Article recommendation method, device and equipment

Abstract

The invention discloses an article recommendation method, device and equipment, and belongs to the technical field of computers. Comprising the steps that target implicit features are obtained, the target implicit features are preference features of a target user for multiple articles, and the target implicit features are determined based on interaction behavior data of the target user and the multiple articles; based on a first explicit feature of the target user, and second explicit features and target implicit features of the plurality of articles, determining a plurality of target scores, the first explicit feature being used for representing attribute information of the target user, the second explicit feature being used for representing attribute information of the plurality of articles, and the target implicit feature being used for representing attribute information of the plurality of articles; the plurality of target scores are used for representing preference degrees of the target user to the plurality of articles; and based on the plurality of target scores, performing article recommendation from the plurality of articles. By integrating the explicit features and the implicit features, the preference degree of the target user for the multiple articles can be accurately determined, so that accurate recommendation is performed according to the preference of the target user, and the user experience is improved.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to an item recommendation method, device and equipment. Background Art

[0002] With the development of technology, personalized recommendation systems can provide users with customized product recommendations by analyzing user behavior data, which improves user experience.

[0003] However, users expect to quickly find items that meet their needs, and accurate recommendations can shorten the user's decision-making process and improve purchasing efficiency. Due to the variability of user needs and the diversity of items, the accuracy of item recommendations is poor. Therefore, there is an urgent need for an item recommendation method that can improve the accuracy of recommendations. Summary of the invention

[0004] The present application provides an item recommendation method, apparatus, device and storage medium, which can recommend preferred items to target users based on explicit features of target users and items and implicit features of targets, thereby improving the accuracy of item recommendations and further enhancing user experience. The technical solution is as follows:

[0005] In a first aspect, a method for recommending an item is provided, the method comprising:

[0006] Acquire a target implicit feature, where the target implicit feature is a preference feature of a target user for a plurality of items, and the target implicit feature is determined based on interaction behavior data between the target user and the plurality of items;

[0007] Determining a plurality of target ratings based on a first explicit feature of the target user, a second explicit feature of the plurality of items, and the target implicit feature, wherein the first explicit feature is used to represent the attribute information of the target user, the second explicit feature is used to represent the attribute information of the plurality of items, and the plurality of target ratings are used to represent the degree of preference of the target user for the plurality of items respectively;

[0008] Based on the multiple target ratings, item recommendations are made from the multiple items.

[0009] In the present application, the target implicit feature is first obtained, that is, the target user's preference feature for multiple items is first obtained. Then, based on the first explicit feature of the target user, the second explicit feature of multiple items, and the target implicit feature, multiple target scores are determined, that is, based on the target user's age, gender, hobbies and other attribute information, the brand, price, use and other attribute information of multiple items, and the target user's preference features for multiple items, the target user's preference for multiple items is determined, and then based on multiple target scores, items are recommended from multiple items. Since the attribute information is generally a shallow feature of the target user and multiple items, the implicit feature can indicate the target user's preference for the item due to the characteristics of the item, so that the implicit feature is a deep feature that indicates the target user's preference. By combining explicit features and implicit features, that is, combining features at different levels to jointly determine the target score, the target user's preference for multiple items can be accurately determined, thereby achieving accurate recommendations based on the target user's preferences, which can improve the accuracy of item recommendations and thus improve user experience.

[0010] Optionally, the acquiring target implicit features includes:

[0011] Acquiring interaction behavior data between the target user and the multiple objects;

[0012] Based on the interaction behavior data, determining a first partial order matrix, where the first partial order matrix is ​​used to indicate the preference ranking of the target user for the multiple items;

[0013] Constructing a user feature matrix and an item feature matrix, wherein the user feature matrix is ​​used to represent the implicit features of the target user, and the item feature matrix is ​​used to represent the implicit features of the plurality of items;

[0014] The target implicit feature is determined based on the first partial order matrix, the user feature matrix and the item feature matrix.

[0015] Optionally, the plurality of items have a plurality of item types, and after determining the first partial order matrix based on the interaction behavior data, the method further includes:

[0016] Determining at least one target item type from the plurality of item types based on the type of the target user's vehicle;

[0017] Determining at least one second partial order matrix based on the at least one target item type and the first partial order matrix;

[0018] Optionally, determining the target implicit feature based on the first partially ordered matrix, the user feature matrix, and the item feature matrix includes:

[0019] The target implicit feature is determined based on the at least one second partially ordered matrix, the user feature matrix, and the item feature matrix corresponding to the at least one target item type.

[0020] Optionally, determining the target implicit feature based on the at least one second partially ordered matrix, the user feature matrix, and the item feature matrix corresponding to the at least one target item type includes:

[0021] For any one of the at least one target item type, multiplying the user feature matrix by a transposed matrix of an item feature matrix corresponding to the target item type to obtain a reference ranking matrix corresponding to the target item type;

[0022] Based on the difference between the reference sorting matrix and the second partially ordered matrix corresponding to the target item type, the user feature matrix and the item feature matrix corresponding to the target item type are updated, and the updated item feature matrix is ​​the reference implicit feature corresponding to the target item type;

[0023] In the case where there is a target item type, determining the reference implicit feature as the target implicit feature;

[0024] When there are at least two target item types, the reference implicit features corresponding to the at least two target item types are concatenated to obtain the target implicit feature.

[0025] Optionally, determining a plurality of target scores based on the first explicit feature of the target user, the second explicit features of the plurality of items, and the target implicit feature comprises:

[0026] Performing feature fusion on the first explicit feature and the second explicit feature to obtain a first sequence feature;

[0027] Performing sequence encoding on the first sequence feature to obtain a second sequence feature;

[0028] Performing attention encoding on the second sequence feature and the target implicit feature to obtain a target attention feature;

[0029] Based on the second sequence feature and the target attention feature, the multiple target scores are determined.

[0030] Optionally, performing attention encoding on the second sequence feature and the target implicit feature to obtain the target attention feature includes:

[0031] Concatenating the second sequence feature with the target implicit feature to obtain a third sequence feature;

[0032] Based on the multi-head attention mechanism, the third sequence feature is attention-encoded to obtain multiple reference attention features;

[0033] The multiple reference attention features are concatenated to obtain the target attention feature.

[0034] Optionally, determining the multiple target scores based on the second sequence feature and the target attention feature includes:

[0035] Combining the second sequence feature and the target attention feature to obtain a fourth sequence feature;

[0036] The fourth sequence feature is activated to obtain the multiple target scores.

[0037] Optionally, recommending items from the multiple items based on the multiple target scores includes:

[0038] Based on the multiple target scores, determining multiple target items from the multiple items, the multiple target items being items among the multiple items whose target scores are greater than a preset score threshold;

[0039] The multiple target objects are displayed in descending order according to the target scores.

[0040] In a second aspect, an item recommendation device is provided, the device comprising:

[0041] An acquisition module, used for acquiring a target implicit feature, wherein the target implicit feature is a preference feature of a target user for a plurality of items, and the target implicit feature is determined based on interaction behavior data between the target user and the plurality of items;

[0042] a determination module, configured to determine a plurality of target ratings based on a first explicit feature of the target user, a second explicit feature of the plurality of items, and the target implicit feature, wherein the first explicit feature is used to represent attribute information of the target user, the second explicit feature is used to represent attribute information of the plurality of items, and the plurality of target ratings are used to represent the degree of preference of the target user for the plurality of items respectively;

[0043] A recommendation module is used to recommend items from the multiple items based on the multiple target scores.

[0044] Optionally, the acquisition module is used to:

[0045] Acquiring interaction behavior data between the target user and the multiple objects;

[0046] Based on the interaction behavior data, determining a first partial order matrix, where the first partial order matrix is ​​used to indicate the preference ranking of the target user for the multiple items;

[0047] Constructing a user feature matrix and an item feature matrix, wherein the user feature matrix is ​​used to represent the implicit features of the target user, and the item feature matrix is ​​used to represent the implicit features of the plurality of items;

[0048] The target implicit feature is determined based on the first partial order matrix, the user feature matrix and the item feature matrix.

[0049] Optionally, the multiple items have multiple item types, and the acquisition module is further used to:

[0050] Determining at least one target item type from the plurality of item types based on the type of the target user's vehicle;

[0051] Determining at least one second partial order matrix based on the at least one target item type and the first partial order matrix;

[0052] Optionally, the acquisition module is used to:

[0053] The target implicit feature is determined based on the at least one second partially ordered matrix, the user feature matrix, and the item feature matrix corresponding to the at least one target item type.

[0054] Optionally, the acquisition module is used to:

[0055] For any one of the at least one target item type, multiplying the user feature matrix by a transposed matrix of an item feature matrix corresponding to the target item type to obtain a reference ranking matrix corresponding to the target item type;

[0056] Based on the difference between the reference sorting matrix and the second partially ordered matrix corresponding to the target item type, the user feature matrix and the item feature matrix corresponding to the target item type are updated, and the updated item feature matrix is ​​the reference implicit feature corresponding to the target item type;

[0057] In the case where there is a target item type, determining the reference implicit feature as the target implicit feature;

[0058] When there are at least two target item types, the reference implicit features corresponding to the at least two target item types are concatenated to obtain the target implicit feature.

[0059] Optionally, the determining module is used to:

[0060] Performing feature fusion on the first explicit feature and the second explicit feature to obtain a first sequence feature;

[0061] Performing sequence encoding on the first sequence feature to obtain a second sequence feature;

[0062] Performing attention encoding on the second sequence feature and the target implicit feature to obtain a target attention feature;

[0063] Based on the second sequence feature and the target attention feature, the multiple target scores are determined.

[0064] Optionally, the determining module is used to:

[0065] Concatenating the second sequence feature with the target implicit feature to obtain a third sequence feature;

[0066] Based on the multi-head attention mechanism, the third sequence feature is attention-encoded to obtain multiple reference attention features;

[0067] The multiple reference attention features are concatenated to obtain the target attention feature.

[0068] Optionally, the determining module is used to:

[0069] Combining the second sequence feature and the target attention feature to obtain a fourth sequence feature;

[0070] The fourth sequence feature is activated to obtain the multiple target scores.

[0071] Optionally, the recommendation module is used to:

[0072] Based on the multiple target scores, determining multiple target items from the multiple items, the multiple target items being items among the multiple items whose target scores are greater than a preset score threshold;

[0073] The multiple target objects are displayed in descending order according to the target scores.

[0074] In a third aspect, a computer device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program implements the above-mentioned item recommendation method when executed by the processor.

[0075] In a fourth aspect, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned item recommendation method is implemented.

[0076] In a fifth aspect, a computer program product comprising instructions is provided, which, when executed on a computer, enables the computer to execute the steps of the above-mentioned item recommendation method.

[0077] It can be understood that the beneficial effects of the second, third, fourth and fifth aspects mentioned above can be found in the relevant description of the first aspect mentioned above, and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] 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 ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0079] Figure 1 is a scenario diagram of an item recommendation method provided in an embodiment of the present application;

[0080] Figure 2 is a flow chart of an item recommendation method provided in an embodiment of the present application;

[0081] Figure 3 It is a flow chart of constructing a behavior feedback matrix provided in an embodiment of the present application;

[0082] Figure 4 It is a flowchart of a method for determining implicit features of a target provided in an embodiment of the present application;

[0083] Figure 5 It is a flowchart of a multi-head attention mechanism provided in an embodiment of the present application;

[0084] Figure 6 is a model schematic diagram of a target scoring model provided in an embodiment of the present application;

[0085] Figure 7 is a structural schematic diagram of an item recommendation device provided in an embodiment of the present application;

[0086] Figure 8 It is a structural diagram of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0087] In order to make the objectives, technical solutions and advantages of the present application clearer, the implementation methods of the present application will be further described in detail below with reference to the accompanying drawings.

[0088] It should be understood that the "multiple" mentioned in this application refers to two or more. In the description of this application, unless otherwise specified, " / " means or, for example, A / B can mean A or B; "and / or" in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships, for example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. In addition, in order to facilitate the clear description of the technical solution of this application, the words "first" and "second" are used to distinguish between the same items or similar items with basically the same functions and effects. Those skilled in the art can understand that the words "first" and "second" do not limit the quantity and execution order, and the words "first" and "second" do not limit them to be different.

[0089] Before describing the item recommendation method provided in the embodiment of the present application, the application scenario of the embodiment of the present application is first described.

[0090] For example, Figure 1 is a schematic diagram of a scenario of an item recommendation method provided in an embodiment of the present application, see Figure 1 , Figure 1 The product browsing interface 101 includes a product browsing interface 101 , wherein a plurality of items 102 are displayed in the product browsing interface 101 .

[0091] For example, if a user wants to buy car supplies, the user can open a car supplies APP (Application), which is specifically used to sell car supplies. After the user opens the car supplies APP, the car supplies APP can display a product browsing interface 101, and the product browsing interface 101 can display multiple items 102, and the user can browse the items 102 on the product browsing interface 101.

[0092] Generally, the display order of the multiple items 102 on the product browsing interface 101 can be preset. In this case, when the user enters the product browsing interface 101 to browse, it may take a long time to see the items he wants to buy. Or the user needs to search for the items he wants to buy in the search box.

[0093] Therefore, when users want to purchase items, it will not be convenient enough and the experience will be poor.

[0094] To this end, an embodiment of the present application provides an item recommendation method, which can recommend corresponding items to users based on user preferences, thereby providing convenience for users to purchase items, increasing users' purchase rate, and further improving user experience.

[0095] Specifically, the target implicit features are first obtained, that is, the target user's preference features for multiple items are first obtained. The target implicit features can be determined based on the interaction behavior data between the target user and multiple items. Then, based on the target user's explicit features (attribute information), the explicit features (attribute information) of multiple items, and the target implicit features, the target user's preference for multiple items is first determined. Then, based on the target user's preference for multiple items, the target user's preferred items are determined from the multiple items and recommended to the target user.

[0096] Since attribute information is generally shallow-level features of the target user and multiple items, implicit features can indicate the characteristics of the item that the target user prefers, so implicit features are deep-level features that indicate the target user's preferences. By combining explicit and implicit features, that is, combining features at different levels to jointly determine the target score, the target user's preference for multiple items can be accurately determined, thereby achieving accurate recommendations based on the target user's preferences, which can improve the accuracy of item recommendations and thus enhance user experience.

[0097] The item recommendation method provided in the embodiment of the present application is explained in detail below.

[0098] Figure 2 is a flowchart of an item recommendation method provided in an embodiment of the present application. Figure 2 , the method comprises the following steps.

[0099] Step 201: Obtain target implicit features, where the target implicit features are preference features of a target user for a plurality of items, and the target implicit features are determined based on interaction behavior data between the target user and the plurality of items.

[0100] The multiple items may be multiple vehicle supplies. In an embodiment of the present application, the multiple items may also be items displayed on a recommendation page in a vehicle supplies APP.

[0101] The target user's interactive behavior data with the multiple items may include the target user's browsing data, collection data, shopping cart data, purchase data, etc. of the multiple items, that is, whether the target user has purchased the multiple items, browsed the multiple items, collected the multiple items, added some of the multiple items to the shopping cart, etc. Of course, the target user's interactive behavior data with the multiple items may also include the target user's sharing data for the multiple items, etc., which is not limited in the embodiments of the present application.

[0102] Implicit features are features that are not easy to observe directly, but have an important impact on prediction and decision-making. For example, there are some potential factors that make the target user buy an item, which cannot be represented by the basic explicit features. Therefore, the target implicit features can mine the deep features of the target user's preference for the item.

[0103] In this case, by obtaining the target implicit features, we can learn the target user's preference features for the multiple items, that is, we can find out which potential features of the items make the target user prefer the items. In this way, we can subsequently recommend items based on the target user's preferences for the multiple items, thereby improving the accuracy of item recommendations.

[0104] Specifically, the operation of determining the target implicit features may include the following steps (1) to (4).

[0105] (1) Obtain the interaction behavior data between the target user and the multiple objects.

[0106] Optionally, the interaction behavior data between the target user and the multiple objects may be recorded by means of data embedding.

[0107] Data embedding is a data collection method. In the embodiment of the present application, its main function is to monitor the operations of the target user, so as to realize the recording of the interactive behavior data between the target user and the multiple items. The specific process of data embedding refers to that the technician writes the written recording code into the vehicle supplies APP, and sets the click events for each menu to be available. Therefore, when the target user subsequently clicks on the link of an item to browse the item, or clicks on the favorite menu to collect an item, or clicks on the add to cart menu to add an item to the shopping cart, or clicks on the purchase menu to purchase an item, the target user's various operations on the item can be recorded.

[0108] In this way, the interaction behavior data between the target user and the multiple objects can be easily obtained.

[0109] (2) Based on the interaction behavior data, determine a first partial order matrix.

[0110] The first partial order matrix is ​​used to indicate the target user's preference ranking for the multiple items.

[0111] Optionally, the operation of step (2) may be: based on the interaction behavior data, constructing a behavior feedback matrix of the target user for the multiple items; based on the behavior feedback matrix, determining a first partial order matrix.

[0112] The target user's behavior feedback matrix for the multiple items may include positive feedback and negative feedback. Positive feedback is used to indicate that the target user has an interactive behavior with an item, that is, the target user has purchased, browsed, or collected the item, etc. Negative feedback is used to indicate that the target user has no interactive behavior with an item.

[0113] Table 1 below is an example of a behavior feedback matrix. Referring to Table 1, Table 1 includes the behavior feedback of the target user on multiple items, wherein the target user shows positive feedback on some items and negative feedback on some items. Among them, V1, V2, V3, and V4 represent 4 items, "+" represents positive feedback, and "-" represents negative feedback.

[0114] Table 1

[0115] V1 V2 V3 V4 …… Target Users + + - + ……

[0116] The embodiment of the present application only uses the above Table 1 as an example to exemplify the behavior feedback matrix of the target user for the multiple items, and does not constitute a limitation to the embodiment of the present application.

[0117] Among them, based on the behavioral interaction behavior data, the operation of constructing the target user's behavioral feedback matrix for the multiple items can be: for any one of the multiple items, if there is interactive behavior data of the target user for this item in the behavioral interaction data, the feedback data of the target user for this item in the behavioral feedback matrix is ​​recorded as positive feedback; if there is no interactive behavior data of the target user for this item in the behavioral interaction data, the feedback data of the target user for this item in the behavioral feedback matrix is ​​recorded as negative feedback.

[0118] For example, Figure 3 is a flow chart of constructing a behavior feedback matrix provided in an embodiment of the present application, see Figure 3 , Figure 3 The behavior interaction data 301 and the behavior feedback matrix 302 are included.

[0119] The behavior interaction data 301 records the number of interactions between the target user and the multiple items, and the behavior feedback matrix 302 determines whether the target user gives positive feedback or negative feedback to an item based on the behavior interaction data.

[0120] Among them, based on the behavior feedback matrix, the operation of determining the first partial order matrix can be: recording the positive feedback in the behavior feedback matrix as the first value in the first partial order matrix, and recording the negative feedback in the behavior matrix as the second value in the first partial order matrix.

[0121] The first value and the second value can be preset, for example, the first value can be set to 1 and the second value can be set to 0. Wherein, when the first value is greater than the second value, the target user has a greater preference for items with a matrix value of the first value than for items with a matrix value of the second value.

[0122] For example, the following Table 2 is an example of a first partial order matrix. Referring to the following Table 2, it can be known that the target user prefers item V1, item V2, and item V4.

[0123] Table 2

[0124] V1 V2 V3 V4 …… Target Users 1 1 0 1 ……

[0125] The embodiment of the present application only uses the above Table 2 as an example to exemplify the first partial order matrix, and does not constitute a limitation to the embodiment of the present application.

[0126] (3) Construct user feature matrix and item feature matrix.

[0127] The user feature matrix is ​​used to represent the implicit features of the target user, that is, the potential preference features of the target user, etc. The item feature matrix is ​​used to represent the implicit features of the multiple items, specifically, the implicit features of the target user's preferences for the multiple items. It should be understood that the item feature matrix may include the implicit features of each of the multiple items.

[0128] For the first partial order matrix, we actually get an incomplete feedback matrix because the target user does not interact with some items. This can mean that the target user has shown a preference for some items, but does not mean that the target user has no preference for other items. Therefore, in the embodiment of the present application, the first partial order matrix can be approximated by the product of two small matrices.

[0129] The two small matrices can represent the implicit features of the target user and the implicit features of the target user's preferences for the multiple items. In this case, the two small matrices are multiplied to obtain the first partial order matrix for representing the target user's preference ranking for the multiple items. Therefore, in the embodiment of the present application, the user feature matrix and the item feature matrix can be constructed first.

[0130] In this case, it is equivalent to first initializing a user feature matrix and an item feature matrix to represent the implicit features of the target user and the implicit features of the target user's preferences for the multiple items. Subsequently, the user feature matrix and the item feature matrix are optimized to determine the item feature matrix and the user feature matrix that meet the conditions.

[0131] For example, if the target user has k implicit features, and each of the multiple items also has k implicit features, then the user feature matrix A can be constructed. 1×k =[a 1 , a 2 , a 3 , ..., ak], the item feature matrix is

[0132] (4) Determine the target implicit features based on the first partial order matrix, the user feature matrix and the item feature matrix.

[0133] Since the multiplication of two small matrices can obtain the first partial order matrix, in the embodiment of the present application, the purpose is to make the product of the transposed matrix of the user feature matrix and the item feature matrix close to the first partial order matrix. Based on this principle, a set of user feature matrices and item feature matrices can be determined, and the product of the transposed matrix of the determined user feature matrix and item feature matrix is ​​close to the first partial order matrix, which means that the determined user feature matrix and item feature matrix are relatively accurate, so that relatively accurate target implicit features can be obtained.

[0134] In this case, it is equivalent to decomposing the implicit features of the target user and the multiple items by matrix decomposition, and then obtaining the implicit features of the target user's preferences for the multiple items.

[0135] Specifically, the operation of step (4) can be: multiplying the user feature matrix by the transposed matrix of the item feature matrix to obtain a candidate sorting matrix; based on the difference between the first partial order matrix and the candidate sorting matrix, updating the user feature matrix and the item feature matrix, and the updated item feature matrix is ​​the target implicit feature.

[0136] Since our goal is to make the product of the user feature matrix multiplied by the transposed matrix of the item feature matrix approximate the first partial order matrix, the user feature matrix and the item feature matrix can be updated based on the difference between the first partial order matrix and the candidate ranking matrix.

[0137] In addition, since the first partial order matrix can represent the target user's preference ranking for the multiple items, the user feature matrix and the item feature matrix are updated based on the difference between the first partial order matrix and the candidate ranking matrix, which is equivalent to making the preference ranking predicted based on the implicit features of the target user and the implicit features of the target user's preferences for the multiple items close to the target user's true preference ranking for the multiple items. In this way, when the candidate ranking matrix is ​​close to the first partial order matrix, the user feature matrix and the item feature matrix that meet the conditions can be obtained, then a more accurate user feature matrix and item feature matrix can be determined at this time, so that a more accurate target implicit feature can be determined.

[0138] The difference between the first partial order matrix and the candidate sorting matrix can be realized by the following formula (1).

[0139]

[0140] Where y is the difference between the first partial order matrix and the candidate sorting matrix, which is also the loss value in the training and learning process, and m is the number of items in the item feature matrix, which is also the number of columns in the candidate sorting matrix and the first partial order matrix. i is the i-th matrix value in the candidate sorting matrix, is the matrix value of the corresponding position in the first partial order matrix.

[0141] In this case, the difference between the first partial order matrix and the candidate sorting matrix is ​​characterized by calculating the mean square error between the first partial order matrix and the candidate sorting matrix, and then the user feature matrix and the item feature matrix can be updated based on the difference.

[0142] Furthermore, after knowing the difference between the first partial order matrix and the candidate sorting matrix, it is also necessary to determine how to update the user feature matrix and the item feature matrix. The purpose of updating the user feature matrix and the item feature matrix is ​​to make the difference between the first partial order matrix and the candidate sorting matrix smaller. In the embodiment of the present application, the user feature matrix can be updated by the following formula (2), and the item feature matrix can be updated by the following formula (3).

[0143] A′ 1×k =A 1×k +ηyH T m×k (2)

[0144] H T ' m×k =H T m×k +ηyA 1×k (3)

[0145] Among them, A′ 1×k is the updated user feature matrix, H T ' m×k is the transposed matrix of the updated item feature matrix, that is, the updated item feature matrix is ​​H′ m×k =(H T ' m×k ) T η is the learning rate, which is used to indicate the step size of the update.

[0146] Optionally, after constructing the behavior feedback matrix, a partial order relationship dataset of items of the target user may be constructed. Then, the user feature matrix and the item feature matrix may be updated based on the partial order relationship dataset of items of the target user through a Bayesian personalized sorting method.

[0147] The target user's item partial order relationship dataset may include the target user's preferences for the multiple items.

[0148] Bayesian personalized ranking assumes that the target user prefers items that have been interacted with more than items that have not been interacted with. Then, by maximizing the probability of this preference, the target user's implicit preference for the multiple items can be learned during the training process, that is, a more accurate item feature matrix can be obtained.

[0149] Since the behavior feedback matrix includes positive feedback and negative feedback, the items for which the target user has given positive feedback can be grouped into a set Items that have negative feedback for target users can be grouped into a set In this case, the partial order relationship dataset of items for the target user can be expressed as

[0150] In the embodiment of the present application, the partial order of different items by the target user is independent of each other, so the preference of the target user between positive feedback items and negative feedback items is independent of other items. Based on this, the Bayesian personalized sorting method can cater to this feature to update the matrix, so as to obtain a more accurate user feature matrix and item feature matrix.

[0151] Optionally, in the embodiment of the present application, the multiple items may have multiple item types.

[0152] In the case where the plurality of articles are vehicle articles, the plurality of article types may include off-road articles, commercial vehicle articles, multi-purpose passenger vehicle articles, and the like.

[0153] In the case that the plurality of items have a plurality of item types, at least one second partial order matrix may also be determined based on the first partial order matrix.

[0154] Specifically, based on the type of the target user's vehicle, at least one target item type is determined from the multiple item types; based on the at least one target item type and the first partial order matrix, at least one second partial order matrix is ​​determined.

[0155] In an embodiment of the present application, the target user can fill in the vehicle information about the vehicle he is interested in in the vehicle supplies APP, including the vehicle type of the vehicle.

[0156] The target item type corresponds to the type of the target user's vehicle. For example, if the target user's vehicle type is an off-road vehicle, then the target item type may be an off-road item. Of course, when the target user is interested in at least one vehicle type, the target item type may also be at least one. For example, if the target user's vehicle type is an off-road vehicle and a multi-purpose passenger vehicle, then the target item type may also include an off-road item and a multi-purpose passenger item.

[0157] At least one second partial order matrix corresponds to at least one target item type, that is, the second partial order matrix may include the preference ranking of the target user for items of the target item type in the first partial order matrix.

[0158] In this case, by determining at least one target item type and subsequently determining at least one second partial order matrix based on it, the preference ranking of items of the item type that the target user is interested in among the multiple items can be obtained. In this way, the preferences of the target user can be captured and simulated more accurately in the future.

[0159] In addition, since the second partial order matrix corresponds to the target item type, that is to say, the dimension of the second partial order matrix is ​​reduced, then when determining the user feature matrix and item feature matrix later, the number of items to be processed is reduced, and the amount of calculation is less, which simplifies the difficulty and calculation complexity of matrix decomposition, thereby making the calculation faster.

[0160] Moreover, based on the above, only the items without interactive behaviors are considered in the item types that the target user is concerned about, rather than all items, so that the focus is more on the items that the target user is more likely to be interested in, reducing the interference of irrelevant items. Therefore, this optimization strategy can improve the recommendation efficiency and accuracy, and further improve the efficiency of determining the target implicit features.

[0161] In this case, the operation of step (4) may be: determining the target implicit feature based on at least one second partial order matrix, the user feature matrix and at least one item feature matrix corresponding to the target item type.

[0162] Since the target implicit feature is the target user's preference feature for the multiple items, then based on at least one second partial order matrix, the user feature matrix and the item feature matrix corresponding to at least one target item type, it is equivalent to determining the target implicit feature based on the preference ranking of items of the item type that the target user is concerned about, which is equivalent to filtering the items that the target user does not prefer, which can improve the accuracy of determining the target implicit feature.

[0163] Specifically, based on at least one second partial order matrix, a user feature matrix and at least one item feature matrix corresponding to a target item type, the operation of determining the target implicit feature may include the following steps (a) to (d).

[0164] (a) For any one of the at least one target item type, multiply the user feature matrix by the transposed matrix of the item feature matrix corresponding to the target item type to obtain a reference ranking matrix corresponding to the target item type.

[0165] The reference ranking matrix can be considered as the target user's preference ranking for items corresponding to the target item type predicted based on the user feature matrix and the item feature matrix.

[0166] In this case, it is equivalent to constructing at least one item feature matrix, and at least one item feature matrix corresponds to at least one target item type. Then, the transposed matrix of at least one item feature matrix is ​​multiplied by the user feature matrix to obtain at least one reference sorting matrix, that is, to obtain a reference sorting matrix corresponding to at least one target item type.

[0167] (b) Based on the difference between the reference sorting matrix and the second partial order matrix corresponding to the target item type, the user feature matrix and the item feature matrix corresponding to the target item type are updated, and the updated item feature matrix is ​​the reference implicit feature corresponding to the target item type.

[0168] After determining a reference ranking matrix corresponding to at least one target item type, for each reference ranking corresponding to the target item type, the user feature matrix and the item feature matrix corresponding to the target item type can be updated based on the difference between the reference ranking corresponding to the target item type and the second partial order matrix corresponding to the target item type.

[0169] Since the preferences of the target user for items of different item types are independent, and the preferences of the target user for items of the same item type are correlated, for each of the at least one target item type, the item feature matrix corresponding to the corresponding target item type can be updated respectively. In this way, the reference implicit feature corresponding to the at least one target item type can be accurately determined.

[0170] In an embodiment of the present application, the operation of updating the user feature matrix and the item feature matrix corresponding to the target item type based on the difference between the reference sorting matrix and the second partial order matrix corresponding to the target item type is similar to the operation of updating the user feature matrix and the item feature matrix based on the difference between the first partial order matrix and the candidate sorting matrix, and will not be repeated here.

[0171] (c) In the presence of a target item type, the reference implicit feature is determined as the target implicit feature.

[0172] When there is a target item type, a reference sorting matrix is ​​obtained. In this case, the user feature matrix and the constructed item feature matrix are updated based on the difference between the reference sorting matrix and the second partial order matrix corresponding to the target item type. Thus, the updated item feature matrix (reference implicit feature) can be directly used as the target implicit feature.

[0173] (d) When there are at least two target item types, feature concatenation is performed on reference implicit features corresponding to the at least two target item types to obtain the target implicit feature.

[0174] When there are at least two target item types, at least two reference sorting matrices will be obtained. In this case, the user feature matrix and the item feature matrix corresponding to the corresponding target item type are updated based on the differences between the at least two reference sorting matrices and the second partial order matrices corresponding to the corresponding target item types, respectively, so that at least two reference implicit features can be obtained after the update.

[0175] However, the at least two reference implicit features are the preference features of the target user for the items of the item type concerned, so the at least two reference implicit features can be retained. Then in the embodiment of the present application, the reference implicit features corresponding to the at least two target item types can be feature spliced.

[0176] In this way, by concatenating the reference implicit features corresponding to the at least two target item types, the target user's preference features for items of the item type of interest can be completely preserved, so that more accurate item recommendations can be achieved later.

[0177] Combine the following Figure 4 The method for determining the target implicit feature provided in the embodiment of the present application is exemplified. For example, Figure 4 is a flow chart of a method for determining a target implicit feature provided in an embodiment of the present application, see Figure 4 , Figure 4 The process includes steps 401 to 408.

[0178] Step 401: Obtain interaction behavior data between the target user and the multiple objects.

[0179] Step 402: Based on the interactive behavior data, construct a behavior feedback matrix of the target user for the multiple items.

[0180] Step 403: Determine a first partial order matrix based on the behavior feedback matrix.

[0181] Step 404: Construct a user feature matrix and an item feature matrix.

[0182] Step 405: Multiply the user feature matrix by the item feature matrix to obtain a candidate ranking matrix.

[0183] Step 406: Based on the difference between the first partial order matrix and the candidate ranking matrix, the user feature matrix and the item feature matrix are updated.

[0184] Step 407: Determine whether the loss value of the update process is less than a preset threshold. If the loss value of the update process is less than the preset threshold, execute the following step 408. If the loss value of the update process is greater than or equal to the preset threshold, return to execute the above step 406.

[0185] Step 408: Determine the updated item feature matrix as the target implicit feature.

[0186] It is worth noting that the embodiment of the present application can determine the target implicit features through the above step 201, that is, it can determine the target user's preference features for the multiple items, and then corresponding item recommendations can be made based on this, that is, continue to execute the following steps 202 and 203.

[0187] Step 202: Based on the first explicit feature of the target user, the second explicit feature of the multiple items and the target implicit feature, multiple target ratings are determined, wherein the first explicit feature is used to represent the attribute information of the target user, the second explicit feature is used to represent the attribute information of the multiple items, and the multiple target ratings are used to represent the target user's preference degree for the multiple items respectively.

[0188] The first explicit feature refers to a feature of the target user that can be directly observed. In the embodiment of the present application, the first explicit feature can represent the attribute information of the target user. For example, the first explicit feature can include information such as the target user's age, gender, hobbies, and driving habits.

[0189] The second explicit feature refers to a feature of the multiple items that can be directly observed. In the embodiment of the present application, the second explicit feature can represent the attribute information of the multiple items, for example, the second explicit feature can include the item category, item brand, item price, item purpose, item material, etc. of the multiple items.

[0190] The multiple target ratings correspond to the multiple items. For any one of the multiple target ratings, the target rating may represent the target user's preference for the corresponding item.

[0191] The target implicit features can represent the target user's preference features for the multiple items, that is, what features of the items make the target user prefer the items. They are the deep-level features of the target user's preferences for the multiple items learned through training on the target user's interaction behavior data with the multiple items.

[0192] In this case, the attribute information is generally the shallow features of the target user and multiple items, and the implicit features can indicate the characteristics of the item that the target user prefers, so the implicit features are the deep features that indicate the target user's preference. By combining explicit features and implicit features, that is, combining features at different levels to jointly determine the target score, the target user's preference for multiple items can be accurately determined.

[0193] Specifically, the operation of step 202 may include the following steps (1) to (4).

[0194] (1) The first explicit feature and the second explicit feature are fused to obtain a first sequence feature.

[0195] Since the first explicit feature and the second explicit feature are explicit features of different objects, that is, the first explicit feature belongs to the target user, and the second explicit feature belongs to the multiple items. Therefore, in the subsequent feature processing process, the two features are processed separately, which will increase the computational complexity. Therefore, the first explicit feature and the second explicit feature can be fused first.

[0196] In this case, by fusing the first explicit feature and the second explicit feature, the first sequence feature is subsequently processed, thereby reducing the computational complexity.

[0197] Optionally, the operation of step (1) may be: concatenating the first explicit feature and the second explicit feature to obtain a first sequence feature.

[0198] In this case, by concatenating the first explicit feature and the second explicit feature, the explicit features of the target user and the multiple items are completely retained in the first sequence features obtained subsequently, so the explicit features of the target user and the multiple items can be completely learned subsequently.

[0199] (2) The first sequence feature is sequence-encoded to obtain a second sequence feature.

[0200] In the embodiment of the present application, the operation of step (2) may be: performing word embedding on the first sequence feature to obtain the second sequence feature.

[0201] Word embedding is a sequence encoding method that aims to convert each character into a fixed-length vector that captures the semantic and grammatical information of the word.

[0202] Since the explicit features of the multiple items of the target user are discrete, one-hot encoding is required before processing the discrete data. After one-hot encoding, the feature values ​​of most positions of the discrete data are 0, which will cause data sparsity. Therefore, the first sequence feature can be word embedded to map the data in the first sequence feature to a continuous vector space, so that each character can be mapped to a unique vector representation, thereby obtaining the second sequence feature.

[0203] In this way, by embedding the first sequence features, the second sequence features are obtained, so that the explicit features can be processed more conveniently later and useful feature representations can be better learned.

[0204] Optionally, the operation of performing word embedding on the first sequence feature to obtain the second sequence feature may be: obtaining an embedding matrix; and obtaining a vector of each character in the first sequence feature from the embedding matrix, thereby obtaining the second sequence feature.

[0205] The embedding matrix is ​​a parameter matrix, which is a matrix containing all the vocabulary learned in advance. Each column of this matrix represents the vector represented by different words. Therefore, the vector of each character in the first sequence feature can be obtained from the embedding matrix.

[0206] In this case, by obtaining the embedding matrix, the second sequence features can be obtained more conveniently.

[0207] (3) Perform attention encoding on the second sequence features and the target implicit features to obtain the target attention features.

[0208] Attention coding refers to the use of attention mechanism to extract important features from the second sequence features and target implicit features, thereby obtaining features that play a key role in determining the scores of multiple targets. Attention coding can include self-attention coding, channel attention coding, multi-head attention coding, etc.

[0209] In this case, by performing attention encoding on the second sequence features and the target implicit features, features that play a key role in determining the target score can be obtained, so that a more accurate target score can be determined later based on this feature.

[0210] As an implementation method, the operation of step (3) can be: feature concatenating the second sequence feature and the target implicit feature to obtain a third sequence feature; based on a multi-head attention mechanism, performing attention encoding on the third sequence feature to obtain multiple reference attention features; feature concatenating the multiple reference attention features to obtain a target attention feature.

[0211] In order to facilitate subsequent processing, the second sequence features and the target implicit features can be concatenated to obtain a third sequence feature, which can make it easier to process the features later.

[0212] The multi-head attention mechanism captures different aspects of information through multiple attention heads. Figure 5 The multi-head attention mechanism provided in the embodiment of the present application is described. For example, Figure 5 It is a flow chart of a multi-head attention mechanism provided in an embodiment of the present application.

[0213] like Figure 5 As shown, the multi-head attention mechanism includes multiple heads, each of which receives the third sequence feature, and each head can represent the third sequence feature in a different subspace, so that the multiple heads can focus on different aspects of the third sequence feature, that is, each head has a different focus on the third sequence feature, which makes it possible to learn the relationship between the features in the third sequence feature from different angles.

[0214] Secondly, the multi-head attention mechanism can extract attention features independently through each head, so that multiple heads can subsequently process the third sequence features simultaneously in a parallel processing manner to extract different aspects of the attention features in the third sequence features, that is, to pay attention to different aspects of the features simultaneously in a parallel processing manner.

[0215] Finally, after each head outputs the corresponding reference attention features, the reference attention features of each part are merged into a unified feature for output. The usual practice is to perform feature splicing on the reference attention features output by each head, that is, to integrate the various information learned by different heads to form a final target attention feature.

[0216] In this case, the third sequence feature is encoded by the multi-head attention feature to obtain the target attention feature, so that the feature relationship of different aspects in the third sequence feature can be paid attention to, and the complex relationship in the data can be better captured, so that the key features for determining the target score can be paid attention to from different aspects. In this way, a more accurate target score can be determined based on this feature later.

[0217] In addition, the multi-head attention feature focuses on key features in different aspects through parallel processing, which can improve the efficiency of determining the target score while improving the accuracy of the target score.

[0218] Optionally, the operation of performing feature splicing on multiple reference attention features to obtain a target attention feature may be: obtaining splicing weights corresponding to multiple heads; and performing feature splicing on multiple reference attention features with corresponding splicing weights to obtain a target attention feature.

[0219] Since each head uses a different subspace to represent the third sequence feature, in the embodiment of the present application, the splicing weights of different heads can also be set. The splicing weight is used to indicate the importance of the reference attention features output by different heads. The larger the splicing weight, the more important the reference attention feature output by a head is, that is, the greater the impact on the target score. The smaller the splicing weight, the smaller the importance of the reference attention feature output by a head is, that is, the smaller the impact on the target score is.

[0220] In this case, by splicing multiple reference attention features with corresponding splicing weights, the attention features that have a greater impact on the target score can be highlighted, so that a more accurate target attention feature can be obtained, and then a more accurate target score can be obtained based on this.

[0221] (4) Determine multiple target scores based on the second sequence features and the target attention features.

[0222] The target attention features include features that have an important influence on determining the target score. In this case, based on the second sequence features and the target attention features, multiple relatively accurate target scores can be determined.

[0223] Specifically, the operation of step (4) may be: merging the second sequence feature and the target attention feature to obtain a fourth sequence feature; and activating the fourth sequence feature to obtain multiple target scores.

[0224] Since the second sequence feature is an explicit feature of the target user and the multiple items, it can indicate the global feature of the target user and the multiple items. The target attention feature is an attention feature determined by shallow-level features (global features) and deep-level features (local features (target implicit features)).

[0225] In this case, by merging the target attention feature with the second sequence feature, the key features of the global features of the target user and the multiple items can be highlighted. In this way, the global features can also be paid attention to when determining the target score later, thereby improving the accuracy of determining multiple target scores.

[0226] Optionally, the operation of merging the second sequence feature and the target attention feature to obtain the fourth sequence feature may be: performing a weighted summation of the second sequence feature and the target attention feature to obtain the fourth sequence feature.

[0227] The weights of the second sequence features and the target attention features can represent the degree of influence of the second sequence features and the target attention features on the target score. Optionally, the weights of the second sequence features and the target attention features can be obtained through pre-learning and training. Of course, they can also be set in advance by technical personnel. The embodiments of the present application are not limited to this.

[0228] In this case, by weighted summing the second sequence features and the target attention features, they can be merged accordingly according to the degree of influence on the target score, thereby obtaining a more accurate fourth sequence feature.

[0229] Optionally, the fourth sequence feature can be activated by a softmax function to obtain multiple target scores. Of course, the fourth sequence feature can also be activated by other activation functions to obtain multiple target scores, which is not limited in the embodiment of the present application.

[0230] It is worth noting that the process of the above steps (1) to (4) is equivalent to inputting the first explicit feature, the second explicit feature and the target implicit feature into a target scoring model, and processing the first explicit feature, the second explicit feature and the target implicit feature through the target scoring model to output multiple target scores.

[0231] For example, Figure 6 is a model diagram of a target scoring model, see Figure 6 ,The target scoring model can include a hybrid input layer, an embedding layer, an attention fusion layer, and an output layer.

[0232] The mixed input layer is used to input the first explicit feature, the second explicit feature and the target implicit feature.

[0233] The embedding layer is used to perform sequence encoding on the fusion feature (first sequence feature) of the first explicit feature and the second explicit feature to obtain the second sequence feature.

[0234] The attention fusion layer is used to encode the second sequence features and target implicit features based on the multi-head attention mechanism to obtain the target attention features.

[0235] The output layer is used to determine multiple target scores based on the target attention features and the second sequence features.

[0236] It is worth noting that through the above step 202, multiple target ratings can be determined based on the first explicit feature, the second explicit feature and the target implicit feature, that is, the target user's preference for the multiple items can be predicted. Then, based on the target user's preference for the multiple items, the preferred items can be recommended to the target user from the multiple items, that is, continue to execute the following step 203.

[0237] Step 203: Recommend items from the multiple items based on the multiple target scores.

[0238] In this case, multiple target ratings can indicate the target user's preference for the multiple items, so we can also know which items the target user may want to buy. Based on the multiple target ratings, we can recommend items that the target user may buy, thereby achieving accurate recommendations for items.

[0239] Optionally, the operation of step 203 may be: based on multiple target scores, determining multiple target items from the multiple items; and displaying the multiple target items in descending order of the target scores.

[0240] The multiple target items are items among the multiple items whose target scores are greater than a preset score threshold. In other words, the multiple target items are items that the target user may purchase.

[0241] The preset scoring threshold can be set in advance, and the preset scoring threshold can be set larger.

[0242] In this case, the target score corresponding to an item is greater than the preset score threshold, which means that the larger the target score corresponding to the item, the greater the target user's preference for the item, that is, the target user is likely to purchase the item. Therefore, the item can be determined as a target item and displayed.

[0243] In the embodiment of the present application, by determining multiple target items, it is equivalent to determining the items that the target user may purchase from the multiple items. Then, the multiple target items are displayed in descending order according to the target score, so that the items that the target user is most likely to purchase can be displayed at the front, so that the target user can see the items that are most likely to purchase at the first time. In this way, accurate recommendations can be made to the target user, and the target user may not need to search for items to see the items they want to buy, which provides convenience for the target user to purchase items, increases the purchase rate of the target user, and thus improves the user experience.

[0244] Optionally, the operation of step 203 may be: based on multiple target scores, determining multiple target items from the multiple items; based on the item types of the multiple target items, determining the display order of the target items under each item type; first displaying the target items under the target item type; and then displaying the target items under the remaining item types except the target item type.

[0245] Since the target item type is the item type corresponding to the type of vehicle that the target user is interested in, that is, the target user is more interested in items under the target item type, the target items under the target item type are displayed first, and then the target items under other item types are displayed.

[0246] In this case, by sorting the display order of the target items by item type, the target items under the target item type are displayed first, and then the target items under the remaining item types are displayed, so that the items that the target user is more concerned about can be displayed first. In other words, the items that the target user is more concerned about may be the items that the target user wants to buy, so the purchase rate of the target user can be increased.

[0247] It is worth noting that the item recommendation method provided in the embodiment of the present application can optimize the order of recommended items in the recommendation list. The optimized recommendation list can better meet the actual needs of the target users, and can increase the click-through rate and purchase rate of the target users, thereby bringing higher conversion rates and revenues to the enterprise. Secondly, by providing more personalized and accurate recommendations, the target users' loyalty and activity to the vehicle supplies APP can be enhanced.

[0248] In the embodiment of the present application, the computer device first obtains the target implicit feature, that is, first obtains the target user's preference feature for multiple items. Then, based on the first explicit feature of the target user, the second explicit feature of the multiple items, and the target implicit feature, multiple target scores are determined, that is, based on the target user's age, gender, hobbies and other attribute information, the brand, price, use and other attribute information of the multiple items, and the target user's preference features for multiple items, the target user's preference for multiple items is determined, and then based on the multiple target scores, items are recommended from multiple items. Since the attribute information is generally a shallow feature of the target user and multiple items, the implicit feature can indicate the target user's preference for the item due to the characteristics of the item, so the implicit feature is a deep feature that indicates the target user's preference. By combining explicit features and implicit features, that is, combining features at different levels to jointly determine the target score, the target user's preference for multiple items can be accurately determined, thereby achieving accurate recommendations based on the target user's preferences, which can improve the accuracy of item recommendations and thus improve user experience.

[0249] Figure 7is a schematic diagram of the structure of an item recommendation device provided in an embodiment of the present application. The item recommendation device can be implemented by software, hardware, or a combination of both to form part or all of a computer device, and the computer device can be the following Figure 8 Computer equipment shown. Figure 7 The device includes: an acquisition module 701, a first determination module 702 and a recommendation module 703.

[0250] An acquisition module 701 is used to acquire a target implicit feature, where the target implicit feature is a preference feature of a target user for a plurality of items, and the target implicit feature is determined based on interaction behavior data between the target user and the plurality of items;

[0251] A determination module 702 is used to determine a plurality of target ratings based on a first explicit feature of the target user, a second explicit feature of the plurality of items, and a target implicit feature, wherein the first explicit feature is used to represent attribute information of the target user, the second explicit feature is used to represent attribute information of the plurality of items, and the plurality of target ratings are used to represent the target user's preference for the plurality of items respectively;

[0252] The recommendation module 703 is configured to recommend items from the multiple items based on the multiple target scores.

[0253] Optionally, the acquisition module 701 is used to:

[0254] Obtaining interaction behavior data between the target user and the multiple items;

[0255] Based on the interaction behavior data, determining a first partial order matrix, where the first partial order matrix is ​​used to indicate the preference ranking of the target user for the multiple items;

[0256] Constructing a user feature matrix and an item feature matrix, wherein the user feature matrix is ​​used to represent the implicit features of the target user, and the item feature matrix is ​​used to represent the implicit features of the multiple items;

[0257] Based on the first partial order matrix, the user feature matrix and the item feature matrix, the target implicit features are determined.

[0258] Optionally, the multiple items have multiple item types, and the acquisition module 701 is further used to:

[0259] determining at least one target item type from the plurality of item types based on the type of the target user's vehicle;

[0260] Determining at least one second partial order matrix based on at least one target item type and the first partial order matrix;

[0261] Optionally, the acquisition module 701 is used to:

[0262] The target implicit feature is determined based on at least one second partial order matrix, the user feature matrix, and at least one item feature matrix corresponding to the target item type.

[0263] Optionally, the acquisition module 701 is used to:

[0264] For any one of the at least one target item type, multiplying the user feature matrix by the transposed matrix of the item feature matrix corresponding to the target item type to obtain a reference ranking matrix corresponding to the target item type;

[0265] Based on the difference between the reference sorting matrix and the second partial order matrix corresponding to the target item type, the user feature matrix and the item feature matrix corresponding to the target item type are updated, and the updated item feature matrix is ​​the reference implicit feature corresponding to the target item type;

[0266] In the presence of a target item type, the reference implicit feature is determined as the target implicit feature;

[0267] When there are at least two target item types, the reference implicit features corresponding to the at least two target item types are concatenated to obtain the target implicit features.

[0268] Optionally, the determination module 702 is used to:

[0269] The first explicit feature and the second explicit feature are subjected to feature fusion to obtain a first sequence feature;

[0270] Perform sequence encoding on the first sequence feature to obtain a second sequence feature;

[0271] Perform attention encoding on the second sequence features and the target implicit features to obtain the target attention features;

[0272] Based on the second sequence feature and the target attention feature, a plurality of target scores are determined.

[0273] Optionally, the determination module 702 is used to:

[0274] The second sequence feature and the target implicit feature are concatenated to obtain the third sequence feature;

[0275] Based on the multi-head attention mechanism, the third sequence features are encoded with attention to obtain multiple reference attention features;

[0276] Multiple reference attention features are concatenated to obtain the target attention feature.

[0277] Optionally, the determination module 702 is used to:

[0278] The second sequence feature and the target attention feature are combined to obtain the fourth sequence feature;

[0279] The fourth sequence feature is activated to obtain multiple target scores.

[0280] Optionally, the recommendation module 703 is used to:

[0281] Based on the multiple target scores, determining multiple target items from the multiple items, the multiple target items being items among the multiple items whose target scores are greater than a preset score threshold;

[0282] Display multiple target items in descending order according to their target scores.

[0283] In the embodiment of the present application, the target implicit feature is first obtained, that is, the target user's preference feature for multiple items is first obtained. Then, based on the first explicit feature of the target user, the second explicit feature of the multiple items, and the target implicit feature, multiple target scores are determined, that is, based on the target user's age, gender, hobbies and other attribute information, the brand, price, purpose and other attribute information of the multiple items, and the target user's preference features for the multiple items, the target user's preference for the multiple items is determined, and then based on the multiple target scores, items are recommended from the multiple items. Since the attribute information is generally a shallow feature of the target user and the multiple items, the implicit feature can indicate the target user's preference for the item due to the characteristics of the item, so the implicit feature is a deep feature that indicates the target user's preference. By combining the explicit features and the implicit features, that is, combining the features of different levels to jointly determine the target score, the target user's preference for the multiple items can be accurately determined, thereby achieving accurate recommendations based on the target user's preferences, which can improve the accuracy of item recommendations and thus improve user experience.

[0284] It should be noted that: when the item recommendation device provided in the above embodiment recommends items to the target user, it only uses the division of the above-mentioned functional modules as an example. In actual applications, the above-mentioned functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above.

[0285] The functional units and modules in the above embodiments may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit, and the above integrated units may be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the protection scope of the embodiments of the present application.

[0286] The item recommendation device and the item recommendation method embodiments provided in the above embodiments belong to the same concept. The specific working process of the units and modules in the above embodiments and the technical effects brought about can be found in the method embodiment part and will not be repeated here.

[0287] Figure 8 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present application. Figure 8 As shown, the computer device 8 includes: a processor 80, a memory 81, and a computer program 82 stored in the memory 81 and executable on the processor 80. When the processor 80 executes the computer program 82, the steps in the item recommendation method in the above embodiment are implemented.

[0288] The computer device 8 may be a general-purpose computer device or a dedicated computer device. In a specific implementation, the computer device 8 may be a terminal device such as a desktop computer, a portable computer, a PDA, a mobile phone, a tablet computer, etc. The embodiment of the present application does not limit the type of the computer device 8. Those skilled in the art will understand that Figure 8 It is only an example of the computer device 8 and does not constitute a limitation on the computer device 8. It may include more or fewer components than shown in the figure, or a combination of certain components, or different components, such as input and output devices, network access devices, etc.

[0289] The processor 80 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0290] In some embodiments, the memory 81 may be an internal storage unit of the computer device 8, such as a hard disk or memory of the computer device 8. In other embodiments, the memory 81 may also be an external storage device of the computer device 8, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the computer device 8. Further, the memory 81 may also include both an internal storage unit and an external storage device of the computer device 8. The memory 81 is used to store an operating system, an application program, a boot loader, data, and other programs. The memory 81 may also be used to temporarily store data that has been output or is to be output.

[0291] An embodiment of the present application further provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments can be implemented.

[0292] An embodiment of the present application provides a computer program product, which, when executed on a computer, enables the computer to execute the steps in the above-mentioned method embodiments.

[0293] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present application implements all or part of the processes in the above method embodiments, which can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a computer-readable storage medium. When the computer program is executed by the processor, the steps of the above method embodiments can be implemented. Among them, the computer program includes computer program code, which can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may at least include: any entity or device that can carry the computer program code to the camera / terminal device, recording medium, computer memory, ROM (Read-Only Memory), RAM (Random Access Memory), CD-ROM (Compact Disc Read-Only Memory), magnetic tape, floppy disk and optical data storage device. The computer-readable storage medium mentioned in the present application can be a non-volatile storage medium, in other words, it can be a non-transient storage medium.

[0294] It should be understood that all or part of the steps to implement the above embodiments can be implemented by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. The computer instructions can be stored in the above-mentioned computer readable storage medium.

[0295] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0296] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0297] In the embodiments provided in the present application, it should be understood that the disclosed devices / computer equipment and methods can be implemented in other ways. For example, the device / computer equipment embodiments described above are only schematic, for example, the division of modules or units is only a logical function division, and there may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0298] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0299] The embodiments described above are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. An item recommendation method, characterized in that: The method comprises: Acquire a target implicit feature, where the target implicit feature is a preference feature of a target user for a plurality of items, and the target implicit feature is determined based on interaction behavior data between the target user and the plurality of items; Determining a plurality of target ratings based on a first explicit feature of the target user, a second explicit feature of the plurality of items, and the target implicit feature, wherein the first explicit feature is used to represent attribute information of the target user, the second explicit feature is used to represent attribute information of the plurality of items, and the plurality of target ratings are used to represent the target user's preference for the plurality of items respectively; Based on the multiple target ratings, item recommendations are made from the multiple items.

2. The method according to claim 1, characterized in that The acquiring of target implicit features comprises: Acquiring interaction behavior data between the target user and the multiple objects; Based on the interaction behavior data, determining a first partial order matrix, where the first partial order matrix is ​​used to indicate the preference ranking of the target user for the multiple items; Constructing a user feature matrix and an item feature matrix, wherein the user feature matrix is ​​used to represent the implicit features of the target user, and the item feature matrix is ​​used to represent the implicit features of the plurality of items; The target implicit feature is determined based on the first partial order matrix, the user feature matrix and the item feature matrix.

3. The method according to claim 2, characterized in that The plurality of items have a plurality of item types. After determining the first partial order matrix based on the interaction behavior data, the method further includes: Determining at least one target item type from the plurality of item types based on the type of the target user's vehicle; Determining at least one second partial order matrix based on the at least one target item type and the first partial order matrix; The determining the target implicit feature based on the first partial order matrix, the user feature matrix and the item feature matrix includes: The target implicit feature is determined based on the at least one second partially ordered matrix, the user feature matrix, and the item feature matrix corresponding to the at least one target item type.

4. The method according to claim 3, characterized in that The determining the target implicit feature based on the at least one second partially ordered matrix, the user feature matrix, and the item feature matrix corresponding to the at least one target item type includes: For any one of the at least one target item type, multiplying the user feature matrix by a transposed matrix of an item feature matrix corresponding to the target item type to obtain a reference ranking matrix corresponding to the target item type; Based on the difference between the reference sorting matrix and the second partially ordered matrix corresponding to the target item type, the user feature matrix and the item feature matrix corresponding to the target item type are updated, and the updated item feature matrix is ​​the reference implicit feature corresponding to the target item type; In the case where there is a target item type, determining the reference implicit feature as the target implicit feature; When there are at least two target item types, the reference implicit features corresponding to the at least two target item types are concatenated to obtain the target implicit feature.

5. The method according to claim 1, characterized in that The determining of a plurality of target scores based on the first explicit feature of the target user, the second explicit features of the plurality of items, and the target implicit feature comprises: Performing feature fusion on the first explicit feature and the second explicit feature to obtain a first sequence feature; Performing sequence encoding on the first sequence feature to obtain a second sequence feature; Performing attention encoding on the second sequence feature and the target implicit feature to obtain a target attention feature; Based on the second sequence feature and the target attention feature, the multiple target scores are determined.

6. The method according to claim 5, characterized in that The step of performing attention encoding on the second sequence feature and the target implicit feature to obtain the target attention feature includes: Concatenating the second sequence feature with the target implicit feature to obtain a third sequence feature; Based on the multi-head attention mechanism, the third sequence feature is attention-encoded to obtain multiple reference attention features; The multiple reference attention features are concatenated to obtain the target attention feature.

7. The method according to claim 5, characterized in that The determining of the plurality of target scores based on the second sequence feature and the target attention feature comprises: Combining the second sequence feature and the target attention feature to obtain a fourth sequence feature; The fourth sequence feature is activated to obtain the multiple target scores.

8. The method according to claim 1, characterized in that The recommending items from the multiple items based on the multiple target scores includes: Based on the multiple target scores, determining multiple target items from the multiple items, the multiple target items being items among the multiple items whose target scores are greater than a preset score threshold; The multiple target objects are displayed in descending order according to the target scores.

9. An item recommendation device, characterized in that: The device comprises: A first acquisition module is used to acquire a target implicit feature, wherein the target implicit feature is a preference feature of a target user for a plurality of items, and the target implicit feature is determined based on interaction behavior data between the target user and the plurality of items; a first determination module, configured to determine a plurality of target ratings based on a first explicit feature of the target user, a second explicit feature of the plurality of items, and the target implicit feature, wherein the first explicit feature is used to represent attribute information of the target user, the second explicit feature is used to represent attribute information of the plurality of items, and the plurality of target ratings are used to represent the degree of preference of the target user for the plurality of items respectively; A recommendation module is used to recommend items from the multiple items based on the multiple target scores.

10. A computer device, characterized in that: The computer device comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program implements the method according to any one of claims 1 to 8 when executed by the processor.

Images