A multi-modal product push method and system

Abstract

The present invention discloses a commodity push method and system based on multimodality, which relates to the technical field of commodity push, and specifically includes the following steps: analyzing the collected multimodal feature information, respectively generating the recommendation correlation coefficient and compression loss impact index of each feature information in the multimodal feature information; constructing an important information evaluation model for the generated recommendation correlation coefficient and compression loss impact index of each feature information, respectively generating the importance coefficient of each feature information, evaluating the importance of each feature information in the multimodal feature information to the recommendation result, and dividing each feature information into high-importance feature information, medium-importance feature information and low-importance feature information according to the evaluation result; based on the classification results of each feature information, corresponding data compression measures are respectively taken for feature information of different importance. The present invention solves the problem of preferentially retaining important features in multimodal data compression and optimizes the recommendation accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of commodity push, and in particular to a commodity push method and system based on multimodality. Background Art

[0002] Product push refers to proactively pushing product information that meets the user's preferences to the user through various channels (such as application notifications, emails, or text messages) based on the user's interests, behaviors, and needs, in order to improve the user's shopping experience and the platform's sales conversion rate. Multimodal product push further integrates multiple data sources on this basis, including multimodal data such as text, images, videos, numerical values, and geographic location information, and fuses and analyzes these heterogeneous data through deep learning algorithms, so as to accurately capture the user's personalized needs and preferences and generate more relevant recommendation results. The special feature of this method is that it can comprehensively utilize the user's multi-dimensional information and the multiple characteristics of the product to provide richer push content and more accurate recommendations, significantly improving the relevance of the push and user acceptance.

[0003] Existing multimodal product push technology collects multimodal data of users (such as browsing history, purchase history, text comments, image interaction and geographic location, etc.) and multiple features of products (such as text description, pictures, videos, sales and reviews, etc.), pre-processes and extracts features of these heterogeneous data to form multimodal feature vectors of users and products. Then, through the multimodal feature matching model, the multiple feature data of users and products are integrated and analyzed, the recommendation score of each product to the user is calculated, and the recommendation results are dynamically adjusted in combination with factors such as inventory status and seasonality. Finally, personalized product information is pushed to users through multiple channels (such as in-app notifications, emails or text messages) according to priority, achieving highly relevant and high-precision product recommendations.

[0004] The prior art has the following deficiencies:

[0005] In a multimodal product push system, when a large amount of multimodal data (such as images, texts, and videos) needs to be processed simultaneously, the data compression task will increase dramatically. Since the system cannot distinguish which feature information is more important in real time during the compression process, important features will not be retained first, resulting in the loss of key information. When facing large-scale data processing, the existing technology lacks a dynamic evaluation mechanism for feature importance and cannot effectively optimize the feature retention order and strategy, which will cause the product push results generated by the recommendation system to be inaccurate, reduce the relevance of the recommendation and the user's click-through rate, and further affect the user's shopping experience and the platform's sales conversion rate. If the feature priority retention mechanism fails frequently, it will also reduce the efficiency of system resource utilization and affect the overall effect and performance of the push.

[0006] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not constitute the prior art that is already known to one of ordinary skill in the art. Summary of the invention

[0007] The purpose of the present invention is to provide a multi-modal commodity push method and system to solve the problems in the above-mentioned background technology.

[0008] In order to achieve the above object, the present invention provides the following technical solution: a multi-modal product push method, which specifically includes the following steps:

[0009] Before starting data compression, determine the data processing requirements and multimodal data processing targets in the product push system, and collect the corresponding multimodal feature information after determination;

[0010] Analyze the collected multimodal feature information, and generate the recommended correlation coefficient and compression loss impact index of each feature information in the multimodal feature information;

[0011] An important information evaluation model is constructed for the recommendation correlation coefficient and compression loss impact index of each feature information generated, and the important coefficient of each feature information is generated respectively. After the generation, the importance of each feature information in the multimodal feature information to the recommendation result is evaluated, and each feature information is divided into high-importance feature information, medium-importance feature information and low-importance feature information according to the evaluation results;

[0012] Based on the classification results of each feature information, corresponding data compression measures are taken for feature information of different importance;

[0013] Verify the retention and push effect of compressed feature information, dynamically adjust data compression measures based on verification results, and regularly update the importance parameters of feature information to optimize feature information evaluation and compression strategies;

[0014] Generate personalized product recommendations based on the optimized multimodal feature information and push them to users. At the same time, collect user interaction data on the pushed content, update the user's interest preference model and the importance parameters of feature information, and continuously optimize the feature information evaluation and data compression strategy.

[0015] Preferably, the collected multimodal feature information is analyzed to generate a recommended correlation coefficient and a compression loss influence index for each feature information in the multimodal feature information, which specifically includes the following steps:

[0016] Differentiate and process the collected multimodal feature information, and classify each feature information according to the type of feature information;

[0017] Acquire user interaction information and compression impact information of each feature information in the multimodal feature information in real time, and perform preprocessing after acquisition;

[0018] The user interaction information and compression impact information of each preprocessed feature information are analyzed to generate a recommendation correlation coefficient and a compression loss impact index of each feature information in the multimodal feature information.

[0019] Preferably, the logic for obtaining the recommended correlation coefficient of each feature information is as follows:

[0020] The user interaction information of each preprocessed feature information is extracted, including the number of views of each feature information at different times within a period of time, the difference in click rate before and after recommendation, the difference in purchase conversion rate before and after recommendation, and the entropy value of each feature information, and calibrated as and E i , Indicates the number of times the i-th feature information is viewed at time m within a period of time. It represents the difference in click rate before and after the recommendation of the i-th feature information at time m within a period of time. represents the purchase conversion rate difference before and after the recommendation of the i-th feature information at time m within a period of time, E i represents the entropy value of the i-th feature information, i = 1, 2, 3, ..., k, m = 1, 2, 3, ..., g, k and g are both positive integers;

[0021] Calculate the recommended correlation coefficient of each feature information. The specific calculation formula is as follows:

[0022]

[0023] In the formula, RRC i is the recommendation correlation coefficient of the i-th feature information.

[0024] Preferably, the logic for obtaining the compression loss impact index of each feature information is as follows:

[0025] Extract the compression impact information of each preprocessed feature information, including the initial data volume of each feature information and the information retention rate of each feature information at different times during a period of time in the historical compression task, and mark them as S i and S i represents the initial data volume of the i-th feature information, represents the information retention rate of the i-th feature information at time n in a period of time in the history compression task, i = 1, 2, 3, ..., k, n = 1, 2, 3, ..., h, k and h are both positive integers;

[0026] Extract the click-through rate difference before and after the recommendation of each feature information in the user interaction information at different times within a period of time The difference in purchase conversion rate before and after the recommendation m=1, 2, 3, ..., g, where g is a positive integer;

[0027] Calculate the compression loss impact index of each feature information. The specific calculation formula is as follows:

[0028]

[0029] Where, CLII i is the compression loss impact index of the i-th feature information.

[0030] Preferably, an important information evaluation model is constructed for the recommendation correlation coefficient and compression loss impact index of each feature information generated, and the important coefficient of each feature information is generated respectively, which specifically includes the following steps:

[0031] The recommended correlation coefficients, compression loss impact indexes and corresponding importance coefficients of several feature information generated in the past period of time are collected and calibrated as and x represents the recommended correlation coefficients, compression loss impact indexes and corresponding important coefficient numbers of several characteristic information generated in the past period of time, x=3, 4, 5, ..., d, d is a positive integer, and the data collected in the past period of time is formed into a historical data set;

[0032] The multivariate regression model is selected as the important information evaluation model, and is trained through historical data sets to determine the value of the regression coefficient according to the formula:

[0033]

[0034] Where β0, β1 and β2 are regression coefficients;

[0035] By minimizing the error between the predicted value and the actual value, the regression coefficient is optimized, and finally the values ​​of the regression coefficients β0, β1 and β2 are determined;

[0036] Use the finalized regression coefficient to construct an important information evaluation model and input the recommended correlation coefficient RRC of each feature information generated in real time i and compression loss impact index CLII i , generate the important coefficient ZY of each feature information in real time i .

[0037] Preferably, the important coefficients ZY of each feature information generated are iand the pre-set threshold interval of important coefficients [ZY min , ZY max ] to compare, evaluate the importance of each feature information in the multimodal feature information to the recommendation results according to the comparison results, and divide each feature information into high-importance feature information, medium-importance feature information and low-importance feature information according to the evaluation results. The specific comparison analysis and division are as follows:

[0038] If ZY i <ZY min , if the importance of the feature information in the multimodal feature information to the recommendation result is low, then the feature information is classified as low-importance feature information;

[0039] If ZY min ≤ZY i ≤ZY max , if the importance of the feature information in the multimodal feature information to the recommendation result is medium, then the feature information is classified as feature information of medium importance;

[0040] If ZY i >ZY max , if the importance of the feature information in the multimodal feature information to the recommendation result is high, then the feature information is classified as high-importance feature information.

[0041] Preferably, based on the classification results of each feature information, corresponding data compression measures are taken for feature information of different importance, specifically:

[0042] For low-importance feature information with a low degree of importance, the data compression measure adopted is to use a compression algorithm with a high compression ratio to perform intensive compression on the feature information;

[0043] For medium-importance feature information with a medium degree of importance, the data compression measure adopted is a compression algorithm with a medium compression ratio, and compression is performed under the premise of ensuring the integrity of the feature information;

[0044] For high-importance feature information with a high degree of importance, the data compression measure taken is to use a compression algorithm with a low compression ratio to retain the original quality of the feature information.

[0045] Preferably, a multimodal product push system includes a data demand and feature acquisition module, a feature analysis and coefficient generation module, an importance assessment and classification module, a dynamic compression strategy adjustment module, a compression verification and feedback optimization module, and a personalized recommendation and preference update module;

[0046] The data demand and feature collection module determines the data processing requirements and multimodal data processing targets in the product push system before starting data compression, and collects the corresponding multimodal feature information after determination;

[0047] The feature analysis and coefficient generation module analyzes the collected multimodal feature information and generates the recommended correlation coefficient and compression loss impact index of each feature information in the multimodal feature information;

[0048] The importance evaluation and classification module constructs an important information evaluation model for the recommendation correlation coefficient and compression loss impact index of each feature information generated, generates the importance coefficient of each feature information respectively, and analyzes it after generation to evaluate the importance of each feature information in the multimodal feature information to the recommendation result, and divides each feature information into high-importance feature information, medium-importance feature information and low-importance feature information according to the evaluation results;

[0049] The dynamic compression strategy adjustment module takes corresponding data compression measures for feature information of different importance based on the classification results of each feature information;

[0050] The compression verification and feedback optimization module verifies the retention and push effect of the compressed feature information, dynamically adjusts the data compression measures according to the verification results, and regularly updates the importance parameters of the feature information to optimize the evaluation and compression strategy of the feature information;

[0051] The personalized recommendation and preference update module generates personalized product recommendations based on the optimized multimodal feature information and pushes them to users. At the same time, it collects user interaction data on the pushed content, updates the user's interest preference model and the importance parameters of feature information, and continuously optimizes the evaluation of feature information and data compression strategies.

[0052] In the above technical solution, the technical effects and advantages provided by the present invention are:

[0053] 1. The present invention can accurately evaluate the importance of multimodal feature information to the recommendation results by introducing the recommendation correlation coefficient and the compression loss impact index. This model can not only evaluate the recommendation effect based on the user's interactive behavior, but also dynamically evaluate the information retention rate through historical compression data. Compared with the prior art, it solves the problem that the system cannot distinguish and prioritize important feature information in real time during large-scale data processing, ensures that feature information that has a greater impact on the recommendation effect is prioritized during data compression, and effectively avoids the loss of key information.

[0054] 2. The dynamic compression adjustment mechanism and regularly updated importance parameters in the present invention ensure that the system can dynamically optimize compression measures based on real-time push effects and user feedback. After each compression and push, the system verifies the data retention and push effects, and adjusts the compression strategy when necessary. This mechanism solves the problem of the lack of dynamic evaluation and feedback mechanisms in the prior art. The system can continuously optimize itself based on the recommendation effect and compression effect, improving the accuracy of the recommendation and user experience. At the same time, the dynamic allocation and optimization of resources ensures that the system's resource utilization efficiency is higher and avoids unnecessary waste of resources.

[0055] 3. The present invention can continuously optimize the recommendation strategy by collecting user interaction data and updating the user interest preference model, ensuring that the recommendation system can be adjusted as user behavior changes. The importance parameters of the feature information will also be dynamically updated with the historical data of the push task, and the system can better adapt to changes in user needs. This continuous improvement capability not only enhances the personalization and accuracy of recommendations, but also improves the efficiency of the push system, reduces invalid pushes, and increases user click-through rates and conversion rates. At the same time, the optimized compression strategy ensures that information saves storage and transmission resources while ensuring quality, further improving the overall system performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0057] Figure 1 The present invention is a flowchart of a multi-modal commodity push method and system.

[0058] Figure 2 The module diagram of a multi-modal commodity push method and system of the present invention is shown. DETAILED DESCRIPTION

[0059] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in a variety of forms and should not be construed as limited to the examples set forth herein; rather, these example embodiments are provided so that the description of the present disclosure will be more comprehensive and complete, and the concept of the example embodiments will be fully conveyed to those skilled in the art.

[0060] The present invention provides Figure 1 A multi-modal product push method is shown, which specifically includes the following steps:

[0061] Before starting data compression, determine the data processing requirements and multimodal data processing targets in the product push system, and collect the corresponding multimodal feature information after determination. The multimodal feature information includes image, text, and video data and their quantitative information.

[0062] How to achieve "determine the data processing requirements and multimodal data processing goals in the product push system before starting data compression": This can be achieved through the demand analysis module in the software system. First, collect the historical operation data of product push from multiple data sources such as system logs, historical data, and user feedback, including indicators such as system load, user access frequency, and click-through rate of pushed content. Then use data analysis algorithms to perform statistical analysis on these historical data to predict the system's push needs in different time periods. In addition, combined with specific business scenarios, such as high concurrency that may occur during large-scale promotional activities, the processing goals of multimodal data are derived. These goals can set reasonable compression rates, real-time requirements, and recommendation accuracy goals based on the system's processing capabilities. Through such analysis methods, data processing requirements and multimodal data processing goals are determined to provide guidance for subsequent data collection and processing steps.

[0063] How to achieve "and collect the corresponding multimodal feature information after determination, the multimodal feature information includes image, text and video data and its quantitative information": It can be achieved through the data acquisition module in the software system. First, formulate a collection strategy for multimodal feature information based on the determined processing requirements and goals. The collection strategy can define the type of data to be collected, the source of data, the timeliness requirements of the data, etc. For example, extract the image information of the product from the product database, extract text information from the product description, and extract video data from the video content generated by the user. At the same time, the system can use the data interface to connect with various data sources, automatically perform data capture and preprocessing operations, and generate corresponding quantitative information, such as image resolution, text word count, video duration, etc. These collection processes can be implemented through batch processing or stream processing to ensure that the data can be imported into the system in real time or regularly for subsequent feature analysis and processing.

[0064] Before starting data compression, first determine the processing requirements of the product push system and the processing goals of multimodal data. The purpose is to rationally plan the data collection and processing strategy and ensure that important feature information is retained first in large-scale data compression tasks. As mentioned in the technical issue, the system lacks a dynamic evaluation mechanism for the importance of features when facing large amounts of data processing, which makes it impossible to effectively optimize the order in which features are retained. Therefore, by predetermining the data processing requirements and goals, it is possible to specifically collect and process those data features that have a greater impact on the recommendation results, and reasonably set compression strategies and processing priorities. This method can reduce the waste of system resources and improve the efficiency of data processing, thereby better meeting the real-time and accuracy requirements of push.

[0065] Analyze the collected multimodal feature information, and generate the recommended correlation coefficient and compression loss impact index of each feature information in the multimodal feature information;

[0066] In this embodiment, the collected multimodal feature information is analyzed to generate a recommended correlation coefficient and a compression loss influence index for each feature information in the multimodal feature information, which specifically includes the following steps:

[0067] Differentiate and process the collected multimodal feature information, and classify each feature information according to the type of feature information;

[0068] The collected multimodal feature information can be distinguished and processed by the automatic classification algorithm of the software system, and each feature information can be classified according to the type of feature information (such as image, text and video). The specific implementation method can adopt the feature recognition algorithm of machine learning or the classification method based on data format: first, the system can make a preliminary classification according to the format or data structure of the file (such as image format is JPEG or PNG, text format is TXT or JSON, video format is MP4 or AVI); then, the feature extraction algorithm is used to further analyze the attributes of the data content, such as the resolution of the image, the language characteristics of the text, the frame rate and duration of the video, so as to accurately divide the feature information into different categories. These classification information can be stored in a specific data structure so that different processing methods and analysis algorithms can be applied in a targeted manner in the subsequent analysis steps. The purpose of this is to better cope with the complexity of multimodal data, so that the system can take appropriate processing measures according to different types of features, thereby improving the accuracy of analysis and the efficiency of processing, while ensuring that the processing method for each feature information can be optimized in the subsequent parameter generation process.

[0069] Acquire user interaction information and compression impact information of each feature information in the multimodal feature information in real time, and perform preprocessing after acquisition;

[0070] The data acquisition module and real-time monitoring function in the software system can be used to obtain user interaction information and compression impact information of each feature information in the multimodal feature information in real time. First, the system will connect with the user behavior data source (such as click logs, browsing records, playback statistics, etc.), and use the data interface (such as API or data stream) to continuously obtain the user's interaction data on the feature information, such as the number of image views, text reading frequency, video viewing time, etc. At the same time, it is connected to the data storage system and content management system to obtain data related to feature compression, such as the amount of data before and after compression, quality scores (such as PSNR of images, compression ratio of text, frame rate changes of videos), etc. The system will automatically schedule data acquisition tasks and trigger the data acquisition process in a timed or event-driven manner to ensure the real-time and continuity of the data.

[0071] The purpose of preprocessing is to clean and standardize the acquired data to make it more suitable for subsequent analysis and parameter calculation, and to improve the accuracy and efficiency of data processing. Preprocessing can include operations such as data denoising, format conversion, data normalization, and missing value filling. Specifically, denoising can remove outliers and noise data through filtering algorithms or statistical methods, format conversion can unify data from different sources into a standard structured format (such as unifying the time format and normalizing the text encoding format), and normalization can convert indicators with different dimensions of data into the same dimension range (such as normalizing the value to between 0 and 1). For missing value filling, interpolation, mean value filling, or machine learning algorithms can be used to infer the value of missing data. These preprocessing steps can be automatically performed by the data processing module to ensure data quality and consistency, thereby improving the accuracy of subsequent analysis.

[0072] The user interaction information and compression impact information of each preprocessed feature information are analyzed to generate a recommendation correlation coefficient and a compression loss impact index of each feature information in the multimodal feature information.

[0073] In this embodiment, the logic for obtaining the recommendation correlation coefficient of each feature information is as follows:

[0074] The user interaction information of each preprocessed feature information is extracted, including the number of views of each feature information at different times within a period of time, the difference in click rate before and after recommendation, the difference in purchase conversion rate before and after recommendation, and the entropy value of each feature information, and calibrated as and E i , Indicates the number of times the i-th feature information is viewed at time m within a period of time. It represents the difference in click rate before and after the recommendation of the i-th feature information at time m within a period of time. represents the purchase conversion rate difference before and after the recommendation of the i-th feature information at time m within a period of time, E i represents the entropy value of the i-th feature information, i = 1, 2, 3, ..., k, m = 1, 2, 3, ..., g, k and g are both positive integers;

[0075] Extracting user interaction information from each preprocessed feature information can be done through the log analysis tool or user behavior monitoring module in the software system. First, the system will monitor and record the user's interactive behavior data on the platform in real time, and use the server-side log system to capture the user's operation data when browsing web pages or apps, such as the number of views, click behavior, etc. Then, by analyzing the user behavior log, the number of views, click-through rate and other data of each feature information at different time points can be extracted. At the same time, the system can also extract the purchase conversion rate data before and after the recommendation through the transaction records in the database. For different types of feature information (such as images, texts and videos), the required user interaction information can be obtained by classifying and extracting features of the user's interactive behavior (such as image clicks, video playback, etc.). All these data are standardized and stored in the specified database for subsequent analysis.

[0076] Each quantitative data is obtained in the following ways: 1) The number of views is obtained by analyzing the user's browsing behavior log for a specific feature (such as an image or text). The system will record the timestamp and number of times the user clicks and views the feature each time; 2) The difference in click rate before and after recommendation is obtained by calculating the difference in click rate of the feature information before and after being recommended. The specific steps are to divide the total number of clicks before and after recommendation by the total number of displays, and calculate the difference; 3) The difference in purchase conversion rate before and after recommendation is calculated by analyzing the user's purchase behavior of the product before and after recommendation, obtaining the ratio of the number of purchases and the number of displays of the product before and after recommendation, and calculating the difference; 4) The entropy value is calculated by analyzing the content of the feature data. For example, the color distribution entropy of an image is measured by calculating the frequency distribution of different colors to measure diversity, the lexical entropy of a text is calculated by analyzing the frequency distribution of lexical appearances, and the scene change entropy of a video can be calculated by the frequency of scene switching. All of these data are automatically obtained through feature analysis tools or statistical modules and stored in the database for further analysis and modeling.

[0077] Calculate the recommended correlation coefficient of each feature information. The specific calculation formula is as follows:

[0078]

[0079] In the formula, RRC i is the recommendation correlation coefficient of the i-th feature information.

[0080] When calculating the recommendation relevance coefficient of each feature information, the formula uses weighted average and cube root to ensure that user interaction behavior data in different time periods can be fairly considered. First, the formula sums the feature information of different time periods in segments. The purpose is to incorporate the time dimension into the calculation and consider the user's interaction with the feature information in different time periods. For example, a feature may be more popular during a promotion period, but less viewed at other times. Therefore, by summing up the interaction data of different time periods, long-term user interests can be comprehensively reflected. Secondly, the number of views Click-through rate difference Purchase conversion rate difference Entropy value (E i ) These quantitative data are combined by means of cube roots. The reason for using cube roots is to balance the impact of these different dimensions. Since these data have different dimensions, direct multiplication may cause the impact of some data to be amplified or reduced. Through cube roots, the differences between data can be reduced, making the impact of each indicator on the recommendation effect more balanced in the final recommendation correlation coefficient. Finally, the segmented results are averaged. This is to treat the effects of each time period equally and avoid excessive influence of a certain time period, so that the results can reflect the overall recommendation relevance of feature information over a long period of time.

[0081] Recommendation correlation coefficient RRC of the i-th feature information i The size of directly reflects the importance of the feature information in the recommendation results. The larger the coefficient, the stronger the correlation between the feature information and user interaction behaviors (such as the number of views, click-through rate difference, and purchase conversion rate difference) in multiple time periods. At the same time, the content diversity of the feature (measured by entropy) is also higher, which means that the feature has a greater impact on user decisions in the recommendation. Therefore, the higher the recommendation correlation coefficient, the greater the contribution and influence of the feature information on the recommendation results, and the system should give priority to retaining or highlighting such features. A smaller coefficient indicates that the feature information is less attractive to users, has less influence, and is less important to the recommendation results. The system can appropriately reduce its priority when compressing or optimizing data.

[0082] In this embodiment, the logic for obtaining the compression loss impact index of each feature information is as follows:

[0083] Extract the compression impact information of each preprocessed feature information, including the initial data volume of each feature information and the information retention rate of each feature information at different times during a period of time in the historical compression task, and mark them as S i and S i represents the initial data volume of the i-th feature information, represents the information retention rate of the i-th feature information at time n in a period of time in the history compression task, i = 1, 2, 3, ..., k, n = 1, 2, 3, ..., h, k and h are both positive integers;

[0084] The compression impact information can be extracted through the data management module and historical record monitoring system in the software system. First, the initial data volume can be extracted through the interface with the storage system or data management platform. The system automatically reads the original file size data stored in the database, such as the total number of pixels of the image, the total number of characters of the text, the number of frames of the video, etc. These data are usually recorded when the feature data is first imported or generated, and the data is kept accurate through regular updates. Next, the information retention rate is extracted from the log records of the historical compression tasks. The system monitors the data compression tasks at different times and records the quality changes of the feature information after each compression, such as the peak signal-to-noise ratio (PSNR) of the image, the compression completeness percentage of the text, and the clarity score of the video. By analyzing the historical logs of the compression tasks, the system can extract the retention rate of the feature information by time period for subsequent analysis.

[0085] Initial data volume: For each specific type of feature information, the initial data volume includes: the total number of pixels in the image, the total number of characters in the text, and the number of frames in the video. These data are the basic attributes of the file, and the system can obtain them by directly reading the metadata or file descriptor of the file. For example, for an image, the system reads the resolution (width and height) of the image to calculate the total number of pixels. Information retention rate: This data is obtained by analyzing the change in quality of the feature information after each compression task. For images, the quality difference between the compressed image and the original image is calculated by the PSNR value; for text, the system records the compressed character preservation rate (the ratio of complete characters to original characters); for videos, the clarity score is a score based on changes in video frame rate or resolution. The system extracts these retention rate data by recording the compression task results at different times, and stores them in a historical database for subsequent analysis.

[0086] Extract the click-through rate difference before and after the recommendation of each feature information in the user interaction information at different times within a period of time The difference in purchase conversion rate before and after the recommendation m=1, 2, 3, ..., g, where g is a positive integer;

[0087] These data can be extracted through the user behavior tracking module and recommendation effect monitoring system in the software system. First, the extraction of the click rate difference can be achieved by analyzing the user's click behavior on specific feature information before and after the recommendation. When the recommendation task is executed, the system will record the user's click behavior on the recommended information in real time, and these data are stored in the log file or user behavior database. The click rate can be calculated by counting the number of times a certain feature information is displayed and the number of clicks by users within a period of time. At different times, the system will record the total number of clicks before and after the recommendation and calculate the difference. Secondly, the purchase conversion rate difference can be obtained by connecting with the e-commerce transaction system to obtain the actual purchase behavior data of users before and after the recommendation. The system calculates the purchase conversion rate by the number of goods displayed before and after the recommendation and the number of goods purchased, and saves these records through the historical database. The software will extract these data regularly or event-driven and store them in the data warehouse for subsequent analysis. The click rate difference is obtained by analyzing the user's click behavior on the feature information before and after the recommendation. The click rate before recommendation reflects the user's click behavior before the feature information is recommended. It is the ratio of the number of feature information displays to the number of user clicks; while the click rate after recommendation is the user's click behavior after the feature information is recommended. By comparing the click-through rates before and after the recommendation, the system can evaluate the impact of the recommendation function on the user's click behavior. Similarly, the purchase conversion rate difference is obtained by comparing the user's purchase behavior before and after the recommendation. The purchase conversion rate before the recommendation is the ratio of the user's purchase behavior before the recommendation to the number of impressions, and the purchase conversion rate after the recommendation reflects the purchase behavior after the recommendation. By analyzing the purchase conversion rate before and after the recommendation, the system can evaluate whether the recommendation has increased the user's willingness to buy. These two types of data help the system understand the changes in user behavior before and after the recommendation, and then optimize the recommendation effect.

[0088] Calculate the compression loss impact index of each feature information. The specific calculation formula is as follows:

[0089]

[0090] Where, CLII i is the compression loss impact index of the i-th feature information.

[0091] The calculation formula of the compression loss impact index comprehensively evaluates the compression loss of feature information and its impact on the recommendation results through multiple key steps. First, the initial data volume S in the formula i It reflects the original size of the feature information. The larger the data volume, the more obvious the loss may be during the compression process. Therefore, the initial data volume is multiplied by the average value of the information retention rate. to evaluate the retention quality after compression. By averaging the retention rates of multiple time periods in historical tasks, the performance of feature information in different compression tasks can be more comprehensively reflected. Next, the click-through rate difference and purchase conversion rate difference in the denominator measure the actual impact of compression on the recommendation results. By performing square root processing on the interactive behaviors in multiple time periods and weighted averaging, the data changes in different time periods can be balanced to avoid the impact of abnormal fluctuations at a single moment on the overall evaluation. Finally, the 1+ structure of the denominator ensures that the loss effect of compression is reasonably balanced and will not be amplified by individual factors. The overall formula accurately calculates the comprehensive impact of compression on the recommendation effect by combining the initial data volume, information retention rate, and changes in user behavior.

[0092] The compression loss impact index of the i-th feature information CLII i The size of directly reflects the importance of the information quality of the feature after compression to the recommendation result. The larger the compression loss impact index, the higher the quality of the feature information retained during the data compression process (that is, the higher the information retention rate), and the smaller the negative impact of compression on user behavior (such as click-through rate and purchase conversion rate). Therefore, this type of feature information has a higher retention value in the recommendation system, the recommendation effect is more stable, and the impact on user decision-making is also greater. On the contrary, if the compression loss impact index is low, it means that the feature information has lost more key information during the compression process, resulting in a large difference in the click-through rate and purchase conversion rate before and after the recommendation, and the loss caused by compression has a significant negative impact on the recommendation effect. Therefore, the size of the compression loss impact index can be used as a key indicator to judge the importance of a certain feature information in the recommendation system, helping the system to prioritize the feature information that has a greater impact on the recommendation effect.

[0093] An important information evaluation model is constructed for the recommendation correlation coefficient and compression loss impact index of each feature information generated, and the important coefficient of each feature information is generated respectively. After the generation, the importance of each feature information in the multimodal feature information to the recommendation result is evaluated, and each feature information is divided into high-importance feature information, medium-importance feature information and low-importance feature information according to the evaluation results;

[0094] In this embodiment, an important information evaluation model is constructed based on the recommended correlation coefficient and compression loss impact index of each feature information generated, and the important coefficient of each feature information is generated respectively, which specifically includes the following steps:

[0095] The recommended correlation coefficients, compression loss impact indexes and corresponding importance coefficients of several feature information generated in the past period of time are collected and calibrated as and x represents the recommended correlation coefficients, compression loss impact indexes and corresponding important coefficient numbers of several characteristic information generated in the past period of time, x=3, 4, 5, ..., d, d is a positive integer, and the data collected in the past period of time is formed into a historical data set;

[0096] The collection of recommendation correlation coefficients, compression loss impact indexes, and corresponding importance coefficients of each feature information generated in the past period of time can be achieved through the log management module and database query system of the software system. First, the system will record the recommendation correlation coefficient and compression loss impact index of each feature information in the log file after each recommendation task is executed. These logs will contain specific timestamps, feature information IDs, user behavior data, and model calculation results. Then, through the log management module, the system can store these data in a dedicated historical database on a regular basis (such as daily, weekly, or monthly), and archive them by time period and feature ID. For the importance coefficient, after each training of the important information evaluation model is completed, the system will calculate the information importance coefficient of each feature based on the generated model parameters, and write it to the same database. When collecting data, the system will query the database system in batches according to the time range, feature type, or specific task ID to filter out qualified data (such as all features in the past month). and ), and use data analysis tools to aggregate these historical data for subsequent modeling and evaluation. In this way, the system can efficiently collect and store all required feature information parameters in different time periods.

[0097] The value of x is a positive integer greater than or equal to 3. The reason is that there is only one target equation, and we need to calculate three regression coefficients (β0, β1, and β2), which means that there must be at least three independent equations to form a system of equations in order to solve the specific values ​​of these three regression coefficients. By setting x≥3, we ensure that there are enough historical data points (that is, information on at least three different features) to construct these equations so that the system of equations can be solved by linear regression. In addition, using more data points (x>3) helps to improve the accuracy and stability of the regression model and avoid overfitting or excessive deviation of the model due to insufficient data. At the same time, more data points can also provide the model with more diverse feature representations, ensuring that the regression coefficients can better capture the importance of different features to the recommendation results under different conditions. This not only meets the necessary conditions for mathematically solving the regression coefficients, but also enhances the generalization ability of the model.

[0098] The multivariate regression model is selected as the important information evaluation model, and is trained through historical data sets to determine the value of the regression coefficient according to the formula:

[0099]

[0100] Where β0, β1 and β2 are regression coefficients;

[0101] The multiple regression model is a statistical analysis method used to describe the relationship between a dependent variable and multiple independent variables. In a multiple regression model, the dependent variable is the target variable we want to predict or explain, and the independent variables are the multiple factors that affect this target. By training on historical data, the model can find the linear relationship between the independent variables and the dependent variable, which can be used to predict or evaluate unknown data. In this case, the multiple regression model is selected as the important information evaluation model because the recommended correlation coefficient and compression loss impact index Together they affect the importance of feature information. Therefore, the multivariate regression model can simultaneously consider the impact of these two indicators and generate important coefficients for each feature information. The multivariate regression model can handle multiple independent variables, which is very suitable for evaluating the complex influence relationship of feature information in multimodal recommendation systems.

[0102] The regression coefficients β0, β1, and β2 represent the contribution of different factors to the feature information importance coefficient. Specifically, β0 is the intercept term, which represents the basic value of the feature information importance coefficient when the recommended correlation coefficient and the compression loss impact index are both zero. β1 is the recommended correlation coefficient. The influence coefficient on the feature information importance coefficient reflects the influence of the recommendation relevance of the feature information on its importance; β2 is the compression loss influence index The coefficients of influence on the feature information importance coefficients illustrate the impact of compression loss on the feature information importance. Through the training of the regression model, these regression coefficients can be automatically adjusted to minimize the error between the predicted value and the actual value, thereby providing an accurate assessment.

[0103] By minimizing the error between the predicted value and the actual value, the regression coefficient is optimized, and finally the values ​​of the regression coefficients β0, β1 and β2 are determined;

[0104] The purpose of optimizing the regression coefficient by minimizing the error between the predicted value and the actual value is to ensure that the multivariate regression model can accurately reflect the real impact of the recommended correlation coefficient and the compression loss impact index on the importance of feature information. This process can be achieved through optimization algorithms such as the least squares method (OLS). The core of this method is to find the optimal values ​​of the regression coefficients β0, β1 and β2 so that the sum of square errors between the predicted feature information importance coefficient and the actual value in the historical data is minimized. In the specific implementation process, the software system will first input the recommended correlation coefficient and the compression loss impact index based on the historical data, and continuously adjust the regression coefficient through iterative calculation to gradually reduce the error between the predicted value and the actual value. This process will be executed through an automated model training module, and each iteration will recalculate the error and optimize the regression coefficient until the error converges to an acceptable minimum value. The advantage of this is that the coefficients finally generated by the regression model can adapt to the historical data to the greatest extent, ensuring high prediction accuracy and evaluation capabilities under new data.

[0105] Use the finalized regression coefficient to construct an important information evaluation model and input the recommended correlation coefficient RRC of each feature information generated in real time i and compression loss impact index CLII i , generate the important coefficient ZY of each feature information in real time i .

[0106] In this embodiment, the important coefficients ZY of each feature information generated are i and the pre-set threshold interval of important coefficients [ZY min , ZY max ] to compare, evaluate the importance of each feature information in the multimodal feature information to the recommendation results according to the comparison results, and divide each feature information into high-importance feature information, medium-importance feature information and low-importance feature information according to the evaluation results. The specific comparison analysis and division are as follows:

[0107] If ZY i <ZY min , if the importance of the feature information in the multimodal feature information to the recommendation result is low, then the feature information is classified as low-importance feature information;

[0108] This situation means that the feature information contributes little to the recommendation results, users have little interaction with the information, or the information is lost during the compression process, affecting its contribution to the recommendation effect. The impact is that this type of feature information will be downgraded in the recommendation system and its display priority will be reduced, thereby saving system resources and reducing unnecessary redundant recommendations. The system can compare the ZY of each feature information. iValue and the minimum important coefficient threshold ZY min In the specific implementation, the system will regularly run the data analysis module to analyze the ZY of all feature information. w Calculate if ZY i Less than ZY min , the system automatically marks the feature information as a low-importance feature and saves it in the corresponding data set or classification table. Its display frequency or priority can be reduced in subsequent recommendations.

[0109] If ZY min ≤ZY i ≤ZY max , if the importance of the feature information in the multimodal feature information to the recommendation result is medium, then the feature information is classified as feature information of medium importance;

[0110] This situation means that the feature information has a certain contribution to the recommendation results, but it does not have a significant influence, which is reflected in the moderate user interaction and the retention of a certain quality after compression. The impact of this type of feature information on the recommendation system is to maintain the richness of the basic recommended content, but it does not need to be overly emphasized or prioritized. The system compares the ZY of the feature information. i Values ​​and presets ZY min and ZY max Two thresholds are used for division. min ≤ZY i ≤ZY max , the system will classify the feature information as a medium-importance feature. The specific operation is that the system will calculate ZY i The value is logically judged, and the feature information that meets the interval is marked as medium importance and stored in the corresponding recommendation data structure or database classification.

[0111] If ZY i >ZY max , if the importance of the feature information in the multimodal feature information to the recommendation result is high, then the feature information is classified as high-importance feature information.

[0112] This means that the feature information has a significant contribution to the recommendation results, the user interaction is frequent, and the information retention rate after compression is high, and the influence is large. This type of information has priority display rights in the recommendation system, and the system will tend to display more of these features to ensure the accuracy of the recommendation effect and the matching degree of user interests. The system compares ZY i and the maximum important coefficient threshold ZY max To divide. maxWhen the feature information is displayed, the system will classify it as a high-importance feature and store it in a queue or database table with high priority display. In specific implementation, the system's recommendation engine will prioritize displaying these feature information in the subsequent recommendation process based on the high-priority classification table to ensure the effectiveness of the recommended content and user engagement.

[0113] The pre-set threshold interval of important coefficients can be determined by statistical analysis of historical data and model training results. The specific method is to use the data analysis module to perform a large amount of regression model analysis on past feature information. The system generates recommended correlation coefficients and compression loss impact indexes based on the performance of historical features, thereby calculating the corresponding important coefficients ZY i The software system will respond to these i Perform cluster analysis or distribution analysis to see their overall distribution and determine ZY based on quantiles in the distribution (such as 25%, 75% quantiles) or mean and standard deviation min and ZY max Threshold. In the specific implementation process, the system can use machine learning optimization algorithms, such as K-means clustering or quantile-based automatic classification algorithms, to dynamically generate reasonable threshold intervals. Ultimately, the system will classify feature information into low, medium, and high importance features based on the threshold intervals to ensure that feature information at different levels can be properly processed and displayed with priority.

[0114] Based on the classification results of each feature information, corresponding data compression measures are taken for feature information of different importance;

[0115] In this embodiment, based on the classification results of each feature information, corresponding data compression measures are taken for feature information of different importance, specifically:

[0116] For low-importance feature information with a low degree of importance, the data compression measure adopted is to use a compression algorithm with a high compression ratio to strongly compress the feature information, allowing a large amount of information loss to save storage space and transmission resources;

[0117] For low-importance feature information, a high-compression ratio compression algorithm can be implemented through the compression strategy module in the system. First, the system will identify low-importance feature information based on the pre-classification results, and select high-compression algorithms such as JPEG and MP3 to process these feature information to significantly reduce the amount of data. These algorithms allow the loss of some details, but can significantly save storage space and transmission bandwidth. In specific implementation, the software will automatically call the compression module and pass the low-importance feature information to the high-compression algorithm for processing. The reason for this is that low-importance features have little effect on the recommendation results, and the system can tolerate large information loss, thereby saving resources and optimizing system performance.

[0118] For medium-importance feature information with a medium degree of importance, the data compression measure adopted is a compression algorithm with a medium compression ratio, which is compressed under the premise of ensuring the integrity of the feature information to balance storage resources and information retention;

[0119] For medium-importance feature information, the system will call an algorithm with a medium compression ratio, such as PNG compression or moderate video encoding technology. When implementing, the system first automatically matches the medium compression algorithm based on the result of feature classification to compress these feature information. The selection of the medium compression ratio algorithm ensures the basic integrity of the information, and still retains sufficient quality for recommendation after compression. During the compression process, the system will dynamically adjust the compression rate to ensure the best balance between storage and bandwidth usage. The reason for this is that medium-importance feature information still has a certain impact on the recommendation results, and its key information needs to be retained while optimizing resource consumption.

[0120] For high-importance feature information with a high degree of importance, the data compression measure taken is to use a compression algorithm with a low compression ratio to retain the original quality of the feature information and reduce information loss, so as to ensure the complete display of high-priority feature information in the recommendation system.

[0121] For high-importance feature information, the system will select a compression algorithm with a low compression ratio, such as a lossless compression algorithm (such as PNG lossless compression, lossless audio compression, etc.), to ensure that the quality of the information is retained to the maximum extent. During implementation, the software system automatically calls the corresponding lossless compression algorithm to process the high-importance feature information by identifying the high-importance feature information, ensuring that no important data is lost during the compression process. The reason for this is that high-importance feature information has a greater impact on the recommendation results, and the integrity of the information must be maintained as much as possible to ensure that the recommendation system can present high-quality content to users. This method can maximize the accuracy of the recommendation effect and user experience.

[0122] Verify the retention and push effect of compressed feature information, dynamically adjust data compression measures based on verification results, and regularly update the importance parameters of feature information to optimize feature information evaluation and compression strategies;

[0123] The system can verify the retention of compressed feature information and the push effect through the push effect monitoring module and the data integrity check module. First, after each recommendation task is executed, the software system records the user's interaction with the recommended content in real time, such as click-through rate, dwell time, purchase conversion rate, etc., and measures the push effect through user behavior data. At the same time, the system will also perform a data integrity check on the compressed feature information, and calculate the retention rate (such as the peak signal-to-noise ratio of the image, the frame number comparison of the video, and the character retention rate of the text) by comparing the data before and after compression, so as to evaluate the impact of compression on the information quality. The reason for this is that the quality of the push effect directly determines the effectiveness of the recommendation system, and the information retention affects the actual display quality of the feature data after compression. Therefore, the verification step is a key link to ensure that the pushed content has both high-quality display and reasonable use of resources.

[0124] After the verification is completed, the system can automatically adjust the compression strategy according to the push effect and retention rate through the dynamic compression adjustment module. In specific implementation, the system will compare the effect of each recommendation push task. If the click-through rate and conversion rate of the pushed high-importance feature information are low, it means that the current compression measures may affect the quality of information display. The system will reduce the compression ratio of this type of feature information and give priority to retaining more data. Similarly, if the low-importance feature information does not contribute much to the push effect, its compression ratio can be dynamically increased to save resources. The system dynamically adjusts the compression algorithm selection for various types of feature information by combining historical data and real-time feedback. The reason for this is that the performance of different feature information in recommendations may change over time and user behavior patterns. By dynamically adjusting the compression measures, the system can ensure that the best balance between push effect and resource usage is maintained.

[0125] Regularly updating the importance parameters of feature information can be achieved through the data analysis module combined with the push history data. The system will regularly analyze the performance of each feature information in the push task, and recalculate its importance parameters such as the recommendation correlation coefficient and compression loss impact index. In specific implementation, the software will aggregate the historical recommendation data of all feature information according to a specific time interval (such as daily or weekly), and re-evaluate the importance of the feature information in combination with the latest push effect and user behavior data, and automatically update its parameters. The reason for this is that user interests and preferences will change over time, and regularly updating the importance parameters of feature information can ensure that the recommendation system always reflects the latest user needs, and can more accurately allocate resources and optimize display effects in future push tasks.

[0126] By combining regular verification results and dynamically adjusted feedback, the system can continuously optimize the evaluation and compression strategies of feature information. During implementation, the software will continuously monitor and analyze the compression effect and push performance of each feature information, and gradually correct the deficiencies in the initial compression algorithm. For example, the system may find that the recommendation relevance coefficient of a certain type of feature information is high, but the details lost after compression lead to a decrease in user interaction. At this time, the system will automatically increase the retention rate of this type of information and adjust the corresponding compression strategy. The optimized evaluation model and compression strategy will be further refined through machine learning algorithms to ensure that the optimal compression method is always used in future pushes. The reason for this is that the interactive behaviors of the recommendation system and users are dynamically changing. By continuously optimizing the compression strategy, the system can improve the accuracy of push and the efficiency of resource use.

[0127] Generate personalized product recommendations based on the optimized multimodal feature information and push them to users. At the same time, collect user interaction data on the pushed content, update the user's interest preference model and the importance parameters of feature information, and continuously optimize the feature information evaluation and data compression strategy.

[0128] Personalized product recommendations can be implemented through a recommendation algorithm engine, specifically by using the user's historical behavior data, interest preferences, and optimized multimodal feature information. The system will adjust the recommendation strategy based on the importance parameters of the optimized feature information to ensure that high-importance feature information is recommended first. For example, for different types of feature information such as images, texts, and videos, the system will dynamically calculate their potential appeal to users, and then generate a set of personalized recommended product lists. In specific implementation, the recommendation algorithm will first analyze the user's historical behavior (such as browsing history, purchase history, and click behavior), and match this data with high-priority feature information to generate product recommendations that best meet user needs. When pushing to users, the system will call the communication module to display the recommended products to users through APP, web pages, or other channels. The reason for this is that personalized recommendations can increase user engagement and satisfaction, and by giving priority to displaying the most relevant products to users, the accuracy and effectiveness of the recommendation system are improved.

[0129] The collection of user interaction data can be achieved through the user behavior monitoring module of the system, including clicks, dwell time, favorites, purchases and other interactive behaviors. After each product push, the system will track the user's various behaviors on the platform in real time, and record the user's click rate, browsing time, and product purchase conversion rate through the background log system of the web page or APP. The system can timestamp each user's specific interactive behavior and record the corresponding feature information ID for subsequent analysis of changes in user interest in specific feature information. The reason for collecting this data is that by analyzing the user's actual interaction data, the understanding of user interests and preferences can be dynamically updated, thereby providing a more accurate basis for subsequent recommendations and data compression strategy optimization.

[0130] The update of the interest preference model can be dynamically adjusted through machine learning algorithms. After collecting new user interaction data, the system will input this data into the user interest preference model and retrain the model based on the user's latest behavior. For example, a user may click on a certain type of product or browse a specific type of video more frequently over a period of time, which indicates that the user's interest has changed, and the system will increase the weight of this type of feature information in the model. During implementation, the system will periodically call the training module, combine new user behavior data with historical behavior, and use collaborative filtering, content filtering, or hybrid recommendation techniques to update the user's preference model. The reason for this is that user interests and preferences will change over time, seasons, and market trends. Continuously updating the interest model can ensure that the recommendation system always follows the changes in user needs and provides the most relevant product recommendations.

[0131] The importance parameter of feature information can be updated by combining user behavior feedback and historical push effects. After collecting the user's interaction data, the system will recalculate the recommendation relevance coefficient and compression loss impact index of the feature information, and dynamically adjust the importance parameters of the feature based on these updated data. For example, if a certain type of feature has generated a high click-through rate and conversion rate for users in past recommendation pushes, the system will increase its importance parameter accordingly to ensure that such information is retained first in future pushes. When optimizing the data compression strategy, the system will comprehensively consider the importance parameters of the feature information and dynamically adjust the compression ratio to ensure that a more conservative compression method is used for high-importance features, while a more efficient compression method is used for low-importance features. The reason for this is that the recommendation system needs to maintain a balance between push quality and resource utilization. Updating the importance parameters and compression strategies of feature information can ensure that the system always maintains the best state in different recommendation tasks.

[0132] like Figure 2A multimodal product push system is shown, comprising a data demand and feature acquisition module, a feature analysis and coefficient generation module, an importance evaluation and classification module, a dynamic compression strategy adjustment module, a compression verification and feedback optimization module, and a personalized recommendation and preference update module;

[0133] The data demand and feature collection module determines the data processing requirements and multimodal data processing targets in the product push system before starting data compression, and collects the corresponding multimodal feature information after determination;

[0134] The feature analysis and coefficient generation module analyzes the collected multimodal feature information and generates the recommended correlation coefficient and compression loss impact index of each feature information in the multimodal feature information;

[0135] The importance evaluation and classification module constructs an important information evaluation model for the recommendation correlation coefficient and compression loss impact index of each feature information generated, generates the importance coefficient of each feature information respectively, and analyzes it after generation to evaluate the importance of each feature information in the multimodal feature information to the recommendation result, and divides each feature information into high-importance feature information, medium-importance feature information and low-importance feature information according to the evaluation results;

[0136] The dynamic compression strategy adjustment module takes corresponding data compression measures for feature information of different importance based on the classification results of each feature information;

[0137] The compression verification and feedback optimization module verifies the retention and push effect of the compressed feature information, dynamically adjusts the data compression measures according to the verification results, and regularly updates the importance parameters of the feature information to optimize the evaluation and compression strategy of the feature information;

[0138] The personalized recommendation and preference update module generates personalized product recommendations based on the optimized multimodal feature information and pushes them to users. At the same time, it collects user interaction data on the pushed content, updates the user's interest preference model and the importance parameters of feature information, and continuously optimizes the evaluation of feature information and data compression strategies.

[0139] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0140] The above embodiments can be implemented in whole or in part by software, hardware, firmware or any other combination. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website site, computer, server or data center to another website site, computer, server or data center by wired or wireless (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center containing one or more available media sets. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state hard disk.

[0141] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0142] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0143] In the several embodiments provided in the present application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0144] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0145] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0146] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A multi-modal product push method, characterized in that: The specific steps include: Before starting data compression, determine the data processing requirements and multimodal data processing targets in the product push system, and collect the corresponding multimodal feature information after determination; Analyze the collected multimodal feature information, and generate the recommended correlation coefficient and compression loss impact index of each feature information in the multimodal feature information; An important information evaluation model is constructed for the recommendation correlation coefficient and compression loss impact index of each feature information generated, and the important coefficient of each feature information is generated respectively. After the generation, the importance of each feature information in the multimodal feature information to the recommendation result is evaluated, and each feature information is divided into high-importance feature information, medium-importance feature information and low-importance feature information according to the evaluation results; Based on the classification results of each feature information, corresponding data compression measures are taken for feature information of different importance; Verify the retention and push effect of compressed feature information, dynamically adjust data compression measures based on verification results, and regularly update the importance parameters of feature information to optimize feature information evaluation and compression strategies; Generate personalized product recommendations based on the optimized multimodal feature information and push them to users. At the same time, collect user interaction data on the pushed content, update the user's interest preference model and the importance parameters of feature information, and continuously optimize the feature information evaluation and data compression strategy.

2. A multimodal commodity push method according to claim 1, characterized in that: The collected multimodal feature information is analyzed to generate the recommended correlation coefficient and compression loss impact index of each feature information in the multimodal feature information, which specifically includes the following steps: Differentiate and process the collected multimodal feature information, and classify each feature information according to the type of feature information; Acquire user interaction information and compression impact information of each feature information in the multimodal feature information in real time, and perform preprocessing after acquisition; The user interaction information and compression impact information of each preprocessed feature information are analyzed to generate a recommendation correlation coefficient and a compression loss impact index of each feature information in the multimodal feature information.

3. A multimodal commodity push method according to claim 2, characterized in that: The logic for obtaining the recommended correlation coefficient of each feature information is as follows: The user interaction information of each preprocessed feature information is extracted, including the number of views of each feature information at different times within a period of time, the difference in click rate before and after recommendation, the difference in purchase conversion rate before and after recommendation, and the entropy value of each feature information, and calibrated as and E i , Indicates the number of times the i-th feature information is viewed at time m within a period of time. It represents the difference in click rate before and after the recommendation of the i-th feature information at time m within a period of time. represents the purchase conversion rate difference before and after the recommendation of the i-th feature information at time m within a period of time, E i represents the entropy value of the i-th feature information, i = 1, 2, 3, ..., k, m = 1, 2, 3, ..., g, k and g are both positive integers; Calculate the recommended correlation coefficient of each feature information. The specific calculation formula is as follows: In the formula, RRC i is the recommendation correlation coefficient of the i-th feature information.

4. A multimodal commodity push method according to claim 3, characterized in that: The logic for obtaining the compression loss impact index of each feature information is as follows: Extract the compression impact information of each preprocessed feature information, including the initial data volume of each feature information and the information retention rate of each feature information at different times during a period of time in the historical compression task, and mark them as S i and S i represents the initial data volume of the i-th feature information, represents the information retention rate of the i-th feature information at time n in a period of time in the history compression task, i = 1, 2, 3, ..., k, n = 1, 2, 3, ..., h, k and h are both positive integers; Extract the click-through rate difference before and after the recommendation of each feature information in the user interaction information at different times within a period of time The difference in purchase conversion rate before and after the recommendation m=1, 2, 3, ..., g, where g is a positive integer; Calculate the compression loss impact index of each feature information. The specific calculation formula is as follows: Where, CLII i is the compression loss impact index of the i-th feature information.

5. A multi-modal commodity push method according to claim 4, characterized in that: An important information evaluation model is constructed for the recommended correlation coefficient and compression loss impact index of each feature information generated, and the important coefficient of each feature information is generated respectively, which specifically includes the following steps: The recommended correlation coefficients, compression loss impact indexes and corresponding importance coefficients of several feature information generated in the past period of time are collected and calibrated as and x represents the recommended correlation coefficients, compression loss impact indexes and corresponding important coefficient numbers of several characteristic information generated in the past period of time, x=3, 4, 5, ..., d, d is a positive integer, and the data collected in the past period of time is formed into a historical data set; The multivariate regression model is selected as the important information evaluation model, and is trained through historical data sets to determine the value of the regression coefficient according to the formula: Where β0, β1 and β2 are regression coefficients; By minimizing the error between the predicted value and the actual value, the regression coefficient is optimized, and finally the values ​​of the regression coefficients β0, β1 and β2 are determined; Use the finalized regression coefficient to construct an important information evaluation model and input the recommended correlation coefficient RRC of each feature information generated in real time i and compression loss impact index CLII i , generate the important coefficient ZY of each feature information in real time i .

6. A multi-modal commodity push method according to claim 5, characterized in that: The important coefficients ZY of each feature information generated i and the pre-set threshold interval of important coefficients [ZY min , ZY max ] to compare, evaluate the importance of each feature information in the multimodal feature information to the recommendation results according to the comparison results, and divide each feature information into high-importance feature information, medium-importance feature information and low-importance feature information according to the evaluation results. The specific comparison analysis and division are as follows: If ZY i <ZY min , if the importance of the feature information in the multimodal feature information to the recommendation result is low, then the feature information is classified as low-importance feature information; If ZY min ≤ZY i ≤ZY max , if the importance of the feature information in the multimodal feature information to the recommendation result is medium, then the feature information is classified as feature information of medium importance; If ZY i >ZY max , if the importance of the feature information in the multimodal feature information to the recommendation result is high, then the feature information is classified as high-importance feature information.

7. A multimodal commodity push method according to claim 6, characterized in that: Based on the classification results of each feature information, corresponding data compression measures are taken for feature information of different importance, specifically: For low-importance feature information with a low degree of importance, the data compression measure adopted is to use a compression algorithm with a high compression ratio to perform intensive compression on the feature information; For medium-importance feature information with a medium degree of importance, the data compression measure adopted is a compression algorithm with a medium compression ratio, and compression is performed under the premise of ensuring the integrity of the feature information; For high-importance feature information with a high degree of importance, the data compression measure taken is to use a compression algorithm with a low compression ratio to retain the original quality of the feature information.

8. A multimodal product push system, used to implement a multimodal product push method according to any one of claims 1 to 7, characterized in that: It includes data demand and feature collection module, feature analysis and coefficient generation module, importance assessment and classification module, dynamic compression strategy adjustment module, compression verification and feedback optimization module, and personalized recommendation and preference update module; The data demand and feature collection module determines the data processing requirements and multimodal data processing targets in the product push system before starting data compression, and collects the corresponding multimodal feature information after determination; The feature analysis and coefficient generation module analyzes the collected multimodal feature information and generates the recommended correlation coefficient and compression loss impact index of each feature information in the multimodal feature information; The importance evaluation and classification module constructs an important information evaluation model for the recommendation correlation coefficient and compression loss impact index of each feature information generated, generates the importance coefficient of each feature information respectively, and analyzes it after generation to evaluate the importance of each feature information in the multimodal feature information to the recommendation result, and divides each feature information into high-importance feature information, medium-importance feature information and low-importance feature information according to the evaluation results; The dynamic compression strategy adjustment module takes corresponding data compression measures for feature information of different importance based on the classification results of each feature information; The compression verification and feedback optimization module verifies the retention and push effect of the compressed feature information, dynamically adjusts the data compression measures according to the verification results, and regularly updates the importance parameters of the feature information to optimize the evaluation and compression strategy of the feature information; The personalized recommendation and preference update module generates personalized product recommendations based on the optimized multimodal feature information and pushes them to users. At the same time, it collects user interaction data on the pushed content, updates the user's interest preference model and the importance parameters of feature information, and continuously optimizes the evaluation of feature information and data compression strategies.

Images