A certainty conversion identification method and system based on large-screen advertising

Abstract

The present application relates to the technical field of data analysis, in particular to a certainty conversion identification method and system based on large-screen advertising, comprising collecting viewing behavior characteristics of a large screen, inputting the viewing behavior characteristics into a viewing mode classification model, clustering and matching in a historical viewing database, and outputting the current viewing mode category; based on the viewing mode category, screening out historical viewing consumption data with a matching mode category from the historical viewing database; obtaining advertising material characteristics of a target advertisement and promoted goods, based on the historical viewing consumption data, performing bidirectional demand correlation analysis on the promoted goods to generate a demand matching score; based on the historical viewing consumption data, performing material tendency analysis on the advertising material characteristics to obtain a material tendency score; and comprehensively matching the demand matching score and the material tendency score to identify the conversion probability of the target advertisement under the current viewing behavior characteristics.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, specifically to a deterministic conversion identification method and system based on large-screen advertising. Background Technology

[0002] In the fields of digital marketing and e-commerce, in-screen purchases on large-screen terminals (such as smart TVs and OTT devices) have become an important business channel in the home setting. Existing internet advertising systems typically rely on static attribute data such as device IDs or registered accounts as the basis for decision-making when allocating business resources and predicting conversion rates. However, this approach based on static administrative data faces significant challenges in data processing and decision-making within dynamic business marketing scenarios.

[0003] A discrepancy exists between audience identification data and actual consumer decision-makers, leading to inefficient allocation of marketing resources. Large-screen terminals, as shared household assets, have diverse consumer groups. Existing data processing methods struggle to dynamically analyze real-time interaction data to identify current consumer decision-makers, relying instead on historical account profiles. This results in advertising systems frequently pushing high-value commercial information to non-target audiences—for example, pushing high-priced items to children without purchasing power. This leads to significant waste of advertising inventory resources and inefficient use of marketing budgets, hindering the effective reach and conversion rates of commercial messages.

[0004] How to use data processing technology to accurately predict business intentions through behavioral data clustering while protecting privacy is a technical problem that urgently needs to be solved in the field of digital marketing.

[0005] To address this, a deterministic conversion identification method and system based on large-screen advertising is proposed. Summary of the Invention

[0006] The purpose of this invention is to provide a deterministic conversion identification method and system based on large-screen advertising, which integrates demand matching score and material preference score to identify the conversion probability of target advertisement under the current viewing behavior characteristics.

[0007] To achieve the above objectives, the present invention provides a deterministic conversion identification method based on large-screen advertising, comprising:

[0008] The system collects viewing behavior characteristics on the large screen, including program switching sequence data, program content attribute data, remote control interaction command sequence, and the current time point.

[0009] The movie-watching behavior characteristics are input into the movie-watching mode classification model, clustered and matched from the historical movie-watching database, and the current movie-watching mode category is output; the movie-watching mode category represents the aggregation state of the current audience's behavioral interaction habits and content consumption tendencies in front of the big screen.

[0010] Based on the movie viewing mode category, historical movie viewing consumption data that matches the mode category is filtered from the historical movie viewing database; the historical movie viewing consumption data includes movie viewing behavior characteristics and converted product information under the same movie viewing mode.

[0011] The process involves: acquiring the advertising creative characteristics and promoted products of the target advertisement; conducting a two-way demand correlation analysis on the promoted products based on historical movie-watching consumption data to generate a demand matching score; the two-way demand correlation analysis includes complementarity analysis and substitution analysis; and conducting creative bias analysis on the advertising creative characteristics based on historical movie-watching consumption data to obtain a creative bias score.

[0012] By combining demand matching score and creative preference score, the conversion probability of the target ad is identified under the current movie-watching behavior characteristics.

[0013] The movie-watching behavior characteristics are used to quantify the current interaction state, and their specific data dimensions are defined as follows:

[0014] The program switching timing data includes the channel switching rate, non-linear jump frequency, and average continuous playback duration of a single content within a preset time sliding window, which is used to distinguish between browsing mode and immersive mode.

[0015] The program content attribute data includes the metadata feature vector of the currently playing streaming media content, covering the content's primary category tags, secondary theme tags, and target audience profile tags.

[0016] The remote control interaction command sequence includes remote control button trigger density, button response delay, and function key usage preference;

[0017] The current time point includes the timestamp feature of the current system time and the time period attribute corresponding to the timestamp.

[0018] The construction process of the movie viewing mode classification model includes:

[0019] Collect historical viewing databases from large-screen terminals and extract historical viewing feature vectors; use unsupervised clustering algorithms to perform cluster analysis on the historical viewing feature vectors, dividing the data space into high-density clusters, with each cluster center representing a viewing mode category; the viewing mode categories include, but are not limited to: high-frequency switching exploration mode, high-engagement viewing mode for children, background sound accompaniment mode, and deep immersive viewing mode.

[0020] Calculate the Euclidean distance between the currently collected movie-watching behavior features and each cluster center; select the category label corresponding to the nearest cluster center as the movie-watching mode category; at the same time, calculate the membership degree of the current movie-watching behavior feature to the cluster center. If the membership degree is lower than the preset outlier threshold, mark the current state as a novel mode and trigger the incremental clustering update mechanism.

[0021] The process of obtaining historical movie-watching consumption data based on movie-watching mode categories includes:

[0022] A movie viewing mode index table is established, which records the movie viewing mode category to which each effective conversion behavior in the historical movie viewing database belongs.

[0023] Based on the current movie viewing mode category, a reverse query is performed through the movie viewing mode index table to recall all historical conversion records marked with the same category;

[0024] The historical conversion records of the recall are filtered again based on scene similarity. The program content attribute data of the current viewing behavior characteristics are extracted as the query vector, and the cosine similarity with the historical program content attributes in the recall records is calculated. Records with a cosine similarity greater than the preset association threshold are retained to form historical viewing consumption data.

[0025] The historical movie-watching consumption data includes the list of purchased items, purchase amount, and conversion time interval data generated by users when watching the same movie-watching mode and watching similar content styles.

[0026] A two-way demand correlation analysis is conducted on the promoted products, specifically including:

[0027] Complementarity analysis process: Traverse the converted products in historical movie-watching consumption data and query the co-occurrence probability between the target promotion product and the converted products; the co-occurrence probability represents the tendency for the two products to be consumed together under the same movie-watching mode; calculate the complementarity score based on the weighted co-occurrence probability;

[0028] Substitutability analysis process: Identify converted products in historical movie-watching consumption data that belong to the same category as the target promoted product; obtain the most recent purchase time of the converted products, and calculate the substitutability score at the current moment by combining it with the average repurchase cycle;

[0029] The final demand matching score is obtained by linearly superimposing the complementarity score and the substitution score, and the superposition weight is determined by the conversion preference characteristics of the current movie viewing mode category.

[0030] Conduct material preference analysis on advertising material characteristics, specifically including:

[0031] By analyzing the historical movie-watching consumption data and recording the corresponding historical advertising materials based on conversion rates, we can construct high-response material features for the current movie-watching mode.

[0032] Multimodal feature analysis is performed on the target advertisement to extract visual and auditory elements; the multimodal features of the target advertisement are matched with the features of the high-response material to calculate the matching degree, and the weighted matching degree of the visual and auditory elements is output, which is the material tendency score.

[0033] The process of calculating the conversion probability under the current viewing mode category includes:

[0034] Construct a conversion probability prediction model and input it after normalizing the demand matching score and material preference score;

[0035] A pattern confidence factor is introduced, which is determined by the distance membership degree and is used to characterize the impact of the accuracy of the current movie-watching pattern determination on conversion prediction.

[0036] Based on the normalized demand matching score, material preference score, and pattern confidence factor, the conversion probability under the current movie-watching behavior characteristics is predicted.

[0037] A deterministic conversion recognition system based on large-screen advertising includes:

[0038] The behavior acquisition module collects viewing behavior characteristics on the large screen, including program switching sequence data, program content attribute data, remote control interaction command sequence, and the current time point.

[0039] The pattern recognition module inputs the movie-watching behavior features into the movie-watching pattern classification model, clusters and matches them from the historical movie-watching database, and outputs the current movie-watching pattern category; the movie-watching pattern category represents the aggregation state of the current audience's behavioral interaction habits and content consumption tendencies in front of the big screen.

[0040] The data matching module, based on the movie viewing mode category, filters historical movie viewing consumption data that matches the mode category from the historical movie viewing database; the historical movie viewing consumption data includes movie viewing behavior characteristics and converted product information under the same movie viewing mode.

[0041] The feature analysis module acquires the advertising material features and promoted products of the target advertisement; based on historical movie-watching consumption data, it performs two-way demand correlation analysis on the promoted products and generates a demand matching score; the two-way demand correlation analysis includes complementarity analysis and substitution analysis; based on historical movie-watching consumption data, it performs material preference analysis on the advertising material features and obtains material preference score;

[0042] The conversion prediction module combines demand matching score and creative preference score to identify the conversion probability of the target ad under the current movie-watching behavior characteristics.

[0043] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0044] 1. This invention uses unsupervised clustering of time-series behavioral data, interaction commands, and content attributes to map fuzzy family accounts to specific movie-watching consumption patterns, overcoming the targeting bias problem caused by traditional methods relying on static account identity data. It accurately identifies behavioral states with specific consumption tendencies, effectively cleans up invalid audience traffic without violating user privacy, and ensures that advertising decisions are based on current real consumption scenarios, thereby significantly improving the accuracy of advertising resources in reaching the target audience and the return on investment of marketing activities.

[0045] 2. This invention differs from traditional shallow recommendations based solely on content interests. It innovatively constructs a bidirectional demand association model that includes complementarity analysis and substitution analysis. By processing the co-occurrence probability and product lifecycle characteristics in historical transaction data, it transforms abstract purchase intentions into calculable quantitative scores, which can intelligently identify potential associated purchase opportunities and repurchase or replacement opportunities, greatly improving the confidence of the conversion rate prediction model.

[0046] 3. This invention quantifies the certainty of pattern recognition into a confidence factor and incorporates it as a weight into the calculation path of the conversion rate prediction model. When user behavior features deviate from the typical cluster center, the confidence factor is automatically used to perform penalized smoothing on the initial score output by the model. This mechanism suppresses overfitting caused by noise in the input features or outlier behaviors at the algorithmic level, ensuring that the final output conversion probability value can truly reflect the consistency between the behavioral pattern matching degree and the user's actual intention, thus achieving mathematical calibration of the conversion probability. Attached Figure Description

[0047] Figure 1 This is a flowchart illustrating a deterministic conversion recognition method based on large-screen advertising according to the present invention.

[0048] Figure 2 This is a data logic diagram of a deterministic conversion recognition method based on large-screen advertising according to the present invention;

[0049] Figure 3 This is a schematic diagram of the structure of a deterministic conversion recognition system based on large-screen advertising according to the present invention. Detailed Implementation

[0050] The technical solutions of 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.

[0051] Example 1:

[0052] This invention proposes a deterministic conversion recognition method based on large-screen advertising. The process of the method is as follows: Figure 1 As shown, the data logic of the method is as follows: Figure 2 As shown, it includes:

[0053] The system collects viewing behavior characteristics on the large screen, including program switching sequence data, program content attribute data, remote control interaction command sequence, and the current time point.

[0054] The movie-watching behavior characteristics are input into the movie-watching mode classification model, clustered and matched from the historical movie-watching database, and the current movie-watching mode category is output; the movie-watching mode category represents the aggregation state of the current audience's behavioral interaction habits and content consumption tendencies in front of the big screen.

[0055] Based on the movie viewing mode category, historical movie viewing consumption data that matches the mode category is filtered from the historical movie viewing database; the historical movie viewing consumption data includes movie viewing behavior characteristics and converted product information under the same movie viewing mode.

[0056] The process involves: acquiring the advertising creative characteristics and promoted products of the target advertisement; conducting a two-way demand correlation analysis on the promoted products based on historical movie-watching consumption data to generate a demand matching score; the two-way demand correlation analysis includes complementarity analysis and substitution analysis; and conducting creative bias analysis on the advertising creative characteristics based on historical movie-watching consumption data to obtain a creative bias score.

[0057] By combining demand matching score and creative preference score, the conversion probability of the target ad is identified under the current movie-watching behavior characteristics.

[0058] The movie-watching behavior characteristics are used to quantify the current interaction state, and their specific data dimensions are defined as follows:

[0059] The program switching timing data includes the channel switching rate, non-linear jump frequency, and average continuous playback duration of a single content within a preset time sliding window, which is used to distinguish between browsing mode and immersive mode.

[0060] The program content attribute data includes the metadata feature vector of the currently playing streaming media content, covering the content's primary category tags, secondary theme tags, and target audience profile tags.

[0061] The remote control interaction command sequence includes remote control button trigger density, button response delay, and function key usage preference;

[0062] The current time point includes the timestamp feature of the current system time and the time period attribute corresponding to the timestamp.

[0063] Preferably, the data acquisition process includes:

[0064] Collect program switching timing data: Set a time sliding window, such as five minutes, and count the number of times the user switches channels within the time sliding window to calculate the switching rate; at the same time, identify non-linear jumping behavior, that is, the frequency of users randomly jumping to select channels instead of following the list order, and record the average continuous playback duration of a single content.

[0065] Collect program content attribute data: Read the metadata of the currently playing streaming media, extract its primary category, secondary theme, and the preset target audience tags for the content; among them, the primary category includes movies, TV series, variety shows, animation, etc., the secondary theme includes science fiction, suspense, family, etc., and the preset target audience tags, such as suitable for 3-6 years old, suitable for all ages, etc.

[0066] Collect remote control interaction command sequence: Obtain remote control signal through infrared or Bluetooth interface, count the key trigger density per unit time, record key response delay, that is, the time difference between the appearance of the screen and the user pressing the confirmation key; and count the usage preferences of function keys, such as volume keys and menu keys.

[0067] Collect the current time point: Obtain the current system timestamp and map it to a time period attribute, such as weekday evening rush hour or weekend early morning.

[0068] Non-linear switching frequency refers to the number of times a user changes channels or content drastically within a short period. High-frequency non-linear switching usually indicates that the user is browsing aimlessly, resulting in a lower conversion rate; while low frequency represents immersive viewing. Key response latency refers to the user's reaction speed to screen information and is an important indicator of user focus. The lower the latency, the more focused the user's attention.

[0069] Furthermore, the specific implementation of data acquisition includes:

[0070] Time-sliding window management mechanism: Maintain a fixed-size circular buffer, the size of which corresponds to the maximum number of events within the time-sliding window (e.g., 5 minutes). Whenever a new channel switching event occurs, record the event's timestamp, source channel ID, and target channel ID. If the buffer is full, a first-in, first-out (FIFO) strategy is used to remove the oldest event, and the new event is inserted at the end of the buffer.

[0071] Channel switching rate definition and calculation: Channel switching refers to the event where a user presses the channel up / down buttons on the remote control or directly enters the channel number, causing the playback content to switch from one channel to another. All valid switching events are counted within a time sliding window, and the switching rate is obtained by dividing the total number of events by the window duration. To eliminate noise caused by accidental remote control triggering, the system introduces an anti-shake mechanism: if the time interval between two switching events is less than a preset anti-shake threshold, such as 100 milliseconds, it is considered a single accidental operation and counted as only one switching event.

[0072] Construction and normalization of high-dimensional feature vectors: During a single viewing sampling period, data from four dimensions are collected simultaneously in the following order. After the sampling period ends, each dimension undergoes independent minimum-maximum normalization. After normalization, the feature vector concatenated from the four normalized dimensional data is input into the viewing pattern classification model.

[0073] Measurement and processing of remote control button response latency: When outputting interactive prompts to the user interface, such as displaying pop-up menus or highlighted options, the rendering completion timestamp of the prompt is recorded simultaneously. When a remote control button signal is detected, the arrival timestamp of the button event is recorded; the difference between the two timestamps is the button response latency. If the user does not perform any button operation within 30 seconds after the interactive prompt appears, the interaction is recorded as a no-response state, and the corresponding latency value is set to 30 seconds. For multiple consecutive interactive prompts, the average of all valid latency values ​​is calculated as the button response latency index for this sampling period.

[0074] This invention achieves precise quantification of user states in front of a large screen by constructing a multi-dimensional viewing behavior feature system. It introduces time-series switching data and remote control interaction commands to effectively distinguish between different states such as idle playback, casual browsing, and in-depth viewing. For example, while playing an animated film, frequent volume adjustments and random skipping might indicate that a parent is adjusting the device rather than a child watching. This fine-grained feature capture eliminates a large amount of false traffic noise, providing a solid data foundation for subsequent pattern recognition, thereby significantly improving the accuracy of audience identification and solving the technical challenge of unclear actual viewers behind home devices.

[0075] The construction process of the movie viewing mode classification model includes:

[0076] Collect historical viewing databases from large-screen terminals and extract historical viewing feature vectors; use unsupervised clustering algorithms to perform cluster analysis on the historical viewing feature vectors, dividing the data space into high-density clusters, with each cluster center representing a viewing mode category; the viewing mode categories include, but are not limited to: high-frequency switching exploration mode, high-engagement viewing mode for children, background sound accompaniment mode, and deep immersive viewing mode.

[0077] Calculate the Euclidean distance between the currently collected movie-watching behavior features and each cluster center; select the category label corresponding to the nearest cluster center as the movie-watching mode category; at the same time, calculate the membership degree of the current movie-watching behavior feature to the cluster center. If the membership degree is lower than the preset outlier threshold, mark the current state as a novel mode and trigger the incremental clustering update mechanism.

[0078] The movie viewing mode classification model is constructed using an unsupervised learning algorithm, specifically a density-based clustering algorithm.

[0079] Training process: Export historical viewing feature vectors of large-screen terminals in batches; then, use clustering algorithm to calculate the distribution of these vectors in multidimensional space, and group points that are close to each other into high-density clusters; the algorithm iterates until the cluster center is stable, and each cluster center is defined as a standard viewing mode category, such as high-frequency switching exploration mode or deep immersive viewing mode.

[0080] Recognition process: During online processing, calculate the Euclidean distance between the currently acquired feature vector and all trained cluster centers; select the label represented by the nearest cluster center as the current mode;

[0081] Incremental update mechanism: Simultaneously calculate the membership degree of the current feature to the cluster center. If the membership degree is lower than the preset outlier threshold, it means that the current behavior does not belong to any known pattern. The state is marked as a novel pattern and stored in the cache. When the cache accumulates a certain amount of data, it triggers incremental clustering update to generate new cluster centers.

[0082] Euclidean distance is the straight-line distance between two points in a multidimensional space, used to mathematically measure the similarity between the current behavior and typical historical behaviors; distance membership represents the probability or degree that the current sample belongs to a certain cluster, used to judge the reliability of pattern recognition and prevent errors caused by forced classification;

[0083] High-frequency switching exploration mode: This mode corresponds to the state where users don't know what to watch and are frequently switching channels to find something, i.e., browsing mode; users quickly browse the channel list without focusing on any particular content, and their attention is in a wandering state.

[0084] Technical features include:

[0085] Program switching sequence: The channel switching rate is extremely high, and the non-linear jumping frequency is high, that is, the channel selection jumps across a large range.

[0086] Playback duration: The average continuous playback duration of a single piece of content is very short, such as a few seconds to a dozen seconds;

[0087] Remote control interaction: The button trigger density is extremely high, and users frequently operate the up / down or back buttons;

[0088] At this point, although the user's attention is on the screen, they are in the filtering stage and have low tolerance for interstitial ads, making them likely to swipe away directly. This is suitable for extremely short, visually impactful brand exposure ads, rather than long story ads.

[0089] High-engagement viewing mode for children: This corresponds to the state of children watching cartoons or specific children's programs; viewers are usually very focused, highly engaged, and rarely switch once they have selected content.

[0090] Technical characteristics:

[0091] Program content attributes: In the metadata characteristics of the content being played, the primary category is animation or children's, and the target audience profile tag is clearly defined as suitable for young children aged 3-6.

[0092] Program switching sequence: The switching rate is extremely low, the continuous playback time of a single content is very long, and the viewing experience is immersive.

[0093] Remote control interaction: button triggering density is extremely low.

[0094] The main decision-makers may not be in front of the screen, or they may just be accompanying the child; it is suitable to target products that children drive to buy, such as toys and snacks, or educational products.

[0095] Background noise mode: This mode is for users who only listen and don't watch; the TV acts as background noise in the home. Users might be cleaning, using their phones, or cooking, while the TV plays news, music shows, or long dramas.

[0096] Technical characteristics:

[0097] Remote control interaction: A key indicator is extremely high button response latency. For example, when interactive pop-ups or advertisements appear on the screen, the user takes a long time to respond.

[0098] Program content attributes: Usually, it is content with rich auditory information, such as news, music variety shows, etc.

[0099] Program switching sequence: Long periods of inactivity, long playback duration, and no fine-grained operations such as pause or fast forward.

[0100] Visual advertising is almost ineffective. Identify this pattern and in subsequent material preference analysis, prioritize matching advertising materials with high scores in auditory elements to reach the target audience through sound.

[0101] Deep Immersive Viewing Mode: Corresponds to the state of watching TV programs at a deep intensity;

[0102] Technical characteristics:

[0103] Current time point: usually occurs within the timestamp of weekday evening rush hour or weekend;

[0104] Program content attributes: The content is usually variety shows, movies or sports events suitable for all ages;

[0105] Remote control interaction: Low button response latency, user focus, and may be accompanied by moderate preference for using function keys such as volume adjustment or pause.

[0106] Furthermore, the Mini-Batch K-Means algorithm is used for incremental updates. When the number of samples in the storage area reaches a threshold, such as 1000, the data in the storage area is iteratively updated using the old cluster center as the initial centroid. A supplementary elimination mechanism is added, introducing a time decay factor for the cluster center. If a cluster does not absorb new samples within M consecutive update cycles, it is hidden from the model.

[0107] This invention employs an unsupervised clustering modeling strategy combined with incremental updates, significantly enhancing the system's adaptability to complex family scenarios. Traditional supervised learning requires a large amount of manually labeled data and struggles to handle emerging behavioral patterns. This solution automatically discovers the inherent structure within the data through cluster analysis, not only identifying viewing patterns but also uncovering implicit habits. More importantly, the incremental clustering update mechanism enables the model to evolve, automatically learning new behavioral paradigms as family structure changes or holiday schedules adjust, ensuring robust recognition over long-term operation and preventing model failure due to data drift.

[0108] The process of obtaining historical movie-watching consumption data based on movie-watching mode categories includes:

[0109] A movie viewing mode index table is established, which records the movie viewing mode category to which each effective conversion behavior in the historical movie viewing database belongs.

[0110] Based on the current movie viewing mode category, a reverse query is performed through the movie viewing mode index table to recall all historical conversion records marked with the same category;

[0111] The historical conversion records of the recall are filtered again based on scene similarity. The program content attribute data of the current viewing behavior characteristics are extracted as the query vector, and the cosine similarity with the historical program content attributes in the recall records is calculated. Records with a cosine similarity greater than the preset association threshold are retained to form historical viewing consumption data.

[0112] The historical movie-watching consumption data includes the list of purchased items, purchase amount, and conversion time interval data generated by users when watching the same movie-watching mode and watching similar content styles.

[0113] Index Query: Maintain a movie viewing mode index table, with the mode category as the key and the historical valid conversion record ID as the value. First, based on the currently determined mode category, retrieve all historical purchase records that occurred under that mode through the index table.

[0114] Secondary filtering: To improve accuracy, the system extracts attribute data of the currently playing content, such as science fiction or action, as a query vector. Then, it calculates the cosine similarity between this vector and the attributes of the content played at that time in the recall record.

[0115] Data restructuring: A correlation threshold is set, and only records with a cosine similarity greater than the threshold are retained. This set of records constitutes the final historical movie-watching consumption data, which includes the purchase list, amount, and conversion time interval at that time.

[0116] Cosine similarity: It evaluates the similarity between two vectors by calculating the cosine of the angle between them. In text or tag feature comparison, cosine similarity reflects the consistency of the content theme better than distance and is not affected by the magnitude of the value.

[0117] This invention constructs a high-quality reference dataset through a two-stage filtering mechanism: initial pattern screening and refined content screening. Relying solely on viewing patterns may be too broad; for example, an immersive viewing mode could involve watching a tragedy or a comedy, with completely different consumer mindsets. This solution, based on locking onto behavioral states, further introduces cosine similarity matching of content attributes, ensuring that historical data highly fits the current scenario in both state and context dimensions. It eliminates a large amount of interfering data that, while sharing the same pattern, has irrelevant content context, allowing subsequent demand predictions to be based on highly similar parallel spatiotemporal histories. This significantly improves the reference value of the prediction results and the accuracy of conversion rate estimation.

[0118] A two-way demand correlation analysis is conducted on the promoted products, specifically including:

[0119] Complementarity analysis process: Traverse the converted products in historical movie-watching consumption data and query the co-occurrence probability between the target promotion product and the converted products; the co-occurrence probability represents the tendency for the two products to be consumed together under the same movie-watching mode; calculate the complementarity score based on the weighted co-occurrence probability;

[0120] Substitutability analysis process: Identify converted products in historical movie-watching consumption data that belong to the same category as the target promoted product; obtain the most recent purchase time of the converted products, and calculate the substitutability score at the current moment by combining it with the average repurchase cycle;

[0121] The final demand matching score is obtained by linearly superimposing the complementarity score and the substitution score, and the superposition weight is determined by the conversion preference characteristics of the current movie viewing mode category.

[0122] Complementarity scoring method: Iterate through the filtered historical movie-watching consumption data; for the target promotional product, count the number of times it and historically converted products appear simultaneously in the same order or within the same time window. Calculate the co-occurrence probability, which is the probability of both occurring simultaneously divided by the probability of each occurring individually. Use this probability as a weight to calculate the complementarity score.

[0123] Substitutability analysis algorithm: Identify whether historical data contains purchased items belonging to the same category as the target product, such as detergents. If so, obtain the most recent purchase time and retrieve the average repurchase cycle for that category. Calculate the ratio of the time elapsed since the last purchase to the average cycle to generate a substitutability score.

[0124] Linear overlay: Based on the characteristics of the current viewing mode, such as the exploration mode being more inclined towards new products and the daily mode being more inclined towards repeat purchases, the complementary score and the replacement score are added together.

[0125] Co-occurrence probability: The statistical probability of two events occurring simultaneously. Based on association rule mining, it is used to discover potential related demands.

[0126] Average repurchase cycle: The average time interval between users repeatedly purchasing the same type of product. Based on the product life cycle theory, it is used to predict users' existing consumption and repurchase opportunities.

[0127] This solution, through complementarity analysis, can keenly capture potential cross-category pairing needs; through substitution analysis, it can accurately predict the repurchase window for daily necessities. This in-depth analysis, combining product correlation and lifecycle, ensures that advertising is no longer blindly targeted, but rather precisely pushed to users at crucial moments when they need to upgrade or make bundled purchases, greatly improving users' purchase intentions and the success rate of ad conversion.

[0128] Furthermore, in conducting bidirectional demand correlation analysis for promoted products, a spatiotemporal product knowledge graph can be used for deep reasoning. This spatiotemporal product knowledge graph includes product nodes, time nodes, and scene nodes. The complementarity analysis no longer relies solely on co-occurrence probability but instead executes the following graph reasoning steps: mapping the current viewing mode category to scene nodes in the graph, and mapping the current time point to a time node; using scene nodes and time nodes as starting anchors, performing multi-hop path walks in the graph to find target promoted products with a strong "scene-function-time" correlation path with historically converted product nodes; calculating the semantic correlation strength of each path, and comprehensively generating a deep demand matching score that includes explicit complementarity and implicit causal relationships, replacing the complementarity score based on statistical probability.

[0129] Model architecture: A knowledge graph is pre-built; nodes include not only products, but also scenarios, such as watching a game late at night, or family activities on weekends, as well as time and seasons, such as summer and winter.

[0130] Reasoning process: Assuming the current identification is a late-night sports event viewing mode, and the time is summer, traditional statistics may only know that beer is bought; but graph reasoning can follow the path: watching the game late at night -> fatigue -> need to be refreshed -> energy drinks / coffee, or summer -> more mosquitoes -> mosquito repellent.

[0131] Logical details: Path search is performed using graph neural networks to discover implicit needs that are not obvious in statistical data but are logically valid; the DeepWalk algorithm based on random walks is used to calculate node embeddings, and the association strength is determined by calculating cosine similarity; or the P-PageRank algorithm is used to calculate the transition probability from scene nodes to product nodes.

[0132] This solution overcomes the limitations of traditional collaborative filtering, which is based on historical statistics and can only recommend items that others have purchased. By introducing a spatiotemporal knowledge graph, it gains logical reasoning capabilities, enabling it to discover implicit causal needs across product categories. This is particularly important for new product promotion, as it can deduce potential needs based on scenario logic without accumulating large amounts of historical data, greatly expanding the breadth and surprise factor of ad recommendations.

[0133] Conduct material preference analysis on advertising material characteristics, specifically including:

[0134] By analyzing the historical movie-watching consumption data and recording the corresponding historical advertising materials based on conversion rates, we can construct high-response material features for the current movie-watching mode.

[0135] Multimodal feature analysis is performed on the target advertisement to extract visual and auditory elements; the multimodal features of the target advertisement are matched with the features of the high-response material to calculate the matching degree, and the weighted matching degree of the visual and auditory elements is output, which is the material tendency score.

[0136] High-response feature construction: This involves analyzing the characteristics of ad creatives that successfully drive conversions based on historical movie-watching data. For example, it was found that ads with warm colors and fast-paced music had the highest conversion rates under this model, thus constructing a high-response creative feature model.

[0137] Multimodal analysis and matching: Perform multimodal feature analysis on the target advertisement. Visually, extract the color square map and brightness distribution of keyframes; auditorily, extract the dynamic range of volume and BPM (beats per minute).

[0138] Matching score calculation: The feature vector of the target ad is compared with the feature vector of the high-response creative, and their weighted matching score is calculated. If the feature distributions of the two have a high degree of overlap, a higher creative propensity score is output.

[0139] Multimodal feature analysis: a technology that simultaneously processes multiple media information such as images and sounds. The effectiveness of advertising depends not only on the product, but also on whether the audiovisual impact resonates with the audience's current emotional state.

[0140] Further, the specific implementation of multimodal feature parsing and matching:

[0141] Visual feature extraction and quantification: The following visual feature extraction operations are performed on the target advertisement. First, if the advertisement is in video format, the system samples keyframes at fixed intervals, for a total of N keyframes; if the advertisement is a static image, then N=1.

[0142] For each keyframe, perform the following extraction steps:

[0143] Color Features: Keyframes are converted from RGB color space to HSV color space; histograms (bin count = 32) are calculated for the H (hue), S (saturation), and V (luminance) channels respectively, resulting in a 3×32=96-dimensional color feature vector. The calculation method is as follows: For each channel, the pixel value range [0,255] is uniformly divided into 32 bins, the number of pixels in each bin is counted, and the histogram distribution of that channel is obtained after normalization.

[0144] Spatial features: Extract SIFT (Scale Invariant Feature Transform) keypoints from keyframes, retain the 100 keypoints with the highest scores, and construct a 100-dimensional spatial feature vector.

[0145] Brightness distribution characteristics: Calculate the histogram (bin number = 32) of the grayscale image of the keyframe, and calculate its first moment (average brightness) and second moment (contrast) to form a 34-dimensional brightness feature vector. Average the visual feature vectors of all N keyframes to obtain the final visual features.

[0146] Auditory feature extraction and quantification: The audio track of the advertisement is processed as follows: First, the audio data is subjected to a short-time Fourier transform, with a frame length of 2048 samples and a frame shift of 512 samples, to obtain a time-spectrum graph; then, the following feature extraction is performed on the time-spectrum graph:

[0147] Rhythmic characteristics: The energy envelope of the time-frequency spectrum is calculated. By summing the energy of each frequency component, the autocorrelation function is used to identify the dominant rhythmic period and convert it to BPM (beats per minute). If the identification result is within the range of 30-300 BPM, the BPM value is retained; otherwise, it is set to 0, indicating that the rhythm is not obvious.

[0148] Volume dynamic range: The maximum energy value E_max and minimum energy value E_min of all frames in the spectrum during calculation. The dynamic range is defined as: 20*log10(E_max / E_min)dB.

[0149] Frequency distribution: The frequency range is divided into four intervals: low frequency (0-250Hz), low-mid frequency (250-1000Hz), mid-high frequency (1000-4000Hz), and high frequency (4000-22050Hz). The proportion of each interval in the total energy is calculated to obtain a 4-dimensional frequency distribution vector.

[0150] Human voice proportion: Apply a pre-trained human voice detection model to the audio spectrum, calculate the proportion of time frames containing obvious human voice features to the total number of frames, and obtain a scalar value [0,1]; concatenate the above features to obtain the final auditory features.

[0151] Construction of High-Response Creative Features: During the model training phase, all advertising records that generated conversions in historical movie-watching consumption data were statistically analyzed. A conversion event was defined as a user clicking or purchasing within 30 seconds of an ad being displayed. Ads corresponding to all conversion events were selected, and their conversion rates (number of conversion events / total ad impressions) were calculated. Ads were sorted from highest to lowest conversion rate, and the top 20% were selected as the high-response ad set. Visual and auditory features were extracted from all ads in this set. The mean and covariance of all visual feature vectors were calculated to obtain the distribution parameters of the high-response visual features; similarly, the distribution parameters of the high-response auditory features were obtained. These distribution parameters represent the ad feature style preferred by users in this movie-watching mode.

[0152] Matching score calculation and creative preference score generation: In the prediction phase, the similarity between the visual features of the ad to be delivered and the distribution of high-response visual features is calculated using the Mahalanobis distance formula; similarly, auditory similarity is calculated. The two similarities are weighted and combined to obtain the final creative preference score, where the weights are set according to the user's sensory sensitivity for that viewing mode. For example, in the deeply immersive viewing mode, the user is fully engaged, and the audiovisual experience is balanced, so both are set to 0.5.

[0153] This invention addresses the low conversion rate caused by a mismatch between advertising creatives and the viewing atmosphere through creative preference analysis. Users' acceptance of advertising formats varies significantly across different viewing modes. For example, in a late-night immersive viewing mode, noisy, hard-sell advertising may annoy users or even cause them to shut down their devices; while in a more engaging viewing mode for children, lively audiovisual elements are more likely to attract attention. This solution quantifies the audiovisual characteristics of historically successful creatives and matches them with upcoming ads, ensuring that the ads not only match the product but also the creative style. This refined creative matching mechanism effectively reduces user resistance to advertising, improving ad completion rates and ultimately, conversion rates.

[0154] Furthermore, the analysis of advertising material characteristics to determine material preference also includes visual saliency heatmap alignment; extracting the last frame of the currently playing content, generating a residual attention heatmap using a visual saliency detection algorithm to determine the area where the viewer's gaze lingers on the screen; simultaneously extracting the coordinates of the core visual elements of the target advertisement from the first frame, calculating the spatial overlap rate between the core visual elements and the residual attention heatmap; adding the spatial overlap rate as a correction factor to the material preference score; if the overlap rate is lower than a preset threshold, i.e., the key information of the advertisement appears in the viewer's blind spot, then the material preference score is reduced through the correction factor.

[0155] Furthermore, to evaluate the visual continuity of advertising creatives at the moment the current viewing ends, this embodiment also introduces an attention alignment mechanism based on visual saliency. This mechanism corrects the creative score by generating a heatmap and calculating the spatial overlap rate. The specific implementation steps are as follows:

[0156] Step 1: Generate a residual attention heatmap of the content;

[0157] As the current playback content is about to end, the last frame is captured as the source image; to adapt to the real-time computing needs of large-screen terminals, a calculation method based on frequency domain residuals is adopted, specifically including:

[0158] Image preprocessing: Convert the source image to a grayscale image and downsample it to a preset low-resolution size to reduce the amount of subsequent computation;

[0159] Spectrum Transformation: Performing a two-dimensional discrete Fourier transform on the preprocessed grayscale image converts the image data from the spatial domain to the frequency domain, thereby obtaining the amplitude spectrum and phase spectrum;

[0160] Extracting the frequency domain residual: First, calculate the natural logarithm of the amplitude spectrum to construct the logarithmic spectrum; second, smooth the logarithmic spectrum using a mean filter (e.g., a 3×3 matrix) to obtain the average spectrum; finally, subtract the average spectrum from the logarithmic spectrum, and the difference is the frequency domain residual. This step is based on the statistical regularities of natural images, namely that background textures usually conform to a specific spectral distribution, while salient targets exhibit abrupt changes in the spectrum.

[0161] Reconstructing the heatmap: By combining the frequency domain residual with the original phase spectrum, and calculating the inverse discrete Fourier transform of the sum of the exponentially increased frequency domain residual and the phase spectrum, and taking the square of its modulus, a preliminary saliency map in the spatial domain is reconstructed. This preliminary saliency map is then Gaussian smoothed and normalized, mapped to the 0-1 interval, ultimately yielding a content residual attention heatmap. Regions with higher values ​​in this heatmap represent a higher probability of viewer attention remaining at the moment of image transition.

[0162] Step 2: Determine the coordinates of the core visual elements of the advertisement;

[0163] For the first frame of the target advertisement to be evaluated, obtain the coordinate regions of its core visual elements, such as brand logo and product body. These coordinate regions can be obtained by reading the metadata preset in the advertisement material, or by applying the above-mentioned saliency detection algorithm to the first frame of the advertisement to extract connected regions with saliency values ​​greater than a preset threshold, and using the bounding rectangle of the connected region as the coordinate region.

[0164] Step 3: Calculate the spatial overlap rate;

[0165] The coordinate regions of the identified core visual elements of the advertisement are mapped onto the aforementioned residual attention heatmap. Spatial overlap rate is defined as the ratio of the sum of the salience values ​​of all pixels located within the coordinate region of the core visual element in the heatmap to the sum of the salience values ​​of all pixels in the entire heatmap. This metric quantifies whether the key information of the advertisement appears in the high-probability area where the user's current gaze lingers.

[0166] Step 4: Adjust the material preference score based on the overlap rate;

[0167] A blind spot threshold is preset, for example, 0.15; the calculated spatial overlap rate is compared with the blind spot threshold.

[0168] If the spatial overlap rate is greater than or equal to the blind spot threshold, the original material tendency score remains unchanged, or a positive weighting with a preset range is given.

[0169] If the spatial overlap rate is less than the blind spot threshold, the key information in the advertisement is determined to be in the viewer's blind spot, triggering a penalty mechanism to lower the creative's bias score. At this point, the score is updated using a correction factor.

[0170] The calculation logic of the correction factor is as follows: First, calculate the ratio of spatial overlap rate to blind spot threshold, and subtract the ratio from 1 to obtain the normalization deviation; then multiply the normalization deviation by the preset penalty intensity coefficient; finally, subtract the above product from 1 to obtain the final correction factor. The original material tendency score multiplied by the correction factor is the updated final score.

[0171] For example, a user has just finished watching a movie, and the end credits are scrolling on the left side of the screen while the right side is black. At this moment, the user's attention is highly focused on the left side. A heatmap is generated by calculating the salience of the movie's ending sequence. If the upcoming advertisement is for a car brand and the car appears on the right side of the screen, the user may not immediately see the main focus of the ad.

[0172] This solution utilizes computer vision technology to address the blind spot problem during scene transitions. Existing advertising only focuses on content matching, neglecting the user's physiological visual inertia; users' gaze does not immediately shift during scene changes. By using computational visual persistence heatmaps, the degree of overlap between the user's attention focus and the core area of ​​the advertisement is identified, ensuring accurate identification of conversion probability.

[0173] The process of calculating the conversion probability under the current viewing mode category includes:

[0174] Construct a conversion probability prediction model and input it after normalizing the demand matching score and material preference score;

[0175] A pattern confidence factor is introduced, which is determined by the distance membership degree and is used to characterize the impact of the accuracy of the current movie-watching pattern determination on conversion prediction.

[0176] Based on the normalized demand matching score, material preference score, and pattern confidence factor, the conversion probability under the current movie-watching behavior characteristics is predicted.

[0177] Normalization: The calculated demand matching score and material preference score are mapped to the range of zero to one to eliminate the difference in dimensions.

[0178] Introducing a confidence factor: Extract the calculated distance membership degree and convert it into a pattern confidence factor. If the current behavioral feature is far from the cluster center, it indicates that the pattern judgment is inaccurate, and the confidence factor will be a small value.

[0179] Prediction Calculation: A neural network model is constructed, using the normalized score as the feature input and the pattern confidence factor as the adjustment weight. The model outputs a probability value between 0 and 1, which is the final conversion probability. Specifically, the conversion probability prediction model uses a multilayer perceptron network, including an input layer, two hidden layers, and an output layer; the number of nodes in the hidden layers are 128 and 64, respectively, and the ReLU activation function is used; the output layer uses the Sigmoid activation function; and the mean squared error loss function is used for model training.

[0180] Pattern confidence factor: A coefficient used to penalize or reward prediction results; based on Bayesian inference logic, when the prior conditions (pattern recognition) are uncertain, the confidence of the posterior probability (transformation prediction) should be reduced for risk control.

[0181] This invention quantifies the certainty of pattern recognition into a confidence factor and incorporates it as a weight into the calculation path of the conversion rate prediction model. When user behavior features deviate from the typical cluster center, the confidence factor is automatically used to perform penalized smoothing on the initial score output by the model. This mechanism suppresses overfitting caused by noise in the input features or outlier behaviors at the algorithmic level, ensuring that the final output conversion probability value can truly reflect the consistency between the behavioral pattern matching degree and the user's actual intention, thus achieving mathematical calibration of the conversion probability.

[0182] Furthermore, the process of calculating the transformation probability also incorporates the feedback loop of online reinforcement learning;

[0183] A state-action-reward model is constructed, where viewing mode category is the environment state, advertising strategy is the action, and real-time user interaction feedback is the reward signal. The reward signal includes not only positive rewards for clicks or purchases but also negative penalties for quick escape behavior. Quick escape behavior is defined as shutting down, changing channels, or muting within a specified time after ad display. The Q-Learning algorithm is used to update the Q-value table of various ad creatives under different viewing modes in real time. When predicting conversion probability, the Q-value corresponding to the target ad in the current mode is queried and used as a historical lesson correction coefficient, which is then added to the final probability. If the Q-value is negatively high, the ad's delivery is forcibly blocked.

[0184] Scenario Description: An ad with a spooky sound effect was displayed in background noise mode, and the user immediately pressed the mute button; the action of mute was captured and defined as a high-weight negative reward. The Q-value table was updated through reinforcement learning, and "background noise mode + spooky sound effect = extremely poor policy" was marked.

[0185] Strategy Adjustment: The next time the same pattern is encountered, the conversion probability will be automatically reduced and the campaign will be rejected due to the extremely low Q value, so as to avoid harassing users again.

[0186] This solution introduces a reinforcement learning mechanism with self-correcting capabilities. Traditional conversion models are typically trained offline, resulting in delayed updates. This invention, through online reinforcement learning, can capture users' aversion in real time and transform it into the system's muscle memory; this not only optimizes conversion rates but, more importantly, protects the long-term user experience of large-screen media, preventing traffic loss due to excessive interruptions.

[0187] Example 2:

[0188] This invention proposes a deterministic conversion recognition system based on large-screen advertising, the system structure of which is as follows: Figure 3 As shown, it includes:

[0189] The behavior acquisition module collects viewing behavior characteristics on the large screen, including program switching sequence data, program content attribute data, remote control interaction command sequence, and the current time point.

[0190] The current physical time is 23:30 on Friday night. The user turned on the TV screen at 23:00, and after 3 minutes of browsing and switching programs, is currently watching a live football match from 23:03 to the present. The behavior collection module has collected specific viewing behavior feature vectors containing the following dimensions of data:

[0191] Current time point: Collect the system timestamp and map it to a time period attribute [Friday, 23:30-24:00, the eve of the late-night holiday].

[0192] Program switching timing data: Operation records within the sliding window over the past 10 minutes were read. The channel switching rate was found to be 0 times / minute, meaning the user did not change channels. The non-linear jump frequency was 0, and the average continuous playback duration of a single content session was 27 minutes. This data set exhibits typical characteristics of low-frequency switching and high continuity.

[0193] Program content attribute data: Read the metadata of the currently playing streaming media; extract the primary category as sports, the secondary theme as football / live broadcast, and the target audience profile tag as 25-45 years old / sports fans.

[0194] Remote control interaction command sequence: Statistics on button data in the past 10 minutes show that the button trigger density is extremely low, with the volume only being adjusted once at the beginning. The button response delay is high, there is no pop-up interaction, no data was measured, and the default is passive viewing mode.

[0195] The pattern recognition module inputs the movie-watching behavior features into the movie-watching pattern classification model, clusters and matches them from the historical movie-watching database, and outputs the current movie-watching pattern category; the movie-watching pattern category represents the aggregation state of the current audience's behavioral interaction habits and content consumption tendencies in front of the big screen.

[0196] The pattern recognition module receives the above-mentioned movie-watching behavior features and performs the following steps:

[0197] Clustering matching: The movie-watching behavior features are mapped to a pre-trained high-dimensional feature space, and the Euclidean or Mahalanobis distance between them and the center of each cluster is calculated; the system finds that the vector is closest to the center of the deep immersive movie-watching mode cluster.

[0198] Confidence calculation: The distance membership degree from the current vector to the cluster center is calculated. The result is 0.85, which is higher than the outlier threshold of 0.4, indicating that the current user behavior is very typical and the pattern recognition result is reliable. The system outputs the current pattern category as a deep immersive movie-watching mode and converts the distance membership degree into a pattern confidence factor, which is then passed to subsequent modules.

[0199] Preferably, in this embodiment, the DBSCAN algorithm is used for unsupervised clustering; this algorithm identifies high-density clusters based on the concept of density and has good recognition ability for clusters of arbitrary shapes.

[0200] The core parameters of DBSCAN are:

[0201] The neighborhood radius eps, defined as 0.5, is obtained by F1-Score verification on historical test data in the normalized feature space. This value achieves the optimal balance between clustering effect and novel sample recognition ability.

[0202] The minimum number of points, min_pts, is defined as 50, indicating that a valid cluster must contain at least 50 samples; clusters of points below this threshold are marked as noise or outliers.

[0203] Distance membership is calculated using the following formula: membership = exp(-d² / σ²), where d is the distance from the current sample to the cluster center, and σ is the scaling parameter of the covariance matrix of the cluster.

[0204] Training set preparation and initialization: The system periodically retrains the entire model. Before each retraining, 100,000 viewing feature vectors are randomly sampled from the historical viewing database. To ensure the representativeness and balance of the samples, the system weights the sampled data according to their frequency of occurrence, ensuring that long-tail patterns, such as viewing behavior during special periods like the New Year, are fully represented. The sampled feature vectors are input into the DBSCAN algorithm, which iterates until the cluster centers stabilize, i.e., the distance between the cluster centers in two adjacent iterations is less than 10^-5, resulting in several initial clusters. Each cluster is labeled as a viewing mode category.

[0205] The data matching module, based on the viewing mode category, filters historical viewing consumption data that matches the mode category from the historical viewing database; the historical viewing consumption data includes viewing behavior characteristics and converted product information under the same viewing mode; the data matching module performs two-level filtering based on the viewing mode category:

[0206] First-level index recall: Query the movie viewing mode index table to recall the set of valid conversion record IDs marked as deep immersive movie viewing mode in the historical movie viewing database, totaling 5,000 records; the historical movie viewing database can include all user data of the entire region to ensure the volume of data.

[0207] Secondary content filtering: Extract the attribute vector of the current content, including sports, football, competition category, etc.; calculate the cosine similarity between this vector and the attributes of the content played at that time in 5000 historical records.

[0208] Data generation: The association threshold was set to 0.7, and records of watching tragic movies or documentaries in the same mode were removed. Only 1,200 high-similarity records of watching sports events or competitive games were retained. This subset constituted the historical movie-watching consumption data on which this calculation was based.

[0209] The feature analysis module acquires the ad creative features and promoted products of the target advertisement; based on historical movie-watching consumption data, it performs two-way demand correlation analysis on the promoted products and generates a demand matching score; the two-way demand correlation analysis includes complementarity analysis and substitution analysis; based on historical movie-watching consumption data, it performs creative preference analysis on the ad creative features and obtains a creative preference score.

[0210] Identify the target advertisement and determine that the promoted product is craft beer; first, search the user's latest purchase record and find that 15 minutes before the current time, at 23:15, the account linked to the large screen just placed an order for spicy peanuts (a snack); based on the above information, perform a two-way demand correlation analysis:

[0211] Complementarity Analysis: The system traverses the historical movie-watching consumption database, filtering out all order sets marked as deeply immersive viewing modes and containing snack items. Within these order sets, the number of joint orders also including beer / beverages is counted. Calculations show that in the scenario of watching sports late at night, 82% of users who purchased snacks also purchased beer. The system maps this co-occurrence probability of 0.82 to a complementarity score.

[0212] Cycle Prediction: Historical data revealed that this user group has an average repurchase cycle of 7 days for "beer / beverages". The history shows the last purchase was 6 days ago. The replacement score is calculated by dividing the last purchase interval by the average repurchase cycle.

[0213] Linear superposition: According to the preset weight table, in the deep immersion mode, users are more likely to be driven by physiological needs rather than simply repeat purchases; therefore, the complementarity weight is set to 0.7 and the substitutability weight is set to 0.3, and the demand matching score is calculated.

[0214] Ad creative preference analysis: Statistical analysis of historical movie-watching consumption data revealed that high-converting creatives in this model often feature fast-paced and highly saturated characteristics; based on similarity identification between the target ad and high-converting creatives in this model, the initial creative score was 0.8;

[0215] Visual saliency alignment: Capture the last frame of a live football match, such as the scoreboard displayed in the upper left corner of the screen; use a frequency domain residual algorithm to process the image, perform a Fourier transform, extract the logarithmic spectrum residual, perform an inverse Fourier transform to reconstruct the image, and generate a content residual attention heatmap; the heatmap shows that the user's gaze is highly focused on the score area in the upper left corner.

[0216] Coordinate extraction and overlap calculation were performed to extract the core visual elements of the first frame of the target advertisement, such as a beer can. Assuming the beer can is located in the lower right corner of the screen, the calculation showed that the spatial overlap rate between the core area of ​​the advertisement and the highlighted area of ​​the heat map was 0.05.

[0217] Score Correction: Set a blind spot threshold of 0.15; since 0.05 < 0.15, the ad is determined to be in a blind spot. Calculate the correction factor and obtain the final creative score based on the correction factor.

[0218] The conversion prediction module combines demand matching score and creative preference score to identify the conversion probability of the target ad under the current viewing behavior characteristics. The conversion prediction module makes a final decision based on the above data.

[0219] Construct a conversion probability prediction model and input it after normalizing the demand matching score and material preference score;

[0220] A pattern confidence factor is introduced, which is determined by the distance membership degree and is used to characterize the impact of the accuracy of the current movie-watching pattern determination on conversion prediction.

[0221] Based on the normalized demand matching score, material preference score, and pattern confidence factor, the conversion probability under the current movie-watching behavior characteristics is predicted.

[0222] Normalization and splicing: Normalize the demand score and material score. Splice the one-hot encoded vector of the current viewing mode and time features to construct the input vector.

[0223] Neural network prediction: The input vector is fed into a multilayer perceptron (MLP) model. This model consists of an input layer, two hidden layers (128 and 64 nodes respectively, with ReLU activation function), and an output layer (Sigmoid). The model combines input features and pattern confidence factors as weights to adjust the output, and outputs the conversion probability.

[0224] Furthermore, the construction and training of the conversion probability prediction model includes:

[0225] Definition and transformation of pattern confidence factor: The value range of distance membership degree is [0,1]. The closer its value is to 1, the closer the current sample is to a certain cluster center, that is, the more reliable the pattern recognition. The pattern confidence factor is derived from the distance membership degree through a function such as the Sigmoid function, and it exhibits an S-shaped curve characteristic in the middle interval. It can effectively emphasize cases with extremely reliable pattern recognition, while strongly penalizing cases with unreliable recognition.

[0226] Neural network model architecture and training process: A multilayer perceptron is used for conversion probability prediction. The specific architecture is as follows:

[0227] Input layer: Receives 7 features: demand matching score (1-dimensional), material preference score (1-dimensional), normalized mode confidence factor (1-dimensional), and one-hot encoding of the current viewing mode (4-dimensional, corresponding to four viewing modes). Total input dimensions = 7.

[0228] The first hidden layer has 128 neurons and uses ReLU as the activation function. Dropout is added after this layer with a dropout rate of 0.3 to prevent overfitting.

[0229] The second hidden layer has 64 neurons with the ReLU activation function, and Dropout is also added with a dropout rate of 0.2.

[0230] Output layer: 1 neuron, activation function is sigmoid, output range [0,1], representing the conversion probability.

[0231] Training Data Preparation and Model Training: Valid ad-user-conversion result triplet data were randomly sampled from the historical movie viewing database. For each triplet, the label was defined as follows: if the user clicked or made a purchase within 30 seconds of the ad display, the label = 1; otherwise, the label = 0. The dataset was randomly divided into training, validation, and test sets in a 7:1.5:1.5 ratio. All features in the training set were Z-score standardized, and the same training set statistics were used to standardize the validation and test sets to avoid data leakage.

[0232] Training is performed using mean squared error as the loss function, with gradient descent using the Adam optimizer. The initial learning rate is 0.001, the batch size is 128, and the maximum number of iterations is 100 epochs. After each epoch, the system evaluates the model's performance on the validation set, calculating the validation loss and AUC. If the validation loss does not improve after 10 consecutive epochs, an early stopping mechanism is triggered, halting training and loading the optimal model parameters.

[0233] The pattern confidence factor is integrated as follows: It is input into the input layer as one of the features and concatenated with other dimensional features. An additional adjustment step is added after the second hidden layer and before the output layer: the initial prediction value of the output layer is multiplied by the pattern confidence factor to obtain the final conversion probability prediction. This ensures that when pattern recognition is unreliable, regardless of how high the initial prediction value is, the final output conversion probability will be reduced accordingly, thus achieving risk control.

[0234] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A deterministic conversion identification method based on large-screen advertising, characterized in that, include: The system collects viewing behavior characteristics on the large screen, including program switching sequence data, program content attribute data, remote control interaction command sequence, and the current time point. The movie-watching behavior features are input into the movie-watching pattern classification model, clustered and matched from the historical movie-watching database, and the current movie-watching pattern category is output. The viewing mode category represents the aggregation state of the current audience's behavioral interaction habits and content consumption tendencies in front of the big screen; Based on the movie viewing mode category, historical movie viewing consumption data that matches the mode category is filtered from the historical movie viewing database; the historical movie viewing consumption data includes movie viewing behavior characteristics and converted product information under the same movie viewing mode. The process of obtaining historical movie-watching consumption data based on movie-watching mode categories includes: A movie viewing mode index table is established, which records the movie viewing mode category to which each effective conversion behavior in the historical movie viewing database belongs. Based on the current movie viewing mode category, a reverse query is performed through the movie viewing mode index table to recall all historical conversion records marked with the same category; The historical conversion records of the recall are filtered again based on scene similarity. The program content attribute data of the current viewing behavior characteristics are extracted as the query vector, and the cosine similarity with the historical program content attributes in the recall records is calculated. Records with a cosine similarity greater than the preset association threshold are retained to form historical viewing consumption data. The historical movie-watching consumption data includes the list of purchased items, purchase amount, and conversion time interval data generated by users when watching the same movie-watching mode and watching similar content styles; The process involves: acquiring the advertising creative characteristics and promoted products of the target advertisement; conducting a two-way demand correlation analysis on the promoted products based on historical movie-watching consumption data to generate a demand matching score; the two-way demand correlation analysis includes complementarity analysis and substitution analysis; and conducting creative bias analysis on the advertising creative characteristics based on historical movie-watching consumption data to obtain a creative bias score. By combining demand matching score and creative preference score, the conversion probability of the target ad is identified under the current movie-watching behavior characteristics.

2. The deterministic conversion recognition method based on large-screen advertising according to claim 1, characterized in that: The movie-watching behavior characteristics are used to quantify the current interaction state, and their specific data dimensions are defined as follows: The program switching timing data includes the channel switching rate, non-linear jump frequency, and average continuous playback duration of a single content within a preset time sliding window, which is used to distinguish between browsing mode and immersive mode. The program content attribute data includes the metadata feature vector of the currently playing streaming media content, covering the content's primary category tags, secondary theme tags, and target audience profile tags. The remote control interaction command sequence includes remote control button trigger density, button response delay, and function key usage preference; The current time point includes the timestamp feature of the current system time and the time period attribute corresponding to the timestamp.

3. The deterministic conversion recognition method based on large-screen advertising according to claim 1, characterized in that: The construction process of the movie viewing mode classification model includes: Collect historical viewing databases from large-screen terminals and extract historical viewing feature vectors; use unsupervised clustering algorithms to perform cluster analysis on the historical viewing feature vectors, dividing the data space into high-density clusters, with each cluster center representing a viewing mode category; the viewing mode categories include, but are not limited to: high-frequency switching exploration mode, high-engagement viewing mode for children, background sound accompaniment mode, and deep immersive viewing mode. Calculate the Euclidean distance between the currently collected movie-watching behavior features and each cluster center; select the category label corresponding to the nearest cluster center as the movie-watching mode category; at the same time, calculate the membership degree of the current movie-watching behavior feature to the cluster center. If the membership degree is lower than the preset outlier threshold, mark the current state as a novel mode and trigger the incremental clustering update mechanism.

4. The deterministic conversion recognition method based on large-screen advertising according to claim 1, characterized in that: A two-way demand correlation analysis is conducted on the promoted products, specifically including: Complementarity analysis process: Traverse the converted products in historical movie-watching consumption data and query the co-occurrence probability between the target promotion product and the converted products; the co-occurrence probability represents the tendency for the two products to be consumed together under the same movie-watching mode; calculate the complementarity score based on the weighted co-occurrence probability; Substitutability analysis process: Identify converted products in historical movie-watching consumption data that belong to the same category as the target promoted product; obtain the most recent purchase time of the converted products, and calculate the substitutability score at the current moment by combining it with the average repurchase cycle; The final demand matching score is obtained by linearly superimposing the complementarity score and the substitution score, and the superposition weight is determined by the conversion preference characteristics of the current movie viewing mode category.

5. The deterministic conversion recognition method based on large-screen advertising according to claim 1, characterized in that: Conduct material preference analysis on advertising material characteristics, specifically including: By analyzing the historical movie-watching consumption data and recording the corresponding historical advertising materials based on conversion rates, we can construct high-response material features for the current movie-watching mode. Multimodal feature analysis is performed on the target advertisement to extract visual and auditory elements; the multimodal features of the target advertisement are matched with the features of the high-response material to calculate the matching degree, and the weighted matching degree of the visual and auditory elements is output, which is the material tendency score.

6. The deterministic conversion recognition method based on large-screen advertising according to claim 1, characterized in that: The process of calculating the conversion probability under the current movie-watching behavior characteristics includes: Construct a conversion probability prediction model and input it after normalizing the demand matching score and material preference score; A pattern confidence factor is introduced, which is determined by the distance membership degree and is used to characterize the impact of the accuracy of the current movie-watching pattern determination on conversion prediction. Based on the normalized demand matching score, material preference score, and pattern confidence factor, the conversion probability under the current movie-watching behavior characteristics is predicted.

7. A deterministic conversion recognition system based on large-screen advertising, characterized in that, include: The behavior acquisition module collects viewing behavior characteristics on the large screen, including program switching sequence data, program content attribute data, remote control interaction command sequence, and the current time point. The pattern recognition module inputs the movie-watching behavior features into the movie-watching pattern classification model, clusters and matches them from the historical movie-watching database, and outputs the current movie-watching pattern category. The viewing mode category represents the aggregation state of the current audience's behavioral interaction habits and content consumption tendencies in front of the big screen; The data matching module, based on the movie viewing mode category, filters historical movie viewing consumption data that matches the mode category from the historical movie viewing database; the historical movie viewing consumption data includes movie viewing behavior characteristics and converted product information under the same movie viewing mode. The process of obtaining historical movie-watching consumption data based on movie-watching mode categories includes: A movie viewing mode index table is established, which records the movie viewing mode category to which each effective conversion behavior in the historical movie viewing database belongs. Based on the current movie viewing mode category, a reverse query is performed through the movie viewing mode index table to recall all historical conversion records marked with the same category; The historical conversion records of the recall are filtered again based on scene similarity. The program content attribute data of the current viewing behavior characteristics are extracted as the query vector, and the cosine similarity with the historical program content attributes in the recall records is calculated. Records with a cosine similarity greater than the preset association threshold are retained to form historical viewing consumption data. The historical movie-watching consumption data includes the list of purchased items, purchase amount, and conversion time interval data generated by users when watching the same movie-watching mode and watching similar content styles; The feature analysis module acquires the advertising material features and promoted products of the target advertisement; based on historical movie-watching consumption data, it performs two-way demand correlation analysis on the promoted products and generates a demand matching score; the two-way demand correlation analysis includes complementarity analysis and substitution analysis; based on historical movie-watching consumption data, it performs material preference analysis on the advertising material features and obtains material preference score; The conversion prediction module combines demand matching score and creative preference score to identify the conversion probability of the target ad under the current movie-watching behavior characteristics.

Images