Intelligent commodity recommendation system based on live stream real-time analysis

Abstract

The invention relates to the technical field of live broadcast e-commerce intelligent recommendation, in particular to an intelligent commodity recommendation system based on live broadcast stream real-time analysis, which comprises the following steps: analyzing live broadcast video data, live broadcast audio data and user interaction behavior data to generate a live broadcast event sequence, and extracting anchor behavior characteristics based on the live broadcast event sequence; establishing an anchor behavior sequence model to identify an anchor behavior state corresponding to the current live broadcast stage; meanwhile, a live broadcast commodity evolution path model is constructed according to historical live broadcast data, candidate commodity paths are determined, and a candidate commodity set is generated; and according to the user interaction behavior data, constructing a user interest propagation graph and identifying a user interest propagation path, screening a target commodity set from the candidate commodity set, and generating a commodity recommendation result. The system can combine the real-time live broadcast behavior of the anchor and the user interest propagation relationship to realize intelligent commodity recommendation in the live broadcast scene.

Description

An intelligent product recommendation system based on real-time analysis of live streams Technical Field

[0001] This invention relates to the field of intelligent recommendation technology for live-streaming e-commerce, and in particular to an intelligent product recommendation system based on real-time analysis of live streams. Background Technology

[0002] With the development of live streaming technology, live e-commerce has gradually become an important way to sell goods. During a live stream, the host introduces products through video demonstrations, audio explanations, and interaction with users, thereby promoting purchases. To improve the efficiency of live streaming sales, some live streaming platforms have begun to introduce product recommendation technology, recommending potentially interesting products to viewers by analyzing user behavior data or historical sales data. For example, some existing technologies build user interest models by analyzing user click records, browsing history, or purchase records, thereby generating a product recommendation list; other technologies recommend products by analyzing historical product sales data or product similarity information. However, in actual live streaming scenarios, the live streaming process has obvious time-series characteristics and host behavior characteristics. Hosts typically showcase different types of products at different stages of the live stream. Therefore, relying solely on historical user behavior data or static product data for recommendations often fails to fully reflect the real-time changes during the live stream.

[0003] Existing live streaming recommendation technologies typically generate recommendation results based primarily on users' historical behavior data or static product relationships. Therefore, it is difficult to dynamically adjust recommended products according to the streamer's behavior during the current live streaming phase, thus affecting the adaptability of product recommendations in live streaming scenarios. Summary of the Invention

[0004] To overcome the above shortcomings, this invention provides an intelligent product recommendation system based on real-time analysis of live streams, aiming to improve the problem of difficulty in dynamically adjusting recommended products according to the behavior of the streamer during the current live stream, thereby affecting the adaptability of product recommendations in live streaming scenarios.

[0005] This invention provides the following technical solution: an intelligent product recommendation system based on real-time analysis of live streams, comprising the following modules: a live stream data parsing module, used to acquire live stream video data, live stream audio data, and user interaction behavior data, and to parse the live stream video data, live stream audio data, and user interaction behavior data to generate a live stream event sequence arranged in chronological order; a streamer behavior state recognition module, which extracts streamer behavior features based on the live stream event sequence and establishes a streamer behavior sequence model based on historical live stream data to identify the streamer behavior state corresponding to the current live stream stage; and a product evolution path modeling module, which constructs a live stream product evolution path model based on historical live stream data, wherein the live stream product... The product evolution path model includes product nodes and path connecting edges representing the product display order. The path connecting edges contain the transition weights corresponding to the product display order, and candidate product paths are determined based on the anchor's behavior state and product information in the live event sequence. The candidate product generation module extracts product nodes from the paths based on the candidate product paths to generate a candidate product set. The group interest propagation analysis module constructs a user interest propagation graph based on user interaction behavior data and identifies user interest propagation paths in the user interest propagation graph. The target product determination module selects corresponding products from the candidate product set based on the user interest propagation paths to determine the target product set and generates product recommendation results.

[0006] By adopting the above technical solution, a live event sequence is generated by parsing live video data, live audio data, and user interaction behavior data. Based on the live event sequence, the anchor's behavioral features are extracted, and a anchor behavior sequence model is established by combining historical live data. This identifies the anchor's behavioral state corresponding to the current live stage, determines the candidate product path based on the anchor's behavioral state, and generates a candidate product set. This enables dynamic adjustment of recommended products according to the anchor's behavior in the current live stage, solving the problem that existing live recommendation technologies are unable to dynamically adjust recommended products according to the anchor's behavior in the current live stage, thus affecting the adaptability of product recommendations in live scenarios.

[0007] Preferably, in the live streaming data parsing module, generating a live streaming event sequence arranged in chronological order includes: performing image recognition on the live streaming video data to extract product display information from the live streaming screen and generating product display events; performing speech recognition on the live streaming audio data to extract semantic information from the anchor's explanation and generating anchor semantic events; parsing user interaction behavior data to identify user click behavior, comment behavior, or like behavior and generating user interaction events; and sorting the product display events, anchor semantic events, and user interaction events according to the event occurrence time to obtain the live streaming event sequence.

[0008] Preferably, in the anchor behavior state recognition module, the step of extracting anchor behavior features based on the live event sequence includes: dividing the live event sequence into time windows to obtain multiple consecutive event segments; statistically analyzing the order information of the anchor displaying products and the time interval information of the products appearing in each event segment; constructing anchor behavior sequence features based on the product display order information and the time interval information; and performing feature encoding processing on the anchor behavior sequence features to generate anchor behavior features.

[0009] Preferably, in the anchor behavior state recognition module, the step of establishing an anchor behavior sequence model based on historical live streaming data includes: extracting historical live streaming event sequences from historical live streaming data and sorting the historical live streaming event sequences according to the event occurrence time; identifying historical product display events based on the historical live streaming event sequences and constructing an anchor historical behavior sequence according to the product display order; statistically analyzing the transition relationship between product display orders based on the anchor historical behavior sequence and calculating the transition probability between adjacent product displays, and constructing an anchor behavior sequence model based on the transition probability.

[0010] Preferably, in the product evolution path modeling module, the step of constructing a live-stream product evolution path model based on historical live-stream data includes: extracting historical product display records from historical live-stream data and sorting the historical product display records according to the product display time; identifying adjacent product display relationships based on the sorted historical product display records and establishing sequential association relationships between products; establishing path connection edges between product nodes based on the sequential association relationships to construct a product evolution path structure; and generating a live-stream product evolution path model based on the product nodes and path connection edges.

[0011] Preferably, in the product evolution path modeling module, the path connection edge includes the transfer weights corresponding to the product display order, which includes: extracting historical product display sequences from historical live data; identifying adjacent product display relationships based on the historical product display sequences and counting the number of adjacent displays between product nodes; counting the transfer relationships between product nodes based on the number of adjacent displays between product nodes; calculating the transfer probability between product nodes based on the transfer relationship, and using the transfer probability as the transfer weight of the path connection edge.

[0012] Preferably, in the product evolution path modeling module, determining the candidate product path based on the anchor's behavior state and product information in the live event sequence includes: extracting the product node corresponding to the currently displayed product from the live event sequence; determining the behavior state identifier corresponding to the current live stage based on the anchor's behavior state; using the product node as the starting node in the live product evolution path model, and retrieving the path connection edge connected to the product node based on the behavior state identifier; determining the corresponding product node sequence based on the transition weight of the path connection edge, and determining the product node sequence as the candidate product path.

[0013] Preferably, in the candidate product generation module, the step of extracting product nodes from the path based on the candidate product path includes: identifying a path node sequence in the candidate product path and obtaining a set of product nodes corresponding to the path node sequence; performing node traversal processing on the product node set to extract product nodes from the path; establishing a product node list based on the extracted product nodes, and generating a candidate product set based on the product node list.

[0014] Preferably, in the group interest propagation analysis module, the step of constructing a user interest propagation graph based on user interaction behavior data includes: extracting user interaction records from user interaction behavior data and identifying the user identifier and product identifier corresponding to the interaction records; establishing user nodes based on user identifiers and establishing product interest nodes based on product identifiers; establishing the association relationship between user nodes and product interest nodes based on user interaction records and generating user interest association edges; and constructing a user interest propagation graph based on user nodes, product interest nodes, and user interest association edges.

[0015] Preferably, in the group interest propagation analysis module, identifying the user interest propagation path in the user interest propagation graph includes: determining the starting node corresponding to the product interest node in the user interest propagation graph; using the starting node as the starting point, retrieving user interest association edges connected to the starting node in the user interest propagation graph; identifying nodes adjacent to the starting node based on the user interest association edges, and constructing a node propagation sequence; and determining the user interest propagation path based on the node propagation sequence.

[0016] The present invention has the following beneficial effects: 1. In the present invention, by analyzing live video data, live audio data and user interaction behavior data, a live event sequence arranged in chronological order is generated, and the anchor behavior features are extracted based on the live event sequence and an anchor behavior sequence model is established, thereby realizing the identification of the anchor behavior status in the current live stage and solving the problem that existing live recommendation systems are difficult to combine with the anchor's real-time live behavior to recommend products.

[0017] 2. In this invention, a live-stream product evolution path model containing product nodes and path connection edges is constructed by utilizing historical live-stream data, and the transition weights corresponding to the product display order are set for the path connection edges. This achieves structured modeling of the product display order relationship during the live-stream process, solving the problem that existing product recommendation processes cannot reflect the evolutionary pattern of live-stream product display.

[0018] 3. In this invention, by constructing a user interest propagation graph based on user interaction behavior data and identifying user interest propagation paths, the propagation relationship of user interests among user groups is analyzed. Combined with the candidate product set, the target product set is selected, solving the problem that existing live streaming recommendation systems are unable to combine the propagation relationship of user group interests for product recommendation. Attached Figure Description

[0019] Figure 1 is a flowchart of an intelligent product recommendation system based on real-time analysis of live streams proposed in this invention. Detailed Implementation

[0020] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1: In the first embodiment of the present invention, the present invention provides an intelligent product recommendation system based on real-time analysis of live streams, as shown in Figure 1, including the following modules: a live stream data parsing module, used to acquire live stream video data, live stream audio data, and user interaction behavior data, and to parse the live stream video data, live stream audio data, and user interaction behavior data to generate a live stream event sequence arranged in chronological order; further, in the live stream data parsing module, generating the live stream event sequence arranged in chronological order includes: performing image recognition on the live stream video data to extract product display information from the live stream screen and generating product display events; performing speech recognition on the live stream audio data to extract semantic information of the anchor's explanation and generate anchor semantic events; parsing the user interaction behavior data to identify user click behavior, comment behavior, or like behavior and generate user interaction events; and sorting the product display events, anchor semantic events, and user interaction events according to the event occurrence time to obtain the live stream event sequence.

[0022] Specifically, the live streaming data parsing module is used to uniformly acquire and analyze multimodal data generated during the live streaming process, thereby constructing a structured live streaming event sequence. Live streaming video data, live streaming audio data, and user interaction behavior data can be acquired in real time through the live streaming platform's data acquisition interface. For example, video streams can be acquired from the live streaming system, audio streams from the audio processing module, and user interaction behavior records from the platform's log system. All the data from these different sources have timestamp information. The system uses a unified time synchronization mechanism to align the multi-source data in time, enabling different modal data to be uniformly analyzed on the same timeline, thus providing basic data for subsequent event generation and sorting. Live streaming video data first undergoes image recognition processing to identify product display information in the live stream. One possible approach is to use a convolutional neural network-based object detection model to identify video frames, for example, by using a single-stage object detection network to locate and classify product objects in video frames. The live streaming video stream is divided into a continuous sequence of video frames, denoted as... in This represents the time index of the video frame; the object detection model processes each frame... Output the product inspection result set Each element in the detection result represents a identified product target, denoted as... ,in This indicates the product number in the frame; each detection result includes a product category label. Detection confidence and detection timestamp When the detection confidence level is higher than a preset threshold, the recognition result is recorded as valid product display information, and a product display event is generated; the product display event can be represented as an event structure. ,in Indicates product identification, This indicates the time of the product display event; the live audio data is processed through speech recognition to extract semantic information from the host's explanation; the live audio signal first undergoes speech signal preprocessing, including speech framing, noise reduction, and speech enhancement, to obtain stable speech signal features; then, an automatic speech recognition model converts the speech signal into a text sequence, and keyword recognition methods are used to identify product names, product numbers, or product category information in the text; when semantic content related to the product is identified, this semantic information is associated with the corresponding timestamp, thereby generating a host semantic event; the host semantic event can be represented as... ,in Represents semantic content information, This indicates the time of occurrence of a semantic event; user interaction behavior data is generated into user interaction events through a behavior log parsing process; user interaction behavior data typically comes from the interaction logs of the live streaming platform, such as records of user actions like clicking product links, sending comments, or performing like actions; the system identifies the type of user interaction behavior by parsing the interaction logs and extracts the time of occurrence of the behavior and related product identification information; user interaction events can be represented as... ,in Indicates user identifier, Indicates the type of user interaction behavior. This indicates the time when the action occurred. After generating product display events, anchor semantic events, and user interaction events, all events need to be uniformly sorted according to their occurrence time to construct a live event sequence. Let the event set consisting of all events be denoted as . ,in Represents the total number of events, events Includes event type identifier and event timestamp The system sorts the event set according to timestamps, thus obtaining a chronological sequence of events. Its representation is as follows: ;in This indicates the number after sorting by time. An event, and satisfying Through the above sorting process, data events from different sources are uniformly mapped onto the same timeline, thereby forming a structured live event sequence. This event sequence can be used for subsequent extraction of anchor behavior features and analysis of product evolution paths.

[0023] The streamer behavior state recognition module extracts streamer behavior features based on the live event sequence and establishes a streamer behavior sequence model based on historical live data to identify the streamer behavior state corresponding to the current live stage. Further, the extraction of streamer behavior features based on the live event sequence in the streamer behavior state recognition module includes: dividing the live event sequence into time windows to obtain multiple consecutive event segments; statistically analyzing the order in which the streamer displays products and the time interval information of product appearance in each event segment; constructing streamer behavior sequence features based on the product display order information and time interval information; and performing feature encoding processing on the streamer behavior sequence features to generate streamer behavior characteristics.

[0024] Furthermore, in the anchor behavior state recognition module, the anchor behavior sequence model based on historical live streaming data includes: extracting historical live streaming event sequences from historical live streaming data and sorting the historical live streaming event sequences according to the event occurrence time; identifying historical product display events based on the historical live streaming event sequences and constructing the anchor's historical behavior sequence according to the product display order; statistically analyzing the transition relationship between product display orders based on the anchor's historical behavior sequence and calculating the transition probability between adjacent product displays, and constructing the anchor behavior sequence model based on the transition probability.

[0025] Specifically, the anchor behavior state recognition module is used to extract anchor behavior features that reflect the anchor's product display behavior patterns based on the live event sequence, and to build an anchor behavior sequence model using historical live data, thereby identifying the anchor behavior state corresponding to the current live stage. The live event sequence is generated by the aforementioned live data parsing module, and each event includes event type and event occurrence time information. By further analyzing this event sequence, the order and timing of the anchor's product display during the live stream can be obtained. First, the live event sequence is divided into time windows, so that continuous live events are segmented according to time intervals. For example, the live event sequence can be divided into sliding windows according to a fixed time length, and each time window forms an event segment; let the first time window be the first time window. The event segments corresponding to each time window are ,but: ;in Indicates the first event segment One event, This indicates the number of events contained in the event segment; this method allows continuous live streaming behavior to be divided into multiple local behavior segments, enabling the analysis of the anchor's product display behavior at different time stages; product display events are identified in each event segment, and the order information of the anchor's product display and the time interval information of product appearance are statistically analyzed; let the product display sequence identified in a certain event segment be... ,in Indicates the first Each product label is displayed. This indicates the number of products displayed within that time segment; the corresponding product display time is denoted as... ,in Indicates goods The display time of adjacent products; the display time interval between adjacent products can be expressed as: ;in Indicates goods With goods The display time interval between items; based on the order in which the products are displayed. and time interval sequences The system can construct a sequence of anchor behavior features and map product identifiers and time intervals into a unified-dimensional behavioral feature vector using feature encoding to describe the anchor's product display behavior patterns within that time segment. To establish the anchor behavior sequence model, historical live streaming data is used for behavioral pattern statistics. The system extracts historical live streaming event sequences from the historical live streaming data and sorts them according to the event occurrence time to obtain historical live streaming behavior sequences. Subsequently, historical product display events are identified, and the anchor's historical behavior sequence is constructed according to the product display order. Let the historical behavior sequence be... ,in Indicates the first [number] in the historical live broadcast process The product logo displayed next time. This represents the total number of times a product was displayed during historical live streams; based on this, the transition relationships between product display orders are calculated, i.e., product statistics are compiled. Then the products were displayed. Number of times And count the goods Total number of occurrences in the historical sequence Based on the statistical results, the transition probability between product displays is calculated, thus forming a streamer behavior sequence model. The calculation method is as follows: ;in Indicates that the product is being displayed Then display the products The conditional probability, Indicates goods Following the product The number of times it appears, Indicates goods The frequency of occurrence in historical behavior sequences; by statistically analyzing the transition probabilities between different products, a behavior sequence model describing the pattern of the anchor's product display behavior can be formed, which can then be used to identify the anchor's behavior status in the current live broadcast stage based on the current live broadcast event sequence.

[0026] The product evolution path modeling module constructs a live-stream product evolution path model based on historical live-stream data. This model includes product nodes and path connections representing the product display order. Each path connection contains transition weights corresponding to the product display order. Candidate product paths are determined based on the streamer's behavior and product information in the live-stream event sequence. Further, the construction of the live-stream product evolution path model based on historical live-stream data includes: extracting historical product display records from historical live-stream data and sorting these records according to their display time; identifying adjacent product display relationships based on the sorted historical product display records and establishing sequential associations between products; establishing path connections between product nodes based on these sequential associations to construct the product evolution path structure; and generating the live-stream product evolution path model based on the product nodes and path connections.

[0027] Furthermore, in the product evolution path modeling module, the path connection edges include the transition weights corresponding to the product display order, which include: extracting historical product display sequences from historical live data; identifying adjacent product display relationships based on historical product display sequences and counting the number of adjacent displays between product nodes; counting the transition relationships between product nodes based on the number of adjacent displays between product nodes; calculating the transition probability between product nodes based on the transition relationship, and using the transition probability as the transition weight of the path connection edges.

[0028] Furthermore, in the product evolution path modeling module, determining candidate product paths based on the anchor's behavioral state and product information in the live event sequence includes: extracting the product node corresponding to the currently displayed product from the live event sequence; determining the behavioral state identifier corresponding to the current live stage based on the anchor's behavioral state; using the product node as the starting node in the live product evolution path model, and retrieving the path connection edges connected to the product node based on the behavioral state identifier; determining the corresponding product node sequence based on the transition weights of the path connection edges, and identifying the product node sequence as candidate product paths.

[0029] Specifically, the product evolution path modeling module is used to construct a live-stream product evolution path model based on historical live-stream data, and to use this model to describe the evolutionary relationship between the product display order during the live stream. Historical live-stream data can be derived from the live-streaming platform's historical live-stream database, which contains product display records generated during the live stream and corresponding time information. The system first extracts historical product display records from the historical live-stream data and sorts them according to the product display time, thus forming a historical product display sequence arranged in chronological order. Let the historical product display sequence be... ,in Indicates the first The product logo displayed next time. This represents the total number of times a product was displayed during the historical live stream. By sorting the historical product display sequence, the temporal relationship between product displays can be clarified, providing basic data for constructing the product evolution path structure. After obtaining the historical product display sequence, it is necessary to identify the relationship between adjacent product displays to establish the sequential association between products. When a product... The product display follows the product display. At that time, it is considered that the goods With goods There is a sequential relationship between products; the system abstracts each product as a product node and establishes path connection edges between adjacent product nodes, thus forming a product evolution path structure; in this structure, nodes represent product entities, and path connection edges represent the product display order relationship; by statistically analyzing multiple product display sequences in historical live broadcasts, a complete set of product nodes and path connection edges can be gradually constructed, thereby generating a live broadcast product evolution path model; after the path connection edges are established, it is necessary to further calculate the transition weight corresponding to each path connection edge to describe the degree of correlation between product display orders; the system counts the number of times adjacent product nodes are displayed from historical product display sequences, assuming the product set is... ,in Indicates the quantity of different types of goods; when goods The product display follows the product display. When this occurs, it is counted as a sequential transfer and recorded as... Simultaneously, statistics on goods The total number of times a product appears as a preceding item in historical live streams is denoted as . Based on the above statistical results, the transition weight matrix between commodity nodes can be calculated. , where matrix elements Indicates from goods To the goods The transfer weight is calculated as follows: ;in Indicates goods To goods Path connection weights Indicates goods Then the products were displayed. Number of times, Indicates goods The sum of all possible product display times is then used to obtain the complete product transition weight matrix, which is stored in the product evolution path model. After constructing the product evolution path model, candidate product paths need to be determined based on the current live stream status. The system first extracts the currently displayed product information from the live stream event sequence and determines the corresponding product node. Then, based on the streamer behavior status output by the streamer behavior status recognition module, the behavior status identifier corresponding to the current live stream stage is determined. In the product evolution path model, the current product node is used as the starting node, and the path connection edge connected to this node is retrieved based on the behavior status identifier. The system performs path search on the candidate product nodes based on the transition weight of the path connection edge, and a weight-based path expansion method can be used to expand the product nodes layer by layer. In one possible implementation, the candidate path score can be calculated by accumulating the path weight, assuming the candidate path is... The path score can then be expressed as: ;in The path score represents the path to the candidate product. This represents the transfer weight between adjacent product nodes in the path. This represents the path length. By calculating the path score for different paths, possible product node sequences can be obtained, and these product node sequences can be identified as candidate product paths, thus providing path input data for the subsequent candidate product generation module.

[0030] The candidate product generation module extracts product nodes from the path based on the candidate product path to generate a candidate product set. Further, the extraction of product nodes from the path based on the candidate product path in the candidate product generation module includes: identifying the path node sequence in the candidate product path and obtaining the product node set corresponding to the path node sequence; performing node traversal processing on the product node set to extract product nodes from the path; establishing a product node list based on the extracted product nodes; and generating a candidate product set based on the product node list.

[0031] Specifically, the candidate product generation module parses the candidate product paths output by the product evolution path modeling module, extracts product nodes from the paths, and generates a candidate product set based on the extraction results. A candidate product path typically consists of multiple sequentially connected product nodes, each corresponding to a product identifier or product instance identifier. Therefore, the module's operation is based on converting the path structure into a product set structure that can be used for subsequent propagation analysis and target product determination. In one possible implementation, the candidate product path can be represented as a path sequence. ,in Indicates the first in the candidate product path 1 path node This represents the number of path nodes. Each path node is mapped to a product node, allowing the system to parse the corresponding set of product nodes from the path structure. The system first performs path node identification on candidate product paths to obtain a path node sequence. This step can be implemented based on an ordered storage structure of the path; for example, if the candidate product paths are stored in the order of node connections during generation, each node in the path can be directly read as the path node sequence. If the candidate product paths are stored in a graph structure, the node sequence can be extracted in connection order through path traversal. After path node identification is completed, the system obtains the corresponding product nodes based on the mapping relationship between path nodes and product nodes. The product node set includes information such as product identifier, product category, and product location index within the path, ensuring a clear data foundation for subsequent traversal and set generation. After obtaining the product node set, it is necessary to perform node traversal to extract valid product nodes from the path. This traversal can be sequential, visiting each node in the order it appears in the candidate product path and reading the corresponding product information. In one possible implementation, to prevent duplicate product nodes from being counted repeatedly in the candidate product set, the system can perform a uniqueness check on the product nodes during the traversal. Let the product node set be... ,in Indicates the first There are 10 product nodes; if you need to generate a list of deduplicated product nodes, you can define the list of deduplicated nodes as follows: ,in , This indicates the node that is retained after node traversal and uniqueness filtering. There are 10 product nodes; node uniqueness filtering can be based on product identifiers. That is, if the product identifier corresponding to the currently traversed node does not appear in the node list, the node is added to the node list; if the product identifier corresponding to the currently traversed node already exists in the node list, the node is skipped, thus forming a product node list that can represent candidate product paths. After completing the construction of the product node list, a candidate product set is generated based on the product node list. The candidate product set represents all the candidate products associated with the current candidate product path and serves as input for the subsequent group interest propagation analysis module and target product determination module. The candidate product set can be represented as... ,in Indicates the first item in the candidate item set. Each product element This represents the number of candidate products. In one possible implementation, there is a one-to-one correspondence between the product node list and the candidate product set, meaning that each product node in the node list maps to a product element in the candidate product set. If the product node also contains product category information, path location index, or path source identifier, this information can also be written into the candidate product set so that subsequent modules can call it when filtering the propagation path. Through the above process of path node identification, node traversal, and node list construction, the candidate product path can be converted into a structured candidate product set, thus completing the conversion from a path structure to a product set structure.

[0032] The group interest propagation analysis module constructs a user interest propagation graph based on user interaction behavior data and identifies user interest propagation paths within the graph. Further, the construction of the user interest propagation graph based on user interaction behavior data within this module includes: extracting user interaction records from the data and identifying the corresponding user and product identifiers; establishing user nodes based on user identifiers and product interest nodes based on product identifiers; establishing associations between user nodes and product interest nodes based on user interaction records and generating user interest association edges; and constructing the user interest propagation graph based on the user nodes, product interest nodes, and user interest association edges.

[0033] Furthermore, in the group interest propagation analysis module, identifying the user interest propagation path in the user interest propagation graph includes: determining the starting node corresponding to the product interest node in the user interest propagation graph; using the starting node as the starting point, retrieving the user interest association edge connected to the starting node in the user interest propagation graph; identifying the nodes adjacent to the starting node based on the user interest association edge, and constructing the node propagation sequence; and determining the user interest propagation path based on the node propagation sequence.

[0034] Specifically, the group interest propagation analysis module is used to construct a user interest propagation graph based on user interaction behavior data and identify user interest propagation paths within the graph. User interaction behavior data can originate from click logs, comment logs, like logs, or add-to-cart logs recorded by the live streaming platform. Each user interaction behavior record includes at least a user identifier, a product identifier, the behavior type, and the time of occurrence. During runtime, the group interest propagation analysis module first parses the user interaction behavior data, extracts user interaction records, and maps different types of interaction behaviors into a calculable user-product interaction relationship for subsequent graph structure construction and propagation path identification. After extracting user interaction records, the system establishes user nodes based on user identifiers, product interest nodes based on product identifiers, and establishes associations between user nodes and product interest nodes based on the same interaction record. One possible implementation is to represent the user interest propagation graph as a bipartite graph structure. ;in, Represents a set of user nodes. Indicates the first One user node, Indicates the number of user nodes; Represents a set of product interest nodes. Indicates the first Product interest nodes, Indicates the number of interest nodes for the product; Represents the set of edges associated with user interests, where each edge... Represents user node Related to product interest nodes There are interactive relationships between them; when a user clicks, comments, or likes a product, the system establishes a user interest association edge between the corresponding user node and the product interest node, thus obtaining the basic topology of the user interest propagation graph; in order for the user interest propagation graph to represent the degree of association between different interactive behaviors, edge weights can be assigned to the user interest association edges according to user interaction records; one possible way is to calculate the edge weights based on behavior type and behavior frequency; let user node... Related to product interest nodes The edge weight between them is denoted as Then the following formula can be used for calculation; ;in, Represents user node Related to product interest nodes The strength of the correlation between them; Indicates user For products The number of interaction records; Indicates the first The behavior type corresponding to each interaction record; Indicates behavior type The corresponding behavior weights; different behavior types can correspond to different weights, for example, click behavior, comment behavior, and like behavior can be assigned different behavior weights respectively; in this way, the user interest propagation graph not only contains node connection relationships, but also a structured representation of the user's degree of interest in the product; after completing the construction of the user interest propagation graph, it is necessary to identify the user interest propagation path in the graph; the system first determines the corresponding product interest node as the starting node based on the candidate product set or the current analysis target; then, starting from the starting node, it retrieves the user interest association edges connected to the starting node in the user interest propagation graph, and identifies the user nodes adjacent to the starting node; after identifying the adjacent user nodes, it continues to retrieve other product interest nodes associated with them from the user node, thus forming a node propagation sequence that expands along the direction of "product interest node - user node - product interest node"; one possible implementation is to use a breadth-first search algorithm for path retrieval; the breadth-first search algorithm is a graph search algorithm that expands nodes layer by layer according to the hierarchy. Its processing method is to first visit all adjacent nodes of the starting node, and then visit the nodes of the next layer in turn, thus obtaining the propagation sequence starting from the starting node; for any propagation path, it can be represented as ;in, Indicates the starting node. Indicates the first in the propagation path 1 node This indicates the number of nodes contained in the path. When a node sequence satisfies the condition that there are user interest-related edges between adjacent nodes in the graph, the node sequence is identified as a user interest propagation path. After obtaining the user interest propagation path, the system can extract the product interest nodes in the path for use in the subsequent selection process of the target product set. Since the propagation path reflects the correlation and diffusion relationship of product interests among user nodes, the product interest nodes appearing in the path can be used as data input for subsequent product selection and recommendation decisions. Through the above processes of user interaction record extraction, node construction, edge generation, propagation graph establishment, and propagation path identification, the group interest propagation analysis module completes the structured transformation from user interaction behavior data to user interest propagation paths, enabling the group interest relationships in the live broadcast process to participate in the subsequent recommendation process in the form of graph structure and path structure.

[0035] The target product determination module filters corresponding products from the candidate product set based on the user interest propagation path to determine the target product set and generates product recommendation results.

[0036] Specifically, the target product determination module filters corresponding products from the candidate product set based on the user interest propagation path, forming a target product set, thereby generating product recommendation results. The candidate product set originates from the product node set output by the candidate product generation module, while the user interest propagation path originates from the node propagation sequence identified by the group interest propagation analysis module. The target product determination module parses the product interest nodes in the propagation path and matches them with the candidate product set to determine the final range of recommended products. The system first parses the user interest propagation path to extract product interest nodes; the propagation path can be represented as a node sequence. ;in Indicates the first in the path 1 node This indicates the number of nodes in the path; since the propagation path contains both user nodes and product interest nodes, the system extracts the set of product interest nodes by identifying the node type. ;in Indicates the first in the propagation path Product interest nodes, This represents the number of product interest nodes; this set is used to represent the range of products involved in the dissemination of group interests; after obtaining the product interest node set, it is matched and filtered with the candidate product set; let the candidate product set be... ;in Indicates the first One candidate product, This indicates the number of candidate products; the system determines the correspondence between the two through product identifier matching and generates a target product set. The target product set can be represented as: ;in Represents the target set of goods. This represents the product elements in the candidate product set. When a candidate product appears in both the product interest node set and the target product set, it is included in the target product set. After generating the target product set, the system converts it into product recommendation results. The recommendation results can be output in a product list structure and include data fields such as product identifier, product name, or product display information for display on the live streaming interface or for use by the recommendation module. Through the above-mentioned propagation path parsing, product matching and filtering, and set generation process, the process of generating product recommendation results from the candidate product set can be realized, thus completing the operation flow of the target product determination module.

[0037] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An intelligent product recommendation system based on real-time analysis of live streams, characterized in that, It includes the following modules: a live data parsing module, which is used to acquire live video data, live audio data, and user interaction behavior data, and to parse the live video data, live audio data, and user interaction behavior data to generate a live event sequence arranged in chronological order; The anchor behavior status recognition module extracts anchor behavior features based on the live event sequence and builds an anchor behavior sequence model based on historical live data to identify the anchor behavior status corresponding to the current live stage. The product evolution path modeling module constructs a live product evolution path model based on historical live streaming data. The live product evolution path model includes product nodes and path connection edges that represent the product display order relationship. The path connection edges contain the transition weights corresponding to the product display order and determine candidate product paths based on the anchor's behavior state and product information in the live event sequence. The candidate product generation module extracts product nodes from the path based on the candidate product path and generates a candidate product set; the group interest propagation analysis module constructs a user interest propagation graph based on user interaction behavior data and identifies user interest propagation paths in the user interest propagation graph; the target product determination module selects corresponding products from the candidate product set based on the user interest propagation path to determine the target product set and generates product recommendation results.

2. The intelligent product recommendation system based on real-time analysis of live streams according to claim 1, characterized in that, In the live streaming data parsing module, generating a live streaming event sequence arranged in chronological order includes: performing image recognition on the live streaming video data to extract product display information from the live streaming screen and generating product display events; performing speech recognition on the live streaming audio data to extract semantic information from the host's explanation and generating host semantic events; parsing user interaction behavior data to identify user click behavior, comment behavior, or like behavior and generating user interaction events; and sorting the product display events, host semantic events, and user interaction events according to the event occurrence time to obtain the live streaming event sequence.

3. The intelligent product recommendation system based on real-time analysis of live streams according to claim 1, characterized in that, In the anchor behavior state recognition module, the step of extracting anchor behavior features based on the live event sequence includes: dividing the live event sequence into time windows to obtain multiple consecutive event segments; statistically analyzing the order information of the anchor displaying products and the time interval information of the products appearing in each event segment; constructing anchor behavior sequence features based on the product display order information and the time interval information; and performing feature encoding processing on the anchor behavior sequence features to generate anchor behavior features.

4. The intelligent product recommendation system based on real-time analysis of live streams according to claim 1, characterized in that, In the anchor behavior state recognition module, the step of establishing an anchor behavior sequence model based on historical live streaming data includes: extracting historical live streaming event sequences from historical live streaming data and sorting the historical live streaming event sequences according to the event occurrence time; identifying historical product display events based on the historical live streaming event sequences and constructing an anchor historical behavior sequence according to the product display order; statistically analyzing the transition relationship between product display orders based on the anchor historical behavior sequence and calculating the transition probability between adjacent product displays, and constructing an anchor behavior sequence model based on the transition probability.

5. The intelligent product recommendation system based on real-time analysis of live streams according to claim 1, characterized in that, In the product evolution path modeling module, the step of constructing a live-stream product evolution path model based on historical live-stream data includes: extracting historical product display records from historical live-stream data and sorting the historical product display records according to the product display time; identifying adjacent product display relationships based on the sorted historical product display records and establishing sequential association relationships between products; establishing path connection edges between product nodes based on the sequential association relationships to construct a product evolution path structure; and generating a live-stream product evolution path model based on the product nodes and path connection edges.

6. The intelligent product recommendation system based on real-time analysis of live streams according to claim 1, characterized in that, In the product evolution path modeling module, the path connection edges include the transfer weights corresponding to the product display order, which include: extracting historical product display sequences from historical live data; identifying adjacent product display relationships based on historical product display sequences and counting the number of adjacent displays between product nodes; counting the transfer relationships between product nodes based on the number of adjacent displays between product nodes; calculating the transfer probability between product nodes based on the transfer relationship, and using the transfer probability as the transfer weight of the path connection edges.

7. The intelligent product recommendation system based on real-time analysis of live streams according to claim 1, characterized in that, In the product evolution path modeling module, determining candidate product paths based on the anchor's behavior state and product information in the live event sequence includes: extracting the product node corresponding to the currently displayed product from the live event sequence; determining the behavior state identifier corresponding to the current live stage based on the anchor's behavior state; using the product node as the starting node in the live product evolution path model, and retrieving the path connection edge connected to the product node based on the behavior state identifier; determining the corresponding product node sequence based on the transition weight of the path connection edge, and determining the product node sequence as the candidate product path.

8. The intelligent product recommendation system based on real-time analysis of live streams according to claim 1, characterized in that, In the candidate product generation module, the step of extracting product nodes from the path based on the candidate product path includes: identifying the path node sequence in the candidate product path and obtaining the product node set corresponding to the path node sequence; performing node traversal processing on the product node set to extract the product nodes in the path; establishing a product node list based on the extracted product nodes, and generating a candidate product set based on the product node list.

9. The intelligent product recommendation system based on real-time analysis of live streams according to claim 1, characterized in that, In the group interest propagation analysis module, the step of constructing a user interest propagation graph based on user interaction behavior data includes: extracting user interaction records from user interaction behavior data and identifying the user identifier and product identifier corresponding to the interaction records; establishing user nodes based on user identifiers and establishing product interest nodes based on product identifiers; establishing the association relationship between user nodes and product interest nodes based on user interaction records and generating user interest association edges; and constructing a user interest propagation graph based on user nodes, product interest nodes, and user interest association edges.

10. The intelligent product recommendation system based on real-time analysis of live streams according to claim 1, characterized in that, In the group interest propagation analysis module, identifying the user interest propagation path in the user interest propagation graph includes: determining the starting node corresponding to the product interest node in the user interest propagation graph; using the starting node as the starting point, retrieving user interest association edges connected to the starting node in the user interest propagation graph; identifying nodes adjacent to the starting node based on the user interest association edges, and constructing a node propagation sequence; and determining the user interest propagation path based on the node propagation sequence.

Images