Model generation method and device, article recommendation method and device, electronic equipment and medium

Abstract

The embodiment of the invention discloses a model generation method and device, an article recommendation method and device, electronic equipment and a medium. A specific embodiment of the method comprises the steps of obtaining training data; generating initial article interest information by utilizing an initial article interest information generation model according to the user data and the article data; generating initial position deviation information by utilizing an initial position deviation information generation model according to the article recommendation position information; according to the initial article interest information and the initial position deviation information, article interest information for a target article is generated, and the target article is an article corresponding to the article data; and generating an article interest information generation model and a position deviation information generation model according to the article interest information. The implementation mode is related to artificial intelligence, the obtained article interest information generation model can be utilized, and the article interest information representing the interest degree of the user in the article can be accurately generated under the condition that the position deviation is eliminated.

Description

Technical Field

[0001] The embodiments of this disclosure relate to the field of computer technology, and more specifically to model generation methods, item recommendation methods, apparatus, electronic devices, and media. Background Technology

[0002] Currently, with the continuous development of artificial intelligence, the application of AI models in the recommendation field is becoming increasingly mature. Item recommendation based on user interests is a commonly used technique. In the process of obtaining item interest information, eliminating biases such as location, popularity, and exposure has become one of the main development directions. To eliminate biases, the common approach is as follows: First, during training, the sample's location is used as a one-dimensional feature and combined with other features into the item interest model. Then, during testing, the location feature is uniformly set to a default value to remove the influence of location and generate item interest information.

[0003] However, the inventors discovered that the following technical problems often arise when using the above method:

[0004] Using the location of a sample as a feature and incorporating it into other features during training can lead to premature crossover between features in the item interest model. Even if the location is set to the default value during subsequent testing, it is difficult to completely eliminate the influence of the location. As a result, the item interest information is often affected by the location deviation, and the generated item interest information is not accurate enough.

[0005] The information disclosed in this background section is only intended to enhance the understanding of the background of the inventive concept, and therefore may contain information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The summary portion of this disclosure is intended to provide a brief overview of the concepts, which will be described in detail in the detailed description portion later. This summary portion is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.

[0007] Some embodiments of this disclosure propose model generation methods, item recommendation methods, apparatuses, electronic devices, and media to address the technical problems mentioned in the background section above.

[0008] In a first aspect, some embodiments of this disclosure provide a model generation method, comprising: acquiring training data, wherein the training data includes user data and item data, the item data including item recommendation location information; generating initial item interest information by generating a model using initial item interest information based on the user data and the item data; generating initial location deviation information by generating a model using initial location deviation information based on the item recommendation location information; generating item interest information for a target item based on the initial item interest information and the initial location deviation information, wherein the target item is an item corresponding to the item data; and generating an item interest information generation model and a location deviation information generation model based on the item interest information.

[0009] Optionally, generating item interest information for a target item based on the initial item interest information and the initial position deviation information includes: obtaining item statistics data corresponding to the target item; generating initial position deviation gating information using an initial position deviation gating model based on the item statistics data and the initial item interest information; and generating the item interest information based on the initial item interest information, the initial position deviation information, and the initial position deviation gating information.

[0010] Optionally, the above-mentioned method of generating initial position deviation information based on the recommended item position information using an initial position deviation information generation model includes: performing word embedding processing on the recommended item position information to generate an item recommended position vector; performing position smoothing processing on the recommended item position vector to generate a smooth position vector; and inputting the smooth position vector into the initial position deviation information generation model to generate the initial position deviation information.

[0011] Optionally, the aforementioned initial item interest information generation model includes: an initial multi-head attention mechanism model and an initial expert network model; and the generation of initial item interest information based on the aforementioned user data and the aforementioned item data using the initial item interest information generation model includes: performing data vector transformation on the aforementioned user data to generate user vectors; performing data vector transformation on the aforementioned item data to generate item vectors; inputting the aforementioned user vectors and the aforementioned item vectors into the aforementioned initial multi-head attention mechanism model to generate initial user attention vectors and initial item attention vectors; and inputting the aforementioned initial user attention vectors and the aforementioned initial item attention vectors into the aforementioned initial expert network model to generate the aforementioned initial item interest information.

[0012] Optionally, generating the item interest information based on the initial item interest information, the initial position deviation information, and the initial position deviation gating information includes: multiplying the initial position deviation information and the initial position deviation gating information to generate multiplied information; performing a random information discarding operation on the multiplied information to generate discarded information; and adding the discarded information and the initial item interest information to generate item interest information.

[0013] Optionally, the aforementioned item statistics include at least one of the following: item exposure information, item click volume information, and item click-through rate information; and the generation of initial position deviation gating information based on the aforementioned item statistics and the aforementioned initial item interest information using the initial position deviation gating model includes: vectorizing at least one of the aforementioned item exposure information, the aforementioned item click volume information, and the aforementioned item click-through rate information to generate a vector set; vectorizing the aforementioned initial item interest information to generate an interest vector; concatenating each vector in the aforementioned vector set with the aforementioned interest vector to generate a concatenated vector; and inputting the aforementioned concatenated vector into the aforementioned initial position deviation gating model to generate the aforementioned initial position deviation gating information.

[0014] Secondly, some embodiments of this disclosure provide a model generation apparatus, comprising: a first acquisition unit configured to acquire training data, wherein the training data includes user data and item data, the item data including item recommendation location information; a first generation unit configured to generate initial item interest information by generating a model using initial item interest information based on the user data and the item data; a second generation unit configured to generate initial location deviation information by generating a model using initial location deviation information based on the item recommendation location information; a third generation unit configured to generate item interest information for a target item based on the initial item interest information and the initial location deviation information, wherein the target item is an item corresponding to the item data; and a fourth generation unit configured to generate an item interest information generation model and a location deviation information generation model based on the item interest information.

[0015] Optionally, the third generation unit can be configured to: obtain the item statistics data corresponding to the target item; generate initial position deviation gating information based on the item statistics data and the initial item interest information using the initial position deviation gating model; and generate the item interest information based on the initial item interest information, the initial position deviation information, and the initial position deviation gating information.

[0016] Optionally, the second generation unit can be configured to: perform word embedding processing on the above-mentioned item recommendation location information to generate an item recommendation location vector; perform position smoothing processing on the above-mentioned item recommendation location vector to generate a smoothed location vector; and input the above-mentioned smoothed location vector into the above-mentioned initial location deviation information generation model to generate the above-mentioned initial location deviation information.

[0017] Optionally, the initial item interest information generation model includes: an initial multi-head attention mechanism model and an initial expert network model; and performs data vector transformation on the user data to generate user vectors; performs data vector transformation on the item data to generate item vectors; inputs the user vectors and item vectors into the initial multi-head attention mechanism model to generate initial user attention vectors and initial item attention vectors; and inputs the initial user attention vectors and initial item attention vectors into the initial expert network model to generate the initial item interest information.

[0018] Optionally, the third generation unit can be configured to: multiply the initial position deviation information and the initial position deviation gating information to generate multiplied information; perform a random information discarding operation on the multiplied information to generate discarded information; and add the discarded information and the initial item interest information to generate item interest information.

[0019] Optionally, the aforementioned item statistics include at least one of the following: item exposure information, item click volume information, and item click-through rate information; and the third generation unit can be configured to: vectorize at least one of the aforementioned item exposure information, the aforementioned item click volume information, and the aforementioned item click-through rate information to generate a vector set; vectorize the aforementioned initial item interest information to generate an interest vector; concatenate each vector in the aforementioned vector set and the aforementioned interest vector to generate a concatenated vector; and input the aforementioned concatenated vector into the aforementioned initial position deviation gating model to generate the aforementioned initial position deviation gating information.

[0020] Thirdly, some embodiments of this disclosure provide an item recommendation method, including: acquiring an item dataset and target user data; for each item data in the item dataset, performing the following generation steps: vectorizing the item data to generate an item vector, which serves as a target item vector; vectorizing the target user data to generate a target user vector; inputting the target item vector and the target user vector into a pre-trained item interest information generation model to generate item interest information, wherein the item interest information generation model is generated based on an item recommendation method; according to the obtained item interest information set, selecting item information to be recommended to the target user from the item information set as recommended item information, obtaining at least one recommended item information, wherein the target user corresponds to the target user data, and the item information in the item information set corresponds to the item data in the item dataset; and performing item recommendation for the at least one recommended item information.

[0021] Optionally, the above-mentioned item interest information generation model includes: a multi-head attention mechanism model and an expert network model; and the above-mentioned inputting the target item vector and the target user vector into the pre-trained item interest information generation model to generate item interest information includes: inputting the target item vector and the target user vector into the multi-head attention mechanism model to generate user attention vector and item attention vector; and inputting the user attention vector and the item attention vector into the expert network model to generate the item interest information.

[0022] Fourthly, some embodiments of this disclosure provide an item recommendation apparatus, comprising: a second acquisition unit configured to acquire an item dataset and target user data; a first execution unit configured to perform the following generation steps for each item data in the item dataset: vectorizing the item data to generate an item vector as a target item vector; vectorizing the target user data to generate a target user vector; inputting the target item vector and the target user vector into a pre-trained item interest information generation model to generate item interest information, wherein the item interest information generation model is generated based on a model generation method; a filtering unit configured to filter item information to be recommended to the target user from the item information set according to the obtained item interest information set, as recommended item information, to obtain at least one recommended item information, wherein the target user corresponds to the target user data, and the item information in the item information set corresponds to the item data in the item dataset; and a second execution unit configured to perform item recommendation for the at least one recommended item information.

[0023] Optionally, the above-mentioned item interest information generation model includes: a multi-head attention mechanism model and an expert network model; and the first execution unit can be configured to: input the above-mentioned target item vector and the above-mentioned target user vector into the above-mentioned multi-head attention mechanism model to generate user attention vector and item attention vector; input the above-mentioned user attention vector and the above-mentioned item attention vector into the above-mentioned expert network model to generate the above-mentioned item interest information.

[0024] Fifthly, some embodiments of this disclosure provide an electronic device, including: one or more processors; and a storage device having one or more programs stored thereon, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any of the implementations of the first and third aspects.

[0025] Sixthly, some embodiments of this disclosure provide a computer-readable medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the method as described in any of the implementations of the first and third aspects.

[0026] In a seventh aspect, some embodiments of this disclosure provide a computer program product, including a computer program that, when executed by a processor, implements the method described in any of the implementations of the first aspect above.

[0027] The above embodiments of this disclosure have the following beneficial effects: Through the model generation methods of some embodiments of this disclosure, a model can be generated using the obtained item interest information, accurately generating item interest information representing the user's degree of interest in items while eliminating positional bias. Specifically, the reason why the relevant item interest information is not accurate enough is that: merging the sample position as a feature into other features for training leads to premature crossover between features in the item interest model. Even if the position is set to a default value during subsequent testing, it is difficult to completely eliminate the influence of position, resulting in the item interest information often being affected by positional bias, and the generated item interest information being inaccurate. Based on this, the model generation method of some embodiments of this disclosure first acquires training data, wherein the training data includes: user data and item data, and the item data includes: item recommendation location information. Here, acquiring training data facilitates the training of the initial item interest information generation model and the initial positional bias information generation model. Here, extracting item recommendation location information allows for full consideration of item recommendation location features in subsequent training of the item interest information generation model, so as to fully consider the impact of recommendation location offset on output accuracy during the training process of the item interest information generation model. Then, based on the aforementioned user data and item data, the initial item interest information generation model can accurately generate initial item interest information to obtain item interest information before the impact of positional offset is eliminated. Next, based on the aforementioned item recommendation position information, the initial positional deviation information generation model can accurately generate initial positional deviation information to facilitate the elimination of the impact of recommendation positional offset on the initial item interest information. Furthermore, based on the aforementioned initial item interest information and the aforementioned initial positional deviation information, item interest information for the target item can be accurately generated. The obtained item interest information can minimize the impact of positional offset to generate more accurate information representing the user's degree of interest in the item. Here, the aforementioned target item is the item corresponding to the aforementioned item data. Finally, based on the aforementioned item interest information, an item interest information generation model and a positional deviation information generation model are generated. In summary, during the training process of the initial item interest information generation model, the positional deviation information generation model can fully learn the impact information brought about by the recommendation positional offset to accurately generate item interest information representing the user's degree of interest in the item. Attached Figure Description

[0028] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and elements are not necessarily drawn to scale.

[0029] Figure 1This is a schematic diagram illustrating an application scenario of a model generation method according to some embodiments of the present disclosure;

[0030] Figure 2 This is a flowchart of some embodiments of the model generation method according to this disclosure;

[0031] Figure 3 These are flowcharts of other embodiments of the model generation method according to this disclosure;

[0032] Figure 4 This is a flowchart of some embodiments of the item recommendation method according to this disclosure;

[0033] Figure 5 This is a schematic diagram of the structure of some embodiments of the model generation apparatus according to the present disclosure;

[0034] Figure 6 These are schematic diagrams illustrating the structure of some embodiments of the item recommendation device according to this disclosure;

[0035] Figure 7 This is a schematic diagram of the structure of an electronic device suitable for implementing some embodiments of the present disclosure. Detailed Implementation

[0036] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0037] It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other.

[0038] It should be noted that the concepts of "first" and "second" mentioned in this disclosure are used only to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0039] It should be noted that the terms "a" and "a plurality of" used in this disclosure are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0040] The names of messages or information exchanged between multiple devices in the embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of such messages or information.

[0041] Before performing any of the operations involving the collection, storage, or use of user personal information (such as user data) disclosed in this disclosure, the relevant organizations or individuals shall fulfill their obligations, including conducting personal information security impact assessments, informing personal information subjects, and obtaining prior authorization and consent from personal information subjects.

[0042] This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0043] Figure 1 This is a schematic diagram illustrating an application scenario of a model generation method according to some embodiments of the present disclosure.

[0044] exist Figure 1 In this application scenario, firstly, the electronic device 101 can acquire training data. This training data includes user data 102 and item data 103. The item data includes item recommendation location information 104. In this application scenario, user data 102 can be: "User age: 18; User purchase behavior: Purchased target item in *year*month; User gender: Male; User add-to-cart behavior: Added target item in *year*month." Item data 103 can be: "Item category: First-level category; Item value: 188 yuan; Item origin: *province*city; Item weight: 188g; Item recommendation location: 4th item in the 7th row of the page." Item recommendation location information 104 can be: "Item recommendation location information: 4th item in the 7th row of the page." Then, based on the user data 102 and item data 103, the electronic device 101 can use the initial item interest information generation model 105 to generate initial item interest information 106. In this application scenario, initial item interest information 106 can be: "Initial item interest: 56 points." Next, the electronic device 101 can generate initial position deviation information 108 based on the recommended item location information 104 and the initial position deviation information generation model 107. In this application scenario, the initial position deviation information 108 can be: "Initial position deviation information: 12". Furthermore, the electronic device 101 can generate item interest information 109 for the target item based on the initial item interest information 106 and the initial position deviation information 108. The target item is the item corresponding to the item data. In this application scenario, the item interest information 109 can be: "Item interest: 66 points". Finally, the electronic device 101 can generate an item interest information generation model and a position deviation information generation model based on the item interest information 109.

[0045] It should be noted that the aforementioned electronic device 101 can be either hardware or software. When the electronic device is hardware, it can be implemented as a distributed cluster consisting of multiple servers or terminal devices, or as a single server or a single terminal device. When the electronic device is software, it can be installed in the hardware devices listed above. It can be implemented as, for example, multiple software programs or software modules used to provide distributed services, or as a single software program or software module. No specific limitations are made here.

[0046] It should be understood that Figure 1 The number of electronic devices shown is merely illustrative. Any number of electronic devices can be used depending on the implementation requirements.

[0047] Continue to refer to Figure 2 The diagram illustrates a flow 200 of some embodiments of a model generation method according to the present disclosure. The model generation method includes the following steps:

[0048] Step 201: Obtain training data.

[0049] In some embodiments, the execution entity of the above model generation method (e.g. Figure 1 The electronic device 101 shown can acquire training data via wired or wireless connection. The training data includes user data and item data, with the item data including item recommendation location information. The training data is used for subsequent model training. User data can be user-related data. In practice, user data can include data corresponding to user behavior characteristics and data corresponding to user attributes. For example, in e-commerce scenarios, user behavior characteristics can be, but are not limited to, at least one of the following: user browsing characteristics, user click characteristics, user order placement characteristics, user favorites characteristics, and shopping cart addition characteristics. User attributes can be, but are not limited to, at least one of the following: user age, user gender, and user device identifier. Item data can be item-related data. In practice, in e-commerce scenarios, items can be commodities. In practice, item data can include, but is not limited to, at least one of the following: item category, item title, and item image. Item recommendation location information can be the position of the item in the item recommendation queue. Similarly, item recommendation location information can represent the degree of recommendation for the item. Specifically, the higher the item's position in the item recommendation queue, the greater the recommendation strength. Items ranked higher in the recommendation queue appear more prominently on the recommendation page, and are more likely to become popular. The user and item data in the training data can be pre-processed and in a predetermined format.

[0050] In addition, item recommendation location information can be the actual exposure position of the item on the page. Specifically, item recommendation location information includes page and index. Page indicates which page the item appears on, and index indicates the position of the item on the page, using different location IDs to represent the actual position on the page. For example, item recommendation location information could be page * 10 + index.

[0051] Step 202: Based on the user data and item data mentioned above, generate initial item interest information using the initial item interest information generation model.

[0052] In some embodiments, the aforementioned execution entity can generate initial item interest information based on the aforementioned user data and item data using an initial item interest information generation model. This initial item interest information generation model can be an item interest information generation model that has not yet completed training. The item interest information generation model can be a neural network model that generates item interest information. Item interest information can represent the degree of user interest in items. Item interest information can be information in various forms. Similarly, item interest information can also reflect its degree of interest through the indicator features of at least one predetermined metric. For example, item interest information can reflect its degree of interest through item click-through rate. In addition, item interest information can be information in the form of scores. The higher the score, the higher the corresponding item click-through rate, and the higher the degree of user interest. The item interest information generation model can be trained using conventional neural network training methods. In practice, the item interest information generation model can be a dual-tower model. The initial item interest information can be interest information that has not yet undergone recommendation position offset elimination.

[0053] As an example, firstly, the aforementioned execution entity can perform embedding processing on the user data and item data to generate user data vectors and item data vectors. Then, the user data vectors and the aforementioned item data vectors are input into the initial item interest information generation model to generate initial item interest information.

[0054] Step 203: Based on the above recommended location information of the items, generate the initial location deviation information by using the initial location deviation information to generate the model.

[0055] In some embodiments, the execution entity can generate initial position deviation information based on the recommended item location information using an initial position deviation information generation model. The initial position deviation information generation model can be a position deviation information generation model that has not yet finished training. The position deviation information generation model can be a neural network model that generates position deviation information. Position deviation information can characterize the deviation of the item's corresponding recommended location. That is, position deviation information can characterize the degree of deviation of the item's corresponding recommended location. For example, position deviation information could be "the degree of deviation of the item's corresponding recommended location is the first degree." In practice, the position deviation information generation model can be a model including at least one fully connected layer. The recommended item location can be the placement of the item on the recommendation page.

[0056] As an example, the aforementioned execution entity can vectorize the recommended location information of the aforementioned items to generate a location vector. Then, the location vector is input into the aforementioned initial location deviation information generation model to generate initial location deviation information.

[0057] In practice, existing deep learning-based recommendation systems commonly suffer from positional bias. This bias prevents the model from scoring products based on the user's true interest, leading to popular products appearing prominently in the user's recommendation list, exacerbating the Matthew effect and negatively impacting the user experience. Therefore, a positional offset information generation model can accurately generate positional offset information.

[0058] Step 204: Generate item interest information for the target item based on the initial item interest information and the initial position deviation information.

[0059] In some embodiments, the executing entity can generate item interest information for a target item based on the initial item interest information and the initial position deviation information. The target item is the item corresponding to the item data. The item interest information may also include interest information that recommends position offset effect elimination.

[0060] As an example, firstly, the executing entity can obtain the preset positional deviation weight corresponding to the target item. Then, it multiplies the positional deviation weight with the initial positional deviation information to generate multiplied information. Finally, it adds the initial item interest information with the multiplied information to generate item interest information.

[0061] In some optional implementations of certain embodiments, based on the recommended location information of the items, the executing entity can generate a model using the initial location deviation information to generate the initial location deviation information, which may include the following steps:

[0062] The first step is to perform word embedding processing on the above-mentioned item recommendation location information to generate an item recommendation location vector. This item recommendation location vector can be a vector representing the semantic information of the recommendation features corresponding to the item recommendation location information.

[0063] As an example, the aforementioned execution entity can input the item recommendation location information into the word embedding model to generate an item recommendation location vector.

[0064] The second step is to perform position smoothing on the recommended position vectors of the above items to generate smooth position vectors.

[0065] As an example, firstly, the executing entity can identify at least one item that is geographically adjacent to the target item. Then, it determines the recommended position vector for each of the at least one item, resulting in at least one recommended position vector. Next, it performs vector averaging on the at least one recommended position vector and the other recommended position vector to generate an average vector, which serves as a smoothed position vector. The number of items corresponding to at least one item can be pre-set.

[0066] In real-world recommendation scenarios, subsequent links employ complex recommendation strategies. Therefore, the magnitude of positional bias does not decrease monotonically with increasing position ID; instead, it exhibits jagged fluctuations. To enable the model to learn a more generalized positional bias (when the distribution of positional bias changes due to changes in subsequent recommendation strategies), rather than a highly precise one, a positional smoothing method is employed. This method merges positions with different recommendation strategies from adjacent links, preventing the model from overly relying on the precise location of item appearance and mitigating the impact of changes in subsequent recommendation strategies on the model's bias reduction performance.

[0067] The third step is to input the smoothed position vector into the initial position deviation information generation model to generate the initial position deviation information.

[0068] As an example, the aforementioned execution entity can generate initial position deviation information by invoking a graphics processor.

[0069] Step 205: Based on the above item interest information, generate an item interest information generation model and a location deviation information generation model.

[0070] In some embodiments, the execution entity can generate an item interest information generation model and a location deviation information generation model based on the item interest information. The item interest information generation model and the location deviation information generation model can be models that have already been trained.

[0071] Here, using training data, a dual-tower model, consisting of an initial item interest information generation model and an initial positional deviation information generation model, is trained as a whole. This allows the trained positional deviation information generation model to accurately generate the positional deviation of the corresponding item, and also enables the trained item interest information generation model to generate item interest information after positional deviation correction. Through the training of the dual-tower model, while training the initial positional deviation information generation model, the initial item interest information generation model can potentially learn the data features of items without positional deviation information during training, thus generating item interest information without positional deviation.

[0072] As an example, firstly, the aforementioned execution entity obtains the recommendation tags corresponding to the target item. Then, using a pre-set loss function, it determines the loss value between the item interest information and the recommendation tags. Finally, in response to the numerical change of the loss value sequence corresponding to the aforementioned loss value tending towards the target value, the initial item interest information generation model is determined as the item interest information generation model, and the initial location deviation information generation model is determined as the location deviation information generation model. In response to the numerical change of the loss value sequence corresponding to the aforementioned loss value not tending towards the target value, training data is re-acquired to train the initial location deviation information generation model and the initial item interest information generation model, continuously training the models until the numerical change of the corresponding loss value sequence tends towards the target value.

[0073] In some optional implementations of certain embodiments, the execution entity can generate item interest information for the target item based on the initial item interest information and the initial position deviation information, which may include the following steps:

[0074] The first step is to obtain the item statistics data corresponding to the target item. This item statistics data can be the statistical content of various indicator characteristics corresponding to the target item. In practice, these indicator characteristics may include: impressions, clicks, and click-through rate (CTR). In practice, the item statistics data can also be the statistical data of the target item within a predetermined historical time period.

[0075] The second step involves generating initial position deviation gating information based on the aforementioned item statistics and initial item interest information, using an initial position deviation gating model. This initial position deviation gating model can be a position deviation gating model that has not yet completed training. The position deviation gating model can be a neural network model that generates the position deviation gating information. The position deviation gating information can characterize the degree of deviation in the position deviation information. That is, it can characterize not only the accuracy of the position deviation information for the target item, but also how to adjust the position deviation to accurately generate the recommended position deviation for the target item. In practice, the position deviation gating model can be a gating network.

[0076] As an example, firstly, the aforementioned executing entity can combine item statistics and initial item interest information to generate combined data. Then, the combined data is input into the initial position deviation gating model to generate initial position deviation gating information.

[0077] The third step is to generate the above-mentioned item interest information based on the above-mentioned initial item interest information, the above-mentioned initial position deviation information, and the above-mentioned initial position deviation gating information.

[0078] As an example, firstly, the executing entity can multiply the initial position deviation information and the initial position deviation gating information to generate deviation multiplication information. Then, it can add the deviation multiplication information and the initial item interest information to generate item interest information.

[0079] In some optional implementations of certain embodiments, generating the item interest information based on the initial item interest information, the initial position deviation information, and the initial position deviation gating information may include the following steps:

[0080] The first step is to multiply the initial position deviation information and the initial position deviation gating information to generate multiplied information.

[0081] The second step involves performing a random information discarding operation on the multiplied information to generate discarded information. This random information discarding operation can be a Dropout operation. The probability value corresponding to this operation is a pre-set value. For example, the probability value could be 15%. The discarded information can be information that falls within a target value based on the corresponding probability value. For example, the target value could be the value 1. That is, the discarded information has a 15% chance of being set to the value 1.

[0082] The third step is to add the discarded information and the initial item interest information together to generate item interest information.

[0083] In some optional implementations of certain embodiments, the aforementioned item statistics include at least one of the following: item exposure information, item click count information, and item click-through rate information. In practice, item exposure information can be the number of times an item is exposed.

[0084] Optionally, based on the aforementioned item statistics and initial item interest information, and utilizing the initial position deviation gating model, the executing entity can generate initial position deviation gating information, which may include the following steps:

[0085] The first step is to vectorize at least one of the above-mentioned item exposure information, item click volume information, and item click-through rate information to generate a vector set. That is, the vector set includes at least one of the following: the information vector corresponding to the item exposure information, the information vector corresponding to the item click volume information, and the information vector corresponding to the item click-through rate.

[0086] The second step is to vectorize the initial item interest information to generate interest vectors.

[0087] As an example, the aforementioned execution entity can perform embedding processing on the initial item interest information to generate an interest vector.

[0088] The third step is to concatenate the vectors in the above vector set with the above interest vector to generate a concatenated vector.

[0089] The fourth step is to input the spliced ​​vector into the initial position deviation gating model to generate the initial position deviation gating information.

[0090] The above embodiments of this disclosure have the following beneficial effects: Through the model generation methods of some embodiments of this disclosure, a model can be generated using the obtained item interest information, accurately generating item interest information representing the user's degree of interest in items while eliminating positional bias. Specifically, the reason why the relevant item interest information is not accurate enough is that: merging the sample position as a feature into other features for training leads to premature crossover between features in the item interest model. Even if the position is set to a default value during subsequent testing, it is difficult to completely eliminate the influence of position, resulting in the item interest information often being affected by positional bias, and the generated item interest information being inaccurate. Based on this, the model generation method of some embodiments of this disclosure first acquires training data, wherein the training data includes: user data and item data, and the item data includes: item recommendation location information. Here, acquiring training data facilitates the training of the initial item interest information generation model and the initial positional bias information generation model. Here, extracting item recommendation location information allows for full consideration of item recommendation location features in subsequent training of the item interest information generation model, so as to fully consider the impact of recommendation location offset on output accuracy during the training process of the item interest information generation model. Then, based on the aforementioned user data and item data, the initial item interest information generation model can accurately generate initial item interest information to obtain item interest information before the impact of positional offset is eliminated. Next, based on the aforementioned item recommendation position information, the initial positional deviation information generation model can accurately generate initial positional deviation information to facilitate the elimination of the impact of recommendation positional offset on the initial item interest information. Furthermore, based on the aforementioned initial item interest information and the aforementioned initial positional deviation information, item interest information for the target item can be accurately generated. The obtained item interest information can minimize the impact of positional offset to generate more accurate information representing the user's degree of interest in the item. Here, the aforementioned target item is the item corresponding to the aforementioned item data. Finally, based on the aforementioned item interest information, an item interest information generation model and a positional deviation information generation model are generated. In summary, during the training process of the initial item interest information generation model, the positional deviation information generation model can fully learn the impact information brought about by the recommendation positional offset to accurately generate item interest information representing the user's degree of interest in the item.

[0091] Further reference Figure 3 The diagram illustrates a flow 300 of another embodiment of the model generation method according to this disclosure. The model generation method includes the following steps:

[0092] Step 301: Obtain training data.

[0093] Step 302: Perform data vector transformation on the above user data to generate user vectors.

[0094] In some embodiments, the executing entity (e.g. Figure 1 The electronic device 101 shown can perform data vector transformation on the aforementioned user data to generate user vectors. The user vector can be a vector representing the semantic features of the corresponding user data. The initial item interest information generation model includes an initial multi-head attention mechanism model and an initial expert network model. The initial multi-head attention mechanism model can be a multi-head attention mechanism model that has not yet finished training. In practice, the multi-head attention mechanism model can be a Transformer model. The initial expert network model can be an expert network model that has not yet finished training. In practice, the expert network model can be a network model for each execution task. Each execution task has a corresponding task scenario. For example, the execution task can be an order placement prediction task. The expert network model can be an MMoE model (Multi-gate Mixture-of-Experts), including multiple expert networks. That is, the multi-head attention mechanism model consists of a multi-head attention network layer, a feedforward neural network layer, and a residual normalization network layer. In practice, the multiple expert networks can include: an expert network predicting conversion rate scores and an expert network predicting click-through rate scores.

[0095] As an example, the aforementioned execution entity can perform embedding processing on user data to generate user vectors.

[0096] Step 303: Perform data vector transformation on the above item data to generate item vectors.

[0097] In some embodiments, the execution entity may perform data vector transformation on the item data to generate item vectors. The item vector may be a vector representing the semantic features of the corresponding item data.

[0098] As an example, the aforementioned execution entity can perform embedding processing on item data to generate item vectors.

[0099] Step 304: Input the above user vector and the above item vector into the above initial multi-head attention mechanism model to generate the initial user attention vector and the initial item attention vector.

[0100] In some embodiments, the execution entity may input the user vector and the item vector into the initial multi-head attention mechanism model to generate an initial user attention vector and an initial item attention vector.

[0101] Step 305: Input the initial user attention vector and the initial item attention vector into the initial expert network model to generate the initial item interest information.

[0102] In some embodiments, the execution entity inputs the initial user attention vector and the initial item attention vector into the initial expert network model to generate the initial item interest information.

[0103] Step 306: Based on the above recommended location information of the items, generate the initial location deviation information by using the initial location deviation information to generate the model.

[0104] Step 307: Generate item interest information for the target item based on the initial item interest information and the initial position deviation information.

[0105] Step 308: Based on the above item interest information, generate an item interest information generation model and a location deviation information generation model.

[0106] from Figure 3 It can be seen from this that, with Figure 2 Compared to the description of some corresponding embodiments, Figure 3 In some corresponding embodiments, the model generation method process 300 utilizes a multi-head attention mechanism model to more fully extract user feature information from user data and user feature information from item data. Using an expert network model, the outputs of different tasks (click tasks, order tasks, etc.) can be fused for different scenarios (e.g., click scenarios, order scenarios) to accurately generate initial item interest information.

[0107] Continue to refer to Figure 4 The diagram illustrates a flow 400 of some embodiments of a model recommendation method according to the present disclosure. This model recommendation method includes the following steps:

[0108] Step 401: Obtain the item dataset and target user data.

[0109] In some embodiments, the execution entity of the above-described model recommendation method (e.g., an electronic device) can acquire an item dataset and target user data. The item dataset, corresponding to each item, can be the items to be recommended and ranked. The target user data can be user data corresponding to a target user, where the target user can be the user to whom the item is recommended.

[0110] Step 402: For each item in the above item dataset, perform the following generation steps:

[0111] Step 4021: Vectorize the above item data to generate item vectors, which will serve as the target item vectors.

[0112] In some embodiments, the aforementioned execution entity may perform Embedding processing on the item data to generate an item vector, which serves as the target item vector.

[0113] Step 4022: Vectorize the above target user data to generate target user vectors.

[0114] In some embodiments, the aforementioned execution entity may perform embedding processing on the target user data to generate a user vector, which serves as the target user vector.

[0115] Step 4023: Input the target item vector and the target user vector into the pre-trained item interest information generation model to generate item interest information.

[0116] In some embodiments, the execution entity may input the target item vector and the target user vector into a pre-trained item interest information generation model to generate item interest information. The item interest information generation model may be generated based on a model generation method.

[0117] In some optional implementations of certain embodiments, after step 4023, the steps further include:

[0118] The recommended location information of the items in the above item data is input into a pre-trained location deviation information generation model to generate location deviation information for the items corresponding to the above item data. The location deviation information generation model is trained based on a model generation method.

[0119] In some optional implementations of certain embodiments, the above-mentioned item interest information generation model includes: a multi-head attention mechanism model and an expert network model.

[0120] Optionally, the above-mentioned inputting the target item vector and the target user vector into a pre-trained item interest information generation model to generate item interest information may include the following steps:

[0121] The first step is to input the target item vector and the target user vector into the multi-head attention mechanism model to generate user attention vector and item attention vector.

[0122] The second step is to input the aforementioned user attention vector and item attention vector into the aforementioned expert network model to generate the aforementioned item interest information.

[0123] Step 403: Based on the obtained set of item interest information, filter out the item information to be recommended to the target user from the item information set, and obtain at least one set of recommended item information.

[0124] In some embodiments, the executing entity can filter out item information to be recommended to target users from the item information set based on the obtained item interest information set, and obtain at least one recommended item information. Here, the target user corresponds to the target user data, and the item information in the item information set corresponds to the item data in the item dataset.

[0125] As an example, the aforementioned executing entity can filter out item information from the item information set whose corresponding interest information scores are greater than a preset value, and use them as recommended item information to obtain at least one recommended item information.

[0126] Step 404: Perform item recommendation based on at least one of the above recommended item information.

[0127] In some embodiments, the aforementioned executing entity may perform item recommendations based on the aforementioned at least one recommended item information.

[0128] The above embodiments of this disclosure have the following beneficial effects: by using the model generation method of some embodiments of this disclosure to generate a model using item interest information, item recommendations can be achieved accurately and efficiently.

[0129] Further reference Figure 5 As an implementation of the methods shown in the above figures, this disclosure provides some embodiments of a model generation apparatus, which are similar to... Figure 2 Corresponding to the method embodiments shown, this model generation apparatus can be specifically applied to various electronic devices.

[0130] like Figure 5 As shown, a model generation device 500 includes: a first acquisition unit 501, a first generation unit 502, a second generation unit 503, a third generation unit 504, and a fourth generation unit 505. The first acquisition unit 501 is configured to acquire training data, including user data and item data, wherein the item data includes item recommendation location information. The first generation unit 502 is configured to generate initial item interest information based on the user data and the item data, using initial item interest information to generate a model. The second generation unit 503 is configured to generate initial position deviation information based on the item recommendation location information, using initial position deviation information to generate a model. The third generation unit 504 is configured to generate item interest information for a target item based on the initial item interest information and the initial position deviation information, wherein the target item is an item corresponding to the item data. The fourth generation unit 505 is configured to generate an item interest information generation model and a position deviation information generation model based on the item interest information.

[0131] In some optional implementations of certain embodiments, the third generation unit 504 may be configured to: obtain the item statistics data corresponding to the target item; generate initial position deviation gating information using an initial position deviation gating model based on the item statistics data and the initial item interest information; and generate the item interest information based on the initial item interest information, the initial position deviation information, and the initial position deviation gating information.

[0132] In some optional implementations of some embodiments, the second generation unit 503 may be configured to: perform word embedding processing on the above-mentioned item recommendation location information to generate an item recommendation location vector; perform position smoothing processing on the above-mentioned item recommendation location vector to generate a smoothed location vector; and input the above-mentioned smoothed location vector into the above-mentioned initial location deviation information generation model to generate the above-mentioned initial location deviation information.

[0133] In some optional implementations of certain embodiments, the initial item interest information generation model includes: an initial multi-head attention mechanism model and an initial expert network model; and the first generation unit 502 can be configured to: perform data vector transformation on the user data to generate a user vector; perform data vector transformation on the item data to generate an item vector; input the user vector and the item vector into the initial multi-head attention mechanism model to generate an initial user attention vector and an initial item attention vector; and input the initial user attention vector and the initial item attention vector into the initial expert network model to generate the initial item interest information.

[0134] In some optional implementations of some embodiments, the third generation unit 504 may be configured to: multiply the initial position deviation information and the initial position deviation gating information to generate multiplied information; perform a random information discarding operation on the multiplied information to generate discarded information; and add the discarded information and the initial item interest information to generate item interest information.

[0135] In some optional implementations of certain embodiments, the aforementioned item statistics data includes at least one of the following: item exposure information, item click volume information, and item click-through rate information; and the third generation unit 504 can be configured to: vectorize at least one of the aforementioned item exposure information, the aforementioned item click volume information, and the aforementioned item click-through rate information to generate a vector set; vectorize the aforementioned initial item interest information to generate an interest vector; concatenate each vector in the aforementioned vector set and the aforementioned interest vector to generate a concatenated vector; and input the concatenated vector into the aforementioned initial position deviation gating model to generate the aforementioned initial position deviation gating information.

[0136] It is understandable that the units described in the model generation apparatus 500 and the reference Figure 2 The steps in the described method correspond to each other. Therefore, the operations, features, and beneficial effects described above for the method also apply to the model generation device 500 and the units contained therein, and will not be repeated here.

[0137] Further reference Figure 6 As an implementation of the methods shown in the above figures, this disclosure provides some embodiments of an item recommendation device, which are similar to... Figure 4 Corresponding to the method embodiments shown, the recommended device can be specifically applied to various electronic devices.

[0138] like Figure 6 As shown, a model generation device 600 includes: a second acquisition unit 601, a first execution unit 602, a filtering unit 603, and a second execution unit 604. The second acquisition unit 601 is configured to acquire an item dataset and target user data; the first execution unit 602 is configured to perform the following generation steps for each item data in the item dataset: vectorizing the item data to generate an item vector, which serves as the target item vector; vectorizing the target user data to generate a target user vector; inputting the target item vector and the target user vector into a pre-trained item interest information generation model to generate item interest information, wherein the item interest information generation model is generated based on a model generation method; the filtering unit 603 is configured to filter item information to be recommended to the target user from the item information set according to the obtained item interest information set, as recommended item information, to obtain at least one recommended item information, wherein the target user corresponds to the target user data, and the item information in the item information set corresponds to the item data in the item dataset; the second execution unit 604 is configured to perform item recommendation for the at least one recommended item information.

[0139] In some optional implementations of some embodiments, the above-mentioned item interest information generation model includes: a multi-head attention mechanism model and an expert network model; and the first execution unit 602 may be further configured to: input the above-mentioned target item vector and the above-mentioned target user vector into the above-mentioned multi-head attention mechanism model to generate user attention vector and item attention vector; input the above-mentioned user attention vector and the above-mentioned item attention vector into the above-mentioned expert network model to generate the above-mentioned item interest information.

[0140] It is understandable that the units described in the model generation apparatus 600 and the reference Figure 4The steps in the described method correspond accordingly. Therefore, the operations, features, and beneficial effects described above for the method also apply to the model generation device 600 and the units contained therein, and will not be repeated here.

[0141] The following is for reference. Figure 7 It illustrates electronic devices suitable for implementing some embodiments of this disclosure (e.g., Figure 1 A schematic diagram of the structure of electronic device 101)700 in the middle. Figure 7 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of this disclosure.

[0142] like Figure 7 As shown, the electronic device 700 may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) 701, which can perform various appropriate actions and processes according to a program stored in a read-only memory 702 or a program loaded from a storage device 708 into a random access memory 703. The random access memory 703 also stores various programs and data required for the operation of the electronic device 700. The processing unit 701, the read-only memory 702, and the random access memory 703 are interconnected via a bus 704. An input / output interface 705 is also connected to the bus 704.

[0143] Typically, the following devices can be connected to the input / output interface 705: input devices 706 including, for example, a touchscreen, touchpad, keyboard, mouse, camera, microphone, accelerometer, gyroscope, etc.; output devices 707 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 708 including, for example, magnetic tape, hard disk, etc.; and communication devices 709. Communication device 709 allows electronic device 700 to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 7 An electronic device 700 with various devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively. Figure 7 Each box shown can represent a device or multiple devices as needed.

[0144] In particular, according to some embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, some embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 709, or installed from a storage device 708, or installed from a read-only memory 702. When the computer program is executed by the processing device 701, it performs the functions defined in the methods of some embodiments of this disclosure.

[0145] It should be noted that, in some embodiments of this disclosure, the computer-readable medium described above may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In some embodiments of this disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In some embodiments of this disclosure, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0146] In some implementations, clients and servers can communicate using any currently known or future-developed network protocol such as HTTP (Hypertext Transfer Protocol) and can interconnect with digital data communication (e.g., communication networks) of any form or medium. Examples of communication networks include local area networks (“LANs”), wide area networks (“WANs”), the Internet (e.g., the Internet of Things), and peer-to-peer networks (e.g., ad hoc peer-to-peer networks), as well as any currently known or future-developed networks.

[0147] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device. The aforementioned computer-readable medium carries one or more programs that, when executed by the electronic device, cause the electronic device to: acquire training data, wherein the training data includes: user data and item data, wherein the item data includes: item recommendation location information; generate initial item interest information using an initial item interest information generation model based on the user data and the item data; generate initial location deviation information using an initial location deviation information generation model based on the item recommendation location information; generate item interest information for a target item based on the initial item interest information and the initial location deviation information, wherein the target item is an item corresponding to the aforementioned item data; and generate an item interest information generation model and a location deviation information generation model based on the item interest information. Obtain an item dataset and target user data; for each item in the item dataset, perform the following generation steps: vectorize the item data to generate an item vector, which serves as the target item vector; vectorize the target user data to generate a target user vector; input the target item vector and the target user vector into a pre-trained item interest information generation model to generate item interest information, wherein the item interest information generation model is generated based on a model generation method; based on the obtained item interest information set, select item information to be recommended to the target user from the item information set, as recommended item information, to obtain at least one recommended item information, wherein the target user corresponds to the target user data, and the item information in the item information set corresponds to the item data in the item dataset; perform item recommendation for the at least one recommended item information.

[0148] Computer program code for performing operations of some embodiments of this disclosure can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0149] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0150] The units described in some embodiments of this disclosure can be implemented in software or hardware. The described units can also be housed in a processor; for example, a processor may be described as including a first acquisition unit, a first generation unit, a second generation unit, a third generation unit, and a fourth generation unit. The names of these units do not necessarily limit the specific unit; for example, the first acquisition unit may also be described as a "unit for acquiring training data."

[0151] The functions described above in this document can be performed, at least in part, by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: Field Programmable Gate Arrays (FPGAs), Application-Specific Integrated Circuits (ASICs), Application Standard Products (ASSPs), System-on-Chip (SoCs), Complex Programmable Logic Devices (CPLDs), and so on.

[0152] Some embodiments of this disclosure also provide a computer program product, including a computer program that, when executed by a processor, implements any of the model generation methods or item recommendation methods described above.

[0153] The above description is merely a selection of preferred embodiments of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of this disclosure.

Claims

1. A model generation method, comprising: Acquire training data, wherein the training data includes: user data and item data, and the item data includes: item recommendation location information; Based on the user data and the item data, an initial item interest information is generated using an initial item interest information generation model. Based on the recommended location information of the item, a model is generated using the initial location deviation information to generate the initial location deviation information; Based on the initial item interest information and the initial position deviation information, item interest information for the target item is generated, wherein the target item is the item corresponding to the item data; Based on the item interest information, an item interest information generation model and a location deviation information generation model are generated.

2. The method according to claim 1, wherein, The step of generating item interest information for the target item based on the initial item interest information and the initial position deviation information includes: Obtain the item statistics data corresponding to the target item; Based on the item statistics and the initial item interest information, initial position deviation gating information is generated using the initial position deviation gating model; The item interest information is generated based on the initial item interest information, the initial position deviation information, and the initial position deviation gating information.

3. The method according to claim 1, wherein, The step of generating initial position deviation information based on the recommended location information of the item and using the initial position deviation information to generate a model includes: The item recommendation location information is processed by word embedding to generate an item recommendation location vector; The recommended position vector of the item is smoothed to generate a smooth position vector; The smoothed position vector is input into the initial position deviation information generation model to generate the initial position deviation information.

4. The method according to claim 1, wherein, The initial item interest information generation model includes: an initial multi-head attention mechanism model and an initial expert network model; and The step of generating initial item interest information based on the user data and the item data using an initial item interest information generation model includes: The user data is transformed into a data vector to generate a user vector; The item data is transformed into an item vector to generate an item vector; The user vector and the item vector are input into the initial multi-head attention mechanism model to generate the initial user attention vector and the initial item attention vector; The initial user attention vector and the initial item attention vector are input into the initial expert network model to generate the initial item interest information.

5. The method according to claim 2, wherein, The step of generating the item interest information based on the initial item interest information, the initial position deviation information, and the initial position deviation gating information includes: The initial position deviation information and the initial position deviation gating information are multiplied together to generate multiplied information; Perform a random information discarding operation on the multiplied information to generate discarded information; The discarded information and the initial item interest information are added together to generate item interest information.

6. The method according to claim 2, wherein, The item statistics include at least one of the following: item exposure information, item click volume information, and item click-through rate information; as well as The step of generating initial position deviation gating information based on the item statistics data and the initial item interest information, using an initial position deviation gating model, includes: Vectorize at least one of the item exposure information, the item click volume information, and the item click-through rate information to generate a vector set; The initial item interest information is vectorized to generate an interest vector; The vectors in the vector set and the interest vector are concatenated to generate a concatenated vector; The spliced ​​vector is input into the initial position deviation gating model to generate the initial position deviation gating information.

7. A method for recommending items, including: Acquire item datasets and target user data; For each item in the item dataset, perform the following generation steps: The item data is vectorized to generate an item vector, which serves as the target item vector. The target user data is vectorized to generate a target user vector; The target item vector and the target user vector are input into a pre-trained item interest information generation model to generate item interest information, wherein the item interest information generation model is generated based on the method described in any one of claims 1-6; Based on the obtained set of item interest information, item information to be recommended to the target user is filtered out from the item information set and used as recommended item information to obtain at least one recommended item information. The target user corresponds to the target user data, and the item information in the item information set corresponds to the item data in the item dataset. Item recommendation is then performed for the at least one recommended item information.

8. The method according to claim 7, wherein, The item interest information generation model includes: a multi-head attention mechanism model and an expert network model; and The step of inputting the target item vector and the target user vector into a pre-trained item interest information generation model to generate item interest information includes: The target item vector and the target user vector are input into the multi-head attention mechanism model to generate user attention vector and item attention vector; The user attention vector and the item attention vector are input into the expert network model to generate the item interest information.

9. A model generation apparatus, comprising: The first acquisition unit is configured to acquire training data, wherein the training data includes user data and item data, and the item data includes item recommendation location information. The first generation unit is configured to generate initial item interest information based on the user data and the item data, using an initial item interest information generation model; The second generation unit is configured to generate initial position deviation information by generating a model based on the recommended position information of the item and using initial position deviation information. The third generation unit is configured to generate item interest information for a target item based on the initial item interest information and the initial position deviation information, wherein the target item is an item corresponding to the item data; The fourth generation unit is configured to generate an item interest information generation model and a location deviation information generation model based on the item interest information.

10. An item recommendation device, comprising: The second acquisition unit is configured to acquire item datasets and target user data; The first execution unit is configured to perform the following generation steps for each item data in the item dataset: vectorizing the item data to generate an item vector, which serves as the target item vector; vectorizing the target user data to generate a target user vector; and inputting the target item vector and the target user vector into a pre-trained item interest information generation model to generate item interest information, wherein the item interest information generation model is generated based on the method described in any one of claims 1-6. The filtering unit is configured to filter out item information to be recommended to the target user from the item information set based on the obtained item interest information set, and obtain at least one recommended item information, wherein the target user corresponds to the target user data, and the item information in the item information set corresponds to the item data in the item dataset; The second execution unit is configured to perform item recommendations based on the at least one recommended item information.

11. An electronic device, comprising: One or more processors; Storage device, on which one or more programs are stored, When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-8.

12. A computer-readable medium having a computer program stored thereon, wherein, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-8.

13. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1-8.

Images