An interactive information management method and system for a smart cloud platform

Abstract

The present invention relates to the field of data processing, and specifically to an interactive information management method and system for a smart cloud platform, which obtains user behavior data, context information, product information, and user static attribute data, and uses data preprocessing, feature extraction, dynamic feature weight calculation and other technologies, combined with deep neural network modeling and training, to generate personalized product recommendation results in real time. The dynamic adjustment of feature weights is not only based on changes in user behavior, but also takes into account the influence of context information, thereby improving the real-time and accuracy of the recommendation system. The present invention realizes personalized recommendations on a smart cloud platform through a dynamic feature weight adjustment method based on adaptive context perception combined with a deep neural network.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and in particular to an interactive information management method and system for a smart cloud platform. Background Art

[0002] With the rapid development of information technology, smart cloud platforms play an increasingly important role in all walks of life, especially in the field of e-commerce. By integrating advanced technologies such as big data, artificial intelligence, and cloud computing, smart cloud platforms can efficiently process and analyze massive amounts of user data and provide personalized services and recommendations. As a core component of the smart cloud platform, interactive information covers various interactive behaviors between users and the platform, such as browsing products, clicking links, searching keywords, adding products to shopping carts, and making purchases. However, existing interactive information management methods face significant technical challenges in the personalized recommendation system of smart cloud platforms. Traditional methods usually use fixed feature weights or feature fusion strategies based on static rules, which cannot adapt to the dynamic changes of user behaviors and interests in real time.

[0003] First, fixed feature weights cannot reflect the changes in user needs at different times and in different situations, resulting in a mismatch between recommended content and user current interests, reducing the accuracy and real-time performance of the recommendation system. Second, rule-based feature fusion methods lack flexibility and have difficulty handling complex multi-dimensional interactive relationships, limiting the recommendation system's deep understanding and response to user needs. In addition, with the increase in user base and interactive behavior, existing methods face huge challenges in computational efficiency and resource utilization, making it difficult to achieve efficient and accurate real-time recommendations. Summary of the invention

[0004] In response to the above-mentioned problem that efficient and accurate real-time recommendations cannot be achieved, in the first aspect, the present invention proposes an interactive information management method for a smart cloud platform, including: obtaining user behavior data and user static attribute information in the smart e-commerce cloud platform and extracting user feature vectors; obtaining product information data and extracting product feature vectors; obtaining context information and extracting context feature vectors; obtaining a weight matrix through back propagation and setting a bias vector; calculating the product of the weight matrix and the context feature vector, and the sum of the product and the bias vector constitutes a feature weight vector; splitting the feature weight vector into a first vector and a second vector; multiplying the first vector and the user feature vector bit by bit to form a weighted user feature vector; multiplying the second vector and the product feature vector bit by bit to form a weighted product feature vector; the weighted user feature vector, the weighted product feature vector and the context feature vector together constitute a fused feature vector, and the fused feature vector is used as the input of the DNN model to complete model training; obtaining real-time data of users in the smart e-commerce cloud platform, and inputting the real-time data into the trained DNN model to obtain the recommendation score of the product; and displaying the products on the platform page in descending order according to the recommendation score.

[0005] The present invention combines user behavior data, context information and product information to generate a comprehensive fusion feature vector, and achieves precision recommendation results by dynamically adjusting weights. Compared with the existing technology, this method not only integrates the multi-dimensional information of users, products and contexts, but also enables the recommendation system to adapt to changes in user interests in real time through context-aware dynamic feature weight allocation, while enhancing the system's ability to handle complex scenarios. This technology effectively overcomes the lack of personalization and response delay problems caused by static weight allocation in traditional recommendation algorithms, thereby significantly improving user experience and the accuracy of system recommendations.

[0006] Furthermore, the feature weight vector also includes a dynamic adjustment method, specifically:

[0007] ;

[0008] ;

[0009] where ω t represents the feature weight vector at time point t; a represents the weight decay coefficient; ω t-1 represents the feature weight vector at time point t-1; f(b t ) indicates that based on the current user behavior b t The feature weight adjustment function of m represents the set weight factor corresponding to the user behavior m; e m represents the set feature influence vector corresponding to user behavior m; M represents the total number of user behaviors.

[0010] The dynamic adjustment method combines context features and the user's current behavior features to design a weight decay coefficient and behavior priority adjustment mechanism to achieve flexible and dynamic weight allocation. Compared with the traditional fixed weight allocation method, this method enables the recommendation system to quickly capture real-time changes in user interests and can still provide highly relevant recommendation results in scenarios where user behavior changes. This design significantly improves the flexibility and personalization of recommendations while ensuring a balance between computational efficiency and real-time responsiveness.

[0011] Furthermore, the method for obtaining the feature weight vector is specifically as follows:

[0012] ;

[0013] ;

[0014] ;

[0015] Where ω represents the feature weight vector; Sigmoid () represents the activation function; Represents the context feature vector F c The set of all elements in the feature weight space; W c represents the weight matrix; F c represents the context feature vector; γ c represents the bias vector; Represents a collection of elements The nth element in ; exp() represents the natural exponential function.

[0016] By using the Sigmoid activation function to generate feature weight vectors, the weight values ​​are limited to a specific range, with smooth distribution and strong robustness, thus avoiding the problem of model instability caused by excessive or too small weights. In addition, the sensitivity of weight calculation to changes in contextual features has been optimized. Compared with the traditional linear mapping method, this design can more effectively distinguish the degree of feature influence in different scenarios, greatly improving the accuracy and stability of weight calculation.

[0017] Furthermore, the method for splitting the feature weight vector is specifically as follows:

[0018] ω user =Sliceω[1,d u ];

[0019] ω item =Sliceω[d u +1,dω];

[0020] where ω user represents the first vector; ω itemrepresents the second vector; Sliceω[ ] represents the slicing function of the feature weight vector ω; 1 represents that the starting point of the slice is the first element of the feature weight vector ω; d u Indicates that the slice endpoint is the dimension of the set first vector; dω indicates that the slice endpoint is the dimension of the feature weight vector ω, and the dimension of the feature weight vector is the sum of the first vector dimension and the second vector dimension.

[0021] Furthermore, obtaining user behavior data and user static attribute information in the smart e-commerce cloud platform and extracting user feature vectors also includes: the user feature vector is denoted as F u , specifically:

[0022] F u =[age,gen,occ,pla,tot,fre,bra,pro,key,pag];

[0023] Integrate SDK at the front end of the smart e-commerce cloud platform to capture user behavior data in real time, and obtain user ID, product ID, and behavior type; the behavior data includes browsing behavior, clicking behavior, searching behavior, adding to cart behavior, purchasing behavior, and timestamps of each behavior; based on the behavior data, obtain the total purchase amount tot within the user's set time period, the product category fre with the most purchases within the set time period, the brand bra with the most purchases within the set time period, the concatenation vector pro of the product IDs purchased in the last 5 times, the concatenation vector key of the keywords searched in the last 5 times, and the page type pag with the last 5 clicks; obtain user static attribute information from the platform account management system, including age age, gender gen, occupation occ, and permanent residence pla.

[0024] By using user behavior data and static attribute information to extract user feature vectors, a method for integrating behavioral features with user static features is designed to make the construction of user portraits more detailed and comprehensive. By integrating multi-dimensional features such as age group, occupation category, and purchase behavior statistics, this method can more comprehensively characterize user characteristics compared to traditional methods of single feature extraction. At the same time, it can significantly improve the dynamic update capability of user portraits by combining dynamic adjustment with contextual environment.

[0025] Furthermore, obtaining product information data and extracting product feature vectors also includes: the product feature vector is denoted as F i , specifically:

[0026] F i =[cat,log,pri,qua,sco,num,sal,tex];

[0027] Use the CDC database change capture tool Debezium to obtain and synchronize the latest information data of products from the platform's product database in real time, including: product category cat, brand log, price pri, inventory quantity qua, evaluation score sco, number of evaluations num, and sales popularity sal.

[0028] Furthermore, obtaining context information and extracting a context feature vector also includes: the context feature vector is denoted as F c , specifically:

[0029] F c =[tim,geo,dev];

[0030] Use GeoIP technology and User-Agent analysis to obtain the user's context information, and obtain time information, geographic location and device type; the time information tim includes: month and hour; the geographic location geo includes: country and city; the device type dev includes: mobile phone, tablet, PC.

[0031] By combining GeoIP technology and User-Agent parsing methods to extract context features, and using multi-dimensional representations of time, geographic location, and device type, the recommendation system's ability to perceive the user's current environment is enhanced. Compared with traditional methods, this design can generate more accurate context features based on the diversity of scenarios, thereby ensuring that the recommendation results are highly matched with the user's actual needs, especially in mobile devices or cross-regional recommendation scenarios, showing higher applicability.

[0032] Furthermore, the fused feature vector is used as the input of the DNN model and the model training is completed, which also includes: inputting the fused feature vector into a multi-layer fully connected neural network, and realizing modeling through non-linear activation function ReLU and hidden layer stacking; using a cross entropy loss function during training; using an Adam optimizer to iteratively update model parameters to minimize the loss function; and using cross validation to evaluate the model.

[0033] Furthermore, real-time data of users in the smart e-commerce cloud platform is obtained, and the real-time data is input into the trained DNN model to obtain recommendation scores of products, which also includes: sending the user ID, related product ID and current context information of the user when browsing and searching on the platform to the back-end recommendation service through the front-end; the back-end recommendation service performs feature extraction and dynamic feature weight update according to the corresponding data received to generate a weighted user feature vector and a weighted product feature vector; the weighted user feature vector, weighted product and context feature vector are concatenated to obtain a fused feature vector; the fused feature vector is input into the deployed DNN model to generate a recommendation score for each product.

[0034] In a second aspect, the present invention provides an interactive information management system for a smart cloud platform, comprising a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the interactive information management method for a smart cloud platform of the present invention is implemented.

[0035] The technical effects of the present invention are:

[0036] The present invention significantly improves the real-time, accuracy and personalization capabilities of the recommendation system through a context-aware dynamic feature weight adjustment mechanism and deep fusion of multi-source data. Compared with the prior art, the present invention has the following advantages: First, the dynamic weight allocation enables the system to adapt to instant changes in user interests; second, the combination of multi-dimensional feature fusion and deep learning models enhances the recommendation system's ability to handle complex scenarios; third, real-time data synchronization and front-end and back-end collaborative optimization ensure the timeliness and high relevance of recommendation results. The present invention provides an efficient, flexible and robust solution for the recommendation service of the smart cloud platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] By reading the following detailed description with reference to the accompanying drawings, the above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily understood. In the accompanying drawings, several embodiments of the present invention are shown in an exemplary and non-limiting manner, and the same or corresponding reference numerals represent the same or corresponding parts, wherein:

[0038] Figure 1 is a flow chart schematically showing an interactive information management method for a smart cloud platform in an embodiment of the present invention;

[0039] Figure 2 It is a block diagram schematically showing the structure of an interactive information management system for a smart cloud platform according to an embodiment of the present invention. DETAILED DESCRIPTION

[0040] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0041] The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0042] Embodiment of the interactive information management method for the smart cloud platform:

[0043] like Figure 1 As shown, the interactive information management method for the smart cloud platform of the present invention includes:

[0044] S1. Obtain user behavior data, context information, product information data, and user static attribute information; perform preprocessing operations on the acquired data and complete storage.

[0045] In the smart cloud platform, the interaction between users and the platform is frequent and diverse. For example, the interactive information contained in the smart e-commerce cloud platform includes: browsing products, clicking links, searching keywords, adding products to the shopping cart, making purchases, etc. These interactive behaviors not only reflect the interests and preferences of users, but are also affected by various contextual factors such as time, geographic location, and device usage. However, traditional interactive information management methods usually use fixed feature weights or simple rules when it comes to model training and personalized recommendations, which makes it difficult to capture and respond to dynamic changes in user interests in real time, resulting in insufficient accuracy and real-time performance of the recommendation system. Therefore, in this embodiment, a dynamic feature weight adjustment method based on adaptive context perception can be introduced to improve the accuracy and real-time performance of the recommendation system in this embodiment.

[0046] S1.1. Obtain user behavior data, context information, product information data, and user static attribute information.

[0047] In this embodiment, Kafka message queues can be used to obtain user behavior data, context information, and product information data from the smart e-commerce cloud platform. When collecting user behavior data, SDK can be integrated in the front end of the cloud platform (including web pages and mobile APPs) to capture user behavior operations in real time. The acquired data includes: user ID, product ID, behavior type (browse, click, search, add to shopping cart, purchase), and behavior timestamp; when collecting context information, GeoIP technology and User-Agent parsing can be used to obtain user context information, including: time information: month and hour; geographic location: country and city; device type: mobile phone, tablet, PC; when collecting product information data, CDC database change capture tool Debezium can be used to obtain and synchronize the latest information of products from the platform's product database in real time, including: product category, brand, price, inventory quantity, user evaluation, and sales popularity. User static attribute information can also be obtained from the platform account management system, including: age group, gender, occupation, and permanent residence.

[0048] S1.2. Preprocess the acquired data and complete storage.

[0049] After obtaining the above-mentioned various types of initial information, if it is directly used for feature extraction and subsequent model training, performance may be degraded. Therefore, in this embodiment, the data is preprocessed to improve the data quality and ensure the accuracy and stability of subsequent model training.

[0050] First, data cleaning can be used to filter abnormal data, handle missing values, and handle outliers. Specifically, robot traffic can be identified and excluded by filtering behavior logs with abnormally fast access speeds. For missing contextual information, such as geographic locations that cannot be resolved, the mean filling method can be used to fill them. Furthermore, the 3σ principle can be used to filter out outliers in numerical data, such as price, sales popularity, etc., and supplemented with the mean of two adjacent values.

[0051] After data cleaning, data conversion can be further implemented, including: converting all timestamps into a unified time format, such as the 24-hour digital format of year-month-day-hour-minute-second; standardizing all geographic location information into a unified coding format, such as converting it into a coding format represented by longitude and latitude; standardizing device types, such as encoding various types of equipment into digital form, for example: mobile phones are represented as 00, tablets are represented as 01, and computers are represented as 02.

[0052] After the data conversion is completed, it is also necessary to standardize the numerical data such as price and inventory, and the Z-score algorithm can be used for standardization. Finally, the preprocessed data is stored. In this embodiment, the data can be stored in a distributed database Hbase, and an efficient index can be established based on the user ID and the product ID to support fast query and subsequent feature extraction. The above data preprocessing operations belong to the known technology, and the specific implementation method will not be repeated here.

[0053] S2. Perform user feature extraction, product feature extraction, and context feature extraction; obtain weighted user feature vectors and product feature vectors and calculate dynamic feature weights; and construct a fused feature vector.

[0054] After completing data acquisition and preprocessing in step S1, multidimensional features that are helpful for personalized recommendation are subsequently extracted from the preprocessed data, and these features are dynamically weighted through a dynamic feature weight adjustment method based on adaptive context perception to improve the accuracy and real-time performance of the recommendation system.

[0055] S2.1. Perform user feature extraction, product feature extraction, and context feature extraction.

[0056] High-quality features are the key to improving the performance of recommendation systems. By extracting multi-dimensional features of users, products, and contexts from preprocessed data, we can more comprehensively describe user interests and product attributes, thereby improving the accuracy and relevance of recommendations.

[0057] First, user feature extraction F u, including user basic features, behavior statistical features, and dynamic behavior features. When extracting user basic features, the user's age group and gender can be obtained and represented by One-Hot encoding respectively; the user's occupation category can also be converted into a high-dimensional vector using an embedding layer; finally, the user's permanent location is converted into a vector representation using geoembedding. In this embodiment, the user's age group, gender, occupation category, and permanent location after feature extraction are recorded as age, gen, occ, and pla, respectively.

[0058] When extracting behavioral statistical features, the total amount of all purchases made by the user within 30 days can be calculated; then the product categories that the user browsed most frequently in the past 30 days can be obtained, and the categories with the highest frequency can be One-Hot encoded; finally, the brands that the user purchased most times in the past 30 days can be obtained and converted into vector representation through the embedding layer. In this embodiment, the total purchase amount after feature extraction, the product category with the highest frequency, and the brand with the most purchases are recorded as: tot, fre, and bra, respectively.

[0059] When extracting dynamic behavior features, the product IDs of the user's most recent 10 browsings can be recorded, and converted into vectors through item embedding (ItemEmbedding) for splicing; further record the keywords of the user's most recent 5 searches, and convert them into vectors through word embedding (WordEmbedding) for splicing; finally, record the page types of the user's most recent 5 clicks, and convert them into vectors through page type embedding (PageEmbedding) for splicing. In this embodiment, the spliced ​​product IDs, keywords, and page types after feature extraction are recorded as: pro, key, and pag, respectively. The specific pre-training process of each of the above embedding layers belongs to the well-known technology and will not be repeated here. At this point, the user feature vector F can be obtained. u , and there are:

[0060] F u =[age,gen,occ,pla,tot,fre,bra,pro,key,pag];

[0061] Then, product feature extraction F i, including basic features of the product, evaluation and popularity features, and text features respectively. When obtaining the basic features of the product, the category and brand of the product can be extracted and converted into a vector representation through the embedding layer. The price and inventory quantity of the product can also be extracted. In this embodiment, the category, brand, price, and inventory quantity after feature extraction are respectively recorded as: cat, log, pri, and qua; when obtaining evaluation and popularity features, the average user evaluation score, number of evaluations, and sales popularity of the product can be extracted. In this embodiment, the evaluation score, number of evaluations, and sales popularity after feature extraction are respectively recorded as: sco, num, and sal; when obtaining text features, the pre-trained BERT model can be used to encode the product description to extract a semantic vector representation. The input of the model is the text marked with WordPiece, and the output is a high-dimensional vector. In this embodiment, the text features after feature extraction are recorded as: tex. The pre-training process of the above-mentioned BERT model belongs to the well-known technology and will not be repeated here. At this point, the product feature vector F can be obtained. i , and there are:

[0062] F i =[cat,log,pri,qua,sco,num,sal,tex];

[0063] Finally, context feature extraction F c , respectively obtain the time information, geographic location and device type in the preprocessed context information. In this embodiment, the time information, geographic location and device type after feature extraction are recorded as: tim, geo and dev respectively. So far, the context feature vector F can be obtained c , and there are:

[0064] F c =[tim,geo,dev];

[0065] It should be noted that the letter elements of the above feature vectors do not mean that they are only one-dimensional vectors. For example, the user gender element after feature extraction is recorded as gen, but one possible representation of gender after One-Hot encoding is: male gender is encoded as [1,0]; female gender is encoded as [0,1], that is, the user gender vector may be two-dimensional. Therefore, the user feature vector F u 、Product feature vector F i And the context feature vector F c The dimension of the user feature vector F is not fixed. In this embodiment, the user feature vector F u The dimension is denoted as d u , the product feature vector F i The dimension is denoted as d i , and the context feature vector F cThe dimension is denoted as d c .

[0066] S2.2. Obtain the weighted user feature vector and product feature vector and calculate the dynamic feature weight.

[0067] Contextual features reflect the user's current environment and situational factors, and have an important impact on the user's purchasing decision. The user's interests and behavior patterns may change rapidly over time and with contextual factors. Feature weights need to be dynamically adjusted to reflect the latest user preferences and ensure the accuracy and real-time nature of the recommendation results. Therefore, a dynamic feature weight adjustment method based on adaptive context perception is introduced to dynamically calculate and adjust feature weights through real-time context information.

[0068] According to the context feature vector, a trainable weight matrix and bias vector are used to calculate the feature weight vector using the Sigmoid activation function to reflect the importance of different features in the current context. Specifically:

[0069] ;

[0070] ;

[0071] ;

[0072] Where ω represents the feature weight vector; Sigmoid ( ) represents the activation function, which is used to map all elements mapped to the feature weight space to the range of (0,1); Represents the context feature vector F c The set of all elements in the feature weight space; ω c Represents the weight matrix, which is used to transform the context feature vector F c Mapped to feature weight space; F c represents the context feature vector; γ c represents the bias vector; Represents a collection of elements The nth element in ; exp() represents the natural exponential function.

[0073] The above weight matrix W c It is automatically learned through backpropagation, and its dimension depends on the context feature vector F c The dimension d c And the dimension dω of the feature weight vector ω, an exemplary explanation: there is a context feature vector F c The dimension d c is 52, the user feature vector F u The dimension d u is 124, and the product feature vector Fi Dimension i is 64, and the dimension dω of the feature weight vector ω is set to: d u +d i =124+64=192, then the weight matrix W c The dimension is , that is At the same time, the dimension of the bias vector is the same as the dimension dω of the feature weight vector ω, and the value of the initial bias vector can be randomly obtained through Gaussian distribution.

[0074] By using the Sigmoid activation function to generate feature weight vectors, the weight values ​​are limited to a specific range, with smooth distribution and strong robustness, thus avoiding the problem of model instability caused by excessive or too small weights. In addition, the sensitivity of weight calculation to changes in contextual features has been optimized. Compared with the traditional linear mapping method, this design can more effectively distinguish the degree of feature influence in different scenarios, greatly improving the accuracy and stability of weight calculation.

[0075] After obtaining the feature weight vector, split it into a user feature weight vector and a product feature weight vector, and apply them to the user feature vector and the product feature vector respectively to realize dynamic weighted fusion. The specific method is as follows:

[0076] ω user =Sliceω[1,d u ];

[0077] ω item =Sliceω[d u +1,dω];

[0078] where ω user represents the user feature weight vector; ω item represents the product feature weight vector; Sliceω[ ] represents the slicing function of the feature weight vector ω, which is used to extract the weight of a specific part from the overall feature weight vector; 1 means that the starting point of the slice is the first element of the feature weight vector ω; d u Indicates that the slice endpoint is the dimension of the user feature vector; dω indicates that the slice endpoint is the dimension of the feature weight vector ω. Among them, the user feature weight vector can also be recorded as the first vector, and the product feature weight vector can also be recorded as the second vector.

[0079] Further, through the element-by-element multiplication operation, the feature weight vector is applied to the user and product feature vectors to achieve dynamic weighted fusion of features, highlight the most relevant features in the current context, and suppress irrelevant features. The specific operation method is as follows:

[0080] F u ,ω=F u ⊙ω user;

[0081] F i ,ω=F i ⊙ω item ;

[0082] Among them, F u, ω represents the weighted user feature vector, that is, the weighted user feature vector; F i, ω represents the weighted commodity feature vector, that is, the weighted commodity feature vector; ω user represents the user feature weight vector; ω item Represents the user feature weight vector; ⊙ represents element-by-element multiplication. User behavior and interests change over time, and feature weights need to be dynamically adjusted based on the latest user behavior to ensure that the recommendation system can quickly respond to changes in user interests. Finally, the feature weights are dynamically adjusted by combining historical weights and current behavior through weighted averaging to ensure the model's sensitivity to the user's latest behavior.

[0083] ;

[0084] ;

[0085] where ω t represents the feature weight at time point t; a represents the weight decay coefficient, which is used to control the degree of retention of historical weights. Its value range is (0,1). In this embodiment, it can be set to an empirical value of 0.5; ω t-1 represents the feature weight at time point t-1; f(b t ) indicates that based on the current behavior b t The feature weight adjustment function of m represents the weight factor of behavior type m; e m represents the feature influence vector of behavior type m; M represents the total number of behavior types.

[0086] In this embodiment, the weight factor β of each behavior type can be pre-set according to business needs and historical data analysis. m Example: Purchase behavior β purchase Due to its high value, the weight factor can be set to 0.5, and the add to cart action β add The weight factor can be set to 0.3, and the search behavior β search The weight factor can be set to 0.1, and the click behavior β click The weight factor can be set to 0.05, and the browsing behavior β view The weight factor can be set to 0.05.

[0087] According to different behavior types, corresponding feature influence vectors can also be set to reflect the specific impact of behavior types on feature weight adjustment. Specifically, since purchase behavior significantly increases the weights of all related features, the purchase behavior influence vector can be set to e purchase =[1,1,1,…,1]; At the same time, the behavior of adding to the shopping cart has a significant increase in the weights of all related features, and the influence vector of the behavior of adding to the shopping cart can be set to e add =[0.8,0.8,0.8,…,0.8]; Search behavior has a moderate improvement on feature weights, and the click behavior influence vector can be set to e search =[0.6,0.6,0.6,…,0.6]; Clicking behavior has a moderate improvement on feature weight, but is slightly lower than searching behavior. Here, the influence vector of click behavior can be set to e click =[0.4,0.4,0.4,…,0.4]; Browsing behavior has a small impact on feature weights, so the browsing behavior impact vector can be set to e view =[0.2, 0.2, 0.2, …, 0.2]. It should be noted that in this embodiment, the dimension of the feature influence vector corresponding to each of the above-mentioned behavior types is dω, so as to ensure the consistency of the format with the feature weight.

[0088] The dynamic adjustment method combines context features and the user's current behavior features to design a weight decay coefficient and behavior priority adjustment mechanism to achieve flexible and dynamic weight allocation. Compared with the traditional fixed weight allocation method, this method enables the recommendation system to quickly capture real-time changes in user interests and can still provide highly relevant recommendation results in scenarios where user behavior changes. This design significantly improves the flexibility and personalization of recommendations while ensuring a balance between computational efficiency and real-time responsiveness.

[0089] S2.3. Construct fusion feature vector.

[0090] By integrating the weighted user features, product features, and context features, a comprehensive fusion feature vector is formed, which provides a comprehensive information basis for subsequent deep neural network modeling. The specific fusion process is as follows:

[0091]

[0092] in represents the fused feature vector; F u, ω represents the weighted user feature vector; F i, ω represents the weighted product feature vector; F c Represents a comprehensive context feature vector; Concat() represents a concatenation function, which is used to concatenate multiple vectors into a longer vector.

[0093] S3. Complete the modeling and training of deep neural network DNN and make real-time recommendations; display personalized recommendation results.

[0094] The weighted user features and product features are modeled through a deep neural network (DNN), complex nonlinear interactions are captured, and recommendation scores are generated for sorting recommendation lists.

[0095] S3.1. Complete the modeling and training of deep neural network DNN and perform real-time recommendation.

[0096] First, the fusion feature vector The input is fed into a multi-layer fully connected neural network, and through the nonlinear activation function ReLU and the hidden layer stacking, the deep modeling of the features is realized to generate the recommendation score; then the loss function during training can adopt the cross entropy loss function; at the same time, the Adam optimizer can be used to iteratively update the model parameters to minimize the loss function; finally, the model can be evaluated using cross validation to ensure the generalization ability and recommendation performance of the model on the training set and the validation set. The above-mentioned deep neural network DNN belongs to the well-known technology, and the specific training process will not be described here.

[0097] When users browse or search, the system needs to generate a personalized recommendation list in real time to improve user experience and platform sales conversion rate. In this embodiment, the trained DNN model is deployed in an efficient online inference service framework (TensorFlow Serving) to ensure low-latency model inference capabilities; at the same time, load balancing and high-availability architecture need to be configured to ensure stable services under high concurrency. The final real-time recommendation process is as follows:

[0098] When a user browses, searches or performs other operations on the platform, the front-end system first sends the user ID, current context information and related product IDs to the back-end recommendation service; then the back-end recommendation service performs feature extraction and dynamic feature weight update in step S2 based on the received user behavior data and context information to generate weighted user feature vectors and product feature vectors; further concatenates the weighted user feature vector, product feature vector and context feature vector to generate a fused feature vector; then inputs the fused feature vector into the deployed DNN model to generate a recommendation score; finally, the candidate products are sorted according to the recommendation score, and several products with the highest scores are selected as the recommendation list.

[0099] S3.2. Display of personalized recommendation results.

[0100] The sorted recommendation list is transmitted to the front-end through the API interface. The front-end system receives the recommendation list returned by the back-end recommendation service, displays it in recommendation modules such as "Guess you like" and "Recommended for you", and dynamically updates the user interface to display recommended products.

[0101] When the online inference service is unavailable or the response delay is too high, the system automatically switches to the preset rule recommendation mode, such as the recommendation list based on popular products, to ensure the continuous availability of the recommendation function. The performance indicators of the recommendation service, such as response time, are monitored in real time. When the response time is too long, the alarm system is used to respond and handle service anomalies in a timely manner.

[0102] The present invention combines user behavior data, context information and product information to generate a comprehensive fusion feature vector, and achieves precision recommendation results by dynamically adjusting weights. Compared with the existing technology, this method not only integrates the multi-dimensional information of users, products and contexts, but also enables the recommendation system to adapt to changes in user interests in real time through context-aware dynamic feature weight allocation, while enhancing the system's ability to handle complex scenarios. This technology effectively overcomes the lack of personalization and response delay problems caused by static weight allocation in traditional recommendation algorithms, thereby significantly improving user experience and the accuracy of system recommendations.

[0103] Interactive information management system embodiment for smart cloud platform:

[0104] On the other hand, the present invention also provides an interactive information management system for a smart cloud platform. Figure 2 As shown, an interactive information management system for a smart cloud platform includes a processor and a memory, wherein the memory stores computer program instructions. When the computer program instructions are executed by the processor, an interactive information management method for a smart cloud platform according to the first aspect of the present invention is implemented.

[0105] The interactive information management system for the smart cloud platform also includes other components familiar to those skilled in the art, such as communication interfaces, and their settings and functions are known in the art, so they will not be described in detail here.

[0106] In the present invention, the aforementioned memory may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, apparatus or device. For example, a computer-readable storage medium may be any appropriate magnetic storage medium or magneto-optical storage medium, such as a resistive random access memory RRAM (Resistive Random Access Memory), a dynamic random access memory DRAM (Dynamic Random Access Memory), a static random access memory SRAM (Static Random-Access Memory), an enhanced dynamic random access memory EDRAM (Enhanced Dynamic Random Access Memory), a high-bandwidth memory HBM (High-Bandwidth Memory), a hybrid memory cube HMC (Hybrid Memory Cube), etc., or any other medium that can be used to store the required information and can be accessed by an application, a module, or both. Any such computer storage medium may be part of a device or accessible or connectable to a device. Any application or module described in the present invention may be implemented using computer-readable / executable instructions that may be stored or otherwise maintained by such a computer-readable medium.

[0107] In the description of this specification, "plurality" or "several" means at least two, such as two, three or more, etc., unless otherwise clearly and specifically defined.

[0108] Although this specification has shown and described a number of embodiments of the present invention, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Those skilled in the art will conceive of many modifications, changes and alternatives without departing from the ideas and spirit of the present invention. It should be understood that in the practice of the present invention, various alternatives to the embodiments of the present invention described herein may be employed.

Claims

1. A method for managing interactive information for a smart cloud platform, characterized in that: The method comprises: Obtain user behavior data and user static attribute information in the smart e-commerce cloud platform and extract user feature vectors; obtain product information data and extract product feature vectors; obtain context information and extract context feature vectors; Obtain a weight matrix and set a bias vector through back propagation; calculate the product of the weight matrix and the context feature vector, and the sum of the product and the bias vector constitutes a feature weight vector; The feature weight vector is split into a first vector and a second vector; the first vector is multiplied bit by bit with the user feature vector to form a weighted user feature vector; the second vector is multiplied bit by bit with the product feature vector to form a weighted product feature vector; the weighted user feature vector, the weighted product feature vector and the context feature vector together form a fused feature vector, and the fused feature vector is used as the input of the DNN model to complete the model training; Obtain real-time data of users in the smart e-commerce cloud platform, input the real-time data into the trained DNN model to obtain the recommended scores of the products; display the products on the platform page in descending order according to the recommended scores; The feature weight vector also includes a dynamic adjustment method, specifically: ； ； where ω t represents the feature weight vector at time point t; a represents the weight decay coefficient; ω t-1 represents the feature weight vector at time point t-1; f(b t ) indicates that based on the current user behavior b t The feature weight adjustment function of m represents the set weight factor corresponding to the user behavior m; e m represents the set feature influence vector corresponding to user behavior m; M represents the total number of user behaviors; The method for obtaining the feature weight vector is specifically as follows: ； ； ； Where ω represents the feature weight vector; Sigmoid ( ) represents the activation function; Represents the context feature vector F c The set of all elements in the feature weight space; W c represents the weight matrix; F c represents the context feature vector; γ c represents the bias vector; Represents a collection of elements The nth element in ; exp() represents the natural exponential function.

2. The interactive information management method for a smart cloud platform according to claim 1, characterized in that: The method for splitting the feature weight vector is specifically as follows: oh user =Sliceω[1,d u ]; oh item =Sliceω[d u +1,dω]; where ω user represents the first vector; ω item represents the second vector; Sliceω[ ] represents the slicing function of the feature weight vector ω; 1 represents that the slicing starting point is the first element of the feature weight vector ω; d u Indicates that the slice endpoint is the dimension of the set first vector; dω indicates that the slice endpoint is the dimension of the feature weight vector ω, and the dimension of the feature weight vector is the sum of the first vector dimension and the second vector dimension.

3. The interactive information management method for a smart cloud platform according to claim 1, characterized in that: Obtain user behavior data and user static attribute information in the smart e-commerce cloud platform and extract user feature vectors, including: The user feature vector is denoted as F u , specifically: F u =[age,gen,occ,pla,tot,fre,bra,pro,key,pag]; Integrate SDK at the front end of the smart e-commerce cloud platform to capture user behavior data in real time, obtain user ID, product ID, and behavior type; the behavior data includes browsing behavior, clicking behavior, searching behavior, adding to shopping cart behavior, purchasing behavior, and timestamps of each behavior; Based on the behavior data, the user's total purchase amount tot within a set time period, the product category fre with the most purchases within the set time period, the brand bra with the most purchases within the set time period, the concatenation vector pro of the product IDs purchased in the last five times, the concatenation vector key of the keywords searched in the last five times, and the page type pag with the last five clicks are obtained; Obtain user static attribute information from the platform account management system, including age, gender, occupation, and place of residence.

4. The interactive information management method for a smart cloud platform according to claim 1, characterized in that: Obtain product information data and extract product feature vectors, including: The product feature vector is denoted as F i , specifically: F i =[cat,log,pri,qua,sco,num,sal,tex]; Use the CDC database change capture tool Debezium to obtain and synchronize the latest information data of products from the platform's product database in real time, including: product category cat, brand log, price pri, inventory quantity qua, evaluation score sco, number of evaluations num, sales popularity sal, and text features tex.

5. The interactive information management method for a smart cloud platform according to claim 1, characterized in that: Obtain context information and extract context feature vectors, including: The context feature vector is denoted as F c , specifically: F c =[tim,geo,dev]; Use GeoIP technology and User-Agent analysis to obtain the user's context information, including time information, geographic location, and device type; The time information tim includes: month and hour; The geographical location geo includes: country and city; The device types dev include: mobile phone, tablet, and PC.

6. The interactive information management method for a smart cloud platform according to claim 1, characterized in that: The fused feature vector is used as the input of the DNN model and the model training is completed, including: The fused feature vector is input into a multi-layer fully connected neural network, and modeling is achieved through nonlinear activation function ReLU and hidden layer stacking; The cross entropy loss function is used during training; Use the Adam optimizer to iteratively update the model parameters to minimize the loss function; The models were evaluated using cross validation.

7. The interactive information management method for a smart cloud platform according to claim 6, characterized in that: Obtain real-time data of users in the smart e-commerce cloud platform, and input the real-time data into the trained DNN model to obtain the recommended rating of the product, including: The front-end sends the user ID, related product ID, and current context information when the user browses and searches on the platform to the back-end recommendation service; The backend recommendation service extracts features and updates dynamic feature weights based on the corresponding data received to generate weighted user feature vectors and weighted product feature vectors; Concatenating the weighted user feature vector, the weighted product feature vector, and the context feature vector to obtain a fused feature vector; The fused feature vector is input into the deployed DNN model to generate the recommendation score for each product.

8. An interactive information management system for a smart cloud platform, characterized in that: It includes a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, an interactive information management method for a smart cloud platform as described in any one of claims 1 to 7 is implemented.

Images