A combination object recommendation method, device, system and storage medium

Abstract

The present disclosure relates to a combination object recommendation method, device, system and storage medium. In at least one embodiment of the present disclosure, when recommending a combination object, different object types are distinguished, a corresponding recommendation model is established for each object type, and then for a recommended user, a recommendation ranking list of objects in each object type is obtained by using the corresponding recommendation model of each object type. It can be seen that, since different objects of the same object type have the same attribute characteristics, there is no case of empty feature value, so that the training features of the recommendation model become dense from sparse, which is beneficial to accelerate the training convergence of the recommendation model. Moreover, the redundant features are removed, the noise interference is reduced, and the recommendation ranking result of the recommendation model is more accurate. In addition, by selecting objects from the corresponding recommendation ranking list of each object type, different recommendation ranking lists are fused into a combination object recommendation result, improving the effectiveness and diversity of object recommendation.

Description

TECHNICAL FIELD

[0001] Embodiments of the present disclosure relate to the technical field of machine learning, in particular to a combined object recommendation method, device, system and storage medium. BACKGROUND

[0002] With the increasing maturity of machine learning algorithms in industrial applications, more and more companies begin to build their own product recommendation systems to meet the needs of different users for personalized product recommendations, so that users can more conveniently obtain their favorite products, so that products can obtain higher click rates in the exposure and display stage, improve client user activity, and thus improve product purchase rates.

[0003] In the product recommendation system, machine learning needs to be applied to collect the attribute features of the recommended products as part of the training features of the model. When recommending different types of products (i.e., combined product recommendation), since different types of products have different features, the attribute features of multiple types of products are usually directly concatenated as training features.

[0004] However, when a product does not have the features possessed by the remaining products, the feature values will be taken as empty. The empty values of the multi-dimensional feature dimensions will cause the product feature vector to become sparse, which not only increases the calculation and storage of the model, but also introduces unnecessary noise, reducing the effect of the model. Therefore, the traditional single model recommendation method has poor effect when recommending combined products. SUMMARY

[0005] To solve at least one problem existing in the prior art, at least one embodiment of the present disclosure provides a combined object recommendation method, device, system and storage medium.

[0006] In a first aspect, embodiments of the present disclosure provide a combined object recommendation method, wherein the combined object includes a plurality of objects, and the plurality of objects belong to at least two object types, and the method comprises:

[0007] establishing a corresponding recommendation model for each object type;

[0008]

[0008] For a recommended user, using the recommendation model corresponding to each object type to obtain a recommended ranking list of the objects in each object type, and selecting objects from the recommended ranking list corresponding to each object type to obtain a combined object recommendation result.

[0009] In some embodiments, before establishing a corresponding recommendation model for each object type, the method further comprises:

[0010] obtaining one or more attribute features of each object;

[0011] All objects with the same number of attributes and the same attribute definitions are classified into the same object type.

[0012] In some embodiments, establishing a corresponding recommendation model for each object type includes:

[0013] Obtain a collection of user behavior data, where each piece of user behavior data describes a specific action performed by a user on an object;

[0014] The user behavior data set is divided into datasets corresponding to different object types based on the different object types involved;

[0015] Based on datasets of different object types, construct training samples for each object type;

[0016] Based on the model training samples for each object type, a preset model is trained to obtain the recommendation model corresponding to each object type.

[0017] In some embodiments, each piece of user behavior data includes: user ID, object ID, behavior occurrence time, and set of exposed objects;

[0018] The model training samples include: user ID, object ID, time of behavior occurrence, and label.

[0019] In some embodiments, for a recommended user, obtaining the recommended ranking list of objects in each object type using the recommendation model corresponding to each object type includes:

[0020] The matching score between the object ID and the recommended user ID in each object type is output using the recommendation model corresponding to each object type;

[0021] Sort the matching scores of object IDs and recommended user IDs in each object type from high to low to obtain a recommended sorted list of objects in each object type.

[0022] In some embodiments, selecting objects from the recommendation ranking lists corresponding to each object type to obtain a combined object recommendation result includes:

[0023] Obtain the behavioral data set of the recommended users;

[0024] Based on the behavioral data set of the recommended users, determine the user type of the recommended users;

[0025] Based on the user type of the recommended user, objects are selected from the recommendation ranking list corresponding to each object type to obtain the combined object recommendation result.

[0026] In some embodiments, determining the user type of the recommended user based on the behavioral data set of the recommended user includes:

[0027] determine an object type to which each object in which the behavior occurs in the behavior data set of the recommended user belongs, and count the number of object types;

[0028] determine a user type of the recommended user based on the number of object types.

[0029] In some embodiments, the selecting the objects from the recommendation ranking list corresponding to each object type to obtain the combined object recommendation result based on the user type of the recommended user comprises:

[0030] obtain one or more attribute feature data of each user in each user type;

[0031] determine one or more statistical feature data of each user based on the behavior data of each user in each user type;

[0032] select the objects from the recommendation ranking list corresponding to each object type to obtain the combined object recommendation result based on the one or more attribute feature data of each user, the one or more statistical feature data of each user, and the user type of the recommended user.

[0033] In some embodiments, the selecting the objects from the recommendation ranking list corresponding to each object type to obtain the combined object recommendation result based on the one or more attribute feature data of each user, the one or more statistical feature data of each user, and the user type of the recommended user comprises:

[0034] determine at least two feature mean vectors based on the one or more attribute feature data of each user and the one or more statistical feature data of each user, and the number of the at least two feature mean vectors is the same as the number of the at least two object types;

[0035] determine a distance between the recommended user and each feature mean vector;

[0036] determine a preference probability of each object type of the recommended user based on the distance between the recommended user and each feature mean vector;

[0037] select the objects from the recommendation ranking list corresponding to each object type multiple times based on the preference probability of each object type of the recommended user to obtain the combined object recommendation result.

[0038] In some embodiments, the selecting the objects from the recommendation ranking list corresponding to each object type multiple times based on the preference probability of each object type of the recommended user comprises:

[0039] for each time of selection:

[0040] generating one or more random probabilities, and based on the one or more random probabilities, filtering a preference probability from the preference probabilities of all object types, and taking an object from the recommended ranking list of the object type corresponding to the filtered preference probability.

[0041] In some embodiments, generating one or more random probabilities, and based on the one or more random probabilities, filtering a preference probability from the preference probabilities of all object types comprises:

[0042] generating a random probability;

[0043] filtering the preference probability greater than the random probability to obtain a filtered set;

[0044] determining whether the number in the filtered set is greater than 1;

[0045] if greater than 1, generating a random probability again, and updating the filtered set;

[0046] repeating the steps of determining, generating again, and updating until the number in the filtered set is equal to 1, and stopping generating the random probability.

[0047] In a second aspect, the embodiments of the present disclosure further provide a combined object recommendation device, wherein the combined object comprises a plurality of objects, and the plurality of objects belong to at least two object types, and the device comprises:

[0048] a establishing unit configured to establish a corresponding recommendation model for each object type;

[0049] a recommendation unit configured to obtain a recommended ranking list of the objects in each object type by using the corresponding recommendation model of each object type for a recommended user, and select objects from the recommended ranking list of each object type to obtain a combined object recommendation result.

[0050] In some embodiments, the device further comprises:

[0051] an obtaining unit configured to obtain one or more attribute features of each object;

[0052] a dividing unit configured to divide all objects with the same number of attribute features and the same definition of attribute features into the same object type;

[0053] the establishing unit is configured to establish a corresponding recommendation model for each object type after the dividing unit divides the object types.

[0054] In some embodiments, the establishing unit is configured to:

[0055] obtain a set of user behavior data, wherein each piece of user behavior data is data describing a specific behavior of a user on an object;

[0056] dividing the user behavior data set into data sets corresponding to different object types according to the different object types involved;

[0057] constructing a model training sample for each object type based on the data set of the different object types;

[0058] training a preset model based on the model training sample of each object type to obtain a recommendation model corresponding to each object type.

[0059] In some embodiments, each piece of user behavior data includes: a user ID, an object ID, a behavior occurrence time, and a set of exposed objects.

[0060] The model training sample includes: a user ID, an object ID, a behavior occurrence time, and a label.

[0061] In some embodiments, the recommendation unit obtains a recommendation ranking list of objects in each object type for a recommended user using the recommendation model corresponding to each object type respectively includes:

[0062] The recommendation unit outputs a matching score of an object ID in each object type and a recommended user ID using the recommendation model corresponding to each object type.

[0063] The recommendation unit sorts the matching scores of the object IDs in each object type and the recommended user ID from high to low to obtain a recommendation ranking list of objects in each object type.

[0064] In some embodiments, the recommendation unit selects objects from the recommendation ranking list corresponding to each object type to obtain a combined object recommendation result includes:

[0065] The recommendation unit obtains a behavior data set of a recommended user.

[0066] The recommendation unit determines a user type of the recommended user based on the behavior data set of the recommended user.

[0067] The recommendation unit selects objects from the recommendation ranking list corresponding to each object type to obtain a combined object recommendation result based on the user type of the recommended user.

[0068] In some embodiments, the recommendation unit determines a user type of the recommended user based on the behavior data set of the recommended user includes:

[0069] The recommendation unit determines the object type to which each object in the behavior data set of the recommended user belongs, and counts the number of object types.

[0070] The recommendation unit determines the user type of the recommended user based on the number of object types.

[0071] In some embodiments, the recommendation unit selects objects from the recommendation ranking list corresponding to each object type to obtain the combined object recommendation result based on the user type of the recommended user includes:

[0072] The recommendation unit obtains one or more attribute feature data of each user in each user type;

[0073] The recommendation unit determines one or more statistical feature data of each user based on the behavior data of each user in each user type;

[0074] The recommendation unit selects objects from the recommendation ranking list corresponding to each object type to obtain the combined object recommendation result based on the one or more attribute feature data of each user, the one or more statistical feature data of each user, and the user type of the recommended user.

[0075] In some embodiments, the recommendation unit selects objects from the recommendation ranking list corresponding to each object type to obtain the combined object recommendation result based on the one or more attribute feature data of each user, the one or more statistical feature data of each user, and the user type of the recommended user includes:

[0076] The recommendation unit determines at least two feature mean vectors based on the one or more attribute feature data of each user and the one or more statistical feature data of each user, and the number of the at least two feature mean vectors is the same as the number of the at least two object types;

[0077] The recommendation unit determines the distance between the recommended user and each feature mean vector;

[0078] The recommendation unit determines the preference probability of the recommended user for each object type based on the distance between the recommended user and each feature mean vector;

[0079] The recommendation unit selects objects from the recommendation ranking list corresponding to each object type multiple times based on the preference probability of the recommended user for each object type to obtain the combined object recommendation result.

[0080] In some embodiments, the recommendation unit selects objects from the recommendation ranking list corresponding to each object type multiple times based on the preference probability of the recommended user for each object type includes:

[0081] For each selection:

[0082] The recommendation unit generates one or more random probabilities, and selects a preference probability from all the preference probabilities of the object types based on the one or more random probabilities, and selects an object from the recommendation ranking list of the object type corresponding to the selected preference probability.

[0083] In some embodiments, the recommendation unit generates one or more random probabilities, and filters a preference probability from the preference probabilities of all object types based on the one or more random probabilities, including:

[0084] generating a random probability;

[0085] filtering the preference probability greater than the random probability to obtain a filtered set;

[0086] determining whether the number in the filtered set is greater than 1;

[0087] if greater than 1, generating a random probability again, and updating the filtered set;

[0088] repeating the steps of determining, generating again, and updating until the number in the filtered set is equal to 1, and stopping generating the random probability.

[0089] In a third aspect, the embodiments of the present disclosure further provide a system including at least one computing device and at least one storage device storing instructions, wherein the instructions, when executed by the at least one computing device, cause the at least one computing device to perform the steps of the combined object recommendation method according to any embodiment of the first aspect.

[0090] In a fourth aspect, the embodiments of the present disclosure further provide a non-transitory computer-readable storage medium for storing a program or instructions, which, when executed by at least one computing device, cause the at least one computing device to perform the steps of the combined object recommendation method according to any embodiment of the first aspect.

[0091] As can be seen, in at least one embodiment of the present disclosure, when recommending a combined object, different object types are distinguished, a corresponding recommendation model is established for each object type, and then for a recommended user, a recommendation ranking list of objects in each object type is obtained by using the corresponding recommendation model of each object type. As can be seen, since different objects of the same object type have the same attribute characteristics, there is no case of empty feature values, so that the training features of the recommendation model become dense from sparse, which is beneficial to accelerate the training convergence of the recommendation model and reduce the calculation amount and storage amount of the recommendation model.

[0092] Moreover, the recommendation model corresponds to different objects of the same object type, and when training the recommendation model, the redundant features of objects not belonging to the object type are removed, noise interference is reduced, and then the recommendation effect is improved, so that the recommendation ranking result of the recommendation model is more accurate.

[0093] In addition, different recommendation ranking lists are fused into a combined object recommendation result by selecting objects from the recommendation ranking lists corresponding to each object type, so that the multi-model result can be used at the same time, the objects of multiple types can be effectively ranked, and the objects of multiple types are combined for recommendation, thereby improving the effectiveness and diversity of object recommendation. BRIEF DESCRIPTION OF DRAWINGS

[0094] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present disclosure, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0095] Figure 1 is an exemplary flowchart of a combined object recommendation method provided by an embodiment of the present disclosure;

[0096] Figure 2 is an exemplary block diagram of a combined object recommendation device provided by an embodiment of the present disclosure;

[0097] Figure 3 is an exemplary block diagram of a system including at least one computing device and at least one storage device storing instructions provided by an embodiment of the present disclosure. DETAILED DESCRIPTION

[0098] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present disclosure, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0099] It should be noted that, in this paper, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations.

[0100] For the combined object recommendation scenario, there are usually multiple types of objects, including real goods and virtual goods. Real goods are tangible goods such as clothing, electrical appliances, and wearable goods, and virtual goods are intangible goods such as videos, music, content, and services (take-out). Among them, clothing and electrical appliances are different types of objects. The same type of object may also include subtypes, for example, clothing includes two subtypes of clothes and trousers.

[0101] However, in the traditional single model recommendation method, when recommending a combination object, for example, the combination object is a sports shoe and a refrigerator, the attribute features of the sports shoe include size, and the attribute features of the refrigerator include volume, in the training of the single model, the attribute features of multiple types of objects are directly spliced as training features, for example, the size of the sports shoe and the volume of the refrigerator are spliced, since the sports shoe does not have volume, the volume of the sports shoe is empty, similarly, the size of the refrigerator is empty, it can be seen that the multi-dimensional feature dimension value is empty, which will make the feature vector sparse, which not only increases the calculation and storage of the model, but also introduces some unnecessary noise, for example, the sports shoe has the volume feature itself is a kind of noise, which reduces the effect of the model.

[0102] Therefore, for the combination object recommendation scene, the embodiment of the present disclosure provides a combination object recommendation method, device, system and storage medium. In at least one embodiment of the present disclosure, when recommending a combination object, different object types are distinguished, a corresponding recommendation model can be established for each object type, and then for a recommended user, the recommendation ranking list of the objects in each object type is obtained by using the corresponding recommendation model of each object type. It can be seen that since different objects of the same object type have the same attribute features, there is no case of empty feature value, which makes the training features of the recommendation model from sparse to dense, which is beneficial to speed up the training convergence of the recommendation model and reduce the calculation and storage of the recommendation model.

[0103] In addition, the recommendation model corresponds to different objects of the same object type, and when training the recommendation model, the redundant features of the objects that do not belong to the object type are removed, the noise interference is reduced, and the recommendation effect is improved, so that the recommendation ranking result of the recommendation model is more accurate.

[0104] In addition, by selecting objects from the recommendation ranking list corresponding to each object type, the different recommendation ranking lists are fused into a combination object recommendation result, so that the multi-model result can be used at the same time, the multiple types of objects can be effectively sorted, and the multiple types of objects are combined and recommended, thereby improving the effectiveness and diversity of object recommendation.

[0105] Figure 1 An exemplary flowchart of a combination object recommendation method provided by the embodiment of the present disclosure is provided, wherein the combination object includes multiple objects, and the multiple objects belong to at least two object types. For example, the combination object is a sports shoe and a refrigerator, the sports shoe belongs to the object type "shoes", and the refrigerator belongs to the object type "refrigerator". As shown in Figure 1 The combination object recommendation method includes steps 101 and 102:

[0106] In step 101, a corresponding recommendation model is established for each object type.

[0107] For example, a corresponding shoe recommendation model is established for the object type of "shoes", and during training of the shoe recommendation model, different objects (sports shoes, leather shoes, high-heeled shoes, etc.) of the object type of "shoes" have the same attribute features (size, brand, material, etc.), and there is no case of empty feature values, so that the training features of the shoe recommendation model become dense from sparse, which is beneficial to accelerate the training convergence of the shoe recommendation model and reduce the calculation amount and storage amount of the shoe recommendation model.

[0108] In addition, the shoe recommendation model corresponds to different objects of the object type of "shoes", and during training of the shoe recommendation model, the redundant features of objects not belonging to "shoes" are removed, noise interference is reduced, and the recommendation effect is further improved, so that the recommendation ranking result of the shoe recommendation model is more accurate.

[0109] For example, a corresponding refrigerator recommendation model is established for the object type of "refrigerator", and during training of the refrigerator recommendation model, different objects (stand-type refrigerator, horizontal-type refrigerator, table-type refrigerator, etc.) of the object type of "refrigerator" have the same attribute features (capacity, cooling method, shape, purpose, etc.), and there is no case of empty feature values, so that the training features of the refrigerator recommendation model become dense from sparse, which is beneficial to accelerate the training convergence of the refrigerator recommendation model and reduce the calculation amount and storage amount of the refrigerator recommendation model.

[0110] In addition, the refrigerator recommendation model corresponds to different objects of the object type of "refrigerator", and during training of the refrigerator recommendation model, the redundant features of objects not belonging to "refrigerator" are removed, noise interference is reduced, and the recommendation effect is further improved, so that the recommendation ranking result of the refrigerator recommendation model is more accurate.

[0111] In step 102, for a recommended user, the recommendation ranking list of the objects in each object type is obtained by using the corresponding recommendation model of each object type, and the combined object recommendation result is obtained by selecting objects from the recommendation ranking list corresponding to each object type.

[0112] For example, the shoe recommendation model is used to obtain the recommendation ranking list of the objects in the object type of "shoes", and the refrigerator recommendation model is used to obtain the recommendation ranking list of the objects in the object type of "refrigerator". After obtaining the recommendation ranking list of "shoes" and the recommendation ranking list of "refrigerator", the objects can be selected from the two recommendation ranking lists, and the two recommendation ranking lists are fused into a combined object recommendation result, so that the two recommendation ranking lists can be used simultaneously, and the objects of the object type of "shoes" and the objects of the object type of "refrigerator" can be effectively ranked, thereby combining the objects of the object types of "shoes" and "refrigerator" for recommendation, and improving the effectiveness and diversity of object recommendation.

[0113] In some embodiments, the present disclosure provides a manner of dividing object types, according to attribute features possessed by objects. Specifically, before establishing a corresponding recommendation model for each object type in step 101, one or more attribute features of each object are obtained; and all objects with the same number of attribute features and the same definition of attribute features are divided into the same object type.

[0114] For example, sports shoes have three attribute features: size, brand, and material; and high heels also have three attribute features: size, brand, and material. Since sports shoes and high heels have the same number of attribute features (both are three) and the same definition of attribute features (both are size, brand, and material), sports shoes and high heels are divided into the same object type, i.e., “footwear”.

[0115] For another example, upright refrigerators have four attribute features: capacity, cooling method, shape, and purpose; and horizontal refrigerators have four attribute features: capacity, cooling method, shape, and purpose. Since upright refrigerators and horizontal refrigerators have the same number of attribute features (both are four) and the same definition of attribute features (both are capacity, cooling method, shape, and purpose), upright refrigerators and horizontal refrigerators are divided into the same object type, i.e., “refrigerator”.

[0116] It should be noted that in the present embodiment, any attribute feature of different object types is different, for example, any attribute feature of the three attribute features (size, brand, and material) of the object type “footwear” is different from any attribute feature of the four attribute features (capacity, cooling method, shape, and purpose) of the object type “refrigerator”.

[0117] In addition, if the number of attribute features of two objects is different, or if the number of attribute features of two objects is the same but there is at least one attribute feature with different definition of attribute features, the two objects belong to different object types.

[0118] In some embodiments, the establishment of a corresponding recommendation model for each object type in step 101 specifically includes steps 1011 to 1014:

[0119] 1011, obtaining a set of user behavior data, wherein each piece of user behavior data is data describing a specific behavior of a user on an object. The specific behavior is, for example, a selection operation such as clicking.

[0120] In this embodiment, user behavior data can be collected in real time through a client front-end product page, for example, each piece of user behavior data includes: user ID, object ID, behavior occurrence time, and exposed object set. The exposed object can be understood as an object displayed on the client front-end product page, and the exposed object set includes objects clicked by the user and objects exposed to the user but not clicked. The obtained user behavior data set is shown in Table 1, and in Table 1, the click time is the behavior occurrence time.

[0121] Table 1: User behavior data set

[0122] User ID Object ID Clicking time Exposed object set U1 P1 T1 P1, P2, P3, P4, P5 U2 P2 T2 P1, P2, P3, P4, P5 U2 P3 T3 P1, P2, P3, P4, P5 U3 P3 T4 P1, P2, P3, P4, P5 U3 P4 T5 P1, P2, P3, P4, P5 U4 P5 T5 P1, P2, P3, P4, P5

[0123] 1012. The user behavior data set is divided into data sets corresponding to different object types according to different object types involved.

[0124] In this embodiment, considering that the obtained user behavior data set includes objects belonging to different object types, in order to facilitate the establishment of a corresponding recommendation model for each object type, it is necessary to organize the user behavior data set to obtain data sets corresponding to different object types.

[0125] For example, in Table 1, if P1, P2, and P3 belong to an object type A, and P4 and P5 belong to an object type B, the user behavior data set shown in Table 1 can be divided into a data set corresponding to the object type A and a data set corresponding to the object type B. The data set of the object type A is shown in Table 2, and the data set of the object type B is shown in Table 3.

[0126] Table 2: Data set of object type A

[0127] User ID Object ID Clicking time Exposed object set U1 P1 T1 P1, P2, P3, P4, P5 U2 P2 T2 P1, P2, P3, P4, P5 U2 P3 T3 P1, P2, P3, P4, P5 U3 P3 T4 P1, P2, P3, P4, P5

[0128] Table 3: Data set of object type B

[0129] User ID Object ID Clicking time Exposed object set U3 P4 T5 P1, P2, P3, P4, P5 U4 P5 T5 P1, P2, P3, P4, P5

[0130] 1013. Based on the data sets of different object types, model training samples of each object type are constructed.

[0131] In this embodiment, if each piece of data in the data set of the object type includes: user ID, object ID, behavior occurrence time, and exposed object set, the constructed model training sample includes: user ID, object ID, behavior occurrence time, and label. The behavior occurrence time is, for example, the click time, and the label takes a value of 1 or 0.

[0132] If the label value is 1, the constructed model training sample is a positive sample, indicating an exposed and clicked object by the user; if the label value is 0, the constructed model training sample is a negative sample, indicating an exposed and non-clicked object by the user.

[0133] For example, based on the object type A data set shown in Table 2, the model training sample of the object type A can be constructed as shown in Table 4. Based on the object type B data set shown in Table 3, the model training sample of the object type B can be constructed as shown in Table 5.

[0134] Table 4 Model training sample of object type A

[0135] User ID Object ID Clicking time Label U1 P1 T1 1 U1 P2 T1 0 U1 P3 T1 0 U2 P1 T2 0 U2 P2 T2 1 U2 P3 T2 0 U3 P1 T4 1 U3 P2 T4 0 U3 P3 T4 0

[0136] Table 5 Model training sample of object type B

[0137] User ID Object ID Clicking time Label U3 P4 T5 1 U3 P5 T5 0 U4 P4 T5 0 U4 P5 T5 1

[0138] It should be noted that the model training sample of the object type A shown in Table 4 is only a part of the sample, not all samples, and the model training sample of the object type B shown in Table 5 is only a part of the sample, not all samples.

[0139] In some embodiments, in addition to including the user ID, object ID, behavior occurrence time and label, the model training sample further includes attribute feature information of the user and attribute feature information of the object. The attribute features of the user include but are not limited to gender, age, occupation, marital status, annual income, city (or home address), etc.

[0140] In constructing the model training sample, the attribute feature information of the user and the attribute feature information of the object can be added to the model training sample by splicing the user information table and the object information table. The user information table includes the attribute feature information of the user, and the object information table includes the attribute feature information of the object.

[0141] For example, after splicing the user information table and the object information table, the model training sample of the object type A shown in Table 4 is as shown in Table 6. After splicing the user information table and the object information table, the model training sample of the object type B shown in Table 5 is as shown in Table 7.

[0142] Table 6 Model training sample of object type A

[0143] User ID User's attribute feature Object ID Object's attribute feature Clicking time Label U1 [F11 to F61] P1 [f11 to f31] T1 1 U1 [F11 to F61] P2 [f12 to f32] T1 0 U1 [F11 to F61] P3 [f13 to f33] T1 0 U2 [F12 to F62] P1 [f11 to f31] T2 0 U2 [F12 to F62] P2 [f12 to f32] T2 1 U2 [F12 to F62] P3 [f13 to f33] T2 0 U3 [F13 to F63] P1 [f11 to f31] T4 1 U3 [F13 to F63] P2 [f12 to f32] T4 0 U3 [F13 to F63] P3 [f13 to f33] T4 0

[0144] Table 7 Model training sample of object type B

[0145] User ID User's attribute feature Product ID Object's attribute feature Clicking time Label U3 [F13 to F63] P4 [f44 to f74] T5 1 U3 [F13 to F63] P5 [f45 to f75] T5 0 U4 [F14 to F64] P4 [f44 to f74] T5 0 U4 [F14 to F64] P5 [f45 to f75] T5 1

[0146] In Table 6 and Table 7, F1 to F6 represent 6 attribute features of a user, for example: gender, age, occupation, marital status, annual income and city. The subscript of F1 to F6 is used to distinguish different users. f1 to f3 represent attribute features of object type A, f4 to f7 represent attribute features of object type B, and f1 to f7 are all different. The subscript of f1 to f7 is used to distinguish different objects.

[0147] 1014、Based on the model training sample of each object type, the preset model is trained to obtain a recommendation model corresponding to each object type.

[0148] In this embodiment, the preset model can be a tree model such as a GBDT (Gradient Boosting Decision Tree) model, or a regression model such as a LR (Logistic Regression) model, or a classification model.

[0149] In training the preset model, the model training sample of the object type is input into the preset model, and the parameters of the preset model itself are adjusted until the training target is reached. The training target includes two kinds: for positive samples, the output of the preset model converges to 1, and for negative samples, the output of the preset model converges to 0.

[0150] After training the preset model to obtain a recommendation model corresponding to each object type, for a recommendation model corresponding to a certain object type, for a recommended user, the recommendation model can be used to estimate the matching score of the object of the corresponding object type and the recommended user. The matching score value is greater than 0 and less than 1. Specifically, the recommended user ID, the attribute features of the recommended user, the object ID and the attribute features of the object are input into the recommendation model, and the matching score (score) of the recommended user ID and the object ID is estimated by the recommendation model.

[0151] In some embodiments, in step 102, for a recommended user, the recommendation ranking list of the objects in each object type is obtained by using the recommendation model corresponding to each object type, which specifically includes the following steps 1021 and 1022:

[0152] 1021, the matching score of the object ID and the recommended user ID in each object type is output by using the recommendation model corresponding to each object type.

[0153] 1022, the matching scores of the object ID and the recommended user ID in each object type are sorted from high to low to obtain the recommendation ranking list of the objects in each object type.

[0154] After obtaining the recommended ranking list of objects in each object type, a combination object recommendation result can be obtained by selecting objects from the recommended ranking list corresponding to each object type.

[0155] In some embodiments, the step 102 of obtaining the combination object recommendation result by selecting objects from the recommended ranking list corresponding to each object type specifically includes steps 1023-1025 as follows:

[0156] 1023. Obtain a behavior data set of the recommended user. The behavior data set of the recommended user can include historical behavior data of the recommended user, and can also include real-time behavior data of the recommended user.

[0157] 1024. Determine the user type of the recommended user based on the behavior data set of the recommended user.

[0158] In this embodiment, the object type to which each object in the behavior data set of the recommended user belongs is determined, and the number of object types is counted, and the user type of the recommended user is determined based on the number of object types.

[0159] For example, object type A and object type B can be divided according to the attribute characteristics of the objects, and accordingly, four user types can be divided as shown in Table 8.

[0160] Table 8 User type division

[0161] User type User behavior Y1 Object behavior only to A Y2 Object behavior only to B Y3 Object behavior to both A and B Y4 No object behavior to both A and B

[0162] It should be noted that this embodiment takes two object types as an example, and four user types can be divided, and if the number of object types is more than two, more user types can be divided based on the number of object types in which behaviors occur.

[0163] 1025. Based on the user type of the recommended user, a combination object recommendation result can be obtained by selecting objects from the recommended ranking list corresponding to each object type. Specifically, it can include steps S1-S3 as follows:

[0164] S1. Obtain one or more attribute feature data of each user in each user type. In this embodiment, the attribute feature data of the user is, for example, F1-F6 shown in Table 6 and Table 7.

[0165] S2. Determine one or more statistical feature data of each user based on the behavior data of each user in each user type.

[0166] In this embodiment, one or more statistical feature data of each user can be determined based on a set of historical user behavior data. The statistical feature data is, for example, a sliding window statistical feature of the number of clicks of an object in the past 3 days, 7 days, 14 days, 30 days, or 60 days, denoted as H1 to H5. The sliding window can be understood as 7 days being 4 days added to 3 days, 14 days being 7 days added to 7 days, and so on. In some embodiments, the statistical feature data can also be counted in other ways, and the specific counting method is not limited in this embodiment.

[0167] S3, selecting an object from the recommendation ranking list corresponding to each object type to obtain a combined object recommendation result based on the one or more attribute feature data of each user, the one or more statistical feature data of each user, and the user type of the recommended user.

[0168] In some embodiments, step S3 can include steps S31 to S34 as follows:

[0169] S31, determining at least two feature mean vectors based on the one or more attribute feature data of each user and the one or more statistical feature data of each user, and the number of the at least two feature mean vectors being the same as the number of the at least two object types.

[0170] Taking two object types (A and B) as an example, the user type division is shown in Table 8, which includes four types Y1 to Y4. Then, the five sliding window features of the objects of A for all users in Y1 and Y2 are calculated, denoted as Ha1 to Ha5, and the five sliding window features of the objects of B for all users in Y1 and Y2 are calculated, denoted as Hb1 to Hb5.

[0171] After obtaining Ha1 to Ha5 and Hb1 to Hb5, the Ha1 to Ha5 and Hb1 to Hb5 are standardized (for example, Min-Max standardization) to obtain the numerical representation of Ha1 to Ha5 and Hb1 to Hb5. In this way, the average values of Y1 and Y2 on F1 to F6, Ha1 to Ha5, and Hb1 to Hb5 can be calculated to obtain two feature mean vectors V1 and V2.

[0172] For example, user u1 in Y1 has only three features F1, Ha1, and Hb1, with values of 0, 1.2, and 0.7, respectively; user u2 in Y1 has only three features F1, Ha1, and Hb1, with values of 1, 0.9, and 0.5, respectively; then, V1 is [0.5, 1.05, 0.6], as shown in Table 9.

[0173] Table 9 Calculation of feature mean vector V1

[0174] F1 Ha1 Hb1 u1 0 1.2 0.7 u2 1 0.9 0.5 Mean value 0.5 1.05 0.6

[0175] S32, determine the distance between the recommended user and each feature mean vector.

[0176] For example, for any one user u in Y3, the Euclidean distance calculation method is used to calculate the distance of u to V1 and V2, denoted as d1 and d2, and the calculation example is as follows:

[0177] Suppose one user u1 in Y3 only has three features F1, Ha1 and Hb1, and the corresponding feature vector is [0, 1.2, 0.7], and the feature mean vector V1 is [0.5, 1.05, 0.6], then the distance d1 of user u1 to V1 is: √(0-0.5) 2 +(1.2-1.05) 2 +(0.7-0.6)2=0.5315.

[0178] S33, based on the distance between the recommended user and each feature mean vector, determine the preference probability of the recommended user for each object type.

[0179] For example, the preference probability of the recommended user for object type A is: d1 / (d1+d2); and the preference probability of the recommended user for object type B is: d2 / (d1+d2).

[0180] S34, based on the preference probability of the recommended user for each object type, select objects from the recommended ranking list corresponding to each object type multiple times to obtain a combined object recommendation result.

[0181] For any selection:

[0182] Generate one or more random probabilities, and based on the one or more random probabilities, filter one preference probability from the preference probabilities of all object types, and take one object from the recommended ranking list of the object type corresponding to the filtered preference probability.

[0183] Specifically, for any selection, first generate a random probability; second, filter the preference probability greater than the random probability to obtain a filtered set (the filtered set is a set of preference probabilities); further determine whether the number of preference probabilities in the filtered set is greater than 1; if greater than 1, generate a random probability again and update the filtered set; repeat the steps of determining, generating again and updating until the number in the filtered set is equal to 1, and stop generating random probability, so that the only preference probability in the filtered set is the filtered preference probability.

[0184] In this embodiment, the object with the highest ranking is taken from the recommended ranking list of the object type corresponding to the filtered preference probability. In this way, after multiple selections of objects, a combined object recommendation result can be obtained.

[0185] It should be noted that, taking the user types Y1 to Y3 as an example, the combination object recommendation result of the user type Y3 is obtained through the above steps S31 to S34, that is, when the user type of the recommended user is Y3, the combination object recommendation result is obtained through the above steps S31 to S34.

[0186] If the user type of the recommended user is Y1 (or Y2), the recommendation ranking list corresponding to A (or B) is the final recommendation result; if the user type of the recommended user is Y4, the object sales are calculated by using the object transaction table, and the final recommendation result is obtained according to the object sales from high to low.

[0187] It should be noted that, for the foregoing method embodiments, in order to simply describe, they are all expressed as a series of action combinations, but those skilled in the art can understand that the embodiments of the present disclosure are not limited by the action order described, because according to the embodiments of the present disclosure, certain steps can be performed in other order or simultaneously. In addition, those skilled in the art can understand that the embodiments described in the specification all belong to optional embodiments.

[0188] Figure 2 An exemplary block diagram of a combination object recommendation device provided by the embodiments of the present disclosure is shown in FIG. 2, wherein the combination object includes a plurality of objects, and the plurality of objects belong to at least two object types. As shown in FIG. 2, the combination object recommendation device includes an establishing unit 21 and a recommendation unit 22. Figure 2 The establishing unit 21 is configured to establish a corresponding recommendation model for each object type.

[0189] The establishing unit 21 is configured to establish a corresponding recommendation model for each object type.

[0190] The recommendation unit 22 is configured to obtain a recommendation ranking list of the objects in each object type by using the corresponding recommendation model of each object type for a recommended user, and select objects from the recommendation ranking list of each object type to obtain a combination object recommendation result.

[0191] In some embodiments, the device further includes:

[0192] The obtaining unit 23 is configured to obtain one or more attribute features of each object.

[0193] The dividing unit 24 is configured to divide all objects with the same number of attribute features and the same definition of attribute features into the same object type.

[0194] The establishing unit 21 is configured to establish a corresponding recommendation model for each object type after the object types are divided by the dividing unit 24.

[0195] In some embodiments, the establishing unit 21 is configured to:

[0196] obtain a set of user behavior data, wherein each piece of user behavior data is data describing a specific behavior of a user on an object;

[0197] divide the set of user behavior data into data sets corresponding to different object types according to the different object types involved;

[0198] construct a model training sample for each object type based on the data set of the different object types;

[0199] train a preset model based on the model training sample of each object type to obtain a recommendation model corresponding to each object type.

[0200] In some embodiments, each piece of user behavior data includes: a user ID, an object ID, a behavior occurrence time, and a set of exposed objects;

[0201] The model training sample includes: a user ID, an object ID, a behavior occurrence time, and a label.

[0202] In some embodiments, the recommendation unit 22 obtains, for a recommended user, a recommended ranking list of objects in each object type by using the recommendation model corresponding to each object type to include:

[0203] The recommendation unit 22 outputs a matching score of an object ID in each object type and a recommended user ID by using the recommendation model corresponding to each object type;

[0204] The recommendation unit 22 sorts the matching scores of the object IDs in each object type and the recommended user ID from high to low to obtain a recommended ranking list of objects in each object type.

[0205] In some embodiments, the recommendation unit 22 selects objects from the recommended ranking list corresponding to each object type to obtain a combined object recommendation result to include:

[0206] The recommendation unit 22 obtains a set of behavior data of a recommended user;

[0207] The recommendation unit 22 determines a user type of the recommended user based on the set of behavior data of the recommended user;

[0208] The recommendation unit 22 selects objects from the recommended ranking list corresponding to each object type based on the user type of the recommended user to obtain a combined object recommendation result.

[0209] In some embodiments, the recommendation unit 22 determines a user type of a recommended user based on a set of behavior data of the recommended user to include:

[0210] The recommendation unit 22 determines the object type to which each object on which a behavior occurs in the set of behavior data of the recommended user belongs, and counts the number of object types.

[0211] The recommendation unit 22 determines a user type of the recommended user based on the number of object types.

[0212] In some embodiments, the recommendation unit 22 selects objects from the recommendation ranking list corresponding to each object type to obtain the combined object recommendation result based on the user type of the recommended user includes:

[0213] The recommendation unit 22 obtains one or more attribute feature data of each user in each user type;

[0214] The recommendation unit 22 determines one or more statistical feature data of each user based on the behavior data of each user in each user type;

[0215] The recommendation unit 22 selects objects from the recommendation ranking list corresponding to each object type to obtain the combined object recommendation result based on the one or more attribute feature data of each user, the one or more statistical feature data of each user, and the user type of the recommended user.

[0216] In some embodiments, the recommendation unit 22 selects objects from the recommendation ranking list corresponding to each object type to obtain the combined object recommendation result based on the one or more attribute feature data of each user, the one or more statistical feature data of each user, and the user type of the recommended user includes:

[0217] The recommendation unit 22 determines at least two feature mean vectors based on the one or more attribute feature data of each user and the one or more statistical feature data of each user, and the number of the at least two feature mean vectors is the same as the number of the at least two object types;

[0218] The recommendation unit 22 determines the distance between the recommended user and each feature mean vector;

[0219] The recommendation unit 22 determines the preference probability of the recommended user for each object type based on the distance between the recommended user and each feature mean vector;

[0220] The recommendation unit 22 selects objects from the recommendation ranking list corresponding to each object type multiple times to obtain the combined object recommendation result based on the preference probability of the recommended user for each object type.

[0221] In some embodiments, the recommendation unit 22 selects objects from the recommendation ranking list corresponding to each object type multiple times based on the preference probability of the recommended user for each object type includes:

[0222] For each selection:

[0223] The recommendation unit 22 generates one or more random probabilities, and based on the one or more random probabilities, filters a preference probability from the preference probabilities of all object types, and takes an object from the recommendation ranking list of the object type corresponding to the filtered preference probability.

[0224] In some embodiments, the recommendation unit 22 generates one or more random probabilities, and based on the one or more random probabilities, filtering a preference probability from the preference probabilities of all object types includes:

[0225] generating a random probability;

[0226] filtering a preference probability greater than the random probability to obtain a filtered set;

[0227] determining whether the number in the filtered set is greater than 1;

[0228] if greater than 1, generating a random probability again, and updating the filtered set;

[0229] repeating the steps of determining, generating again, and updating until the number in the filtered set is equal to 1, and stopping generating the random probability.

[0230] The specific details of the combination object recommendation device can refer to the specific details of the combination object recommendation method, and will not be repeated here.

[0231] In some embodiments, the division of each unit in the combination object recommendation device is only a logical functional division, and there can be another division way in actual implementation, for example, at least two units in the combination object recommendation device can be implemented as one unit; each unit in the combination object recommendation device can also be divided into multiple sub-units. It can be understood that each unit or sub-unit can be realized by electronic hardware, or a combination of computer software and electronic hardware. Whether the functions are executed in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application.

[0232] Figure 3 is an exemplary block diagram of a system including at least one computing device and at least one storage device storing instructions provided by an embodiment of the present disclosure. In some embodiments, the system can be used for big data processing, and the at least one computing device and the at least one storage device can be distributedly deployed, so that the system is a distributed data processing cluster.

[0233] As Figure 3As shown, the system includes: at least one computing device 301, at least one storage device 302 storing instructions. It can be understood that the storage device 302 in this embodiment can be a volatile memory or a non-volatile memory, or can include both volatile and non-volatile memories.

[0234] In some embodiments, the storage device 302 stores elements, executable units or data structures, or a subset thereof, or an extended set thereof: an operating system and an application program.

[0235] The operating system includes various system programs, such as a framework layer, a core library layer, a driver layer, etc., for implementing various basic tasks and processing hardware-based tasks. The application program includes various application programs, such as a media player, a browser, etc., for implementing various application tasks. The program for implementing the combination object recommendation method provided by the embodiments of the present disclosure can be included in the application program.

[0236] In the embodiments of the present disclosure, the at least one computing device 301 can execute the steps of each embodiment of the combination object recommendation method provided by the embodiments of the present disclosure by invoking the program or instruction stored in the at least one storage device 302, specifically, the program or instruction stored in the application program.

[0237] The combination object recommendation method provided by the embodiments of the present disclosure can be applied to or implemented by the computing device 301. The computing device 301 can be an integrated circuit chip with signal processing capability. In the implementation process, each step of the above method can be completed by integrated logic circuits of hardware in the computing device 301 or instructions in the form of software. The computing device 301 described above can be a general-purpose processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0238] The steps of the combination object recommendation method provided by the embodiments of the present disclosure can be directly embodied as hardware coding processor execution completion, or combined execution completion by hardware and software units in the coding processor. The software units can be located in random access memory, flash memory, read-only memory, programmable read-only memory, or electrically erasable programmable memory, registers, and other mature storage media in the art. The storage media is located in the storage device 302, and the computing device 301 reads the information in the storage device 302, and combines the hardware to complete the steps of the method.

[0239] The embodiments of the present disclosure also provide a non-transitory computer readable storage medium storing programs or instructions, which, when executed by at least one computing device, cause the at least one computing device to perform the steps of the combination object recommendation method. For the sake of brevity, the description will not be repeated here. The computing device can be the computing device 301 shown in the figure. Figure 3

[0240] It should be noted that in this paper, the term "include", "contain" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or includes elements inherent to such process, method, article or device. Without more limitations, the elements defined by the statement "include" do not exclude the presence of other identical elements in the process, method, article or device including the element.

[0241] Those skilled in the art can understand that although some embodiments described herein include certain features included in other embodiments rather than other features, the combination of features of different embodiments means that it is within the scope of the present disclosure and forms different embodiments.

[0242] Those skilled in the art can understand that the description of each embodiment is focused on, and the parts not described in detail in a certain embodiment can refer to the relevant description of other embodiments.

[0243] Although the embodiments of the present disclosure are described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the present disclosure, and such modifications and variations fall within the scope defined by the appended claims.​

Claims

1. A method for recommending composite objects, wherein, The composite object includes multiple objects, and the multiple objects belong to at least two object types. The method includes: Develop a corresponding recommendation model for each object type; For a recommended user, the recommendation models corresponding to each object type are used to obtain the recommended ranking list of objects in each object type, and objects are selected from the recommended ranking list corresponding to each object type and merged to obtain the combined object recommendation result. Different objects of the same object type have the same attribute characteristics. The combined object recommendation result includes the recommendation result of the recommendation model corresponding to each object type. The establishment of a corresponding recommendation model for each object type includes: Obtain a set of user behavior data, wherein each piece of user behavior data describes a user's specific behavior toward an object, and the specific behavior is a selection operation; The user behavior data set is divided into datasets corresponding to different object types based on the different object types involved; Based on the datasets of the different object types, construct model training samples for each object type; Based on the model training samples for each object type, a preset model is trained to obtain a recommendation model corresponding to each object type.

2. The method according to claim 1, wherein, Before establishing a corresponding recommendation model for each object type, the method further includes: Retrieve one or more attribute characteristics for each object; All objects with the same number of attributes and the same attribute definitions are classified into the same object type.

3. The method according to claim 1, wherein, Each piece of user behavior data includes: user ID, object ID, behavior occurrence time, and set of exposed objects; The model training samples include: user ID, object ID, time of behavior occurrence, and label.

4. The method according to claim 1, wherein, For a given user, obtaining the recommended ranking list of objects in each object type using the recommendation model corresponding to each object type includes: The matching score between the object ID and the recommended user ID in each object type is output using the recommendation model corresponding to each object type; Sort the matching scores of object IDs and recommended user IDs in each object type from high to low to obtain a recommended sorted list of objects in each object type.

5. The method according to claim 1, wherein, The step of selecting objects from the recommendation ranking list corresponding to each object type to obtain the combined object recommendation result includes: Obtain the set of behavioral data of the recommended user; Based on the behavioral data set of the recommended users, the user type of the recommended users is determined; Based on the user type of the recommended user, objects are selected from the recommendation ranking list corresponding to each object type to obtain the combined object recommendation result.

6. The method according to claim 5, wherein, The process of determining the user type of the recommended user based on the behavioral data set of the recommended user includes: Determine the object type of each object that performs a behavior in the behavioral data set of the recommended user, and count the number of object types. The user type of the recommended user is determined based on the number of object types.

7. The method according to claim 5, wherein, The step of selecting objects from the recommendation ranking list corresponding to each object type based on the user type of the recommended user to obtain the combined object recommendation result includes: Retrieve one or more attribute feature data for each user in each user type; Based on the behavioral data of each user in each user type, determine one or more statistical characteristic data for each user; Based on one or more attribute feature data of each user, one or more statistical feature data of each user, and the user type of the recommended user, objects are selected from the recommendation ranking list corresponding to each object type to obtain the combined object recommendation result.

8. The method according to claim 7, wherein, The step of selecting objects from the recommendation ranking list corresponding to each object type to obtain the combined object recommendation result based on one or more attribute feature data of each user, one or more statistical feature data of each user, and the user type of the recommended user includes: Based on one or more attribute feature data and one or more statistical feature data of each user, at least two feature mean vectors are determined, and the number of the at least two feature mean vectors is the same as the number of the at least two object types; Determine the distance between the recommended user and each of the feature mean vectors; Based on the distance between the recommended user and each of the feature mean vectors, the probability of the recommended user's preference for each object type is determined; Based on the recommended user's preference probability for each object type, objects are selected multiple times from the recommendation ranking list corresponding to each object type to obtain a combined object recommendation result.

9. The method according to claim 8, wherein, The step of selecting objects multiple times from the recommendation ranking list corresponding to each object type based on the recommended user's preference probability for each object type includes: For any given selection: Generate one or more random probabilities, and based on the one or more random probabilities, select a preference probability from the preference probabilities of all object types, and select an object from the recommended ranking list of object types corresponding to the selected preference probability.

10. The method according to claim 9, wherein, The step of generating one or more random probabilities and, based on the one or more random probabilities, selecting a preference probability from the preference probabilities of all object types includes: Generate random probabilities; Select the preference probabilities that are greater than the random probability to obtain the filter set; Determine whether the number in the filter set is greater than 1; If the value is greater than 1, then generate random probabilities again and update the selection set; Repeat the steps of judgment, regeneration, and update until the number in the filter set equals 1, then stop generating random probabilities.

11. A combined object recommendation device, wherein, The combined object includes multiple objects, and the multiple objects belong to at least two object types. The device includes: Establish a unit to build a corresponding recommendation model for each object type; The recommendation unit is used to obtain a recommendation ranking list of objects in each object type using the recommendation model corresponding to each object type for a recommended user, and to select objects from the recommendation ranking list corresponding to each object type and merge them to obtain a combined object recommendation result. Different objects of the same object type have the same attribute characteristics. The combined object recommendation result includes the recommendation result of the recommendation model corresponding to each object type. The establishment unit is used for: Obtain a set of user behavior data, wherein each piece of user behavior data describes a user's specific behavior toward an object, and the specific behavior is a selection operation; The user behavior data set is divided into datasets corresponding to different object types based on the different object types involved; Based on the datasets of the different object types, construct model training samples for each object type; Based on the model training samples for each object type, a preset model is trained to obtain a recommendation model corresponding to each object type.

12. The apparatus according to claim 11, wherein, The device also includes: A retrieval unit is used to retrieve one or more attribute characteristics of each object; A partitioning unit is used to divide all objects with the same number of attribute features and the same attribute feature definitions into the same object type. The establishment unit is used to establish a corresponding recommendation model for each object type after the object type is divided by the division unit.

13. The apparatus according to claim 11, wherein, Each piece of user behavior data includes: user ID, object ID, behavior occurrence time, and set of exposed objects; The model training samples include: user ID, object ID, time of behavior occurrence, and label.

14. The apparatus according to claim 11, wherein, For a given user, the recommendation unit uses the recommendation model corresponding to each object type to obtain a recommended ranking list of objects in each object type, including: The recommendation unit uses the recommendation model corresponding to each object type to output the matching score between the object ID and the recommended user ID in each object type; The recommendation unit sorts the matching scores between the object IDs and the recommended user IDs in each object type from high to low, thus obtaining a recommended ranking list of objects in each object type.

15. The apparatus according to claim 11, wherein, The recommendation unit selects objects from the recommendation ranking list corresponding to each object type to obtain the combined object recommendation result, including: The recommendation unit obtains the behavioral data set of the recommended user; The recommendation unit determines the user type of the recommended user based on the behavioral data set of the recommended user; The recommendation unit selects objects from the recommendation sorting list corresponding to each object type based on the user type of the recommended user to obtain a combined object recommendation result.

16. The apparatus according to claim 15, wherein, The recommendation unit determines the user type of the recommended user based on the behavioral data set of the recommended user, including: The recommendation unit determines the object type of each object that has a behavior in the behavioral data set of the recommended user, and counts the number of object types. The recommendation unit determines the user type of the recommended user based on the number of object types.

17. The apparatus according to claim 15, wherein, The recommendation unit selects objects from the recommendation sorting list corresponding to each object type based on the user type of the recommended user to obtain a combined object recommendation result, including: The recommendation unit obtains one or more attribute feature data for each user in each user type; The recommendation unit determines one or more statistical feature data for each user based on the behavioral data of each user in each user type; The recommendation unit selects objects from the recommendation ranking list corresponding to each object type based on one or more attribute feature data of each user, one or more statistical feature data of each user, and the user type of the recommended user to obtain the combined object recommendation result.

18. The apparatus according to claim 17, wherein, The recommendation unit selects objects from the recommendation ranking list corresponding to each object type based on one or more attribute feature data of each user, one or more statistical feature data of each user, and the user type of the recommended user to obtain the combined object recommendation result, including: The recommendation unit determines at least two feature mean vectors based on one or more attribute feature data and one or more statistical feature data of each user, and the number of the at least two feature mean vectors is the same as the number of the at least two object types; The recommendation unit determines the distance between the recommended user and each of the feature mean vectors; The recommendation unit determines the probability of the recommended user's preference for each object type based on the distance between the recommended user and each of the feature mean vectors; The recommendation unit selects objects multiple times from the recommendation ranking list corresponding to each object type based on the preference probability of the recommended user for each object type, and obtains a combined object recommendation result.

19. The apparatus according to claim 18, wherein, The recommendation unit selects objects multiple times from the recommendation ranking list corresponding to each object type based on the probability of the recommended user's preference for each object type, including: For any given selection: The recommendation unit generates one or more random probabilities, and based on the one or more random probabilities, selects a preference probability from the preference probabilities of all object types, and selects an object from the recommendation ranking list of the object type corresponding to the selected preference probability.

20. The apparatus according to claim 19, wherein, The recommendation unit generates one or more random probabilities, and based on the one or more random probabilities, selects a preference probability from the preference probabilities of all object types, including: Generate random probabilities; Select the preference probabilities that are greater than the random probability to obtain the filter set; Determine whether the number in the filter set is greater than 1; If the value is greater than 1, then generate random probabilities again and update the selection set; Repeat the steps of judgment, regeneration, and update until the number in the filter set equals 1, then stop generating random probabilities.

21. A system comprising at least one computing device and at least one storage device for storing instructions, wherein, When the instruction is executed by the at least one computing device, it causes the at least one computing device to perform the steps of the combined object recommendation method as described in any one of claims 1 to 10.

22. A non-transitory computer-readable storage medium, wherein, The non-transitory computer-readable storage medium stores a program or instructions that, when executed by at least one computing device, cause the at least one computing device to perform the steps of the combined object recommendation method as described in any one of claims 1 to 10.

Images