Image recommendation methods, apparatus, devices, and computer-readable storage media

Abstract

This invention discloses an image recommendation method, apparatus, device, and computer-readable storage medium. The method includes: acquiring the exposure volume of an object, behavioral data corresponding to multiple images, and user image preference features; the multiple images representing the displayed images of the object, and the behavioral data representing the user's feedback on the exposed images; determining the quality score of each image based on the object's exposure volume and the behavioral data corresponding to the multiple images, with the quality score reflecting the image display effect; ranking the multiple images according to the user's image preference features and the quality scores of each image to obtain an image ranking result; and recommending images according to the image ranking result, ensuring that the image ranking seen by each user corresponds to their own preferences, thereby improving the accuracy of image recommendation.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to an image recommendation method, apparatus, device, and computer-readable storage medium. Background Technology

[0002] With the continuous development of big data and internet technology, online consumption is becoming more and more widespread, online shopping is becoming increasingly popular, and more and more online e-commerce platforms are emerging.

[0003] Existing technology calculates the similarity between purchased items and items in a database by analyzing users' historical browsing and purchasing behavior on e-commerce websites, and then recommends items with high similarity to the user.

[0004] However, existing technologies only recommend different items to different users, lacking a personalized distribution strategy for the images under the items. For the same item, different users often see the same item image, which lacks differentiated matching of user interests and reduces the accuracy of image recommendations. Summary of the Invention

[0005] This invention aims to provide an image recommendation method, apparatus, device, and computer-readable storage medium. By calculating the quality score of each image based on the exposure data of the object and the exposure data of each image, and combining the image preference features of each user, multiple images are sorted so that the image ranking seen by each user corresponds to their own preferences, thereby improving the accuracy of image recommendation.

[0006] The technical solution of this invention is implemented as follows:

[0007] In a first aspect, embodiments of the present invention provide an image recommendation method, the method comprising: acquiring the exposure volume of an object, behavioral data corresponding to multiple images, and user image preference features, wherein the multiple images represent display images introducing the object, and the behavioral data represent user feedback on the exposed images; determining the quality score of each image based on the exposure volume of the object and the behavioral data corresponding to the multiple images; sorting the multiple images according to the user image preference features and the quality scores of each image to obtain an image sorting result, and recommending images according to the image sorting result.

[0008] Secondly, embodiments of the present invention provide an image recommendation device, the device comprising: an acquisition module, configured to acquire the exposure volume of an object, behavioral data corresponding to multiple images, and image preference features of a user, wherein the multiple images represent display images introducing the object, and the behavioral data represent user feedback on the exposed images; a determination module, configured to determine the quality score of each image based on the exposure volume of the object and the behavioral data corresponding to the multiple images; and a sorting module, configured to sort the multiple images based on the image preference features of the user and the quality scores of each image to obtain an image sorting result, and to recommend images according to the image sorting result.

[0009] Thirdly, embodiments of the present invention provide an image recommendation device, the device including a memory for storing executable instructions and a processor for executing the executable instructions stored in the memory to implement the above-mentioned image recommendation method.

[0010] Fourthly, embodiments of the present invention provide a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, implement the above-described image recommendation method.

[0011] This invention provides an image recommendation method, apparatus, device, and computer-readable storage medium. According to the solution provided by this invention, the exposure volume of an object, behavioral data corresponding to multiple images, and the user's image preference characteristics are obtained. The multiple images represent the displayed images of the object, and the behavioral data represents the user's feedback on the exposed images. Based on the object's exposure volume and the behavioral data corresponding to the multiple images, a quality score is determined for each image, thereby completing the image exploration and utilization process. The quality score reflects the image display effect; the higher the quality score, the better the image display effect. Based on the user's image preference characteristics and the quality scores of each image, the multiple images are sorted to obtain an image ranking result. Image recommendations are then performed according to the image ranking result, ensuring that the image ranking seen by each user corresponds to their own preferences, thus improving the accuracy of image recommendations. Attached Figure Description

[0012] Figure 1 A flowchart illustrating optional steps of an image recommendation method provided in an embodiment of the present invention;

[0013] Figure 2 A flowchart of optional steps for another image recommendation method provided in an embodiment of the present invention;

[0014] Figure 3A This is an exemplary schematic diagram of a candidate main image provided in an embodiment of the present invention;

[0015] Figure 3B An exemplary schematic diagram of another candidate main image provided in an embodiment of the present invention;

[0016] Figure 3C An exemplary schematic diagram of another candidate main image provided for an embodiment of the present invention;

[0017] Figure 3D This is an exemplary schematic diagram of yet another candidate main image provided in an embodiment of the present invention;

[0018] Figure 4 A flowchart illustrating optional steps of another image recommendation method provided in an embodiment of the present invention;

[0019] Figure 5 A flowchart illustrating optional steps of another image recommendation method provided in an embodiment of the present invention;

[0020] Figure 6 This is a schematic diagram of the structure of an image recommendation device provided in an embodiment of the present invention;

[0021] Figure 7 This is a schematic diagram of the structure of an image recommendation device provided in an embodiment of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It should be understood that some embodiments described herein are merely illustrative of the technical solutions of the present invention and are not intended to limit the scope of the present invention.

[0023] To facilitate understanding of this solution, the relevant technologies in the embodiments of the present invention will be described before describing the embodiments of the present invention.

[0024] In the e-commerce sector, taking items as an example, personalized distribution technology for items is relatively mature. Item images are recommended to users during the personalized distribution process. As the most direct way to present an item, the image conveys more information to users or consumers in a shorter time compared to the title. High-quality images are especially important in e-commerce. However, currently, in various item recommendation scenarios, the same item is often presented with the same image, lacking differentiation based on user interests and reducing the accuracy of image recommendations.

[0025] In this embodiment of the invention, by extracting user preference features for different images, all image materials introducing a certain object are personalized and sorted. This upgrades static, identical images into dynamic, personalized recommendations that match user needs and personalizes traffic distribution for object images. The personalized distribution strategy for object images better stimulates user expression of needs, improves shopping efficiency and user experience, and enhances the accuracy of image recommendations.

[0026] The image recommendation method provided in this embodiment of the invention is also applicable to recommendation schemes for different image display formats such as video recommendation and animation recommendation. This image recommendation method can be applied to object recommendation scenarios and film and television recommendation scenarios in e-commerce platforms, and this embodiment of the invention does not limit these applications.

[0027] This invention provides an image recommendation method, such as... Figure 1 As shown, Figure 1 This is an optional flowchart of an image recommendation method provided in an embodiment of the present invention. The image recommendation method includes the following steps:

[0028] S101. Obtain the exposure of the object, behavioral data corresponding to multiple images, and the user's image preference characteristics. The multiple images represent the display images of the object, and the behavioral data represents the user's feedback on the exposed images.

[0029] In this embodiment of the invention, exposing an object to a user means displaying multiple images introducing the object to the user. By using user behavior data such as clicking on images and placing orders, the exposure and click count of each image are statistically analyzed to obtain the exposure of the object, the exposure of multiple images, and the click count of multiple images. The exposure of the object is equal to the sum of the exposure of the multiple images introducing the object. The behavior data corresponding to the image includes the exposure and click count of the image.

[0030] In this embodiment of the invention, images in e-commerce can be displayed in various ways, such as different background colors (e.g., white background, red background), the presence or absence of text, the proportion of text in the image, and the presence or absence of models. User image preference features are used to characterize the features of images that users tend to prefer when browsing, clicking, or placing an order. By combining user image preference features, images that match their preferences can be recommended to users, thereby improving the user experience. In this embodiment of the invention, preference can also be understood as liking, interest, inclination, etc.

[0031] In some embodiments, the user's image preference features in S101 above can be obtained in the following ways: Acquire user behavior data, which includes at least one of the following: the number of historical images clicked by the user, the dwell time of the user browsing historical images, and multiple images included in the objects corresponding to historical completed orders; extract features from the user's behavior data to obtain the user's image preference features, which include image scene type preferences and image content composition preferences.

[0032] In this embodiment of the invention, taking an item as an example, each user, as a living individual, has different interests, preferences, purchasing power, and other factors. Therefore, the acceptance level of the same image for the same item varies among different users and in different contexts. User behavior data represents user feedback on images. Clicked historical images can include clicked images, downloaded images, browsed images, saved images, liked images, and zoomed-in images. The time a user spends browsing historical images reflects their interest in those images. Before placing an order for an item, a user will at least browse some images describing the item. The images under the item corresponding to the completed orders in the past, to a certain extent, represent the user's interest in the images.

[0033] In this embodiment of the invention, user image preference features are obtained by extracting features from user behavior data. Feature extraction can be understood as classifying the composition of images. Image preference features include image scene type preferences (e.g., the image scene is the actual application scenario of the object, the image background is red, the image scene is to enhance the atmosphere of the image, etc.) and image content composition preferences (e.g., whether the image has text, whether the image has a model, whether the image has a comparison image, etc.).

[0034] In this embodiment of the invention, by extracting features from user behavior data, the accuracy of user image preference features is improved, so as to recommend images that are suitable for the user based on the user's image preference features.

[0035] S102. Determine the quality score of each image based on the exposure of the object and the behavioral data corresponding to multiple images.

[0036] In this embodiment of the invention, the recommendation system continuously recommends images to the user and collects user feedback information on multiple images during the recommendation process. This feedback information includes, but is not limited to, the exposure of the object, the exposure of multiple images, and the number of clicks on multiple images. Based on the recommendation algorithm and the aforementioned feedback information, the quality score of each image can be calculated, thereby completing the image exploration and utilization process.

[0037] It should be noted that the quality score mentioned above reflects the display effect of the corresponding image. The higher the quality score, the more exposure traffic (exposure times) is allocated to the image. The recommendation algorithm can be an E&E algorithm, which is an exploration and exploitation family of algorithms, including but not limited to one or more of the following algorithms: epsilon-greedy (ε-greedy), upper confidence bound (UCB), ranked bandits, contextual bandits, Thompson sampling, naive bandit, and softmax. This embodiment of the invention does not limit the specific algorithms used.

[0038] In some embodiments, the behavioral data corresponding to the multiple images includes the exposure of the multiple images and the number of clicks on the multiple images. The above S102 may also include S1021 and S1022.

[0039] S1021. Based on the exposure of the object and the behavioral data corresponding to multiple images, determine the click-through rate and traffic index of each image respectively.

[0040] S1022. Determine the quality score of each image based on its click-through rate and traffic index.

[0041] In this embodiment of the invention, when calculating the quality score of each image, the click-through rate and the traffic index of the image are determined based on the exposure of the object and the corresponding behavioral data. The click-through rate reflects the average revenue of the image, which can also be understood as the expected return of the image. The traffic index reflects the feedback information after the image is exposed, which can also be understood as the degree of exploration of the image. The click-through rate and the traffic index are added together to obtain the quality score of the image.

[0042] In this embodiment of the invention, the expected reward is "utilization" in the recommendation algorithm; the degree of exploration is "exploration" in the recommendation algorithm, which can be understood as the degree of uncertainty of the expected reward. If there is only "utilization," it is easy to get trapped in local extrema. The significance of "exploration" is that if an image is exposed too few times, the feedback information of the image is less, the confidence of the image's expected reward is very low, the uncertainty is very high, and the confidence interval is very large, which can also be understood as a large variance. The recommendation system does not believe that the expected reward of the image is its true expected reward. Therefore, it needs to expose the image to obtain more feedback information, repeating the process to complete the exploration and utilization process.

[0043] In some embodiments, S1021 described above can also be implemented in the following manner: The click-through rate of each image is determined based on the ratio of the click count to the exposure count of each image; the traffic index of each image is determined based on the exposure count of the object and the exposure count of each image, wherein the exposure count of the object is the sum of the exposure counts of multiple images.

[0044] In this embodiment of the invention, the click-through rate of each image is determined based on the ratio of the number of clicks to the number of exposures of each image; the logarithm of the sum of the exposures of the object and a first preset value is calculated based on the exposures of the object, where the exposure of the object is equal to the sum of the exposures of multiple images; the product of the logarithm and a second preset value is used as the numerator, and the sum of the exposures of each image and the first preset value is used as the denominator; the square root of the fraction formed by the numerator and the denominator is used as the traffic index of each image, where the traffic index reflects the degree of exploration of the image, and the degree of exploration characterizes the uncertainty of the expected return.

[0045] In this embodiment of the invention, E&E modeling is performed on all images of the introduced object, and the exposure and click volume of each image are counted in real time. The exposure volume of the object is equal to the sum of the exposure volumes of multiple images introducing the object. Based on the UCB mechanism of the E&E algorithm, all images of the introduced object are scored to obtain a quality score for each image. Specifically, in the exploration phase of the E&E algorithm, all images are guaranteed to have the opportunity to be displayed to the user and obtain corresponding feedback information; in the exploitation phase of the E&E algorithm, more exposure traffic is allocated to the images that have performed well so far (i.e., images with higher quality scores).

[0046] For example, the UCB mechanism of the E&E algorithm is used as an example for illustration. The UCB formula is shown in formula (1).

[0047]

[0048] In formula (1), ucb(i,j) represents the quality score of image j under object i, and sku_pv i Represents the cumulative exposure of object i to date, img_ctr j This represents the average revenue per unit of image j under object i to date. Average revenue per unit can be understood as click-through rate (CTR) or expected return. CTR is equal to the ratio of the cumulative clicks to the cumulative impressions of image j. (img_pv) j It represents the cumulative exposure of image j under object i up to date.

[0049] In this embodiment of the invention, the expected return is represented by the click-through rate and the exploration level is represented by the traffic index, as an example. The first part of formula (1) is img_ctrj This represents the expected return for image j under object i, which is the "exploitation" in the E&E algorithm; the latter half... The degree of exploration, or "exploration" in the E&E algorithm, can be understood as the degree of uncertainty of the expected reward. If only the first half of formula (1) exists, it is a pure exploitation, i.e., a greedy strategy, which is prone to getting trapped in local extrema. The second half of formula (1) can be understood as an indicator used to measure the image feedback information. The less feedback information there is, the larger the value corresponding to the second half. With the addition of this second half indicator, the UCB algorithm is curious. When there is not enough feedback information for an image, it will be selected, even if the expected reward of the image is very low.

[0050] It should be noted that the above formula (1) is illustrated with the first preset value being 1 and the second preset value being 2 as an example. The values ​​of the first preset value and the second preset value can also be other values. Specifically, those skilled in the art can make appropriate settings according to actual needs, as long as the degree of exploration can be guaranteed. This embodiment of the present invention does not limit this.

[0051] In this embodiment of the invention, the UCB mechanism based on the E&E algorithm scores all images introducing a certain object based on feedback information from multiple images, obtaining a quality score for each image. This achieves automatic traffic exploration of images during the cold start phase. The exploration phase favors images with lower cumulative exposure among all images of the same object. Once all images have been sufficiently exposed to the user, the utilization phase tends to display images with higher click-through rates, improving the accuracy of image recommendations.

[0052] S103. Based on the user's image preference features and the quality score of each image, sort the multiple images to obtain the image sorting results, and recommend images according to the image sorting results.

[0053] In this embodiment of the invention, after obtaining the quality scores of each image, the image ranking result can be determined in various ways by combining the user's image preference features. For example, it can be done through a preset calculation method or a neural network model (NN). When using a neural network model, the user's image preference features and the quality scores of each image are input into an image personalization model, which then ranks the multiple images and outputs the image ranking result.

[0054] For example, taking items as an example, the image sorting result is based on the image sorting score. In practical applications, the image with the highest sorting score is placed at the top of the item details page, and so on, completing the image recommendation for the user. Since the image sorting result corresponding to each user is related to their image preference characteristics, the order of images seen in the image recommendations is different for different users, improving the personalized display of image recommendations.

[0055] It should be noted that the aforementioned image personalization model can be a pre-trained image personalization model. This pre-trained model has the function of taking the user's image preference features and the quality scores of each image as input, and outputting the corresponding image ranking results. The specific implementation and processing of the image personalization model are not limited here, as long as it can output image ranking results based on the user's image preference features and the quality scores of each image. In practical applications, models such as Convolutional Neural Networks (CNN), LeNet, AlexNet, VGG, GoogLeNet, and ResNet are all applicable, and this embodiment of the invention does not impose any limitations on them.

[0056] In some embodiments, after S103 above, the image recommendation method further includes the following steps: Obtaining user feedback data on image recommendations, the feedback data including the user's click-through rate for each image in the image recommendations and the user's order completion data, the order completion data reflecting the exposure and click-through rates of multiple images included in the order's corresponding object; updating the object's exposure, the exposure of multiple images, and the click-through rates of multiple images based on the image recommendation feedback data, obtaining the updated exposure, exposure, and click-through rates of the object, multiple images, and multiple images; determining the updated quality score of each image based on the updated exposure, exposure, and click-through rates of the object, multiple images, and multiple images; using the updated quality scores of each image for the next image sorting, and determining the convergence speed of the object during the multiple determinations of the quality scores of each image.

[0057] In this embodiment of the invention, the image exploration and utilization process is a cyclical process. During this process, images are recommended to the user based on their quality scores, and user feedback data on these recommendations is obtained. Based on this feedback data, the exposure of the object, the exposure of multiple images, and the click count of multiple images are statistically analyzed. The exposure and click counts used in the next image exploration and utilization process (i.e., the process of determining image quality scores) are then updated. Based on the updated exposure of the object, the updated exposure of multiple images, and the updated click count of multiple images, the quality score of each image is calculated, and then images are recommended to the user again. This further improves the accuracy of image recommendations and prepares for the integration of new test images, solving the cold start problem for new test images.

[0058] In this embodiment of the invention, the click-through rate of each image in the image recommendations for a certain object is considered. The click-through rate is composed of factors including, but not limited to, clicks, downloads, views, favorites, likes, and zooms. The feedback data from the image recommendations can serve as the user's historical behavior data, and features can be extracted from this data to update the user's image preference characteristics. Before completing an order, a user will at least browse some of the images introducing the object. The user's order completion data, to a certain extent, reflects the exposure and click-through rate of the multiple images included in the order's corresponding object.

[0059] In this embodiment of the invention, the UCB mechanism based on the E&E algorithm no longer treats the expected return of an image as a fixed number, but rather as a distribution. The UCB mechanism means that when recommending images, it does not recommend based on the point with the maximum expected return, but rather selects based on the upper bound of the expected return of the distribution. For each image, the expected return point that it is most likely to reach is selected for recommendation. Over time, as feedback information increases, the distribution curve becomes narrower and narrower, eventually converging to a fixed value. Assuming that an image's expected return is not high, meaning its performance may not be optimal, this situation will not persist indefinitely under the UCB mechanism. After several explorations, its distribution curve will quickly converge. When other images are found to have better display effects, this image will no longer be included. However, this method will not miss images with truly good display effects, because images with good display effects, without feedback information, have a very wide distribution and a large quality score, so there is always a chance to select them. After selection, the distribution will quickly converge to the actual, definite expected return. Therefore, after exploring and utilizing the image, the UCB mechanism can also obtain the convergence speed of the image, and determine the convergence speed of the object based on the convergence speed of multiple images of the object.

[0060] In some embodiments, the image includes a new test image, and the number of exposed objects is multiple, during the exploration and utilization of the image (i.e., in...). Figure 1 Prior to step S101, the image recommendation method further includes the following steps: Based on the exposure of multiple objects, select multiple first objects from among the multiple objects, where the exposure of the first objects is greater than a preset exposure; based on the convergence speed of the multiple first objects obtained after determining the quality scores of each image last time, and the preset number of objects connected to new test images within a preset time period, determine multiple second objects from among the multiple first objects; expose the multiple new test images connected to each second object, and acquire the behavioral data of the multiple new test images; the number of multiple new test images connected to each second object is a preset number, and the behavioral data corresponding to the multiple images includes the behavioral data of the multiple new test images.

[0061] In this embodiment of the invention, by setting a preset number of objects that can access new test images within a preset time period, the number of second objects accessing new test images is controlled. By setting the number of new test images that each second object can access to a preset number, the number of new test images explored is controlled. The new test images accessed by each second object are exposed, thereby obtaining the exposure volume and click volume of the new test images accessed by each second object, i.e., the behavioral data of multiple new test images.

[0062] In this embodiment of the invention, the head object (i.e., multiple first objects) is selected according to the exposure of the object. For example, according to the sku_pv in the above formula (1). i Select the head object, whose exposure is greater than the preset exposure. Because the head object has a large exposure, only feed new test images to the head object to ensure the stability of feeding new test images.

[0063] In this embodiment of the invention, taking a preset time period of one day as an example, based on the convergence speed of the head objects obtained during the previous exploration and utilization process, and the preset number of objects that can access new test images each day, objects that can access new test images are selected from the head objects, thereby controlling the number of objects that can explore new test images based on the recommendation algorithm.

[0064] For example, the number of objects that can access new test images each day is pre-set to not exceed N, and the convergence speed is greater than the preset speed. The objects that can access new test images each day are selected from the header objects. The number of objects that can access new test images is controlled to ensure the stability of the test, thereby improving the recommendation speed of new test images.

[0065] In this embodiment of the invention, not only is the number of objects that can access new test images within a preset time period pre-set, but the number of new test images that can be accessed is also pre-set, thereby controlling the number of new test images explored based on the recommendation algorithm.

[0066] For example, each object can access a maximum of M new test images per day (e.g., hundreds of thousands). By controlling the exposure traffic of new test images, the flow of new test images is limited, thereby ensuring the stability of the test and improving the recommendation speed of new test images.

[0067] After identifying the objects that can be connected to new test images, and the new test images connected to those objects, the new test images, along with the images retained from the previous image exploration and utilization process (these images are those with good recommendation performance), are exposed. The exposure and click counts of the new test images are statistically analyzed using user behavior data such as image clicks and order placements. Similarly, the exposure and click counts of the retained images are also statistically analyzed. Combining these statistical results, the behavioral data for multiple images in S101 is obtained, namely, the exposure and click counts of multiple images. Specifically, the exposure count of multiple images includes the exposure count of the new test images and the exposure count of the retained images, and the click count of multiple images includes the click count of the new test images and the click count of the retained images.

[0068] It should be noted that the preset exposure, preset time period, preset number of objects to be added to the new test image, and preset number of new test images mentioned above can all be appropriately set by those skilled in the art according to actual needs. Alternatively, the thresholds for the large number of exposures, time periods, number of objects, and number of images can be determined through analysis during the exploration of a large number of new test images, as long as the stability of the test can be guaranteed. This embodiment of the invention does not impose any limitations on this.

[0069] In some embodiments, during the exploration and utilization of images (i.e., in Figure 1 Following step S102), the image recommendation method further includes the following steps: When the number of clicks on an object is greater than a first threshold, images whose click rates are less than or equal to a preset click rate and whose exposures are greater than a second threshold are automatically removed; wherein, the number of clicks on an object is the sum of the number of clicks on multiple images, and the click rate of each image is the ratio between the number of clicks on each image and the exposure of each image.

[0070] In this embodiment of the invention, the cumulative clicks of an object and the cumulative exposure of each image introducing the object over a recent period are statistically analyzed. If the cumulative clicks of an object are greater than a first threshold, it indicates that the multiple images introducing the object have sufficiently received user feedback. If the cumulative exposure of an image is greater than a second threshold, it indicates that the image has been sufficiently exposed to the user. When the click-through rate (CTR) of an image is less than or equal to a preset CTR, it indicates that the recommendation effect of that image is poor, and these images are automatically removed, retaining images with a CTR greater than the preset CTR. This increases the exposure traffic of images with a CTR greater than the preset CTR in subsequent exploration and utilization processes.

[0071] In this embodiment of the invention, a low quality score for an image does not necessarily indicate poor recommendation performance. It could be that the object associated with the image has not received sufficient user feedback, or that the image has not been adequately exposed to the user. Therefore, a rejection condition is set (the object's click count is greater than a first threshold, and the image's exposure is greater than a second threshold). If an image meets the rejection condition and its click-through rate is less than or equal to a preset click-through rate, the image is automatically rejected. Based on user feedback on image recommendations, images with better recommendation performance are retained, while images with poorer performance are rejected. This reduces competition for exposure traffic with other images (retained images and new test images), achieving a reasonable allocation of test traffic and improving its utilization rate.

[0072] It should be noted that the aforementioned preset click-through rate can be appropriately set by those skilled in the art according to actual needs. Alternatively, it can be determined through analysis of a large number of click-through rate thresholds used during the exploration of a large number of new test images, as long as the full utilization of test traffic can be ensured. This embodiment of the invention does not impose any limitations on this.

[0073] In related technologies, taking items as an example, after a new test image is added to the traffic field, it faces a cold start problem due to the lack of historical user behavior data. To test the new test image, manual operation and forced exposure are often used. Under the stimulation of more traffic, the new test image may become an image with better recommendation performance, and some are replaced and some new test images are added every day. The above manual operation method not only leads to increased costs, but also has a certain impact on user experience. The lack of a traffic operation mechanism for new test images reduces the recommendation speed of new test images.

[0074] In this embodiment of the invention, the E&E algorithm is used to automatically explore traffic for new test images during the cold start phase. The exploration prioritizes images with lower cumulative exposure among all images of the same object; in other words, the recommendation system performs sufficient exploration and exploitation on each image. Once all images have sufficient exposure, images with higher click-through rates are preferentially displayed, thereby identifying truly effective images (i.e., proven images) and gradually eliminating less desirable images, thus correcting the limitations of manual selection by operations personnel. Simultaneously, traffic control is applied to both newly introduced test images and previously tested images to ensure the stability of the testing process and full utilization of test traffic, thereby improving the recommendation speed of new test images.

[0075] In some embodiments, prior to S101 described above, the image recommendation algorithm further includes a process of determining multiple images of the introductory object, such as... Figure 2 As shown, Figure 2 The following is an optional flowchart of another image recommendation method provided in an embodiment of the present invention, wherein the image recommendation algorithm includes steps S201-S205.

[0076] S201. Obtain the initial image.

[0077] In this embodiment of the invention, the initial image is the original image that has not been processed. Taking an object as an example, the initial image can be understood as the main body of the object obtained by the acquisition device from the object, without any other related descriptions about the object. Related descriptions may include text, image layout, etc.

[0078] S202. Generate a first candidate image based on the preset image display template and the initial image. The image display template represents a static general template used to create the candidate image.

[0079] In this embodiment of the invention, images, as one of the important attributes of object recommendation, are often manually generated and uploaded by merchants and displayed to users with the help of object recommendation. By generating images manually, the efficiency of object image production is reduced.

[0080] In this embodiment of the invention, the preset image display template can be a static general template designed by a designer. The static general template reserves areas for the main object to be displayed, text, etc., which can be programmatically filled during the image generation stage, improving the efficiency of object image production. The generated first candidate image may include one or more images, but this embodiment of the invention does not limit this.

[0081] S203. Obtain the object's main image and multiple candidate main images from the object's description information.

[0082] S204. Based on the similarity between the object's main image and multiple candidate main images, determine the second candidate image from the multiple candidate main images.

[0083] It should be noted that the execution order of S201-S202 and S203-S204 is not important; they can be executed simultaneously or sequentially. This embodiment of the invention does not impose any restrictions on this.

[0084] S205. The first candidate image, the main image of the object, and the second candidate image are treated as multiple images.

[0085] In this embodiment of the invention, the object main image can be understood as the object entry image, an image representing the object's function or overall appearance; images showing details (i.e., partial images) are not suitable as entry images. For example... Figures 3A-3D As shown, Figures 3A-3D The present invention provides exemplary schematic diagrams of four candidate main images for embodiments thereof. Figures 3A-3C The images in the image all reflect the complete picture of the object and can be used as candidate main images for that object; while Figure 3D The image in the image is a partial view of the object, that is, it shows a detailed view. Figures 3A-3C A portion of an image is unsuitable as a candidate main image for the object. The object's description information often places the main image first to allow users to quickly understand the object. When the display page or information is limited, the main image is usually displayed first, and detailed information about the object is then shown when the user clicks on it. The object's description information includes not only the main image but also multiple other images. These other images are used as candidate main images, and their image similarity is calculated to select a second candidate image. The second candidate image may include one or more images; this is not limited in this embodiment of the invention.

[0086] In this embodiment of the invention, the first and second candidate images obtained above, combined with the main image of the object, are used as multiple images to introduce a certain object, and are exposed to the user. The exposure and click volume of each image are statistically analyzed, and feedback information such as the exposure volume of the object is obtained. Based on the feedback information, the above-mentioned actions are performed. Figure 1 Image recommendation methods in [the context of the text].

[0087] In related technologies, in order to improve the effectiveness of recommended objects and ensure user experience, the object images need to be of high quality. However, high-quality images are often generated manually and require a certain level of design ability. Therefore, there are problems such as low generation efficiency and high investment costs, making it difficult to achieve large-scale operation and reducing the production efficiency of object images.

[0088] In this embodiment of the invention, by designing image templates and mining similarity images, low-cost large-scale image generation is achieved. Under the premise of ensuring a certain image quality, the object image material pool is rapidly constructed, thereby improving the production efficiency of object images.

[0089] In some embodiments, S204 is implemented as follows: Calculate the vector similarity between the embedding vector corresponding to the main image of the object and the embedding vectors corresponding to multiple candidate main images to obtain multiple vector similarities; determine the candidate range of similarity based on the numerical distribution of the multiple vector similarities; and determine the second candidate image based on the candidate main image corresponding to the vector similarity that falls within the candidate range of similarity among the multiple vector similarities.

[0090] In this embodiment of the invention, an embedding vector refers to converting (or reducing the dimensionality of) an image into a fixed-size feature representation (e.g., a vector) to facilitate processing and computation (e.g., calculating similarity, distance). This embodiment of the invention does not limit the dimension of the embedding vector. Candidate main images are obtained from the object details page (i.e., the description information). Embedding vectors for the current object's main image and the candidate main images are constructed respectively. The vector similarity between the embedding vectors of all candidate main images and the embedding vector of the object's main image is calculated. Candidate main images within the similarity candidate range are used as the image material for that object, thereby achieving the mining of similar images.

[0091] It should be noted that the vector similarity calculated above can be represented in, but is not limited to, the following ways: cosine similarity, Euclidean distance, Manhattan distance, Pearson correlation coefficient (PC), Spearman Rank Correlation (SRC), Jaccard distance, Hamming distance, etc.

[0092] For example, let's take a scenario where the embedded vector has a dimension of 3 and the vector similarity is cosine similarity. Figures 3A-3D As shown, assuming Figure 3A The image in the middle is the main image of the current object. Figure 3B Chinese images and Figure 3C The image in the middle is a candidate main image in the object details page, and the embedding vectors of the three are [0.5230, 1.4063, -1.0689], [0.2645, 2.1730, -0.5476] and [-0.8288, 0.8689, -0.6818], respectively. Calculate according to formula (2). Figure 3BChinese images and Figure 3C Chinese images and Figure 3A Cosine similarity between images.

[0093]

[0094] In this embodiment of the invention, A and B in the above formula (2) represent the embedding vectors, n represents the total number of dimensions of the embedding vectors, such as the embedding vector [0.5230, 1.4063, -1.0689] having a dimension of 3, and i represents the dimension of the embedding vector. When calculating vector similarity, This means calculating the dot product of two embedding vectors in the same dimension, and then summing the dot products in multiple dimensions. The result is obtained according to formula (2) above. Figure 3B Chinese images and Figure 3A The cosine similarity between the images is 0.909. Figure 3C Chinese images and Figure 3A The cosine similarity between the images is 0.596. If the candidate similarity range is [0.9-0.98], then... Figure 3B The image in the middle is determined as the second candidate image. If the similarity candidate range is [0.7-0.9], then... Figure 3C The image in the middle was identified as the second candidate image.

[0095] In this embodiment of the invention, by calculating the vector similarity between the embedding vectors, the candidate main image located within the similarity candidate range is selected as the second candidate image, thereby improving the accuracy of the second candidate image.

[0096] In some embodiments, if the numerical distribution of multiple vector similarities indicates that multiple vector similarities are clustered at a first similarity, then the similarity candidate range is the first similarity candidate range, and the lower limit of the first similarity candidate range is greater than the first similarity; if the numerical distribution of multiple vector similarities indicates that multiple vector similarities are clustered at a second similarity, then the similarity candidate range is the second similarity candidate range, and the upper limit of the second similarity candidate range is less than the second similarity; wherein, the first similarity is less than the second similarity, and the lower limit of the first similarity candidate range is greater than or equal to the upper limit of the second similarity candidate range.

[0097] For example, if multiple vector similarities tend to be high, meaning the numerical distribution of multiple vector similarities indicates that multiple vector similarities cluster in a high-similarity region, then the similarity candidate range can be set lower. For instance, if most of the multiple vector similarities are above 0.9, it indicates that the scenes or backgrounds corresponding to these images are highly similar. Therefore, the similarity candidate range can be set to 0.7-0.9, and candidate main images with similarities between 0.7 and 0.9 can be used as supplementary image materials to reflect the differences between the images included in the selected second candidate images.

[0098] For example, if multiple vector similarities tend to be low, meaning the numerical distribution of multiple vector similarities indicates that multiple vectors are clustered in low-similarity regions, then the similarity candidate range should be set higher. For instance, if most vector similarities are below 0.9, it indicates that the scenes or backgrounds corresponding to these images have very low similarity, meaning there is too much interfering data. Therefore, the similarity candidate range can be set to 0.9-0.98, and candidate main images with similarities between 0.9 and 0.98 can be used as supplementary image materials to filter out interfering data, thereby reflecting the similarity between the images included in the selected second candidate images.

[0099] In this embodiment of the invention, the similarity candidate range is customized according to the actual use scenario, which not only ensures the image quality, but also reflects the differences between images and improves the usability of the images included in the second candidate image.

[0100] In some embodiments, Figure 1 Following S103, this image recommendation method also includes S104-S106.

[0101] S104. Obtain user feedback data on target images in image recommendations. The feedback data on target images includes the user's click-through rate on the target image and the order completion data of the object introduced by the target image. The target image represents an image generated based on a preset image display template.

[0102] S105. Analyze the feedback data from the target object to obtain information on changes in the user's image preferences.

[0103] S106. Based on the user's image preference change information, modify the image display template to obtain an updated image display template. The updated image display template is used to generate the updated first candidate image.

[0104] In this embodiment of the invention, the feedback data of users on target images (i.e., images generated based on preset image display templates) in image recommendations includes the user's click-through rate on the target image. The feedback data also includes user information, including but not limited to user identity document (ID), user preferences, and other information.

[0105] In this embodiment of the invention, the feedback data can be obtained in real time, or it can be obtained according to the surrounding environment or according to preset conditions. This embodiment of the invention does not limit the method of obtaining the feedback data.

[0106] It should be noted that user behavior data refers to the user's responses to the above. Figure 2In S204, the initial user behavior data does not include click and order data for the main image and the second candidate image. Therefore, after exposing the selected first candidate image (i.e., the target image) to the user, it is necessary to obtain user feedback data on the target image in the image recommendation. Then, based on this feedback data, the user's image preference changes for the target image are analyzed. These changes reflect the user's inclination towards the target image, which corresponds to different image display templates. The image display templates can then be modified based on these changes. This process is repeated iteratively, continuously iterating the image display templates and creating images based on the modified templates, thereby improving the richness and display effect of images created from the preset templates.

[0107] The following will describe an exemplary application of the embodiments of the present invention in a practical application scenario.

[0108] This invention also provides an image recommendation method, such as... Figure 4 As shown, Figure 4 The following is a flowchart of optional steps in another image recommendation method provided in an embodiment of the present invention, taking an item as an example, including steps S401-S403.

[0109] S401, Image Material Access.

[0110] S402, Personalized Image Distribution.

[0111] S403, Image recommendation effect optimized.

[0112] The image material access in S401 above is preliminary data preparation, providing rich image materials for the personalized distribution of item images in S402. The image materials in S401 include two sources: automated template creation and similar image mining. The image materials include those mentioned above. Figure 2 The first candidate image, the main image of the item, and the second candidate image are detailed above. For specific creation methods and discovery techniques, please refer to the above. Figure 2 The steps described in S201-S205 will not be repeated here.

[0113] In this embodiment of the invention, the image personalized distribution process described in S402 above provides a traffic access and personalized distribution strategy for new test images. It automatically explores traffic for new test images and calculates the quality score of each image based on a recommendation algorithm. For details, please refer to the above. Figure 1The steps described in S102 are not repeated here. The user's image preference features and the quality scores of each image are input into the image personalization model to output image ranking results. For example, the image ranking results can be personalized ranking of all images under the same item according to their ranking scores, selecting the image with the highest ranking score for display, or selecting images with ranking scores above a threshold for display, thereby achieving personalized image recommendations. This embodiment of the invention does not limit the specific method of displaying images in image recommendations.

[0114] In this embodiment of the invention, the image effect optimization step in S403 provides an optimization feedback mechanism. Based on the user feedback data on image recommendations in the image recommendation results of S403 (user clicks and order information for images created based on preset image display templates), the changes in user interest characteristics for different images are analyzed, providing a reference for the automated template creation (i.e., the creation of image display templates) in image materials. This feedback loop returns to the creation of image display templates, dynamically optimizing them to improve the quality of image materials. For details, please refer to the steps described in S104-S106 above, which will not be repeated here.

[0115] In some embodiments, the present invention also provides an image recommendation method, such as... Figure 5 As shown, Figure 5 A flowchart illustrating optional steps of another image recommendation method provided in this embodiment of the invention. Figure 5 The above Figure 4 The specific details of step S402, "Personalized Image Distribution," include S501-S503.

[0116] S501. Image materials are accessed, and user feedback information on the exposed image materials is collected (i.e., user behavior data). User behavior data can be understood as historical behavior data, including but not limited to: user click behavior on images (e.g., number of clicks, dwell time), historical order completion status, etc. By extracting user interest features for different image tags, which include but are not limited to: image type (e.g., white background image, scene image), presence of models, presence of text, etc., a user profile feature for image tags is established to obtain the user's personalized image features (i.e., image preference features).

[0117] S502. Image material access includes the access of new test images. After the image materials are accessed, modeling is performed based on the E&E algorithm, and the new test images and each image are scored to obtain the quality score of each image.

[0118] S503. When performing image personalization modeling, the user's image personalization features and the quality scores of each image are used as samples, and user behavior data such as clicking on images or placing orders are used as ground truth labels. The image personalization model is trained based on the samples and ground truth labels. During training, the parameters in the image personalization model are iteratively updated until the training termination condition is met, such as reaching a preset number of training iterations, or the image ranking results output by the image personalization model being within a preset deviation range, to obtain the trained image personalization model.

[0119] This invention provides personalized recommendations for both items and their images. By stitching elements from image display templates and mining similar images, it achieves low-cost, large-scale image generation, improving image production efficiency. Through the E&E algorithm, it automatically explores traffic for new test images during the cold start phase, while simultaneously mining user interest features in images and performing personalized modeling of image materials. This enables personalized distribution of item images and improves the accuracy of image recommendations.

[0120] Based on the image recommendation method of this invention, this invention also provides an image recommendation device, such as... Figure 6 As shown, Figure 6 This is a schematic diagram of the structure of an image recommendation device provided in an embodiment of the present invention. The image recommendation device 60 includes: an acquisition module 601, used to acquire the exposure of an object, behavioral data corresponding to multiple images, and the user's image preference features. The multiple images represent the display images of the object being introduced, and the behavioral data represent the user's feedback on the exposed images.

[0121] The determination module 602 is used to determine the quality score of each image based on the exposure of the object and the behavioral data corresponding to multiple images;

[0122] The sorting module 603 is used to sort multiple images based on the user's image preference features and the quality scores of each image, obtain the image sorting results, and recommend images according to the image sorting results.

[0123] In some embodiments, the determining module 602 is further configured to determine the click-through rate and traffic index of each image based on the exposure of the object and the behavioral data corresponding to the multiple images; and to determine the quality score of each image based on the click-through rate and traffic index of each image.

[0124] In some embodiments, the behavioral data corresponding to the multiple images includes: the exposure of the multiple images and the number of clicks on the multiple images;

[0125] The determining module 602 is also used to determine the click-through rate of each image based on the ratio of the clicks of each image to the exposure of each image; and to determine the traffic index of each image based on the exposure of the object and the exposure of each image, wherein the exposure of the object is the sum of the exposures of multiple images.

[0126] In some embodiments, the image recommendation device 60 further includes an update module;

[0127] The acquisition module 601 is also used to acquire user feedback data on image recommendations. The feedback data on image recommendations includes the click-through rate of each image in the image recommendations and the user's order completion data. The order completion data reflects the exposure and click-through rate of multiple images included in the order corresponding to the object.

[0128] The update module is also used to update the exposure of the object, the exposure of multiple images, and the clicks of multiple images based on the feedback data of image recommendations, so as to obtain the updated exposure of the object, the updated exposure of multiple images, and the updated clicks of multiple images.

[0129] The determination module 602 is also used to determine the quality score of each image after the update based on the exposure of the object after the update, the exposure of multiple images after the update, and the click volume of multiple images after the update; the quality score of each image after the update is used for the next image sorting, and the convergence speed of the object is determined in the process of determining the quality score of each image multiple times.

[0130] In some embodiments, the image includes: a new test image, wherein the number of exposed objects is multiple; the image recommendation device 60 further includes an access module;

[0131] The access module is used to select multiple first objects from multiple objects based on their exposure levels, where the exposure level of the first objects is greater than a preset exposure level; based on the convergence speed of the multiple first objects obtained after determining the quality scores of each image last time, and the number of objects accessing new test images within a preset time period, multiple second objects are determined from the multiple first objects; the multiple new test images accessed by each second object are exposed, and the behavioral data of the multiple new test images is acquired; the number of multiple new test images accessed by each second object is a preset number, and the behavioral data corresponding to the multiple images includes the behavioral data of the multiple new test images.

[0132] In some embodiments, the image recommendation device 60 further includes a rejection module;

[0133] The rejection module is used to automatically reject images whose click rate is less than or equal to a preset click rate and whose exposure is greater than a second threshold when the click count of an object is greater than a first threshold. Here, the click count of an object is the sum of the click counts of multiple images, and the click rate of each image is the ratio between the click count of each image and the exposure of each image.

[0134] In some embodiments, the image recommendation device 60 further includes a candidate module;

[0135] The acquisition module 601 is also used to acquire the initial image;

[0136] The candidate module is used to generate a first candidate image based on a preset image display template and an initial image. The image display template represents a static general template used to create the candidate image.

[0137] The acquisition module 601 is also used to acquire the object's main image and multiple candidate main images from the object's description information;

[0138] The candidate module is also used to determine a second candidate image from multiple candidate main images based on the similarity between the object main image and multiple candidate main images; and to treat the first candidate image, the object main image and the second candidate image as multiple images.

[0139] In some embodiments, the candidate module is further configured to calculate the vector similarity between the embedding vector corresponding to the main image of the object and the embedding vectors corresponding to multiple candidate main images, thereby obtaining multiple vector similarities; determine the similarity candidate range based on the numerical distribution of the multiple vector similarities; and determine the second candidate image based on the candidate main image corresponding to the vector similarity that is located within the similarity candidate range among the multiple vector similarities.

[0140] In some embodiments, if the numerical distribution of multiple vector similarities indicates that multiple vector similarities are clustered at a first similarity, then the similarity candidate range is the first similarity candidate range, and the lower limit of the first similarity candidate range is greater than the first similarity; if the numerical distribution of multiple vector similarities indicates that multiple vector similarities are clustered at a second similarity, then the similarity candidate range is the second similarity candidate range, and the upper limit of the second similarity candidate range is less than the second similarity; wherein, the first similarity is less than the second similarity, and the lower limit of the first similarity candidate range is greater than or equal to the upper limit of the second similarity candidate range.

[0141] In some embodiments, the acquisition module 601 is further configured to acquire user feedback data on target images in image recommendations. The feedback data on target images includes the user's click-through rate on the target image and the user's order completion data for the object introduced by the target image. The target image represents an image generated based on a preset image display template.

[0142] The update module is also used to analyze the feedback data of the target object to obtain information on changes in the user's image preferences; based on the information on changes in the user's image preferences, the image display template is modified to obtain an updated image display template, which is used to generate the updated first candidate image.

[0143] In some embodiments, the image recommendation device 60 further includes an extraction module;

[0144] The acquisition module 601 is also used to acquire user behavior data, which includes at least one of the following: the number of times the user has clicked on historical images, the dwell time of the user when browsing historical images, and the multiple images included in the objects corresponding to historical completed orders;

[0145] The extraction module is also used to extract features from user behavior data to obtain user image preference features, which include image scene type preferences and image content composition preferences.

[0146] It should be noted that the image recommendation device provided in the above embodiments is only illustrated by the division of the above-described program modules when performing image recommendation. In practical applications, the above processing can be assigned to different program modules as needed, that is, the internal structure of the device can be divided into different program modules to complete all or part of the processing described above. Furthermore, the image recommendation device and image recommendation method embodiments provided in the above embodiments belong to the same concept, and their specific implementation process and beneficial effects are detailed in the method embodiments, which will not be repeated here. For technical details not disclosed in this device embodiment, please refer to the description of the method embodiments of the present invention for understanding.

[0147] In an embodiment of the present invention, Figure 7 This is a schematic diagram of the composition structure of the image recommendation device proposed in an embodiment of the present invention, as shown below. Figure 7 As shown, the device 70 proposed in the embodiments of the present invention may further include a processor 701 and a memory 702 storing instructions executable by the processor 701. In some embodiments, the image recommendation device 70 may further include a communication interface 703 and a bus 704 for connecting the processor 701, the memory 702 and the communication interface 703.

[0148] In this embodiment of the invention, the processor 701 can be at least one of the following: Application-Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Digital Signal Processing Device (DSPD), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), Central Processing Unit (CPU), Controller, Microcontroller, and Microprocessor. It is understood that for different devices, the electronic device used to implement the above-mentioned processor function can also be other types, and this embodiment of the invention does not specifically limit the specific types.

[0149] In this embodiment of the invention, bus 704 is used to connect communication interface 703, processor 701 and memory 702 and the mutual communication between these devices.

[0150] In this embodiment of the invention, the processor 701 is used to acquire the exposure of an object, behavioral data corresponding to multiple images, and the user's image preference features. The multiple images represent the display images of the object, and the behavioral data represents the user's feedback on the exposed images. Based on the exposure of the object and the behavioral data corresponding to the multiple images, the quality score of each image is determined. Based on the user's image preference features and the quality scores of each image, the multiple images are sorted to obtain an image sorting result, and images are recommended according to the image sorting result.

[0151] In the image recommendation device 70, the memory 702 can be connected to the processor 701. The memory 702 stores executable program code and data, including computer operation instructions. The memory 702 may contain high-speed RAM or non-volatile memory, such as at least two disk drives. In practical applications, the memory 702 can be volatile memory, such as random-access memory (RAM); or non-volatile memory, such as read-only memory (ROM), flash memory, hard disk drive (HDD), or solid-state drive (SSD); or a combination of the above types of memory, providing instructions and data to the processor 701.

[0152] Furthermore, in the embodiments of the present invention, the functional modules can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional module.

[0153] If the integrated unit is implemented as a software functional module and is not sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this embodiment, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the method of this embodiment. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0154] This invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the image recommendation method as described in any of the above embodiments.

[0155] For example, the program instructions corresponding to an image recommendation method in this embodiment can be stored on a storage medium such as an optical disc, hard disk, or USB flash drive. When the program instructions corresponding to an image recommendation method in the storage medium are read or executed by an electronic device, the image recommendation method as described in any of the above embodiments can be implemented.

[0156] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0157] This invention is described with reference to schematic and / or block diagrams illustrating the implementation of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the schematic and / or block diagrams can be implemented by computer program instructions, as well as combinations of blocks in the schematic and / or block diagrams. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the schematic and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0158] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in the implementation flow diagram. Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0159] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0160] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention.

Claims

1. An image recommendation method, characterized in that, The method includes: Get the initial image; A first candidate image is generated based on a preset image display template and the initial image, wherein the image display template represents a static general template used to create the candidate image; Retrieve the object's main image and multiple candidate main images from the object's description information; Based on the similarity between the main object image and the plurality of candidate main images, a second candidate image is determined from the plurality of candidate main images; The first candidate image, the main image of the object, and the second candidate image are treated as multiple images; The system acquires the exposure volume of the object, the behavioral data corresponding to the multiple images, and the user's image preference characteristics. The multiple images represent the display images introducing the object, and the behavioral data represents the user's feedback on the exposed images. The object is an item. The behavioral data corresponding to the multiple images includes the exposure volume of the multiple images and the number of clicks on the multiple images. The click-through rate of each image is determined based on the ratio of the number of clicks to the number of exposures of each image. Based on the exposure of the object and the exposure of each image, the traffic index of each image is determined, and the exposure of the object is the sum of the exposures of the multiple images; The quality score of each image is determined based on its click-through rate and traffic index. Based on the user's image preference features and the quality scores of each image, the multiple images are sorted to obtain an image sorting result, and image recommendations are made according to the image sorting result.

2. The method according to claim 1, characterized in that, After performing image recommendation based on the image sorting results, the method further includes: Obtain the user's feedback data on the image recommendations. The feedback data includes the user's click-through rate on each image in the image recommendations and the user's order completion data. The order completion data reflects the exposure and click-through rate of multiple images included in the order's corresponding object. Based on the feedback data of the image recommendation, the exposure of the object, the exposure of the multiple images, and the clicks of the multiple images are updated to obtain the updated exposure of the object, the updated exposure of the multiple images, and the updated clicks of the multiple images. Based on the updated exposure of the object, the updated exposure of the plurality of images, and the updated clicks of the plurality of images, the updated quality score of each image is determined; the updated quality score of each image is used for the next image sorting, and the convergence speed of the object is determined during the process of determining the quality score of each image multiple times.

3. The method according to claim 1 or 2, characterized in that, The image includes: a new test image, in which multiple objects are exposed; Before acquiring the exposure of the object, the behavioral data corresponding to multiple images, and the user's image preference features, the method further includes: Based on the exposure of multiple objects, select multiple first objects from among the multiple objects, where the exposure of the first object is greater than the preset exposure. Based on the convergence speed of the plurality of first objects obtained after determining the quality scores of each image last time, and the preset number of objects that are connected to new test images within a preset time period, a plurality of second objects are determined among the plurality of first objects. The multiple new test images connected to each second object are exposed, and the behavioral data of the multiple new test images is obtained; the number of the multiple new test images connected to each second object is a preset number, and the behavioral data corresponding to the multiple images includes the behavioral data of the multiple new test images.

4. The method according to claim 1 or 2, characterized in that, After determining the quality score of each image, the method further includes: When the number of clicks on the object is greater than the first threshold, the images whose click rate is less than or equal to the preset click rate and whose exposure is greater than the second threshold are automatically removed. The click count of the object is the sum of the click counts of the multiple images, and the click rate of each image is the ratio between the click count of each image and the exposure of each image.

5. The method according to claim 1, characterized in that, The step of determining a second candidate image from the plurality of candidate main images based on the similarity between the object main image and the plurality of candidate main images includes: Calculate the vector similarity between the embedding vector corresponding to the main image of the object and the embedding vectors corresponding to the multiple candidate main images to obtain multiple vector similarities; Based on the numerical distribution of the similarity of the multiple vectors, the similarity candidate range is determined; The second candidate image is determined based on the candidate main image corresponding to the vector similarity that falls within the candidate similarity range among the plurality of vector similarities.

6. The method according to claim 5, characterized in that, The step of determining the similarity candidate range based on the numerical distribution of the multiple vector similarities includes: If the numerical distribution of the multiple vector similarities indicates that the multiple vector similarities are clustered at the first similarity, then the similarity candidate range is the first similarity candidate range, and the lower limit of the first similarity candidate range is greater than the first similarity. If the numerical distribution of the multiple vector similarities indicates that the multiple vector similarities are clustered in the second similarity, then the similarity candidate range is the second similarity candidate range, and the upper limit of the second similarity candidate range is less than the second similarity. Wherein, the first similarity is less than the second similarity, and the lower limit of the first similarity candidate range is greater than or equal to the upper limit of the second similarity candidate range.

7. The method according to claim 1, characterized in that, After performing image recommendation based on the image sorting results, the method further includes: Obtain user feedback data on target images in the image recommendations. The feedback data for target images includes the user's click-through rate on the target image and the user's order completion data for the object introduced by the target image. The target image represents an image generated based on a preset image display template. Based on the feedback data from the target object, information on changes in the user's image preferences is obtained; Based on the user's image preference change information, the image display template is modified to obtain an updated image display template, which is used to generate an updated first candidate image.

8. The method according to claim 1 or 2, characterized in that, The method further includes: Obtain the user's behavioral data, which includes at least one of the following: the number of times the user clicked on historical images, the dwell time of the user when browsing historical images, and multiple images included in the objects corresponding to historical completed orders; Feature extraction is performed on the user's behavioral data to obtain the user's image preference features, which include image scene type preference and image content composition preference.

9. An image recommendation device, characterized in that, The device includes: The acquisition module is used to acquire the initial image; The candidate module is used to generate a first candidate image based on a preset image display template and the initial image, wherein the image display template represents a static general template used to create the candidate image; The acquisition module is also used to acquire the object's main image and multiple candidate main images from the object's description information; The candidate module is further configured to determine a second candidate image from the plurality of candidate main images based on the similarity between the object main image and the plurality of candidate main images; and to treat the first candidate image, the object main image, and the second candidate image as a plurality of images; The acquisition module is further configured to acquire the exposure of the object, the behavioral data corresponding to the multiple images, and the user's image preference features. The multiple images represent the display images introducing the object, and the behavioral data represents the user's feedback on the exposed images. The object is an item. The behavioral data corresponding to the multiple images includes the exposure of the multiple images and the number of clicks on the multiple images. The determination module is used to determine the click-through rate of each image based on the ratio of the clicks to the exposures of each image; to determine the traffic index of each image based on the exposures of the object and the exposures of each image, wherein the exposure of the object is the sum of the exposures of the multiple images; and to determine the quality score of each image based on the click-through rate and the traffic index of each image. The sorting module is used to sort the multiple images according to the user's image preference features and the quality scores of each image, obtain the image sorting result, and recommend images according to the image sorting result.

10. An image recommendation device, characterized in that, The device includes a memory and a processor, the memory storing a computer program that can run on the processor, the processor executing the program to implement the steps of the method according to any one of claims 1-8.

11. A computer-readable storage medium, characterized in that, It stores executable instructions that, when executed by a processor, implement the method described in any one of claims 1-8.

Images