface are connected via a system bus, and the communication interface, the display unit and the input device are connected to the system bus via the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be implemented through WIFI, a mobile cellular network, NFC (near field communication) or other technologies. When the computer program is executed by the processor, an item recommendation method is implemented. The display unit of the computer device is used to form a visually visible image, and can be a display screen, a projection device or a virtual reality imaging device. The display screen can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer covered on the display screen, or a button, trackball or touchpad set on the computer device casing, or an external keyboard, touchpad or mouse, etc.

[0147] Those skilled in the art will understand that Fig. 9 The structure shown in the figure is only a block diagram of a partial structure related to the present application scheme, and does not constitute a limitation on the computer device to which the present application scheme is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0148] Based on the same inventive concept, an embodiment of the present application also provides a computer-readable storage medium, which may include: a read-only memory (ROM), a random access memory (RAM), a disk or an optical disk, etc.

[0149] Since the computer program stored in the computer-readable storage medium can execute any one of the item recommendation methods provided in the embodiments of the present application, the beneficial effects that can be achieved by any one of the item recommendation methods provided in the embodiments of the present application can be achieved. Please refer to the previous embodiments for details and will not be repeated here.

[0150] Based on the same inventive concept, the embodiment of the present application further provides a computer program product or a computer program, which includes a computer instruction, and the computer instruction is stored in a computer-readable storage medium. The processor of the computer device reads the computer instruction from the computer-readable storage medium, and the processor executes the computer instruction, so that the computer device executes the method provided in various optional implementations in the above embodiments.

[0151] It should be noted that the object data (including but not limited to user device information, user personal information, etc.) and conversation data involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data need to comply with relevant laws, regulations and standards of relevant countries and regions. A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods.

[0152] Any reference to the memory, database or other media used in the embodiments provided in the present application may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (Re RAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory may include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0153] The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. The non-relational database may include a distributed database based on blockchain, etc., but is not limited thereto. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but is not limited thereto.

[0154] In the above-mentioned item recommendation device, model processing device, computer-readable storage medium, computer device, and computer program product embodiments, the description of each embodiment has its own emphasis. For parts not described in detail in a certain embodiment, reference can be made to the relevant description of other embodiments. Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working process and beneficial effects of the above-described item recommendation device, computer-readable storage medium, computer program product, computer device and its corresponding units can refer to the description of the item recommendation method in the above embodiment, and will not be repeated here.

[0155] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0156] The above is a detailed introduction to an item recommendation method and model processing method, computer equipment and storage medium provided in the embodiments of the present application. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea; at the same time, for technical personnel in this field, according to the ideas of the present application, there will be changes in the specific implementation methods and application scopes. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. An item recommendation method, characterized in that: The method comprises: Determining user characteristics of a target user, item characteristics of candidate items, and interaction behaviors between the target user and multiple items on a network platform; Inputting the user features and the item features into a first network of a first model to predict the interest of the target user in the candidate item; Inputting the item features and the interaction behavior into the second network of the first model to predict a first probability of the target user interacting with the candidate item; A recommended item is selected from the candidate items to be recommended to the target user according to the interest degree and the first probability.

2. The method according to claim 1, characterized in that The selecting, according to the interest level and the first probability, a recommended item to be recommended to the target user from the candidate items includes: Determine weight coefficients corresponding to the first network and the second network in the first model; Determining a recommendation score for the target user corresponding to the candidate item according to the weight coefficient, the interest degree and the first probability; Based on the recommendation score, a recommended item is selected from the candidate items to be recommended to the target user.

3. The method according to claim 1, characterized in that Before selecting a recommended item to be recommended to the target user from the candidate items according to the interest level and the first probability, the method further includes: Inputting the user features, the item features and the interaction behaviors into a third network of the first model to obtain a predicted behavior sequence; the predicted behavior sequence includes predicted interaction behaviors between the target user and a plurality of items among the candidate items; The selecting, according to the interest level and the first probability, a recommended item to be recommended to the target user from the candidate items includes: Determining, from the candidate items, a target item corresponding to the interactive behavior according to the predicted interactive behavior included in the predicted behavior sequence; A recommended item is selected from the target items to be recommended to the target user according to the interest degree and the first probability.

4. The method according to any one of claims 1 to 3, characterized in that: The determining of the user characteristics of the target user, the item characteristics of the candidate items, and the interaction behavior between the target user and the multiple items on the network platform includes: Acquire initial user features of a target user, initial item features of candidate items, and interaction behaviors between the target user and multiple items on the network platform; Inputting the interaction behavior into a language model to obtain a user description text of the target user; Generate a target question sentence associated with the candidate item according to a preset question template, and input the target question sentence into a language model to obtain an item description text of the candidate item; Performing enhancement processing on the initial user features according to the user description text to obtain user features of the target user; The initial item features are enhanced according to the item description text to obtain item features of the candidate item.

5. A model processing method, characterized in that: The method comprises: Determining sample user characteristics of sample users, sample item characteristics of sample items, and sample interaction behaviors between the sample target users and a plurality of items; Inputting the sample user features and the sample item features into a first initial network of an initial model to obtain the predicted interest of the sample user in the sample item; Inputting the sample item features and the sample interaction behavior into a second initial network of the initial model to obtain a predicted probability of the sample user interacting with the sample item; The initial model is updated according to the predicted interest and the predicted probability.

6. The method according to claim 5, characterized in that The updating of the initial model according to the predicted interest and the predicted probability includes: Determining a recommended tag between the sample user and the sample item; Determining a first loss value corresponding to the first initial network according to the predicted interest level and a first preset value corresponding to the recommended tag; Determine a second loss value corresponding to the second initial network according to the predicted probability and a second preset value corresponding to the sample interaction behavior; The initial model is updated according to the first loss value and the second loss value until a target model is obtained.

7. The method according to claim 6, characterized in that The updating of the initial model according to the first loss value and the second loss value includes: Determining the number of updates of the initial model; Determine a first weight coefficient corresponding to the first initial network and a second weight coefficient corresponding to the second initial network according to the number of updates; Determining a model loss value of the initial model according to the first weight coefficient, the second weight coefficient, the first loss value, and the second loss value; The initial model is updated according to the model loss value.

8. The method according to claim 5 or 6, characterized in that: The initial model also includes a third initial network, and the method further includes: Inputting the sample user features, the sample item features and the sample interaction behavior queue into the third initial network to obtain a sample predicted behavior sequence; The updating of the initial model according to the predicted interest and the predicted probability includes: The initial model is updated according to the predicted interest, the predicted probability and a third loss value corresponding to the sample predicted behavior sequence.

9. A computer device, characterized in that: It includes a processor and a memory, wherein the memory stores a plurality of instructions; the processor loads instructions from the memory to execute the steps of the item recommendation method as described in any one of claims 1 to 4 or the steps of the model processing method as described in any one of claims 5 to 8.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a plurality of instructions, which are suitable for being loaded by a processor to execute the steps of the item recommendation method as described in any one of claims 1 to 4 or the steps of the model processing method as described in any one of claims 5 to 8.

Images