Deterministic conversion identification method and system based on large-screen advertisement

Abstract

The invention relates to the technical field of data analysis, in particular to a large-screen advertisement-based deterministic transformation recognition method and system, which comprises the following steps of: collecting film watching behavior characteristics of a large screen, inputting the film watching behavior characteristics into a film watching mode classification model, clustering and matching from a historical film watching database, and outputting a current film watching mode category; on the basis of the film watching mode category, screening out historical film watching consumption data matched with the mode category from a historical film watching database; obtaining advertisement material characteristics of the target advertisement and a promotion commodity, and based on historical film watching consumption data, performing bidirectional demand association analysis on the promotion commodity to generate a demand matching score; performing material tendency analysis on the advertisement material features based on historical film watching consumption data to obtain a material tendency score; and synthesizing the demand matching score and the material tendency score, and identifying the conversion probability of the target advertisement under the current film watching behavior characteristics.

Description

TECHNICAL FIELD

[0001] 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. BACKGROUND

[0002] In the field of digital marketing and e-commerce, in-screen purchasing of large-screen terminals (such as smart TVs and OTT devices) has become an important commercial channel in the home scenario. Existing Internet advertising delivery systems usually rely on device IDs or static attribute data of registered accounts as decision-making basis when allocating commercial resources and estimating conversion rates. However, this processing method based on static administrative data faces serious data processing and decision-making problems in dynamic business marketing scenarios.

[0003] There is a mapping deviation between audience recognition data and actual consumption decision-making subjects, resulting in low efficiency of marketing resource allocation. As a public asset in the family, the consumption subject behind the large-screen terminal is variable. The existing data processing method cannot dynamically analyze the current consumption decision-making unit according to real-time interaction behavior data, and can only mechanically follow the historical portrait of the account holder. This leads to the fact that the advertising delivery system often pushes high-value commercial information to non-target audiences, such as pushing high-priced goods to children who have no purchasing decision-making power, resulting in a great waste of advertising inventory resources and an inefficient consumption of marketing budgets, making it difficult to improve the effective reach rate and final conversion rate of commercial information.

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

[0005] Therefore, a certainty conversion identification method and system based on large-screen advertising are proposed. SUMMARY

[0006] The purpose of the present application is to provide a certainty conversion identification method and system based on large-screen advertising, which combines demand matching scores and material inclination scores to identify the conversion probability of target advertising under current viewing behavior characteristics.

[0007] To achieve the above purpose, the present application provides a certainty conversion identification method based on large-screen advertising, comprising: Collecting viewing behavior characteristics of the large screen, the viewing behavior characteristics including program switching timing data, program content attribute data, remote control interaction instruction sequence, and current time point; Inputting the viewing behavior characteristics into a viewing mode classification model, clustering and matching from a historical viewing database, and outputting the current viewing mode category; the viewing mode category represents the aggregated state of the behavior interaction habits and content consumption inclination of the current audience group in front of the large screen; Filtering historical viewing consumption data matching the mode category from a historical viewing database based on the viewing mode category; the historical viewing consumption data includes viewing behavior characteristics under the same viewing mode and converted commodity information; Obtaining the advertising material features of the target advertisement and the promoted commodity; based on the historical viewing consumption data, performing bidirectional demand correlation analysis on the promoted commodity to generate a demand matching score; the bidirectional demand correlation analysis includes complementarity analysis and substitution analysis; based on the historical viewing consumption data, performing material tendency analysis on the advertising material features to obtain a material tendency score; 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.

[0008] The viewing behavior characteristics are used to quantify the current interaction state, and the specific data dimension definition is as follows: The program switching time sequence data includes the channel switching rate, the nonlinear jump frequency, and the average continuous playing time of single content in a preset time sliding window, and is used to distinguish between browsing mode and immersion mode; The program content attribute data includes the metadata feature vector of the current playing streaming media content, covering the first-level classification label, the second-level theme label, and the target audience portrait label of the content; The remote control interaction instruction sequence includes the remote control key trigger density, the key response delay, and the function key use preference; The current time point includes the timestamp feature of the current system time and the time period attribute corresponding to the timestamp.

[0009] The construction process of the viewing mode classification model includes: Collecting a large-screen terminal historical viewing database and extracting historical viewing feature vectors; using an unsupervised clustering algorithm to perform clustering analysis on the historical viewing feature vectors, dividing the data space into high-density clusters, and each cluster center represents a viewing mode category; the viewing mode category includes but is not limited to: high-frequency switching exploration mode, children's high-stickiness viewing mode, background sound accompanying mode, and deep immersion viewing mode; Calculating the Euclidean distance between the current collected viewing behavior characteristics and each cluster center; selecting the category label corresponding to the nearest cluster center as the viewing mode category; at the same time, calculating the distance membership degree of the current viewing behavior characteristics to the cluster center, if the membership degree is lower than a preset outlier threshold, the current state is marked as a new mode and an incremental clustering update mechanism is triggered.

[0010] The process of matching the historical viewing consumption data according to the viewing mode category includes: An index table of viewing modes is established, which records the viewing mode category to which each valid conversion behavior in the historical viewing database belongs; Based on the current viewing mode category, the viewing mode index table is queried in reverse to recall all historical conversion records marked with the same category; The recalled historical conversion records are subjected to secondary filtering of scene similarity, the program content attribute data of the current viewing behavior characteristics are extracted as query vectors, and the cosine similarity of the historical program content attributes in the recalled records is calculated; records with a cosine similarity greater than a preset correlation threshold are retained to form historical viewing consumption data; The historical viewing consumption data includes the user's purchased item list, purchase amount, and conversion time interval data generated when the user watches content with similar styles under the same viewing mode.

[0011] Bidirectional demand correlation analysis is performed on the promoted goods, specifically including: Complementary analysis process: traverse the converted goods in the historical viewing consumption data, query the co-occurrence probability between the target promoted goods and the converted goods; the co-occurrence probability represents the tendency of the two goods being consumed together under the same viewing mode; the complementary score is calculated based on the co-occurrence probability weighting; Replacement analysis process: identify the converted goods in the historical viewing consumption data that belong to the same category as the target promoted goods; obtain the latest purchase time of the converted goods, and calculate the replacement score at the current time based on the average re-purchase period; The final demand matching score is obtained by linearly superimposing the complementary score and the replacement score, and the superimposition weight is determined by the conversion preference feature of the current viewing mode category.

[0012] Material tendency analysis is performed on the advertising material features, specifically including: Statistical analysis of the feature distribution of the historical advertising materials corresponding to the conversion rate records in the historical viewing consumption data to construct high-response material features under the current viewing mode; Multi-modal feature analysis is performed on the target advertisement to extract visual elements and auditory elements; the multi-modal features of the target advertisement are matched with the high-response material features to calculate the weighted matching degree of the visual elements and auditory elements, which is the material tendency score.

[0013] The process of calculating the conversion probability under the current viewing mode category includes: A conversion probability prediction model is constructed, and the demand matching score and the material tendency score are normalized and input; A mode confidence factor is introduced, which is determined by the distance membership and used to represent the influence of the accuracy of the current viewing mode determination on the conversion prediction; Based on the normalized demand matching score, the material tendency score, and the pattern confidence factor, a conversion probability under the current viewing behavior characteristics is predicted.

[0014] A deterministic conversion identification system based on large-screen advertising, comprising: A behavior collection module collects viewing behavior characteristics of the large screen, the viewing behavior characteristics including program switching timing data, program content attribute data, remote control interaction instruction sequences, and a current time point; A pattern recognition module inputs the viewing behavior characteristics into a viewing mode classification model, clusters and matches in a historical viewing database, and outputs a current viewing mode category; the viewing mode category represents the aggregated state of the behavior interaction habits and content consumption tendencies of the current audience group in front of the large screen; A data matching module filters historical viewing consumption data matching the pattern category from the historical viewing database based on the viewing mode category; the historical viewing consumption data includes viewing behavior characteristics and converted product information under the same viewing mode; A feature analysis module obtains advertising material features and promoted products of a target advertisement; based on the historical viewing consumption data, bidirectional demand correlation analysis is performed on the promoted products to generate a demand matching score; the bidirectional demand correlation analysis includes complementarity analysis and substitution analysis; based on the historical viewing consumption data, material tendency analysis is performed on the advertising material features to obtain a material tendency score; A conversion prediction module integrates the demand matching score and the material tendency score to identify a conversion probability of the target advertisement under the current viewing behavior characteristics.

[0015] Compared with the prior art, the present application has the following advantages: 1. The present application maps the fuzzy family account to a specific viewing consumption mode through unsupervised clustering processing of timing behavior data, interaction instructions, and content attributes, overcomes the directional bias problem caused by the dependence of traditional methods on static account identity data, accurately locks the behavior state with specific consumption tendencies, effectively cleans the invalid audience traffic without touching the user privacy red line, ensures that the advertisement placement decision is made based on the current real consumption scenario, and thus greatly improves the coverage accuracy of advertisement resources on target audiences and the input-output ratio of marketing activities.

[0016] 2. The present application differs from the traditional shallow recommendation based on content interest, and innovatively constructs a bidirectional demand correlation model including complementarity analysis and substitution analysis; by processing the co-occurrence probability and product life cycle characteristics in the historical transaction data, the abstract purchase intention is converted into a calculable quantitative score, which can intelligently identify potential associated purchase opportunities and repurchase or replacement opportunities, greatly improving the confidence of the conversion rate estimation model.

[0017] 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

[0018] Figure 1 This is a flowchart illustrating a deterministic conversion recognition method based on large-screen advertising according to the present invention. Figure 2 This is a data logic diagram of a deterministic conversion recognition method based on large-screen advertising according to the present invention; 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

[0019] 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.

[0020] Example 1: 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: 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 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. 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. An advertisement material feature and a promoted commodity of the target advertisement are acquired; based on historical viewing consumption data, bidirectional demand correlation analysis is performed on the promoted commodity to generate a demand matching score; the bidirectional demand correlation analysis includes complementarity analysis and substitution analysis; based on historical viewing consumption data, material tendency analysis is performed on the advertisement material feature to obtain a material tendency score; The demand matching score and the material tendency score are combined to identify a conversion probability of the target advertisement under the current viewing behavior feature.

[0021] The viewing behavior feature is used to quantify the current interaction state, and the specific data dimension definition is as follows: The program switching timing data includes channel switching rate, nonlinear jump frequency, and average continuous playing time of single content in a preset time sliding window, and is used to distinguish between browsing mode and immersion mode; The program content attribute data includes a metadata feature vector of the current playing streaming media content, covering the first-level classification label, the second-level theme label, and the target audience portrait label of the content; The remote control interaction instruction sequence includes remote control key trigger density, key response delay, and function key use preference; The current time point includes a timestamp feature of the current system time and a time period attribute corresponding to the timestamp.

[0022] Preferably, the data acquisition process includes: Acquiring program switching timing data: set a time sliding window, for example, five minutes, count the number of times the user switches channels in the time sliding window, and calculate the switching rate; at the same time, identify nonlinear jump behavior, i.e., the frequency of the user randomly jumping to the channel without following the list order, and record the average continuous playing time of single content.

[0023] Acquiring program content attribute data: reading the metadata of the current playing streaming media, extracting the first-level classification, second-level theme, and target audience label of the content; wherein the first-level classification includes movies, TV series, variety shows, animation, etc., the second-level theme includes science fiction, suspense, parenting, etc., and the target audience label is pre-set, such as suitable for 3-6 years old, suitable for all ages, etc.

[0024] Acquiring remote control interaction instruction sequence: acquiring remote control signals through infrared or Bluetooth interface, counting the key trigger density in unit time, recording the key response delay, i.e., the time difference between the appearance of the picture and the user pressing the confirmation key; and counting the use preference of function keys such as volume keys and menu keys.

[0025] Acquiring the current time point: acquiring the timestamp of the current system, and mapping it to the time period attribute, such as workday evening peak, weekend early morning.

[0026] Among them, the nonlinear jump frequency refers to the number of times the user changes the channel or content in a short time; high-frequency nonlinear jumps usually mean that the user is in a purposeless browsing state, and the conversion rate is usually low; and low frequency represents immersive viewing. The key response delay refers to the user's reaction speed to screen information, which is an important indicator of user focus. The lower the delay, the more focused the user's attention.

[0027] Further, the specific implementation of data collection includes: Time sliding window management mechanism: maintain a fixed-size ring buffer, whose size corresponds to the maximum number of events corresponding to the time sliding window (for example, 5 minutes). Whenever a new channel switching event occurs, record the timestamp, source channel ID, and target channel ID. If the buffer is full, remove the oldest event using the first-in, first-out strategy, and insert the new event at the end of the buffer.

[0028] Definition and calculation of channel switching rate: channel switching refers to the event that the user presses the channel plus or minus key or directly inputs the channel number using the remote control, causing the playing content to switch from one channel to another. Within the time sliding window, all valid switching events are counted, and the total number of events is divided by the window length to obtain the switching rate. To eliminate noise caused by accidental triggering of the remote control, the system introduces a de-bouncing mechanism: if the time interval between two switching events is less than a pre-set de-bouncing threshold, for example, 100 milliseconds, it is considered as a single accidental operation and only counted as one switching event.

[0029] Construction and normalization of high-dimensional feature vector: within a single viewing sampling period, four-dimensional data is collected simultaneously in the following order. After the sampling period ends, each dimension is independently normalized by the minimum and maximum value. After normalization, the four normalized dimension data is concatenated into a feature vector, which is input into the viewing mode classification model.

[0030] Measurement and processing of remote control key response delay: when outputting an interactive prompt to the user interface, such as displaying a pop-up menu or highlighting an option, the rendering completion timestamp of the prompt is recorded simultaneously. When a remote control key signal is detected, the arrival timestamp of the key event is recorded. The difference between the two timestamps is the key response delay. If the user does not produce any key operation within 30 seconds after the interactive prompt appears, the interaction is recorded as a non-response state, and the corresponding delay value is set to 30 seconds. For multiple consecutive interactive prompts, the average of all valid delay values is calculated as the key response delay indicator for the current sampling period.

[0031] The application realizes accurate quantification of the user state in front of a large screen by constructing a multi-dimensional viewing behavior feature system; time sequence switching data and remote control interaction instructions are introduced, which can effectively distinguish different states such as on-hold playing, random browsing and deep viewing. For example, if an animation is played, and high-frequency volume adjustment and random jumping are accompanied, it may mean that a parent is debugging the device rather than a child watching. Through this fine-grained feature capture, a large amount of false traffic noise can be removed, providing a solid data foundation for subsequent pattern recognition, thereby significantly improving the accuracy of audience identification and solving the technical problem of unclear actual viewing subject behind a home device.

[0032] The construction process of the viewing mode classification model includes: A large-screen terminal historical viewing database is collected, and a historical viewing feature vector is extracted; an unsupervised clustering algorithm is used to analyze the historical viewing feature vector, and the data space is divided into high-density clusters, and each cluster center represents a viewing mode category; the viewing mode category includes but is not limited to: high-frequency switching exploration mode, children's high-stickiness viewing mode, background sound accompanying mode and deep immersion viewing mode; The Euclidean distance between the current collected viewing behavior feature and each cluster center is calculated; the category label corresponding to the nearest cluster center is selected as the viewing mode category; at the same time, the distance membership degree of the current viewing behavior feature to the cluster center is calculated, and if the membership degree is lower than a preset outlier threshold, the current state is marked as a new mode and an incremental clustering update mechanism is triggered.

[0033] The viewing mode classification model uses an unsupervised learning algorithm to construct a classification model, and specifically uses a density-based clustering algorithm; Training process: batch export of historical viewing feature vectors of large-screen terminals; then, the clustering algorithm is used to calculate the distribution of these vectors in a multi-dimensional space, and points with similar distances are merged into high-density clusters; the algorithm is iterated 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 immersion viewing mode; Identification process: when online processing, the Euclidean distance between the current collected feature vector and all trained cluster centers is calculated; the label represented by the nearest cluster center is selected as the current mode; Incremental update mechanism: the distance membership degree of the current feature to the cluster center is calculated synchronously, and if the membership degree is lower than a preset outlier threshold, it means that the current behavior does not belong to any known mode, and the state is marked as a new mode and stored in a cache area; when a certain amount of data is accumulated in the cache area, an incremental clustering update is triggered to generate a new cluster center; Euclidean distance is the straight-line distance between two points in a multi-dimensional space, which is used to measure the similarity between the current behavior and the historical typical behavior; Distance membership degree represents the probability or degree of the current sample belonging to a certain cluster, which is used to judge the reliability of pattern recognition and prevent errors caused by forced classification; High-frequency switching exploration mode: This mode corresponds to the state that the user does not know what to watch and is frequently changing channels to find, that is, the browsing type state; The user quickly browses the channel list and does not fix on a certain content, and the attention is in a free state.

[0034] Technical features include: Program switching timing: The channel switching rate is extremely high, and the non-linear jump frequency is high, that is, large-span random jump channel selection; Playing time: The average continuous playing time of a single content is very short, for example, a few seconds to tens of seconds; Remote control interaction: The key trigger density is extremely high, and the user frequently operates the up and down keys or the back key; At this time, the user's attention is on the screen, but in the screening stage, the tolerance to interstitial advertising is low, and it is easy to directly swipe away, which is suitable for extremely short and visually powerful brand exposure advertising, rather than long story type advertising.

[0035] Children's high-stickiness viewing mode: Corresponding to the state of children watching animation or specific children's programs; The viewer is usually very focused, high-stickiness, and rarely switches once the content is selected.

[0036] Technical features include: Program content attributes: In the metadata characteristics of the playing content, the first-level classification is animation or children, and the target audience portrait label is clear for the low-age segment such as 3-6 years old.

[0037] Program switching timing: The switching rate is extremely low, the single content continuous playing time is very long, and the immersion viewing is very long.

[0038] Remote control interaction: The key trigger density is extremely low.

[0039] The subject of consumption decision may not be in front of the screen, or just accompany; It is suitable for the placement of toys, snacks and other child-driven purchases, or educational products.

[0040] Background sound accompanying mode: This mode corresponds to the state that the user only listens but does not watch, and the TV serves as the background noise of the family. The user may be cleaning, playing with the phone or cooking, and the TV is on playing news, music variety shows or long episodes.

[0041] Technical features include: Remote control interaction: The key indicator is that the key response delay is extremely high. For example, the screen pops up an interactive pop-up window or an advertisement, and the user does not react for a long time.

[0042] Program content attribute: usually news information, music variety show and other auditory information rich content.

[0043] Program switching timing: long time no operation, long playing time, and no pause, fast forward and other fine operations.

[0044] Visual advertising is almost ineffective, identify this mode and in subsequent material tendency analysis, preferentially match the advertising material with high auditory element score, and reach the audience through sound.

[0045] Deep immersion viewing mode: corresponding to the state of deep watching TV programs; Technical features: Current time point: usually occurs in the time stamp of the time period attribute as weekday evening peak or weekend; Program content attribute: the content is usually suitable for all age groups of variety shows, movies or sports events; Remote control interaction: low key response delay, user attention, and may be accompanied by moderate volume adjustment or pause function key use preference.

[0046] Further, the Mini-Batch K-Means algorithm is used for incremental updating, and when the number of samples in the storage area reaches a threshold, such as 1000, the old cluster center is used as the initial centroid, and the storage area data is iteratively updated. The elimination mechanism is supplemented, and the time decay factor of the cluster center is introduced, and when a cluster does not absorb new samples for M consecutive update periods, it is hidden from the model.

[0047] The modeling strategy of unsupervised clustering combined with incremental updating greatly improves the adaptability of the system to complex family scenes. Traditional supervised learning requires a large amount of manually labeled data and is difficult to cope with new behavior patterns. This scheme automatically discovers the internal structure in the data through clustering analysis, which not only identifies viewing patterns, but also extracts implicit habits; more importantly, the incremental clustering update mechanism enables the model to evolve, automatically learning new behavior paradigms as the family structure changes or the holiday schedule adjusts, ensuring the recognition robustness of long-term operation and avoiding model failure due to data drift.

[0048] The process of matching historical viewing consumption data according to the viewing mode category includes: A viewing mode index table is established, which records the viewing mode category to which each valid conversion behavior in the historical viewing database belongs; Based on the current viewing mode category, reverse query is performed through the viewing mode index table to recall all historical conversion records marked as the same category; The scene similarity secondary filtering is performed on the historical conversion record of the recall, program content attribute data of the current viewing behavior feature is extracted as a query vector, and the cosine similarity of the historical program content attribute in the recall record is calculated; the record with the cosine similarity greater than a preset correlation threshold is retained to form historical viewing consumption data; The historical viewing consumption data includes the purchased commodity list, the purchase amount and the conversion time interval data generated by the user when watching the content style under the same viewing mode.

[0049] Index query: a viewing mode index table is maintained, the table takes the mode category as the key and takes the historical effective conversion record ID as the value. First, all historical records of the purchase behavior under the mode are recalled in reverse through the index table according to the currently determined mode category.

[0050] Secondary filtering: in order to improve the precision, the attribute data of the current playing content, such as science fiction, action, etc., is extracted as a query vector. Then, the cosine similarity of the vector and the attribute of the playing content in the recall record is calculated.

[0051] Data reorganization: a correlation threshold is set, and only the records with the cosine similarity greater than the threshold are retained. The record set constitutes the final historical viewing consumption data, which includes the purchase list, the amount and the conversion time interval at that time.

[0052] Cosine similarity: the cosine value of the included angle between two vectors is calculated to evaluate their similarity. In the comparison of text or label features, the cosine similarity can better reflect the consistency of the content theme than the distance, and is not affected by the numerical size.

[0053] The two-level filtering mechanism of mode primary screening and content fine screening is used to construct a high-quality reference data set. Simply relying on the viewing mode may be too broad, for example, the immersive viewing mode may be watching a tragedy movie or a comedy variety show, and the consumption mentality of the two is completely different. On the basis of locking the behavior state, the cosine similarity matching of the content attribute is further introduced in the scheme, so that the historical data is highly fitted with the current scene in the state and context two dimensions; a large amount of irrelevant data with the same mode but different content context is excluded, so that the subsequent demand prediction is based on the highly similar parallel space-time history, thereby greatly improving the reference value of the prediction result and the accuracy of the conversion rate estimation.

[0054] Bidirectional demand correlation analysis is performed on the promoted commodity, specifically including: Complementary analysis process: the converted commodities in the historical viewing consumption data are traversed, and the co-occurrence probability between the target promoted commodity and the converted commodity is queried; the co-occurrence probability represents the tendency of the two commodities being consumed together under the same viewing mode; the complementary score is calculated based on the co-occurrence probability weighting; Substitutability analysis process: identify the converted goods in the historical viewing consumption data that belong to the same category as the target promoted goods; obtain the latest purchase time of the converted goods, combine the average re-purchase cycle, and calculate the substitutability score at the current time; The final demand matching score is obtained by linear superposition of the complementarity score and the substitutability score, and the superposition weight is determined by the conversion preference characteristics of the current viewing mode category.

[0055] Complementarity method: traverse the filtered historical viewing consumption data; for the target promoted goods, count the number of times it and the historical converted goods appear in the same order or the same time window. Calculate the co-occurrence probability, i.e. the probability of both occurring divided by the probability of each occurring, and use this probability as a weight to calculate the complementarity score.

[0056] Substitutability analysis algorithm: identify whether the historical data contains goods that belong to the same category as the target goods, such as purchased goods that belong to detergents. If so, obtain the time of the last purchase and retrieve the average re-purchase cycle for that category. Calculate the ratio of the current time to the last purchase duration to the average cycle to generate a substitutability score.

[0057] Linear superposition: according to the characteristics of the current viewing mode, such as the exploration mode preferring new products and the daily mode preferring re-purchase, set the weight and add the complementarity score and the substitution score.

[0058] Co-occurrence probability: the statistical probability of two events occurring simultaneously, based on association rule mining, used to discover potential associated needs.

[0059] Average re-purchase cycle: the average time interval for users to repeatedly purchase similar goods, based on the theory of product life cycle, used to predict the inventory consumption and re-purchase timing of users.

[0060] This scheme can accurately capture potential matching needs across categories through complementarity analysis, and accurately predict the re-purchase time window of daily necessities through substitutability analysis. This combination of product correlation and life cycle analysis enables targeted advertising at the right time when users need to replace or purchase matching products, greatly improving user purchase willingness and the success rate of advertising conversion.

[0061] Further, in the process of promoting commodities for bidirectional demand correlation analysis, a spatio-temporal commodity knowledge graph can also be used for deep reasoning; the spatio-temporal commodity knowledge graph comprises commodity nodes, time nodes and scene nodes; the complementarity analysis no longer depends on co-occurrence probability, but performs the following graph reasoning steps: mapping the current viewing mode category to a scene node in the graph, and mapping the current time point to a time node; taking the scene node and the time node as starting anchor points, performing multi-hop path walking in the graph, and finding a target promotion commodity that has a strong "scene-function-time" correlation path with a historical converted commodity node; calculating the semantic correlation strength of each path, and comprehensively generating a deep demand matching score comprising explicit complementarity and implicit causality, to replace the complementarity score based on statistical probability.

[0062] Model architecture: a knowledge graph is constructed in advance; nodes not only have commodities, but also scenes, such as late-night sports viewing mode, weekend family outings, and time seasons such as summer and winter.

[0063] Reasoning process: assuming that the current recognition is a late-night sports viewing mode, and the time is summer, the traditional statistics may only know to buy beer; but the graph reasoning can walk through the path: late-night sports viewing -> easy fatigue -> need to refresh -> functional beverages / coffee, or summer -> many mosquitoes -> mosquito repellent.

[0064] Logical details: path search is performed through a graph neural network to find implicit demands that are not obvious in statistical data but logically valid; node embedding is calculated using the DeepWalk algorithm based on random walking, and the correlation strength is determined by calculating the cosine similarity; or the transition probability from the scene node to the commodity node is calculated using the P-PageRank algorithm.

[0065] The present scheme breaks through the limitations of traditional collaborative filtering based on historical statistics, i.e., only recommending things that others have bought. By introducing a spatio-temporal knowledge graph, it has the ability of logical reasoning and can discover cross-category implicit causal demands. This is particularly important for new product promotion, as it can deduce potential demands based on scene logic without the need for accumulating a large amount of historical data, greatly expanding the correlation range and surprise degree of advertisement recommendation.

[0066] Material tendency analysis is performed on the features of the advertisement material, specifically including: Statistical analysis of the feature distribution of the corresponding historical advertisement material under the conversion rate record in the historical viewing consumption data, to construct high-response material features under the current viewing mode; Multi-modal feature analysis is performed on the target advertisement to extract visual elements and auditory elements; the multi-modal features of the target advertisement are matched with the high-response material features to calculate the weighted matching degree of the visual elements and the auditory elements, and the weighted matching degree is the material tendency score.

[0067] High-response feature construction: Statistics in the history of viewing consumption data, successful conversion of advertising material features. For example, statistics show that in this mode, warm color, fast-paced music ads have the highest conversion rate, and high-response material feature models are constructed.

[0068] Multi-modal analysis and matching: Multi-modal feature analysis of target ads. Visually, extract the color histogram and brightness distribution of the key frame; audibly, extract the volume dynamic range and rhythm BPM (beats per minute).

[0069] Matching degree calculation: Compare the feature vector of the target ad with the high-response material feature vector, and calculate the weighted matching degree. If the feature distribution overlap is high, output a higher material tendency score.

[0070] Multi-modal feature analysis: A technology that simultaneously processes multiple media information such as images and sounds. The effectiveness of an ad depends not only on the product, but also on whether the visual and auditory impact resonates with the viewer's current emotional state.

[0071] Further, the specific implementation of multi-modal feature analysis and matching: Visual feature extraction and quantification: The following visual feature extraction operations are performed on the target ad. First, if the ad is in video format, the system samples key frames at fixed intervals, a total of N key frames; if the ad is a static image, N = 1.

[0072] For each key frame, the following extraction steps are performed: Color feature: Convert the key frame from the RGB color space to the HSV color space; calculate the histogram (bin number = 32) of the H (hue), S (saturation), and V (brightness) channels respectively to obtain a 3x32=96-dimensional color feature vector. The calculation method is as follows: for each channel, divide the pixel value range [0, 255] uniformly into 32 bins, count the number of pixels in each bin, and normalize to obtain the histogram distribution of the channel.

[0073] Spatial feature: Extract the SIFT (Scale Invariant Feature Transform) key points of the key frame, retain the top 100 key points, and construct a 100-dimensional spatial feature vector.

[0074] Brightness distribution feature: Calculate the histogram (bin number = 32) of the key frame grayscale image and its first moment (average brightness) and second moment (contrast), and construct a 34-dimensional brightness feature vector. Average the visual feature vectors of all N key frames to obtain the final visual feature.

[0075] Extraction and quantification of auditory features: the following processing is performed on the audio track of the advertisement; first, the audio data is subjected to a short-time Fourier transform, frame length = 2048 samples, frame shift = 512 samples, to obtain a time-frequency spectrogram; then, the following feature extraction is performed on the time-frequency spectrogram: Rhythm feature: the energy envelope of the time-frequency spectrogram is calculated, the dominant period of the rhythm is identified by summing the energy of each frequency component using the autocorrelation function, and is converted to BPM (beats per minute). If the identification result is in the range of 30-300 BPM, the BPM value is retained; otherwise, it is set to 0, indicating that the rhythm is not obvious.

[0076] Dynamic range of volume: the maximum E_max and minimum E_min of the energy of all frames in the time-frequency spectrogram are calculated, and the dynamic range is defined as: 20*log10(E_max / E_min)dB.

[0077] Frequency distribution: the frequency range is divided into four intervals: low frequency (0-250Hz), medium-low frequency (250-1000Hz), medium-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.

[0078] Vocal proportion: a pre-trained vocal detection model is applied to the audio spectrum to calculate the proportion of time frames containing obvious vocal features to the total number of frames, resulting in a scalar value [0,1]; the above features are concatenated to obtain the final auditory features.

[0079] Construction of high-response material features: in the model training stage, all advertisements in the historical viewing consumption data that result in conversion are counted. A conversion event is defined as a user clicking or purchasing within 30 seconds after the advertisement is displayed. All conversion events corresponding to the advertisements are filtered out, and the conversion rate (number of conversion events / total number of displays of the advertisement) is calculated. The conversion rates are sorted from high to low, and the top 20% of the advertisements with the highest conversion rates are selected as the high-response advertisement set. The visual features and auditory features of all advertisements in the set are extracted, the mean and covariance of all visual feature vectors are calculated, and the distribution parameters of the high-response visual features are obtained; similarly, the distribution parameters of the high-response auditory features are obtained. These distribution parameters represent the preferred advertisement feature style of the user in this viewing mode.

[0080] Calculation of matching degree and generation of material tendency score: in the prediction stage, the similarity between the visual features of the advertisement to be launched and the high-response visual feature distribution is calculated using the Mahalanobis distance formula; similarly, the auditory similarity is calculated. The two similarities are combined by weighting to obtain the final material tendency score, where the weights are set according to the sensory sensitivity of the user in this viewing mode. For example, in the deep immersion viewing mode, the user is fully engaged, and the audio-visual balance is balanced, so both are set to 0.5.

[0081] The application solves the problem of low conversion rate caused by the incoordination between the advertising materials and the viewing atmosphere through material tendency analysis. In different viewing modes, the acceptance of users to the advertising form is quite different. For example, in the immersive viewing mode at midnight, noisy hard advertising materials may cause users to be disgusted and even close the device; and in the children stickiness mode, active audiovisual elements are more likely to attract attention. The scheme quantifies the audiovisual features of historical successful materials and matches them with the advertising to be launched, ensuring that the advertising not only matches the goods but also matches the material style. This fine material matching mechanism effectively reduces the advertising resistance of users and improves the viewing completion rate and the final conversion effect of the advertising.

[0082] Further, the material tendency analysis on the advertising material features further includes visual saliency heat map alignment; the last frame of the current playing content is extracted, a content residual attention heat map is generated by using a visual saliency detection algorithm, and a user's gaze area on the screen is determined; meanwhile, the coordinates of the advertising core visual elements of the first frame of the target advertising are extracted, and the spatial overlap rate of the advertising core visual elements and the content residual attention heat map is calculated; the spatial overlap rate is used as a correction factor to join the material tendency score; if the overlap rate is lower than a preset threshold, that is, the key information of the advertising appears in the user's gaze blind area, the material tendency score is reduced through the correction factor.

[0083] Further, in order to evaluate the visual connection effect of the advertising material at the moment when the current viewing ends, the embodiment further introduces an attention alignment mechanism based on visual saliency. The mechanism corrects the material score by generating a heat map and calculating a spatial overlap rate, and the specific implementation steps are as follows: Step one: generating a content residual attention heat map; When the current playing content is about to end, the last frame of picture is intercepted as a source image; in order to adapt to the real-time calculation demand of a large-screen terminal, a calculation method based on frequency domain residual is adopted, which specifically includes: Image preprocessing: converting the source image into a gray image and down-sampling it to a preset low-resolution size to reduce the subsequent calculation amount; Spectrum transformation: performing two-dimensional discrete Fourier transform on the preprocessed gray image, converting the image data from the spatial domain to the frequency domain, and thus obtaining an amplitude spectrum and a phase spectrum; Extracting frequency domain residual: first, calculating the natural logarithm of the amplitude spectrum to construct a logarithmic spectrum; second, using a mean filter (such as a 3x3 matrix) to smooth the logarithmic spectrum to obtain an average spectrum; finally, subtracting the average spectrum from the logarithmic spectrum to obtain the difference value, which is the frequency domain residual. This step is based on the statistical law of natural images, that is, the background texture usually conforms to a specific frequency spectrum distribution, and the salient target shows a mutation of the frequency spectrum; Reconstructing the heat map: the frequency domain residual is combined with the original phase spectrum, and the preliminary saliency map in the spatial domain is reconstructed by calculating the inverse discrete Fourier transform of the sum of the exponential frequency domain residual and the phase spectrum, and taking the square of its modulus. The preliminary saliency map is subjected to Gaussian smoothing filtering and normalization processing, and is mapped to the interval of 0 to 1, and finally the content residual attention heat map is obtained. The higher the value in the heat map, the higher the probability of residual attention of the audience's visual line at the moment of picture switching.

[0084] Step two: determining the coordinates of the core visual elements of the advertisement; For the target advertisement first frame picture to be evaluated, the coordinates of its core visual elements, such as brand logo and product main body, are obtained. The coordinate region 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 picture of the advertisement, extracting the connected region with a saliency value greater than a preset threshold, and taking the circumscribed rectangle of the connected region as the coordinate region.

[0085] Step three: calculating the spatial overlap rate; The determined coordinate region of the core visual elements of the advertisement is mapped into the above-mentioned content residual attention heat map. The spatial overlap rate is defined as the ratio of the sum of the saliency values of all pixel points located in the coordinate region of the core visual elements of the advertisement in the heat map to the total sum of the saliency values of all pixel points in the entire heat map. This index is used to quantify whether the key information of the advertisement appears in the high probability of the current visual line of the user.

[0086] Step four: correcting the material tendency score based on the overlap rate; A visual line blind area threshold is preset, for example, 0.15; the calculated spatial overlap rate is compared with the visual line blind area threshold: If the spatial overlap rate is greater than or equal to the visual line blind area threshold, the original material tendency score is kept unchanged, or a positive weighting of a preset amplitude is given.

[0087] If the spatial overlap rate is less than the visual line blind area threshold, it is determined that the key information of the advertisement is in the visual line blind area of the audience, and a penalty mechanism is triggered to reduce the material tendency score. At this time, the score is updated by using a correction factor.

[0088] The calculation logic of the correction factor is as follows: first, calculate the ratio of the spatial overlap rate to the visual line blind area threshold, subtract 1 from the ratio to get the normalized deviation; then multiply the normalized deviation by a preset penalty intensity coefficient; finally, subtract 1 from the product to get the final correction factor, and multiply the original material tendency score by the correction factor to get the updated final score.

[0089] 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 time, the user's gaze is highly concentrated on the left side. Perform saliency calculation on the movie ending picture to generate a heat map. If the upcoming advertisement is for a certain brand of car, and the car appears on the right side of the picture, the user will not be able to see the key point at first glance.

[0090] The present solution solves the transition blind area problem using computer vision technology; existing advertisement placement only focuses on content matching, ignoring the user's physiological visual inertia; the user's gaze will not immediately move during scene switching. By calculating the visual residual heat map, the overlap between the user's attention focus and the core area of the advertisement is identified, ensuring accurate identification of conversion probability.

[0091] The process of calculating the conversion probability under the current viewing mode category includes: A conversion probability prediction model is constructed, and the demand matching score and material tendency score are normalized and inputted; A mode confidence factor is introduced, which is determined by the distance membership and used to represent the influence of the accuracy of the current viewing mode determination on the conversion prediction; Based on the normalized demand matching score, material tendency score, and mode confidence factor, the conversion probability under the current viewing behavior characteristics is predicted.

[0092] Normalization processing: map the demand matching score and material tendency score calculated above to the interval of zero to one, eliminating the dimensional difference.

[0093] Introducing a confidence factor: extract the calculated distance membership and convert it into a mode confidence factor. If the current behavior characteristics are far away from the cluster center, it means that the mode determination is not accurate, and the confidence factor will be a small value.

[0094] Prediction calculation: construct a neural network model, input the normalized scores as features, and input the mode confidence factor as a regulation 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 multi-layer perception network, including an input layer, two hidden layers, and an output layer; the number of nodes in the hidden layer is 128 and 64 respectively, using ReLU activation function; the output layer uses Sigmoid activation function; the model is trained using mean square error loss function.

[0095] Mode confidence factor: a coefficient used to punish or reward the prediction result; based on Bayesian inference logic, when the prior condition (mode recognition) is uncertain, the credibility of the posterior probability (conversion prediction) should be reduced for risk control.

[0096] The present application quantifies the certainty degree of pattern recognition as a confidence factor, and introduces it as a weight into the calculation path of the conversion rate prediction model. When the user behavior characteristics deviate from the typical cluster center, the confidence factor is automatically used to punish and smooth the preliminary score output by the model. This mechanism suppresses the prediction overfitting caused by input feature noise or outlier behavior from the algorithm level, ensures that the final output conversion probability value can truly reflect the consistency of behavior pattern matching degree and user actual intention, and realizes the numerical calibration of conversion probability.

[0097] Further, the process of calculating the conversion probability also combines the feedback loop of online reinforcement learning. A state-action-reward model is constructed, the viewing mode category is taken as the environment state, the advertising placement strategy is taken as the action, and the user's instant interaction feedback is taken as the reward signal; the reward signal includes not only the positive reward of clicking or purchasing, but also the negative punishment of rapid escape behavior; the rapid escape behavior is defined as the power-off, channel change or mute operation within a specified time after the advertisement is displayed; the Q-Learning algorithm is used to update the Q value table of each type of advertising material in different viewing modes in real time; when predicting the conversion probability, the Q value corresponding to the target advertisement in the current mode is queried, which is taken as a historical lesson correction coefficient and superimposed into the final probability; if the Q value is a negative high value, the placement of the advertisement is forcibly blocked.

[0098] Scene description: an advertisement with a thrilling sound effect was placed in the background sound accompanying mode, and the user immediately pressed the mute key; the mute action is captured, defined as a high-weight negative reward, and the Q value table is updated through reinforcement learning, marking that "background sound mode + thrilling sound effect = poor strategy".

[0099] Strategy correction: next time the same mode is encountered, the conversion probability will also be automatically lowered due to the extremely low Q value, and the placement will be refused to avoid disturbing the user again.

[0100] The present scheme introduces a reinforcement learning mechanism with self-correction ability. Traditional conversion models are usually offline trained, and the update is lagging. The present application can capture the user's aversion emotion in real time through online reinforcement learning, and convert it into the system's muscle memory; not only the conversion rate is optimized, but more importantly, the long-term user experience of the large screen media is protected, and the traffic loss caused by excessive disturbance is prevented.

[0101] Embodiment two: The present application proposes a deterministic conversion recognition system based on large screen advertising, the system structure is as shown in Figure 3 The system structure is as shown in The behavior acquisition module acquires the viewing behavior characteristics of the large screen, and the viewing behavior characteristics include program switching timing data, program content attribute data, remote control interaction instruction sequence and current time point. The current physical time is Friday night 23:30, the user turns on the TV large screen from 23:00, after 3 minutes of program browsing switching, from 23:03 to the current time is watching a live football match; the behavior collection module collects the specific viewing behavior feature vector containing the following dimension data: Current time point: collect system timestamp and map to time period attribute [Friday, 23:30-24:00, eve of holiday before night].

[0102] Program switching timing data: read the operation records in the past 10 minutes sliding window, and statistically obtain the channel switching rate of 0 times / min, i.e. the user does not change the channel, the nonlinear jump frequency is 0, and the average continuous playing time of single content is 27 minutes. This group of data has the typical characteristics of low switching frequency and high continuity.

[0103] Program content attribute data: read the current playing streaming media metadata; extract the first-level classification as sports, the second-level theme as football / live, and the target audience portrait label as 25-45 years old / sports fan.

[0104] Remote control interaction instruction sequence: statistically the key data in the past 10 minutes, the key trigger density is extremely low, only once adjusting the volume at the start of the game, the key response delay is high, there is no pop-up interaction, and no data is measured, which is in the default passive watching state.

[0105] The mode recognition module inputs the viewing behavior features into the viewing mode classification model, clusters and matches in the historical viewing database, and outputs the current viewing mode category; the viewing mode category represents the aggregated state of the behavior interaction habits and content consumption tendency of the current audience group in front of the large screen; The mode recognition module receives the above feature viewing behavior features, and executes the following steps: Cluster matching: map the viewing behavior features to a pre-trained high-dimensional feature space, calculate the Euclidean distance or Mahalanobis distance between them and each cluster center; the system finds that the vector is closest to the cluster center of the deep immersion viewing mode.

[0106] Confidence calculation: calculate the distance membership of the current vector to the cluster center, and the calculation result is 0.85, which is higher than the outlier threshold 0.4, indicating that the current user behavior is very typical, and the mode recognition result is reliable. The system outputs the current mode category deep immersion viewing mode, and converts the distance membership into a mode confidence factor and transmits it to the subsequent module.

[0107] Preferably, in the present embodiment, DBSCAN algorithm is used for unsupervised clustering; the algorithm identifies high-density clusters based on the concept of density, and has good recognition ability for clusters of any shape.

[0108] The core parameters of DBSCAN are: Neighborhood radius eps, defined as 0.5, is verified by F1-Score on historical test data in normalized feature space, which makes the clustering effect and the ability to identify new and different samples achieve the optimal balance; Minimum point number min_pts, defined as 50, represents that an effective cluster must contain at least 50 samples; points below this threshold are marked as noise or outliers.

[0109] Distance membership is calculated by the following formula: membership = exp(-d² / σ²), where d is the distance from the current sample to the cluster center, and σ is the scale parameter of the covariance matrix of the cluster.

[0110] Training set preparation and initialization: the system periodically performs full model retraining. 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 performs weighted sampling on the sampled data according to their frequency of occurrence, ensuring that long-tail patterns such as viewing behavior during special periods such as the New Year are fully represented. The sampled feature vectors are input into the DBSCAN algorithm, and the algorithm is iterated until the cluster center is stable, i.e. the distance moved by the cluster center in the adjacent two iterations is less than 10^-5, obtaining a number of initial clusters, each of which is labeled as a viewing mode category.

[0111] The data matching module filters the historical viewing consumption data that matches the viewing mode category from the historical viewing database based on the viewing mode category; 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: First-level index recall: query the viewing mode index table to recall the set of valid conversion record IDs in the historical viewing database labeled as deep immersion viewing mode, totaling 5000; the historical viewing database can include all user data in the entire region to ensure the volume of data.

[0112] Second-level content filtering: extract the attribute vector of the current content, including sports, football, and competition categories; calculate the cosine similarity between this vector and the attribute of the content played at the time in the 5000 historical records.

[0113] Data generation: set the association threshold to 0.7, remove records of watching tragic movies or documentaries under the same mode, and only keep 1200 high-similarity records of watching sports events or competitive games, which constitute the historical viewing consumption data for this calculation.

[0114] The feature analysis module obtains advertisement material features and promoted goods of the target advertisement; based on historical viewing consumption data, bidirectional demand correlation analysis is performed on the promoted goods to generate a demand matching score; the bidirectional demand correlation analysis includes complementarity analysis and substitution analysis; based on historical viewing consumption data, material tendency analysis is performed on the advertisement material features to obtain a material tendency score; The target advertisement is identified, and it is determined that the promoted goods are craft beer; first, the latest consumption record of the user is retrieved, and it is found that 15 minutes before the current time point, i.e., 23:15, the account bound to the large screen has just completed an order of spicy peanuts (snack food); based on the above information, bidirectional demand correlation analysis is performed: Complementarity analysis: the system traverses the historical viewing consumption database, filters out all order sets marked as deep immersion viewing mode and containing snack food goods. In the order set, the number of joint orders containing beer / water goods is counted; it is calculated that in the late-night football watching scenario, 82% of users who buy snacks also buy beer. The system maps the co-occurrence probability 0.82 to a complementarity score.

[0115] Period prediction: query historical data to find that the average repurchase period of the user group for “beer / water” is 7 days. The historical record shows that the last purchase was 6 days ago. According to the last purchase interval divided by the average repurchase period, the substitution score is calculated.

[0116] Linear superposition: according to the preset weight table, in the deep immersion mode, users are more likely to be driven by physiological needs than simple repurchase; therefore, the complementarity weight is set to 0.7, and the substitution weight is set to 0.3, and the demand matching score is calculated.

[0117] Advertisement material tendency analysis: statistics of historical viewing consumption data show that high-conversion materials in this mode are mostly fast-paced and high-saturation; based on similarity identification of the target advertisement and high-conversion materials in this mode, the initial material score is 0.8; Visual saliency alignment: the last frame of the football match live broadcast is intercepted, such as the score board displayed in the upper left corner of the screen; using a frequency domain residual algorithm, the image is subjected to Fourier transform, log spectrum residual is extracted, inverse Fourier transform is performed for reconstruction, and a content residual attention heat map is generated; the heat map shows that the user's gaze is highly concentrated in the upper left corner of the score area.

[0118] Coordinate extraction and overlap calculation: the core visual elements of the first frame of the target advertisement are extracted, such as the beer can; it is assumed that the beer can is located in the lower right corner of the screen; it is found that the spatial overlap rate of the advertisement core area and the highlighted area of the heat map is 0.05.

[0119] Score correction: Set the threshold of line-of-sight blind area as 0.15; since 0.05 < 0.15, it is determined that the advertisement is in the blind area. Calculate the correction factor, and calculate the final material score based on the correction factor.

[0120] The conversion prediction module integrates the demand matching score and the material tendency score to identify the conversion probability of the target advertisement under the current viewing behavior characteristics. The conversion prediction module integrates the above data to make a final decision: A conversion probability prediction model is constructed, and the demand matching score and the material tendency score are normalized and inputted; A pattern confidence factor is introduced, which is determined by the distance membership and used to represent the influence of the accuracy of the current viewing pattern judgment on the conversion prediction; Based on the normalized demand matching score, the material tendency score, and the pattern confidence factor, the conversion probability under the current viewing behavior characteristics is predicted.

[0121] Normalization and splicing: normalize the demand score and the material score. Splice the One-hot encoding vector of the current viewing pattern and the time characteristics to construct the input vector.

[0122] Neural network prediction: input the input vector into a multi-layer perceptron (MLP) model. The model includes an input layer, two hidden layers (with 128 and 64 nodes respectively, and ReLU activation function), and an output layer (Sigmoid). The model combines the input features and the pattern confidence factor as a weight adjustment, and outputs the conversion probability.

[0123] Further, the construction and training of the conversion probability prediction model include: Definition and conversion of pattern confidence factor: the distance membership has a value range of [0, 1], and the closer the value is to 1, the closer the current sample is to a cluster center, i.e., the more reliable the pattern recognition is. The pattern confidence factor is converted from the distance membership through a function such as Sigmoid, and shows an S-shaped curve characteristic in the middle interval, which can effectively emphasize those cases with extremely reliable pattern recognition, and strongly penalize cases with unreliable recognition.

[0124] Architecture and training process of neural network model: a multi-layer perceptron is used for conversion probability prediction, and the specific architecture is as follows: Input layer: receives 7 features, including demand matching score (1 dimension), material tendency score (1 dimension), normalized pattern confidence factor (1 dimension), one-hot encoding of current viewing pattern (4 dimensions, corresponding to four viewing patterns). Total input dimension = 7.

[0125] The first hidden layer: 128 neurons, activation function ReLU, and add Dropout after this layer, dropout rate = 0.3, to prevent overfitting.

[0126] The second hidden layer: 64 neurons, activation function ReLU, and also add Dropout, dropout rate = 0.2.

[0127] The output layer: 1 neuron, activation function Sigmoid, output range [0, 1], representing the conversion probability.

[0128] Training data preparation and model training: randomly sample valid ad-user-conversion result triple data from the historical viewing database; for each triple, the label is defined as: if the user produces a click or purchase behavior within 30 seconds after the ad is displayed, then the label = 1; otherwise, the label = 0. The data set is randomly divided into training set, validation set and test set according to the ratio of 7:1.5:1.5; all features in the training set are standardized by Z-score, and the same training set statistics are used to standardize the validation set and test set to avoid data leakage.

[0129] Using mean square error as the loss function for training, using Adam optimizer for gradient descent, initial learning rate = 0.001, batch size = 128, and maximum iteration number is 100 epochs. After each epoch, the system evaluates the performance of the model on the validation set, calculates the validation loss and AUC index; if the validation loss of 10 consecutive epochs does not improve, the early stopping mechanism is triggered, the training is stopped and the optimal model parameters are loaded.

[0130] The integration method of pattern confidence factor: the pattern confidence factor is one of the input features into the input layer, and is spliced with other dimensional features. After the second hidden layer and before the output layer, an additional adjustment step is added: multiply the preliminary prediction value of the output layer by the pattern confidence factor to get the final conversion probability prediction. It ensures that when the pattern recognition is unreliable, no matter how high the preliminary prediction value is, the final output conversion probability will be correspondingly reduced, thereby achieving risk control.

[0131] Although embodiments of the present application have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and alterations can be made thereto without departing from the principles and spirit of the present application, 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. Obtain the ad creative characteristics and promoted products of the target ad; based on historical movie-watching consumption data, conduct two-way demand correlation analysis on the promoted products and generate demand matching scores; The bidirectional demand correlation analysis includes complementarity analysis and substitution analysis; based on historical movie-watching consumption data, the characteristics of advertising materials are analyzed to obtain material preference scores; 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: 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.

5. 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.

6. 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.

7. 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 viewing mode category 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.

8. 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 feature analysis module obtains 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 demand matching scores. The bidirectional demand correlation analysis includes complementarity analysis and substitution analysis; based on historical movie-watching consumption data, the characteristics of advertising materials are analyzed to obtain material preference scores; 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