Content recommendation method and apparatus, readable storage medium, and electronic device

Abstract

The present disclosure relates to a content recommendation method, device, readable storage medium and electronic equipment. The method comprises: in response to receiving a content recommendation request of a target object, obtaining multi-type behavior information of the target object on recommended content, content feature information of to-be-recommended content and attribute information of the target object, wherein the multi-type behavior information comprises an explicit feedback sequence and an implicit feedback sequence; determining a feedback level of each type of behavior information in the multi-type behavior information; determining behavior feedback information of the target object according to the feedback level, the multi-type behavior information, the content feature information and the attribute information of the target object; and determining target recommended content recommended to the target object according to the behavior feedback information and the content feature information. The accuracy of the determined behavior feedback information of the target object is improved, high-precision content recommendation is achieved, and user experience is improved.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of computer, in particular, to a content recommendation method and device, readable storage medium and electronic equipment. BACKGROUND

[0002] Personalized recommendation system has been widely applied in e-commerce, news and video and many other fields. The model in the existing recommendation system is usually a Click-Through Rate Prediction (CTR) model based on deep neural network. Exposure click data is used as sample for learning during training, and all the candidate contents are scored during prediction, and the highest one is recommended to the user.

[0003] How to accurately determine the recommended content to be recommended to the user is the core problem of the recommendation system. The existing method mainly relies on the implicit feedback behavior (such as click behavior) of the user to predict the recommended content to be recommended to the user. For example, different network structures are used to extract the user's historical browsing content from the historical click sequence, and then predict the content that the user is likely to click next. However, the implicit feedback behavior often contains a lot of noise, and there is a gap between the user's real viewing content, which will cause the problem of inaccurate prediction of the recommended content. For example, in many recommendation scenarios, users often click due to the attraction of the title or picture, but if the user is not satisfied with the content, the user will quickly close the page. SUMMARY

[0004] The purpose of the present disclosure is to provide a content recommendation method, device, readable storage medium and electronic equipment to solve the problems in the related art.

[0005] To achieve the above purpose, the first aspect of the present disclosure provides a content recommendation method, comprising:

[0006] In response to receiving a content recommendation request of a target object, obtaining multi-class behavior information of the target object on the recommended content, content feature information of the recommended content and attribute information of the target object, wherein the multi-class behavior information includes explicit feedback sequence and implicit feedback sequence;

[0007] Determining the feedback level of each class of behavior information in the multi-class behavior information;

[0008] According to the feedback level, the multi-class behavior information, the content feature information and the attribute information of the target object, determining the behavior feedback information of the target object;

[0009] According to the behavior feedback information and the content feature information, determining the target recommended content recommended to the target object.

[0010] Optionally, the determining the behavior feedback information of the target object according to the feedback level, the multi-type behavior information, the content feature information, and the attribute information of the target object comprises:

[0011] determining a cross-feature vector of each type of behavior information according to the feedback level, the multi-type behavior information, the content feature information, and a deep cross-feature extraction network of a prediction model;

[0012] determining the behavior feedback information of the target object according to the attribute information of the target object and the cross-feature vector of each type of behavior information.

[0013] Optionally, the determining the cross-feature vector of each type of behavior information according to the feedback level, the multi-type behavior information, the content feature information, and a deep cross-feature extraction network of a prediction model comprises:

[0014] inputting each type of behavior information into a Transform module in the deep cross-feature extraction network respectively to obtain a high-order vector corresponding to each type of behavior information;

[0015] obtaining a cross-feature vector corresponding to each type of behavior information in sequence according to the high-order vector corresponding to each type of behavior information, the content feature information, and an Attention module in the deep cross-feature extraction network according to the size of the feedback level.

[0016] Optionally, the multi-type behavior information comprises a positive behavior feedback sequence and a negative behavior feedback sequence; the feedback level comprises a first feedback level corresponding to the positive behavior feedback sequence determined according to the positive behavior feedback sequence, and a second feedback level corresponding to the negative behavior feedback sequence determined according to the negative behavior feedback sequence;

[0017] the obtaining the cross-feature vector corresponding to each type of behavior information in sequence according to the high-order vector corresponding to each type of behavior information, the content feature information, and an Attention module in the deep cross-feature extraction network according to the size of the feedback level comprises:

[0018] for each type of behavior information, inputting the high-order vector corresponding to the type of behavior information and the content feature information into the Attention module in the deep cross-feature extraction network to obtain a first cross-feature vector of the type of behavior information;

[0019] respectively determining the first feedback level and the second feedback level as a target level, and performing the following steps:

[0020] sort each type of behavior information corresponding to the target level according to the size of the target level, to obtain a sorted list corresponding to the target level;

[0021] input the high-order vector corresponding to the second type of behavior information located at the second position in the sorted list and the first cross-feature vector corresponding to the first type of behavior information located at the first position in the sorted list into the Attention module, to obtain a second cross-feature vector corresponding to the second type of behavior information; and

[0022] input the high-order vector corresponding to the i-th type of behavior information located at the i-th position in the sorted list, the first cross-feature vector corresponding to the first type of behavior information, and the second cross-feature vector corresponding to each type of behavior information located between the first position and the i-th position in the sorted list into the Attention module, to obtain a second cross-feature vector corresponding to the i-th type of behavior information, where i ranges from 3 to N, and N is the number of categories of the multi-type behavior information.

[0023] Optionally, the determining of the behavior feedback information of the target object according to the attribute information of the target object and the cross-feature vector of each type of behavior information comprises:

[0024] determining, according to the first cross-feature vector corresponding to the first type of behavior information located at the first position in the sorted list corresponding to the target level and the second cross-feature vector corresponding to the behavior information located at other positions except the first position, a first type of behavior feedback vector of the target object corresponding to the target level; and

[0025] determining, according to the first cross-feature vector corresponding to the behavior information located at other positions except the first position in the sorted list corresponding to the target level, a second type of behavior feedback vector of the target object corresponding to the target level;

[0026] determining the behavior feedback information of the target object according to the first type of behavior feedback vector, the second type of behavior feedback vector, and the attribute information of the target object.

[0027] Optionally, when the target level is the first feedback level, the first type of behavior feedback vector is a first positive feedback vector, and the second type of behavior feedback vector is a second positive feedback vector; when the target level is the second feedback level, the first type of behavior feedback vector is a first negative feedback vector, and the second type of behavior feedback vector is a second negative feedback vector; and the like degree represented by the first positive feedback vector is greater than the like degree represented by the second positive feedback vector, and the dislike degree represented by the first negative feedback vector is greater than the dislike degree represented by the second negative feedback vector.

[0028] Optionally, the explicit feedback sequence comprises a sharing behavior feedback sequence, a collection behavior feedback sequence and a dislike behavior feedback sequence; and the implicit feedback sequence comprises a click behavior feedback sequence, an exposure without click behavior feedback sequence and a short-time exit behavior feedback sequence.

[0029] The second aspect of the present disclosure provides a content recommendation device, comprising:

[0030] a first determining module configured to determine a feedback level of each type of behavior information in the multi-type behavior information;

[0031] a first determining module configured to determine a feedback level of each type of behavior information in the multi-type behavior information;

[0032] a second determining module configured to determine behavior feedback information of the target object according to the feedback level, the multi-type behavior information, the content feature information and the attribute information of the target object;

[0033] a third determining module configured to determine target recommendation content recommended to the target object according to the behavior feedback information and the content feature information.

[0034] Optionally, the second determining module comprises:

[0035] a first determining sub-module configured to determine a cross-feature vector of each type of behavior information according to the feedback level, the multi-type behavior information, the content feature information and a deep cross-feature extraction network of an estimation model;

[0036] a second determining sub-module configured to determine behavior feedback information of the target object according to the attribute information of the target object and the cross-feature vector of each type of behavior information.

[0037] Optionally, the first determining sub-module comprises:

[0038] a first input sub-module configured to input each type of behavior information into a Transform module in the deep cross-feature extraction network respectively to obtain a high-order vector corresponding to each type of behavior information;

[0039] a second input sub-module configured to obtain a cross-feature vector corresponding to each type of behavior information in turn according to the high-order vector corresponding to each type of behavior information, the content feature information and an Attention module in the deep cross-feature extraction network according to the size of the feedback level.

[0040] Optionally, the multiple types of behavior information comprises a positive behavior feedback sequence and a negative behavior feedback sequence; the feedback level comprises a first feedback level corresponding to the positive behavior feedback sequence determined according to the positive behavior feedback sequence, and a second feedback level corresponding to the negative behavior feedback sequence determined according to the negative behavior feedback sequence; and the second input submodule comprises:

[0041] a third input submodule configured to input, for each type of behavior information, a high-order vector corresponding to the type of behavior information and the content feature information into an Attention module in the deep cross-feature extraction network to obtain a first cross-feature vector of the type of behavior information;

[0042] a performing submodule configured to determine the first feedback level and the second feedback level as target levels respectively, and perform the following steps:

[0043] sorting each type of behavior information corresponding to the target levels according to sizes of the target levels to obtain a sorting list corresponding to the target levels;

[0044] inputting a high-order vector corresponding to a second type of behavior information located at a second position in the sorting list and a first cross-feature vector corresponding to a first type of behavior information located at a first position in the sorting list into the Attention module to obtain a second cross-feature vector corresponding to the second type of behavior information; and

[0045] inputting a high-order vector corresponding to an i-th type of behavior information located at an i-th position in the sorting list, the first cross-feature vector corresponding to the first type of behavior information, and a second cross-feature vector corresponding to each type of behavior information located between the first position and the i-th position in the sorting list into the Attention module to obtain a second cross-feature vector corresponding to the i-th type of behavior information, where i ranges from 3 to N, and N is a number of types of the multiple types of behavior information.

[0046] Optionally, the second determining submodule comprises:

[0047] a third determining submodule configured to determine, according to a first cross-feature vector corresponding to a first type of behavior information located at a first position in a sorting list corresponding to the target levels and a second cross-feature vector corresponding to behavior information located at other positions except the first position, a first type of behavior feedback vector of the target object corresponding to the target levels; and

[0048] a fourth determining submodule configured to determine, according to a first cross-feature vector corresponding to behavior information located at other positions except the first position in the sorting list corresponding to the target levels, a second type of behavior feedback vector of the target object corresponding to the target levels;

[0049] The fifth determining sub-module is configured to determine the behavior feedback information of the target object according to the first type of behavior feedback vector, the second type of behavior feedback vector, and the attribute information of the target object.

[0050] Optionally, the first type of behavior feedback vector is a first positive feedback vector, and the second type of behavior feedback vector is a second positive feedback vector when the target level is the first feedback level; the first type of behavior feedback vector is a first negative feedback vector, and the second type of behavior feedback vector is a second negative feedback vector when the target level is the second feedback level; and the like degree represented by the first positive feedback vector is greater than the like degree represented by the second positive feedback vector, and the dislike degree represented by the first negative feedback vector is greater than the dislike degree represented by the second negative feedback vector.

[0051] Optionally, the explicit feedback sequence includes a sharing behavior feedback sequence, a collection behavior feedback sequence, and a dislike behavior feedback sequence; and the implicit feedback sequence includes a click behavior feedback sequence, an exposure and non-click behavior feedback sequence, and a short-time exit behavior feedback sequence.

[0052] The third aspect of the present disclosure provides a computer readable storage medium having a computer program stored thereon, the program being executed by a processor to implement the steps of the method provided in the first aspect of the present disclosure.

[0053] The fourth aspect of the present disclosure provides an electronic device, comprising:

[0054] a memory having a computer program stored thereon;

[0055] a processor configured to execute the computer program in the memory to implement the steps of the method provided in the first aspect of the present disclosure.

[0056] By the above technical solution, the rich explicit feedback sequence and implicit feedback sequence of the target object are fused to accurately determine the behavior feedback information of the target object, and the noise problem in the implicit feedback sequence can be solved. Moreover, when determining the behavior feedback information of the target object, the feedback level of each type of behavior information in the multiple types of behavior information is also referred to, the behavior information with a greater feedback level is used to guide the behavior information with a smaller feedback level, the effective behavior feedback information of the target object is extracted, and the accuracy of the determined behavior feedback information of the target object is improved. Moreover, when recommending content for the target object based on the behavior feedback information with a higher accuracy, high-precision content recommendation can be achieved, and the user experience is improved.

[0057] Other features and advantages of the present disclosure will be described in detail in the following specific embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0058] The accompanying drawings are included to provide a further understanding of the disclosure and constitute a part of the specification, illustrate the present disclosure and, together with the specific embodiments described below, serve to explain the present disclosure, but do not constitute a limitation on the present disclosure. In the drawings:

[0059] Figure 1 is a flowchart of a content recommendation method according to an exemplary embodiment.

[0060] Figure 2 is a schematic diagram of generating behavior information i k according to an exemplary embodiment.

[0061] Figure 3 is a flowchart of determining behavior feedback information of a target object according to an exemplary embodiment.

[0062] Figure 4 is a structural schematic diagram of a prediction model according to an exemplary embodiment.

[0063] Figure 5 is a block diagram of a content recommendation device according to an exemplary embodiment.

[0064] Figure 6 is a block diagram of an electronic device according to an exemplary embodiment.

[0065] Figure 7 is a block diagram of an electronic device according to an exemplary embodiment. DETAILED DESCRIPTION

[0066] In related technologies, the following three methods are generally used to solve the noise problem of determining the to-be-recommended content to be recommended to the user according to implicit feedback behavior: the first method is a false positive sample identification method, in which some behavior-related auxiliary information (such as page stay time) or content attribute features (such as content length) are used to identify samples that are false positives (False Positive) in the click samples, and remove these samples from the training set. The second method is an adaptive sample weighting method, which dynamically adjusts the weight of the sample during training, and reduces the weight of the sample with a larger loss during training to reduce the interference of noisy samples. However, the loss of difficult samples will also be larger, so this method will weaken the learning effect of difficult samples. The third method is to introduce multiple feedback behavior methods, which do not remove noisy data in the training set in advance, but add additional feedback behavior of the user to the model for training, such as skipping, completing browsing, and other additional feedback information, using collaborative filtering or multi-task learning to recommend content to the user.

[0067] However, the data on which the two manners of false positive sample identification and adaptive sample weighting are based is still the click behavior of the user, but other feedback (such as negative feedback such as dislike) of the user is difficult to represent by the click behavior, and these feedbacks are crucial for improving the user experience. However, there are still some shortcomings in introducing multiple feedback behavior manners, for example: the combined feedback behaviors are not comprehensive enough, such as only adding the click, exposure without click and dislike behaviors, without considering the interest intensity difference expressed by different feedback behaviors and the interaction between feedback behaviors. Therefore, the related technologies cannot well solve the noise problem in the implicit feedback sequence, cannot accurately determine the to-be-recommended content to be recommended to the user, and result in poor user experience.

[0068] Therefore, the present disclosure provides a content recommendation method and device, a readable storage medium and an electronic device to improve the accuracy of the behavior feedback information of the target object determined, implement high-precision content recommendation, and improve the user experience.

[0069] The specific embodiments of the present disclosure are described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only used to illustrate and explain the present disclosure, and are not used to limit the present disclosure.

[0070] It should be noted that all the actions of obtaining signals, information (for example, the multi-type behavior information used below, the content feature information of the to-be-recommended content, the attribute information of the target object, etc.) or data in the present disclosure are performed under the premise of complying with the corresponding data protection regulations and policies of the country where the device is located, and with the authorization given by the corresponding device owner.

[0071] Figure 1 is a flowchart of a content recommendation method according to an exemplary embodiment. As shown in Figure 1 the content recommendation method can include the following steps.

[0072] In step S11, in response to receiving a content recommendation request of a target object, multi-type behavior information of the target object to the recommended content, content feature information of the to-be-recommended content and attribute information of the target object are obtained.

[0073] It should be noted that the target object can be a user, or an account logged in when the user views the recommended content, etc., and the present disclosure does not make specific limitations thereon. Taking the target object as a user as an example, the user can manually input a content recommendation request in a human-computer interaction interface or voice input a content recommendation request, etc. In addition, the content recommendation method provided by the present disclosure can be applied to a server, or can be applied to an electronic device, and the present disclosure does not make specific limitations thereon.

[0074] In the present disclosure, the multi-type behavior information of the target object to the recommended content can include explicit feedback sequences and implicit feedback sequences. The explicit feedback sequence is a behavior sequence explicitly representing the preference of the target object to the recommended content, and the explicit feedback sequence can further include an explicit positive feedback sequence representing the preference of the target object to the recommended content and / or an explicit negative feedback sequence representing the disliking of the target object to the recommended content. For example, the explicit positive feedback sequence can include but is not limited to a sharing behavior feedback sequence and a collection behavior feedback sequence. The explicit negative feedback information can include but is not limited to a disliking behavior feedback sequence.

[0075] Similarly, the implicit feedback sequence is a behavior sequence not explicitly representing the preference of the target object to the recommended content, and the implicit feedback sequence can further include an implicit positive feedback sequence representing the preference of the target object to the recommended content and / or an implicit negative feedback sequence representing the disliking of the target object to the recommended content. For example, the implicit positive feedback sequence can include but is not limited to a clicking behavior feedback sequence, and the implicit negative feedback information can include but is not limited to an exposure non-clicking behavior feedback sequence and a short-time exit behavior feedback sequence. In addition, the implicit negative feedback sequence can also include a user closing notification behavior feedback sequence and an uninstalling APP behavior feedback sequence, etc.

[0076] Among them, it is found in data statistics that the number of implicit feedback sequences is large, but the noise contained is large, and the intensity of the behavior feedback information reflecting the user is low. The number of explicit feedback sequences is small, but the noise contained is small, and the intensity of the behavior feedback information reflecting the user is high. Different types of feedback sequences reflect different aspects of user behavior feedback, so in the present disclosure, the obtained multi-type behavior information includes explicit feedback sequences and implicit feedback sequences, so as to complement each other and be beneficial to comprehensively depicting the unbiased feedback of the user.

[0077] It is worth noting that in the present disclosure, the specific content of the multi-type behavior information is not limited, which can include explicit positive feedback sequences and implicit positive feedback sequences, can also include explicit negative feedback sequences and implicit negative feedback sequences, can also include explicit positive feedback sequences, implicit positive feedback sequences, explicit negative feedback sequences and implicit negative feedback sequences, etc. The following will be mainly described in detail taking the multi-type behavior information including explicit positive feedback sequences, implicit positive feedback sequences, explicit negative feedback sequences and implicit negative feedback sequences as an example.

[0078] Among them, the behavior information i k can include but is not limited to the attribute information a k of the recommended content and the context information c k of the behavior. Taking the recommended content as a commodity as an example, the attribute information a k can include a merchant ID, a secondary category ID of the commodity, a commercial circle ID, etc. The context information ck The behavior occurrence time can include a time difference between the behavior occurrence time and a current time, a day of the week in which the behavior occurs, an hour of the day in which the behavior occurs, and the like.

[0079] For example, Figure 2 is a schematic diagram of generating behavior information i k according to an example embodiment. As Figure 2 shown, attribute information a k of the recommended content and context information c k of the behavior are obtained through a fusion function to obtain behavior information i k = fusion(a k ,c k ). Wherein i k represents the kth behavior information.

[0080] Still taking the goods as an example, the content information of the content to be recommended can include the name of the goods, the merchant ID, the secondary category ID of the goods, the commercial district ID, and the like. It should be noted that the recommended content and the content to be recommended mentioned in the present disclosure can also be articles, videos, music, and the like, and the present disclosure does not make specific limitations thereto.

[0081] In addition, the attribute information of the target object can include but is not limited to: user age, occupation, hobby, and the like.

[0082] In step S12, the feedback level of each type of behavior information in the multiple types of behavior information is determined.

[0083] In the present disclosure, the feedback level is used to represent the feedback degree of the behavior information to the recommended content. For example, the feedback level is a first feedback level used to represent the like degree to the recommended content, and the greater the first feedback level represents the higher the like degree to the recommended content. For another example, the feedback level is a second feedback level used to represent the dislike degree to the recommended content, and the greater the second feedback level represents the higher the dislike degree to the recommended content.

[0084] In a possible manner, the feedback level of each type of behavior information can be determined according to a preset correspondence between the feedback level and the behavior information. For example, the feedback level of each type of behavior information determined according to the positive behavior feedback sequence of the multiple types of behavior information (i.e., the first feedback level) is in descending order as follows: the first feedback level of the sharing behavior feedback sequence > the first feedback level of the collection behavior feedback sequence > the first feedback level of the clicking behavior feedback sequence. The feedback level of each type of behavior information determined according to the negative behavior feedback sequence of the multiple types of behavior information (i.e., the second feedback level) is in descending order as follows: the second feedback level of the dislike behavior feedback sequence > the second feedback level of the short-time exit behavior feedback sequence > the second feedback level of the exposure and non-clicking behavior feedback sequence.

[0085] In step S13, the behavior feedback information of the target object is determined according to the feedback level, the multiple types of behavior information, the content feature information, and the attribute information of the target object.

[0086] In the present disclosure, the feedback level of each type of behavior information is referred to when determining the behavior feedback information of the target object, so that the behavior information with a larger feedback level can be used to guide the behavior information with a smaller feedback level, and the effective behavior feedback information of the target object can be extracted. The specific implementation of determining the behavior feedback information of the target object will be described in detail below.

[0087] In step S14, the target recommendation content recommended to the target object is determined according to the behavior feedback information and the content feature information.

[0088] First, the estimated click rate of the target object to the to-be-recommended content is determined according to the behavior feedback information and the content feature information, and then the target recommendation content recommended to the target object is determined according to the estimated click rate.

[0089] For example, the behavior feedback information and the content feature information of the to-be-recommended content can be input into a model for estimating the click rate, so as to obtain the estimated click rate of each to-be-recommended content output by the model. The model for estimating the click rate can be trained by using a machine learning model technology in the related art, and the present disclosure does not make a specific limitation on this.

[0090] In the present disclosure, the to-be-recommended content with the largest estimated click rate can be determined as the target recommendation content, or multiple to-be-recommended contents with an estimated click rate greater than a preset threshold can be determined as the target recommendation content. After the target recommendation content is determined, the target recommendation content is recommended to the target object. The present disclosure does not make a specific limitation on the number of the determined target recommendation content.

[0091] The technical solution is adopted to fuse rich explicit feedback sequences and implicit feedback sequences of the target object to accurately determine the behavior feedback information of the target object, and can solve the noise problem in the implicit feedback sequence. In addition, the feedback level of each type of behavior information in the multi-type behavior information is also referred to in determining the behavior feedback information of the target object, the behavior information with a larger feedback level is used to guide the behavior information with a smaller feedback level, the effective behavior feedback information of the target object is extracted, and the accuracy of the determined behavior feedback information is improved. In addition, when the content is recommended for the target object based on the behavior feedback information with a higher accuracy, high-precision content recommendation can be realized, and the user experience is improved.

[0092] The specific implementation of determining the behavior feedback information of the target object according to the feedback level, the multi-type behavior information, the content feature information, and the attribute information of the target object is described below.

[0093] Figure 3 is a flowchart of determining the behavior feedback information of the target object according to an example embodiment. As shown in Figure 3 Figure 1 The step S13 can include the following steps.

[0094] In step S131, the cross-feature vector of each type of behavior information is determined according to the feedback level, the multi-type behavior information, the content feature information, and the deep cross-feature extraction network of the estimation model.

[0095] First, each type of behavior information is input into the Transform module in the deep cross-feature extraction network to obtain the high-order vector corresponding to each type of behavior information.

[0096] For example, each behavior information in each type of behavior information can be represented by a sequence vector. For each behavior information in each type of behavior information, the Transformer module is used to extract the fine-grained interaction information of the sequence to obtain the high-order vector corresponding to each type of behavior information. The Transformer module includes a multi-head attention (Multi-head attention), a feed-forward network (Feed-forward network), a residual connection (Residual connections), and a layer normalization (Layer normalization) unit.

[0097] For example, as shown in Figure 4 each behavior sequence included in the sharing behavior feedback sequence, the collection behavior feedback sequence, the click behavior feedback sequence, the exposure and non-click behavior feedback sequence, the short-time exit behavior feedback sequence, and the dislike behavior feedback sequence is input into the Transform module in the deep cross-feature extraction network to obtain the high-order vector corresponding to each behavior sequence. Figure 4 ​The input Transformer module is used to obtain a high-order vector corresponding to each row of sequence (as shown by the rectangular frame at the bottom of the figure). It is worth noting that each row of sequence corresponds to a high-order vector, and each type of behavior information includes multiple sequences of behavior, so the number of high-order vectors corresponding to each type of behavior information is multiple.

[0098] Then, according to the size of the feedback level, the high-order vector corresponding to each type of behavior information, the content feature information, and the Attention module in the deep cross feature extraction network are used to obtain the cross feature vector corresponding to each type of behavior information in turn.

[0099] It is worth noting that in the present disclosure, in addition to the type of behavior information with the maximum feedback level including one cross feature vector (denoted as the first cross feature vector), the other types of behavior information each include two cross feature vectors, which are referred to as the first cross feature vector and the second cross feature vector.

[0100] The first cross feature vector is generated in the following manner: for each type of behavior information, the high-order vector corresponding to the type of behavior information and the content feature information are input into the Attention module in the deep cross feature extraction network to obtain the first cross feature vector of the type of behavior information. For example, as shown in Figure 4 The content feature information of the content to be recommended is taken as a target item, and the target item is taken as a query Query input into the Attention module. In this way, for each type of behavior information, the query Query and the high-order vector corresponding to the type of behavior information are input into the Attention module to obtain the first cross feature vector of the type of behavior information. The first cross feature vector of the type of behavior information is generated only according to the type of behavior information and the content feature information. In this way, the first cross feature vector corresponding to each type of behavior information can be obtained.

[0101] It is worth noting that in the present disclosure, the cross feature vector is determined for the positive behavior feedback sequence and the negative behavior feedback sequence respectively, and the negative behavior feedback sequence is not considered when determining the cross feature vector corresponding to the positive behavior feedback sequence, and the positive behavior feedback sequence is not considered when determining the cross feature vector corresponding to the negative behavior feedback sequence. Therefore, when the multiple types of behavior information include the positive behavior feedback sequence and the negative behavior feedback sequence, the feedback level includes a first feedback level corresponding to the positive behavior feedback sequence determined according to the positive behavior feedback sequence, and a second feedback level corresponding to the negative behavior feedback sequence determined according to the negative behavior feedback sequence.

[0102] Accordingly, the second cross-feature vector is generated by determining the first feedback level and the second feedback level as the target level respectively, and performing the following steps: sorting each type of behavior information corresponding to the target level according to the size of the target level to obtain a sorting list corresponding to the target level; inputting the high-order vector corresponding to the second type of behavior information located at the second position in the sorting list and the first cross-feature vector corresponding to the first type of behavior information located at the first position in the sorting list into the Attention module to obtain the second cross-feature vector corresponding to the second type of behavior information; and inputting the high-order vector corresponding to the ith type of behavior information located at the ith position in the sorting list, the first cross-feature vector corresponding to the first type of behavior information, and the second cross-feature vector corresponding to each type of behavior information located between the first position and the ith position in the sorting list into the Attention module to obtain the second cross-feature vector corresponding to the ith type of behavior information, where i ranges from 3 to N, and N is the number of categories of the multiple types of behavior information.

[0103] It should be understood that in the present disclosure, the second cross-feature vector corresponding to each type of behavior information between the first position and the ith position does not include the second cross-feature vector corresponding to the behavior information located at the first position and the ith position.

[0104] For example, assuming that the target level is the first feedback level corresponding to the positive behavior feedback sequence, and the size of the first feedback level is: the first feedback level of the sharing behavior feedback sequence is greater than the first feedback level of the collection behavior feedback sequence, and the first feedback level of the collection behavior feedback sequence is greater than the first feedback level of the click behavior feedback sequence, therefore, the sorting list corresponding to the first feedback level is: the sharing behavior feedback sequence, the collection behavior feedback sequence, and the click behavior feedback sequence.

[0105] Referring to Figure 4 , first, the high-order vector corresponding to the collection behavior feedback sequence and the first cross-feature vector corresponding to the sharing behavior feedback sequence are input into the Attention module to obtain the second cross-feature vector corresponding to the collection behavior feedback sequence. For example, the first cross-feature vector corresponding to the sharing behavior feedback sequence is taken as the query Query input into the Attention module, and the high-order vector corresponding to the collection behavior feedback sequence and the query Query are input into the Attention module to obtain the second cross-feature vector corresponding to the collection behavior feedback sequence.

[0106] Then, the high-order vector corresponding to the click behavior feedback sequence, the first cross-feature vector corresponding to the sharing behavior feedback sequence, and the second cross-feature vector corresponding to the collection behavior feedback sequence are input into the Attention module to obtain the second cross-feature vector corresponding to the click behavior feedback sequence. For example, the first cross-feature vector corresponding to the sharing behavior feedback sequence and the second cross-feature vector corresponding to the collection behavior feedback sequence are taken as the query Query of the Attention module, and the high-order vector corresponding to the click behavior feedback sequence and the query Query are input into the Attention module to obtain the second cross-feature vector corresponding to the click behavior feedback sequence.

[0107] In this way, the first cross-feature vector of the first type of behavior information located at the first position in the ranking list corresponding to the first feedback level, and the first cross-feature vector and the second cross-feature vector of the other types of behavior information located at other positions are obtained.

[0108] For example, it is assumed that the target level is the second feedback level corresponding to the negative behavior feedback sequence, and the size of the second feedback level is: the second feedback level of the dislike behavior feedback sequence is greater than the second feedback level of the short-time exit behavior feedback sequence, and the second feedback level of the short-time exit behavior feedback sequence is greater than the second feedback level of the exposure non-click behavior feedback sequence. Therefore, the ranking list corresponding to the second feedback level is: the dislike behavior feedback sequence, the short-time exit behavior feedback sequence, and the exposure non-click behavior feedback sequence.

[0109] Referring to Figure 4 First, the high-order vector corresponding to the short-time exit behavior feedback sequence and the first cross-feature vector corresponding to the dislike behavior feedback sequence are input into the Attention module to obtain the second cross-feature vector corresponding to the short-time exit behavior feedback sequence. For example, the first cross-feature vector corresponding to the dislike behavior feedback sequence is taken as the query Query of the Attention module, and the high-order vector corresponding to the short-time exit behavior feedback sequence and the query Query are input into the Attention module to obtain the second cross-feature vector corresponding to the short-time exit behavior feedback sequence.

[0110] Then, the high-order vector of the exposure non-click behavior feedback sequence, the first cross feature vector corresponding to the dislike behavior feedback sequence, and the second cross feature vector corresponding to the short-time exit behavior feedback sequence are input into the Attention module to obtain the second cross feature vector corresponding to the exposure non-click behavior feedback sequence. For example, the first cross feature vector corresponding to the dislike behavior feedback sequence and the second cross feature vector corresponding to the short-time exit behavior feedback sequence are taken as the query Query of the Attention module, and the high-order vector corresponding to the exposure non-click behavior feedback sequence and the query Query are input into the Attention module to obtain the second cross feature vector corresponding to the exposure non-click behavior feedback sequence.

[0111] In this way, the first cross feature vector of the first type of behavior information located at the first position in the ranking list corresponding to the second feedback level and the first cross feature vector and the second cross feature vector of other types of behavior information located at other positions can be obtained.

[0112] By using the above technical solution, when key useful information of a type of behavior information with a slightly lower feedback level is extracted, one or more types of behavior information located before the type of behavior information in the ranking list are taken as a reference to more comprehensively and stereoscopically depict the unbiased feedback of the target object, and the accuracy of the determined behavior feedback information of the target object can be improved when the behavior feedback information of the target object is determined based on the cross feature vector.

[0113] Return Figure 3 In step S132, the behavior feedback information of the target object is determined according to the attribute information of the target object and the cross feature vector of each type of behavior information.

[0114] The specific implementation of step S132 is as follows: first, the first type of behavior feedback vector corresponding to the target level of the target object is determined according to the first cross feature vector corresponding to the first type of behavior information located at the first position in the ranking list corresponding to the target level and the second cross feature vector corresponding to the behavior information located at other positions except the first position; and the second type of behavior feedback vector corresponding to the target level of the target object is determined according to the first cross feature vector corresponding to the behavior information located at other positions except the first position in the ranking list corresponding to the target level; and the behavior feedback information of the target object is determined according to the first type of behavior feedback vector, the second type of behavior feedback vector, and the attribute information of the target object.

[0115] wherein, when the target level is the first feedback level, the first type of behavior feedback vector is a first positive feedback vector, and the second type of behavior feedback vector is a second positive feedback vector; when the target level is the second feedback level, the first type of behavior feedback vector is a first negative feedback vector, and the second type of behavior feedback vector is a second negative feedback vector; and the like degree represented by the first positive feedback vector is greater than the like degree represented by the second positive feedback vector, and the dislike degree represented by the first negative feedback vector is greater than the dislike degree represented by the second negative feedback vector.

[0116] For example, assuming that the target level is the first feedback level, as shown in FIG. 1, the first positive feedback vector e1 corresponding to the first feedback level of the target object is determined according to the first cross-feature vector corresponding to the sharing behavior feedback sequence, the second cross-feature vector corresponding to the collection behavior feedback sequence, and the second cross-feature vector corresponding to the clicking behavior feedback sequence. For example, the first positive feedback vector e1 is obtained by inputting the first cross-feature vector corresponding to the sharing behavior feedback sequence, the second cross-feature vector corresponding to the collection behavior feedback sequence, and the second cross-feature vector corresponding to the clicking behavior feedback sequence into the aggregation module Aggregator. And the second positive feedback vector e2 corresponding to the first feedback level of the target object is determined according to the first cross-feature vector corresponding to the collection behavior feedback sequence and the first cross-feature vector corresponding to the clicking behavior feedback sequence. For example, the second positive feedback vector e2 is obtained by inputting the first cross-feature vector corresponding to the collection behavior feedback sequence and the first cross-feature vector corresponding to the clicking behavior feedback sequence into the aggregation module Aggregator. Wherein, the like degree represented by the first positive feedback vector e1 is greater than the like degree represented by the second positive feedback vector e2. Figure 4 For another example, assuming that the target level is the second feedback level, as shown in FIG. 2, the first negative feedback vector e1 corresponding to the second feedback level of the target object is determined according to the first cross-feature vector corresponding to the sharing behavior feedback sequence, the second cross-feature vector corresponding to the collection behavior feedback sequence, and the second cross-feature vector corresponding to the clicking behavior feedback sequence. For example, the first negative feedback vector e1 is obtained by inputting the first cross-feature vector corresponding to the sharing behavior feedback sequence, the second cross-feature vector corresponding to the collection behavior feedback sequence, and the second cross-feature vector corresponding to the clicking behavior feedback sequence into the aggregation module Aggregator. And the second negative feedback vector e2 corresponding to the second feedback level of the target object is determined according to the first cross-feature vector corresponding to the collection behavior feedback sequence and the first cross-feature vector corresponding to the clicking behavior feedback sequence. For example, the second negative feedback vector e2 is obtained by inputting the first cross-feature vector corresponding to the collection behavior feedback sequence and the first cross-feature vector corresponding to the clicking behavior feedback sequence into the aggregation module Aggregator. Wherein, the dislike degree represented by the first negative feedback vector e1 is greater than the dislike degree represented by the second negative feedback vector e2.

[0117] Figure 4 ​As shown, based on the first cross-feature vector corresponding to the dislike behavior feedback sequence, the second cross-feature vector corresponding to the short-term exit behavior feedback sequence, and the second cross-feature vector corresponding to the exposure-without-click behavior feedback sequence, a first negative feedback vector e4 corresponding to the second feedback level of the target object is determined. For example, the first cross-feature vector corresponding to the dislike behavior feedback sequence, the second cross-feature vector corresponding to the short-term exit behavior feedback sequence, and the second cross-feature vector corresponding to the exposure-without-click behavior feedback sequence are input into the aggregation module Aggregator for aggregation to obtain the first negative feedback vector e4. And, based on the first cross-feature vector corresponding to the short-term exit behavior feedback sequence and the first cross-feature vector corresponding to the exposure-without-click behavior feedback sequence, a second negative feedback vector e3 corresponding to the second feedback level of the target object is determined. For example, the first cross-feature vector corresponding to the short-term exit behavior feedback sequence and the first cross-feature vector corresponding to the exposure-without-click behavior feedback sequence are input into the aggregation module Aggregator for aggregation to obtain the second negative feedback vector e3. The degree of dislike represented by the first negative feedback vector e4 is greater than the degree of dislike represented by the second negative feedback vector e3.

[0118] Then, the target object's attribute information e0, the first positive feedback vector e1, the second positive feedback vector e2, the first negative feedback vector e4, and the second negative feedback vector e3 can be aggregated through the aggregation module Aggregator to obtain the target object's behavior feedback information u = f(e0, e1, e2, e3, e4).

[0119] like Figure 4 As shown, after determining the behavioral feedback information of the target object, the behavioral feedback information of the target object and the content feature information of the content to be recommended are input into the CTR Predictor module of the evaluation model to obtain the CTR Score of the target object for the content to be recommended, output by the CTR Predictor module. Finally, based on this estimated click-through rate (CTR) score... Determine the target content to recommend to the target audience.

[0120] Compared to solutions that determine the target object's behavioral feedback information solely based on click behavior feedback sequences, the solution provided in this disclosure improves the accuracy of determining behavioral feedback information by 0.36%. Therefore, the solution provided in this disclosure can achieve high-precision content recommendation and enhance the user experience.

[0121] Based on the same inventive concept, this disclosure also provides a content recommendation device. Figure 5 This is a block diagram illustrating a content recommendation apparatus according to an exemplary embodiment. Figure 5As shown, the content recommendation apparatus 500 can include:

[0122] an acquisition module 501 configured to acquire, in response to receiving a content recommendation request of a target object, multi-type behavior information of the target object on recommended content, content feature information of to-be-recommended content, and attribute information of the target object, wherein the multi-type behavior information includes explicit feedback sequences and implicit feedback sequences;

[0123] a first determination module 502 configured to determine a feedback level of each type of behavior information in the multi-type behavior information;

[0124] a second determination module 503 configured to determine behavior feedback information of the target object according to the feedback level, the multi-type behavior information, the content feature information, and the attribute information of the target object;

[0125] a third determination module 504 configured to determine target recommended content recommended to the target object according to the behavior feedback information and the content feature information.

[0126] Optionally, the second determination module 503 includes:

[0127] a first determination sub-module configured to determine a cross feature vector of each type of behavior information according to the feedback level, the multi-type behavior information, the content feature information, and a deep cross feature extraction network of an estimation model;

[0128] a second determination sub-module configured to determine the behavior feedback information of the target object according to the attribute information of the target object and the cross feature vector of each type of behavior information.

[0129] Optionally, the first determination sub-module includes:

[0130] a first input sub-module configured to input each type of behavior information into a Transform module in the deep cross feature extraction network respectively to obtain a high-order vector corresponding to each type of behavior information;

[0131] a second input sub-module configured to obtain, according to the high-order vector corresponding to each type of behavior information, the content feature information, and an Attention module in the deep cross feature extraction network, in sequence, a cross feature vector corresponding to each type of behavior information according to a size of the feedback level.

[0132] Optionally, the multiple types of behavior information comprises a positive behavior feedback sequence and a negative behavior feedback sequence; the feedback level comprises a first feedback level corresponding to the positive behavior feedback sequence determined according to the positive behavior feedback sequence, and a second feedback level corresponding to the negative behavior feedback sequence determined according to the negative behavior feedback sequence; and the second input submodule comprises:

[0133] a third input submodule configured to input, for each type of behavior information, a high-order vector corresponding to the type of behavior information and the content feature information into an Attention module in the deep cross-feature extraction network to obtain a first cross-feature vector of the type of behavior information;

[0134] a performing submodule configured to determine the first feedback level and the second feedback level as target levels respectively, and perform the following steps:

[0135] sorting each type of behavior information corresponding to the target levels according to sizes of the target levels to obtain a sorting list corresponding to the target levels;

[0136] inputting a high-order vector corresponding to a second type of behavior information located at a second position in the sorting list and a first cross-feature vector corresponding to a first type of behavior information located at a first position in the sorting list into the Attention module to obtain a second cross-feature vector corresponding to the second type of behavior information; and

[0137] inputting a high-order vector corresponding to an i-th type of behavior information located at an i-th position in the sorting list, the first cross-feature vector corresponding to the first type of behavior information, and a second cross-feature vector corresponding to each type of behavior information located between the first position and the i-th position in the sorting list into the Attention module to obtain a second cross-feature vector corresponding to the i-th type of behavior information, where i ranges from 3 to N, and N is a number of types of the multiple types of behavior information.

[0138] Optionally, the second determining submodule comprises:

[0139] a third determining submodule configured to determine, according to a first cross-feature vector corresponding to a first type of behavior information located at a first position in a sorting list corresponding to the target levels and a second cross-feature vector corresponding to behavior information located at other positions except the first position, a first type of behavior feedback vector of the target object corresponding to the target levels; and

[0140] a fourth determining submodule configured to determine, according to a first cross-feature vector corresponding to behavior information located at other positions except the first position in the sorting list corresponding to the target levels, a second type of behavior feedback vector of the target object corresponding to the target levels;

[0141] a fifth determining sub-module, configured to determine behavior feedback information of the target object according to the first type of behavior feedback vector, the second type of behavior feedback vector, and attribute information of the target object.

[0142] Optionally, the first type of behavior feedback vector is a first positive feedback vector, and the second type of behavior feedback vector is a second positive feedback vector when the target level is the first feedback level; the first type of behavior feedback vector is a first negative feedback vector, and the second type of behavior feedback vector is a second negative feedback vector when the target level is the second feedback level; and the like degree represented by the first positive feedback vector is greater than the like degree represented by the second positive feedback vector, and the dislike degree represented by the first negative feedback vector is greater than the dislike degree represented by the second negative feedback vector.

[0143] Optionally, the explicit feedback sequence includes a sharing behavior feedback sequence, a collection behavior feedback sequence, and a dislike behavior feedback sequence; and the implicit feedback sequence includes a click behavior feedback sequence, an exposure without click behavior feedback sequence, and a short-time exit behavior feedback sequence.

[0144] As to the apparatus in the above-described embodiments, the specific manners in which various modules perform operations have been described in detail in the embodiments of the method, and thus will not be described here in detail.

[0145] Figure 6 is a block diagram of an electronic device according to an exemplary embodiment. As shown in Figure 6 the electronic device 700 can include a processor 701, a memory 702. The electronic device 700 can also include one or more of a multimedia component 703, an input / output interface 704, and a communication component 705.

[0146] The processor 701 is configured to control overall operations of the electronic device 700 to complete all or part of the steps of the content recommendation method described above. The memory 702 is configured to store various types of data to support operations of the electronic device 700, which can include, for example, instructions for operating any application or method on the electronic device 700, and application-related data, such as contact data, transmitted and received messages, pictures, audio, video, and the like. The memory 702 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic storage, a flash memory, a magnetic disk, or an optical disk. The multimedia component 703 can include a screen and an audio component. The screen can be, for example, a touch screen, and the audio component is configured to output and / or input audio signals. For example, the audio component can include a microphone configured to receive external audio signals. The received audio signals can be further stored in the memory 702 or transmitted through the communication component 705. The audio component further includes at least one speaker configured to output audio signals. The input / output interface 704 provides an interface between the processor 701 and other interface modules, which can be a keyboard, a mouse, a button, and the like. The buttons can be virtual buttons or physical buttons. The communication component 705 is configured to perform wired or wireless communication between the electronic device 700 and other devices. The wireless communication, such as Wi-Fi, Bluetooth, near field communication (NFC), 2G, 3G, 4G, NB-IOT, eMTC, or other 5G, and the like, or a combination of one or more of them, is not limited herein. Therefore, the communication component 705 can include, for example, a Wi-Fi module, a Bluetooth module, an NFC module, and the like.

[0147] In an exemplary embodiment, the electronic device 700 can be implemented by one or more Application Specific Integrated Circuits (ASICs), Digital Signal Processors (DSPs), Digital Signal Processing Devices (DSPDs), Programmable Logic Devices (PLDs), Field Programmable Gate Arrays (FPGAs), controllers, micro-controllers, microprocessors, or other electronic elements for performing the above-described content recommendation method.

[0148] In another exemplary embodiment, a computer-readable storage medium including program instructions that, when executed by a processor, implement the steps of the above-described content recommendation method is also provided. For example, the computer-readable storage medium can be the above-described memory 702 including program instructions that are executable by the processor 701 of the electronic device 700 to complete the above-described content recommendation method.

[0149] Figure 7 is a block diagram of an electronic device according to an exemplary embodiment. For example, the electronic device 1900 can be provided as a server. Referring to Figure 7 , the electronic device 1900 includes a processor 1922, the number of which can be one or more, and a memory 1932 for storing a computer program executable by the processor 1922. The computer program stored in the memory 1932 can include one or more modules each corresponding to a set of instructions. In addition, the processor 1922 can be configured to execute the computer program to perform the above-described content recommendation method.

[0150] In addition, the electronic device 1900 can further include a power supply component 1926 that can be configured to perform power management of the electronic device 1900, and a communication component 1950 that can be configured to implement communication of the electronic device 1900, for example, wired or wireless communication. In addition, the electronic device 1900 can further include an input / output (I / O) interface 1958. The electronic device 1900 can operate based on an operating system stored in the memory 1932, for example, Windows Server TM , Mac OS X TM , Unix TM , Linux TM , etc.

[0151] In another exemplary embodiment, a computer-readable storage medium including program instructions is also provided, which when executed by a processor implement the steps of the content recommendation method described above. For example, the computer-readable storage medium can be the memory 1932 described above including program instructions executable by the processor 1922 of the electronic device 1900 to complete the content recommendation method described above.

[0152] In another exemplary embodiment, a computer program product is also provided, which contains a computer program executable by a programmable apparatus, the computer program having code portions for performing the content recommendation method described above when executed by the programmable apparatus.

[0153] The preferred embodiments of the present disclosure are described in detail above with reference to the accompanying drawings, but the present disclosure is not limited to the specific details in the above-described embodiments. Various simple modifications can be made to the technical solutions of the present disclosure within the scope of the technical concepts of the present disclosure, and all these simple modifications shall fall within the protection scope of the present disclosure.

[0154] It should be further noted that each of the various technical features described in the above specific embodiments can be combined in any suitable manner, provided that there is no contradiction. In order to avoid unnecessary repetition, the present disclosure will not make further descriptions on various possible combinations.

[0155] Furthermore, any combination of the various different embodiments of the present disclosure can also be made, provided that it does not deviate from the idea of the present disclosure, and it shall be considered as the disclosed content of the present disclosure.

Claims

1. A content recommendation method characterized by, The method comprises the following steps: in response to receiving a content recommendation request of a target object, obtaining multi-type behavior information of the target object on recommended content, content feature information of to-be-recommended content, and attribute information of the target object, wherein the multi-type behavior information comprises explicit feedback sequences and implicit feedback sequences; determining feedback levels of each type of behavior information in the multi-type behavior information; determining behavior feedback information of the target object according to the feedback levels, the multi-type behavior information, the content feature information, and the attribute information of the target object; determining target recommended content recommended to the target object according to the behavior feedback information and the content feature information; the step of determining the behavior feedback information of the target object according to the feedback levels, the multi-type behavior information, the content feature information, and the attribute information of the target object comprises the following steps: determining a cross-feature vector of each type of behavior information according to the feedback levels, the multi-type behavior information, the content feature information, and a deep cross-feature extraction network of a prediction model; determining the behavior feedback information of the target object according to the attribute information of the target object and the cross-feature vector of each type of behavior information; the step of determining the cross-feature vector of each type of behavior information according to the feedback levels, the multi-type behavior information, the content feature information, and the deep cross-feature extraction network of the prediction model comprises the following steps: inputting each type of behavior information into a Transform module in the deep cross-feature extraction network respectively to obtain a high-order vector corresponding to each type of behavior information; obtaining a cross-feature vector corresponding to each type of behavior information in turn according to the high-order vector corresponding to each type of behavior information, the content feature information, and an Attention module in the deep cross-feature extraction network according to the size of the feedback levels; the multi-type behavior information comprises positive behavior feedback sequences and negative behavior feedback sequences; the feedback levels comprise a first feedback level corresponding to the positive behavior feedback sequences determined according to the positive behavior feedback sequences, and a second feedback level corresponding to the negative behavior feedback sequences determined according to the negative behavior feedback sequences; the step of obtaining the cross-feature vector corresponding to each type of behavior information in turn according to the high-order vector corresponding to each type of behavior information, the content feature information, and the Attention module in the deep cross-feature extraction network according to the size of the feedback levels comprises the following steps: for each type of behavior information, inputting the high-order vector corresponding to the type of behavior information and the content feature information into the Attention module in the deep cross-feature extraction network to obtain a first cross-feature vector of the type of behavior information; respectively determining the first feedback level and the second feedback level as a target level, and performing the following steps: sorting each type of behavior information corresponding to the target level according to the size of the target level to obtain a sorting list corresponding to the target level; inputting the high-order vector corresponding to the second type of behavior information located at the second position in the ranking list and the first cross-feature vector corresponding to the first type of behavior information located at the first position in the ranking list into the Attention module to obtain a second cross-feature vector corresponding to the second type of behavior information; and inputting the high-order vector corresponding to the i-th type of behavior information located at the i-th position in the ranking list, the first cross-feature vector corresponding to the first type of behavior information, and the second cross-feature vector corresponding to each type of behavior information located between the first position and the i-th position in the ranking list into the Attention module to obtain a second cross-feature vector corresponding to the i-th type of behavior information, where i ranges from 3 to N, and N is the number of categories of the multi-type behavior information.

2. The method of claim 1, wherein, The behavior feedback information of the target object is determined according to the attribute information of the target object and the cross-feature vectors of each type of behavior information, including: determining, according to the first cross-feature vector corresponding to the first type of behavior information located at the first position in the ranking list corresponding to the target level and the second cross-feature vectors corresponding to the behavior information located at positions other than the first position, a first type of behavior feedback vector of the target object corresponding to the target level; and determining, according to the first cross-feature vectors corresponding to the behavior information located at positions other than the first position in the ranking list corresponding to the target level, a second type of behavior feedback vector of the target object corresponding to the target level; determining the behavior feedback information of the target object according to the first type of behavior feedback vector, the second type of behavior feedback vector, and the attribute information of the target object.

3. The method of claim 2, wherein, When the target level is the first feedback level, the first type of behavior feedback vector is a first positive feedback vector, and the second type of behavior feedback vector is a second positive feedback vector; when the target level is the second feedback level, the first type of behavior feedback vector is a first negative feedback vector, and the second type of behavior feedback vector is a second negative feedback vector; and the like degree represented by the first positive feedback vector is greater than the like degree represented by the second positive feedback vector, and the dislike degree represented by the first negative feedback vector is greater than the dislike degree represented by the second negative feedback vector.

4. The method according to any one of claims 1-3, characterized in that, The explicit feedback sequence includes a sharing behavior feedback sequence, a collection behavior feedback sequence, and a dislike behavior feedback sequence; and the implicit feedback sequence includes a click behavior feedback sequence, an exposure and non-click behavior feedback sequence, and a short-time exit behavior feedback sequence.

5. A content recommendation apparatus characterized by comprising: The method comprises: The acquisition module is configured to, in response to receiving a content recommendation request of a target object, acquire multi-type behavior information of the target object on recommended content, content feature information of to-be-recommended content, and attribute information of the target object, where the multi-type behavior information includes explicit feedback sequences and implicit feedback sequences. The first determination module is configured to determine feedback levels of each type of behavior information in the multi-type behavior information. The behavior feedback information of the target object is determined according to the attribute information of the target object and the cross-feature vectors of each type of behavior information, including: determining, according to the first cross-feature vector corresponding to the first type of behavior information located at the first position in the ranking list corresponding to the target level and the second cross-feature vectors corresponding to the behavior information located at positions other than the first position, a first type of behavior feedback vector of the target object corresponding to the target level; and determining, according to the first cross-feature vectors corresponding to the behavior information located at positions other than the first position in the ranking list corresponding to the target level, a second type of behavior feedback vector of the target object corresponding to the target level; determining the behavior feedback information of the target object according to the first type of behavior feedback vector, the second type of behavior feedback vector, and the attribute information of the target object. When the target level is the first feedback level, the first type of behavior feedback vector is a first positive feedback vector, and the second type of behavior feedback vector is a second positive feedback vector; when the target level is the second feedback level, the first type of behavior feedback vector is a first negative feedback vector, and the second type of behavior feedback vector is a second negative feedback vector; and the like degree represented by the first positive feedback vector is greater than the like degree represented by the second positive feedback vector, and the dislike degree represented by the first negative feedback vector is greater than the dislike degree represented by the second negative feedback vector. The explicit feedback sequence includes a sharing behavior feedback sequence, a collection behavior feedback sequence, and a dislike behavior feedback sequence; and the implicit feedback sequence includes a click behavior feedback sequence, an exposure and non-click behavior feedback sequence, and a short-time exit behavior feedback sequence. The method comprises: The acquisition module is configured to, in response to receiving a content recommendation request of a target object, acquire multi-type behavior information of the target object on recommended content, content feature information of to-be-recommended content, and attribute information of the target object, where the multi-type behavior information includes explicit feedback sequences and implicit feedback sequences. The first determination module is configured to determine feedback levels of each type of behavior information in the multi-type behavior information. The second determining module is configured to determine behavior feedback information of the target object according to the feedback level, the multi-type behavior information, the content feature information, and attribute information of the target object; The third determining module is configured to determine target recommended content recommended to the target object according to the behavior feedback information and the content feature information; The determining of the behavior feedback information of the target object according to the feedback level, the multi-type behavior information, the content feature information, and the attribute information of the target object includes: The cross feature vector of each type of behavior information is determined according to the feedback level, the multi-type behavior information, the content feature information, and a deep cross feature extraction network of a prediction model; The behavior feedback information of the target object is determined according to the attribute information of the target object and the cross feature vector of each type of behavior information; the determining of the cross feature vector of each type of behavior information according to the feedback level, the multi-type behavior information, the content feature information, and the deep cross feature extraction network of the prediction model includes: Each type of behavior information is input into a Transform module in the deep cross feature extraction network to obtain a high-order vector corresponding to each type of behavior information; According to the high-order vector corresponding to each type of behavior information, the content feature information, and an Attention module in the deep cross feature extraction network, a cross feature vector corresponding to each type of behavior information is sequentially obtained according to the size of the feedback level; the multi-type behavior information includes a positive behavior feedback sequence and a negative behavior feedback sequence; the feedback level includes a first feedback level corresponding to the positive behavior feedback sequence determined according to the positive behavior feedback sequence, and a second feedback level corresponding to the negative behavior feedback sequence determined according to the negative behavior feedback sequence; The sequentially obtaining of the cross feature vector corresponding to each type of behavior information according to the high-order vector corresponding to each type of behavior information, the content feature information, and the Attention module in the deep cross feature extraction network according to the size of the feedback level includes: For each type of behavior information, the high-order vector corresponding to the type of behavior information and the content feature information are input into the Attention module in the deep cross feature extraction network to obtain a first cross feature vector of the type of behavior information; The first feedback level and the second feedback level are respectively determined as a target level, and the following steps are performed: According to the size of the target level, each type of behavior information corresponding to the target level is sorted to obtain a sorting list corresponding to the target level; inputting the high-order vector corresponding to the second type of behavior information located at the second position in the ranking list and the first cross-feature vector corresponding to the first type of behavior information located at the first position in the ranking list into the Attention module to obtain a second cross-feature vector corresponding to the second type of behavior information; and inputting the high-order vector corresponding to the i-th type of behavior information located at the i-th position in the ranking list, the first cross-feature vector corresponding to the first type of behavior information, and the second cross-feature vector corresponding to each type of behavior information located between the first position and the i-th position in the ranking list into the Attention module to obtain a second cross-feature vector corresponding to the i-th type of behavior information, where i is in a range of 3 to N, and N is a type quantity of the multi-type behavior information.

6. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program, when executed by a processor, implements the steps of the method of any one of claims 1-4.

7. An electronic device, comprising: comprising: a memory having a computer program stored thereon; a processor configured to execute the computer program in the memory to implement the steps of the method of any one of claims 1-4.

Images