Advertisement audience portrait construction and analysis device
By integrating user behavior signals across devices and analyzing them in real time, the advertising audience analysis device solves the problems of fragmented and lagging user behavior data, enabling accurate advertising decisions and matching user intent.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-16
- Publication Date
- 2026-04-03
AI Technical Summary
Existing technologies cannot effectively solve the problem of fragmented user behavior data across anonymous devices and over-reliance on lagging static tags, resulting in a low degree of matching between advertising decisions and users' current intentions.
By integrating cross-device behavioral signals through an advertising audience analysis device, the probability distribution of users' cognitive states and intent entropy are determined, generating dynamic advertising decisions. This device includes a signal acquisition module, a cross-device fusion module, a cognitive state inference module, and an intent entropy calculation module, utilizing meta-behavioral signals and content signals to generate advertising decisions.
It enables the real-time determination of a user's multi-dimensional internal state without relying on user identity identification, and generates advertising decisions that are highly matched with the user's current intent, avoiding data fragmentation and lag issues caused by device switching.
Smart Images
Figure CN121788166A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of online advertising technology, specifically to an advertising audience profiling and analysis device. Background Technology
[0002] In today's digital advertising ecosystem, the key to achieving precise ad targeting lies in building accurate and dynamic user profiles. Traditional advertising systems heavily rely on collecting users' historical behavioral data, such as browsing history, search history, and click behavior. By statistically analyzing this long-term data, users are labeled with a series of static tags, such as interests and demographic attributes. Ad placement decisions are then primarily based on these fixed tags.
[0003] However, this technological approach reveals its inherent flaws when faced with the complex and ever-changing usage scenarios of modern users. On the one hand, as users commonly own and frequently switch between multiple anonymous devices such as smartphones, tablets, and personal computers, their complete behavioral trajectories are artificially fragmented into multiple isolated data segments. Without a unified identity identifier, existing technologies struggle to correlate these scattered data points back to the same user entity, resulting in user profiles that are one-sided and incomplete.
[0004] On the other hand, even when users are active on a single device, their interaction intentions are far from static. Static tags based on historical data are essentially a lagging summary; they cannot capture the user's true state at the present moment. For example, a user whose long-term tag is basketball enthusiast might be focused on researching work-related information at the present moment. If the system still pushes sports equipment ads based on this long-term tag, it will not only waste advertising resources but also disrupt the user's normal task flow, resulting in a negative user experience. Existing technologies generally lack effective means to perceive users' real-time inner states in a refined manner, often attributing different interaction patterns to vague interests, making it difficult to distinguish whether a user is in a targeted search state or a casual exploratory browsing state, thus failing to generate advertising strategies that highly match the user's current intentions and receptiveness. Summary of the Invention
[0005] To address the shortcomings of existing technologies, this invention provides an advertising audience profile construction and analysis device. This device solves the problem that existing technologies, due to the fragmentation of user behavior data across anonymous devices and over-reliance on lagging static tags, cannot accurately perceive the user's real-time, multi-dimensional internal state, thus leading to a low degree of matching between advertising decisions and the user's current intent.
[0006] To achieve the above objectives, the present invention provides an advertising audience profile construction and analysis device, which includes a base, a stand, a display, and an advertising audience analysis and processing system disposed inside the stand. The advertising audience analysis and processing system determines the probability distribution of the user's cognitive state and intent entropy by fusing cross-device behavioral signals, and combines the two to generate advertising decisions.
[0007] The advertising audience analysis and processing system includes:
[0008] The signal acquisition module is used to acquire meta-behavioral signals and content signals from at least one anonymous device terminal;
[0009] A cross-device fusion module, connected to the signal acquisition module, is used to extract and compare composite behavioral features of different anonymous device terminals based on the meta-behavioral signals and content signals, and to anonymously fuse signals from the same user to obtain a fused behavioral stream.
[0010] The cognitive state inference module, connected to the cross-device fusion module, is used to determine the probability distribution of the user's cognitive state in a preset cognitive state space based on the meta-behavioral signals in the fused behavior stream.
[0011] The intent entropy calculation module, connected to the cross-device fusion module, is used to calculate the intent entropy, which characterizes the concentration of user goals, based on the content signals in the fused behavior stream.
[0012] The dynamic decision generation module is connected to the cognitive state inference module and the intent entropy calculation module, and is used to generate advertising decisions by combining the cognitive state probability distribution and the intent entropy.
[0013] In one specific embodiment, the cross-device fusion module includes:
[0014] The composite behavior feature extraction unit is used to extract a composite behavior feature for each anonymous device terminal within a preset time window based on the meta-behavioral signal and content signal obtained from the signal acquisition module.
[0015] A feature similarity calculation unit is used to calculate the similarity of the composite behavioral features between different anonymous device terminals;
[0016] The signal stream merging unit is used to merge the corresponding meta-behavioral signal and content signal according to timestamps when the similarity exceeds a preset threshold, and output the fused behavioral stream.
[0017] Preferably, the composite behavioral features extracted by the composite behavioral feature extraction unit include a meta-behavioral pattern vector and a cognitive state transition sequence. The meta-behavioral pattern vector is the result of statistical aggregation (e.g., mean or median calculation) of meta-behavioral signals within a preset time window; the cognitive state transition sequence is a preliminary cognitive state transition sequence obtained by mapping the meta-behavioral signal sequence within the time window using a preset classifier.
[0018] In one specific embodiment, the cognitive state inference module includes:
[0019] The prior distribution prediction unit is used to obtain the prior cognitive state probability distribution at the current moment based on the posterior cognitive state probability distribution and state transition model of the previous moment.
[0020] The posterior distribution update unit is used to adopt an observation model and take the meta-behavioral signals in the fused behavioral stream received by the cognitive state inference module as observation input to correct the prior cognitive state probability distribution and obtain the posterior cognitive state probability distribution at the current time, which is used as the cognitive state probability distribution.
[0021] Specifically, the prior distribution prediction unit calculates the probability distribution of prior cognitive states using the following formula:
[0022] S t|t-1 =A T S t-1|t-1 ;
[0023] Wherein: S t|t-1 A is the probability distribution vector of the prior cognitive state predicted at time t; A is the state transition probability matrix, whose elements A ij S represents the probability of transitioning from cognitive state i to cognitive state j; t-1|t-1 It is the posterior cognitive state probability distribution vector updated at time t-1; the superscript T indicates matrix transpose.
[0024] Preferably, the observation model used by the posterior distribution update unit defines the probability of observing a specific meta-behavioral signal given any cognitive state. This probability is modeled using a multivariate Gaussian distribution, and its probability density function is as follows:
[0025]
[0026] in:
[0027] P(M fused (t)|S t =c i S represents the true cognitive state at time t. t For the i-th cognitive state c i Under these conditions, the meta-behavioral signal M was observed.fused The conditional probability of (t);
[0028] M fused (t) represents the meta-behavioral signal feature vector obtained from the fused behavioral stream at time t;
[0029] D M The dimension of the feature vector of the meta-behavioral signal is represented;
[0030] μ i Indicates the relationship with the i-th cognitive state c i The mean vector of the associated multivariate Gaussian distribution;
[0031] Σ i Indicates the relationship with the i-th cognitive state c i The covariance matrix of the associated multivariate Gaussian distribution;
[0032] |Σ i | represents the covariance matrix Σ i The determinant of;
[0033] Represents the covariance matrix Σ i The inverse matrix;
[0034] The superscript T indicates the transpose operation of a vector;
[0035] exp(·) denotes the natural exponential function.
[0036] In one specific embodiment, the intent entropy calculation module includes:
[0037] The topic frequency statistics unit is used to map the content signal in the fused behavior stream received by the intent entropy calculation module to a discrete topic set within a preset time window, and to count the frequency of each topic in the discrete topic set.
[0038] An entropy calculation unit is used to calculate and output the intended entropy based on the frequency using the Shannon entropy formula.
[0039] Specifically, the entropy calculation unit calculates the intended entropy using the following formula:
[0040]
[0041] in:
[0042] H t This represents the intention entropy calculated at time t;
[0043] m represents the total number of topics in the discrete topic set;
[0044] o jThis represents the j-th topic in the discrete topic set;
[0045] p(o j ) indicates the topic o j Normalized frequency of occurrence within a preset time window;
[0046] This represents summing all the themes in the set;
[0047] log2 represents the logarithmic operation with base 2.
[0048] In one specific embodiment, the dynamic decision generation module includes:
[0049] The dominant state determination unit is used to determine the cognitive state with the highest probability value as the dominant cognitive state from the cognitive state probability distribution received by the dynamic decision generation module.
[0050] The decision instruction mapping unit is used to match and output a specific advertising decision from a preset set of advertising strategies based on the combination of the dominant cognitive state and the intent entropy received by the dynamic decision generation module.
[0051] In one specific embodiment, the signal acquisition module includes:
[0052] The content signal acquisition unit is used to acquire signals that represent the specific content of user interaction, such as search terms, page addresses viewed, or product icons clicked.
[0053] The metabehavioral signal acquisition unit is used to acquire metabehavioral signals that characterize the user's interaction methods.
[0054] Specifically, the meta-behavioral signals include at least one or more of the following: page scrolling rate, cursor or touch trajectory, text input rhythm, or device motion posture data.
[0055] Specifically, the cognitive state space consists of a set of abstract cognitive patterns decoupled from specific content topics, and the abstract cognitive patterns include at least one or more of the following: goal-oriented search, exploratory browsing, transaction execution, or information consumption.
[0056] This invention provides an apparatus for constructing and analyzing advertising audience profiles. It has the following beneficial effects:
[0057] 1. This invention extracts composite behavioral features consisting of meta-behavioral pattern vectors and cognitive state transition sequences through a cross-device fusion module, and fuses signals from different anonymous device terminals based on the similarity of these features. Without relying on user identification, it merges scattered signal streams originating from the same user into a more complete fused behavioral stream, overcoming the technical problem of fragmented user behavior data caused by device switching.
[0058] 2. This invention, through a cognitive state inference module and an intent entropy calculation module, determines the user's cognitive state probability distribution and intent entropy in real time based on meta-behavioral signals and content signals, respectively. These dual-dimensional indicators directly characterize the user's current interaction pattern and target concentration, overcoming the lag of traditional static labels. This allows the dynamic decision generation module to make decisions based on the user's immediate internal state, rather than on outdated historical preferences.
[0059] 3. This invention separately calculates the probability distribution of cognitive states reflecting interaction methods and the intent entropy reflecting target concentration, and finally combines the two by the dynamic decision generation module to generate a decision. This approach avoids the one-sidedness of relying solely on content signals or behavioral signals for judgment, and can distinguish between various refined scenarios such as targeted browsing and targeted searching, providing a more accurate technical basis for generating advertising decisions that highly match the user's current intent. Attached Figure Description
[0060] Figure 1 This is a perspective view of the present invention;
[0061] Figure 2 This is a block diagram of the functional modules of the advertising audience analysis and processing system of the present invention;
[0062] Figure 3 This is a framework diagram of the cognitive state inference loop process of the present invention;
[0063] Figure 4 This is a diagram of the instruction mapping framework for the dynamic decision generation module of the present invention.
[0064] The components are: 1. base; 2. upright; 3. monitor. Detailed Implementation
[0065] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0066] See attached document Figure 1-4The device includes a base 1, a vertical plate 2, and a display 3. The vertical plate 2 is fixedly connected to the top of the base 1, and the display 3 is disposed in the middle of the vertical plate 2.
[0067] The internal structure of panel 2 houses an advertising audience analysis and processing system. This system integrates behavioral signals from across devices to determine the probability distribution of users' cognitive states and intent entropy, and combines these two factors to generate advertising decisions. This system can be deployed on servers, cloud computing platforms, or dedicated hardware devices containing processors and memory.
[0068] See attached document Figure 2 , Figure 2 This is a functional block diagram of an advertising audience analysis and processing system according to an embodiment of the present invention. The system includes: a signal acquisition module 10, a cross-device fusion module 20, a cognitive state inference module 30, an intent entropy calculation module 40, and a dynamic decision generation module 50.
[0069] The signal acquisition module 10 is used to acquire raw behavioral data from at least one anonymous device terminal, the raw behavioral data including meta-behavioral signals and content signals.
[0070] The cross-device fusion module 20 is connected to the output of the signal acquisition module 10. It receives meta-behavioral signals and content signals, and extracts and compares composite behavioral features of different anonymous device terminals based on these signals. When it is determined that different anonymous device terminals originate from the same user, the corresponding signals are anonymized and fused to obtain a fused behavioral stream, which is then output to the cognitive state inference module 30 and the intent entropy calculation module 40.
[0071] The cognitive state inference module 30 is connected to the output of the cross-device fusion module 20. It receives the fused behavior stream and, based on the meta-behavioral signals in the fused behavior stream, determines the user's cognitive state probability distribution in a preset cognitive state space, and then outputs the probability distribution to the dynamic decision generation module 50.
[0072] The intent entropy calculation module 40 is connected to the output of the cross-device fusion module 20. It receives the fused behavior stream and calculates the intent entropy, which represents the concentration of user goals, based on the content signals in the fused behavior stream. Then, it outputs the intent entropy to the dynamic decision generation module 50.
[0073] The dynamic decision generation module 50 is connected to the output of the cognitive state inference module 30 and the output of the intent entropy calculation module 40, respectively. It receives the cognitive state probability distribution and intent entropy, and combines these two inputs to generate the final advertising decision.
[0074] The signal acquisition module 10 in the advertising audience analysis and processing system is the data input endpoint of the entire system. Its function is to acquire raw behavioral signals from one or more anonymous device terminals in real time. In one specific embodiment, this module includes a content signal acquisition unit and a meta-behavioral signal acquisition unit.
[0075] The content signal acquisition unit is used to acquire signals that characterize the specific content of user interactions. These signals directly reflect the semantic information that the user is interested in. Specifically, content signals include, but are not limited to: the text query string entered by the user in the search box; the Uniform Resource Locator (URL) of the webpage visited by the user; the unique identifier (ID) of the goods or services clicked or viewed by the user in the e-commerce application; and the title of the video watched or the article read by the user. These signals are acquired through scripts or software development kits (SDKs) deployed in the application or webpage.
[0076] The meta-behavioral signal acquisition unit is used to acquire signals that characterize the user's interaction methods. These signals are decoupled from the specific interaction content. Specifically, meta-behavioral signals include one or more of the following:
[0077] Page scroll rate is calculated by monitoring scroll events and determining the difference between the time difference between two events and the vertical displacement of the page, quantifying it as the number of scroll pixels per unit time. This rate reflects the speed and patience of a user when browsing information.
[0078] The cursor or touch trajectory is recorded as a series of two-dimensional coordinate points (x, y, z) with timestamps over a period of time. t ,y t The sequence of coordinates is used to calculate the instantaneous velocity, acceleration, and path curvature of the trajectory.
[0079] The text input rhythm is obtained by listening to keyboard events. It records the time interval sequence between consecutive keystrokes, or the sequence of the duration of each keystroke being pressed and released, rather than the specific character content being input.
[0080] The device's motion attitude data originates from the inertial measurement unit (IMU) built into the mobile device, primarily including readings from the three-axis accelerometer and three-axis gyroscope. The accelerometer provides acceleration data of the device along three orthogonal axes (α, β, γ). x ,a y ,a z The gyroscope provides data on the device's angular velocity around three orthogonal axes (ω). x ,w y ,w z ).
[0081] Finally, the signal acquisition module 10 forms two timestamp-aligned data streams from the acquired content signal and the meta-behavioral signal, and transmits these two data streams to the cross-device fusion module 20 as input for subsequent processing.
[0082] The cross-device fusion module 20 receives meta-behavioral signals and content signals from the signal acquisition module 10. Its function is to identify and merge signal data streams originating from the same anonymous user by analyzing user behavior patterns without relying on any personally identifiable identifiers. In one specific embodiment, the cross-device fusion module 20 includes a composite behavioral feature extraction unit, a feature similarity calculation unit, and a signal stream merging unit.
[0083] The composite behavioral feature extraction unit generates a composite behavioral feature for each independent anonymous device terminal within a preset, non-overlapping time window. This feature, as a summary of user behavior patterns over a time period, consists of two parts: a meta-behavioral pattern vector and a cognitive state transition sequence. The meta-behavioral pattern vector is a numerical vector obtained by statistically aggregating meta-behavioral signals within the time window; for example, calculating the mean and variance of page scrolling rate, the median cursor or touch trajectory speed, and the average keystroke interval for text input rhythm within that time period. The cognitive state transition sequence is a preliminary, discrete cognitive state transition sequence mapped from the meta-behavioral signal subsequences within the time window using a preset classification model.
[0084] The feature similarity calculation unit receives composite behavioral features from different anonymous device terminals and calculates the similarity between them. This calculation process consists of two steps. First, the similarity of corresponding parts of the two composite behavioral features is calculated separately. The similarity between meta-behavioral pattern vectors is calculated using the cosine similarity formula:
[0085]
[0086] Where: Sim vec (V A V B ) represents the meta-behavioral pattern vector V A and V B Similarity score between them;
[0087] V A and V B These represent the meta-behavioral pattern vectors from device A and device B, respectively.
[0088] V A,o and V B,o They are vectors V A and V B The o-th component;
[0089] n is the dimension of the meta-behavioral pattern vector;
[0090] V A ·V B Represents vector V A and V B The dot product;
[0091] Represents the dot product of two vectors;
[0092] It is a vector V A The specific calculation method for the modulus is as follows: Square each component of the vector, sum them up, and then take the square root.
[0093] It is a vector V B The specific calculation method for the modulus is as follows: Square each component of the vector, sum them up, and then take the square root.
[0094] ||V A ‖ and ‖V B ‖ represent vectors V A and V B The Euclidean norm (L2 norm).
[0095] Sim similarity between cognitive state transition sequences seq The similarity is calculated using normalized edit distance. Edit distance refers to the minimum number of single-element edits (insertions, deletions, or replacements) required to transform one sequence into another. The similarity formula is:
[0096]
[0097] in:
[0098] Sim seq (S A ,S B S represents the sequence of cognitive state transitions. A and S B Similarity score between them;
[0099] S A and S B These represent the cognitive state transition sequences from device A and device B, respectively.
[0100] LD(S A ,S B ) is the sequence S A and S B The Levenstein distance between them;
[0101] len(S A ) and len(S B )) are sequences SA and S B Length;
[0102] The max(·) function returns the maximum of two sequence lengths and is used for normalization.
[0103] Secondly, the two similarity scores mentioned above are combined into a total similarity score, Sim, through a weighted summation. total :
[0104] Sim total =α·Sim vec +(1-α)·Sim seq ;
[0105] Where: Sim total This represents the final similarity score between two composite behavioral features;
[0106] α is a preset weighting coefficient within the interval [0,1].
[0107] Sim vec It is the similarity of the meta-behavioral pattern vectors obtained from the aforementioned calculation;
[0108] Sim seq It is the similarity of the cognitive state transition sequence obtained from the aforementioned calculation.
[0109] The signal stream merging unit converts the overall similarity score Sim output by the feature similarity calculation unit. total With a preset similarity threshold θ sim Compare. If Sim total Greater than or equal to θ sim If the two signal streams originate from the same user, then the unit determines that they originate from the same user. At this point, the unit strictly sorts and merges the meta-behavioral signals and content signals corresponding to the two device terminals according to their respective timestamps, forming a single, continuous, and time-ordered fused behavioral stream. This fused behavioral stream is then output to the cognitive state inference module 30 and the intent entropy calculation module 40. If Sim... total Less than θ sim If the condition is met, then no merge operation will be performed.
[0110] See attached document Figure 2 and attached Figure 3 The cognitive state inference module 30 receives the fused behavior stream from the cross-device fusion module 20. Its function is to infer the probability distribution of the user in a preset cognitive state space in real time based on the meta-behavioral signals in the fused behavior stream. In one specific embodiment, the inference process of this module is based on a Bayesian filtering framework, which includes a prior distribution prediction unit and a posterior distribution update unit.
[0111] A cognitive state space is a set of predefined, abstract cognitive patterns decoupled from specific interactive content. Each cognitive pattern in this set represents a specific user interaction method. Specifically, abstract cognitive patterns include, but are not limited to: goal-oriented search, whose corresponding meta-behavioral signal characteristics are fast and stable page scrolling, direct and precise cursor or touch trajectories, and a relatively fast text input rhythm; exploratory browsing, whose corresponding meta-behavioral signal characteristics are slow and irregular page scrolling, and cursor or touch trajectories that move a wide range and hover frequently; transaction execution, whose corresponding meta-behavioral signal characteristics are small-range, precise clicks or touch operations within a specific page area (such as a form or button); and information consumption, whose corresponding meta-behavioral signal characteristics are prolonged page stillness or extremely slow, uniform scrolling.
[0112] The cognitive state inference module uses a Bayesian filtering framework, which dynamically updates the probability distribution of the cognitive state at each time step t through a prediction and update loop.
[0113] The prior distribution prediction unit is responsible for executing the prediction step in this framework. It predicts the prior cognitive state probability distribution at time t based on the posterior cognitive state probability distribution at time t-1 and a pre-defined state transition model. This process is achieved through the following matrix operations:
[0114] S t|t-1 =A T S t-1|t-1 ;
[0115] Among them, S t|t-1 Let be a column vector representing the probability distribution of the cognitive state predicted at time t based on all observation data up to time t-1, i.e., the prior probability distribution, where t|t-1 indicates that the state is a prediction of time t based on the information at time t-1; A is a state transition probability matrix, where the elements A... ij This represents the single-step transition probability of a user moving from the i-th cognitive state to the j-th cognitive state; T represents the transpose of a matrix.
[0116] S t-1|t-1 Let be a column vector representing the probability distribution of cognitive state updated at time t-1 by incorporating all observation data up to time t-1, i.e., the posterior probability distribution of the previous time step.
[0117] The posterior distribution update unit is responsible for executing the update steps in this framework. It employs an observation model, taking the meta-behavioral signal M at the current time t as its basis. fused (t) is used as the observation input to correct the prior probability distribution obtained from the prior distribution prediction unit, thereby obtaining the posterior cognitive state probability distribution at the current time. The observation model defines the probability distribution of the current cognitive state c given by any given cognitive state c. iBelow, a specific meta-behavioral signal M was observed. fused The conditional probability P(M) of (t) fused (t)|S t =c i ).
[0118] In one specific embodiment, this conditional probability is given by a probability density function of a multivariate Gaussian distribution:
[0119]
[0120] Where P(·|·) represents the conditional probability density function;
[0121] M fused (t) is a D M A column vector of dimension, representing the meta-behavioral signal feature vector obtained from the fused behavioral stream at time t;
[0122] S t Let be a random variable representing the user's true cognitive state at time t;
[0123] c i Let be a constant, representing the i-th cognitive state in the cognitive state space;
[0124] π is a mathematical constant, also known as Pi (the mathematical constant for a circle).
[0125] D M Let M be a scalar representing the feature vector M of the meta-behavior signal. fused The dimension of (t);
[0126] exp(·) represents the natural exponential function, that is, an exponential function with the natural constant e as the base;
[0127] μ i For a D M A column vector of dimension c, representing the relationship with the i-th cognitive state c. i The mean vector of the associated multivariate Gaussian distribution;
[0128] Σ i For a D M ×D M The matrix represents the relationship between the i-th cognitive state c. i The covariance matrix of the associated multivariate Gaussian distribution;
[0129] |Σ i | is a scalar representing the covariance matrix Σ i The determinant of;
[0130] For a D M ×D M The matrix Σ represents the covariance matrix.i The inverse matrix;
[0131] The superscript T indicates the transpose operation of a vector or matrix.
[0132] After calculating the observation probabilities for all cognitive states, the posterior distribution update unit, according to Bayes' theorem, multiplies the predicted prior probabilities element-wise with the observation probabilities and normalizes them to obtain the final posterior cognitive state probability distribution S. t|t The posterior distribution S t|t This is the output of the cognitive state inference module at time t, and it is transmitted to the dynamic decision generation module 50.
[0133] The intent entropy calculation module 40 receives the fused behavior stream from the cross-device fusion module 20. The function of this module is to calculate a numerical indicator, i.e., intent entropy, based on the content signal in the fused behavior stream within a preset time window, to quantify the concentration of user objectives. In one specific embodiment, this module includes a topic frequency statistics unit and an entropy value calculation unit.
[0134] The topic frequency statistics unit maps discrete content signals within a time window to a pre-defined, finite set of discrete topics. This mapping process can be implemented using various techniques, such as matching search keywords or URL paths in the content signal with a pre-defined topic dictionary; or using a pre-trained text classification model to map the text description of the content signal to a pre-defined topic category. After completing the mapping, this unit counts the frequency of each topic in the discrete topic set within the current time window and converts it into a normalized frequency.
[0135] The entropy calculation unit receives the normalized frequencies of each topic output by the topic frequency statistics unit and calculates the intent entropy within the current time window using the Shannon entropy formula. The calculation process is as follows:
[0136]
[0137] Wherein: H t This represents the intention entropy value calculated within the time window at time t;
[0138] m is the total number of topics in the discrete topic set;
[0139] This represents summing all the themes in the set;
[0140] o j Represents the j-th topic in the discrete topic set;
[0141] p(o j (This is the topic) jThe normalized frequency of occurrence within this time window is calculated as topic o j The number of occurrences is divided by the total number of all content signals within the time window, and satisfies the following conditions:
[0142] log2 represents the logarithmic function with base 2.
[0143] The calculated intention entropy H t It is a non-negative real number. The magnitude of this value is inversely proportional to the concentration of the user's goals: when the user's interaction content is highly concentrated on a few topics, p(o) j The distribution of H will be very uneven, and the calculated H t The value is lower; conversely, when the user's interactive content is widely distributed across multiple different topics, p(o) is higher. j The distribution of H will tend to be uniform, and the calculated H t The value is relatively high.
[0144] Finally, the intent entropy calculation module 40 calculates the intent entropy value H. t Output to the dynamic decision generation module 50.
[0145] See attached document Figure 2 The dynamic decision generation module 50 is connected to the cognitive state inference module 30 and the intent entropy calculation module 40, respectively. Its function is to receive and integrate information from the user's cognitive state and intent entropy, and generate a specific advertising decision instruction based on this. In one specific embodiment, this module includes a dominant state determination unit and a decision instruction mapping unit.
[0146] The dominant state determination unit is used to receive the posterior cognitive state probability distribution vector S output by the cognitive state inference module 30. t|t The goal of this unit is to generate cognitive states from a pre-defined set C = {c1, c2, ..., c...}. k In the process of selecting the cognitive state with the highest probability at the current time t, the dominant cognitive state is chosen. This selection process is determined by the following formula:
[0147]
[0148] in: This represents the dominant cognitive state determined at time t;
[0149] The operation represents the process of finding, within a predefined set of cognitive state spaces C, the preset cognitive state space parameters c that maximize the objective function p(·). i In other words, this operation will iterate through all elements (c1, c2, ..., c) in set C. k), and return the element that gives the highest probability value;
[0150] p(S t =c i |M fused (1:t) is the posterior cognitive state probability distribution vector S. t|t The i-th cognitive state in the equation represents the cognitive state at time t, given all meta-behavioral signal observations from time 1 to time t. i The probability of;
[0151] M fused (1:t) represents the meta-behavioral signal observation sequence from the initial time 1 to the current time t, after cross-device fusion.
[0152] The decision instruction mapping unit receives the dominant cognitive state output by the dominant state determination unit. and the intent entropy value H output by the intent entropy calculation module 40 t This unit is based on With H t The system combines various factors to match and output a specific advertising decision from a pre-defined set of advertising strategies. This matching process can be defined by a pre-defined decision matrix or a set of conditional rules.
[0153] For example, in one specific embodiment, the intention entropy H is first... t The value and the preset entropy threshold H threshold The intention entropy levels are compared and categorized into low and high levels. Then, based on the combination of the dominant cognitive state and the intention entropy level, a decision is selected and output from a set of advertising strategies, including:
[0154] Conversion ads: Ads that display specific product information, prices, and direct purchase links.
[0155] Discovery ads: Display category or brand information related to user interests, guiding users to discover new content.
[0156] Brand impression advertising: primarily visual displays, aiming to deepen brand recall, without emphasizing immediate interaction.
[0157] Reduce ad frequency: Reduce the number of times ads are displayed within a specified time period or do not display ads at all.
[0158] The specific mapping rules for this decision instruction mapping unit can be set as follows:
[0159] When the dominant cognitive state is goal-oriented search and the intent entropy is low, output conversion ads.
[0160] Discovery ads are output when the dominant cognitive state is exploratory browsing and the intent entropy is low.
[0161] When the dominant cognitive state is exploratory browsing and the intent entropy is high, brand impression ads are output.
[0162] When the dominant cognitive state is transaction execution, regardless of the intent entropy level, the output is to reduce the frequency of advertisements.
[0163] The advertising decision output by the dynamic decision generation module 50 is the final output of the system and is used to guide the subsequent operations of the advertising delivery system.
Claims
1. An advertising audience profiling and analysis device, comprising a base (1), a stand (2), and a display (3), characterized in that, The top of the base (1) is fixedly connected to a vertical plate (2), and a display (3) is provided in the middle of the vertical plate (2). An advertising audience analysis and processing system is provided inside the vertical plate (2). The advertising audience analysis and processing system is used to integrate cross-device behavioral signals, determine the user's cognitive state probability distribution and intention entropy, and generate advertising decisions by combining the cognitive state probability distribution and intention entropy. The advertising audience analysis and processing system includes: The signal acquisition module is used to acquire meta-behavioral signals and content signals from at least one anonymous device terminal; The cross-device fusion module extracts and compares the composite behavioral features of different anonymous device terminals based on the meta-behavioral signals and content signals, and performs anonymous fusion of signals from the same user to obtain a fused behavioral stream; The cognitive state inference module determines the probability distribution of the user's cognitive state in a preset cognitive state space based on the meta-behavioral signals in the fused behavioral stream. The intent entropy calculation module calculates the intent entropy, which characterizes the concentration of user goals, based on the content signals in the fused behavior stream. The dynamic decision generation module is used to generate advertising decisions by combining the probability distribution of the cognitive state with the intent entropy.
2. The advertising audience profiling construction and analysis device according to claim 1, characterized in that, The cross-device fusion module includes: The composite behavior feature extraction unit extracts a composite behavior feature for each anonymous device terminal within a preset time window, based on the meta-behavioral signal and the content signal. A feature similarity calculation unit is used to calculate the similarity of the composite behavioral features between different anonymous device terminals; The signal stream merging unit is used to merge the corresponding meta-behavioral signal and content signal according to timestamps when the similarity exceeds a preset threshold, and output the fused behavioral stream.
3. The advertising audience profiling construction and analysis device according to claim 1, characterized in that, The cognitive state inference module includes: The prior distribution prediction unit obtains the prior cognitive state probability distribution at the current moment based on the posterior cognitive state probability distribution and state transition model of the previous moment. The posterior distribution update unit is used to adopt an observation model, taking the meta-behavioral signals in the fused behavioral stream as observation inputs, to correct the prior cognitive state probability distribution and obtain the posterior cognitive state probability distribution at the current moment, which is then used as the cognitive state probability distribution.
4. The advertising audience profiling construction and analysis device according to claim 1, characterized in that, The intent entropy calculation module includes: The topic frequency statistics unit is used to map the content signal to a discrete topic set within a preset time window, and to count the frequency of each topic in the discrete topic set. An entropy calculation unit is used to calculate and output the intended entropy based on the frequency using the Shannon entropy formula.
5. The advertising audience profiling and analysis device according to claim 1, characterized in that, The dynamic decision generation module includes: The dominant state determination unit is used to determine the cognitive state with the highest probability value from the cognitive state probability distribution as the dominant cognitive state. The decision instruction mapping unit is used to match and output a specific advertising decision from a preset set of advertising strategies based on the combination of the dominant cognitive state and the intent entropy.
6. The advertising audience profiling construction and analysis device according to claim 1, characterized in that, The signal acquisition module includes: Content signal acquisition unit is used to acquire signals that represent specific content of user interaction; The metabehavioral signal acquisition unit is used to acquire metabehavioral signals that characterize the user's interaction methods.
7. The advertising audience profiling and analysis device according to claim 3, characterized in that, The observation model used by the posterior distribution update unit defines the probability of observing a specific meta-behavioral signal given any cognitive state.
8. The advertising audience profiling construction and analysis device according to claim 1, characterized in that, The meta-behavioral signals include at least one or more of the following: page scrolling rate, cursor, touch trajectory, text input rhythm, and device motion posture data.
9. The advertising audience profiling construction and analysis device according to claim 1, characterized in that, The cognitive state space consists of a set of abstract cognitive patterns decoupled from specific content themes, and the abstract cognitive patterns include at least one or more of the following: goal-oriented search, exploratory browsing, transaction execution, and information consumption.
10. The advertising audience profiling construction and analysis device according to claim 2, characterized in that, The composite behavioral feature includes: a meta-behavioral pattern vector generated based on the meta-behavioral signal within the preset time window; A cognitive state transition sequence generated within the preset time window based on changes in the user's cognitive state.