A website illustration recommendation method, device and equipment and readable storage medium
A website illustration recommendation model trained with deep learning and convolutional neural networks calculates recommendation scores based on user history logs and content tags, solving the problem of lack of personalization in website illustration display and improving click-through rates and user satisfaction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BANK OF CHINA
- Filing Date
- 2023-06-30
- Publication Date
- 2026-07-24
AI Technical Summary
In existing technologies, the display of website illustrations lacks personalization and cannot provide personalized recommendations based on the interests and needs of different users, resulting in low click-through rates.
By acquiring the access requests and identifiers of target users, and utilizing a pre-trained website illustration recommendation model, a recommendation score is calculated based on the user's historical data to recommend website illustrations that match their interests. This model employs deep learning and convolutional neural networks, combining user historical behavior logs and content tags, to calculate target and auxiliary loss functions for model training, thereby improving recommendation accuracy.
It enables personalized recommendations of website illustrations based on user interests and needs, thereby increasing click-through rates and preventing user churn.
Smart Images
Figure CN116842256B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of data processing technology, specifically to a method, apparatus, device, and readable storage medium for recommending website illustrations. Background Technology
[0002] In website page layout, website illustrations (banner images) often occupy an important position on the website page. Website illustrations can attract users to click, promote user retention, and thus generate revenue.
[0003] Currently, website illustrations are displayed according to preset website illustration display rules. However, different users have different needs for website content; some like finance, while others like travel. Therefore, there is an urgent need for a website illustration display method to increase the click-through rate of website illustrations. Summary of the Invention
[0004] In view of this, this application provides a method, apparatus, device, and readable storage medium for recommending website illustrations, which enables the displayed website illustrations to meet the interests and needs of users and increase the click-through rate of website illustrations.
[0005] To solve the above problems, the technical solution provided in this application is as follows:
[0006] In a first aspect, this application provides a method for recommending website illustrations, the method comprising obtaining an access request from a target user, the access request from the target user including a target user identifier;
[0007] Based on the target user identifier, a pre-trained website illustration recommendation model is determined;
[0008] Obtain illustration data from websites to be recommended;
[0009] The website illustration data to be recommended is input into the pre-trained website illustration recommendation model, which is trained based on the target user's historical data.
[0010] Obtain the recommendation score output by the website recommendation model;
[0011] If the recommended score is greater than the preset score, then the website illustration data to be recommended is recommended to the target user.
[0012] In one possible implementation, the website illustration recommendation model is trained as follows:
[0013] Acquire training data, which is the historical behavior log of the target user on the target website illustration. The historical behavior log of the target user on the target website illustration includes: the historical dwell time of the target user on the target website illustration, the historical browsing time of the target user on the detail page of the target website illustration, the content tags corresponding to the target website illustration, and the identification information of the target website illustration.
[0014] The content tags corresponding to the website illustrations are assigned values using the random vector assignment method in deep learning, and the first target value corresponding to the content tags is determined.
[0015] Based on the target user's historical dwell time on the target website illustration, the identification information of the target website illustration, the target user's historical browsing time on the target website illustration details page, and the first target value corresponding to the content tag, a target embedding behavior sequence is determined;
[0016] The target website illustration is adjusted to the website illustration to be input based on preset pixel adjustment values;
[0017] Input the website illustration to be input into the convolutional neural network model;
[0018] Obtain the embedding content sequence output by the output layer of the convolutional neural network model;
[0019] Calculate the target loss function based on the target embedding behavior sequence and the embedding content sequence;
[0020] Calculate the auxiliary loss function based on the content tags corresponding to the illustrations on the target website;
[0021] The target loss function and the auxiliary loss function are used to train the model to be trained, thereby obtaining the website illustration recommendation model.
[0022] In one possible implementation, calculating the target loss function based on the target embedding behavior sequence and the embedding content sequence includes:
[0023] The target embedding behavior sequence and the embedding content sequence are concatenated to obtain the input vector;
[0024] The input vector is fed into a neural network model to obtain the target click-through rate.
[0025] If the target click-through rate is greater than the preset click-through rate standard value, then the target loss function is calculated based on the target click-through rate and the behavioral tags of the target user.
[0026] In one possible implementation, calculating the auxiliary loss function based on the content tags corresponding to the illustrations on the target website includes:
[0027] Obtain the target user's historical behavior logs on the comparison website illustrations. The target user's historical behavior logs on the comparison website illustrations include: the target user's historical dwell time on the comparison website illustrations, the target user's historical browsing time on the details page of the comparison website illustrations, the content tags corresponding to the target user's comparison website illustrations, and the identification information of the comparison website illustrations.
[0028] The content tags corresponding to the illustrations on the comparison website are assigned values using the random vector assignment method in deep learning, and the second target value corresponding to the content tags is determined.
[0029] Based on the target user's historical dwell time on the illustrations on the comparison website, the identification information of the illustrations on the comparison website, the target user's historical browsing time on the illustration details page of the comparison website, and the second target value corresponding to the content tag, a comparison embedding behavior sequence is determined;
[0030] Calculate the cosine similarity between the comparison embedding behavior sequence and the target embedding behavior sequence;
[0031] The degree of matching between the comparison embedding behavior sequence and the target embedding behavior sequence is determined using the cosine similarity.
[0032] Based on the matching degree, the log relationship between the target user's historical behavior logs on the comparison website illustrations and the target user's historical behavior logs on the target website illustrations is determined;
[0033] The auxiliary loss function is calculated based on the log relationships.
[0034] In one possible implementation, training the model to obtain the website illustration recommendation model using the target loss function and the auxiliary loss function includes:
[0035] The target loss function and the auxiliary loss function are weighted and fused to obtain a comprehensive loss function;
[0036] The model to be trained is trained using the comprehensive loss function to obtain the pre-trained website illustration recommendation model.
[0037] Secondly, this application also provides a website illustration recommendation device, the device comprising:
[0038] The first acquisition module is used to acquire the access request of the target user, wherein the access request of the target user includes the target user identifier;
[0039] The determination module is used to determine a pre-trained website illustration recommendation model based on the target user identifier;
[0040] The second acquisition module is used to acquire illustration data from the websites to be recommended.
[0041] The input module is used to input the website illustration data to be recommended into the pre-trained website illustration recommendation model, which is trained based on the historical data of the target user.
[0042] The third acquisition module is used to acquire the recommendation score output by the website recommendation model;
[0043] The recommendation module is used to recommend the website illustration data to the target user if the recommendation score is greater than a preset score.
[0044] In one possible implementation, the website illustration recommendation model is trained as follows:
[0045] Acquire training data, which is the historical behavior log of the target user on the target website illustration. The historical behavior log of the target user on the target website illustration includes: the historical dwell time of the target user on the target website illustration, the historical browsing time of the target user on the detail page of the target website illustration, the content tags corresponding to the target website illustration, and the identification information of the target website illustration.
[0046] The content tags corresponding to the website illustrations are assigned values using the random vector assignment method in deep learning, and the first target value corresponding to the content tags is determined.
[0047] Based on the target user's historical dwell time on the target website illustration, the identification information of the target website illustration, the target user's historical browsing time on the target website illustration details page, and the first target value corresponding to the content tag, a target embedding behavior sequence is determined;
[0048] The target website illustration is adjusted to the website illustration to be input based on preset pixel adjustment values;
[0049] Input the website illustration to be input into the convolutional neural network model;
[0050] Obtain the embedding content sequence output by the output layer of the convolutional neural network model;
[0051] Calculate the target loss function based on the target embedding behavior sequence and the embedding content sequence;
[0052] Calculate the auxiliary loss function based on the content tags corresponding to the illustrations on the target website;
[0053] The target loss function and the auxiliary loss function are used to train the model to be trained, thereby obtaining the website illustration recommendation model.
[0054] In one possible implementation, calculating the target loss function based on the target embedding behavior sequence and the embedding content sequence includes:
[0055] The target embedding behavior sequence and the embedding content sequence are concatenated to obtain the input vector;
[0056] The input vector is fed into a neural network model to obtain the target click-through rate.
[0057] If the target click-through rate is greater than the preset click-through rate standard value, then the target loss function is calculated based on the target click-through rate and the behavioral tags of the target user.
[0058] In one possible implementation, calculating the auxiliary loss function based on the content tags corresponding to the illustrations on the target website includes:
[0059] Obtain the target user's historical behavior logs on the comparison website illustrations. The target user's historical behavior logs on the comparison website illustrations include: the target user's historical dwell time on the comparison website illustrations, the target user's historical browsing time on the details page of the comparison website illustrations, the content tags corresponding to the target user's comparison website illustrations, and the identification information of the comparison website illustrations.
[0060] The content tags corresponding to the illustrations on the comparison website are assigned values using the random vector assignment method in deep learning, and the second target value corresponding to the content tags is determined.
[0061] Based on the target user's historical dwell time on the illustrations on the comparison website, the identification information of the illustrations on the comparison website, the target user's historical browsing time on the illustration details page of the comparison website, and the second target value corresponding to the content tag, a comparison embedding behavior sequence is determined;
[0062] Calculate the cosine similarity between the comparison embedding behavior sequence and the target embedding behavior sequence;
[0063] The degree of matching between the comparison embedding behavior sequence and the target embedding behavior sequence is determined using the cosine similarity.
[0064] Based on the matching degree, the log relationship between the target user's historical behavior logs on the comparison website illustrations and the target user's historical behavior logs on the target website illustrations is determined;
[0065] The auxiliary loss function is calculated based on the log relationships.
[0066] In one possible implementation, training the model to obtain the website illustration recommendation model using the target loss function and the auxiliary loss function includes:
[0067] The target loss function and the auxiliary loss function are weighted and fused to obtain a comprehensive loss function;
[0068] The model to be trained is trained using the comprehensive loss function to obtain the pre-trained website illustration recommendation model.
[0069] Thirdly, this application provides a website illustration recommendation device, including: a processor and a memory;
[0070] The memory is used to store one or more programs;
[0071] The processor is configured to implement the website illustration recommendation method as described in the first aspect above when the one or more programs are executed by the processor.
[0072] Fourthly, this application provides a computer-readable storage medium storing instructions that, when executed on a device, cause the device to perform the website illustration recommendation method described in the first aspect.
[0073] Therefore, this application has the following beneficial effects:
[0074] This application provides a website illustration recommendation method, apparatus, device, and readable storage medium. First, it obtains an access request from a target user, the access request including a target user identifier. Then, based on the target user identifier, it determines a pre-trained website illustration recommendation model, obtains website illustration data to be recommended, and inputs the website illustration data to be recommended into the pre-trained website illustration recommendation model. The pre-trained website illustration recommendation model is trained based on the target user's historical data. It obtains a recommendation score output by the website recommendation model. If the recommendation score is greater than a preset score, the website illustration data to be recommended is recommended to the target user. In this way, a recommendation score can be calculated for the website illustration data to be recommended using a pre-trained website illustration recommendation model corresponding to the user identifier. Website illustrations with recommendation scores meeting the preset score are then recommended to the user, ensuring that the recommended website illustrations are sufficiently attractive to the user. This makes the display of website illustrations match the user's interests and needs, increases the click-through rate of website illustrations, and prevents users from leaving due to not finding content of interest. Attached Figure Description
[0075] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0076] Figure 1 A schematic flowchart of a website illustration recommendation method provided in this application embodiment;
[0077] Figure 2 This is a schematic diagram of the device structure of a website illustration recommendation device provided in an embodiment of this application. Detailed Implementation
[0078] To facilitate understanding and explanation of the technical solutions provided in the embodiments of this application, the following will first describe the terms that may be involved in the embodiments of this application and the background technology of this application:
[0079] Banner images, also known as website illustrations, refer to multiple images displayed in a module or window, accessible by swiping with a finger on a mobile device or clicking with a mouse on a computer. These images are called banner images, and the module itself is called a banner image module.
[0080] Embedding (vector): This is a way to convert discrete variables or images into continuous vector representations. In neural networks, it not only reduces the spatial dimensionality of discrete variables but also provides a meaningful representation of the variable or image.
[0081] Banner images (website illustrations) often occupy prominent positions on web pages, playing a crucial role in attracting user clicks, promoting user retention, and ultimately generating revenue. Different users have different needs for website content; some are interested in finance, others in travel. Therefore, a banner image that is sufficiently attractive to users can match their interests, achieving a personalized experience for each individual, thus ensuring good conversion rates.
[0082] Currently, the most widely used method is to display banner images (website illustrations) manually. This means that the same website illustration is displayed to different users, which lacks personalization. As a result, when the displayed website illustration is not related to the content that users want to browse, users will ignore it directly, and the promotional effect will not be achieved.
[0083] To address these issues, this application provides a website illustration recommendation method, apparatus, device, and readable storage medium. First, it obtains an access request from a target user, including a target user identifier. Then, based on the target user identifier, it determines a pre-trained website illustration recommendation model, obtains website illustration data to be recommended, and inputs this data into the pre-trained model. The pre-trained model is trained based on the target user's historical data. It then obtains a recommendation score output by the model. If the recommendation score is greater than a preset score, the website illustration data to be recommended is recommended to the target user. In this way, a pre-trained website illustration recommendation model corresponding to the user identifier can calculate a recommendation score for the website illustration data to be recommended, and then recommend website illustrations with scores meeting the preset score to the user. This ensures that the recommended website illustrations are sufficiently attractive to the user, thereby making the display of website illustrations match the user's interests and needs, increasing the click-through rate of website illustrations, and preventing users from being lost due to not finding content of interest.
[0084] To facilitate understanding of the technical solutions provided in the embodiments of this application, the following description, in conjunction with the accompanying drawings, illustrates a behavior recognition method and apparatus provided in the embodiments of this application.
[0085] First, see Figure 1 The figure is a schematic flowchart of a website illustration recommendation method provided in an embodiment of this application, which specifically includes S101-S106.
[0086] S101: Obtain the access request of the target user, wherein the access request of the target user includes the target user identifier.
[0087] In this step, after receiving the user's access request, the user identifier included in the access request can be further obtained. The user identifier can be used for user authentication. Each different user has a different user identifier, thus ensuring the uniqueness of the user's identity.
[0088] S102: Determine the pre-trained website illustration recommendation model based on the target user identifier.
[0089] After obtaining the user identifier, the pre-trained website illustration recommendation model can be determined based on the user identifier. In other words, the pre-trained website illustration recommendation model corresponds to the user who made the access request. It should be noted that since the user frequently visits this webpage, a website illustration recommendation model needs to be built for this user. This ensures that the website illustration recommendations received by the user match the user's browsing habits on the website, improving the matching degree between website illustration recommendations and user interests, thereby increasing the click-through rate of website illustrations and preventing users from leaving due to not finding content of interest.
[0090] S103: Obtain illustration data from the websites to be recommended.
[0091] In this embodiment, website illustration data is obtained, which can be used to push information to users. The website illustration data can be obtained through web scraping or extracted directly from existing website illustration data. No specific limitations are imposed on the method of obtaining the website illustration data.
[0092] Understandably, in this step, a website image (banner image) library is established, selecting poster images from recent event launches or news and information as the recommended online image data. These images are then tagged with content-related tags, such as credit card tags, investment fund tags, and travel information tags. The image library needs to be updated regularly to ensure the timeliness of the image data.
[0093] S104: Input the website illustration data to be recommended into the pre-trained website illustration recommendation model, which is trained based on the target user's historical data.
[0094] After obtaining website illustration data, the data can be input into a pre-trained website illustration recommendation model. This model is trained based on the user's historical data, meaning it can output website illustration recommendation information relevant to the user.
[0095] After establishing the website illustration library, sub-images belonging to the website illustration (banner image) are further generated. For example, a website illustration (banner image) with a size of 1000*800*3 pixels cannot be directly input into the trained website illustration recommendation model because its size is larger than 224. Therefore, the image is first horizontally flipped 180 degrees. Then, 10 new images of 224*224*3 pixels each are randomly cropped from the original website illustration (banner image) and the flipped image and stored in the sub-illustration library belonging to the website illustration (banner image). For website illustrations (banner images) with a size smaller than 224 pixels, the images are flipped 90 degrees, 180 degrees, and 270 degrees, and zeros are added to the pixels around the images to expand the image size to 224*224*3, generating a total of 3 sub-images. Then, the generated sub-images are input into the trained website illustration recommendation model.
[0096] This application provides a training method for a website illustration recommendation model, as detailed below.
[0097] S105: Obtain the recommendation score output by the website recommendation model.
[0098] After inputting the website illustration data into the website recommendation model, the recommendation score output by the website recommendation model is obtained. It can be understood that the recommendation score output by the website recommendation model can be used to determine whether to recommend the website illustration data to the user.
[0099] S106: If the recommended score is greater than the preset score, then the website illustration data to be recommended is recommended to the target user.
[0100] In this embodiment, a pre-set score is compared with the recommendation score output by the website recommendation model. If the recommendation score output by the website recommendation model is greater than the pre-set score, the website illustration data can be recommended to the user; if the recommendation score output by the website recommendation model is less than the pre-set score, the website illustration data cannot be recommended to the user. For example, if the pre-set score is 0.5, the website illustration data will be recommended to the user if the recommendation score is greater than 0.5. If the recommendation score is less than 0.5, the website illustration data will not be recommended to the user. In this way, the website illustrations recommended to the user match the user's interests and needs, increasing the click-through rate of the website illustrations and achieving personalized display of website illustrations (banner images) for each user.
[0101] Based on the relevant content in S101-S106, a pre-trained website illustration recommendation model corresponding to user identifiers can be used to calculate recommendation scores for website illustration data to be recommended. Then, website illustrations with recommendation scores that meet the preset scores are recommended to users. This ensures that the website illustrations recommended to users are sufficiently attractive to them, thereby making the display of website illustrations match users' interests and needs, increasing the click-through rate of website illustrations, and preventing users from being unable to find content they are interested in and thus leaving the site.
[0102] This application provides a training method for a website illustration recommendation model, which is trained in the following manner, as shown in steps A1-A9:
[0103] A1: Obtain training data, which is the historical behavior log of the target user on the target website illustration. The historical behavior log of the target user on the target website illustration includes: the historical dwell time of the target user on the target website illustration, the historical browsing time of the target user on the detail page of the target website illustration, and the content tags corresponding to the target website illustration.
[0104] In this embodiment, training data can be collected using web crawler technology. Alternatively, a data acquisition system can be used to collect the training data. No limitation is placed on the method of acquiring the training data.
[0105] Understandably, the collected training data could be user click behavior logs on website banner images over 180 days, including the duration of time spent viewing the banner image page, the duration of time viewed on the banner image detail page, and the content tags corresponding to the banner image. The content tags corresponding to the banner image could be tags such as credit card tags, financial fund tags, or travel information tags. A2: The content tags corresponding to the banner image are assigned values using a random vector assignment method in deep learning to determine the first target value corresponding to the content tags.
[0106] Then, using the random vector assignment method in deep learning, an initial vector is assigned to the website illustrations that already have tags. This vector follows a Gaussian distribution.
[0107] A3: Based on the target user's historical dwell time on the target website illustration, the identification information of the target website illustration, the target user's historical browsing time on the target website illustration details page, and the first target value corresponding to the content tag, determine the target embedding behavior sequence.
[0108] In one possible implementation, the user's historical dwell time on website illustrations, the website illustration's identification information, the user's historical browsing time on the website illustration's details page, and the first target value corresponding to the content tag of the website illustration are initialized with random vectors to obtain an embedding behavior sequence.
[0109] A4: Adjust the target website illustration to the website illustration to be input based on the preset pixel adjustment value.
[0110] Resize the website illustration to a new image with a size of 224*224*3 pixels.
[0111] A5: Input the website illustration to be input into the convolutional neural network model.
[0112] Load the pre-trained convolutional neural network model ResNet50 and input a new image of size 224*224*3 pixels into the ResNet50 convolutional neural network model. A6: Obtain the embedding content sequence output by the output layer of the convolutional neural network model.
[0113] Obtain the embedding sequence of the output layer of the ResNet50 convolutional neural network model.
[0114] A7: Calculate the target loss function based on the target embedding behavior sequence and the embedding content sequence.
[0115] The target embedding behavior sequence and the embedding content sequence are concatenated to obtain the input vector;
[0116] The input vector is fed into a neural network model, and the target click-through rate (CTR), i.e., the predicted CTR value, can be obtained through the CTR formula in the neural network model. The specific formula is as follows:
[0117] y = DNN(concat(E) behavior E content ))
[0118] Among them is E behavior Behavior vector, E content y is the content vector, and y is the estimated click-through rate.
[0119] If the estimated click-through rate (CTR) is greater than the preset CTR standard value, it can be predicted that the user has clicked on the website illustration; if the estimated CTR is less than the preset CTR standard value, it can be predicted that the user has not clicked on the website illustration. Then, the user's behavioral tags for the website illustration are obtained. These behavioral tags can be used to determine whether the user clicked on the website illustration, and the main loss function is calculated.
[0120] A8: Calculate the auxiliary loss function based on the content tags corresponding to the illustrations on the target website.
[0121] Obtain the target user's historical behavior logs on the comparison website illustrations. The target user's historical behavior logs on the comparison website illustrations include: the target user's historical dwell time on the comparison website illustrations, the target user's historical browsing time on the details page of the comparison website illustrations, the content tags corresponding to the target user's comparison website illustrations, and the identification information of the comparison website illustrations.
[0122] The content tags corresponding to the illustrations on the comparison website are assigned values using the random vector assignment method in deep learning, and the second target value corresponding to the content tags is determined.
[0123] Based on the target user's historical dwell time on the illustrations on the comparison website, the identification information of the illustrations on the comparison website, the target user's historical browsing time on the illustration details page of the comparison website, and the second target value corresponding to the content tag, a comparison embedding behavior sequence is determined;
[0124] Calculate the cosine similarity between the comparison embedding behavior sequence and the target embedding behavior sequence;
[0125] The cosine similarity is used to determine the matching degree between the comparison embedding behavior sequence and the target embedding behavior sequence. Specifically, it is determined whether the cosine similarity between the comparison embedding behavior sequence and the target embedding behavior sequence is greater than a preset cosine similarity standard value. If the cosine similarity is greater than the preset cosine similarity standard value, then the historical behavior logs of the target website illustration and the historical behavior logs of the comparison website illustration are considered to have a positive sample relationship. An auxiliary loss function (InfoNCE loss function value) is then calculated based on this positive sample relationship.
[0126] If the cosine similarity is less than the preset cosine similarity standard value, then the illustrations on the target website and the illustrations on the comparison website are negative samples of each other, and the auxiliary loss function (InfoNCE loss function value) is calculated based on this negative sample relationship.
[0127] It is understood that in this embodiment, the matching degree between the two sampling results is modeled by unsupervised contrastive learning. In this way, the historical behavior logs of the target website illustration and the historical behavior logs of the comparison website illustration can be defined as having a positive sample relationship or a negative sample relationship. Based on the positive sample relationship or the negative sample relationship, an auxiliary loss function (InfoNCE loss function value) is calculated.
[0128] A9: Using the target loss function and the auxiliary loss function, train the model to be trained to obtain the website illustration recommendation model.
[0129] The target loss function (main loss function) and the auxiliary loss function are weighted and fused to obtain a comprehensive loss function, which is the loss function of the final trained model. The model is then trained using this comprehensive loss function to obtain the website illustration recommendation model.
[0130] It should be noted that the website illustration recommendation method and apparatus provided in this application can be used in the financial field or other fields. The above is only an example and does not limit the application field of the website illustration recommendation method and apparatus provided in this application.
[0131] The specific implementation processes and derivative methods of the above embodiments are all within the protection scope of this application.
[0132] and Figure 1 Corresponding to the method described above, this application also provides a website illustration recommendation device for recommending website illustrations. Figure 1 The specific implementation of the method, the website illustration recommendation device provided in this application embodiment, can be applied to computer terminals or various mobile devices, and its structural schematic diagram is as follows. Figure 2 As shown, the website's illustration recommendation device specifically includes:
[0133] The first acquisition module 201 is used to acquire the access request of the target user, wherein the access request of the target user includes the target user identifier;
[0134] The determination module 202 is used to determine a pre-trained website illustration recommendation model based on the target user identifier;
[0135] The second acquisition module 203 is used to acquire illustration data of the website to be recommended.
[0136] Input module 204 is used to input the website illustration data to be recommended into the pre-trained website illustration recommendation model, which is trained based on the historical data of the target user;
[0137] The third acquisition module 205 is used to acquire the recommendation score output by the website recommendation model;
[0138] The recommendation module 206 is used to recommend the website illustration data to the target user if the recommendation score is greater than a preset score.
[0139] In one possible implementation, the website illustration recommendation model is trained as follows:
[0140] Acquire training data, which is the historical behavior log of the target user on the target website illustration. The historical behavior log of the target user on the target website illustration includes: the historical dwell time of the target user on the target website illustration, the historical browsing time of the target user on the detail page of the target website illustration, and the content tags corresponding to the target website illustration.
[0141] The content tags corresponding to the website illustrations are assigned values using the random vector assignment method in deep learning, and the first target value corresponding to the content tags is determined.
[0142] Based on the target user's historical dwell time on the target website illustration, the identification information of the target website illustration, the target user's historical browsing time on the target website illustration details page, and the first target value corresponding to the content tag, a target embedding behavior sequence is determined;
[0143] The target website illustration is adjusted to the website illustration to be entered based on preset pixel adjustment values;
[0144] Input the website illustration to be input into the convolutional neural network model;
[0145] Obtain the embedding content sequence output by the output layer of the convolutional neural network model;
[0146] Calculate the target loss function based on the target embedding behavior sequence and the embedding content sequence;
[0147] Calculate the auxiliary loss function based on the content tags corresponding to the illustrations on the target website;
[0148] The target loss function and the auxiliary loss function are used to train the model to be trained, thereby obtaining the website illustration recommendation model.
[0149] In one possible implementation, calculating the target loss function based on the target embedding behavior sequence and the embedding content sequence includes:
[0150] The target embedding behavior sequence and the embedding content sequence are concatenated to obtain the input vector;
[0151] The input vector is fed into a neural network model to obtain the target click-through rate.
[0152] If the target click-through rate is greater than the preset click-through rate standard value, then the target loss function is calculated based on the target click-through rate and the behavioral tags of the target user.
[0153] In one possible implementation, calculating the auxiliary loss function based on the content tags corresponding to the illustrations on the target website includes:
[0154] Obtain the target user's historical behavior logs on the comparison website illustrations. The target user's historical behavior logs on the comparison website illustrations include: the target user's historical dwell time on the comparison website illustrations, the target user's historical browsing time on the details page of the comparison website illustrations, the content tags corresponding to the target user's comparison website illustrations, and the identification information of the comparison website illustrations.
[0155] The content tags corresponding to the illustrations on the comparison website are assigned values using the random vector assignment method in deep learning, and the second target value corresponding to the content tags is determined.
[0156] Based on the target user's historical dwell time on the illustrations on the comparison website, the identification information of the illustrations on the comparison website, the target user's historical browsing time on the illustration details page of the comparison website, and the second target value corresponding to the content tag, a comparison embedding behavior sequence is determined;
[0157] Calculate the cosine similarity between the comparison embedding behavior sequence and the target embedding behavior sequence;
[0158] The degree of matching between the comparison embedding behavior sequence and the target embedding behavior sequence is determined using the cosine similarity.
[0159] Based on the matching degree, the log relationship between the target user's historical behavior logs on the comparison website illustrations and the target user's historical behavior logs on the target website illustrations is determined;
[0160] The auxiliary loss function is calculated based on the log relationships.
[0161] In one possible implementation, training the model to obtain the website illustration recommendation model using the target loss function and the auxiliary loss function includes:
[0162] The target loss function and the auxiliary loss function are weighted and fused to obtain a comprehensive loss function;
[0163] The model to be trained is trained using the comprehensive loss function to obtain the pre-trained website illustration recommendation model.
[0164] Thirdly, this application provides a website illustration recommendation device, including: a processor and a memory;
[0165] The memory is used to store one or more programs;
[0166] The processor is configured to implement the website illustration recommendation method as described in any of the above embodiments when the one or more programs are executed by the processor.
[0167] Fourthly, this application provides a computer-readable storage medium storing instructions that, when executed on a device, cause the device to perform the website illustration recommendation method described in any of the above embodiments.
[0168] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0169] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for recommending website illustrations, characterized in that, The method includes: Obtain the access request of the target user, wherein the access request of the target user includes the target user identifier; Based on the target user identifier, a pre-trained website illustration recommendation model is determined; Obtain illustration data from websites to be recommended; The website illustration data to be recommended is input into the pre-trained website illustration recommendation model, which is trained based on the target user's historical data. Obtain the recommendation score output by the website illustration recommendation model; If the recommended score is greater than the preset score, then the website illustration data to be recommended is recommended to the target user; The website illustration recommendation model is trained in the following manner: Acquire training data, which is the historical behavior log of the target user on the target website illustration. The historical behavior log of the target user on the target website illustration includes: the historical dwell time of the target user on the target website illustration, the historical browsing time of the target user on the detail page of the target website illustration, the content tags corresponding to the target website illustration, and the identification information of the target website illustration. The content tags corresponding to the website illustrations are assigned values using the random vector assignment method in deep learning, and the first target value corresponding to the content tags is determined. Based on the target user's historical dwell time on the target website illustration, the identification information of the target website illustration, the target user's historical browsing time on the target website illustration details page, and the first target value corresponding to the content tag, a target embedding behavior sequence is determined; The target website illustration is adjusted to the website illustration to be input based on preset pixel adjustment values; Input the website illustration to be input into the convolutional neural network model; Obtain the embedding content sequence output by the output layer of the convolutional neural network model; Calculate the target loss function based on the target embedding behavior sequence and the embedding content sequence; Auxiliary loss function is calculated based on the content tags corresponding to the illustrations on the target website. This calculation includes: obtaining the target user's historical behavior logs on the illustrations on the comparison website, including: the target user's historical dwell time on the illustrations, the target user's historical browsing time on the details page of the illustrations on the comparison website, the content tags corresponding to the illustrations on the comparison website, and the identification information of the illustrations on the comparison website; assigning values to the content tags corresponding to the illustrations on the comparison website using a random vector assignment method in deep learning to determine a second target value corresponding to the content tags; and determining the comparison value based on the target user's historical dwell time on the illustrations on the comparison website, the identification information of the illustrations on the comparison website, the target user's historical browsing time on the details page of the illustrations on the comparison website, and the second target value corresponding to the content tags. The process involves embedding a behavior sequence; calculating the cosine similarity between the comparison embedding behavior sequence and the target embedding behavior sequence; using the cosine similarity to determine the matching degree between the comparison embedding behavior sequence and the target embedding behavior sequence; determining whether the cosine similarity between the comparison embedding behavior sequence and the target embedding behavior sequence is greater than a preset cosine similarity standard value; if the cosine similarity is greater than the preset cosine similarity standard value, then the historical behavior logs of the target website illustration and the historical behavior logs of the comparison website illustration are considered to have a positive sample relationship, and the auxiliary loss function is calculated based on the positive sample relationship; if the cosine similarity is less than the preset cosine similarity standard value, then the target website illustration and the comparison website illustration are considered to have a negative sample relationship, and the auxiliary loss function is calculated based on the negative sample relationship. The target loss function and the auxiliary loss function are used to train the model to be trained, thereby obtaining the website illustration recommendation model.
2. The method according to claim 1, characterized in that, The step of calculating the target loss function based on the target embedding behavior sequence and the embedding content sequence includes: The target embedding behavior sequence and the embedding content sequence are concatenated to obtain the input vector; The input vector is fed into a neural network model to obtain the target click-through rate. If the target click-through rate is greater than the preset click-through rate standard value, then the target loss function is calculated based on the target click-through rate and the behavioral tags of the target user.
3. The method according to claim 1, characterized in that, The process of training the model to be trained using the target loss function and the auxiliary loss function to obtain the website illustration recommendation model includes: The target loss function and the auxiliary loss function are weighted and fused to obtain a comprehensive loss function; The model to be trained is trained using the comprehensive loss function to obtain the pre-trained website illustration recommendation model.
4. A website illustration recommendation device, characterized in that, The device includes: The first acquisition module is used to acquire the access request of the target user, wherein the access request of the target user includes the target user identifier; The determination module is used to determine a pre-trained website illustration recommendation model based on the target user identifier; The second acquisition module is used to acquire illustration data from the websites to be recommended. The input module is used to input the website illustration data to be recommended into the pre-trained website illustration recommendation model, which is trained based on the historical data of the target user. The third acquisition module is used to acquire the recommendation score output by the website illustration recommendation model; The recommendation module is used to recommend the website illustration data to the target user if the recommendation score is greater than a preset score. The website illustration recommendation model is trained in the following manner: Acquire training data, which is the historical behavior log of the target user on the target website illustration. The historical behavior log of the target user on the target website illustration includes: the historical dwell time of the target user on the target website illustration, the historical browsing time of the target user on the detail page of the target website illustration, the content tags corresponding to the target website illustration, and the identification information of the target website illustration. The content tags corresponding to the website illustrations are assigned values using the random vector assignment method in deep learning, and the first target value corresponding to the content tags is determined. Based on the target user's historical dwell time on the target website illustration, the identification information of the target website illustration, the target user's historical browsing time on the target website illustration details page, and the first target value corresponding to the content tag, a target embedding behavior sequence is determined; The target website illustration is adjusted to the website illustration to be input based on preset pixel adjustment values; Input the website illustration to be input into the convolutional neural network model; Obtain the embedding content sequence output by the output layer of the convolutional neural network model; Calculate the target loss function based on the target embedding behavior sequence and the embedding content sequence; Auxiliary loss function is calculated based on the content tags corresponding to the illustrations on the target website. This calculation includes: obtaining the target user's historical behavior logs on the illustrations on the comparison website, including: the target user's historical dwell time on the illustrations, the target user's historical browsing time on the details page of the illustrations on the comparison website, the content tags corresponding to the illustrations on the comparison website, and the identification information of the illustrations on the comparison website; assigning values to the content tags corresponding to the illustrations on the comparison website using a random vector assignment method in deep learning to determine a second target value corresponding to the content tags; and determining the comparison value based on the target user's historical dwell time on the illustrations on the comparison website, the identification information of the illustrations on the comparison website, the target user's historical browsing time on the details page of the illustrations on the comparison website, and the second target value corresponding to the content tags. The process involves embedding a behavior sequence; calculating the cosine similarity between the comparison embedding behavior sequence and the target embedding behavior sequence; using the cosine similarity to determine the matching degree between the comparison embedding behavior sequence and the target embedding behavior sequence; determining whether the cosine similarity between the comparison embedding behavior sequence and the target embedding behavior sequence is greater than a preset cosine similarity standard value; if the cosine similarity is greater than the preset cosine similarity standard value, then the historical behavior logs of the target website illustration and the historical behavior logs of the comparison website illustration are considered to have a positive sample relationship, and the auxiliary loss function is calculated based on the positive sample relationship; if the cosine similarity is less than the preset cosine similarity standard value, then the target website illustration and the comparison website illustration are considered to have a negative sample relationship, and the auxiliary loss function is calculated based on the negative sample relationship. The target loss function and the auxiliary loss function are used to train the model to be trained, thereby obtaining the website illustration recommendation model.
5. The apparatus according to claim 4, characterized in that, The step of calculating the target loss function based on the target embedding behavior sequence and the embedding content sequence includes: The target embedding behavior sequence and the embedding content sequence are concatenated to obtain the input vector; The input vector is fed into a neural network model to obtain the target click-through rate. If the target click-through rate is greater than the preset click-through rate standard value, then the target loss function is calculated based on the target click-through rate and the behavioral tags of the target user.
6. A website illustration recommendation device, characterized in that, include: Processor and memory; The memory is used to store one or more programs; The processor is configured to implement the method as described in any one of claims 1-3 when the one or more programs are executed by the processor.
7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed on the device, cause the device to perform the method as described in any one of claims 1-3.