Live broadcast method and device for free matching combined type e-commerce operation

Abstract

The invention discloses a live broadcast method and device for free matching combined e-commerce operation, and the method comprises the steps: S1, collecting and analyzing user data, and building a user portrait; s2, based on the user portrait and the real-time behavior data, generating and dynamically updating a personalized commodity recommendation list; s3, in the live broadcast process, the anchor displays a commodity combination effect and provides a real-time interaction function; s4, virtual reality and augmented reality technologies are adopted to realize virtual commodity display and trial; s5, analyzing the real-time behavior data of the user by using an improved collaborative filtering algorithm, and recommending related commodities; s6, analyzing the emotional tendency of the user in interaction by adopting a multi-modal fusion emotion analysis algorithm, and providing personalized interaction suggestions for the anchor; and S7, providing an integrated online shopping function, and supporting the user to seamlessly complete the purchase operation. Through the technical steps, the personalized shopping experience of the user is improved, the live broadcast interactivity is enhanced, and the method has the beneficial effects of high recommendation accuracy, strong participation sense of the user and the like.

Description

Technical Field

[0001] The present invention relates to the field of e-commerce live broadcast technology, and in particular to a live broadcast method and device for free combination e-commerce operations. Background Art

[0002] In the modern e-commerce industry, live streaming e-commerce, as an emerging sales and marketing tool, has gradually become an important channel to attract consumers. Through live streaming, merchants can intuitively display the functions, characteristics and usage effects of products, while interacting with consumers in real time, answering questions and enhancing users' confidence in purchasing decisions. However, these existing technologies still have some limitations and challenges:

[0003] First, the existing live e-commerce system mainly relies on the personal performance of the anchor and static product display, lacking in-depth personalized recommendations and interactive functions. When watching live broadcasts, consumers can usually only passively accept recommended content and lack the space to actively choose according to their own interests and preferences. In addition, the existing technology mostly uses static pictures or videos in the display of goods, which cannot provide a dynamic and intuitive product experience. Although some platforms have introduced virtual reality and augmented reality technologies, the application scope and effect of these technologies are still limited, and it is impossible to provide a comprehensive and realistic virtual product trial experience during the live broadcast.

[0004] Second, in terms of personalized recommendations, traditional recommendation algorithms mostly rely on historical user behavior data, such as browsing history and purchase history, which makes it difficult to reflect changes in user interests in real time. This static recommendation method often leads to inaccurate recommendation results and low user satisfaction. Existing recommendation systems are unable to cope with real-time changes in user interests and behaviors and are unable to provide dynamic personalized recommendations.

[0005] Third, in terms of interaction and sentiment analysis, the application of existing technologies is relatively preliminary, and most of them only stay at simple emotion classification, lacking in-depth understanding and analysis of users' complex emotional states. Traditional sentiment analysis technologies mostly use single-modal data for analysis, ignoring the emotional associations between multimodal data. This makes it impossible for anchors to accurately grasp users' emotional states and needs when interacting with viewers, making it difficult to provide personalized interaction suggestions.

[0006] Therefore, how to provide a live broadcast method and device for freely combining e-commerce operations is an urgent problem that technicians in this field need to solve. Summary of the invention

[0007] One object of the present invention is to propose a live broadcast method and device for free combination e-commerce operation. The present invention describes in detail the technical solution for intelligently realizing product recommendation, sentiment analysis and virtual product display by adopting a meta-path embedding recommendation algorithm, an improved collaborative filtering algorithm and an adaptive multimodal fusion sentiment analysis algorithm. Specifically, the present invention implements personalized product recommendations based on user portraits and real-time behavior data, and provides a display and trial experience of virtual products in combination with virtual reality and augmented reality technologies. At the same time, the present invention uses multimodal sentiment analysis to provide personalized interactive suggestions for anchors. Compared with traditional live broadcast e-commerce systems, the present invention can dynamically adjust recommendation lists and interactive content, and has the advantages of high recommendation accuracy, strong personalized user experience and high interactive participation, thereby significantly improving the user's shopping experience and satisfaction.

[0008] A free-to-match combined e-commerce operation live broadcast method according to an embodiment of the present invention includes the following steps:

[0009] S1. Build user portraits by collecting and analyzing user data;

[0010] S2. Based on user portraits and real-time behavior data, a meta-path embedding recommendation algorithm is used to generate a personalized product recommendation list, which is dynamically updated during the live broadcast;

[0011] S3. During the live broadcast, the anchor will display the diversified combination effects of the products in real time and provide real-time interactive functions;

[0012] S4. Use virtual reality and augmented reality technologies to integrate virtual functions into the live broadcast interface;

[0013] S5. Use an improved collaborative filtering algorithm to analyze users’ real-time behavior data and recommend related products in real time based on the purchase records of similar users.

[0014] S6. Adopting the sentiment analysis algorithm of adaptive multimodal fusion to analyze the emotional tendency of users in real-time interaction and provide personalized interaction suggestions for anchors;

[0015] S7. Provides integrated and convenient online shopping functions, allowing users to seamlessly complete purchase operations while watching live broadcasts.

[0016] Optionally, the S2 specifically includes:

[0017] S21. Through the user portrait generation process, collect the user's browsing history, purchase record, search keywords and social media interaction data, and construct the user behavior matrix U = (u i，j ), where u i，j represents the interaction weight of user i on product i;

[0018] S22, based on the user behavior matrix U and the product feature matrix P = (p j，k ), where p j，k The feature k representing the product j is extracted by using the meta-path embedding recommendation algorithm. The meta-path embedding recommendation algorithm constructs a multi-dimensional relationship path between users and products to achieve a high-dimensional embedding representation of multimodal features.

[0019] S23, after generating the product feature embedding vector φ(P), using the feature selection mechanism based on deep learning, combined with the user behavior matrix U, calculate the user feature embedding vector φ(u);

[0020] S24. Calculate the inner product of the product feature embedding vector φ(P) and the user feature embedding vector φ(U) in the high-dimensional vector space to obtain the product recommendation score r i，j , where r i，j It represents the recommendation score of user i for product j, and optimizes the recommendation result through the weighted adjustment mechanism;

[0021] S25. Based on real-time behavior data and user feedback, the recommendation model is learned and iteratively optimized online, and the recommendation list is dynamically updated.

[0022] Optionally, the S22 specifically includes:

[0023] S221, through the user behavior matrix U = (u i，j ) and product feature matrix P = (p j，k ), construct a user-item relationship graph G = (V, E), where V represents the set of users and items, and E represents the interaction relationship between users and items;

[0024] S222: In the user-product relationship graph G, define multiple meta-paths P = {P1, P2, ..., P n}, each meta-path P k Represent different types of user-product interaction relationships. By traversing these meta-paths on the graph G, the multimodal features of users and products can be mined.

[0025] S223, through the meta-path embedding recommendation algorithm, each meta-path P in the user-product relationship graph G is k Mapped to a high-dimensional embedding space, generating the corresponding feature vector φ k (G), eigenvector φ k (G) indicates the user and product in meta-path P k The relationship between

[0026] S224. For all meta-paths P k The eigenvector φ k(G) performs adaptive weighted fusion to generate the final multimodal feature embedding vector φ(G) of users and products:

[0027]

[0028] Among them, α k Represents the meta-path P k The adaptive weight coefficient, σ is the activation function, σ(φ k (G)) represents the feature vector φ for different meta-paths k (G) performs activation and captures complex nonlinear relationships, where ⊙ represents the element-by-element multiplication operation;

[0029] S225. Use the deep learning model to further extract features and reduce the dimension of the generated multimodal feature embedding vector φ(G) to obtain the final product feature embedding vector φ(P), which will be used to calculate the user feature embedding vector φ(U) and determine the recommendation score.

[0030] Optionally, the S4 specifically includes:

[0031] S41. Integrate virtual reality and augmented reality technologies into the live broadcast interface to provide users with personalized virtual experience;

[0032] S42, real-time capture and analysis of the user's physical environment characteristics, using deep learning models to generate adaptive virtual backgrounds, so that the display effect of virtual products is highly integrated with the user's actual environment, enhancing the realism of AR display;

[0033] S43, using a 3D reconstruction algorithm to accurately model the user's face and body to generate a highly detailed virtual image of the user, the 3D reconstruction algorithm combines the user's real-time video stream and a pre-set 3D template to generate a 3D model that matches the user's actual image through shape matching and texture mapping technology;

[0034] S44. Based on the intelligent interaction module, the virtual goods are dynamically adjusted and personalized. Users can adjust the size, color, and style of virtual goods through voice commands and touch screen interface. The system updates and displays the adjusted effect in real time. The intelligent interaction module includes emotion recognition function, which dynamically adjusts the recommended content and display method by analyzing the user's voice tone;

[0035] S45. Use blockchain technology to achieve secure storage and privacy protection of user virtual interaction data. The user's historical records, preference data and interaction records during the virtual experience process will be encrypted and stored in a distributed ledger, while supporting multi-platform data synchronization and personalized services;

[0036] S46: Generate a personalized recommendation vector R=(r1, r2, r n ), where r i Indicates the user's interest in the i-th product category. The product recommendation list and display order in the live broadcast are dynamically adjusted based on this vector.

[0037] Optionally, the S5 specifically includes:

[0038] S51, collect the user's real-time behavior data, and construct the user behavior matrix A=(a i，j ), where a i，j represents the specific behavior weight of user i at time i;

[0039] S52, using the improved collaborative filtering algorithm, according to the user behavior matrix A and the behavior data of similar users, calculate the similarity matrix S = (s i,k );

[0040] S53, based on the similarity matrix S, generate a user recommendation weight vector R = (r1, r2, r m ), where r j Indicates the recommendation score of product j for the target user:

[0041] r j =∑ k∈N(i) s i,k ·(a k，j +δ j );

[0042] Among them, s i，k represents the similarity between user i and user k, N(i) is the set of users with the highest similarity to user i, δ j is the novelty adjustment factor of product j;

[0043] S54, update the recommendation weight vector R in real time, dynamically adjust the recommendation results according to the real-time behavior data of the user during the live broadcast, and use the time decay function λ(t) to make the recent behavior have a higher impact on the recommendation results:

[0044]

[0045] in, is the new recommendation score, λ(t)=exp(-γt), γ is the time decay rate, and t is the time difference from the occurrence of the behavior to the present;

[0046] S55, compare the real-time updated recommendation weight vector R with the product feature vector T = (t1, t2, t n ) to integrate user preferences and product features to achieve multi-dimensional personalized recommendations;

[0047] S56. Optimize recommendation results through multi-head attention mechanism;

[0048] S57. Display the final recommendation results in the live broadcast interface, and continuously optimize the recommendation algorithm and recommendation list based on the real-time feedback and interactive behavior of users.

[0049] Optionally, the S52 specifically includes:

[0050] S521. Calculate user behavior weight w j , dynamically adjust the weights based on the frequency, time and type of user interaction behavior. The weight calculation combines the behavior influencing factors and user preference deviations:

[0051]

[0052] Among them, f j represents the frequency of behavior j, t j Indicates the time difference from the behavior to the present, p j represents the user preference score for behavior j, is the average preference score of the user, α, β, γ are adjustment coefficients, δ and ζ are nonlinear adjustment parameters;

[0053] S522, introducing user bias term b i,k , the calculation of the bias term combines behavioral consistency and preference volatility:

[0054]

[0055] Among them, μ i，k and ν i，k are the preference stability scores of users i and k, respectively, i，j represents the specific behavior weight of user i at time j, a k，j represents the specific behavior weight of user k at time j, is the preference volatility among users;

[0056] S523, based on weight w j and the bias term b i，k The comprehensive effect of the multi-level weight aggregation algorithm is used to calculate the user similarity s i，k :

[0057]

[0058] Among them, s i，k represents the similarity between user i and user k, a i，j represents the specific behavior weight of user i at time j, a k，j represents the specific behavior weight of user k at time j, w j represents the weight of behavior j, represents the average preference of user i for a specific behavior, represents the average preference of user k for a specific behavior, b i，k is the bias term of user i and k, η j is the hierarchical weight coefficient indicating the importance of behavior j at different levels, and φ is the bias adjustment coefficient;

[0059] S524, using an adaptive learning algorithm to adjust the parameters α, β, γ, δ, ζ and η j Optimize

[0060] S525, the calculated similarity matrix S=(s i,k ) is applied to the recommendation system, combining the user's historical behavior data and real-time interaction behavior to generate a personalized product recommendation list.

[0061] Optionally, the S6 specifically includes:

[0062] S61, collect and preprocess the real-time interactive data of users to form a multimodal data set M = {m1, m2, ..., m n}, where m i represents i interactive data;

[0063] S62. Adopt the adaptive multimodal fusion sentiment analysis algorithm to extract features and analyze sentiment of user interaction data. Dynamically allocate the weight of each modality through the improved sentiment complementary enhancement mechanism, and generate the fusion feature vector F by combining the cross-modal sentiment feature interaction model. i ;

[0064] S63, using a multi-layer deep sentiment classifier to classify the fusion feature vector F i Perform emotion recognition and generate a high-dimensional emotion vector E;

[0065] S64, based on the historical emotion data and the high-dimensional emotion vector E, calculating the user's emotion stability vector C, where the emotion stability vector C reflects the change of the user's emotion by measuring the fluctuation degree of the current emotion score;

[0066] S65, using the emotion stability vector C, generating personalized interaction strategy suggestions, which are delivered to the anchor in real time;

[0067] S66. Through emotion dynamics tracking, continuously monitor and record the user's emotion changes, and store the data in an emotion dynamics database.

[0068] Optionally, the S6 specifically includes:

[0069] S621, preprocessing the user text chat data to generate a text embedding vector stem;

[0070] S622: Extract features from the user's voice comment data, capture the emotional changes in the voice signal through a recursive neural network, and generate a voice feature vector V i ;

[0071] S623, extracting multi-dimensional behavior features from the user's non-verbal behavior to generate a behavior feature vector N i , and capture the implicit emotional trends in user behavior sequences through behavioral pattern decoding algorithms;

[0072] S624, using the improved sentiment complementary enhancement mechanism, the text embedding vector T i , speech feature vector V i and the behavioral feature vector N j Fusion is performed to generate a fusion feature vector F j :

[0073]

[0074] Among them, θ ij is the emotional complementarity weight, represents the sentiment complementation operation between features, T j ★V j ★N j represents the joint interaction of modal features, κ(E j ) is the sentiment enhancement function, ⊙ represents the dot product operation;

[0075] S625. Optimizing F through cross-modal emotional feature interaction model i The emotion recognition capability is combined with a nonlinear emotion conversion layer to identify the complex and multiple emotional states of users;

[0076] S626, F i Perform standardization.

[0077] A freely collocation type live broadcast device for e-commerce operation according to an embodiment of the present invention includes:

[0078] The data collection module is used to collect users’ browsing history, purchase records, search keywords, social media interaction data, and real-time interaction data of users during live broadcasts to form a multimodal data set;

[0079] The user portrait generation module uses the meta-path embedding recommendation algorithm to build user portraits based on the collected user data. It constructs a user behavior matrix and a product feature matrix to achieve a high-dimensional embedding representation of multimodal features.

[0080] The product recommendation module uses an improved collaborative filtering algorithm and an adaptive multimodal fusion sentiment analysis algorithm to generate a personalized product recommendation list based on user portraits and real-time behavior data. The product recommendation module combines the user's historical behavior and real-time interaction data to dynamically adjust the recommendation results and use a time decay function to optimize the recommendation;

[0081] The sentiment analysis module includes text embedding generation, speech feature extraction, and behavior feature extraction submodules. It generates fusion feature vectors through an improved sentiment complement enhancement mechanism, uses a cross-modal sentiment feature interaction model for sentiment recognition, generates high-dimensional sentiment vectors, and calculates the user's sentiment stability vector.

[0082] Virtual reality and augmented reality module integrates virtual reality and augmented reality technologies into the live broadcast interface to provide users with personalized display and trial functions of virtual goods. The virtual reality and augmented reality module combines deep learning models to generate adaptive virtual backgrounds and supports 3D reconstruction of user virtual images.

[0083] The interactive recommendation module generates personalized interactive strategy suggestions in real time based on sentiment analysis results and personalized recommendation vectors;

[0084] The data storage and privacy protection module uses blockchain technology to realize encrypted storage of user virtual interaction data.

[0085] The beneficial effects of the present invention are:

[0086] (1) Improve the accuracy of personalized recommendations: By using the meta-path embedding recommendation algorithm and the improved collaborative filtering algorithm, combined with user portraits and real-time behavior data, personalized product recommendation lists can be generated and dynamically updated more accurately. This fusion of multimodal data and analysis of real-time behavior makes the recommendation results more closely aligned with the user's current interests and needs, significantly improving the accuracy of recommendations and user satisfaction.

[0087] (2) Enhance the user's interactive experience: During the live broadcast, the present invention displays the diversified combination effects of commodities through the anchor and provides real-time interactive functions, such as chatting, liking and barrage. The adaptive multimodal fusion sentiment analysis algorithm can analyze the user's emotional tendency in real time and provide personalized interactive suggestions for the anchor, thus enhancing the user's sense of participation and interactive experience. This emotion-driven interactive method makes it easier for users to engage in the shopping process and increases the possibility of purchase.

[0088] (3) Providing an immersive shopping experience: By introducing virtual reality and augmented reality technologies, the present invention realizes the display and trial functions of virtual goods. Users can experience the various functions and effects of goods through virtual reality, including adjusting the size, color and style of goods, etc., to obtain a more realistic and comprehensive shopping experience. The use of three-dimensional reconstruction technology and adaptive virtual background makes the virtual display more realistic and enhances the user's sense of immersion.

[0089] (4) Improving the timeliness and relevance of real-time recommendations: The present invention introduces a time decay function into the recommendation algorithm, so that recent behavior has a greater impact on the recommendation results. By analyzing the user's behavior data in real time, the recommended content is dynamically adjusted to ensure the timeliness and relevance of the recommendation results and meet the user's immediate needs. This dynamic adjustment mechanism enables the recommendation system to respond to changes in user behavior more quickly. BRIEF DESCRIPTION OF THE DRAWINGS

[0090] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0091] Figure 1 A flowchart of a free-combination combined e-commerce operation live broadcast method and device proposed by the present invention;

[0092] Figure 2 This is an overview of the system architecture of a free-combination combined e-commerce operation live broadcast method and device proposed by the present invention; DETAILED DESCRIPTION

[0093] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0094] refer to Figure 1-2 , a free combination e-commerce operation live broadcast method and device, comprising the following steps:

[0095] S1. Build user portraits by collecting and analyzing user data;

[0096] S2. Based on user portraits and real-time behavior data, a meta-path embedding recommendation algorithm is used to generate a personalized product recommendation list, which is dynamically updated during the live broadcast;

[0097] S3. During the live broadcast, the anchor will display the diversified combination effects of the products in real time and provide real-time interactive functions;

[0098] S4. Use virtual reality and augmented reality technologies to integrate virtual functions into the live broadcast interface;

[0099] S5. Use an improved collaborative filtering algorithm to analyze users’ real-time behavior data and recommend related products in real time based on the purchase records of similar users.

[0100] S6. Adopting the sentiment analysis algorithm of adaptive multimodal fusion to analyze the emotional tendency of users in real-time interaction and provide personalized interaction suggestions for anchors;

[0101] S7. Provides integrated and convenient online shopping functions, allowing users to seamlessly complete purchase operations while watching live broadcasts.

[0102] In this implementation, S2 specifically includes:

[0103] S21. Through the user portrait generation process, collect the user's browsing history, purchase record, search keywords and social media interaction data, and construct the user behavior matrix U = (u i，j ), where u i，j represents the interaction weight of user i on product i;

[0104] S22, based on the user behavior matrix U and the product feature matrix P = (p j，k ), where P j，k The feature k representing the product j is extracted by using the meta-path embedding recommendation algorithm. The meta-path embedding recommendation algorithm constructs a multi-dimensional relationship path between users and products to achieve a high-dimensional embedding representation of multimodal features.

[0105] S23, after generating the product feature embedding vector φ(P), using the feature selection mechanism based on deep learning, combined with the user behavior matrix U, calculate the user feature embedding vector φ(U);

[0106] S24. Calculate the inner product of the product feature embedding vector φ(P) and the user feature embedding vector φ(U) in the high-dimensional vector space to obtain the product recommendation score r i，j , where r i，j It represents the recommendation score of user i for product j, and optimizes the recommendation result through the weighted adjustment mechanism;

[0107] S25. Based on real-time behavior data and user feedback, the recommendation model is learned and iteratively optimized online, and the recommendation list is dynamically updated.

[0108] In this implementation manner, the S22 specifically includes:

[0109] S221, through the user behavior matrix U = (u i，j ) and product feature matrix P = (p j，k), construct a user-item relationship graph G = (V, E), where V represents the set of users and items, and E represents the interaction relationship between users and items;

[0110] S222: In the user-product relationship graph G, define multiple meta-paths P = {P1, P2, ..., P n}, each meta-path P k Represent different types of user-product interaction relationships. By traversing these meta-paths on the graph G, the multimodal features of users and products can be mined.

[0111] S223, through the meta-path embedding recommendation algorithm, each meta-path P in the user-product relationship graph G is k Mapped to a high-dimensional embedding space, generating the corresponding feature vector φ k (G), eigenvector φ k (G) indicates the user and product in meta-path P k The relationship between

[0112] S224. For all meta-paths P k The eigenvector φ k (G) performs adaptive weighted fusion to generate the final multimodal feature embedding vector φ(G) of users and products:

[0113]

[0114] Among them, α k represents the adaptive weight coefficient of the meta-path Px, σ is the activation function, σ(φ k (G)) represents the feature vector φ for different meta-paths k (G) performs activation and captures complex nonlinear relationships, where ⊙ represents the element-by-element multiplication operation;

[0115] S225. Use the deep learning model to further extract features and reduce the dimension of the generated multimodal feature embedding vector φ(G) to obtain the final product feature embedding vector φ(P), which will be used to calculate the user feature embedding vector φ(U) and determine the recommendation score.

[0116] In this implementation manner, the S4 specifically includes:

[0117] S41. Integrate virtual reality and augmented reality technologies into the live broadcast interface to provide users with personalized virtual experience;

[0118] S42, real-time capture and analysis of the user's physical environment characteristics, using deep learning models to generate adaptive virtual backgrounds, so that the display effect of virtual products is highly integrated with the user's actual environment, enhancing the realism of AR display;

[0119] S43, using a 3D reconstruction algorithm to accurately model the user's face and body to generate a highly detailed virtual image of the user, the 3D reconstruction algorithm combines the user's real-time video stream and a pre-set 3D template to generate a 3D model that matches the user's actual image through shape matching and texture mapping technology;

[0120] S44. Based on the intelligent interaction module, the virtual goods are dynamically adjusted and personalized. Users can adjust the size, color, and style of virtual goods through voice commands and touch screen interface. The system updates and displays the adjusted effect in real time. The intelligent interaction module includes emotion recognition function, which dynamically adjusts the recommended content and display method by analyzing the user's voice tone;

[0121] S45. Use blockchain technology to achieve secure storage and privacy protection of user virtual interaction data. The user's historical records, preference data and interaction records during the virtual experience process will be encrypted and stored in a distributed ledger, while supporting multi-platform data synchronization and personalized services;

[0122] S46: Generate a personalized recommendation vector R=(r1, r2, r n ), where r i Indicates the user's interest in the i-th product category. The product recommendation list and display order in the live broadcast are dynamically adjusted based on this vector.

[0123] In this implementation manner, S5 specifically includes:

[0124] S51, collect the user's real-time behavior data, and construct the user behavior matrix A=(a i，j ), where a i，j represents the specific behavior weight of user i at time j;

[0125] S52, using the improved collaborative filtering algorithm, according to the user behavior matrix A and the behavior data of similar users, calculate the similarity matrix S = (s i，k );

[0126] S53, based on the similarity matrix S, generate a user recommendation weight vector R = (r1, r2, r m ), where r j Indicates the recommendation score of product j for the target user:

[0127] r j =∑ k∈N(i) s i,k ·(a k，j +δ j );

[0128] Among them, s i，krepresents the similarity between user i and user k, N(i) is the set of users with the highest similarity to user i, δ j is the novelty adjustment factor of product j;

[0129] S54, real-time update of the recommendation weight vector R, according to the real-time behavior data of users during the live broadcast, dynamically adjust the recommendation results, using the time decay function λ (t) Make recent behavior have a higher impact on recommendation results:

[0130]

[0131] in, is the new recommendation score, λ(t)=exp(-γt), γ is the time decay rate, and t is the time difference from the occurrence of the behavior to the present;

[0132] S55, compare the real-time updated recommendation weight vector R with the product feature vector T = (t1, t2, t n ) to integrate user preferences and product features to achieve multi-dimensional personalized recommendations;

[0133] S56. Optimize recommendation results through multi-head attention mechanism;

[0134] S57. Display the final recommendation results in the live broadcast interface, and continuously optimize the recommendation algorithm and recommendation list based on the real-time feedback and interactive behavior of users.

[0135] In this implementation manner, the S52 specifically includes:

[0136] S521. Calculate user behavior weight w j , dynamically adjust the weights based on the frequency, time and type of user interaction behavior. The weight calculation combines the behavior influencing factors and user preference deviations:

[0137]

[0138] Among them, f j represents the frequency of behavior j, t j Indicates the time difference from the behavior to the present, p j represents the user preference score for behavior j, is the average preference score of the user, α, β, γ are adjustment coefficients, δ and ζ are nonlinear adjustment parameters;

[0139] S522, introducing user bias term b i ,k,The calculation of the bias term combines behavioral consistency and preference volatility:

[0140]

[0141] Among them, μ i，k and ν i，k are the preference stability scores of users i and k, respectively, i，j represents the specific behavior weight of user i at time j, a k，j represents the specific behavior weight of user k at time j, is the preference volatility among users;

[0142] S523, based on weight w j and the bias term b i，k The comprehensive effect of the multi-level weight aggregation algorithm is used to calculate the user similarity s i，k :

[0143]

[0144] Among them, s i，k represents the similarity between user i and user k, a i，j represents the specific behavior weight of user i at time j, a k，j represents the specific behavior weight of user k at time j, w j represents the weight of behavior j, represents the average preference of user i for a specific behavior, represents the average preference of user k for a specific behavior, b i，k is the bias term of user i and k, η j is the hierarchical weight coefficient indicating the importance of behavior j at different levels, and φ is the bias adjustment coefficient;

[0145] S524, using an adaptive learning algorithm to adjust the parameters α, β, γ, δ, ζ and η j Optimize

[0146] S525, the calculated similarity matrix S=(s i,k ) is applied to the recommendation system, combining the user's historical behavior data and real-time interaction behavior to generate a personalized product recommendation list.

[0147] In this implementation manner, S6 specifically includes:

[0148] S61, collect and preprocess the real-time interactive data of users to form a multimodal data set M = {m i , m2, ..., m n}, where m i represents i interactive data;

[0149] S62. Adopt the adaptive multimodal fusion sentiment analysis algorithm to extract features and analyze sentiment of user interaction data. Dynamically allocate the weight of each modality through the improved sentiment complementary enhancement mechanism, and generate the fusion feature vector F by combining the cross-modal sentiment feature interaction model. i ;

[0150] S63, using a multi-layer deep sentiment classifier to classify the fusion feature vector F i Perform emotion recognition and generate a high-dimensional emotion vector E;

[0151] S64, based on the historical emotion data and the high-dimensional emotion vector E, calculating the user's emotion stability vector C, where the emotion stability vector C reflects the change of the user's emotion by measuring the fluctuation degree of the current emotion score;

[0152] S65, using the emotion stability vector C, generating personalized interaction strategy suggestions, which are delivered to the anchor in real time;

[0153] S66. Through emotion dynamics tracking, continuously monitor and record the user's emotion changes, and store the data in an emotion dynamics database.

[0154] In this implementation manner, S6 specifically includes:

[0155] S621, preprocessing the user text chat data to generate a text embedding vector stem;

[0156] S622: Extract features from the user's voice comment data, capture the emotional changes in the voice signal through a recursive neural network, and generate a voice feature vector V i ;

[0157] S623, extracting multi-dimensional behavior features from the user's non-verbal behavior to generate a behavior feature vector N i , and capture the implicit emotional trends in user behavior sequences through behavioral pattern decoding algorithms;

[0158] S624, using the improved sentiment complementary enhancement mechanism, the text embedding vector T i , speech feature vector V i and the behavioral feature vector N i Fusion is performed to generate a fusion feature vector F i :

[0159]

[0160] Among them, θ ij is the emotional complementarity weight, represents the sentiment complementation operation between features, T j ★V j ★N jrepresents the joint interaction of modal features, κ(E j ) is the sentiment enhancement function, ⊙ represents the dot product operation;

[0161] S625. Optimizing F through cross-modal emotional feature interaction model i The emotion recognition capability is combined with a nonlinear emotion conversion layer to identify the complex and multiple emotional states of users;

[0162] S626, F i Perform standardization.

[0163] A freely collocation type live broadcast device for e-commerce operation according to an embodiment of the present invention includes:

[0164] The data collection module is used to collect users’ browsing history, purchase records, search keywords, social media interaction data, and real-time interaction data of users during live broadcasts to form a multimodal data set;

[0165] The user portrait generation module uses the meta-path embedding recommendation algorithm to build user portraits based on the collected user data. It constructs a user behavior matrix and a product feature matrix to achieve a high-dimensional embedding representation of multimodal features.

[0166] The product recommendation module uses an improved collaborative filtering algorithm and an adaptive multimodal fusion sentiment analysis algorithm to generate a personalized product recommendation list based on user portraits and real-time behavior data. The product recommendation module combines the user's historical behavior and real-time interaction data to dynamically adjust the recommendation results and use a time decay function to optimize the recommendation;

[0167] The sentiment analysis module includes text embedding generation, speech feature extraction, and behavior feature extraction submodules. It generates fusion feature vectors through an improved sentiment complement enhancement mechanism, uses a cross-modal sentiment feature interaction model for sentiment recognition, generates high-dimensional sentiment vectors, and calculates the user's sentiment stability vector.

[0168] Virtual reality and augmented reality module integrates virtual reality and augmented reality technologies into the live broadcast interface to provide users with personalized display and trial functions of virtual goods. The virtual reality and augmented reality module combines deep learning models to generate adaptive virtual backgrounds and supports 3D reconstruction of user virtual images.

[0169] The interactive recommendation module generates personalized interactive strategy suggestions in real time based on sentiment analysis results and personalized recommendation vectors;

[0170] The data storage and privacy protection module uses blockchain technology to realize encrypted storage of user virtual interaction data.

[0171] Embodiment 1:

[0172] In Company B, a large online retail company in Shanghai, China, a free-to-match combination e-commerce operation live broadcast method and device was implemented. Company B is a leading e-commerce platform with annual sales of more than 50 billion yuan, thousands of brand partners and millions of users. In order to improve the user's shopping experience and the platform's sales conversion rate, Company B decided to adopt the live broadcast method and device of the present invention to enhance the user's interactive experience through personalized recommendations, sentiment analysis and virtual reality technology.

[0173] Every Friday night, Company B holds a themed live broadcast event, inviting well-known anchors to showcase new fashion clothing and accessories. During the live broadcast, users can not only watch the anchor's product display, but also interact with the anchor in real time, ask questions or make comments. In order to optimize the user experience and improve the sales conversion rate, Company B deployed the live broadcast system of the present invention. Before the live broadcast begins, the system establishes a detailed user portrait through the user's browsing history, purchase history and search keywords. Through the meta-path embedding recommendation algorithm, the system generates a personalized product recommendation list for each user. These recommended products will be dynamically updated according to the user's interactive behavior in the live broadcast (such as likes, comments, etc.). For example, when a user likes a certain product, the system will adjust the recommendation list in real time and give priority to other products with similar styles to the product. During the live broadcast, the anchor showed a variety of fashionable clothing and provided users with a virtual try-on experience through virtual reality technology. Users can choose the products they are interested in, and the system will use three-dimensional reconstruction technology to generate the user's virtual image and simulate the wearing effect of the product. At the same time, users can also adjust the size, color and style of virtual products through voice commands to obtain a personalized try-on experience. The system also uses an adaptive multimodal fusion sentiment analysis algorithm to analyze users' text and voice comments in real time and identify their emotional states. By analyzing users' emotional tendencies, the system provides personalized interaction suggestions for anchors. For example, when the system detects that a user has positive emotions towards a certain product, the anchor will further introduce the details and unique features of the product to enhance the user's desire to buy.

[0174] During the three-month test period, Company B conducted 12 live broadcasts using the system of the present invention, involving more than 500 products. Data from the test period showed that compared with the live broadcasts of the same period without the system of the present invention, the average number of viewers after using the system of the present invention increased by 35% and the interaction rate increased by 50%. The average sales conversion rate of each live broadcast increased from 5% to 12%, and sales increased by approximately RMB 2.4 million.

[0175] Table 1 Live broadcast effect data comparison table

[0176]

[0177]

[0178] As can be seen from Table 1, after implementing the system of the present invention, the live broadcast activities of Company B have significantly improved in terms of user participation and sales conversion rate. The system successfully converts user interest into actual purchases through personalized recommendations and sentiment analysis. At the same time, the application of virtual reality technology enables users to intuitively experience the effects of products during the live broadcast, further enhancing users' confidence in purchasing decisions.

[0179] In summary, the free-combination combination e-commerce operation live broadcast method and device of the present invention not only significantly improves the user's personalized shopping experience and interactive participation, but also effectively improves the platform's sales conversion rate and sales, proving the important value and broad application prospects of the present invention in e-commerce operations.

[0180] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A free combination e-commerce operation live broadcast method, characterized in that: The following steps are involved: S1. Build user portraits by collecting and analyzing user data; S2. Based on user portraits and real-time behavior data, a meta-path embedding recommendation algorithm is used to generate a personalized product recommendation list, which is dynamically updated during the live broadcast; S3. During the live broadcast, the anchor will display the diversified combination effects of the products in real time and provide real-time interactive functions; S4. Use virtual reality and augmented reality technologies to integrate virtual functions into the live broadcast interface; S5. Use an improved collaborative filtering algorithm to analyze users’ real-time behavior data and recommend related products in real time based on the purchase records of similar users. S6. Adopting the sentiment analysis algorithm of adaptive multimodal fusion to analyze the emotional tendency of users in real-time interaction and provide personalized interaction suggestions for anchors; S7. Provides integrated and convenient online shopping functions, allowing users to seamlessly complete purchase operations while watching live broadcasts.

2. A free-to-match combined e-commerce operation live broadcast method according to claim 1, characterized in that: The S2 specifically includes: S21. Through the user portrait generation process, collect the user's browsing history, purchase record, search keywords and social media interaction data, and construct the user behavior matrix U = (u i,j ), where u i,j represents the interaction weight of user i on product j; S22, based on the user behavior matrix U and the product feature matrix P = (p j,k ), where p j,k The feature k representing the product j is extracted by using the meta-path embedding recommendation algorithm. The meta-path embedding recommendation algorithm constructs a multi-dimensional relationship path between users and products to achieve a high-dimensional embedding representation of multimodal features. S23, after generating the product feature embedding vector φ(P), using the feature selection mechanism based on deep learning, combined with the user behavior matrix U, calculate the user feature embedding vector φ(U); S24, perform inner product calculation on the product feature embedding vector φ(P) and the user feature embedding vector φ(U) in the high-dimensional vector space to obtain the product recommendation score r i,j , where r i,j It represents the recommendation score of user i for product j, and optimizes the recommendation result through the weighted adjustment mechanism; S25. Based on real-time behavior data and user feedback, the recommendation model is learned and iteratively optimized online, and the recommendation list is dynamically updated.

3. A free-to-match combined e-commerce operation live broadcast method according to claim 2, characterized in that: The S22 specifically includes: S221, through the user behavior matrix U = (u i,j ) and product feature matrix P = (p j,k ), construct a user-item relationship graph G = (V, E), where V represents the set of users and items, and E represents the interaction relationship between users and items; S222. In the user-product relationship graph G, define multiple meta-paths P = {P1, P2, ..., P n }, each meta-path P k Represent different types of user-product interaction relationships. By traversing these meta-paths on the graph G, the multimodal features of users and products can be mined. S223, through the meta-path embedding recommendation algorithm, each meta-path P in the user-product relationship graph G is k Mapped to a high-dimensional embedding space, generating the corresponding feature vector φ k (G), eigenvector φ k (G) indicates the user and product in the meta-path P k The relationship between S224. For all meta-paths P k The eigenvector φ k (G) performs adaptive weighted fusion to generate the final multimodal feature embedding vector φ(G) of users and products: Among them, α k Represents the meta-path P k The adaptive weight coefficient, σ is the activation function, σ(φ k (G)) represents the feature vector φ for different meta-paths k (G) performs activation and captures complex nonlinear relationships, where ⊙ represents the element-by-element multiplication operation; S225. Use the deep learning model to further extract features and reduce the dimension of the generated multimodal feature embedding vector φ(G) to obtain the final product feature embedding vector φ(P), which will be used to calculate the user feature embedding vector φ(U) and determine the recommendation score.

4. A free-to-match combined e-commerce operation live broadcast method according to claim 1, characterized in that: The S4 specifically includes: S41. Integrate virtual reality and augmented reality technologies into the live broadcast interface to provide users with personalized virtual experience; S42, real-time capture and analysis of the user's physical environment characteristics, using deep learning models to generate adaptive virtual backgrounds, so that the display effect of virtual products is highly integrated with the user's actual environment, enhancing the realism of AR display; S43, using a 3D reconstruction algorithm to accurately model the user's face and body to generate a highly detailed virtual image of the user, the 3D reconstruction algorithm combines the user's real-time video stream and a pre-set 3D template to generate a 3D model that matches the user's actual image through shape matching and texture mapping technology; S44. Based on the intelligent interaction module, the virtual goods are dynamically adjusted and personalized. Users can adjust the size, color, and style of virtual goods through voice commands and touch screen interface. The system updates and displays the adjusted effect in real time. The intelligent interaction module includes emotion recognition function, which dynamically adjusts the recommended content and display method by analyzing the user's voice tone; S45. Use blockchain technology to achieve secure storage and privacy protection of user virtual interaction data. The user's historical records, preference data and interaction records during the virtual experience process will be encrypted and stored in a distributed ledger, while supporting multi-platform data synchronization and personalized services; S46, based on the virtual user behavior data and preference information, generate a personalized recommendation vector R = r1, r2,, r n ), where r i Indicates the user's interest in the i-th product category. The product recommendation list and display order in the live broadcast are dynamically adjusted based on this vector.

5. A free-to-match combined e-commerce operation live broadcast method according to claim 1, characterized in that: The S5 specifically includes: S51, collect the user's real-time behavior data, and construct the user behavior matrix A=(a i,j ), where a i,j represents the specific behavior weight of user i at time j; S52, using the improved collaborative filtering algorithm, according to the user behavior matrix A and the behavior data of similar users, calculate the similarity matrix S = (s i,k ); S53, based on the similarity matrix S, generate a user recommendation weight vector R = (r1, r2,, r m ), where r j Indicates the recommendation score of product j for the target user: r j =∑ k∈N(i) s i,k ·(a k,j +d j ); Among them, s i,k represents the similarity between user i and user k, N(i) is the set of users with the highest similarity to user i, δ j is the novelty adjustment factor of product j; S54, update the recommendation weight vector R in real time, dynamically adjust the recommendation results according to the real-time behavior data of the user during the live broadcast, and use the time decay function λ(t) to make the recent behavior have a higher impact on the recommendation results: in, is the new recommendation score, λ(t)=exp(-γt), γ is the time decay rate, and t is the time difference from the occurrence of the behavior to the present; S55, compare the real-time updated recommendation weight vector R with the product feature vector T = (t1, t2,, t n ) to integrate user preferences and product features to achieve multi-dimensional personalized recommendations; S56. Optimize recommendation results through multi-head attention mechanism; S57. Display the final recommendation results in the live broadcast interface, and continuously optimize the recommendation algorithm and recommendation list based on the real-time feedback and interactive behavior of users.

6. A free-to-match combined e-commerce operation live broadcast method according to claim 5, characterized in that: The S52 specifically includes: S521. Calculate user behavior weight w j , dynamically adjust the weights based on the frequency, time and type of user interaction behavior. The weight calculation combines the behavior influencing factors and user preference deviations: Among them, f j represents the frequency of behavior j, t j Indicates the time difference from the behavior to the present, p j represents the user preference score for behavior j, is the average preference score of the user, α, β, γ are adjustment coefficients, δ and ζ are nonlinear adjustment parameters; S522, introducing user bias term b i,k , the calculation of the bias term combines behavioral consistency and preference volatility: Among them, μ i,k and ν i,k are the preference stability scores of users i and k, respectively, i,j represents the specific behavior weight of user i at time j, a k,j represents the specific behavior weight of user k at time j, is the preference volatility among users; S523, based on weight w j and the bias term b i,k The comprehensive effect of the multi-level weight aggregation algorithm is used to calculate the user similarity s i,k : Among them, s i,k represents the similarity between user i and user k, a i,j represents the specific behavior weight of user i at time j, a k,j represents the specific behavior weight of user k at time j, w j represents the weight of behavior j, represents the average preference of user i for a specific behavior, represents the average preference of user k for a specific behavior, b i,k is the bias term of user i and k, η j is the hierarchical weight coefficient indicating the importance of behavior j at different levels, and φ is the bias adjustment coefficient; S524, using an adaptive learning algorithm to adjust the parameters α, β, γ, δ, ζ and η j Optimize S525, the calculated similarity matrix S=(s i,k ) is applied to the recommendation system, combining the user's historical behavior data and real-time interaction behavior to generate a personalized product recommendation list.

7. A free-to-match combined e-commerce operation live broadcast method according to claim 1, characterized in that: The S6 specifically includes: S61. Collect and preprocess users’ real-time interactive data to form a multimodal data set M=m1,m2,...,m n }, where m i represents i interactive data; S62. Adopt the adaptive multimodal fusion sentiment analysis algorithm to extract features and analyze sentiment of user interaction data. Dynamically allocate the weight of each modality through the improved sentiment complementary enhancement mechanism, and generate the fusion feature vector F by combining the cross-modal sentiment feature interaction model. i ; S63, using a multi-layer deep sentiment classifier to classify the fusion feature vector F i Perform emotion recognition and generate a high-dimensional emotion vector E; S64, based on the historical emotion data and the high-dimensional emotion vector E, calculating the user's emotion stability vector C, where the emotion stability vector C reflects the change of the user's emotion by measuring the fluctuation degree of the current emotion score; S65, using the emotion stability vector C, generating personalized interaction strategy suggestions, which are delivered to the anchor in real time; S66. Through emotion dynamics tracking, continuously monitor and record the user's emotion changes, and store the data in an emotion dynamics database.

8. A free-to-match combined e-commerce operation live broadcast method according to claim 7, characterized in that: The S6 specifically includes: S621, preprocess the user text chat data to generate a text embedding vector T i ; S622: Extract features from the user's voice comment data, capture the emotional changes in the voice signal through a recursive neural network, and generate a voice feature vector V i ; S623, extracting multi-dimensional behavior features from the user's non-verbal behavior to generate a behavior feature vector N i , and capture the implicit emotional trends in user behavior sequences through behavioral pattern decoding algorithms; S624, using the improved sentiment complementary enhancement mechanism, the text embedding vector T i , speech feature vector V i and the behavioral feature vector N i Fusion is performed to generate a fusion feature vector F i : Among them, θ ij is the emotional complementarity weight, represents the sentiment complementation operation between features, T j ★V j ★N j represents the joint interaction of modal features, κ(E j ) is the sentiment enhancement function, ⊙ represents the dot product operation; S625. Optimizing F through cross-modal emotional feature interaction model i The emotion recognition capability is combined with a nonlinear emotion conversion layer to identify the complex and multiple emotional states of users; S626, F i Perform standardization.

9. A freely collocation type live broadcast device for e-commerce operation, characterized in that: include: The data collection module is used to collect users’ browsing history, purchase records, search keywords, social media interaction data, and real-time interaction data of users during live broadcasts to form a multimodal data set; The user portrait generation module uses the meta-path embedding recommendation algorithm to build user portraits based on the collected user data. It constructs a user behavior matrix and a product feature matrix to achieve a high-dimensional embedding representation of multimodal features. The product recommendation module uses an improved collaborative filtering algorithm and an adaptive multimodal fusion sentiment analysis algorithm to generate a personalized product recommendation list based on user portraits and real-time behavior data. The product recommendation module combines the user's historical behavior and real-time interaction data to dynamically adjust the recommendation results and use a time decay function to optimize the recommendation; The sentiment analysis module includes text embedding generation, speech feature extraction, and behavior feature extraction submodules. It generates fusion feature vectors through an improved sentiment complement enhancement mechanism, uses a cross-modal sentiment feature interaction model for sentiment recognition, generates high-dimensional sentiment vectors, and calculates the user's sentiment stability vector. Virtual reality and augmented reality module integrates virtual reality and augmented reality technologies into the live broadcast interface to provide users with personalized display and trial functions of virtual goods. The virtual reality and augmented reality module combines deep learning models to generate adaptive virtual backgrounds and supports 3D reconstruction of user virtual images. The interactive recommendation module generates personalized interactive strategy suggestions in real time based on sentiment analysis results and personalized recommendation vectors; The data storage and privacy protection module uses blockchain technology to realize encrypted storage of user virtual interaction data.

Images