An advertisement recommendation method and device based on data fusion

Abstract

The application discloses an advertisement recommendation method and device based on data fusion, and relates to the technical field of advertisement recommendation.The method comprises the following steps: collecting user portrait, behavior and context multi-source data; performing structured, unstructured and sequence feature processing on the data to generate structured, text, image and sequence feature vectors; adopting a model-level fusion mode, dynamically weighting and fusing multi-modal features through an expert model and a gate network to generate unified user and advertisement representation vectors; inputting the two vectors into a deep network for interaction and nonlinear transformation, outputting a predicted click rate through a Sigmoid function; and finally, combining advertisement bidding to calculate a ranking score and generating a final recommendation list.The application provides an advertisement recommendation method and device which can fuse multi-dimensional dynamic data, realize real-time interest change sensing and precise matching.

Description

Technical Field

[0001] This invention belongs to the field of computer information technology, and specifically relates to an online advertising recommendation technology, and more particularly to an advertising recommendation method and device based on multi-source data fusion. Background Technology

[0002] With the rapid development of internet and mobile internet technologies, online advertising has become a major form of digital marketing. Accurate ad recommendations are crucial for improving user experience, enhancing ad performance, and increasing platform revenue. Traditional ad recommendation methods typically rely on static user profile modeling and collaborative filtering based on basic user attributes and limited historical behavioral data. This approach struggles to capture the dynamic changes in user interests over time and context, and fails to fully integrate multimodal content information such as text and images from advertisements.

[0003] Therefore, existing technologies often result in recommendations that are out of touch with users' actual current needs, lacking accuracy and timeliness. This invention aims to address these problems by proposing an advertising recommendation method and apparatus that integrates multi-dimensional dynamic data, senses changes in interests in real time, and achieves precise matching. Summary of the Invention

[0004] This invention aims to provide an advertising recommendation method and device based on data fusion. To solve the above-mentioned technical problems, the technical solution provided by this invention is as follows:

[0005] An ad recommendation method based on data fusion includes the following steps:

[0006] S1: Multi-source data collection; the multi-source data includes user profile data, user behavior data, and contextual environment data;

[0007] S2: Perform feature processing based on the multi-source data to generate feature vectors; the feature processing includes structured feature processing, unstructured feature processing, and sequence feature processing; the feature vectors include: structured feature vectors, text feature vectors, image feature vectors, and sequence feature vectors;

[0008] S3: Perform model-level fusion based on the feature vectors to generate a unified user representation vector and a unified advertisement representation vector;

[0009] S4: Perform deep matching and pCTR prediction based on the user unified representation vector and the advertisement unified representation vector;

[0010] S5: Calculate the ranking score of candidate ads based on the pCTR prediction and ad bid, and generate the final recommendation list.

[0011] Furthermore, the step S2 of generating the user feature vector includes:

[0012] S21: Perform structured feature processing on the multi-source data to generate structured feature vectors; the structured features include numerical features and categorical features;

[0013] S22: Perform unstructured feature processing on the multi-source data to generate unstructured feature vectors; the unstructured features include text features and image features;

[0014] S23: Based on the multi-source data, a gated loop unit is used to perform sequence feature processing to generate a sequence feature vector.

[0015] Furthermore, the generation of structured feature vectors in step S21 includes the following steps:

[0016] S211: Normalize the numerical features to generate a numerical feature vector;

[0017] S212: Perform one-hot encoding based on the categorical features to generate a categorical feature vector;

[0018] S213: Concatenate the numerical feature vector and the categorical feature vector to generate a structured feature vector.

[0019] Furthermore, the step S22, which generates unstructured feature vectors, includes the following steps:

[0020] S221: Perform pooling operation on the text data to generate a text feature vector;

[0021] S222: Input the image data into a pre-trained convolutional neural network for feature extraction and perform global average pooling to generate an image feature vector.

[0022] Furthermore, the step S23, which generates a sequence feature vector, includes the following steps:

[0023] S231: Based on the multi-source data, an initial vector sequence is formed by mapping it into a dense vector through an embedding layer;

[0024] S232: Input the initial vector sequence into the gated recurrent unit network, and calculate the reset gate, update gate and candidate hidden state in sequence;

[0025] S233: Calculate the hidden state at the current time step based on the reset gate, update gate, and candidate hidden states;

[0026] S234: Perform a weighted summation based on the hidden state to generate the sequence feature vector.

[0027] Furthermore, in step S3, model-level fusion is performed based on the feature vectors, and the steps include:

[0028] S31: Based on the feature vectors, input them into the expert models for processing to generate the corresponding expert model output vectors;

[0029] S32: Input the feature vector into the gated network and calculate the weights corresponding to each expert model using the Softmax function;

[0030] S33: Based on the output vectors of each expert model, perform a weighted summation with their corresponding weights to generate a unified user representation vector.

[0031] The unified representation vector for advertising is generated through a fusion architecture symmetric to the user representation.

[0032] Furthermore, in step S4, deep matching and pCTR prediction are performed based on the user unified representation vector and the advertisement unified representation vector. The steps include:

[0033] S41: Based on the user unified representation vector and the advertisement unified representation vector, perform feature interaction processing to generate a combined feature vector;

[0034] S42: Based on the combined feature vector, input the multilayer perceptron for deep nonlinear transformation to generate high-order interactive features;

[0035] S43: Based on the higher-order interaction features, click-through rate (CTR) is predicted using the Sigmoid activation function, and the predicted CTR is output.

[0036] Furthermore, in step S5, the ranking score of candidate ads is calculated based on the pCTR prediction and ad bid to generate the final recommendation list. This step includes:

[0037] S51: Generate a candidate advertisement set based on the multi-source data;

[0038] S52: Calculate the ranking score of the candidate advertisement, wherein the ranking score is set as the product of the predicted click-through rate and the advertisement bid;

[0039] S53: Based on the sorting scores, sort the ads in descending order and select the top K ads to generate a recommendation list.

[0040] An advertising recommendation device based on data fusion includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the steps of the advertising recommendation method based on data fusion as described in any one of claims 1 to 9.

[0041] The advantages of this invention are:

[0042] 1. This invention systematically collects and processes structured, unstructured, and sequential data of user profiles, behaviors, contexts, and advertising content, and adopts a model-level fusion mechanism to adaptively weight and fuse features from different modalities to generate a unified user or advertising representation vector that is comprehensive in information and rich in semantics, thus solving the problems of insufficient utilization of multi-source data and poor fusion effect.

[0043] 2. This invention models user historical behavior sequences by introducing sequence models such as gated loop units. This invention can effectively capture the temporal dependencies and dynamic evolution patterns of user interests, distinguish between long-term stable interests and short-term real-time intentions, making user profiles more three-dimensional and vivid, and significantly improving the accuracy of predicting users' future behavior. Attached Figure Description

[0044] Figure 1 This is a flowchart of steps S1-S5 in an advertising recommendation method and device based on a multi-dimensional user profile, according to an embodiment of this application.

[0045] Figure 2 This is a flowchart of steps S21-S23 in an advertising recommendation method and device based on a multi-dimensional user profile according to an embodiment of this application.

[0046] Figure 3 This is a flowchart of steps S211-S213 in an advertising recommendation method and device based on a multi-dimensional user profile according to an embodiment of this application.

[0047] Figure 4 This is a flowchart of steps S221-S222 in an advertising recommendation method and device based on a multi-dimensional user profile according to an embodiment of this application.

[0048] Figure 5 This is a flowchart of steps S231-S234 in an advertising recommendation method and device based on a multi-dimensional user profile according to an embodiment of this application.

[0049] Figure 6 This is a flowchart of steps S31-S31 in an advertising recommendation method and device based on a multi-dimensional user profile according to an embodiment of this application.

[0050] Figure 7 This is a flowchart of steps S41-S43 in an advertising recommendation method and device based on a multi-dimensional user profile according to an embodiment of this application.

[0051] Figure 8 This is a flowchart of steps S51-S53 in an advertising recommendation method and device based on a multi-dimensional user profile according to an embodiment of this application. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0053] This application discloses an advertising recommendation method based on data fusion, referring to... Figure 1 The method includes steps S1 to S5, as follows:

[0054] S1: Multi-source data acquisition.

[0055] Data is collected from multiple sources, including user clients, server logs, and business databases. This multi-source data comprises user profile data, user behavior data, and contextual environment data.

[0056] User profile data includes static and dynamic data. Static data includes information such as the user's age, gender, region, and occupation. Dynamic data includes long-term and short-term interests. Long-term interests include stable interests based on months or even years, while short-term interests include recent, volatile interests based on days or weeks.

[0057] User behavior data includes explicit user data and implicit user data. Explicit user data includes strong, deliberate signals of intent from users, such as clicks, purchases, favorites, and shares. Implicit user data includes potential interests or needs indirectly inferred from browsing time, page scrolling, and search keywords.

[0058] Contextual data includes time data, location data, and scene data. Time data includes the time when the user's behavior occurred; location data includes the user's GPS location, Wi-Fi environment, etc.; and scene data includes the specific scene in which the user's behavior occurred.

[0059] S2: Perform feature processing based on multi-source data to generate feature vectors. For example... Figure 2 As shown, steps S21-S23 are included:

[0060] Multi-source data is inherently diverse, heterogeneous, and disorganized. Feature engineering transforms data from various formats into a unified user feature vector. Feature engineering includes structured features, unstructured features, and sequence features.

[0061] S21: Perform structured feature processing based on multi-source data to generate unstructured feature vectors. For example... Figure 3 As shown, steps S211-S213 are included:

[0062] Structured feature processing includes normalization and one-hot encoding. Normalization scales the values ​​of raw data with large scale differences to a similar scale. One-hot encoding converts discrete class labels into numerical forms that the model can compute.

[0063] S211: Normalize the numerical features to generate numerical feature vectors.

[0064] Variables representing user or item attributes in real-valued form, directly collected from raw data, are typically designated as numerical features, such as a user's age, income, product price, and historical click count. Since these raw features often have different dimensions and value ranges, they can lead to slow convergence, parameter estimation bias, and oversensitivity to large-scale features in machine learning models that rely on distance calculations or gradient optimization. Therefore, the raw data is linearly scaled to the [0, 1] range using a normalization formula, typically expressed as:

[0065] Where X is a value in the original data. It is the minimum value of this feature in the entire dataset. It is the maximum value of this feature across the entire dataset. (Calculated) It is a numerical feature vector generated after normalization.

[0066] S212: Perform one-hot encoding based on categorical features to generate categorical feature vectors.

[0067] Discrete variables representing categories or states of things extracted from the original data are typically designated as categorical features, such as a user's gender, city, or product category. Since models cannot directly process text labels, most categorical feature values ​​do not have an inherent order relationship. Therefore, one-hot encoding is needed to convert each categorical feature into a binary vector equal to the number of categories, where only the category position is 1 and the rest are 0.

[0068] For a categorical feature with N distinct values, let its category set be: C = { }

[0069] The one-hot encoding function expression is: OneHot:C→{0,1}

[0070] The expression means that the input comes from any category in set C. Output a binary vector of length N, where each element can only be 0 or 1, and each input category uniquely corresponds to an output vector.

[0071] For any category ∈C, its one-hot encoded feature vector It is an N-dimensional vector: =[ .

[0072] Each component of the vector Determined by the following formula: For N=1,2,......

[0073] S213: Concatenate numerical feature vectors and categorical feature vectors to generate structured feature vectors.

[0074] Numerical and categorical feature vectors are ultimately concatenated to form a fixed-dimensional vector representing the user's structured attributes. Since the numerical feature vectors have undergone normalization and the categorical feature vectors have undergone one-hot encoding, and Both are one-dimensional vectors, which can be... All elements are arranged in order ,Will All elements are arranged in order =[ ].

[0075] Will and The concatenated result is defined as a structured feature vector. The expression is: = +[ ]

[0076] Structured feature vectors Determined as = = Therefore, it can be determined

[0077] S22: Perform unstructured feature processing on multi-source data to generate unstructured feature vectors. For example... Figure 4 As shown, steps S221-S222 are included:

[0078] Unstructured features refer to semantic information extracted from raw data without a pre-defined format, including text features and image features. Text features use pooling operations to transform linguistic information such as user comments and search terms into numerical vectors that contain semantic meaning; image features use convolutional neural networks to extract visual patterns from advertising creatives and product images and compress them into dense vectors that represent visual concepts.

[0079] S221: Perform pooling operations on text data to generate text feature vectors.

[0080] Because machine learning models can only process numerical data and cannot directly understand text, they can be transformed into text feature vectors. These vectors allow the model to utilize semantic information for recommendation calculations. Generating text feature vectors is a complete transformation process from raw text to a numerical representation, aiming to convert unstructured text data into fixed-length numerical vectors that machine learning models can process.

[0081] The first stage is word vectorization. After word segmentation, the original text is broken down into independent word sequences. Given the original text T, each word is mapped to a dense vector using the pre-trained Word2Vec word embedding model, ultimately outputting an independent word sequence as […]. ,...... ].

[0082] Word2Vec is trained based on the Skip-gram architecture, and its objective function is to maximize the conditional probability of the context words.

[0083] Where w is the head word, c is the context word, D is the set of word pairs in the training corpus, and θ is the model parameters. The purpose of this expression is to find a set of model parameters. That is, the vector representation of all words, such that for each center word appearing in the training corpus V, the context word pair... , context words The conditional probability of occurrence given the headword w The sum of logarithms reaches its maximum value.

[0084] Using vectors The current value is calculated according to the Softmax formula. The expression is:

[0085] in, The vector representation of the center word. V is the vector representation of the context words, and V is the entire vocabulary. Used to calculate the score of a single context word c. This indicates that the entire vocabulary V is traversed. For each word, calculate its dot product with the head word w and take the exponent. Add up the index scores of all the words. The score of a single context word c is divided by the sum of the scores of all words, resulting in a value between 0 and 1, representing the relative probability of the context word c appearing given the head word w.

[0086] Calculated probability Compare the result with the actual situation (the context word c does indeed appear, and its probability should be 1), and calculate the loss, expressed as: Loss = -

[0087] Based on the loss expression, update the word vectors according to the following expression:

[0088]

[0089]

[0090] in, The learning rate controls the step size for each update. The update makes... and By getting closer to each other in the vector space, the calculated P(c|w) will increase slightly in the next calculation, and the loss will decrease accordingly.

[0091] After the model repeats this process on a massive number of word pairs, the word vectors gradually converge to stable values ​​through countless fine-tuning steps. Ultimately, each word... It is converted into a dense real number vector, forming a word vector sequence. ,...... ].

[0092] Next is the text vectorization stage, where pooling operations aggregate the variable-length word vector sequence into a fixed-dimensional text vector. For the input text T, its vectorized representation is: )

[0093] Where T is the input text, It is the vector representation of the i-th word.

[0094] The pooling operation primarily employs mean pooling, capturing the overall semantics through an arithmetic average. The calculation formula is as follows: .

[0095] Mean pooling formula This transforms a sequence of word vectors of indeterminate length [,......] into fixed-dimensional text feature vectors. .

[0096] in, This means representing all word vectors The semantic information is collected and superimposed, and the semantic fragments scattered in various words are integrated together through vector addition. This indicates that each word in the text is assigned an equal weight.

[0097] S222: Input the image data into a pre-trained convolutional neural network for feature extraction and perform global average pooling to generate image feature vectors.

[0098] Since images are unstructured data, their raw pixel matrices cannot be directly used by recommendation algorithms. Extracting image feature vectors from images can transform human-visually understandable pixel information into a numerical representation that can be processed by machine learning models. A ResNet pre-trained model is used to extract visual feature vectors from images. Its unique residual connection structure effectively solves the gradient vanishing problem in deep networks, allowing the network to be built deeper and learn from low-level edge textures to high-level semantic concepts.

[0099] Let the pre-trained ResNet model be a function

[0100] The input is an RGB three-channel image with height H and width W, and the output is a c-dimensional visual feature vector.

[0101] The original image is represented as a 3D pixel matrix (height H × width W × number of channels 3) and input into a pre-trained ResNet network. The network processes the image hierarchically through three convolutional layers: shallow convolutional kernels detect basic visual patterns, the middle layer combines these patterns to form more complex part features, and the deep layer further integrates them into high-level semantic concepts. Each residual block ensures effective gradient propagation through skip connections, enabling the network to learn the residual mapping between input and target features. This process is accomplished by a series of residual blocks, with the core operation of each block being:

[0102]

[0103] in, It is the first Output feature maps of each residual block This is the [number]th [item] flowing into the current residual block. Input data for the layer. It is a residual function, usually composed of three stacked convolutional layers. (...) This indicates a skip connection, directly adding the output of the residual function (...) to the original input. This is a non-linear activation function that passes the summed result to another non-linear activation function, introducing non-linearity into the output of the entire residual block. Through this layer-by-layer pass-through, the original image... It is gradually converted into high-level semantic feature maps. .

[0104] After the last residual block, the network outputs a 3D feature map. The feature map (h, w, C) obtained after convolution and pooling of the input image may still vary in spatial dimensions. To compress the three-dimensional tensor into a one-dimensional feature vector of fixed length C, global average pooling is used, expressed as:

[0105] Here, c represents the c-th feature channel. The c-th element of the final vector originates from the c-th feature channel. i and j represent traversing all heights and widths along that channel of the feature map. The average of all spatial positions (hw) for each channel c of the feature map is calculated, thereby generating a fixed-length image feature vector of length C from the 3D tensor.

[0106] S23: Based on multi-source data, a gated cyclic unit is used to process sequence features and generate sequence feature vectors. For example... Figure 5 As shown, steps S231-S234 are included:

[0107] S231: Based on multi-source data, the data is mapped into dense vectors through an embedding layer to form an initial vector sequence.

[0108] Sequence features utilize recurrent neural networks to model user historical behavior sequences, generating concise temporal representation vectors of dynamic interests. Discrete product ID sequences are converted into numerical sequences that the model can process. A user's historical behavior sequence T=20

[0109] Each product ID is mapped to a dense low-dimensional vector through an embedding layer, expressed as follows:

[0110] Where t=1,2,......,T represents the time step, and T is the sequence length.

[0111] GRU reads the behavior sequence step by step and updates its internal state using a gating mechanism to capture temporal dependencies. At each time step t, a GRU unit receives two inputs: the current item vector and the hidden state from the previous time step. Then, calculate and output the new hidden state according to the following steps.

[0112] S232: Input the initial vector sequence into the gated recurrent unit network, and calculate the reset gate, update gate and candidate hidden state in sequence;

[0113] The reset gate can effectively process sequential data and capture complex dependencies. Through a differentiable learning process, it dynamically and adaptively determines whether to retain or discard historical information from the previous time step. The expression is:

[0114] Here, represents the input vector at the current time step t, representing the user's immediate behavior at the current moment. represents the hidden state at the previous time step t−1, containing all historical information from the beginning of the sequence up to time t−1. and This represents the weight matrix, which is responsible for applying a linear transformation to the current input. Responsible for processing the current input Perform a linear transformation. represents the bias term, used to adjust the model's output baseline and increase the model's flexibility. represents the sigmoid activation function, used to compress the result of the linear transformation into the (0, 1) interval.

[0115] Will come from the current input This information is then integrated with historical context to form a comprehensive signal. The signal is then converted into a forgetting control vector between 0 and 1.

[0116] The update gate, by reconciling old and new information, enables the model to simultaneously capture short-term patterns and long-term dependencies. It allows for rapid response to sudden changes in user interests while maintaining the stability of core interests. This determines the proportion of old states and candidate new information in the new hidden state, expressed as:

[0117] Where and represent the weight matrix, responsible for weighting the current input. Perform linear transformation. Responsible for processing the current input Perform a linear transformation. It is responsible for analyzing the impact of the current input on the state update and for analyzing the continued value of historical states. The value and reliability of information used to assess current behavior Used to assess the stability of historical states and the degree to which they need to be preserved.

[0118] Importance of the current input and historical status The value of this forms a signal for updated decision-making. This decision signal is converted into an update control vector between 0 and 1. .

[0119] The candidate hidden state is generated based on the current input and filtered historical information. By resetting the gate's filtering, the current behavior and relevant historical context are integrated to generate a rich and context-sensitive alternative state, expressed as:

[0120] in, This indicates the useful historical information retained after the gate is reset and filtered. This represents a linear transformation of the filtered historical information. This represents the linear transformation of the current input. This represents the hyperbolic tangent activation function, used to compress the results to the (-1, 1) interval, ensuring numerical stability and providing a nonlinear transformation.

[0121] The formula is derived from Extract current behavior information. Extract relevant historical information and integrate it. The fusion result is subjected to a nonlinear transformation to generate a normalized candidate state. .

[0122] S233: Calculate the hidden state at the current time step based on the reset gate, update gate, and candidate hidden states.

[0123] Formula used:

[0124] in, This indicates the old state, referring to all historical memories and contextual interests up to the previous moment. This represents the candidate hidden state, which refers to new information generated based on the current input and the filtered history. This indicates that old information from the old state will be carried over to the new state. This represents the adoption of new information from the candidate states into the new state. The retained old information is summed with the adopted new information to form a completely new, unified hidden state. .

[0125] S234: Perform weighted summation based on the hidden state to generate a sequence feature vector.

[0126] From t=1 to t=T, GRU processes the entire sequence step by step, generating a hidden state sequence. ], each Each sequence contains information up to the current moment, serving as a summary and overview of the entire sequence history. The expression is:

[0127] S3: Model-level fusion based on feature vectors to generate unified user representation vectors and unified advertisement representation vectors. For example... Figure 6 As shown, steps S31-S33 are included:

[0128] By integrating user feature vectors from different modalities, such as structured feature vectors, text feature vectors, image feature vectors, and sequence feature vectors, a more comprehensive unified user feature vector is formed. This enables the correlation between different data sources, combining user browsing text, viewed images, and click sequences, which helps capture complex and dynamic interest intentions, thereby improving the perception capabilities and recommendation accuracy of the recommendation system. Multimodal data fusion includes feature-level fusion, model-level fusion, and decision-level fusion.

[0129] Feature-level fusion has significant limitations, treating all features equally and failing to dynamically adjust the importance of different modalities based on specific tasks and input data. Therefore, model-level fusion is further employed. This involves using a parallelized expert network to deeply mine user features, then leveraging a gating network to achieve context-aware dynamic weight allocation, and finally generating a unified user representation vector through weighted summation.

[0130] S31: Input the feature vectors into the expert model respectively for processing, and generate the corresponding output vector of the expert model.

[0131] An expert model network is used. The original feature vector is input into the expert model MLP network, and after a series of nonlinear transformations, a completely new vector representation is output. The definition formula of the expert model is:

[0132] in, It is the mathematical representation of the m-th expert model, representing a complete transformation process. This represents the input feature vector. This means receiving an input feature vector and performing a nonlinear transformation to generate a new feature vector.

[0133] S32: Input the feature vector into the gating network and calculate the weights corresponding to each expert model through the Softmax function.

[0134] The gating network dynamically calculates the weight of each expert in the final decision based on the current input features. The formula for calculating the gating weight is as follows:

[0135] M={ , , ,sequence}

[0136] in, This represents a weight vector used to evaluate the value of each dimension in the original features. This indicates the baseline adjuster, used to adjust the baseline for the modality score. Representing a linear transformation, it summarizes the high-dimensional original feature vector into a scalar score. Since the Softmax function requires positive inputs, Used to Map to the interval (0, +∞). The normalized denominator is represented by the Softmax function. M represents the set of all modes. This represents the final generated gating weights. Satisfying 0< <1 and =1.

[0137] S33: Based on the output vectors of each expert model, perform a weighted summation with their corresponding weights to generate a unified user representation vector.

[0138] The outputs of each expert are weighted and combined according to their importance to form the final feature representation user vector. The expression is:

[0139] in, This represents the output characteristics of each expert. Indicates the corresponding gating weight. This indicates a weighted sum. This represents the unified representation after final fusion.

[0140] S4: Deep matching and pCTR prediction based on user unified representation vectors and ad unified representation vectors. For example... Figure 7 As shown, steps S41-S43 are included:

[0141] The unified representation vector A of an advertisement is generated through a multimodal fusion architecture symmetric to the user profile: First, the structured attributes, text descriptions, and visual creatives of the advertisement are extracted and vectorized respectively. Then, these heterogeneous features are input into an expert hybrid model, and the multi-source features are integrated into a unified d-dimensional dense vector through a gated weighted fusion mechanism. This vector condenses the complete semantic information of the advertisement and together with the user representation vector, it forms the semantic space basis for deep matching.

[0142] S41: Based on the unified representation vector of users and the unified representation vector of advertisements, perform feature interaction processing to generate a combined feature vector.

[0143] The generated unified user representation vector u and unified ad representation vector a are input into a deep neural network. Let the unified user representation vector be... The unified advertising representation vector is set as .

[0144] To enable the model to learn the complex relationships between users and ads, rather than just individual features, it's necessary to construct interaction features. A common approach to combining inputs is to concatenate multiple interaction features with the original vector:

[0145] in, Represents a unified user vector. This represents a uniform ad vector. The Hadamard product represents the multiplication of each dimension of the user vector and the ad vector: This means connecting the three vectors u, a, and u⊙a end to end along the dimensional direction. This represents the final input vector formed after concatenation, with a dimension of 3d.

[0146] S42: Based on the input of combined feature vectors into a multilayer perceptron, a deep nonlinear transformation is performed to generate high-order interactive features.

[0147] Interaction features The input is fed into a multilayer perceptron for deep nonlinear transformation to learn complex matching patterns between users and advertisements.

[0148]

[0149]

[0150]

[0151] Where L is the number of hidden layers. This represents the input layer, used to receive the synthesized feature vectors from the feature interaction layer. As input to the entire network. This represents the weight matrix, used to linearly combine features from the previous layer to discover higher-order cross features useful for predicting pCTR. This represents the bias vector, used to adjust the activation threshold. This is used to introduce nonlinearity. The model automatically learns from the original interactive features through multiple layers of nonlinear transformation. The complex functional relationship between the final pCTR prediction and the final prediction.

[0152] S43: Based on high-order interaction features, click-through rate (CTR) is predicted using the Sigmoid activation function, and the predicted CTR is output.

[0153] The final layer uses the Sigmoid activation function to compress the output to the (0,1) interval, outputting the predicted click-through rate (pCTR). The Sigmoid function can smoothly map any real-valued output from the last layer of the neural network to the (0,1) interval, which is beneficial for use as a probability score in subsequent ad ranking calculations. The expression is:

[0154]

[0155] Where pCTR represents the probability predicted by the model that user u will click on advertisement a. and These represent the weight matrix and bias vector of the output layer, respectively. This represents the output feature of the last hidden layer. This represents the Sigmoid activation function. This represents an exponential function.

[0156] The hidden layer features are mapped to unnormalized predicted scores. The Sigmoid function converts the scores in the real field into probability values ​​in the interval (0,1). The differentiable S-shaped curve ensures that the model can be trained end-to-end through the backpropagation algorithm.

[0157] The final output pCTR value is the probability that user u will click on ad a, as predicted by the model. This probability value will serve as the core feature in the subsequent ranking stage, and together with commercial factors such as ad bidding, it will determine the final display order of the ads.

[0158] S5: Calculate the ranking score of candidate ads based on pCTR prediction and ad bids, and generate the final recommendation list. For example... Figure 8 As shown, steps S51-S53 are included:

[0159] S51: Generate a candidate ad set based on multi-source data.

[0160] The system quickly selects a reasonably sized set of candidate ads from a vast ad library based on user profiles, ad targeting rules, and real-time strategies. This ensures the efficiency of subsequent calculations.

[0161] S52: Calculate the ranking score of the candidate ads, which is the product of the predicted click-through rate and the ad bid.

[0162] For the predicted pCTR scores of ads in the candidate set, combined with the ad bids, a ranking algorithm is used to rank the candidate ads, expressed as follows:

[0163] in, This represents the ad clicks predicted by the S4 deep matching model. The probability represents the relevance of the advertisement to the user and the user experience. This indicates the advertiser's bid for this exposure or click, representing the platform's immediate revenue. This indicates the expected return on the ad's exposure and is a core indicator for balancing relevance and return.

[0164] S53: Based on the sorting scores in descending order, select the top K ads to generate a recommendation list.

[0165] According to the score Sort all candidate ads in descending order using the expression: SortedList = argsort({ , , ...... })

[0166] From the sorted list, select Top-K ads (K=5 ads) to generate the final recommendation list.

[0167] The value of K can be selected from 5 to 10 as needed.

[0168] An advertising recommendation device based on data fusion includes a processor and a memory, the memory storing a computer program.

[0169] When the processor executes the computer program, it implements an advertising recommendation method based on data fusion as described in any of the foregoing embodiments, including the following steps:

[0170] S1: Multi-source data acquisition;

[0171] S2: Perform feature processing based on multi-source data to generate structured feature vectors, text feature vectors, image feature vectors, and sequence feature vectors;

[0172] S3: Model-level fusion based on feature vectors to generate unified user representation vectors and unified advertising representation vectors;

[0173] S4: Perform deep matching and predict click-through rate calculation based on user unified representation vector and ad unified representation vector;

[0174] S5: Calculate the ranking score of candidate ads based on the predicted click-through rate and ad bid, and generate the final recommendation list.

[0175] This device can be integrated into servers, cloud platforms, or advertising recommendation systems to achieve efficient and personalized advertising recommendations.

[0176] In this embodiment, the aforementioned data fusion-based advertising recommendation method can be deployed on a server cluster or cloud computing platform. Through a distributed computing framework, it processes massive amounts of user and advertising data to achieve high concurrency and low latency real-time recommendations. This method can be widely applied to advertising systems on e-commerce, content information, social media, and online video platforms. By dynamically capturing users' multi-dimensional interests and deeply matching them with advertising content, it significantly improves ad click-through rates and conversion rates, while simultaneously optimizing platform revenue and user experience.

[0177] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An advertising recommendation method based on data fusion, characterized in that, Including the following steps: S1: Multi-source data collection; the multi-source data includes user profile data, user behavior data, and contextual environment data; S2: Perform feature processing based on the multi-source data to generate feature vectors; The feature processing includes structured feature processing, unstructured feature processing, and sequence feature processing; The feature vectors include: structured feature vectors, text feature vectors, image feature vectors, and sequence feature vectors; S3: Perform model-level fusion based on the feature vectors to generate a unified user representation vector and a unified advertisement representation vector; S4: Perform deep matching and pCTR prediction based on the user unified representation vector and the advertisement unified representation vector; S5: Calculate the ranking score of candidate ads based on the pCTR prediction and ad bid, and generate the final recommendation list.

2. The method according to claim 1, characterized in that, The step S2, which generates the user feature vector, includes the following steps: S21: Perform structured feature processing on the multi-source data to generate structured feature vectors; the structured features include numerical features and categorical features; S22: Perform unstructured feature processing on the multi-source data to generate unstructured feature vectors; the unstructured features include text features and image features; S23: Based on the multi-source data, a gated loop unit is used to perform sequence feature processing to generate a sequence feature vector.

3. The method according to claim 1, characterized in that, The step S21, which generates structured feature vectors, includes the following steps: S211: Normalize the numerical features to generate a numerical feature vector; S212: Perform one-hot encoding based on the categorical features to generate a categorical feature vector; S213: Concatenate the numerical feature vector and the categorical feature vector to generate a structured feature vector.

4. The method according to claim 1, characterized in that, The step S22, which generates unstructured feature vectors, includes the following steps: S221: Perform pooling operation on the text data to generate a text feature vector; S222: Input the image data into a pre-trained convolutional neural network for feature extraction and perform global average pooling to generate an image feature vector.

5. The method according to claim 1, characterized in that, The step S23, which generates a sequence feature vector, includes the following steps: S231: Based on the multi-source data, an initial vector sequence is formed by mapping it into a dense vector through an embedding layer; S232: Input the initial vector sequence into the gated recurrent unit network, and calculate the reset gate, update gate and candidate hidden state in sequence; S233: Calculate the hidden state at the current time step based on the reset gate, update gate, and candidate hidden states; S234: Perform a weighted summation based on the hidden state to generate the sequence feature vector.

6. The method according to claim 1, characterized in that, Step S3, which involves model-level fusion based on the feature vector, includes the following steps: S31: Based on the feature vectors, input them into the expert models for processing to generate the corresponding expert model output vectors; S32: Input the feature vector into the gated network and calculate the weights corresponding to each expert model using the Softmax function; S33: Based on the output vectors of each expert model, a weighted sum is performed on their respective weights to generate a unified user representation vector; The unified representation vector for advertising is generated through a fusion architecture symmetric to the user representation.

7. The method according to claim 1, characterized in that, Step S4, which involves deep matching and pCTR prediction based on the user unified representation vector and the advertisement unified representation vector, includes the following steps: S41: Based on the user unified representation vector and the advertisement unified representation vector, perform feature interaction processing to generate a combined feature vector; S42: Based on the combined feature vector, input the multilayer perceptron for deep nonlinear transformation to generate high-order interactive features; S43: Based on the higher-order interaction features, click-through rate (CTR) is predicted using the Sigmoid activation function, and the predicted CTR is output.

8. The method according to claim 1, characterized in that, Step S5, which calculates the ranking score of candidate ads based on the pCTR prediction and ad bid, and generates the final recommendation list, includes the following steps: S51: Generate a candidate advertisement set based on the multi-source data; S52: Calculate the ranking score of the candidate advertisement, wherein the ranking score is set as the product of the predicted click-through rate and the advertisement bid; S53: Based on the sorting scores, sort the ads in descending order and select the top K ads to generate a recommendation list.

9. An advertising recommendation device based on data fusion, characterized in that, It includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the steps of the data fusion-based advertising recommendation method as described in any one of claims 1 to 8.

Images