DMP data management method based on artificial intelligence

Abstract

The invention discloses a DMP data management method based on artificial intelligence, and relates to the technical field of data management.The method comprises the steps that multi-device user behavior data are collected through a cross-platform SDK, a browser plug-in and an application log, and data standardization is conducted on a central server or an edge computing node; in the user identity matching process, in combination with a device fingerprint and deep learning method, a user cross-device behavior mode is analyzed through a Transform model, static device features and dynamic behavior features are extracted, a self-attention mechanism is adopted to learn cross-device associated features, then the user matching degree is calculated through cosine similarity, and finally the user identity is judged. The stability and accuracy of identity matching are improved; and based on a user identity matching result, carrying out time sequence analysis by utilizing an LSTM (Long Short Term Memory) model, capturing evolution of user interests along with time, dynamically adjusting user portrait weights through reinforcement learning, and optimizing feature weight updating by using a Q learning mechanism, so that the portraits can be adaptively changed.

Description

Technical Field

[0001] The present invention relates to the technical field of data management, and in particular to a DMP data management method based on artificial intelligence. Background Art

[0002] In services such as advertising placement and recommendation systems, the data management platform DMP relies on multi-channel data integration to build user portraits for enterprises. In the past, DMP mainly relied on device identifiers such as third-party Cookies and IDFA for user identity recognition, and analyzed user behavior through a rule engine or a statistical model. In a single-device environment, this method could still work, but in the case where users frequently browse, interact, and consume across devices, it is difficult to effectively match the behavior trajectories by this method.

[0003] As the interaction methods between different terminals become increasingly complex, users may browse products on the mobile terminal, complete orders on the PC terminal, or even interact on other terminals. Traditional DMPs cannot effectively integrate this data, resulting in inaccurate user identity matching and affecting the effect of personalized recommendation. In addition, the data formats and storage architectures of different terminals are not unified, posing compatibility challenges for DMPs during data collection and fusion. Even if some DMPs attempt to supplement with device fingerprint technology, due to factors such as browser upgrades and privacy setting adjustments, the stability of device fingerprints is relatively low, making it difficult to support long-term user behavior tracking.

[0004] Facing this problem, some DMPs have introduced a user identification mapping system based on an Identity Graph. However, this method still relies on external data sources, and the data matching rate is limited by privacy policies and data quality. At the same time, there is a data synchronization delay in the Identity Graph, making it difficult to support real-time update prediction of user interests. As privacy policies become increasingly strict (such as iOS ATT restricting access to IDFA), the tracking method of traditional DMPs based on third-party identifiers is further restricted, making it difficult to effectively convert marketing effects. Therefore, there is an urgent need for a DMP data management solution based on artificial intelligence to solve such problems. Summary of the Invention

[0005] In view of the above existing problems, the present invention is proposed.

[0006] The present invention provides a DMP data management method based on artificial intelligence to solve the problems that traditional DMPs rely on Cookies and IDFA for user tracking, are difficult to effectively track cross-device users, have serious data fragmentation, unstable identity recognition, and limited personalized recommendation effects.

[0007] To solve the above technical problems, the present invention provides the following technical solutions:

[0008] An embodiment of the present invention provides an AI-based DMP data management method, which includes

[0009] Step S1: Collect the behavior data of the user on multiple devices, and perform standardization processing on the behavior data to obtain standardized data;

[0010] Step S2: Based on the standardized data, use device fingerprint recognition and deep learning models to perform user identity matching to obtain a matching result;

[0011] Step S3: Based on the matching result, construct a user portrait, and dynamically update the user interest tags by combining time series analysis and reinforcement learning methods to obtain an updated user portrait;

[0012] The time series analysis is based on the LSTM network to learn the user behavior trend;

[0013] The reinforcement learning method adaptively optimizes the user feature weights;

[0014] Step S4: Based on the updated user portrait, optimize the personalized recommendation strategy, dynamically adjust the recommended content to obtain a recommendation result;

[0015] Step S5: Based on the recommendation result, use federated learning to perform user behavior modeling, perform local calculations on the user device side, and use differential privacy technology to encrypt the behavior data. At the same time, manage the user data access rights through decentralized identity authentication.

[0016] As a preferred solution of the AI-based DMP data management method of the present invention, wherein: the behavior data includes web browsing records, application usage, mouse trajectories, click behaviors, dwell times, and interaction events.

[0017] As a preferred solution of the AI-based DMP data management method of the present invention, wherein: the step of collecting the behavior data of the user on multiple devices and performing standardization processing on the behavior data is

[0018] Monitor the behavior data of the user on multiple devices, collect the behavior data through cross-platform development kit SDK, browser plugins, and application logs, and upload it to the central server or edge computing node;

[0019] Perform preprocessing on the central server or edge computing node:

[0020] Remove abnormal data points, perform normalization processing on continuous numerical data, and use Min-Max standardization.

[0021] Map the high-dimensional data of mouse trajectories and click behaviors into feature vectors, and convert them into discrete categories using equal-frequency binning or equal-width binning methods.

[0022] For cross-device behavior data, align it based on the timestamp;

[0023] Generate a unique user identifier UID for associating data between different devices, and store the standardized data in a distributed database.

[0024] As a preferred solution of the AI-based DMP data management method described in the present invention, wherein: the device fingerprint recognition includes the operating system type, screen resolution, browser configuration, and network environment;

[0025] The deep learning model learns the user's cross-device behavior pattern based on the Transformer architecture and optimizes the matching accuracy through multi-task learning.

[0026] As a preferred solution of the AI-based DMP data management method described in the present invention, wherein: based on the standardized data, the steps of using device fingerprint recognition and deep learning model for user identity matching are as follows:

[0027] Extract the device fingerprint, and represent the characteristics of the device fingerprint f as:

[0028] f = (f1, f2, f3, …, f n )

[0029] wherein, f1 represents the operating system type, f2 represents the screen resolution, f3 represents the browser configuration, f4 represents the network environment, and n represents the total number of device fingerprint features.

[0030] Extract the user behavior characteristics, and define the user behavior vector v:

[0031] v = (v1, v2, v3, …, v m )

[0032] wherein, v1 represents the web browsing pattern, v2 represents the mouse movement pattern, v3 represents the click pattern, v4 represents the device switching pattern, and m represents the total number of behavior characteristics.

[0033] Use a deep learning model based on Transformer for identity matching, and define the input feature as X:

[0034] X = [f, v],

[0035] wherein, X is the comprehensive behavior feature matrix of the user, f is the device fingerprint feature vector, and v is the user behavior feature vector.

[0036] Input it into the Transformer encoding layer for encoding, and the encoding process is expressed as:

[0037] H = TransformerEncoder(X, W),

[0038] where Transformer Encoder is a Transformer encoding function for performing representation learning on input features, W is a Transformer training parameter matrix, and H represents the encoded behavioral feature vector,

[0039] The calculation process of Transformer encoding is as follows:

[0040] H = SelfAttention(X) + FeedForward(SelfAttention(X)),

[0041] where the self-attention layer SelfAttention calculates the correlations between different positions in the input sequence:

[0042]

[0043] where Q = XW Q is the query matrix, K = XW K is the key matrix, V = XW V is the value matrix, W Q , W K , W V are training parameters, and d k is the dimension of the key vector,

[0044] The features are further mapped by the feed-forward neural network FeedForward, and the formula is:

[0045] FeedForward(x) = σ(W1x + b1)W2 + b2,

[0046] where W1, W2 are weight matrices, b1, b2 are bias vectors, and σ is a non-linear activation function,

[0047] The cosine similarity is used to calculate the matching degree of two user data, and the calculation formula is:

[0048]

[0049] where S(u1, u2) represents the similarity between users u1 and u2, are the feature vectors of the users respectively, · represents the dot product, and |H| represents the L2 norm of the vector,

[0050] A threshold τ is set. When S(u1, u2) > τ, it is determined to be the same user; otherwise, it is determined to be different users:

[0051] If S(u1, u2) > τ, then

[0052] If S(u1, u2) ≤ τ, then

[0053] where is the matching result, 1 indicates the same user, 0 indicates different users, and τ is the matching threshold set empirically.

[0054] As a preferred solution of the DMP data management method based on artificial intelligence described in the present invention, wherein: the step of constructing a user portrait and dynamically updating user interest tags by combining time series analysis and reinforcement learning methods is

[0055] Based on the user identity matching result of step S2, define the initial user portrait feature vector P u :

[0056] P u =(p1, p2,..., p d ),

[0057] where P u represents the portrait vector of user u, p1, p2,..., p d respectively represent user attributes in different dimensions, and d is the number of dimensions of the user portrait

[0058] Calculate the initial user portrait according to the aggregated standardized behavior data, and the formula is:

[0059]

[0060] where p i represents the initial weight of the user on the portrait dimension i represents the behavior characteristics of the user at time t, and T is the total length of the data collection time window

[0061] Use the LSTM long short-term memory network to perform time series modeling on the user portrait, and define the input as:

[0062] H t = LSTM(H t-1 , v t ),

[0063] where H t is the hidden state at time t, representing the interest deviation of the user at this moment, H t-1 is the hidden state of the previous moment, and v t is the behavior characteristic collected at the current time t

[0064] The LSTM update formula is:

[0065] H t = σ(W h H t-1 + W v v t + b h ),

[0066] where W h , W v are the LSTM training weights, b h is the bias parameter, and σ is the activation function,

[0067] Dynamically adjust the weights of the user profile using the reinforcement learning method, and define the reward function as R t :

[0068]

[0069] where R t is the total reward at time t, w j is the weight of the j-th dimension of the profile feature, is the j-th dimension of the user profile,

[0070] Update the user profile using Q learning, and the update formula is:

[0071]

[0072] where Q(s t , a t ) represents the Q value of performing action a t in state s t , α is the learning rate, r t is the current reward, γ is the discount factor, s t+1 is the state at the next moment, a′ is the possible action at the next moment,

[0073] is the Q value of the optimal action in the next step,

[0074] The final user profile vector is P′ u :

[0075] P′ u = (p′1, p′2, …, p′ d ),

[0076] where,

[0077] p′ j = p j + Δp j ,

[0078] Δp j = β · (H t + Q(st , a t ))

[0079] Among them, p' j is the updated user portrait feature, Δp j is the portrait adjustment amount, and β is the smoothing factor.

[0080] As a preferred solution of the artificial intelligence-based DMP data management method described in the present invention, wherein: the recommendation strategy includes content recommendation, advertisement placement, and user behavior prediction.

[0081] As a preferred solution of the artificial intelligence-based DMP data management method described in the present invention, wherein: based on the updated user portrait, the steps of optimizing the personalized recommendation strategy and dynamically adjusting the recommended content are as follows:

[0082] Based on the user portrait P' u calculate the recommendation score, and the formula is:

[0083]

[0084] Among them, S i represents the preference score of the user for item i, is the correlation of item i in the user portrait dimension j,

[0085] Adopt Top K sorting to select the recommended items:

[0086] I * = argmax I (S),

[0087] Among them, I * is the optimal recommended item set, I is the candidate recommended item set, and argmax takes the item index corresponding to the maximum score.

[0088] Update the recommendation weight according to the user feedback, and the update formula is:

[0089] p' j = p' j + λ(F - E),

[0090] Among them, F is the actual feedback, E is the expected feedback, and λ is the feedback update rate.

[0091] As a preferred solution of the artificial intelligence-based DMP data management method described in the present invention, wherein: the optimized user behavior model is stored in the DMP system.

[0092] As a preferred solution of the DMP data management method based on artificial intelligence according to the present invention, wherein: the step of using federated learning to model user behavior, performing local calculations on the user device side, encrypting the behavior data using differential privacy technology, and managing user data access rights through decentralized identity authentication is as follows,

[0093] Perform federated learning calculations, define the local device model, expressed as:

[0094]

[0095] Wherein, is the model parameter of device u at time t + 1, is the model parameter of device u at time t, and η is the learning rate, is the loss gradient,

[0096] The server aggregates the updates of multiple users:

[0097]

[0098] Wherein, W t+1 is the global model parameter, and U is the total number of devices,

[0099] Perform differential privacy encryption, add noise to the gradient, and the formula is:

[0100]

[0101] Wherein, is the noisy gradient, is the Gaussian noise,

[0102] Define the user identity hash h u Decentralized identity management:

[0103] h u = H(ID u , K),

[0104] Wherein, h u is the user identity hash value, H(·) is the encryption hash function, ID u is the user unique identifier, and K is the key.

[0105] The beneficial effects of the present invention are:

[0106] In the present invention, multi-device user behavior data is collected through a cross-platform SDK, a browser plugin, and application logs, and data standardization is performed on a central server or an edge computing node to reduce fragmentation problems. During the user identity matching process, a device fingerprint and a deep learning method are combined. The Transformer model is used to analyze the cross-device behavior patterns of users, extract static device features and dynamic behavior features, and the self-attention mechanism is adopted to learn cross-device association features. Then, the cosine similarity is used to calculate the user matching degree, and finally the user identity is determined, avoiding the limitations of traditional DMPs that rely on Cookies and IDFAs, and improving the stability and accuracy of identity matching.

[0107] In the present invention, based on the user identity matching results, the LSTM model is used for time series analysis to capture the evolution of user interests over time, and the weights of the user profile are dynamically adjusted through reinforcement learning. The Q learning mechanism is used to optimize the update of feature weights, enabling the profile to adaptively change, thereby solving the problem that traditional DMP user profiles are static for a long time and difficult to adapt to real-time interest fluctuations.

[0108] In the personalized recommendation stage of the present invention, the user profile is used to calculate the recommendation score, and the optimal recommendation items are selected through Top K sorting. At the same time, the recommendation weights are adjusted based on user feedback (such as click-through rate, dwell time) to optimize the recommended content, enabling the recommendation system to accurately adapt to the current needs of users and enhancing the user experience.

[0109] In the present invention, federated learning is introduced, enabling model training to be performed on the user's local device side, and the updates of multiple users are aggregated through a server to improve the accuracy of personalized modeling. At the same time, differential privacy is combined, and noise is added during gradient updates to protect user data and avoid privacy leakage, and a decentralized identity authentication mechanism is used to ensure the security of user data access rights.

[0110] In summary, the present invention not only improves the accuracy of cross-device user identification, reduces data fragmentation, but also strengthens the dynamic update ability of the user profile, enabling the recommendation system to more accurately adapt to changes in user interests, while taking into account privacy compliance. Brief Description of the Drawings

[0111] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention, and those of ordinary skill in the art can obtain other accompanying drawings without creative efforts based on these drawings.

[0112] Figure 1 It is a schematic flow chart of the DMP data management method based on artificial intelligence of the present invention. Detailed Embodiments

[0113] To make the above objects, features, and advantages of the present invention more apparent and understandable, the following provides a detailed description of the specific embodiments of the present invention with reference to the accompanying drawings of the specification.

[0114] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art can make similar generalizations without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0115] Secondly, the so-called "one embodiment" or "embodiment" herein refers to specific features, structures, or characteristics that may be included in at least one implementation manner of the present invention. The phrase "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments.

[0116] Example 1, referring to Figure 1 , this example provides an AI-based DMP data management method, including the following steps:

[0117] Step S1, collect the behavior data of the user on multiple devices, and perform standardization processing on the behavior data to obtain standardized data;

[0118] The behavior data includes web browsing records, application usage, mouse trajectories, click behaviors, dwell times, and interaction events;

[0119] The steps of collecting the behavior data of the user on multiple devices and performing standardization processing on the behavior data are as follows:

[0120] Monitor the behavior data of the user on multiple devices, collect the behavior data through a cross-platform development kit SDK, browser plug-ins, and application logs, and upload it to a central server or an edge computing node;

[0121] Perform preprocessing on the central server or edge computing node:

[0122] Remove abnormal data points, perform normalization processing on continuous numerical data, and use Min-Max standardization.

[0123] Map the high-dimensional data of mouse trajectories and click behaviors into feature vectors, and convert them into discrete categories using the equal-frequency binning or equal-width binning method.

[0124] For cross-device behavior data, align it based on the timestamp.

[0125] Generate a unique user identifier UID for associating data between different devices, and store the standardized data in a distributed database.

[0126] Step S2: Based on the standardized data, use device fingerprint recognition and a deep learning model to perform user identity matching to obtain a matching result;

[0127] Device fingerprint recognition includes the operating system type, screen resolution, browser configuration, and network environment;

[0128] The deep learning model learns the user's cross-device behavior patterns based on the Transformer architecture and optimizes the matching accuracy through multi-task learning;

[0129] The steps of performing user identity matching using device fingerprint recognition and a deep learning model based on the standardized data are as follows:

[0130] Extract the device fingerprint. The feature composition of the device fingerprint f is expressed as:

[0131] f = (f1, f2, f3, …, f n )

[0132] where f1 represents the operating system type, f2 represents the screen resolution, f3 represents the browser configuration, f4 represents the network environment, and n represents the total number of device fingerprint features;

[0133] Extract the user behavior features. Define the user behavior vector v:

[0134] v = (v1, v2, v3, …, v m )

[0135] where v1 represents the web browsing pattern, v2 represents the mouse movement pattern, v3 represents the click pattern, v4 represents the device switching pattern, and m represents the total number of behavior features;

[0136] Use a deep learning model based on Transformer for identity matching. Define the input feature as X:

[0137] X = [f, v]

[0138] where X is the user's comprehensive behavior feature matrix, f is the device fingerprint feature vector, and v is the user behavior feature vector;

[0139] Input it into the Transformer encoding layer for encoding. The encoding process is expressed as:

[0140] H = TransformerEncoder(X, W)

[0141] Among them, the Transformer Encoder is the Transformer encoding function, which is used for representing and learning the input features. W is the Transformer training parameter matrix, and H represents the encoded behavioral feature vector.

[0142] The calculation process of Transformer encoding is as follows:

[0143] H = SelfAttention(X) + FeedForward(SelfAttention(X)),

[0144] Among them, the self-attention layer SelfAttention calculates the correlation between different positions in the input sequence:

[0145]

[0146] Among them, Q = XW Q is the query matrix, K = XW K is the key matrix, V = XW V is the value matrix, W Q , W K , W V are the training parameters, d k is the dimension of the key vector.

[0147] The features are further mapped by the feed-forward neural network FeedForward, and the formula is:

[0148] FeedForward(x) = σ(W1x + b1)W2 + b2,

[0149] Among them, W1 and W2 are the weight matrices, b1 and b2 are the bias vectors, and σ is the non-linear activation function.

[0150] The cosine similarity is used to calculate the matching degree of two user data, and the calculation formula is:

[0151]

[0152] Among them, S(u1, u2) represents the similarity between users u1 and u2. are the feature vectors of the users respectively, · represents the dot product, and |H| represents the L2 norm of the vector.

[0153] Set the threshold τ. When S(u1, u2) > τ, it is determined to be the same user, otherwise it is a different user:

[0154] If S(u1, u2) > τ, then

[0155] If S(u1, u2) ≤ τ, then

[0156] Among them, is the matching result, 1 indicates the same user, 0 indicates different users, and τ is the matching threshold set empirically.

[0157] Specifically, an identity matching model is constructed here through device fingerprints and user behavior characteristics;

[0158] Fingerprints cover static characteristics such as the operating system, screen resolution, browser configuration, etc., while user behavior characteristics include dynamic characteristics such as web browsing patterns, mouse trajectories, click behaviors, etc. The Transformer is used to encode the behavior patterns, cross-device association characteristics are extracted through the self-attention mechanism, and the cosine similarity is used to calculate the matching degree of different users; a matching threshold is set, and it is determined whether users have the same identity according to the similarity;

[0159] This method combines static device characteristics with dynamic behavior data, improves the accuracy of cross-device identity matching, and enhances the robustness; optimizing the matching process through deep learning helps to accurately identify the activities of the same user on different devices;

[0160] Step S3: Based on the matching result, construct a user portrait, and dynamically update the user interest tags by combining time series analysis and reinforcement learning methods to obtain an updated user portrait;

[0161] Time series analysis is based on the LSTM network to learn the user behavior trend;

[0162] The reinforcement learning method adaptively optimizes the user feature weights;

[0163] The steps of constructing a user portrait and dynamically updating the user interest tags by combining time series analysis and reinforcement learning methods are as follows:

[0164] Based on the user identity matching result in step S2, define the initial user portrait feature vector P u :

[0165] P u =(p1, p2, …, p d ),

[0166] Among them, P u represents the portrait vector of user u, and p1, p2, …, p d represent user attributes in different dimensions respectively, and d is the number of dimensions of the user portrait.

[0167] The initial user portrait is calculated by aggregating and calculating according to the standardized behavior data. The formula is:

[0168]

[0169] Among them, pi Denote the initial weight of the user in the portrait dimension i, Denote the behavioral characteristics of the user at time t, where T is the total length of the data collection time window,

[0170] Use the LSTM long short-term memory network to perform temporal modeling on the user portrait, and define the input as:

[0171] H t = LSTM(H t-1 , v t ),

[0172] where, H t is the hidden state at time t, representing the interest deviation of the user at this moment, H t-1 is the hidden state of the previous moment, and v t is the behavioral characteristic collected at the current time t,

[0173] The LSTM update formula is:

[0174] H t = σ(W h H t-1 + W v v t + b h ),

[0175] where, W h , W v are the LSTM training weights, b h is the bias parameter, and σ is the activation function,

[0176] Use the reinforcement learning method to dynamically adjust the user portrait weight, and define the reward function as R t :

[0177]

[0178] where, R t is the total reward at time t, w j is the weight of the portrait feature dimension j, is the j-th dimension of the user portrait,

[0179] Use Q learning to update the user portrait, and the update formula is:

[0180]

[0181] where, Q(s t , a t ) represents the Q value of performing the action a t in the state s t , α is the learning rate, r tis the current reward, γ is the discount factor, s t+1 is the state at the next moment, a′ is the possible action at the next moment,

[0182] is the Q value of the next optimal action,

[0183] The final user portrait vector is P′ u :

[0184] P′ u =(p′1,p′2,…,p′ d ),

[0185] in,

[0186] p′ j =p j +Δp j ,

[0187] Δp j =β·(H t +Q(s t ,a t )),

[0188] Among them, p′ j is the updated user portrait feature, Δp j is the image adjustment amount, β is the smoothing factor;

[0189] Specifically, user portraits are constructed based on matching results, and the portrait weights are dynamically optimized by combining time series analysis and reinforcement learning;

[0190] The user portrait is initialized by standardized behavior data, and the LSTM network captures the behavior change trend. Reinforcement learning optimizes the weight of the portrait feature through the Q learning mechanism. LSTM processes time series data and represents the evolution of user interests in hidden states. Reinforcement learning adjusts the portrait features based on feedback rewards. The user portrait is continuously optimized based on the behavior changes to more accurately reflect the user's current interests. It can adapt to the long-term and short-term changes in users' interests and improve the accuracy of the portrait.

[0191] Step S4, based on the updated user portrait, optimize the personalized recommendation strategy, dynamically adjust the recommended content, and obtain the recommendation result;

[0192] Recommendation strategies include content recommendation, advertising placement, and user behavior prediction;

[0193] Based on the updated user portrait, the steps to optimize the personalized recommendation strategy and dynamically adjust the recommended content are as follows:

[0194] Based on user profile P′ u Calculate the recommendation score using the formula:

[0195]

[0196] Among them, S i represents the user's preference score for item i, is the relevance of item i in the user profile dimension j,

[0197] Use Top K sorting to select recommended items:

[0198] I * = argmax I (S),

[0199] Among them, I * is the set of optimal recommended items, I is the set of candidate recommended items, and argmax takes the item index corresponding to the maximum score,

[0200] Update the recommendation weight according to the user feedback, and the update formula is:

[0201] p' j = p' j + λ(F - E),

[0202] Among them, F is the actual feedback, E is the expected feedback, and λ is the feedback update rate;

[0203] Specifically, calculate the recommendation score based on the updated user profile here, and use the Top K strategy to screen the optimal recommended content;

[0204] The recommendation score is calculated from the correlation between the user profile and the item features, and the item with the highest score is selected as the recommended item; in addition, the recommendation strategy is dynamically adjusted based on user feedback (such as clicks, dwell time, etc.) to optimize the sorting of recommended content; through adaptive weight adjustment, the personalized recommendation system can respond to changes in user interests, improve the recommendation accuracy; and balance historical interests and current behaviors;

[0205] Step S5, based on the recommendation results, use federated learning to model user behavior, perform local calculations on the user device side, encrypt the behavior data using differential privacy technology, and manage the user data access rights through decentralized identity authentication;

[0206] The optimized user behavior model is stored in the DMP system;

[0207] The steps of using federated learning to model user behavior, performing local calculations on the user device side, encrypting the behavior data using differential privacy technology, and managing the user data access rights through decentralized identity authentication are,

[0208] Perform federated learning calculations, define the local device model, expressed as:

[0209]

[0210] Among them, is the model parameter of device u at time t + 1, is the model parameter of device u at time t, and η is the learning rate. is the loss gradient,

[0211] The server aggregates multiple user updates:

[0212]

[0213] Among them, W t+1 is the global model parameter, U is the total number of devices,

[0214] Perform differential privacy encryption, add noise to the gradient, and the formula is:

[0215]

[0216] Among them, is the noise-added gradient, is the Gaussian noise,

[0217] Define the user identity hash h u Decentralized identity management:

[0218] h u = H(ID u , K),

[0219] Among them, h u is the user identity hash value, H(·) is the encryption hash function, ID u is the user unique identifier, and K is the key;

[0220] Specifically, on the basis of personalization, a federated learning and privacy protection mechanism is introduced here to ensure data security; Federated learning enables user devices to locally update the behavior model, and the server aggregates the updates of multiple users to improve the accuracy of personalized modeling; At the same time, differential privacy technology is adopted to protect user data by adding noise to the gradient and prevent privacy leakage; Decentralized identity management is based on hash encryption to improve the security of user identity data;

[0221] Combining data security and personalized optimization, the recommendation ability of the system in a privacy protection environment is improved.

[0222] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. An AI-based DMP data management method, characterized in that: including, Step S1: Collect the behavioral data of the user on multiple devices, and perform standardization processing on the behavioral data to obtain standardized data; Step S2: Based on the standardized data, use device fingerprint recognition and a deep learning model to perform user identity matching to obtain a matching result; Step S3: Based on the matching result, construct a user profile, and dynamically update the user interest tags by combining time series analysis and reinforcement learning methods to obtain an updated user profile; The time series analysis is based on the LSTM network to learn the user behavior trend; The reinforcement learning method adaptively optimizes the user feature weights; Step S4: Based on the updated user profile, optimize the personalized recommendation strategy, dynamically adjust the recommended content, and obtain a recommendation result; Step S5: Based on the recommendation result, use federated learning to perform user behavior modeling, perform local calculations on the user device side, encrypt the behavioral data using differential privacy technology, and manage the user data access rights through decentralized identity authentication.

2. The method for managing DMP data based on artificial intelligence according to claim 1, wherein: The behavioral data includes web browsing records, application usage, mouse trajectories, click behaviors, dwell times, and interaction events.

3. The method for DMP data management based on artificial intelligence according to claim 2, wherein: The step of collecting the behavioral data of the user on multiple devices and performing standardization processing on the behavioral data is as follows: Monitor the behavioral data of the user on multiple devices, collect the behavioral data through a cross-platform development kit SDK, browser plug-in, and application logs, and upload it to a central server or edge computing node; Perform preprocessing on the central server or edge computing node: Remove abnormal data points, perform normalization processing on continuous numerical data, and use Min-Max standardization; Map the high-dimensional data of mouse trajectories and click behaviors into feature vectors, and convert them into discrete categories using the equal-frequency binning or equal-width binning method; For cross-device behavioral data, align it based on the timestamp; Generate a unique user identifier UID for associating data between different devices, and store the standardized data in a distributed database.

4. The method for DMP data management based on artificial intelligence according to claim 3, characterized in that: The device fingerprint recognition includes the operating system type, screen resolution, browser configuration, and network environment; The deep learning model is based on the Transformer architecture to learn the user's cross-device behavior patterns, and optimize the matching accuracy through multi-task learning.

5. The method for managing DMP data based on artificial intelligence according to claim 4, wherein: Based on the standardized data, the step of using device fingerprint recognition and a deep learning model to perform user identity matching is as follows: Perform device fingerprint extraction, and represent the characteristics of the device fingerprint f as: f = (f1, f2, f3, …, f n ) where f1 represents the operating system type, f2 represents the screen resolution, f3 represents the browser configuration, f4 represents the network environment, and n represents the total number of device fingerprint features; Perform user behavior feature extraction, and define the user behavior vector v: v = (v1, v2, v3, …, v m ) where v1 represents the web browsing pattern, v2 represents the mouse movement pattern, v3 represents the click pattern, v4 represents the device switching pattern, and m represents the total number of behavior features; Use a deep learning model based on Transformer for identity matching, and define the input feature as X: X = [f, v], where X is the comprehensive behavior feature matrix of the user, f is the device fingerprint feature vector, and v is the user behavior feature vector. Encode through the input Transformer encoding layer, and the encoding process is expressed as: H = TransformerEncoder(X, W), where Transformer Encoder is the Transformer encoding function for performing representation learning on the input features, W is the Transformer training parameter matrix, and H represents the encoded behavioral feature vector. The calculation process of Transformer encoding is: H = SelfAttention(X) + FeedForward(SelfAttention(X)), where the self-attention layer SelfAttention calculates the correlation between different positions in the input sequence: where Q = XW Q is the query matrix, K = XW K is the key matrix, V = XW V is the value matrix, W Q , W K , W V are training parameters, d k is the dimension of the key vector The features are further mapped by the feed-forward neural network FeedForward, and the formula is: FeedForward(x) = σ(W1x + b1)W2 + b2, where W1, W2 are weight matrices, b1, b2 are bias vectors, and σ is a non-linear activation function. Calculate the matching degree of two user data using cosine similarity, and the calculation formula is: Among them, S(u1, u2) represents the similarity between users u1 and u2, which are the feature vectors of the users respectively, · represents the dot product, and |H| represents the L2 norm of the vector. Set a threshold τ. When S(u1, u2) > τ, it is determined to be the same user, otherwise it is a different user. If S(u1, u2) > τ, then If S(u1, u2) ≤ τ, then Among them, is the matching result, where 1 indicates the same user, 0 indicates different users, and τ is the matching threshold set empirically.

6. The method for DMP data management based on artificial intelligence according to claim 5, characterized in that: The steps of constructing the user portrait and dynamically updating the user interest tags by combining time series analysis and reinforcement learning methods are as follows: Based on the user identity matching result in step S2, define the initial user portrait feature vector P u : P u = (p1, p2, …, p d ) Among them, P u represents the portrait vector of user u, and p1, p2, …, p d respectively represent user attributes in different dimensions, and d is the number of dimensions of the user portrait. Calculate the initial user portrait based on the aggregation of standardized behavioral data, and the formula is: where p i represents the initial weight of the user in the portrait dimension i, represents the behavioral characteristics of the user at time t, and T is the total length of the data collection time window. Use the LSTM long short-term memory network to perform time series modeling on the user portrait, and define the input as: H t = LSTM(H t-1 , v t ), Among them, H t is the hidden state at time t, representing the user's interest deviation at that moment, H t-1 is the hidden state at the previous moment, v t is the behavioral feature collected at the current time t The LSTM update formula is: H t = σ(W h H t-1 + W v v t + b h ), Among them, W h , W v is the LSTM training weight, b h is the bias parameter, and σ is the activation function Dynamically adjust the weights of user portraits using reinforcement learning methods, and define the reward function as R t : where, R t is the total reward at time t, w j is the weight of the j-th dimension of the image feature, is the j-th dimension of the user portrait, Use Q learning to update the user portrait, and the update formula is: Among them, Q(s t , a t ) represents the Q-value of executing action a t under state s t , α is the learning rate, r t is the current reward, γ is the discount factor, s t+1 is the state at the next moment, a′ is the possible action at the next moment The Q-value for the next optimal action, The final user portrait vector is P' u : P′ u =(p′1, p′2, …, p′ d ) where p′ j = p j + Δp j , Δp j = β·(H t + Q(s t , a t )) where p′ j is the updated user portrait feature, Δp j is the portrait adjustment amount, and β is the smoothing factor.

7. The method for managing DMP data based on artificial intelligence according to claim 6, characterized in that: The recommended strategies include content recommendation, advertisement placement, and user behavior prediction.

8. The method for managing DMP data based on artificial intelligence according to claim 7, wherein: Based on the updated user portrait, the steps of optimizing the personalized recommendation strategy and dynamically adjusting the recommended content are as follows: Based on the user profile P′ u Calculate the recommendation score, and the formula is: Among them, S i represents the preference score of the user for item i, and r j i is the relevance of item i in the user profile dimension j. Use Top K sorting to select recommended items: I * = argmax I (S), Among them, I * is the set of optimal recommended items, I is the set of candidate recommended items, and argmax takes the item index corresponding to the maximum score. Update the recommendation weights according to user feedback, and the update formula is: p′ j = p′ j + λ(F - E), where F is the actual feedback, E is the expected feedback, and λ is the feedback update rate.

9. The method for DMP data management based on artificial intelligence according to claim 8, wherein: The optimized user behavior model is stored in the DMP system.

10. A DMP data management method based on artificial intelligence according to claim 9, characterized in that: The steps of using federated learning to model user behavior, performing local calculations on the user device side, encrypting the behavior data using differential privacy technology, and managing user data access rights through decentralized identity authentication are as follows: Perform federated learning calculations, and define the local device model, which is expressed as: where, is the model parameter of device u at time t+1, is the model parameter of device u at time t, and η is the learning rate, is the loss gradient, The server aggregates multiple user updates: Among them, W t+1 is the global model parameter, U is the total number of devices, Perform differential privacy encryption, add noise to the gradient, and the formula is: Among them, is the noise-added gradient, is the Gaussian noise, Define the user identity hash h u Decentralized identity management: h u = H(ID u , K), where h u is the user identity hash value, H(·) is the cryptographic hash function, ID u is the unique user identifier, and K is the secret key.

Images