A promotion effect intelligent attribution analysis method and system for a live broadcast full link

Abstract

This invention relates to the field of computer data analysis technology, specifically a method and system for intelligent attribution analysis of promotional effects across the entire live streaming journey. It constructs a comprehensive data collection system covering pre-live stream promotion, in-live stream interaction and conversion, and post-live stream repeat purchases and dissemination. Based on a unified user identifier, it achieves cross-platform and cross-device user behavior tracking. A multi-touchpoint attribution model integrating temporal attention mechanisms and causal inference is designed to automatically learn the dynamic contribution weights of different promotional channels and live stream stages in the user conversion path. It quantifies the conversion effects of fine-grained stages such as product explanations, giveaways, and interactive raffles during the live stream. Furthermore, it employs stream computing technology to provide real-time attribution analysis and dynamic optimization suggestions. This invention solves the problems of incomplete attribution paths, rigid models, fragmented data, lack of causality, and insufficient real-time performance in existing technologies, providing comprehensive, accurate, and real-time effect evaluation and decision support for live streaming promotion.

Description

Technical Field

[0001] This invention relates to the field of computer data analysis technology, and in particular to an intelligent attribution analysis method and system for the promotion effect of live streaming across the entire chain. Background Technology

[0002] With the rapid development of the live-streaming e-commerce industry, live-streaming has gradually become a core channel for enterprises to promote products, enhance market influence, and achieve sales conversion. Attribution analysis of promotion effects has become a key link for enterprises to optimize operational strategies and improve return on investment. However, there are still many technical bottlenecks in the field of live-streaming promotion effect attribution analysis. Existing technical solutions are difficult to meet the actual needs of enterprises for full-link, precise, and real-time attribution analysis, which seriously restricts the improvement of live-streaming promotion operation efficiency and the rational allocation of promotion resources.

[0003] Existing attribution analysis methods for live-streaming promotion effectiveness generally suffer from incomplete attribution chains. Most focus solely on the interactive conversion phase during the live stream, neglecting crucial stages such as pre-live stream promotion and post-live stream repeat purchases and word-of-mouth marketing. In practice, user conversion behavior is often a complete chain, from initial exposure to promotional information and generating interest, to making a purchase decision during the live stream, and then to post-live stream repeat purchases and word-of-mouth marketing. Each stage is interconnected and influences the others. Current technologies, by attributing only to a single stage, fail to comprehensively capture the overall value of the promotional campaign, easily leading to misjudgments of promotional effectiveness and wasted promotional resources.

[0004] In the design of attribution models, existing technologies mostly adopt traditional attribution models. These models are inherently rigid and cannot adapt to the complexity and dynamism of user behavior in live streaming scenarios. Traditional models are mostly based on fixed weighting rules for attribution, and cannot automatically learn the dynamic differences in the contributions of different promotion channels and different live streaming stages in the user conversion path. At the same time, they lack causal inference capabilities, making it difficult to distinguish between the true contribution of each promotion touchpoint and the influence of confounding factors. This results in significant biases in the attribution results, failing to provide enterprises with accurate and reliable decision-making basis.

[0005] Current technologies for analyzing live streaming processes mostly remain at the overall level, failing to break down the process into fine-grained segments and thus unable to accurately quantify the conversion contribution of each segment. A live stream comprises multiple content segments, each with significantly different impacts on user conversion. Current technologies cannot identify segments with high and low conversion rates, making it difficult for operators to optimize live stream content in a targeted manner and fully leverage the conversion value of each segment.

[0006] Furthermore, existing attribution analysis methods lack real-time capabilities, often relying on traditional batch processing for data processing and attribution calculations. This fails to enable real-time monitoring and attribution analysis of livestream promotion effectiveness. Livestreaming is a dynamic process where user behavior and promotional results can change at any time. Operations personnel need to adjust promotional strategies and livestream content promptly based on real-time data. Traditional batch processing methods cannot provide timely decision support, often leading to delayed optimization strategies and impacting promotional effectiveness.

[0007] Meanwhile, data from different platforms and stages is scattered across their respective systems, lacking a unified user identification system. This makes it impossible to track user behavior across platforms and devices, resulting in severe data fragmentation. The dispersed data cannot be effectively integrated and correlated, making it difficult to reconstruct the complete user conversion behavior path and significantly hindering comprehensive and accurate attribution analysis.

[0008] This invention aims to solve the technical problems existing in the prior art. To this end, it proposes an intelligent attribution analysis method and system for the promotion effect of live streaming across the entire link. Summary of the Invention

[0009] The purpose of this invention is to provide an intelligent attribution analysis method and system for the promotion effect across the entire live streaming chain, so as to solve the technical problems existing in the prior art.

[0010] This invention provides an intelligent attribution analysis method for promotion effects across the entire live streaming chain, comprising the following steps:

[0011] S1: Build a data collection system covering the entire chain of pre-live broadcast, during live broadcast, and post-live broadcast, and collect user behavior data and promotion channel data at each stage;

[0012] S2: Based on a unified user identifier, the collected multi-source heterogeneous data is correlated and cleaned to generate a complete user conversion behavior path;

[0013] S3: Construct a multi-touchpoint attribution model that integrates temporal attention mechanism and causal inference to calculate the contribution of each promotion touchpoint and live streaming link in the user conversion path;

[0014] S4: Break down the live streaming process into fine-grained steps to accurately quantify the conversion contribution of each step.

[0015] S5: Generates multi-dimensional promotion effect analysis reports based on attribution results and provides real-time suggestions for optimizing promotion strategies.

[0016] Furthermore, the end-to-end data acquisition system in step S1 specifically includes:

[0017] Data collection before live streaming: including exposure, clicks, reposts, and comments from channels such as short video pre-heating, image and text seeding, off-site advertising, and private domain community promotion;

[0018] Data collection during live streams includes behavioral data such as viewership, dwell time, likes, comments, shares, product clicks, adding to cart, placing orders, and payments, as well as data from live stream segments such as host explanations, giveaways, and interactive raffles.

[0019] Post-livestream data collection includes data on user repeat purchases, reviews, sharing and dissemination, refunds and returns, and other follow-up behaviors.

[0020] Furthermore, the data association based on the unified user identifier in step S2 specifically includes:

[0021] Employing multi-dimensional identity recognition technology, it integrates user identity information across different platforms and devices to generate a unique user identifier;

[0022] By stitching together all user behavior data throughout the entire process in chronological order, the complete path from the user's first contact with promotional information to final conversion and subsequent behavior can be reconstructed.

[0023] Clean and filter abnormal data, including invalid clicks, duplicate data, and bot behavior.

[0024] Furthermore, the multi-touchpoint attribution model that integrates temporal attention mechanism and causal inference in step S3 specifically includes:

[0025] Feature extraction is performed on each touchpoint in the user conversion path, including features such as channel type, touchpoint time, touchpoint content, and user interaction behavior;

[0026] A hierarchical temporal attention mechanism is adopted to learn attention weights for different time scales and different touchpoint types, thereby capturing the temporal dependencies and synergistic effects between touchpoints;

[0027] A causal inference method is introduced, and the influence of confounding factors such as user preferences, seasonality, and market environment is eliminated by propensity score matching to calculate the average treatment effect of each touchpoint.

[0028] Using the attention weight as a confidence modulator for the causal effect, the causal effect is weighted and corrected to obtain the comprehensive contribution of each touchpoint to the final conversion.

[0029] Furthermore, the hierarchical temporal attention mechanism extracts multi-level temporal feature weights from the user conversion behavior path during multi-touchpoint attribution calculation through the following steps:

[0030] First, the sequence of touchpoints in the user conversion path is encoded to obtain the initial feature representation of each touchpoint;

[0031] Then, short-term dependencies between adjacent touchpoints are captured through a local attention layer;

[0032] Next, a global attention layer is used to capture the long-term dependencies between all touchpoints throughout the entire path;

[0033] Finally, the importance weights of different types of touchpoints are learned through the type attention layer to obtain the final attention weight of each touchpoint. This weight is used to measure the temporal importance of the touchpoint in the subsequent contribution fusion.

[0034] Furthermore, the fine-grained attribution in step S4 of the live streaming process specifically includes:

[0035] The live stream process is broken down into multiple granular segments according to timeline and content type, including opening warm-up, product explanation, giveaway, interactive lucky draw, limited-time flash sale, and closing summary;

[0036] Each step is timestamped to record the start time, end time, and key events of each step;

[0037] Collect user behavior data at each stage, including the number of people entering the live stream, the number of people leaving, the duration of stay, the number of interactions, the number of product clicks, the number of items added to the cart, and the number of orders placed.

[0038] Based on a multi-touchpoint attribution model, the contribution of each live-streaming segment to the final conversion is calculated, and live-streaming segments with high conversion efficiency and those with low conversion efficiency are identified.

[0039] Furthermore, the real-time attribution and dynamic optimization in step S5 specifically includes:

[0040] Stream computing technology is used to process real-time collected user behavior data to achieve attribution calculations at the second level;

[0041] Monitor the conversion rate of each promotion channel and live streaming session in real time, and issue timely warnings when abnormal results are detected;

[0042] Based on historical and real-time data, machine learning models are used to predict the effectiveness of different promotion strategies and provide dynamic optimization suggestions for live streaming operations.

[0043] Generate multi-dimensional promotion effect analysis reports, including channel effect analysis, process effect analysis, user behavior analysis, ROI analysis, etc.

[0044] This invention also provides an intelligent attribution analysis system for promotion effects across the entire live streaming chain, comprising:

[0045] The end-to-end data acquisition module is used to collect user behavior data and promotion channel data at each stage before, during, and after the live stream.

[0046] The data processing and path generation module is used to correlate and clean the collected multi-source heterogeneous data to generate a complete user conversion behavior path.

[0047] The intelligent attribution calculation module is used to build a multi-touchpoint attribution model that integrates temporal attention mechanism and causal inference to calculate the contribution of each promotion touchpoint and live broadcast segment.

[0048] The fine-grained process analysis module is used to break down the live streaming process into fine-grained stages, enabling precise quantification of the conversion contribution of each stage.

[0049] The real-time analysis and optimization suggestion module is used to generate multi-dimensional promotion effect analysis reports based on attribution results and provide real-time promotion strategy optimization suggestions.

[0050] Furthermore, the end-to-end data acquisition module includes:

[0051] The pre-livestream data collection unit is used to collect data from channels such as short video pre-heating, image and text seeding, off-site advertising, and private domain community promotion.

[0052] The live streaming data collection unit is used to collect user behavior data and live streaming process data in the live streaming room;

[0053] The post-livestream data collection unit is used to collect data on user behavior such as repeat purchases, reviews, sharing and dissemination, refunds and returns.

[0054] Furthermore, the intelligent attribution calculation module includes:

[0055] The feature extraction unit is used to extract features for each touchpoint in the user conversion path;

[0056] The temporal attention computation unit is used to learn the attention weights of different touchpoints using a hierarchical temporal attention mechanism.

[0057] The causal inference unit is used to eliminate the influence of confounding factors through propensity score matching and to calculate the average treatment effect at each touchpoint.

[0058] The contribution calculation unit is used to combine attention weights and causal effects to calculate the overall contribution of each touchpoint to the final conversion.

[0059] Compared with the prior art, the beneficial effects of the present invention are:

[0060] This invention constructs a complete data collection system covering pre-live broadcast promotion and seeding, live broadcast interaction and conversion, and post-live broadcast repeat purchase and dissemination. It can comprehensively evaluate the overall effect of promotion activities and avoid the limitation of focusing on only a single link.

[0061] This invention designs a multi-touchpoint attribution model that integrates temporal attention mechanism and causal inference. It can automatically learn the dynamic contribution weight of different promotion channels and live streaming links in the user conversion path, while eliminating the influence of confounding factors and improving the accuracy and reliability of attribution results.

[0062] This invention breaks down the live streaming process into multiple fine-grained stages, enabling precise quantification of the conversion contribution of each stage. This helps operators identify high-conversion-efficiency and low-conversion-efficiency live streaming stages, providing targeted suggestions for optimizing live streaming content.

[0063] This invention uses stream computing technology to achieve attribution calculation at the second level, enabling real-time monitoring of the conversion effects of various promotion channels and live streaming stages. It provides timely decision support for live streaming operations and helps operators dynamically adjust promotion strategies and live streaming content during the live streaming process.

[0064] This invention enables unified tracking of user behavior across platforms and devices based on a unified user identifier, solving the problem of data fragmentation across different platforms and stages, and providing a data foundation for comprehensive and accurate attribution analysis. Attached Figure Description

[0065] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0066] Figure 1 This is an overall architecture diagram of an intelligent attribution analysis system for the promotion effect of the entire live streaming chain, according to the present invention.

[0067] Figure 2 This is a flowchart of the entire data acquisition and processing process of this invention.

[0068] Figure 3 This is a structural diagram of the multi-touchpoint attribution model that integrates temporal attention mechanism and causal inference in this invention.

[0069] Figure 4 This is a flowchart illustrating the fine-grained attribution process in the live streaming phase of this invention.

[0070] Figure 5 This is a flowchart of the real-time attribution and dynamic optimization system of the present invention. Detailed Implementation

[0071] To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.

[0072] Example 1

[0073] This embodiment provides an intelligent attribution analysis system for promotion effects across the entire live streaming chain, such as... Figure 1 As shown, the system includes a full-link data acquisition module 101, a data processing and path generation module 102, an intelligent attribution calculation module 103, a fine-grained link analysis module 104, and a real-time analysis and optimization suggestion module 105.

[0074] The end-to-end data acquisition module 101 is used to collect user behavior data and promotion channel data at each stage before, during, and after the live broadcast. Specifically, it includes a pre-live broadcast data acquisition unit 1011, a live broadcast data acquisition unit 1012, and a post-live broadcast data acquisition unit 1013.

[0075] The pre-livestream data collection unit 1011 is used to collect data from channels such as short video pre-heating, image and text seeding, off-site advertising, and private domain community promotion. By deploying data collection SDKs or API interfaces on short video platforms, social media platforms, and advertising platforms, it collects user behavior data such as exposure, clicks, reposts, comments, and favorites, as well as information such as the release time, content type, distribution channel, and advertising budget of the promoted content.

[0076] The live streaming data acquisition unit 1012 is used to collect user behavior data and live streaming segment data in the live streaming room. By connecting with the live streaming platform, it collects real-time behavioral data such as users entering the live streaming room, leaving the live streaming room, stay time, likes, comments, shares, clicking on products, adding to cart, placing orders, and making payments. At the same time, it collects live streaming segment data such as the start time, end time, host information, product information, and promotional activity information.

[0077] The post-livestream data collection unit 1013 is used to collect data on user behavior such as repeat purchases, reviews, sharing and dissemination, refunds and returns. Through integration with e-commerce platforms and CRM systems, it collects data on user repurchase behavior, product reviews, sharing and dissemination behavior, refund and return behavior after the livestream ends, as well as user basic information and historical purchase records.

[0078] The data processing and path generation module 102 is used to associate and clean the collected multi-source heterogeneous data to generate a complete user conversion behavior path. Specifically, it includes a user identity association unit 1021, a data cleaning unit 1022, and a behavior path generation unit 1023.

[0079] User identity association unit 1021 employs multi-dimensional identity recognition technology to integrate user identity information across different platforms and devices, generating a unique user identifier. Through the association and matching of various identity information such as mobile phone numbers, email addresses, device IDs, and third-party accounts, user behavior data across different platforms and devices is unified under a single user identifier.

[0080] The data cleaning unit 1022 cleans and filters the collected raw data, removing invalid clicks, duplicate data, robot behavior, and other abnormal data. By setting data quality rules and anomaly detection algorithms, it identifies and filters data that does not meet the requirements, ensuring the accuracy and reliability of the data.

[0081] The behavior path generation unit 1023 stitches together all user behavior data throughout the entire process in chronological order, reconstructing the complete path from the user's first contact with promotional information to final conversion and subsequent behaviors. It generates a conversion path for each user containing all touchpoint and behavioral information, including touchpoint time, touchpoint type, touchpoint content, user interaction behavior, and conversion behavior.

[0082] The intelligent attribution calculation module 103 is used to construct a multi-touchpoint attribution model that integrates temporal attention mechanism and causal inference, and calculates the contribution of each promotion touchpoint and live broadcast link. Specifically, it includes a feature extraction unit 1031, a temporal attention calculation unit 1032, a causal inference unit 1033, and a contribution calculation unit 1034.

[0083] The feature extraction unit 1031 extracts features from each touchpoint in the user conversion path, including channel type, touchpoint time, touchpoint content, and user interaction behavior. It converts the raw data of each touchpoint into feature vectors that the model can process, providing a data foundation for subsequent attribution calculations.

[0084] The temporal attention computation unit 1032 adopts a hierarchical temporal attention mechanism to learn attention weights for different time scales and different touchpoint types, capturing the temporal dependencies and synergistic effects between touchpoints. It captures short-term dependencies between adjacent touchpoints through a local attention layer, captures long-term dependencies between all touchpoints in the entire path through a global attention layer, and learns the importance weights of different types of touchpoints through a type attention layer.

[0085] The causal inference unit 1033 introduces a causal inference method, which eliminates the influence of confounding factors such as user preferences, seasonality, and market environment through propensity score matching, and calculates the average treatment effect of each touchpoint. Users are divided into a treatment group and a control group. By using propensity score matching, control group users with similar characteristics to the treatment group users are found. The conversion differences between the two groups of users are compared to obtain the causal effect of each touchpoint.

[0086] The contribution calculation unit 1034 combines attention weights and causal effects to calculate the overall contribution of each touchpoint to the final conversion. The attention weights obtained from the temporal attention mechanism and the average processing effect obtained from causal inference are weighted and fused to obtain the final contribution of each touchpoint.

[0087] The fine-grained segment analysis module 104 is used to break down the live streaming process into fine-grained segments, enabling precise quantification of the conversion contribution of each segment. Specifically, it includes a segment segmentation unit 1041, a segment data statistics unit 1042, and a segment contribution calculation unit 1043.

[0088] Unit 1041 breaks down the live stream into multiple granular segments based on timeline and content type, including pre-show warm-up, product explanation, giveaways, interactive raffles, flash sales, and closing summary. Through automatic content recognition and manual annotation, the live stream is divided into different segments, and a timestamp is added to each segment.

[0089] The segment data statistics unit 1042 collects user behavior data for each segment, including the number of people entering the live stream, the number of people leaving, the duration of stay, the number of interactions, the number of product clicks, the number of items added to the cart, and the number of orders placed. Based on the timestamp of each segment, the user behavior data that occurred within that segment is filtered and statistically analyzed.

[0090] The stage contribution calculation unit 1043, based on a multi-touchpoint attribution model, calculates the contribution of each live-streaming stage to the final conversion, identifying live-streaming stages with high and low conversion efficiency. Each live-streaming stage is treated as a special touchpoint and input into the multi-touchpoint attribution model to calculate its contribution to the final conversion.

[0091] The real-time analysis and optimization suggestion module 105 is used to generate a multi-dimensional promotion effect analysis report based on the attribution results and provide real-time promotion strategy optimization suggestions. Specifically, it includes a real-time calculation unit 1051, an effect monitoring unit 1052, a strategy optimization unit 1053, and a report generation unit 1054.

[0092] The real-time computing unit 1051 uses stream computing technology to process real-time collected user behavior data, achieving attribution calculations within seconds. Through stream computing frameworks such as Flink and Spark Streaming, it processes and analyzes real-time data streams, updating attribution results in a timely manner.

[0093] The performance monitoring unit 1052 monitors the conversion rates of various promotional channels and live streaming sessions in real time, and issues timely warnings when abnormal results are detected. Performance monitoring metrics and warning thresholds can be set; when a metric exceeds the threshold, a warning is sent to operations personnel via SMS, email, system messages, etc.

[0094] The strategy optimization unit 1053 uses historical and real-time data and a machine learning model to predict the effectiveness of different promotional strategies, providing dynamic optimization suggestions for live streaming operations. Based on attribution results and historical data, it trains a promotional strategy optimization model to predict the conversion rates of different promotional channels, live streaming content, and promotional activities, providing optimization suggestions for operations personnel.

[0095] The report generation unit 1054 generates multi-dimensional promotion effect analysis reports, including channel effect analysis, process effect analysis, user behavior analysis, ROI analysis, etc. The attribution results and analysis data are presented in the form of charts, tables, etc., providing operations personnel with intuitive and comprehensive effect evaluation reports.

[0096] Example 2

[0097] This embodiment provides an intelligent attribution analysis method for promotion effects across the entire live streaming chain, such as... Figure 2 As shown, the method includes the following steps:

[0098] S201: End-to-End Data Collection. Construct a data collection system covering the entire process before, during, and after the live stream, collecting user behavior data and promotional channel data at each stage.

[0099] Pre-livestream data collection includes data on exposure, clicks, shares, and comments from channels such as short video pre-promotion, image and text seeding, external advertising, and private community promotion. For example, when releasing pre-livestream videos on short video platforms, data is collected on video views, likes, comments, shares, and the number of users who clicked the video link to enter the livestream reservation page; when releasing livestream announcements in C-platform communities, data is collected on message views, shares, and reservation numbers within the community.

[0100] Data collection during live streams includes behavioral data such as viewership, dwell time, likes, comments, shares, product clicks, adding to cart, placing orders, and payments, as well as data from various aspects of the live stream, such as the host's explanations, giveaways, and interactive raffles. For example, it collects data on when users enter the live stream, when they leave, the duration of their stay, the content of their comments, the product links they click, the products they add to cart, and the products and amounts they place orders. It also collects data on when the host starts explaining a product, when coupons are distributed, and when interactive raffles are held.

[0101] Post-livestream data collection includes data on user behavior such as repeat purchases, reviews, sharing, refunds, and returns. For example, it collects data on user repeat purchases within 7 days after the livestream ends, reviews of purchased products, sharing of products with friends, and refund / return requests.

[0102] S202: Data Association and Cleaning. Based on a unified user identifier, the collected multi-source heterogeneous data is associated and cleaned to generate a complete user conversion behavior path.

[0103] First, multi-dimensional identity recognition technology is used to integrate user identity information across different platforms and devices to generate a unique user identifier. For example, if a user sees a pre-show video on platform A, registers for a live stream reservation using their mobile phone number, and then places an order in the live stream on platform B, the user's behavioral data on platforms A and B is linked through their mobile phone number to generate a unique user identifier.

[0104] Then, the collected raw data is cleaned and filtered to remove abnormal data such as invalid clicks, duplicate data, and bot behavior. For example, invalid click data from the same user clicking the same link multiple times in a short period of time is filtered out; fake reviews and order data generated by bot accounts are also filtered out.

[0105] Finally, all user behavior data throughout the entire process is pieced together in chronological order to reconstruct the complete path from the user's first exposure to promotional information to final conversion and subsequent behavior. For example, User A's conversion path is as follows: 2023-10-01 10:00 Sees the pre-show video on Platform A → 2023-10-01 10:05 Clicks the video link to enter the live stream reservation page → 2023-10-01 10:06 Reserves a live stream → 2023-10-02 19:00 Receives a live stream reminder → 2023-10-02 19:05 Enters the live stream → 2023-10-02 19:15 Clicks the product link → 2023-10-02 19:16 Adds the product to cart → 2023-10-02 19:20 Places an order and makes payment → 2023-10-05 15:00 Confirms receipt → 2023-10-06 10:00 Leaves a positive review → 2023-10-10 14:00 Repurchases the same product.

[0106] S203: Multi-touchpoint attribution calculation. Construct a multi-touchpoint attribution model that integrates temporal attention mechanism and causal inference to calculate the contribution of each promotion touchpoint and live streaming stage in the user conversion path.

[0107] First, feature extraction is performed on each touchpoint in the user conversion path. For example, for the touchpoint "seeing the pre-show video on platform A", the extracted features include: channel type = short video platform, touchpoint time = 2023-10-01 10:00, content type = product pre-show, user interaction behavior = watching video, video playback duration = 15 seconds, etc.

[0108] Then, a hierarchical temporal attention mechanism is used to learn the attention weights for different touchpoints. The calculation process of the hierarchical temporal attention mechanism is as follows:

[0109] 1. Touchpoint Sequence Encoding: Representing the touchpoint sequence in the user conversion path as... , This represents the overall touchpoint feature sequence vector of the user conversion path. Indicates the first Feature vectors of each contact point This represents the total number of promotional touchpoints in a single user conversion path. The feature vector of each touchpoint is converted into a low-dimensional dense vector through an embedding layer, resulting in an embedding sequence. , This represents the contact feature embedding sequence vector. Indicates the first The low-dimensional feature vectors of each contact point after embedding mapping.

[0110] 2. Local Attention Calculation: A local attention layer is used to capture short-term dependencies between adjacent touchpoints. For each touchpoint... Calculate its relationship with the preceding and following parts. Attention weights for each touchpoint:

[0111]

[0112] in This represents the local attention weight of the i-th touchpoint to the j-th touchpoint within the local window. This indicates that the local attention layer can learn the weight matrix. This represents the half-length parameter of the local attention sliding window. Then, the local attention output is obtained through weighted summation:

[0113]

[0114] This is the feature representation of the i-th touch point after local attention weighting.

[0115] 3. Global Attention Calculation: A global attention layer captures long-term dependencies between all touchpoints throughout the entire path. For each local attention output... Calculate its attention weights with all touchpoints:

[0116]

[0117] in Let be the global attention weight of the i-th touchpoint to the j-th touchpoint. This is the weight matrix of the global attention layer. Then, the global attention output is obtained by weighted summation:

[0118]

[0119] Let be the feature representation of the i-th touch point after global attention weighting.

[0120] 4. Type Attention Calculation: The importance weights of different types of touchpoints are learned through the type attention layer. For each global attention output... According to its contact type Calculate type attention weights:

[0121]

[0122] in The type attention weight for the i-th touchpoint. Let i be the type feature vector corresponding to the type of the i-th contact point. This serves as the channel / link type identifier for the i-th touchpoint. Is a type The learnable vectors are then used to obtain the final attention weights through weighted summation.

[0123]

[0124] The final timing attention weight is the weight for the i-th touchpoint.

[0125] Next, a causal inference method is introduced to eliminate the influence of confounding factors through propensity score matching, and the average treatment effect of each touchpoint is calculated. Propensity score refers to the probability that a user will be exposed to a particular touchpoint given a set of covariates X.

[0126]

[0127] in This indicates that the user has touched the contact point. This indicates that the user did not interact with that touchpoint. A propensity score was estimated using a logistic regression model, and then assigned to each processing group of users ( Match one or more control group users with similar propensity scores. By comparing the conversion probabilities of users in the treatment group and the control group, the average treatment effect of this touchpoint can be obtained:

[0128]

[0129] in This indicates the conversion results for the processing group of users. This represents the conversion results for users in the control group.

[0130] Finally, confidence modulators are applied to the causal effects based on attention weights to calculate the overall contribution:

[0131]

[0132] in It is the first A moderating factor for the timing importance of each contact point in the conversion path. It is the first Average processing effect of each contact point.

[0133] S204: Fine-grained attribution of live streaming segments. This involves breaking down the live streaming process into fine-grained segments to accurately quantify the conversion contribution of each segment.

[0134] First, break down the live stream into multiple granular segments based on timeline and content type. For example, a 2-hour live stream can be divided into: opening warm-up (0-10 minutes), first product introduction (10-30 minutes), first giveaway (30-35 minutes), second product introduction (35-55 minutes), interactive lucky draw (55-65 minutes), third product introduction (65-85 minutes), limited-time flash sale (85-95 minutes), and closing summary (95-120 minutes).

[0135] Then, each stage is timestamped, recording the start time, end time, and key events of each stage. For example, the first product demonstration stage starts at 19:10 and ends at 19:30, with key events including the host introducing product features, showcasing product samples, and answering user questions.

[0136] Next, collect user behavior data for each stage. For example, collect data on the number of people entering the live stream, the number of people leaving the live stream, the average user dwell time, the number of comments, the number of times the product link was clicked, the number of times the product was added to the cart, and the number of times the product was ordered during the first product demonstration stage.

[0137] Finally, based on the multi-touchpoint attribution model, the contribution of each live-streaming segment to the final conversion is calculated. Each live-streaming segment is treated as a specific touchpoint and input into the multi-touchpoint attribution model to calculate its contribution to the final conversion. For example, the contribution of the first product demonstration segment is calculated to be 25%, the contribution of the first benefit distribution segment to be 15%, and the contribution of the second product demonstration segment to be 30%, etc.

[0138] S205: Results Analysis and Optimization Suggestions. Based on the attribution results, a multi-dimensional analysis report of the promotional effect is generated, and real-time suggestions for optimizing promotional strategies are provided.

[0139] Generate multi-dimensional promotional performance analysis reports, including channel performance analysis, stage performance analysis, user behavior analysis, and ROI analysis. For example, the channel performance analysis report shows that the contribution of short video pre-promotion on platform A is 40%, the contribution of community promotion on platform C is 25%, the contribution of off-site advertising is 15%, and the contribution of natural traffic from the live stream is 20%. The stage performance analysis report shows that the conversion efficiency of the second product demonstration stage is the highest, followed by the flash sale stage, and the conversion efficiency of the pre-promotion stage is the lowest.

[0140] We provide real-time optimization suggestions for promotional strategies based on attribution results. For example, based on channel performance analysis, we recommend increasing the budget for short video pre-promotion on platform A and reducing the budget for off-site advertising; based on stage performance analysis, we recommend optimizing the content of the opening pre-promotion stage to increase interactivity and attractiveness and extend user dwell time; based on real-time conversion data, we recommend increasing the explanation time for the second product during the live stream and conducting limited-time flash sales in advance, etc.

[0141] Example 3

[0142] This embodiment provides a specific implementation of a multi-touchpoint attribution model that integrates temporal attention mechanisms and causal inference, such as... Figure 3 As shown, the model includes an input layer 301, an embedding layer 302, a hierarchical temporal attention layer 303, a causal inference layer 304, and an output layer 305.

[0143] The input layer 301 receives user conversion path data. Each user conversion path is represented as a sequence of touchpoints, and each touchpoint includes features such as channel type, touchpoint time, touchpoint content, and user interaction behavior.

[0144] Embedding layer 302 transforms the discrete features of each touchpoint into a low-dimensional dense vector. For categorical features, such as channel type and content type, an embedding matrix is ​​used for transformation; for numerical features, such as touchpoint time and video playback duration, they are directly input after standardization. The output of the embedding layer is a touchpoint embedding sequence. ,in The length of the contact sequence.

[0145] The hierarchical temporal attention layer 303 includes a local attention layer 3031, a global attention layer 3032, and a type attention layer 3033.

[0146] The local attention layer 3031 employs a sliding window mechanism to capture short-term dependencies between adjacent touchpoints. For each touchpoint... Taking it as the center, take the front and back. Each touchpoint forms a local window. The attention weights of each touchpoint within the window are calculated, and then the local attention output is obtained by weighted summation. Local attention layers can capture user behavior patterns in a short period of time. For example, if a user places an order quickly after clicking a product link, there is a strong short-term dependency between these two touchpoints.

[0147] The global attention layer 3032 captures the long-term dependencies between all touchpoints throughout the path. For each local attention output... Calculate its attention weights with all touchpoints, and then obtain the global attention output by weighted summation. The global attention layer can capture user behavior patterns over a longer period of time. For example, if a user sees a promotional video a week ago and then places an order during a live stream, there is a long-term dependency between these two touchpoints.

[0148] The Type Attention Layer 3033 learns the importance weights of different types of touchpoints. Different types of touchpoints have varying degrees of impact on conversion; for example, a product click touchpoint has a greater impact on conversion than a like touchpoint. Based on the type of touchpoint, the Type Attention Layer outputs global attention for each touchpoint. Different weights are assigned to obtain the final attention weights. .

[0149] The causal inference layer 304 includes a propensity score estimation unit 3041, a matching unit 3042, and an average treatment effect calculation unit 3043.

[0150] The propensity score estimation unit 3041 uses a logistic regression model to estimate a user's propensity score when they come into contact with a certain touchpoint. Input features include basic user information (such as age, gender, and region), historical behavior information (such as historical purchase frequency and historical spending amount), and market environment information (such as seasonality and holidays), among other confounding factors.

[0151] Matching unit 3042 uses the nearest neighbor matching method to match one or more control group users with similar propensity scores for each treatment group user. The matching criterion is that the difference in propensity scores is less than a preset threshold, ensuring that the treatment group and control group users have similar distributions in terms of confounding factors.

[0152] The average treatment effect calculation unit 3043 compares the conversion probability differences between the treatment group and the control group, and calculates the average treatment effect at each touchpoint. To improve the accuracy of the estimation, a bootstrap method can be used to perform multiple samplings and calculate the confidence interval for the average treatment effect.

[0153] Attention weights obtained by combining output layer 305 with hierarchical temporal attention layer The average treatment effect obtained from the causal inference layer Calculate the overall contribution of each touchpoint to the final conversion. Then, the contribution of all touchpoints is normalized so that the sum of the contributions of all touchpoints is 1.

[0154] The multi-touchpoint attribution model in this embodiment considers both temporal dependencies and causal effects between touchpoints, enabling a more accurate assessment of the contribution of each promotional touchpoint and live-streaming stage to the final conversion. Compared to traditional attribution models, this model offers higher accuracy and reliability, providing more scientific decision support for live-streaming promotion.

[0155] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0156] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for intelligent attribution analysis of promotion effects across the entire live streaming chain, characterized in that, Includes the following steps: S1: Build a data collection system covering the entire chain of pre-live broadcast, during live broadcast, and post-live broadcast, and collect user behavior data and promotion channel data at each stage; S2: Based on a unified user identifier, the collected multi-source heterogeneous data is correlated and cleaned to generate a complete user conversion behavior path; S3: Construct a multi-touchpoint attribution model that integrates temporal attention mechanism and causal inference to calculate the contribution of each promotion touchpoint and live streaming link in the user conversion path; S4: Break down the live streaming process into fine-grained steps to accurately quantify the conversion contribution of each step. S5: Generates multi-dimensional promotion effect analysis reports based on attribution results and provides real-time suggestions for optimizing promotion strategies.

2. The method according to claim 1, characterized in that, The end-to-end data acquisition system in step S1 specifically includes: Data collection before live streaming: including exposure, clicks, reposts, and comments from channels such as short video pre-heating, image and text seeding, off-site advertising, and private domain community promotion; Data collection during live streams includes behavioral data such as viewership, dwell time, likes, comments, shares, product clicks, adding to cart, placing orders, and payments, as well as data from live stream segments such as host explanations, giveaways, and interactive raffles. Post-livestream data collection includes data on user repeat purchases, reviews, sharing and dissemination, refunds and returns, and other follow-up behaviors.

3. The method according to claim 1, characterized in that, The data association based on the unified user identifier in step S2 specifically includes: Employing multi-dimensional identity recognition technology, it integrates user identity information across different platforms and devices to generate a unique user identifier; By stitching together all user behavior data throughout the entire process in chronological order, the complete path from the user's first contact with promotional information to final conversion and subsequent behavior can be reconstructed. Clean and filter abnormal data, including invalid clicks, duplicate data, and bot behavior.

4. The method according to claim 1, characterized in that, The multi-touchpoint attribution model that integrates temporal attention mechanism and causal inference in step S3 specifically includes: Feature extraction is performed on each touchpoint in the user conversion path, including features such as channel type, touchpoint time, touchpoint content, and user interaction behavior; A hierarchical temporal attention mechanism is adopted to learn attention weights for different time scales and different touchpoint types, thereby capturing the temporal dependencies and synergistic effects between touchpoints. A causal inference method is introduced, and the influence of confounding factors such as user preferences, seasonality, and market environment is eliminated by propensity score matching to calculate the average treatment effect of each touchpoint. Using the attention weight as a confidence modulator for the causal effect, the causal effect is weighted and corrected to obtain the comprehensive contribution of each touchpoint to the final conversion.

5. The method according to claim 4, characterized in that, The hierarchical temporal attention mechanism extracts multi-level temporal feature weights from the user conversion behavior path during multi-touchpoint attribution calculation through the following steps: First, the sequence of touchpoints in the user conversion path is encoded to obtain the initial feature representation of each touchpoint; Then, short-term dependencies between adjacent touchpoints are captured through a local attention layer; Next, a global attention layer is used to capture the long-term dependencies between all touchpoints throughout the entire path; Finally, the importance weights of different types of touchpoints are learned through the type attention layer to obtain the final attention weight of each touchpoint. This weight is used to measure the temporal importance of the touchpoint in the subsequent contribution fusion.

6. The method according to claim 1, characterized in that, The fine-grained attribution in step S4 of the live streaming process specifically includes: The live stream process is broken down into several granular segments according to timeline and content type, including opening warm-up, product explanation, giveaway, interactive lucky draw, limited-time flash sale, and closing summary; Each step is timestamped to record the start time, end time, and key events of each step; Collect user behavior data at each stage, including the number of people entering the live stream, the number of people leaving, the duration of stay, the number of interactions, the number of product clicks, the number of items added to the cart, and the number of orders placed. Based on a multi-touchpoint attribution model, the contribution of each live-streaming segment to the final conversion is calculated, and live-streaming segments with high conversion efficiency and those with low conversion efficiency are identified.

7. The method according to claim 1, characterized in that, The real-time attribution and dynamic optimization in step S5 specifically include: Stream computing technology is used to process real-time collected user behavior data to achieve attribution calculations at the second level; Monitor the conversion rate of each promotion channel and live streaming session in real time, and issue timely warnings when abnormal results are detected; Based on historical and real-time data, machine learning models are used to predict the effectiveness of different promotion strategies and provide dynamic optimization suggestions for live streaming operations. Generate multi-dimensional promotion effect analysis reports, including channel effect analysis, process effect analysis, user behavior analysis, ROI analysis, etc.

8. A smart attribution analysis system for promotion effects across the entire live streaming chain, characterized in that, include: The end-to-end data acquisition module is used to collect user behavior data and promotion channel data at each stage before, during, and after the live stream. The data processing and path generation module is used to correlate and clean the collected multi-source heterogeneous data to generate a complete user conversion behavior path. The intelligent attribution calculation module is used to build a multi-touchpoint attribution model that integrates temporal attention mechanism and causal inference to calculate the contribution of each promotion touchpoint and live broadcast link. The fine-grained segment analysis module is used to break down the live streaming process into fine-grained segments, enabling precise quantification of the conversion contribution of each live streaming segment. The real-time analysis and optimization suggestion module is used to generate multi-dimensional promotion effect analysis reports based on attribution results and provide real-time promotion strategy optimization suggestions.

9. The system according to claim 8, characterized in that, The end-to-end data acquisition module includes: The pre-livestream data collection unit is used to collect data from channels such as short video pre-heating, image and text seeding, off-site advertising, and private domain community promotion. The live streaming data collection unit is used to collect user behavior data and live streaming process data in the live streaming room; The post-livestream data collection unit is used to collect data on user behavior such as repeat purchases, reviews, sharing and dissemination, refunds and returns.

10. The system according to claim 8, characterized in that, The intelligent attribution calculation module includes: The feature extraction unit is used to extract features for each touchpoint in the user conversion path; The temporal attention computation unit is used to learn the attention weights of different touchpoints using a hierarchical temporal attention mechanism. The causal inference unit is used to eliminate the influence of confounding factors through propensity score matching and to calculate the average treatment effect at each touchpoint. The contribution calculation unit combines attention weights and causal effects to calculate the overall contribution of each touchpoint to the final conversion.

Images