Content recommendation method and device, computer equipment and storage medium

Abstract

The invention discloses a content recommendation method and device, computer equipment and a storage medium. The method comprises the steps that first description information of first recommendation content is acquired; encoding the first description information to obtain content features of the first description information and first information, the first information being used for representing the degree of association between different information segments in the first description information; determining the complexity of the first description information according to the first information; optimizing the first description information based on the complexity to obtain second description information; predicting first feedback information of the first recommended content based on the content features and the second description information; and based on the first feedback information, recommending the first recommendation content to the target object to improve the recommendation effect.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a content recommendation method, apparatus, computer device and storage medium (computer-readable storage medium). Background Art

[0002] Driven by digital transformation and big data technology, the advertising industry is constantly changing and developing, gradually moving from traditional models to intelligent and personalized ones. As an important part of the advertising industry, advertising recommendations have also attracted much attention for their development prospects.

[0003] At present, advertising recommendations generally rely on relevant staff to make decisions on information such as recommendation time, recommendation location and recommendation group based on experience and intuition, so as to recommend the designed advertising content to the terminals held by the selected recommendation group at the decided recommendation time and recommendation location. Summary of the invention

[0004] In response to the above technical problems, embodiments of the present application provide a content recommendation method, apparatus, computer device, and storage medium.

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

[0006] Obtaining first description information of the first recommended content;

[0007] Encoding the first description information to obtain content characteristics of the first description information and first information, where the first information is used to characterize the degree of association between different information segments in the first description information;

[0008] Determine the complexity of the first description information according to the first information;

[0009] Optimizing the first description information based on the complexity to obtain second description information;

[0010] Based on the content feature and the second description information, predict first feedback information of the first recommended content;

[0011] Based on the first feedback information, the first recommended content is recommended to the target object.

[0012] In a second aspect, an embodiment of the present application provides a content recommendation device, including:

[0013] An information acquisition module, used to acquire first description information of the first recommended content;

[0014] An information encoding module, used to encode the first description information to obtain content characteristics of the first description information and first information, wherein the first information is used to characterize the degree of association between different information segments in the first description information;

[0015] A complexity determination module, used to determine the complexity of the first description information according to the first information;

[0016] An information optimization module, used for optimizing the first description information based on the complexity to obtain second description information;

[0017] An information prediction module, configured to predict first feedback information of the first recommended content based on the content feature and the second description information;

[0018] The content recommendation module is used to recommend the first recommended content to the target object based on the first feedback information.

[0019] In a third aspect, an embodiment of the present application further provides a computer device, including a memory storing a plurality of instructions; a processor loads instructions from the memory to execute the steps of any content recommendation method provided in the embodiment of the present application.

[0020] In a fourth aspect, an embodiment of the present application further provides a computer-readable storage medium, which stores a plurality of instructions suitable for loading by a processor to execute the steps of any content recommendation method provided in an embodiment of the present application.

[0021] In a fifth aspect, an embodiment of the present application further provides a computer program product, including a computer program or instructions, which, when executed by a processor, implements the steps of any content recommendation method provided in the embodiment of the present application.

[0022] The scheme of the application embodiment is adopted, by obtaining first description information for describing the first recommended content, and encoding the first description information to obtain the content characteristics of the first description information and the first information, so as to analyze the first description information according to the degree of association between different information segments in the first information to clarify the complexity of the first description information, so that the first description information can be optimized according to the level of complexity to obtain second description information with moderate complexity, so that the first feedback information of the first recommended content after the content recommendation can be more accurately predicted through the second description information with moderate complexity. Therefore, when recommending the first recommended content to the target object, a suitable recommendation strategy can be selected based on the first feedback information, which is helpful to improve the recommendation effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0024] Figure 1 is a schematic diagram of an application environment of the content recommendation method provided in an embodiment of the present application;

[0025] Figure 2 is a schematic diagram of an application scenario of the content recommendation method provided in an embodiment of the present application;

[0026] Figure 3 This is a flowchart of an embodiment of the content recommendation method provided in the embodiments of the present application;

[0027] Figure 4 is a structural diagram of a content recommendation device provided in an embodiment of the present application;

[0028] Figure 5 It is a schematic diagram of the internal structure of the computer device provided in the embodiment of the present application. DETAILED DESCRIPTION

[0029] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present application. At the same time, in the description of the embodiments of the present application, the terms "first", "second", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance. Thus, the features defined as "first" and "second" may explicitly or implicitly include one or more features. In the description of the embodiments of the present application, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined.

[0030] In one embodiment of the present application, the content recommendation method can be run on a local terminal device or a server. When the game interaction method is run on a server, the method can be implemented and executed based on a cloud interaction system, wherein the cloud interaction system includes a server and a client device.

[0031] In order to better understand a content recommendation method, device, electronic device and storage medium provided in the embodiments of the present application, the application environment applicable to the embodiments of the present application is first described below.

[0032] See also Figure 1 , Figure 1 A schematic diagram of an application environment of the content recommendation method provided by an embodiment of the present application is shown. As an implementation mode, the content recommendation method and the training method of the neural network model provided by the embodiment of the present application can be applied to the same electronic device. The electronic device can be as follows: Figure 1 The server 110 shown in the figure can be connected to the terminal device 120 through a network. The network is used to provide a medium for a communication link between the server 110 and the terminal device 120. The network can include various connection types, such as a wired communication link, a wireless communication link, etc., which are not limited in the embodiments of the present application. Optionally, in other embodiments, the electronic device can also be a smart phone, a laptop computer, etc.

[0033] It should be understood that Figure 1 The server 110, network and terminal device 120 are merely illustrative. Depending on the implementation requirements, there may be any number of servers, networks and terminal devices. For example, the server 110 may be a physical server or a server cluster composed of multiple servers, and the terminal device 120 may be a mobile phone, a tablet, a desktop computer, a laptop computer, and the like. It is understood that the embodiments of the present application may also allow multiple terminal devices 120 to access the server 110 at the same time.

[0034] In some embodiments, the terminal device 120 can pre-design first description information corresponding to the first recommended content to be recommended. Furthermore, the terminal device 120 sends the first description information corresponding to the first recommended content to the server 110 through the network. After the server 110 receives the first description information corresponding to the first recommended content, the first description information can be processed by the content recommendation method described in the embodiment of the present application.

[0035] As another implementation, the content recommendation method and the training method of the neural network model provided in the embodiment of the present application can be applied to different electronic devices. For example, the content recommendation method is applied to electronic device A, and the training method of the neural network model is applied to electronic device B, etc. The embodiment of the present application does not limit the electronic devices to which the above two methods are applied.

[0036] The following is a detailed description in conjunction with the accompanying drawings. In this embodiment, the execution subject is a terminal device as an example. It should be noted that the description order of the following embodiments is not intended to limit the preferred order of the embodiments. Although the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in an order different from that shown in the accompanying drawings.

[0037] The content recommendation method of this embodiment can be applied to any scenario where content recommendation is required. For example, if the content delivered by this method is advertising content, it can be applied to the following scenarios: online advertising (such as display advertising, search engine advertising), social media advertising, video advertising, in-app advertising, or advertising in game application scenarios, etc. The specific settings can be made according to the needs and are not limited here. In addition, the advertising placement can be automatically purchased through real-time bidding (RTB) and programmatic trading to improve the efficiency and accuracy of content delivery. For example, Figure 2 As shown, advertisements can be recommended when playing videos on major video websites and online video platforms.

[0038] At present, content recommendation generally needs to rely on relevant staff to make decisions on information such as recommendation time, recommendation location and recommendation group based on experience and intuition, so as to recommend the designed advertising content to the terminals held by the selected delivery group at the decided delivery time and location.

[0039] However, when recommendations are made through manual decision-making, the advertising content of the recommended advertisements is already designed, and the corresponding recommendation time, recommendation location, recommendation group and other information have been set. As a result, the recommendation process is relatively rigid and cannot achieve adaptive adjustment and complexity management of the advertising content. As a result, the accuracy of content recommendations is low, resulting in poor content recommendation effects.

[0040] In this regard, the content recommendation method of this embodiment can obtain first description information used to describe the first recommended content, and encode the first description information to obtain the content characteristics of the first description information and the first information, so as to analyze the first description information according to the degree of association between different information segments in the first information to clarify the complexity of the first description information, and thus optimize the first description information according to the level of complexity to obtain second description information with moderate complexity, so that the first feedback information of the first recommended content after the content recommendation can be more accurately predicted through the second description information with moderate complexity. Therefore, when recommending the first recommended content to the target object, a suitable recommendation strategy can be selected based on the first feedback information, which is helpful to improve the recommendation effect.

[0041] Please refer to Figure 3 The specific process of the content recommendation method may be as follows: Steps 301 to 306, wherein:

[0042] Step 301: Obtain first description information of first recommended content.

[0043] The first recommended content is used to indicate the recommended content that can be delivered to the terminal held by the target object in the future, for example, to deliver advertising content to a certain target object in the future. The number of the first recommended content can be one, or two or more. When there are two or more first recommended content, the two or more first recommended content can be different advertisements for the same product, advertisements for different products of the same merchant, and other advertising content with certain relevance. The two or more first recommended content can also be advertising content corresponding to different products that are not related. The specific setting can be based on demand and is not limited here.

[0044] Among them, the above-mentioned target object can be used to specifically refer to a user, or can be used to generally refer to multiple users of a certain type. For example, "technology enthusiasts" can be used as the type of target object to select multiple users. For another example, "young mothers" can be used as the type of target object to select multiple users.

[0045] Specifically, labels indicating user types may be pre-configured for different users, such as "technology enthusiast" or "young mother". When there is a demand for delivery, multiple users under the same label may be selected as target objects based on the user-configured labels, so that the corresponding first recommended content may be delivered to the target objects.

[0046] It should be noted that the description information is used to describe the delivery content, so that the corresponding delivery content can be generated based on the description information. For example, for advertisements, the advertising content when the advertisement is displayed can be generated based on the description information. In the language model, the description information can be input into the language model as a prompt instruction so that the language model outputs the first recommended content; the description information can also be sent to the manual platform so that the corresponding staff can clarify the information contained in the first recommended content based on the description information obtained on the manual platform, or adjust the description information to obtain the delivery content that conforms to the adjusted description information. The specific setting can be made according to the needs and is not limited here.

[0047] Accordingly, the first description information is used to describe the first recommended content, that is, the content displayed in the terminal held by the target object. In order to achieve the description of the first recommended content, the first description information may include description information of at least one type of information, or the first description information may include keywords used to describe the first recommended content, so that relevant personnel or relevant processing devices can clearly identify the corresponding first recommended content based on the keywords in the first description information after obtaining the first description information. In this embodiment, the first recommended content is clarified by obtaining the first description information corresponding to the first recommended content, and then the first recommended content is adjusted by adjusting the first description information.

[0048] It is understandable that different first recommended contents correspond to different first description information.

[0049] The description information of the above-mentioned information types includes, but is not limited to, description text of text type, description image of image type, video material of video type, audio material of audio type, etc.

[0050] It is understandable that in order to achieve better content recommendation effects, the description information of the first recommended content can be optimized. However, since the first recommended content of an advertisement contains description information of different information types, and the description information under the same information type contains information describing different aspects of the product, it will undoubtedly take a lot of time and effort to directly adjust the first recommended content manually. Therefore, in this embodiment, the first recommended content of each information type can be simplified to generate the first description information, so that by optimizing the first description information, the first recommended content can be optimized.

[0051] In some embodiments, the obtaining of the first description information of the first recommended content may be to obtain description information of different information types respectively, so as to form the first description information from the description information of each information type obtained respectively.

[0052] The above-mentioned information acquisition method may be to search for information associated with the first recommended content from a preset database, or to generate corresponding information in real time based on the first recommended content, which may be set according to the needs and is not limited here. In addition, in order to describe the first recommended content more accurately and concisely, the acquired information may be processed so that the processed description information of each information type forms the first description information.

[0053] Exemplarily, the above-mentioned method of processing the acquired information may be to perform content extraction on the description information of an information type to obtain the extracted descriptor information, then splice the descriptor information to obtain the spliced ​​information, and finally perform information enhancement processing on the spliced ​​information to obtain enhanced information, and use the enhanced information as the processed description information of the corresponding information type to form the first description information, that is, the first description information contains enhanced information of at least one information type.

[0054] Specifically, if the first description information includes a description text of a text type, then the first description information of the first recommended content obtained may include:

[0055] First, at least one text description content of the above-mentioned first recommended content can be obtained by searching from a preset text material library or generating it in real time. The at least one text description content includes but is not limited to text describing the product itself, text describing the purpose of the product, and text describing the characteristics of the product audience, etc. The specific setting can be based on demand and is not limited here.

[0056] Then, keywords can be extracted from each text description content respectively, that is, a brief description of each text description content can be obtained, and the corresponding extracted keywords are the brief description, and the extracted keywords are spliced ​​to obtain the second information of the text type of the above-mentioned first recommended content.

[0057] Finally, the content recommendation information of the first recommended content is obtained, and based on the content recommendation information, the second information is enhanced to obtain the third information of the text type of the first recommended content, where the third information is the description text of the text type, and based on the third information, the first description information of the first recommended content is determined, that is, the third information can be used as at least part of the first description information.

[0058] Among them, the above-mentioned content recommendation information includes but is not limited to content recommendation time, content recommendation channel and / or recommendation audience positioning, etc. The content recommendation channel is used to indicate the recommendation scene information, and the recommendation audience positioning is used to indicate the user type of content recommendation, such as the above-mentioned "technology enthusiasts" and "young mothers".

[0059] The above-mentioned information enhancement processing includes but is not limited to information embedding, synonym replacement, sentence conversion, enhancement based on historical performance after content recommendation and / or user feedback on the recommended content, etc. Exemplarily, the second information can be embedded based on at least part of the information in the content recommendation information to obtain embedded information, and the embedded information is further processed to obtain the third information.

[0060] Furthermore, in order to better process the data and improve the efficiency and accuracy of data processing, the descriptive information and / or content recommendation information of at least one of the above information types can be preprocessed, and the data preprocessing method includes but is not limited to a combination of at least one method such as data cleaning, data enhancement and feature engineering.

[0061] Among them, the data cleaning includes but is not limited to filling missing data, data repair, removing outliers in the data, and standardizing data (such as normalization, encoding, etc.); the data enhancement includes but is not limited to data enhancement of each type of descriptive information in a corresponding manner, such as, the text enhancement method includes but is not limited to synonym replacement, sentence conversion, etc., to increase the diversity of the descriptive text, and for example, the image enhancement method includes but is not limited to rotation, cropping, color adjustment, etc., to expand the sample space describing the image; the feature engineering can be feature extraction of the descriptive information of each information type, such as extracting content type, content features, etc., which can be set according to needs and is not limited here.

[0062] Exemplarily, it is assumed that there are two first recommended contents, namely, first recommended content A and first recommended content B. The relevant information of the first recommended content A and the first recommended content B is as follows:

[0063] The relevant descriptive information of the first recommended content A includes: description text "the latest smart watch, monitor your health", description image "watch picture", video material "demonstration video A"; the content recommendation information of the first recommended content A includes: content recommendation time "2024-06-01", content recommendation channel "social media, e-commerce platform", and recommended audience positioning "technology enthusiasts".

[0064] The relevant descriptive information of the first recommended content B includes: description text "Comfortable baby stroller, let the baby enjoy every trip", description image "baby stroller picture", video material "Demonstration video B"; the content recommendation information of the first recommended content A includes: content recommendation time "2024-06-02", content recommendation channel "mother and baby community, family shopping website", and recommended audience positioning "young mothers".

[0065] Firstly, data enhancement is performed on the first recommended content A and the first recommended content B respectively. That is, the description text of the first recommended content A is transformed into “a new smart watch for monitoring health”, the description image “watch picture” is rotated 10 degrees, the brightness is adjusted, and the advertisement type “electronic products” is extracted through feature engineering. Also, the description text of the first recommended content B is transformed into “a baby stroller that allows babies to travel comfortably”, the description image “baby stroller picture” is cropped to different angles, and the advertisement type “maternal and child products” is extracted through feature engineering.

[0066] Secondly, the description information of each information type of the first recommended content A and the first recommended content B is processed respectively to obtain first description information including enhanced information of each information type. This example is explained by taking the description text as an example:

[0067] For the first recommended content A: keyword extraction is performed on "new smart watch for monitoring health" to obtain the second information of the first recommended content A "smart watch, monitoring health", and based on part of the information in the content recommendation information of the first recommended content A, information is embedded in the second information of the first recommended content A "smart watch, monitoring health" to obtain the embedded information of the first recommended content A "the latest smart watch for technology enthusiasts, monitoring your health", and the embedded information of the first recommended content A "the latest smart watch for technology enthusiasts, monitoring your health" is further processed to obtain the enhanced information of the first recommended content A "this latest smart watch is suitable for users who like technology, and it can comprehensively monitor health status", that is, the enhanced information of the first recommended content A "this latest smart watch is suitable for users who like technology, and it can comprehensively monitor health status" is at least part of the information in the first description information of the first recommended content A.

[0068] For the first recommended content B: perform keyword extraction on "baby stroller that makes baby travel comfortably" to obtain the second information of the first recommended content B "baby stroller, comfortable travel", and based on part of the information in the content recommendation information of the first recommended content B, perform information embedding on the second information of the first recommended content B "baby stroller, comfortable travel" to obtain the embedded information of the first recommended content B "comfortable baby stroller for young mothers, allowing babies to enjoy every trip", and further process the embedded information of the first recommended content B "comfortable baby stroller for young mothers, allowing babies to enjoy every trip" to obtain the enhanced information of the first recommended content B "this baby stroller is designed for young mothers, making every trip of the baby more comfortable", that is, the enhanced information of the first recommended content B "this baby stroller is designed for young mothers, making every trip of the baby more comfortable" is at least part of the information in the first description information of the first recommended content B.

[0069] Step 302: Encode the first description information to obtain content characteristics of the first description information and first information, where the first information is used to characterize the degree of association between different information segments in the first description information.

[0070] In this embodiment, the first description information can be encoded based on the attention mechanism to obtain a feature vector of the first description information, which is the content feature of the first description information, and the first information generated by encoding the first description information, which is information for the attention feature dimension and is used to characterize the degree of association between different information segments in the first description information.

[0071] It can be understood that the encoding is achieved by capturing key information in the first description information and converting the captured key information into an internal representation so that the auxiliary neural network model can better understand and process the data in subsequent processing.

[0072] It can be understood that, since the above-mentioned attention mechanism can clarify the correlation or importance between different elements in the sequence corresponding to the first description information, the first information obtained when encoded by the above-mentioned attention mechanism contains the attention allocated to each part of the first description information, that is, the degree of correlation between different information segments in the first description information.

[0073] In some embodiments, the above content features may be presented in the form of an information representation sequence, wherein the information representation sequence includes multiple information feature vectors. If the first description information includes description information of multiple information types, there may be multiple information representation sequences corresponding to the information types, each of which includes an information feature vector of the corresponding information type.

[0074] In some embodiments, the above-mentioned first information may include an attention score matrix. It should be noted that in the process of encoding the first description information, the first description information can be converted into a query vector (query, q), a key vector (key, k) and a value vector (value, v) through linear mapping parameters. These three vectors play different roles in the attention layer with an attention mechanism, such as the query vector is used to represent the target position, and the key vector and the value vector are used to compare and generate outputs. Then, by taking the element of each position in the sequence as q, calculating the similarity score (such as the dot product) between q and all k, an attention score is generated for each position in the sequence corresponding to the first description information. The score reflects the similarity score between different positions in the sequence and the target position, that is, the degree of association. Finally, the calculated attention score can be normalized by the softmax function to obtain the weight of each position, and the weights corresponding to each position in the sequence constitute the attention score matrix.

[0075] It should be noted that when there are multiple attention layers, each attention layer can be processed in parallel or in sequence using different linear mapping parameters to obtain the attention score matrix corresponding to each attention layer.

[0076] In some embodiments, since the above-mentioned first description information may include description information of multiple information types, accordingly, the above-mentioned first information includes fourth information corresponding one-to-one to the above-mentioned multiple information types, and the above-mentioned fourth information is used to characterize the degree of association between different information segments in the description information of the corresponding information type.

[0077] Furthermore, if the first description information includes description information of multiple information types, and multiple attention layers are used in encoding the description information of each information type, then each information type may correspond to multiple attention score matrices.

[0078] In some embodiments, the above-mentioned content recommendation method is implemented through a trained neural network model, the above-mentioned neural network model includes an encoder, and the above-mentioned encoder includes an attention layer; the above-mentioned encoding of the above-mentioned first description information to obtain the content characteristics of the above-mentioned first description information, and the first information, includes: encoding the above-mentioned first description information through the encoder to obtain the content characteristics of the above-mentioned first description information; extracting the first information generated by the above-mentioned attention layer encoding the above-mentioned first description information.

[0079] It should be noted that, since the encoder in the above neural network model includes at least one attention layer, when encoding the first description information, the encoder can use the attention layer to prompt the encoder to focus on a specific part of the first description information when encoding, that is, through the attention layer, the encoder can establish an association between each element in the sequence and other elements when processing the sequence corresponding to the first description information, and can adaptively capture the long-range dependencies between elements by calculating the relative importance between the elements. When multiple attention layers are included in the neural network model, different attention layers can prompt the encoder to focus on different specific parts of the first description information.

[0080] Specifically, since the first description information may include description information of multiple information types, and the encoding methods required for description information of different information types are different, an encoder, such as an autoencoder, can be set to respectively encode the description information of different information types in the first description information. Alternatively, the above-mentioned neural network model can also be set to include encoders corresponding to each information type, such as a text encoder for encoding text (such as a BERT model, a GPT model), an image encoder for encoding images (such as a ResNet model, a ViT model), etc., so as to encode the description information of the corresponding information type through the encoder corresponding to each information type. Since the attention layer is required during encoding, it is equivalent to encoding the description information of the corresponding information type through the attention layer of each encoder. The specific setting can be made according to the needs and is not limited here.

[0081] Exemplarily, based on the examples of the first recommended content A and the first recommended content B mentioned above, the description information of each information type in the first description information of the first recommended content A and the first recommended content B is encoded respectively. This example takes encoding of description text and description image as an example for explanation:

[0082] For the description text of the first recommended content A: a text encoder can be used to encode "This latest smart watch is suitable for users who like technology. It can comprehensively monitor health status", such as performing BERT encoding to obtain a text representation sequence [0.12, 0.34, -0.45, ..., 0.89] of the first recommended content A, that is, a context-related embedding representation of the first recommended content A is obtained. The text representation sequence includes a text feature vector under the text type of the first recommended content A.

[0083] For the description image of the first recommended content A: an image encoder can be used to encode the “watch picture” processed by the first recommended content A, such as ResNet encoding, to obtain an image representation sequence [0.23, -0.34, 0.45, ..., -0.89] of the first recommended content A, that is, the high-dimensional image features of the first recommended content A are obtained. The image representation sequence includes the image feature vector under the image type of the first recommended content A.

[0084] For the description text of the first recommended content B: a text encoder can be used to encode "This baby stroller is designed for young mothers, making every trip of the baby more comfortable as the first recommended content", such as BERT encoding, to obtain the text representation sequence of the first recommended content B [-0.23, 0.56, 0.67, ..., -0.12], that is, the context-related embedding representation of the first recommended content B is obtained, and the text representation sequence includes the text feature vector of the text type of the first recommended content B.

[0085] For the description image of the first recommended content B: an image encoder can be used to encode the “baby carriage picture” processed by the first recommended content B, such as ResNet encoding, to obtain an image representation sequence [-0.45, 0.67, -0.78, ..., 0.12] of the first recommended content B, that is, the high-dimensional image features of the first recommended content B are obtained. The image representation sequence includes an image feature vector under the image type of the first recommended content B.

[0086] Step 303: Determine the complexity of the first description information according to the first information.

[0087] The above complexity is used to characterize the complexity of the first description information when it is represented. Specifically, the complexity can be used to measure the content, hierarchical structure and relevance between information of the first description information.

[0088] It can be understood that, since the above-mentioned first information is used to characterize the degree of association between different information segments in the first description information, the complexity of the first description information can be determined based on the first information.

[0089] Specifically, the complexity can be determined by the length of the computing language required to describe the first description information, the organization method of the information in the first description information, the regularity of the first description information, and the discrete degree of the correlation between the information in the first description information (the discrete degree of the first information). The specific complexity can be set according to the needs and is limited here.

[0090] Exemplarily, the first information includes an attention matrix. The complexity of the first description information can be determined based on the first information. The attention score matrix can be equalized to obtain a parameter representing the degree of discreteness of each data in the attention score matrix. The parameter can be used as the complexity of the first description information, that is, the variance of the attention matrix can be calculated to determine the complexity of the first description information based on the variance. For example, the variance can be directly used as the complexity of the first description information. In addition, the complexity of the first description information can also be determined by calculating parameters such as the standard deviation, the out-of-population ratio, and the interquartile range.

[0091] Exemplarily, it is assumed that there are two first recommended contents, namely, the first recommended content C and the first recommended content D, and the attention score matrices generated by encoding the first description information of the first recommended content C and the first recommended content D in the attention layer are respectively extracted, and the variance of the corresponding attention score matrices is calculated to obtain:

[0092] The variance calculated for the description text "This latest smart watch is suitable for users who like technology, and it can comprehensively monitor health status" in the first description information of the first recommended content C is 0.05; the variance calculated for the description image "Watch images at different angles" in the first description information of the first recommended content C is 0.08.

[0093] The variance calculated for the description text "the latest smart watch, monitor your health" in the first description information of the first recommended content D is 0.15; the variance calculated for the description image "watch picture" in the first description information of the first recommended content D is 0.12.

[0094] In some embodiments, the above-mentioned first description information includes description information of multiple information types. Accordingly, the above-mentioned first information includes fourth information corresponding one-to-one to the above-mentioned multiple information types, and the above-mentioned first complexity may include sub-complexity corresponding one-to-one to the above-mentioned multiple information types; the above-mentioned sub-complexity is determined based on the fourth information of the corresponding information type.

[0095] Step 304: Optimize the first description information based on the complexity to obtain second description information.

[0096] It can be understood that the above-mentioned complexity can make it clear whether the first description information is simple or complex. If the first description information is too simple, that is, the complexity is low, insufficient information will lead to misunderstanding of the first recommended content, resulting in the final generated first recommended content, or the result of predicting the first recommended content based on the first description information, being far from the expectation. If the first description information is too complex, that is, the complexity is high, there will be errors in the focus on the first recommended content, that is, focusing on the less important parts of the first recommended content, which will also lead to the final generated first recommended content, or the result of predicting the first recommended content based on the first description information, being far from the expectation, thereby reducing the advertising effect.

[0097] Therefore, in order to ensure that the first recommended content meets expectations while the corresponding advertising description information is easy to understand and can provide sufficient information to enhance user understanding and participation, in this embodiment, the content or structure of the first description information can be adjusted based on complexity, that is, the content of the first description information can be simplified or enriched to increase or decrease the amount of information in the first description information, so as to obtain second description information with moderate complexity for the first recommended content. For example, when the complexity is high, the focus can be increased by simplifying the content, so as to pay more attention to the core area of ​​the advertising description information.

[0098] Furthermore, by adjusting the first description information to obtain the second description information, the situation in which the complexity of the candidate content cannot be efficiently managed and the inability to flexibly adapt to different scenarios and user needs can be avoided. The first description information is analyzed through the first information and automatically adjusted, which greatly improves the content recommendation effect and efficiency.

[0099] Among them, the above-mentioned method of simplifying the first description information includes but is not limited to removing redundant information, reducing description objects, etc. For example, the descriptive image contained in the first description information can reduce the elements or visual information in the image, and the descriptive text contained in the first description information can reduce modifiers; the method of enriching the first description information includes but is not limited to adding specific functions or context descriptions, etc., which can be set according to needs and are not limited here.

[0100] In some embodiments, when determining the complexity based on the variance of the above-mentioned attention matrix, the information adjustment rule corresponding to the preset threshold interval in which the variance is located can be determined, or the information adjustment rule corresponding to the comparison result of the variance with the preset threshold can be determined, so as to adjust the first description information based on the information adjustment rule to obtain the second description information.

[0101] Specifically, when the variance is greater than a preset threshold, the amount of information in the second description information is greater than the first description information; when the variance is less than or equal to the preset threshold, the amount of information in the second description information is less than the first description information.

[0102] It should be noted that when the variance is used to indicate complexity, the size of the variance reflects the complexity of the first descriptive information. If the variance is large, it means that the amount of information in the first descriptive information is small and the first descriptive information is relatively simple, which means that the model only pays close attention to a few specific parts of the first descriptive information. Therefore, the amount of information in the first descriptive information needs to be increased so that the amount of information in the second descriptive information is greater than that in the first descriptive information. If the variance is small, it means that the amount of information in the first descriptive information is large and the first descriptive information is relatively complex, which means that the model needs to pay balanced attention to all parts of the first descriptive information. Therefore, the amount of information in the first descriptive information needs to be reduced so that the amount of information in the second descriptive information is less than that of the first descriptive information.

[0103] In this regard, in this embodiment, a preset threshold is used to distinguish the complexity of the first description information. The threshold can be determined based on manual experience, experimental data, etc., and is not limited here.

[0104] In some embodiments, since there can be multiple attention layers mentioned above, and each attention layer has a corresponding attention score matrix, after obtaining the variance indicating the complexity of the first description information under each attention layer, the variance can be averaged to calculate the variance mean.

[0105] Furthermore, the variance mean obtained after averaging the variances of each attention layer can determine the information adjustment rule corresponding to the preset threshold interval in which the variance mean is located, or determine the information adjustment rule corresponding to the comparison result between the variance mean and the preset threshold, thereby adjusting the first description information based on the information adjustment rule to obtain the second description information.

[0106] In some embodiments, the above-mentioned first description information includes description information of multiple information types. Accordingly, the above-mentioned first information includes fourth information corresponding one-to-one to the above-mentioned multiple information types, and the above-mentioned first complexity may include sub-complexity corresponding one-to-one to the above-mentioned multiple information types; the above-mentioned sub-complexity is determined based on the fourth information of the corresponding information type.

[0107] Correspondingly, the above-mentioned first description information is optimized based on the above-mentioned complexity to obtain the second description information, including: for the fourth information of each information type, based on the complexity corresponding to the above-mentioned information type, the fourth information corresponding to the information type is adjusted to obtain the second description information.

[0108] Exemplarily, based on the examples of the first recommended content C and the first recommended content D mentioned above, and further setting the above preset threshold to 0.1, that is, the variance is greater than 0.1, indicating that the first description information is relatively simple, and the variance is less than or equal to 0.1, indicating that the first description information is relatively complex, then the description information of each information type in the first description information of the first recommended content C and the first recommended content D can be compared and analyzed with the preset threshold, and the comparison results are as follows:

[0109] For the description text of the first recommended content C: Since 0.05 is less than 0.1, it means that the description text of the first recommended content C is relatively complex. It is necessary to reduce the amount of information in the description text of the first recommended content C to obtain the description text in the second description information of the first recommended content C, that is, simplify the description text "This latest smart watch is suitable for users who like technology. It can comprehensively monitor health status" in the first description information of the first recommended content C to "This smart watch can monitor health". "This smart watch can monitor health" is the description text in the second description information of the first recommended content C.

[0110] For the description image of the first recommended content C: Since 0.08 is less than 0.1, it means that the description image of the first recommended content C is relatively complex. It is necessary to reduce the amount of information in the description image of the first recommended content C to obtain the description image in the second description information of the first recommended content C. For example, reduce the elements or visual information in the description image so that users can focus on the core area and change the complex background into a single background.

[0111] For the description text of the first recommended content D: Since 0.15 is greater than 0.1, it means that the description text of the first recommended content D is relatively simple. It is necessary to increase the amount of information in the description text of the first recommended content D to obtain the description text in the second description information of the first recommended content D, that is, enrich the description text "The latest smart watch, monitor your health" in the first description information of the first recommended content D to "This smart watch can monitor your heart rate and health 24 hours a day". "This smart watch can monitor your heart rate and health 24 hours a day" is the description text in the second description information of the first recommended content D.

[0112] For the description image of the first recommended content D: Since 0.12 is greater than 0.1, it means that the description image of the first recommended content D is relatively simple. It is necessary to increase the amount of information of the description image of the first recommended content D to obtain the description image in the second description information of the first recommended content D. For example, more function display pictures or explanatory labels can be added to the "watch picture".

[0113] Step 305: predicting first feedback information of the first recommended content based on the content feature and the second description information.

[0114] In this embodiment, the first feedback information of the first recommended content may be predicted based on the content features and the second description information, so as to clarify the feedback such as the performance and effect of the first recommended content after recommendation based on the first feedback information.

[0115] The first feedback information includes but is not limited to click rate, conversion rate, reach rate, second-jump rate, etc., which can be set according to needs and are not limited here.

[0116] Specifically, since the above-mentioned content features are content features of the first description information of the first recommended content, the content features can be updated through the second description information to predict the first feedback information of the above-mentioned first recommended content based on the updated content features; or, the above-mentioned content features and the second description information can be predicted respectively to obtain respective first feedback information, and then the final first feedback information can be obtained through mean processing based on the respective first feedback information. The specific setting can be made according to the needs and is not limited here.

[0117] Specifically, the first feedback information of the first recommended content may be predicted based on the content feature and the second description information through a neural network model.

[0118] It is understandable that, since the content features based on which the neural network model makes predictions may include content features corresponding to descriptive information of different information types, it is necessary to fuse the content features corresponding to descriptive information of different information types, such as using a joint attention mechanism to fuse the content features corresponding to descriptive information of different information types to generate a unified multimodal feature representation. The above-mentioned joint attention mechanism can be a cross-modal attention mechanism, such as the CLIP (Contrastive Language-Image Pre-Training) model.

[0119] Exemplarily, based on the above-mentioned examples of the first recommended content A and the first recommended content B, the text representation sequence and the image representation sequence of the first recommended content A and the first recommended content B may be fused as shown below:

[0120] For the first recommended content A: the text representation sequence [0.12, 0.34, -0.45, ..., 0.89] for the first recommended content A and the image representation sequence [0.23, -0.34, 0.45, ..., -0.89] for the first recommended content A are fused with the image features through the joint attention mechanism to obtain the fused content features [0.18, -0.12, 0.03, ..., 0.56] of the first recommended content A.

[0121] For the first recommended content B: the text representation sequence [-0.23, 0.56, 0.67, ..., -0.12] for the first recommended content B and the image representation sequence [-0.45, 0.67, -0.78, ..., 0.12] for the first recommended content B are fused with the text and image features through the joint attention mechanism to obtain the fused content features of the first recommended content B [-0.33, 0.45, 0.56, ..., -0.22].

[0122] In some embodiments, the first recommended content may be associated with a target object, and the terminal to which the target object belongs is used to receive the first recommended content. It can be seen that when recommending content, the content needs to be recommended to the target object, and the content needs to be recommended in a certain environment. Therefore, in order to ensure that the first recommended content can adapt to the user's interests and needs in real time, and that the first recommended content can adapt to the corresponding recommendation environment, improve the personalization of the content, and the accuracy of the content recommendation, so as to achieve better content recommendation effect, in this embodiment, the trained neural network model can be further based on the user characteristics of the target object and / or the recommendation environment characteristics of the first recommended content, combined with the content characteristics of the first description information and the second description information to obtain corresponding prediction results.

[0123] That is to say, in this embodiment, the neural network model further considers the user characteristics and / or the recommendation environment characteristics of the first recommended content during training, so as to promote how to allocate attention through the attention model to obtain the corresponding attention score matrix, and the attention score matrix incorporates the user's focus on the first descriptive information, thereby enhancing interactivity.

[0124] Specifically, the above-mentioned prediction of the first feedback information of the above-mentioned first recommended content based on the above-mentioned content characteristics and the above-mentioned second description information may include: obtaining recommendation-related characteristics of the above-mentioned first recommended content, the above-mentioned recommendation-related characteristics include user characteristics of the above-mentioned target object, and / or recommendation environment characteristics of the above-mentioned first recommended content; through the above-mentioned neural network model, based on the above-mentioned content characteristics, the above-mentioned second description information, and the above-mentioned recommendation-related characteristics, predicting the feedback information after the above-mentioned first delivered content is delivered, so as to obtain the first feedback information of the above-mentioned first delivered content.

[0125] Among them, the above-mentioned user characteristics include but are not limited to user portrait characteristics and user behavior characteristics; the above-mentioned environmental characteristics include but are not limited to recommended time period characteristics, recommended location characteristics, etc., which can be set according to specific needs and are not limited here.

[0126] Specifically, the above-mentioned user characteristics and / or recommended environment characteristics can be pre-stored in a preset space so that they can be directly used when the neural network model needs them; the above-mentioned user characteristics and / or recommended environment characteristics can also be input into the neural network model by inputting the user data of the target object and the scene information related to the first recommended content, so that the neural network model can extract features from the user data and scene information.

[0127] The above user data include but are not limited to basic user information (such as age, gender, region), behavioral data (such as browsing history, click behavior) and social data (such as followed brands, social interactions).

[0128] The above scenario information includes, but is not limited to, the location, weather, time, major events, etc. of the content recommendation, which can be set according to the needs and are not limited here. The above major events can be natural disasters, accidents, disasters, public health events, social cases, and other events that may cause large-scale social impacts.

[0129] Furthermore, in order to better process the data and improve the efficiency and accuracy of data processing, the above-mentioned user data and scene information can be preprocessed, and the data preprocessing method includes but is not limited to a combination of at least one method such as data cleaning and data enhancement.

[0130] Among them, the data cleaning includes but is not limited to filling missing data, data repair, removing outliers in the data, and standardizing data (such as normalizing age, encoding gender, etc.); the data enhancement includes but is not limited to data enhancement of data of different dimensions in a corresponding manner, such as text enhancement methods include but are not limited to synonym replacement, sentence conversion, etc., to increase the diversity of descriptive text, and image enhancement methods include but are not limited to rotation, cropping, color adjustment, etc., to expand the sample space describing the image. In addition, feature engineering can be used to extract features from user data and scene information after data preprocessing to obtain the above-mentioned user features and recommended environment features, which can be set according to needs and are not limited here.

[0131] For example, it is assumed that there are two users, namely user A and user B, and the relevant user data of user A and user B are as follows:

[0132] The user data of user A includes: age "28", gender "male", region "Beijing", browsing history "electronic products" and "sports equipment", click behavior "electronic product advertisements" and "sports equipment advertisements", and social interaction "follow technology brands".

[0133] User B's user data includes age "35", gender "female", region "Shanghai", browsing history "home decoration" and "maternal and child products", click behavior "home advertisements" and "maternal and child products advertisements", and social interaction "follow home brands".

[0134] And, it is assumed that there are two first recommended contents, namely, first recommended content E and first recommended content F, and the scene information corresponding to the first recommended content E and the first recommended content F is as follows:

[0135] The scene information of the first recommended content E: time "2024-06-01", location "Beijing", weather "sunny-temperature 30℃".

[0136] The scene information of the first recommended content F: time "2024-06-02", location "Shanghai", weather "cloudy - temperature 25℃".

[0137] Then, the above user data and scenario information can be preprocessed or feature extracted respectively. For example, the region values ​​of user A and user B can be standardized to "Beijing" and "Shanghai" respectively; or the missing brand information of user A and user B can be filled in respectively.

[0138] Then, feature engineering can be used to extract features from the user data and scenario information after data preprocessing. For example, feature engineering can be used to extract the user portrait "technology enthusiast" from user A's user data; and feature engineering can be used to extract the user portrait "young mother" from user B's user data.

[0139] Also, the encoding weather condition "the encoding of a sunny day" is extracted from the scene information of the first recommended content E through feature engineering; and the encoding weather condition "the encoding of a cloudy day" is extracted from the scene information of the first recommended content F through feature engineering.

[0140] Step 306: recommend the first recommended content to the target object based on the first feedback information.

[0141] In this embodiment, the first feedback information is obtained through prediction, so that when the first recommended content is recommended to the target object, a suitable recommendation strategy can be selected based on the first feedback information, which helps to improve the recommendation effect.

[0142] Exemplarily, the first recommended content for a specific audience can be recommended to the corresponding target object, such as the first recommended content corresponding to "This latest smart watch is suitable for users who like technology" can be recommended to the target object of type "technology enthusiast", and the first recommended content corresponding to "This baby stroller is designed for young mothers" can be recommended to the target object of type "young mother".

[0143] In some embodiments, since the performance of the first recommended content indicated by the first feedback information after recommendation may be better, worse, etc., the above-mentioned recommending the first recommended content to the target object based on the above-mentioned first feedback information may include: if the above-mentioned first feedback information meets the first preset recommendation condition, then based on the above-mentioned second description information, updating the above-mentioned first recommended content, obtaining the second recommended content, and recommending the above-mentioned second recommended content to the above-mentioned target object, thereby recommending the first recommended content while ensuring that the performance of the first recommended content indicated by the first feedback information after recommendation is good.

[0144] The first preset recommendation condition includes but is not limited to: the first feedback information meets the preset feedback threshold, the first feedback information is within the preset feedback threshold range, etc., which can be set according to needs and are not limited here.

[0145] It can be understood that since the second description information is information obtained by adjusting the first description information, the second description information optimizes the description of the first recommended content. Therefore, the first recommended content is updated through the second description information to obtain recommended content with better presentation effect, and then the updated second recommended content is recommended to the target object, which can improve the recommendation effect of the first recommended content.

[0146] Among them, the method of updating the first recommended content based on the second description information can be: input the second description information as a prompt instruction into the language model to make the language model output content, and use the content output by the language model as the updated first recommended content, that is, the second recommended content.

[0147] In some embodiments, the updating of the first recommended content based on the second description information to obtain the second recommended content may include: based on the scene information in the second description information, adding the scene content corresponding to the scene information to the first recommended content to obtain the second recommended content, so that the second recommended content contains environmental information to enhance the ability of the target object to associate the product used in the recommended content with the scene when viewing the second recommended content.

[0148] The above scenario information includes, but is not limited to, the location, weather, time, major events, etc. of the content recommendation or expected recommendation, which can be set according to the needs and are not limited here. The above major events can be natural disasters, accidents, public health events, social cases, and other events that may cause large-scale social impact.

[0149] Exemplarily, the information integrating the second description information and the scene information may be “In sunny weather, wear this smart watch to keep track of your health information at any time”, “Choose this stroller so that your baby can travel comfortably even on cloudy days”.

[0150] Correspondingly, the recommending the second recommended content to the target object includes: if the scene where the target object is located matches the scene indicated by the scene information, recommending the second recommended content to the target object.

[0151] Exemplarily, if the above-mentioned scene information is set to include two groups of information, one group: the weather is "sunny" and the location is "Beijing", and the other group: the weather is "cloudy" and the location is "Shanghai", in this regard, the above-mentioned method of recommending the adjusted first recommended content based on the scene information can be: on sunny days, the first recommended content corresponding to "In sunny weather, bring this smart watch to keep track of your health information at any time" is recommended to the target object of type "technology enthusiast" in "Beijing"; on cloudy days, the first recommended content corresponding to "Your baby can travel comfortably even on cloudy days, choose this stroller" is recommended to the target object of type "young mother" in "Shanghai".

[0152] In some embodiments, the above-mentioned first recommended content is recommended to the target object based on the above-mentioned first feedback information, including: if there are multiple first recommended contents, then at least one second recommended content whose first feedback information meets the second preset recommendation condition is selected from the multiple first recommended contents, and the second recommended content is recommended to the target object based on the second description information corresponding to the second recommended content, so as to ensure the recommendation effect of the recommended first recommended content by only recommending a part of the multiple first recommended contents.

[0153] Among them, the above-mentioned second preset recommendation condition includes but is not limited to: a preset number of recommended contents with the best first feedback information, recommended contents whose first feedback information meets a preset feedback threshold, etc., which can be set according to needs and are not limited here.

[0154] Specifically, recommending the second recommended content to the target object based on the second description information corresponding to the second recommended content may include: updating the second recommended content based on the second description information to obtain updated second recommended content; and recommending the updated second recommended content.

[0155] In some embodiments, the training process of the neural network model may include: obtaining model training data of multiple third recommended contents, wherein the model training data includes third description information and third feedback information of the third recommended contents, and the third feedback information is feedback information within a preset delivery period after the third recommended contents are delivered; based on the model training data of multiple third recommended contents, training the neural network model until the neural network model meets the preset model convergence conditions, thereby obtaining a trained neural network model.

[0156] Among them, the description information of the above-mentioned third recommended content can refer to the method of obtaining the above-mentioned first description information, and the above-mentioned third feedback information can be obtained by real-time monitoring of the data within the preset recommendation period of recommending the above-mentioned third recommended content, such as by real-time monitoring of the click rate and conversion rate within the preset recommendation period of recommending the above-mentioned third recommended content. For example, the click rate of the third recommended content A is 5%, and the conversion rate is 2%, and the click rate of the third recommended content B is 7%, and the conversion rate is 3%.

[0157] Among them, the classification loss and contrast loss can be used to determine whether the neural network model in the training process meets the preset model convergence conditions. For example, the classification loss can use the cross entropy loss for category prediction to measure the first feedback information, such as click-through rate and conversion rate, and use the contrast loss to maintain the distinction between different categories.

[0158] Optionally, the encoder of the neural network model may include multiple attention layers. In order to enable different attention layers to focus on different parts of the third descriptive information when the encoder of the neural network model is encoding, the model training data of multiple third recommended contents can be split into multiple groups of model training data of third recommended contents during training to train different sub-models in the neural network model respectively. Each sub-model can encode each descriptive information in the group for separate training. Finally, the outputs of multiple sub-models are integrated through weighted averaging to serve as the output of the model.

[0159] Optionally, when training the neural network model, gradient descent method and back propagation can be used for model training.

[0160] In some embodiments, a feature database can be constructed for the model training data of the third recommended content, so that data can be obtained from the feature database during the training and application of the neural network model. It can be understood that the feature database can include two types of feature databases, one is a source domain feature library, which contains the encoded features of the descriptive information of the third recommended content and the third feedback information, and the other is a target domain feature library, which only contains the encoded features of the descriptive information of the recommended content without the corresponding third feedback information, for example, the recommended content in a new field that currently needs to be predicted by the neural network model.

[0161] Among them, the expected performance of recommended content in new fields in the target domain feature library can be obtained by aligning the branch model and adjusting the features of the feature library.

[0162] It is understandable that, since the dimensions of source domain features and target domain features may be different, an alignment branch model, such as an image-guided feature adjustment (IFT) model, can be introduced to use contrastive learning methods (such as SimCLR) to align source domain features and target domain features. Among them, in order to improve domain adaptability, the model's adaptability to self-domain features can be enhanced through a loop feedback mechanism, and cross-domain common features can be integrated through feature alignment. In addition, when the neural network model fuses the content features of descriptive information of different information types, it may not be able to be effectively aligned, resulting in inconsistent information presentation and reduced advertising recommendation effects. Therefore, an alignment branch model can also be used to effectively align the content features of descriptive information of different information types and enhance the fusion effect of content features of descriptive information of multiple information types.

[0163] Exemplarily, the above source domain feature library is as follows:

[0164]

[0165] The above target domain feature library is as follows:.

[0166]

[0167] In some embodiments, when the first feedback information is obtained through a trained neural network model based on the content features and the second description information, the real-time predicted recommended content and the feedback information after the recommended content is recommended can be used as new samples to train the neural network model, such as using an incremental learning method to continuously update the model according to new data, so that the description information and the recommendation strategy of the corresponding recommended content can be adjusted in real time according to the feedback information after the content recommendation.

[0168] Specifically, after recommending the first recommended content to the target object based on the first feedback information, it may also include: obtaining second feedback information within a preset recommendation period after the first recommended content is recommended, so as to use the second description information and the content features as samples and the second feedback information as labels, input them into the neural network model, and update the neural network model.

[0169] In some embodiments, in order to ensure the effectiveness of content recommendation, the recommendation effects corresponding to the first description information and the second description information can be evaluated respectively through methods such as A / B testing, so as to optimize the recommended content and recommendation strategy based on the evaluation results, such as adjusting the prompt complexity or changing the recommendation channel.

[0170] Exemplarily, the first description information G is "the latest smart watch, monitor your health"; and the second description information H is "this latest smart watch is suitable for users who like technology, it can comprehensively monitor health status", monitor the click rate and conversion rate of the two groups, compare the effects, and select the description information with better effect based on the comparison results. For example, if the click rate of the second description information H is increased by 2% and the conversion rate is increased by 1% compared with the first description information G, then the second description information H is selected, that is, the more complex enhanced prompts are retained.

[0171] For example, the first description information E is "This stroller is designed for young mothers, making every trip more comfortable for the baby"; and the second description information F is "A comfortable stroller allows the baby to enjoy every trip." The click-through rate and conversion rate of the two groups are monitored, and the effects are compared. The description information with better effect is selected based on the comparison results. For example, if the click-through rate of the second description information F is increased by 1% compared with the first description information E, and the conversion rate has no significant change, the second description information F is selected, that is, the simpler basic prompts are retained.

[0172] In some embodiments, the above content recommendation method may further include the following steps:

[0173] Step 1: The user data, scenario information and description information of the target object and the content to be recommended can be obtained.

[0174] Step 2: Through the neural network model, the user data of the target object, the scene information and the description information of the content to be recommended are encoded respectively to extract the user characteristics of the target object, the recommendation environment characteristics of the scene information, and the content characteristics of the description information of the content to be recommended.

[0175] Step 3: Extract the attention score matrix generated when encoding the description information of the content to be recommended, and determine the complexity of the description information of the content to be recommended based on the attention score matrix, so as to update the description information based on the above complexity to obtain updated description information.

[0176] Step 4: Use the neural network model to predict user characteristics, recommendation environment characteristics, content characteristics and updated description information to obtain feedback information of the content to be recommended.

[0177] Step 5: Recommend the above-mentioned content to be recommended to the target object based on the feedback information.

[0178] It should be understood that, although each step in the flowcharts involved in the above-mentioned embodiments is displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0179] Based on the same inventive concept, the embodiment of the present application also provides a content recommendation device for implementing the content recommendation method involved above, and a content recommendation device for implementing the training method of the neural network model. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in the one or more content recommendation devices and the embodiments of the content recommendation device provided below can refer to the above limitations on the content recommendation method and the training method of the neural network model, and the specific limitations will not be repeated here.

[0180] This embodiment also provides a content recommendation device, which can be integrated into a terminal device. Figure 4 As shown, the content recommendation device may include:

[0181] The information acquisition module 401 is used to acquire first description information of the first recommended content;

[0182] An information encoding module 402 is used to encode the first description information to obtain content characteristics of the first description information and first information, where the first information is used to characterize the degree of association between different information segments in the first description information;

[0183] A complexity determination module 403, configured to determine the complexity of the first description information according to the first information;

[0184] An information optimization module 404 is used to optimize the first description information based on the complexity to obtain second description information;

[0185] An information prediction module 405, configured to predict first feedback information of the first recommended content based on the content feature and the second description information;

[0186] The content recommendation module 406 is configured to recommend the first recommended content to the target object based on the first feedback information.

[0187] In the above-mentioned content recommendation device, the first description information used to describe the first recommended content is obtained by the information acquisition module 401, and the first description information is encoded by the information encoding module 402 to obtain the content characteristics of the first description information and the first information, so that the complexity determination module 403 analyzes the first description information according to the degree of association between different information segments in the first information to clarify the complexity of the first description information, so that the information optimization module 404 can optimize the first description information according to the level of complexity to obtain the second description information with moderate complexity, so that the information prediction module 405 can more accurately predict the first feedback information of the first recommended content after the content recommendation through the second description information with moderate complexity. Therefore, when the content recommendation module 406 recommends the first recommended content to the target object, it can select a suitable recommendation strategy based on the first feedback information, which is helpful to improve the recommendation effect.

[0188] In some embodiments, the information acquisition module 401 is specifically used to:

[0189] Obtaining at least one text description content of the first recommended content;

[0190] Extracting keywords from each type of text description content respectively, and concatenating the extracted keywords to obtain second information of the text type of the first recommended content;

[0191] Obtaining content recommendation information of the first recommended content, and based on the content recommendation information, performing information enhancement processing on the second information to obtain third information of the text type of the first recommended content;

[0192] Based on the third information, first description information of the first recommended content is determined.

[0193] In some embodiments, the content recommendation module 406 is specifically used to:

[0194] If the first feedback information meets the first preset recommendation condition, then based on the second description information, the first recommended content is updated to obtain second recommended content;

[0195] The second recommended content is recommended to the target object.

[0196] In some embodiments, the content recommendation module 406 is specifically used to:

[0197] Based on the scene information in the second description information, adding the scene content corresponding to the scene information to the first recommended content to obtain the second recommended content;

[0198] If the scene where the target object is located matches the scene indicated by the scene information, the second recommended content is recommended to the target object.

[0199] In some embodiments, the content recommendation module 406 is specifically used to:

[0200] If there are multiple first recommended contents, selecting at least one second recommended content whose first feedback information meets a second preset recommendation condition from the multiple first recommended contents;

[0201] Based on the second description information corresponding to the second recommended content, the second recommended content is recommended to the target object.

[0202] In some embodiments, the first information includes an attention matrix, and the complexity determination module 403 is specifically used to:

[0203] Calculate the variance of the above attention matrix;

[0204] The complexity of the first description information is determined based on the variance.

[0205] In some embodiments, when the variance is greater than a preset threshold, the amount of information in the second description information is greater than the first description information; when the variance is less than or equal to the preset threshold, the amount of information in the second description information is less than the first description information.

[0206] In some embodiments, the first feedback information is obtained by a trained neural network model based on the content features and the second description information, and the content recommendation device further includes an updating module, which is specifically used to:

[0207] Obtaining second feedback information within a preset recommendation period after the first recommended content is recommended;

[0208] The second description information and the content features are used as samples, and the second feedback information is used as a label, which are input into the neural network model to update the neural network model.

[0209] Based on the same inventive concept, the embodiment of the present application also provides a computer device, which may be a server or a terminal device, and includes a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps of the above content recommendation method when executing the computer program. Thereby, various functions are realized, such as:

[0210] Obtaining first description information of the first recommended content;

[0211] Encoding the first description information to obtain content characteristics of the first description information and first information, where the first information is used to characterize the degree of association between different information segments in the first description information;

[0212] Determine the complexity of the first description information according to the first information;

[0213] Optimizing the first description information based on the complexity to obtain second description information;

[0214] Based on the content feature and the second description information, predict first feedback information of the first recommended content;

[0215] Based on the first feedback information, the first recommended content is recommended to the target object.

[0216] It can be seen that the computer program can be loaded by the processor to execute any content recommendation method provided in the embodiments of the present application, thereby bringing the following technical effects: the content recommendation effect can be improved.

[0217] The specific implementation of the above operations can be found in the previous embodiments, which will not be described in detail here.

[0218] In one embodiment, the computer device is a terminal device, for example, its internal structure diagram can be as follows: Figure 5As shown. The computer device includes a processor, a memory, an input / output interface, a communication interface, a display unit and an input device. The processor, the memory and the input / output interface are connected via a system bus, and the communication interface, the display unit and the input device are connected to the system bus via the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be implemented through WIFI, a mobile cellular network, NFC (near field communication) or other technologies. When the computer program is executed by the processor, a content recommendation method is implemented. The display unit of the computer device is used to form a visually visible image, and can be a display screen, a projection device or a virtual reality imaging device. The display screen can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer covered on the display screen, or a button, trackball or touchpad set on the computer device casing, or an external keyboard, touchpad or mouse, etc.

[0219] Those skilled in the art will understand that Figure 5 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0220] Based on the same inventive concept, an embodiment of the present application also provides a computer-readable storage medium, which may include: a read-only memory (ROM), a random access memory (RAM), a disk or an optical disk, etc.

[0221] Since the computer program stored in the computer-readable storage medium can execute any one of the content recommendation methods provided in the embodiments of the present application, the beneficial effects that can be achieved by any one of the content recommendation methods provided in the embodiments of the present application can be achieved. Please refer to the previous embodiments for details and will not be repeated here.

[0222] Based on the same inventive concept, the embodiment of the present application further provides a computer program product or a computer program, which includes a computer instruction, and the computer instruction is stored in a computer-readable storage medium. The processor of the computer device reads the computer instruction from the computer-readable storage medium, and the processor executes the computer instruction, so that the computer device executes the method provided in various optional implementations in the above embodiments.

[0223] It should be noted that the object data (including but not limited to user device information, user personal information, etc.) and conversation data involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data need to comply with relevant laws, regulations and standards of relevant countries and regions. A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods.

[0224] Any reference to the memory, database or other media used in the embodiments provided in the present application may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (Re RAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory may include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM may be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0225] The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. The non-relational database may include a distributed database based on blockchain, etc., but is not limited thereto. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but is not limited thereto.

[0226] In the above-mentioned content recommendation device, computer-readable storage medium, computer device, and computer program product embodiments, the description of each embodiment has its own emphasis. For parts not described in detail in a certain embodiment, reference can be made to the relevant description of other embodiments. Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working process and beneficial effects of the above-mentioned content recommendation device, computer-readable storage medium, computer program product, computer device, and its corresponding units can refer to the description of the content recommendation method in the above embodiment, and will not be repeated here.

[0227] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0228] The above is a detailed introduction to a content recommendation method, device, computer equipment, computer-readable storage medium and computer program product provided in the embodiments of the present application. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea; at the same time, for technical personnel in this field, according to the ideas of the present application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. A content recommendation method, characterized in that: The method comprises: Obtaining first description information of the first recommended content; Encoding the first description information to obtain content characteristics of the first description information and first information, where the first information is used to characterize the degree of association between different information segments in the first description information; determining the complexity of the first description information according to the first information; Optimizing the first description information based on the complexity to obtain second description information; Predicting first feedback information of the first recommended content based on the content feature and the second description information; Based on the first feedback information, the first recommended content is recommended to the target object.

2. The method according to claim 1, characterized in that The obtaining of first description information of the first recommended content includes: Obtaining at least one text description content of the first recommended content; Extracting keywords from each type of text description content respectively, and concatenating the extracted keywords to obtain second information of the text type of the first recommended content; Acquire content recommendation information of the first recommended content, and based on the content recommendation information, perform information enhancement processing on the second information to obtain third information of the text type of the first recommended content; Based on the third information, first description information of the first recommended content is determined.

3. The method according to claim 1, characterized in that The recommending the first recommended content to the target object based on the first feedback information includes: If the first feedback information meets the first preset recommendation condition, updating the first recommended content based on the second description information to obtain second recommended content; The second recommended content is recommended to the target object.

4. The method according to claim 3, characterized in that The updating the first recommended content based on the second description information to obtain the second recommended content includes: Based on the scene information in the second description information, adding the scene content corresponding to the scene information to the first recommended content to obtain the second recommended content; The recommending the second recommended content to the target object includes: If the scene where the target object is located matches the scene indicated by the scene information, the second recommended content is recommended to the target object.

5. The method according to claim 1, characterized in that The recommending the first recommended content to the target object based on the first feedback information includes: If there are multiple first recommended contents, selecting at least one second recommended content whose first feedback information meets a second preset recommendation condition from the multiple first recommended contents; Based on the second description information corresponding to the second recommended content, the second recommended content is recommended to the target object.

6. The method according to claim 1, characterized in that The first information includes an attention matrix, and determining the complexity of the first description information according to the first information includes: Calculating the variance of the attention matrix; The complexity of the first description information is determined based on the variance.

7. The method according to claim 6, characterized in that When the variance is greater than a preset threshold, the amount of information in the second description information is greater than that in the first description information; when the variance is less than or equal to the preset threshold, the amount of information in the second description information is less than that in the first description information.

8. The method according to claim 1, characterized in that The first feedback information is obtained by a trained neural network model based on the content feature and the second description information. After the first recommended content is recommended to the target object based on the first feedback information, the method further includes: Acquire second feedback information within a preset recommendation period after the first recommended content is recommended; The second description information and the content features are used as samples, and the second feedback information is used as a label, which are input into the neural network model to update the neural network model.

9. A content recommendation device, characterized in that: The device comprises: An information acquisition module, used to acquire first description information of the first recommended content; an information encoding module, configured to encode the first description information to obtain content characteristics of the first description information and first information, wherein the first information is used to characterize the degree of association between different information segments in the first description information; A complexity determination module, configured to determine the complexity of the first description information according to the first information; An information optimization module, configured to optimize the first description information based on the complexity to obtain second description information; an information prediction module, configured to predict first feedback information of the first recommended content based on the content feature and the second description information; A content recommendation module is used to recommend the first recommended content to a target object based on the first feedback information.

10. A computer device, characterized in that: The invention comprises a processor and a memory, wherein the memory stores a plurality of instructions; the processor loads instructions from the memory to execute the steps of the content recommendation method according to any one of claims 1 to 8.

11. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a plurality of instructions, and the instructions are suitable for being loaded by a processor to execute the steps of the content recommendation method according to any one of claims 1 to 8.

12. A computer software, characterized in that: The computer software includes a computer program, and the computer program is used by a processor to execute the steps of the content recommendation method according to any one of claims 1 to 8.

Images