E-commerce audience expansion and circle selection method based on similar crowd discovery

Abstract

This invention relates to the field of e-commerce technology, and more specifically, to a method for expanding and selecting e-commerce audiences based on similarity group discovery. The method includes: collecting samples to be cleaned from multiple data sources; selecting corresponding modal encoders and modal enhancement strategies based on data source identifiers; achieving cross-modal feature space alignment by constructing a set of positive sample pairs and a set of negative sample pairs and performing comparative learning training; obtaining a final group set through multi-table joint candidate set construction and fine clustering; solving for the hyperspherical prototype center using Riemann gradient descent for each final group; labeling normal or abnormal samples based on comparison results; and outputting the cleaned dataset and the abnormal sample set. This invention solves the problem of incomparability of feature spaces in multi-source heterogeneous data, improves the efficiency of large-scale data processing, maintains the geometric consistency of the manifold space, and achieves adaptive determination of the anomaly detection threshold.

Description

Technical Field

[0001] This invention belongs to the field of e-commerce technology, specifically relating to a method for expanding and selecting e-commerce audiences based on the discovery of similar groups. Background Technology

[0002] Existing data cleaning techniques can be broadly categorized into three types: rule-based methods, statistical methods, and machine learning-based methods. Rule-based methods rely on domain experts to define data quality constraints, including field format validation, numerical range checks, uniqueness constraints, and referential integrity verification. These methods are effective at identifying format errors in structured data, but struggle to capture semantic-level data quality issues. Furthermore, the formulation and maintenance of rules require significant manual intervention and are difficult to adapt to dynamic changes in data patterns. Statistical methods utilize the statistical distribution characteristics of data to identify outliers. Typical techniques include outlier detection based on Mahalanobis distance, density estimation based on local outlier factors, and anomaly scoring based on isolated forests. These methods can automatically detect outliers that deviate from the overall distribution, but they typically assume that the data follows a Gaussian distribution or other specific parametric distributions. However, actual multi-source data often exhibits complex non-parametric distribution characteristics, making it difficult to guarantee the accuracy of the detection results.

[0003] In recent years, deep representation learning techniques have been widely introduced into the field of data quality assessment. Autoencoders learn a low-dimensional representation of data by minimizing reconstruction error, using this error as the basis for anomaly scoring. Variational autoencoders introduce probability distribution constraints in the latent space, utilizing the likelihood probability of samples for anomaly detection. Contrastive learning, as an emerging self-supervised representation learning paradigm, constructs positive and negative sample pairs to optimize the feature encoder, ensuring that semantically similar samples are close to each other in the feature space, while semantically unrelated samples are far apart. The SimCLR framework uses data augmentation to generate different views of the same sample as positive sample pairs, the MoCo framework introduces a momentum encoder and a negative sample queue to improve training stability, and the CLIP framework achieves cross-modal feature alignment for both image and text modalities. However, existing contrastive learning methods are mainly designed for single-modal or bimodal data, lacking a unified processing framework for multiple heterogeneous modalities such as images, text, and sensor signals, making it difficult to effectively support joint cleaning tasks for multi-source heterogeneous data. Regarding efficient processing of large-scale data, locality-sensitive hashing accelerates near-nearest neighbor retrieval by mapping high-dimensional feature vectors to low-dimensional hash codes. Locality-Sensitive Hashing (LSH) based on random hyperplane projection offers theoretical guarantees in preserving angular similarity and is widely used in image retrieval, document deduplication, and recommender systems. However, the recall rate of a single hash table is limited by the randomness of the projection; similar samples may be grouped into different hash buckets due to the random selection of the projection direction. While multi-table joint strategies can alleviate the recall instability problem, existing methods typically use LSH as an independent retrieval component, lacking deep collaborative design with downstream clustering and anomaly detection stages. Summary of the Invention

[0004] The main objective of this invention is to provide an e-commerce audience expansion and selection method based on similar group discovery. It solves the problem of incomparability of feature spaces of multi-source heterogeneous data by cross-modal contrastive learning, significantly improves the processing efficiency of large-scale data by adopting a locality-sensitive hash multi-table joint strategy, maintains the geometric consistency of manifold space by using hyperspherical geodesic distance and Riemann gradient descent, and achieves adaptive determination of anomaly detection threshold based on group-level conformal prediction and provides statistical reliability guarantee.

[0005] To solve the above problems, the technical solution of the present invention is implemented as follows:

[0006] E-commerce audience expansion and targeting methods based on similar demographics include the following steps:

[0007] Step 1: Select users who meet the preset conversion conditions from the user database of the e-commerce platform as seed users, obtain the behavioral event sequence of each seed user, extract features from the behavioral event sequence and input it into the sequence encoder to obtain the seed user representation vector corresponding to the seed user, and construct the seed representation set by combining all the seed user representation vectors.

[0008] Step 2: Construct a conditional diffusion generation model. The conditional diffusion generation model includes a forward noise injection module, a reverse denoising generation module, a generation validity gating module, and a diversity constraint module. The forward noise injection module gradually adds noise to the seed user representation vector to obtain a full noise latent vector. The reverse denoising generation module gradually denoises the initial noise vector with the seed user representation vector as a condition to obtain candidate virtual representation vectors. The generation validity gating module and the diversity constraint module screen the candidate virtual representation vectors to obtain virtual similar user representation vectors. All virtual similar user representation vectors constitute a virtual user representation set.

[0009] Step 3: Merge the seed representation set and the virtual user representation set to obtain the enhanced query representation set. Based on the enhanced query representation set, perform an approximate nearest neighbor search in the vector retrieval index to obtain the initial candidate user set. Perform filtering and constraint processing on the initial candidate user set in sequence to obtain the target audience user set. Adaptively adjust the selection range according to the conversion status data returned by the campaign.

[0010] Furthermore, in step 1, the preset conversion condition is that the purchase of goods is completed within the current statistical period and the purchase amount reaches the preset amount threshold; the method for generating the behavioral event sequence is as follows: obtain the behavioral event records generated by the seed user within the current statistical period. The behavioral event records include product browsing events, product adding to cart events, product favorite events, order submission events, payment completion events, and product review events. They are arranged in ascending order according to the occurrence timestamp of each behavioral event record to form a behavioral event sequence.

[0011] Furthermore, in step 1, the method for feature extraction of the behavioral event sequence is as follows: for each behavioral event record in the behavioral event sequence, extract the product category identifier, product price range identifier, and product brand identifier of the product associated with the behavioral event record, concatenate the product category identifier, product price range identifier, and product brand identifier into an event feature vector, arrange all event feature vectors in chronological order to form an event feature vector sequence, and input the event feature vector sequence into the sequence encoder.

[0012] Furthermore, in step 2, the execution process of the forward noise injection module is as follows: the seed user representation vector is used as the original input vector, the total number of diffusion steps is set to a positive integer T, a noise intensity sequence containing T elements and monotonically increasing from step 1 to step T is pre-constructed, and noise superposition operations are performed sequentially from step 1 to step T. In the noise superposition operation of step k, the sampled noise vector of step k is obtained from the standard normal distribution, and the sampled noise vector of step k is multiplied by the kth noise intensity value in the noise intensity sequence to obtain the scaled noise vector of step k. The intermediate vector output of step k-1 is added to the scaled noise vector of step k to obtain the intermediate vector output of step k. The intermediate vector output of step 0 is the original input vector, and the intermediate vector output of step T is used as the full noise latent vector.

[0013] Furthermore, in step 2, the inverse denoising generation module includes a denoising neural network. The denoising neural network adopts a multilayer perceptron structure, which includes an input concatenation layer, a first hidden layer, a second hidden layer, a third hidden layer, and an output mapping layer connected in series. The input concatenation layer is used to concatenate the vector to be denoised, the conditional vector, and the step encoding vector into a single input vector. The first, second, and third hidden layers all contain linear transformation units and nonlinear activation units. The output mapping layer maps the output of the third hidden layer into a noise estimation vector.

[0014] Furthermore, in step 2, the execution process of the reverse denoising generation module is as follows: select one seed user representation vector from the seed representation set as the condition vector, sample the initial noise vector from the standard normal distribution and use it as the denoising vector to be denoised in step T, perform denoising operations sequentially from step T to step 1, in the denoising operation of step j, perform sinusoidal position encoding on the current step number j to obtain the step number encoding vector, input the denoising vector to be denoised in step j, the condition vector, and the step number encoding vector into the denoising neural network to output the noise estimation vector of step j, multiply the noise estimation vector of step j with the jth noise intensity value in the noise intensity sequence to obtain the scaled noise estimation vector of step j, subtract the scaled noise estimation vector of step j from the denoising vector to be denoised in step j-1 to obtain the denoising vector to be denoised in step j-1, and use the denoising vector to be denoised in step 0 as the candidate virtual representation vector.

[0015] Furthermore, in step 2, the execution process of the validity gating module is as follows: calculate the cosine similarity between the candidate virtual representation vector and the condition vector. If the cosine similarity between the candidate virtual representation vector and the condition vector is less than the preset lower limit threshold for validity similarity, discard the candidate virtual representation vector. If the cosine similarity between the candidate virtual representation vector and the condition vector is greater than or equal to the preset lower limit threshold for validity similarity, further calculate the average cosine similarity between the candidate virtual representation vector and all seed user representation vectors in the seed representation set. If the average cosine similarity is less than the preset lower limit threshold for distribution consistency, discard the candidate virtual representation vector. If the average cosine similarity is greater than or equal to the preset lower limit threshold for distribution consistency, determine that the candidate virtual representation vector passes the validity gating.

[0016] Furthermore, in step 2, the execution process of the diversity constraint module is as follows: for each seed user representation vector, the inverse denoising generation module and the generation validity gating module are repeatedly executed until a preset number of M candidate virtual representation vectors that pass the validity gating are obtained. The M candidate virtual representation vectors that pass the validity gating are used to form a vector group to be screened. The cosine similarity between any two candidate virtual representation vectors in the vector group to be screened is calculated. If the cosine similarity between any two candidate virtual representation vectors is greater than the preset diversity similarity upper limit threshold, one of the two candidate virtual representation vectors is removed, and the remaining candidate virtual representation vectors in the vector group to be screened are retained as virtual similar user representation vectors.

[0017] Furthermore, in step 3, the filtering process includes deduplication, frequency control, and cold start representation compensation. Deduplication involves retaining only one record for each candidate user that appears repeatedly. Frequency control involves removing candidate users whose cumulative number of times they have been reached by advertisements within the current statistical period has reached a preset frequency limit. Cold start representation compensation involves identifying cold start users whose number of behavioral event records is less than a preset minimum record threshold, querying the user profile database for the preferred product category set of active user groups with the same attributes based on the cold start user's registered region and registered age range, performing average pooling on the category embedding vectors of each product category in the preferred product category set to obtain a compensation category vector, and concatenating the original representation vector of the cold start user with the compensation category vector and compressing it into a compensated representation vector through a linear mapping layer.

[0018] Furthermore, in step 3, the constraint processing includes category bias correction processing and regional inventory constraint processing. Category bias correction processing involves: calculating the proportion of each product category in the candidate user set; calculating the excess ratio for product categories whose proportion exceeds the preset category proportion limit; the excess ratio is the difference between the actual proportion of the product category and the preset category proportion limit divided by the actual proportion of the product category; subtracting the excess ratio from 1 gives the retention probability of candidate users associated with product categories whose proportion exceeds the preset category proportion limit; and performing Bernoulli sampling based on the retention probability to determine whether to retain candidate users. Regional inventory constraint processing involves removing users whose delivery addresses are not in the list of deliverable regions. Candidate users with zero inventory in the corresponding region or region; the adaptive adjustment method is as follows: calculate the sliding window conversion rate based on the conversion status data, set the hysteresis interval width value, if the sliding window conversion rate is lower than the difference between the lower limit of the conversion rate threshold and the hysteresis interval width value, increase the ranking threshold N of the near nearest neighbor search to narrow the selection range, if the sliding window conversion rate is higher than the sum of the upper limit of the conversion rate threshold and the hysteresis interval width value, decrease the ranking threshold N to expand the selection range, if the sliding window conversion rate is between the difference between the lower limit of the conversion rate threshold and the hysteresis interval width value and the sum of the upper limit of the conversion rate threshold and the hysteresis interval width value, keep the ranking threshold N unchanged.

[0019] The e-commerce audience expansion and selection method based on similarity group discovery described in this invention has the following beneficial effects: By constructing homologous positive sample pairs and cross-source positive sample pairs, and combining modality enhancement strategies customized for different data sources with a shared projection head network, this invention achieves feature alignment of heterogeneous modal data such as images, text, and sensor signals in a unified semantic space. After contrastive learning training, cross-modal samples with the same entity association key are brought closer to adjacent positions in the feature space, while samples belonging to different entities are pushed further away to different regions, fundamentally solving the technical problem that traditional methods struggle to handle the incomparability of feature spaces for multi-source heterogeneous data.

[0020] This invention employs a multi-table joint locality-sensitive hashing technique to construct a candidate neighbor set, reducing the computational complexity of similarity search from the quadratic level of global sample pairs to the local computational scale within hash buckets, significantly improving the processing efficiency of large-scale datasets. Multiple independent hash tables converge the bucketing results through a union operation, effectively mitigating the recall instability caused by the randomness of random projection in a single hash table, thus controlling the size of the candidate set while ensuring high recall.

[0021] This invention employs the Riemann gradient descent algorithm to iteratively solve for the Fraser mean on a unit hyperspherical manifold as the prototype center for grouping, and uses geodesic distance as a metric for sample deviation. This design strictly preserves the geometric structure of the manifold space containing the normalized feature vectors, avoiding the measurement distortion problems caused by traditional arithmetic mean and Euclidean distance calculation methods, and ensuring that the prototype deviation metric accurately reflects the true degree of deviation of the samples from the typical grouping pattern.

[0022] This invention employs a group-level conformal prediction calibration strategy, independently calculating the inconsistency threshold boundary for each final group. This allows the anomaly detection threshold to adapt to the local data distribution characteristics of different groups, effectively avoiding the problem of judgment bias caused by a globally uniform threshold for groups with different densities. The conformal prediction framework relies only on the sample commutativity assumption without assuming that the data follows a specific parameter distribution, providing coverage guarantees under limited sample conditions at a given confidence level, significantly enhancing the statistical reliability and interpretability of anomaly detection results. Attached Figure Description

[0023] Figure 1 The present invention provides a noise intensity sequence and a curve showing the change in cumulative signal retention rate during the diffusion process, as provided in the embodiments of the present invention.

[0024] Figure 2 This is a schematic diagram illustrating the evolution of the characterization vector distribution during the diffusion process, provided in an embodiment of the present invention.

[0025] Figure 3 The graph showing the changes in the excess ratio and retention probability of the product category during the category bias correction process provided in this embodiment of the invention.

[0026] Figure 4 This is a schematic diagram illustrating how the approximate nearest neighbor retrieval ranking threshold N is adaptively adjusted with the delivery cycle, as provided in an embodiment of the present invention. Detailed Implementation

[0027] E-commerce audience expansion and targeting methods based on similar demographics include the following steps:

[0028] Step 1: Select users who meet the preset conversion conditions from the user database of the e-commerce platform as seed users, obtain the behavioral event sequence of each seed user, extract features from the behavioral event sequence and input it into the sequence encoder to obtain the seed user representation vector corresponding to the seed user, and construct the seed representation set by combining all the seed user representation vectors.

[0029] Step 2: Construct a conditional diffusion generation model. The conditional diffusion generation model includes a forward noise injection module, a reverse denoising generation module, a generation validity gating module, and a diversity constraint module. The forward noise injection module gradually adds noise to the seed user representation vector to obtain a full noise latent vector. The reverse denoising generation module gradually denoises the initial noise vector with the seed user representation vector as a condition to obtain candidate virtual representation vectors. The generation validity gating module and the diversity constraint module screen the candidate virtual representation vectors to obtain virtual similar user representation vectors. All virtual similar user representation vectors constitute a virtual user representation set.

[0030] Step 3: Merge the seed representation set and the virtual user representation set to obtain the enhanced query representation set. Based on the enhanced query representation set, perform an approximate nearest neighbor search in the vector retrieval index to obtain the initial candidate user set. Perform filtering and constraint processing on the initial candidate user set in sequence to obtain the target audience user set. Adaptively adjust the selection range according to the conversion status data returned by the campaign.

[0031] In the actual operation of e-commerce platforms, identifying and expanding high-value user groups is a core step in improving advertising efficiency. This embodiment first selects seed users with high conversion value from a massive user base and encodes their behavioral characteristics into computable vector representations. Then, it uses a conditional diffusion generative model to generate virtual user representations in the representation space that are similar to the seed users but have a certain degree of diversity, thereby achieving intelligent expansion of the audience.

[0032] Seed users are selected based on their actual purchasing behavior on e-commerce platforms. Specifically, user records that have completed purchases within the current statistical period are extracted from the e-commerce platform's user database. The current statistical period can be set to the most recent 30, 60, or 90 days, flexibly adjusted according to business needs. For each user who has completed a purchase, the actual payment amount of all orders within the current statistical period is summed to obtain the user's cumulative purchase amount. The user's cumulative purchase amount is compared with a preset amount threshold. When the user's cumulative purchase amount is greater than or equal to the preset amount threshold, the user is marked as a seed user. The setting of the preset amount threshold needs to be combined with the platform's average order value distribution. For example, on a platform with a median average order value of 200 yuan, the preset amount threshold can be set to 500 yuan. This value is approximately 2.5 times the median average order value, which can filter out users with above-average spending power. In some business scenarios, purchase frequency can also be introduced as an auxiliary screening condition, requiring seed users to have made at least two purchases within the current statistical period to exclude users who make occasional large purchases but lack sustained purchasing intentions.

[0033] After selecting seed users, it is necessary to collect behavioral event records generated by each seed user on the e-commerce platform. These behavioral event records cover various event types generated during user interaction with the platform, including product browsing events, product adding to cart events, product favorites events, order submission events, payment completion events, and product review events. Product browsing events record the user's action of clicking to enter the product details page; product adding to cart events record the user's action of adding the product to the shopping cart; product favorites events record the user's action of adding the product to their favorites list; order submission events record the user's action of clicking to submit the order on the shopping cart page; payment completion events record the user's action of completing the order payment; and product review events record the user's action of posting a review of the purchased product. Each behavioral event record includes a timestamp of the event, an event type identifier, and an associated product identifier. During the data collection phase, all behavioral event records generated by seed users within the current statistical period are extracted from the platform's behavioral log system and sorted in ascending order by timestamp to form the behavioral event sequence corresponding to each seed user. The length of the behavioral event sequence varies depending on the user's activity level; the behavioral event sequence for active users may contain hundreds or even thousands of records, while the behavioral event sequence for inactive users may only contain dozens of records.

[0034] Feature extraction of behavioral event sequences aims to transform raw event records into numerical vectors suitable for neural network processing. For each behavioral event record in the sequence, three types of attribute information related to the associated product are extracted: product category identifier, product price range identifier, and product brand identifier. The product category identifier reflects the category level to which the product belongs. E-commerce platforms typically use a multi-level category system; for example, "clothing-women's clothing-dresses" constitutes a three-level category path. During feature extraction, the finest-grained leaf category is taken as the product category identifier. The product price range identifier maps the actual selling price of the product to discrete price ranges. For example, 0 to 50 yuan is mapped to price range 1, 50 to 100 yuan to price range 2, 100 to 200 yuan to price range 3, 200 to 500 yuan to price range 4, and above 500 yuan to price range 5. This discretization reduces the impact of price fluctuations on feature stability. The product brand identifier records the brand name of the product; for unbranded products, a generic brand identifier is used. The product category identifier, product price range identifier, and product brand identifier are each converted into fixed-dimensional embedding vectors using independent embedding lookup tables. The dimensions of the embedding vectors can be set to 64 or 128. The three embedding vectors are concatenated along their feature dimensions to form the event feature vector corresponding to the behavioral event record. Assuming each embedding vector is 64-dimensional, the concatenated event feature vector will have a dimension of 192. This feature extraction operation is performed on all behavioral event records in the behavioral event sequence, and the records are arranged in chronological order to form an event feature vector sequence.

[0035] The event feature vector sequence needs to be converted into a fixed-length user representation vector by a sequence encoder. The sequence encoder adopts a Transformer architecture based on a self-attention mechanism, which can effectively capture the correlation between different events in the behavioral event sequence. Before being input into the sequence encoder, the event feature vector sequence is first truncated or padded to unify the sequence length to a preset maximum sequence length. The maximum sequence length can be set to 256 or 512. When the original sequence length exceeds the maximum sequence length, the event record with the most recent timestamp is retained and the earlier records are truncated; when the original sequence length is less than the maximum sequence length, zero vectors are padded at the end of the sequence. The reason why the truncation strategy prioritizes retaining recent behavioral records is that the user's recent behavior better reflects their current interests and purchase intentions, while earlier behavior may be outdated.

[0036] The internal structure of a sequence encoder includes an embedding layer, a positional encoding layer, multiple self-attention encoding layers, and a pooling layer. The embedding layer maps the input event feature vector from 192 dimensions to the hidden layer dimension of the encoder through a linear transformation; the hidden layer dimension can be set to 256 or 512 dimensions. The positional encoding layer generates a positional encoding vector for each position in the sequence, using a sine-cosine positional encoding method. For the i-th positional encoding vector in the sequence... The position of the position encoding vector is the position of the position. Calculation method and location of each component and component index Related, among which This represents the index of the event within the sequence, with values ​​ranging from 0 to the maximum sequence length minus 1. This represents the dimension index of the location encoding vector. The vector output from the embedding layer is added to the location encoding vector at the corresponding position to obtain the input representation of the fused location information.

[0037] The multi-layer self-attention encoding layer is the core component of the sequence encoder, containing multiple stacked Transformer encoder layers, with the number of encoder layers adjustable to 4 or 6. Each Transformer encoder layer contains a multi-head self-attention sublayer and a feedforward neural network sublayer. The multi-head self-attention sublayer uses 8 or 16 attention heads, each independently calculating the attention relationships between positions in the input sequence. For each position in the input sequence, the multi-head self-attention sublayer weights and aggregates the representations of all positions based on their relevance to all positions in the sequence, thus enabling the output representation of each position to incorporate information from other positions in the sequence. This mechanism allows the sequence encoder to capture long-range dependencies in user behavior sequences, such as the behavior pattern of users browsing a certain type of product before making a purchase several days later. The feedforward neural network sublayer contains two linear transformation layers and a non-linear activation function, further transforming the features of the multi-head self-attention sublayer's output. The output of each sublayer is added to the sublayer input via residual connections and then normalized to improve the stability of model training.

[0038] After processing through multiple self-attention encoding layers, a context-aware vector sequence of the same length as the input sequence is obtained, where each vector incorporates global contextual information. The pooling layer compresses the variable-length context-aware vector sequence into a single fixed-length vector. This embodiment uses mean pooling to calculate the arithmetic mean of all non-padded position vectors in the context-aware vector sequence, yielding the seed user representation vector corresponding to the seed user. The dimension of the seed user representation vector is consistent with the hidden layer dimension of the sequence encoder, being either 256 or 512 dimensions. In optional implementations, max pooling or attention pooling can also be used. Max pooling takes the maximum value across all dimensions, while attention pooling introduces a learnable query vector to weighted aggregate the positions in the sequence.

[0039] Perform the aforementioned behavioral event sequence collection, feature extraction, and sequence encoding operations on all seed users to obtain a seed user representation vector for each seed user. Collect all seed user representation vectors to form a seed representation set. Assuming the number of selected seed users is 10,000, the seed representation set contains 10,000 seed user representation vectors, each with a dimension of 256 or 512.

[0040] After obtaining the seed representation set, a conditional diffusion generative model is used to generate virtual similar user representation vectors in the representation space. The core idea of ​​the conditional diffusion generative model is to transform random noise into the target data distribution by learning the noise reduction and denoising process of the data distribution. Compared with the simple extension method of directly sampling randomly in the neighborhood of the seed user representation vector, the conditional diffusion generative model can learn the overall distribution characteristics of the seed user representation vector, and the generated virtual representation vectors are more consistent with the potential distribution of high-value users, rather than just a small perturbation of the existing vectors.

[0041] The purpose of forward noise injection is to gradually transform the seed user representation vector into a fully noisy latent vector that approximates a standard normal distribution. This process provides a training target for inverse denoising generation. The total number of diffusion steps is set to a positive integer T. The value of the total number of diffusion steps affects the balance between generation quality and computational efficiency. A larger total number of diffusion steps allows for finer noise control but increases computational overhead, while a smaller total number of diffusion steps is computationally efficient but may affect generation quality. In practical applications, the total number of diffusion steps can be set to 100, 500, or 1000 steps. A noise intensity sequence containing T elements is pre-constructed. The nth element in the noise intensity sequence... The elements are denoted as , indicating the first The noise intensity value corresponding to the step diffusion, where The value range is from 1 to T. The noise intensity sequence is pre-set according to a monotonically increasing pattern from step 1 to step T. The noise intensity in the early steps is smaller to preserve the main characteristics of the original signal, and the noise intensity in the later steps is larger so that the vector gradually approximates the pure noise distribution. The specific values ​​of the noise intensity sequence can be linearly increasing, for example... Set to 0.0001. The value is set to 0.02, and the noise intensity values ​​of each intermediate step are obtained by linear interpolation between the two endpoints. In an optional implementation, a cosine increment method or a square increment method can also be used to construct the noise intensity sequence. The cosine increment method results in a slower increase in noise intensity in the early steps and an accelerated increase in the later steps, while the square increment method exhibits a non-linear growth characteristic of being slow at first and then fast.

[0042] refer to Figure 1 The horizontal axis of the graph represents the number of diffusion steps. The value ranges from 0 to 1000, and the vertical axis represents the cumulative signal retention rate. The value ranges from 0 to 1.0. The figure contains two curves, corresponding to the changes in the cumulative signal retention rate with the number of diffusion steps under two different noise intensity sequence construction methods: linear noise scheduling and cosine noise scheduling. The curve corresponding to linear noise scheduling shows a relatively smooth monotonically decreasing trend, gradually decreasing from a cumulative signal retention rate of 1.0 when the number of diffusion steps is 0, and approaching 0 when the number of diffusion steps reaches 1000. The curve corresponding to cosine noise scheduling also shows a monotonically decreasing trend, but its rate of decrease differs significantly between the early and late stages of the diffusion process; the decrease is relatively gentle in the early stage, and accelerates in the later stage. A horizontal reference line is drawn at 0.5 on the vertical axis, dividing the entire figure into a signal-dominated region and a noise-dominated region. When the cumulative signal retention rate is higher than 0.5, it indicates that the signal component of the original representation vector dominates in the intermediate vector; when the cumulative signal retention rate is lower than 0.5, it indicates that the noise component gradually dominates. The figure also marks key sampling points at diffusion steps of 100, 250, 500, 750, and 1000, facilitating observation of the specific values ​​of cumulative signal retention rate at different diffusion stages. (Cumulative signal retention rate) Reflects the process The parameter represents the degree to which signal components are preserved in the original input vector after noise superposition. A larger value indicates that more of the original signal is preserved, while a smaller value indicates that the original signal is more obscured by noise. This figure provides an intuitive understanding of the physical meaning of the forward noise injection process in the conditional diffusion generation model. As the number of diffusion steps increases, the user feature information contained in the seed user representation vector is gradually diluted, eventually transforming into a fully noise-laden vector that approximates a standard normal distribution after all diffusion steps are completed.

[0043] The execution process of forward noise injection is as follows: The seed user representation vector is denoted as the original input vector. ,in This represents the coordinates of the seed user's representation vector in the representation space. The noise superposition operation is performed sequentially from step 1 to step T. In the... In the noise superposition operation, the first step is to sample the first noise vector from the standard normal distribution, which has the same dimension as the original input vector. Step sampling noise vector, denoted as ,in Each component independently follows a normal distribution with a mean of 0 and a variance of 1. The th component... Step sampling noise vector With the noise intensity sequence of the th Noise intensity value Multiply to get the first... Scaling the noise vector step by step. intermediate vector of step output With the The noise vector is scaled step by step and then vector addition is performed to obtain the first step. intermediate vector of step output For the noise superposition operation in step 1, the intermediate vector output in step 0... This is the original input vector. After completing the T-th step of noise superposition, the intermediate vector output at step T is... This serves as the fully noise-laden vector corresponding to the original input vector. After T steps of noise superposition, the distribution of the fully noise-laden vector approximates a standard normal distribution, and the user feature information in the original input vector is gradually covered up.

[0044] The purpose of inverse denoising generation is to start from random noise and, using the seed user representation vector as a condition, progressively generate virtual representation vectors similar to the seed user. The inverse denoising process can be understood as the reverse of the forward noise injection process, using a neural network to learn which noise components should be removed at each step. The denoising neural network uses a multilayer perceptron structure. Compared to convolutional neural networks or recurrent neural networks, multilayer perceptrons are more suitable for processing one-dimensional representation vectors and have higher computational efficiency. The multilayer perceptron structure includes a sequentially connected input concatenation layer, a first hidden layer, a second hidden layer, a third hidden layer, and an output mapping layer. The function of the input concatenation layer is to concatenate the vector to be denoised, the condition vector, and the step-count encoding vector along the feature dimension into a single input vector. The dimension of the vector to be denoised is the same as the dimension of the seed user representation vector, which is 256 or 512 dimensions. The condition vector is the seed user representation vector used as the generation condition, also with a dimension of 256 or 512 dimensions. The step count encoding vector indicates to the denoising neural network which step it is currently in the denoising process. Different steps correspond to different noise levels, and the denoising neural network needs to adjust its denoising strategy according to the current step. Assuming the seed user representation vector has a dimension of 256 and the step count encoding vector has a dimension of 64, then the output dimension of the input concatenation layer is 256 + 256 + 64 = 576 dimensions.

[0045] The step count encoding vector is generated using a sinusoidal positional encoding method. For the current step count... The first step of the encoded vector Each component is based on the number of steps. and component index The calculation yielded, where This represents the current denoising step count, with a value ranging from 1 to T. This represents the dimension index of the step-count encoding vector. Sine positional encoding generates a unique encoding vector for each step by combining sine and cosine functions of different frequencies, enabling the denoising neural network to distinguish between different denoising stages.

[0046] The first, second, and third hidden layers have similar structures, each containing a linear transformation unit and a nonlinear activation unit. The linear transformation unit performs matrix multiplication and bias addition operations, mapping the input vector to an output space of a specified dimension. The nonlinear activation unit applies a nonlinear transformation to the output of the linear transformation. In this embodiment, a Gaussian error linear unit is used as the activation function. Compared to the traditional corrected linear unit, the Gaussian error linear unit has smoother gradient characteristics, which is beneficial to the stability of model training. The input dimension of the first hidden layer is 576, and the output dimension can be set to 1024. The input dimension of the second hidden layer is 1024, and the output dimension is also 1024. The input dimension of the third hidden layer is 1024, and the output dimension is 512. The output mapping layer maps the output of the third hidden layer to a noise estimation vector with the same dimension as the vector to be denoised through a linear transformation, with an output dimension of 256 or 512. The noise estimation vector represents the denoising neural network's estimate of the noise components in the current vector to be denoised.

[0047] The reverse denoising generation process is as follows: Select one seed user representation vector from the seed representation set as the condition vector. The condition vector is used to guide the generation process to produce a virtual representation vector similar to that of the seed user. An initial noise vector with the same dimension as the seed user's representation vector is sampled from a standard normal distribution, and this initial noise vector is used as the denoising vector to be denoised in the T-th step. The denoising operation is performed sequentially from step T to step 1, gradually removing noise components from the vector to be denoised. In the... In the step-by-step denoising operation, first, the current step number is... Perform sinusoidal position encoding to obtain the step-count encoding vector. Then, assign the first... Denoising vector Conditional vector The step-count encoding vector is input into the denoising neural network, and the denoising neural network outputs the first step-count encoding vector. Step noise estimation vector This vector represents the neural network's... Estimation of noise components. Obtaining the first noise component from the noise intensity sequence. Noise intensity value , will the Step noise estimation vector With noise intensity value Multiply to get the first... Scale the noise estimation vector step by step. Denoising vector Subtract the first The noise estimation vector is scaled up step by step to obtain the first step. Denoising vector After completing the first step of denoising, the vector to be denoised in step 0... As a condition vector Candidate virtual representation vectors are generated based on the conditions. The introduction of conditional vectors ensures that the generated virtual representation vectors are semantically similar to the seed user, rather than generating random vectors that are completely unrelated to the seed user.

[0048] The purpose of generation validity gating is to filter out candidate virtual representation vectors that are of low quality or deviate too far from the distribution of seed users. Due to the randomness of the diffusion generation process, some generated candidate virtual representation vectors may not meet the expected quality standards and need to be screened through a gating mechanism. The execution process of generation validity gating includes two levels of judgment. The first level of judgment calculates the cosine similarity between the candidate virtual representation vector and the condition vector. Cosine similarity measures the degree of consistency in direction between two vectors, ranging from -1 to 1; the closer the value is to 1, the more consistent the directions of the two vectors. Assuming a preset lower threshold for validity similarity of 0.6 or 0.7, if the cosine similarity between the candidate virtual representation vector and the condition vector is less than the lower threshold, the candidate virtual representation vector is deemed invalid and discarded. This judgment ensures that the generated virtual representation vector maintains sufficient similarity to the seed user representation vector used as a condition, avoiding the generation of noise vectors unrelated to the conditions.

[0049] If a candidate virtual representation vector passes the first-level judgment, it proceeds to the second-level judgment. The second-level judgment calculates the average cosine similarity between the candidate virtual representation vector and all seed user representation vectors in the seed representation set. Specifically, it calculates the cosine similarity between the candidate virtual representation vector and each seed user representation vector in the seed representation set, and then calculates the arithmetic mean of all cosine similarities to obtain the average cosine similarity. A preset distribution consistency lower bound threshold is set to 0.4 or 0.5. If the average cosine similarity is less than the distribution consistency lower bound threshold, the candidate virtual representation vector is considered to deviate from the seed distribution and is discarded. The significance of the second-level judgment is to ensure that the generated virtual representation vector is not only similar to the specific condition vector but also consistent with the overall distribution of the entire seed user group. If a candidate virtual representation vector is only similar to the condition vector but differs significantly from other seed user representation vectors, this vector may represent an isolated outlier rather than a typical high-value user feature. If the average cosine similarity is greater than or equal to the distribution consistency lower bound threshold, the candidate virtual representation vector is considered to have passed the validity gating.

[0050] The purpose of diversity constraints is to ensure sufficient differentiation among multiple virtual representation vectors generated for the same seed user, preventing excessive concentration of generated virtual representation vectors within the smallest neighborhood of the seed user's representation vector. If multiple generated virtual representation vectors are highly similar, highly overlapping candidate user groups will be recalled in subsequent approximate nearest neighbor retrieval, failing to effectively expand the audience coverage. The execution process of diversity constraints is as follows: For each seed user representation vector in the seed representation set, using that seed user representation vector as a condition vector, the process of inverse denoising generation and generation validity gating is repeatedly executed. Each execution resamples the initial noise vector from a standard normal distribution. Due to the randomness of the initial noise vector, the generated candidate virtual representation vectors differ each time. The generation process continues until a preset number of M candidate virtual representation vectors that pass the validity gating are obtained. The preset number M can be set to 5, 10, or 20, adjusted according to the scale requirements of audience expansion.

[0051] refer to Figure 2 The graph contains six subgraphs, each corresponding to a diffusion step. For 0, 100, 300, 500, 800 and The time-representation vector is distributed in the two-dimensional projection space. The horizontal and vertical axes of each subgraph represent representation dimension 1 and representation dimension 2, respectively, with values ​​ranging from -5 to 5. In the diffusion steps... In the corresponding subgraph, the representation vectors exhibit a clear clustering structure, with scattered points concentrated in several compact regions. This reflects the clustering characteristics of the seed user representation vectors in the original representation space, indicating that high-value user groups share similar behavioral preferences, thus forming clusters in the representation space. As the number of diffusion steps increases, in... In the corresponding subplot, it can be observed that the scatter distribution begins to spread outwards, the original cluster boundaries gradually become blurred, but the overall cluster structure is still discernible. In the corresponding subplot, the diffusion of scatter points intensifies further, the distinction between the original cluster centers decreases significantly, and some scatter points have spread into previously empty areas. In the corresponding subplot, the scatter point distribution tends to be more uniform, the original clustering structure basically disappears, and the distribution range of the scatter points in the two-dimensional projection space is significantly expanded. In the corresponding subplot, the scatter points are distributed almost uniformly randomly, making it difficult to discern any structural features from the original data. In the corresponding sub-graphs, the scatter points exhibit a standard normal distribution, and the representation vector has been transformed into a fully noisy latent vector, completely overwriting the user feature information in the original seed user representation vector. Each sub-graph also includes contour lines to illustrate the density distribution of the scatter points; the density of the contour lines reflects the density of scatter points in that area. This diagram provides an intuitive understanding of the gradual change in the representation vector distribution during the forward noise injection process. This change provides a learning target for inverse denoising generation, whose task is to gradually recover virtual similar user representation vectors that conform to the seed user distribution characteristics from the fully noisy state.

[0052] M candidate virtual representation vectors that pass the validity gating are formed into a vector group to be screened. The cosine similarity between any two candidate virtual representation vectors in the vector group to be screened is calculated to form a similarity matrix. All elements in the similarity matrix are traversed; if the cosine similarity between any two candidate virtual representation vectors is greater than a preset diversity similarity upper threshold, then the two vectors are considered too similar, and one of them needs to be removed. The diversity similarity upper threshold can be set to 0.9 or 0.95. This threshold setting needs to strike a balance between diversity and relevance; too low a threshold will lead to the removal of too many candidate vectors, while too high a threshold cannot effectively guarantee diversity. In the removal operation, candidate virtual representation vectors with higher cosine similarity to the condition vector are retained, while candidate virtual representation vectors with lower cosine similarity are removed. After diversity constraint screening, the remaining candidate virtual representation vectors in the vector group to be screened are retained as the virtual similar user representation vectors corresponding to the seed user representation vector.

[0053] The virtual representation vector generation, validity gating, and diversity constraint operations described above are performed on all seed user representation vectors in the seed representation set. Assuming the seed representation set contains 10,000 seed user representation vectors, and each seed user representation vector generates 5 virtual similar user representation vectors, approximately 50,000 virtual similar user representation vectors may be generated in the end. The actual number may be slightly less than this value because some candidate vectors are removed during the diversity constraint process. All virtual similar user representation vectors are then aggregated to form a virtual user representation set. This virtual user representation set, together with the seed representation set, constitutes the enhanced query representation set, used for subsequent similar user recall operations. The virtual similar user representation vectors generated through the conditional diffusion generation model introduce controllable diversity while maintaining similarity to seed users, effectively expanding audience coverage and recalling potential high-value user groups.

[0054] After generating seed user representation vectors and virtual similar user representation vectors based on the conditional diffusion model, these representation vectors need to be used to recall potential target audiences from the full user database. Multi-layer filtering and constraint processing are used to ensure the quality and deliverability of recalled users. Finally, the selection strategy is dynamically adjusted based on the actual delivery results.

[0055] The construction of the enhanced query representation set is fundamental to the recall of similar user groups. The seed representation set and the virtual user representation set are merged by aggregating all representation vectors from both sets into a single set. Assuming the seed representation set contains 10,000 seed user representation vectors and the virtual user representation set contains 45,000 virtual similar user representation vectors, the merged enhanced query representation set will contain 55,000 representation vectors. This enhanced query representation set includes both representation vectors from real high-value users and virtual representation vectors generated through a diffusion model. Both types of vectors serve as query anchors in subsequent approximate nearest neighbor retrieval. The inclusion of virtual representation vectors expands the query coverage, recalling potential users who share similar characteristics with seed users but might have been missed by retrieval based solely on seed vectors.

[0056] Approximate nearest neighbor retrieval relies on a pre-built vector retrieval index. During system initialization or periodic updates, the same representation vector generation operation as for seed users is performed on all registered users of the e-commerce platform. This involves collecting behavioral event sequences for each registered user, extracting features, and inputting them into a sequence encoder to obtain a user representation vector for each registered user. The representation vectors of all registered users are stored in the vector retrieval index, and a mapping relationship between representation vectors and user identifiers is established. The vector retrieval index employs either a hierarchical navigable small-world graph structure or an inverted product quantization structure. Both of these index structures support millisecond-level approximate nearest neighbor retrieval even with hundreds of millions of vectors. The hierarchical navigable small-world graph structure, by constructing a multi-layered graph structure, quickly locates candidate regions in higher-level graphs and accurately searches for nearest neighbors in lower-level graphs, resulting in high retrieval accuracy but significant memory consumption. The inverted product quantization structure, by combining vector quantization and inverted indexing, achieves a balance between memory consumption and retrieval speed, making it suitable for ultra-large-scale vector libraries. Assuming the e-commerce platform has 100 million registered users, and each user's representation vector has 256 dimensions, using 32-bit floating-point storage, the original vector data occupies approximately 100 gigabytes of storage space. After product quantization compression, this can be reduced to less than 10 gigabytes.

[0057] For each representation vector in the enhanced query representation set, an approximate nearest neighbor search based on cosine similarity is performed in the vector retrieval index. Cosine similarity measures the degree of agreement between two vectors in direction for the query vector. and candidate vectors The cosine similarity is calculated by dividing the inner product of the two vectors by the product of their magnitudes, where... This represents a representation vector within the enhanced query representation set. This represents the representation vector of a registered user in the vector retrieval index. The retrieval process returns the top-ranked vectors based on their cosine similarity to the query vector. Candidate users, This is called the ranking threshold, and its initial value can be set to 100 or 200. The setting of the ranking threshold affects the trade-off between the recall scale and the recall precision. A larger ranking threshold can recall more candidate users but may introduce more low-relevance users, while a smaller ranking threshold recalls more relevant users but has limited coverage.

[0058] For each of the 55,000 representation vectors in the enhanced query representation set, an approximate nearest neighbor search is performed sequentially, with a recall ranking threshold for each representation vector. There are 10 candidate users. All search results are aggregated to form an initial candidate user set. Since different query vectors may recall the same candidate users, duplicate records exist in the initial candidate user set. Assume a ranking threshold. If the value is set to 100, the enhanced query representation set contains 55,000 representation vectors, and the theoretical maximum recall is 5.5 million user records. However, due to the overlap of candidate users, the actual size of the initial candidate user set is usually smaller than this value.

[0059] refer to Figure 4 The horizontal axis of the graph represents the campaign period, measured in time windows, with values ​​ranging from 0 to 50. The vertical axis represents the approximate nearest neighbor retrieval ranking threshold. The value ranges from 30 to 210. The ranking threshold is plotted in the graph. The step-like curve changes with the campaign period, starting at 100 on the vertical axis, corresponding to the initial value of the ranking threshold. Each step of the stepped curve represents a threshold adjustment event, and the height of the step corresponds to the preset adjustment step size. When the curve rises, it indicates a ranking threshold. Increasing the selection scope narrows the pool, with each query vector only recalling candidate users with higher similarity. The aim is to improve recall accuracy when conversion rates are poor. A downward curve indicates a ranking threshold. Reducing the number of queries expands the selection scope, allowing each query vector to recall more candidate users. The goal is to broaden audience reach when conversion rates are good. The graph shows horizontal reference lines corresponding to the initial values. The position of the ranking threshold is indicated by the filled area below the curve. The change range relative to the minimum allowable value. The graph also includes textual explanations regarding the ranking threshold. When increasing the selection range, the selection area shrinks to improve accuracy, when the ranking threshold is reached. The selection range expands to increase coverage when the selection is reduced. The ranking threshold can be observed from the overall trend of the curve. The system underwent multiple adjustments, both upward and downward. This dynamic adjustment was triggered by comparing the sliding window conversion rate with a preset threshold, demonstrating the adaptive selection range adjustment mechanism of this invention. Through a hysteresis control strategy, frequent fluctuations in the ranking threshold near the conversion rate critical value were avoided, ensuring the stability of the selection strategy.

[0060] The purpose of deduplication is to eliminate duplicate records in the initial candidate user set, ensuring that each candidate user retains only one record. The deduplication process is as follows: iterate through all user records in the initial candidate user set, using the user identifier as the deduplication key. If the same user identifier appears multiple times in the initial candidate user set, only one record for that user identifier is retained, and the remaining duplicate records are deleted. In an optional implementation, the record with the highest cosine similarity can be retained during deduplication. That is, when the same user is recalled by multiple query vectors, the cosine similarity between the user and each query vector is recorded, and the maximum cosine similarity is retained as the user's similarity score. After deduplication, a deduplicated candidate user set is obtained, where each user corresponds to only one record.

[0061] The purpose of frequency control is to avoid excessively frequent ad exposure to the same user, preventing user annoyance and reducing ad effectiveness. The frequency control process works as follows: It queries the cumulative number of times each candidate user in the deduplicated candidate user set has been reached by ads within the current statistical period. This cumulative number is obtained from the ad delivery log system, recording the number of ad impressions received by each user within the current statistical period. The cumulative number for each candidate user is compared to a preset frequency limit, which can be set to 3 times per day, 7 times per week, or 15 times per month, flexibly adjusted according to ad type and delivery strategy. If a candidate user's cumulative number reaches or exceeds the preset frequency limit, that candidate user is removed from the deduplicated candidate user set. The frequency-controlled candidate user set is obtained after the frequency control process. Frequency control can control the ad exposure density for individual users while ensuring ad coverage, improving user experience and avoiding waste of ad resources.

[0062] Cold start representation compensation processing targets new or inactive users with sparse behavioral data. Because these users have a limited number of behavioral event records, their representation vectors may not accurately reflect their true preferences, leading to incorrect recall or omission in near-nearest neighbor retrieval. The execution process of cold start representation compensation processing is as follows: First, cold start users are identified in the candidate user set after frequency control. The identification criterion is that the number of behavioral event records for a user is less than a preset minimum record count threshold, which can be set to 10 or 20 records. Users with fewer than this threshold are marked as cold start users, and their representation vectors have lower reliability.

[0063] For each cold-start user, their registration information is retrieved, including the registration region and age range. The registration region records the city or province the user entered during registration, and the age range is calculated based on the user's birth date, typically divided into 18-24, 25-34, 35-44, 45-54, and 55 and above. Based on the cold-start user's registration region and age range, the preferred product categories of active user groups with similar attribute combinations are retrieved from the user profile database. The user profile database pre-calculates the consumption preferences of user groups with different attribute combinations. For example, an active user group registered in Beijing and aged 25-34 might have a preferred product category set including digital products, beauty and skincare, and sports and fitness. The preferred product category set is determined based on recent purchase behavior statistics of this user group, selecting several product categories with high purchase frequency or purchase amount.

[0064] After obtaining the set of preferred product categories, the system queries the category embedding table to retrieve the category embedding vector corresponding to each product category. The category embedding table is pre-built during system initialization, mapping each product category to a fixed-dimensional embedding vector. The dimension of the embedding vector is consistent with the dimension of the user representation vector, for example, both being 256 dimensions. An average pooling operation is performed on the category embedding vectors of each product category in the preferred product category set, that is, the arithmetic mean of each category embedding vector is calculated according to its dimension, resulting in a compensated category vector. The compensated category vector integrates the typical preference characteristics of the user group with this attribute combination.

[0065] The original representation vector of the cold-start user is concatenated with the compensated category vector, following the feature dimension end-to-end. If both the original and compensated category vectors are 256-dimensional, the concatenated vector is 512-dimensional. This concatenated vector is then compressed using a linear mapping layer, which performs matrix multiplication to map the 512-dimensional vector back to 256 dimensions, resulting in the compensated representation vector. This compensated representation vector integrates the limited behavioral information of the cold-start user with typical preference information from similar groups, reflecting the user's potential interests more accurately than the original representation vector. The compensated representation vector replaces the original representation vector for the cold-start user in the vector retrieval index, enabling more accurate user representation in subsequent ad placement and audience targeting. The cold-start representation compensation process yields a compensated candidate user set.

[0066] The purpose of category bias correction is to prevent the recalled candidate users from being overly concentrated in certain popular product categories, ensuring a relatively balanced distribution of the target audience across different product categories. In real-world e-commerce scenarios, some popular categories, such as clothing and food, have a large user base. Without bias correction, a large number of recalled candidate users may be concentrated among those who prefer these popular categories, causing potential users in long-tail categories to be overlooked. The execution process of category bias correction is as follows: First, the primary preferred product categories associated with each candidate user in the compensated candidate user set are statistically analyzed. The primary preferred product categories are determined based on the user's behavioral event records, selecting the product category with the most views, add-to-cart, or purchases as the user's primary preferred product category.

[0067] refer to Figure 3 The horizontal axis of the chart represents the actual percentage of product categories. The percentage is expressed as a percentage, ranging from 5% to 45%. The graph contains two vertical axes; the left axis represents the percentage exceeding the limit. The value is expressed as a percentage, ranging from 0% to 75%, with the right vertical axis representing the probability of retention. Expressed as a percentage, the value ranges from 25% to 105%. A vertical reference line is drawn in the chart to indicate the upper limit of the category percentage. In this embodiment, the category percentage is set to a maximum of 15%. (Exceeding this limit...) The corresponding curve represents the actual percentage of the product category. Less than or equal to the category percentage limit If the value remains 0 within the specified range, it indicates that the category has not exceeded the limit and no adjustment is needed. When the actual percentage of the product category... Exceeding the category percentage limit Afterwards, the proportion exceeding the limit It begins to rise, with the upward trend showing a pattern of rapid initial growth followed by a slower pace. (Probability retention) The corresponding curve represents the actual percentage of the product category. Less than or equal to the category percentage limit If the percentage remains at 100% within the specified range, it means that all candidate users associated with that category will be retained. When the actual percentage of the product category... Exceeding the category percentage limit After that, retain the probability. It started to decline, the downward trend and the over-limit ratio The upward trends are complementary. A specific example point is marked in the chart, where the actual percentage of the product category... When it is 30%, the excess ratio Approximately 50%, retention probability Approximately 50%, meaning that each candidate user associated with this category has a 50% probability of being retained. (Exceeding the limit percentage) With retention probability They satisfy a complementary relationship, that is This graph illustrates the quantitative control mechanism of category bias correction. The more severe the actual proportion of a product category exceeds the limit, the lower the retention probability of its associated candidate users. This achieves a balanced adjustment of the category distribution in the candidate user set, preventing the target audience from being overly concentrated in popular categories.

[0068] The percentage of each product category in the compensated candidate user set is calculated. The percentage of a product category is defined as the number of candidate users associated with that product category divided by the total number of users in the compensated candidate user set. This percentage is compared to a preset upper limit for category percentages, which can be set to 10% or 15%, meaning that the number of candidate users associated with any product category should not exceed 10% or 15% of the total number of candidate users. For product categories whose percentage exceeds the preset upper limit, the excess percentage for that product category is calculated. The excess percentage is calculated as follows: Let the actual percentage of the product category be... The preset category percentage limit is The excess ratio Equal to actual percentage With respect to the preset category percentage limit The difference divided by the actual percentage ,in This indicates the actual percentage of product categories in the compensated candidate user set. This indicates the maximum percentage of preset categories. This indicates the percentage exceeding the limit. For example, if the actual percentage of a product category is 25%, and the preset category percentage limit is 15%, then the excess percentage is 0.4, meaning that approximately 40% of the candidate users associated with that category need to be removed.

[0069] For each candidate user in a product category where the association percentage exceeds the preset category percentage limit, calculate the retention probability of that candidate user. The retention probability is calculated as 1 minus the excess percentage. That is, the probability of retention ,in This represents the retention probability of a candidate user. In the example above, the retention probability is 0.6, meaning that each candidate user associated with this category has a 60% probability of being retained. Bernoulli sampling is performed on the candidate users based on the retention probability to determine whether to retain them. Bernoulli sampling is a binary random sampling method for sampling with a retention probability of 0.6. For candidate users, generate a random number uniformly distributed between 0 and 1. If the random number is less than... If the candidate user is selected based on retention probability, then that user is retained; otherwise, they are removed. This sampling method based on retention probability is more reasonable than simple random removal. The more severe the category's exceedance, the higher the removal rate of its associated users, thus making the corrected category distribution more balanced. The category bias correction process yields the corrected candidate user set.

[0070] The purpose of region-based inventory constraint processing is to ensure that recalled candidate users can actually complete the purchase, avoiding resource waste caused by targeting users in regions with no delivery capability or insufficient inventory. The execution process of region-based inventory constraint processing is as follows: The list of deliverable regions for the product associated with the current advertisement, as well as the inventory quantity in each deliverable region, are obtained. The list of deliverable regions is set by the merchant when the product is published, recording the cities or provinces that the product supports delivery to. The inventory quantity in each deliverable region is obtained in real time from the inventory management system, reflecting the available inventory status of the product in each warehouse.

[0071] Iterate through each candidate user in the category-corrected candidate user set to obtain the region of the candidate user's shipping address. The region of the shipping address is extracted from the user's default or most recently used shipping address, specifying the city or province. Match the candidate user's shipping address region with the list of deliverable regions. If the candidate user's shipping address region is not in the list, remove the candidate user from the category-corrected candidate user set. If the candidate user's shipping address region is in the list, further check the inventory quantity in that region. If the inventory quantity in the corresponding region is 0, it means there is no stock available in that region, and the candidate user is also removed from the category-corrected candidate user set. After processing the regional inventory constraints, the target audience user set is obtained, where all users satisfy the constraints of being deliverable and having stock.

[0072] The target audience set is output to the advertising delivery system for ad delivery. The advertising delivery system displays relevant ad content to users when they visit the e-commerce platform, based on the identifiers of each user in the target audience set. The advertising delivery system also records the time of each ad display, the user identifier, and the user's subsequent behavior.

[0073] The system collects conversion status data from the advertising delivery system, recording the behavioral responses of each user in the target audience after an ad is displayed. Conversion status data includes whether the user clicked the ad, added the associated product to their cart, and completed the purchase. Clicking the ad records whether the user clicked to enter the product details page after seeing the ad; adding the associated product to the cart records whether the user performed the add-to-cart action; and completing the purchase records whether the user ultimately completed the order payment. Among these three conversion behaviors, completing the purchase is the most crucial conversion goal, directly related to the commercial value of the advertising campaign.

[0074] The sliding window conversion rate is calculated based on conversion status data. The calculation is based on recent campaign data, rather than cumulative total data; this method reflects recent changes in campaign performance more promptly. Let the sliding window duration be... , It can be set to 24 hours, 48 ​​hours, or 72 hours. It calculates the most recent time window. The number of users who complete the purchase within the specified time is denoted as . ,in This indicates the number of users who received an ad display and completed a purchase within the time window. It also includes statistics for the most recent time window. The total number of users who received ad impressions is denoted as . ,in This represents the total number of users who received ad impressions within the time window. Sliding window conversion rate. Equal to the number of users who completed the purchase Divide by the total number of users targeted ,in This represents the sliding window conversion rate, ranging from 0 to 1. For example, if an ad was shown to 10,000 users in the past 24 hours, and 150 users made a purchase, the sliding window conversion rate is 1.5%.

[0075] The selection range is adaptively adjusted based on the sliding window conversion rate. This adjustment is achieved by modifying the ranking threshold for near-nearest neighbor retrieval. Achieve ranking threshold The larger the value, the more candidate users are recalled per query vector, and the wider the selection range; ranking threshold The smaller the value, the fewer candidate users are recalled, and the smaller the selection range. (Preset conversion rate lower limit threshold) and conversion rate upper limit threshold , It can be set to 0.8%. It can be set to 2.0%, with the specific value determined based on the advertiser's campaign goals and the industry average.

[0076] To avoid frequent fluctuations in the selection range around the conversion rate threshold, a hysteresis interval width value is introduced. , It can be set to 0.2%. The purpose of the hysteresis interval is to set different adjustment trigger points during the conversion rate increase and decrease, preventing frequent adjustments to the ranking threshold due to small fluctuations in the conversion rate. The specific adjustment logic is as follows: if the sliding window conversion rate... Below the lower threshold of conversion rate Subtract the hysteresis interval width value The difference, i.e. This indicates that the current conversion rate is significantly lower than the expected lower limit, and it is necessary to improve the quality of recalled users. In this case, the ranking threshold should be adjusted. The value increases the preset adjustment step size to narrow the selection range. The preset adjustment step size can be set to 10 or 20, that is... Increase the current value by 10 or 20. After narrowing the selection, each query vector will only recall candidate users with higher similarity, which is expected to improve the conversion probability of recalled users. If the sliding window conversion rate... Above the upper limit of conversion rate Add hysteresis interval width value The sum of, i.e. This indicates that the current conversion rate has exceeded the expected upper limit, and there is room to expand the coverage. In this case, the ranking threshold will be adjusted. The value of this setting is reduced by decreasing the preset adjustment step size to expand the selection area. Expanding the selection area allows reaching more potential users, increasing ad reach while maintaining a high conversion rate. (If the sliding window conversion rate...) The value lies between the difference between the lower threshold of conversion rate and the width of the hysteresis interval, and the sum of the upper threshold of conversion rate and the width of the hysteresis interval. This indicates that the current conversion rate is within an acceptable range, and the ranking threshold should be maintained. The value remains unchanged.

[0077] Let's illustrate the operation of the hysteresis mechanism using specific numerical examples. Assume a lower threshold for conversion rate. The upper limit of the conversion rate is 0.8%. The hysteresis interval width is 2.0%. The threshold is 0.2%. When the sliding window conversion rate drops from 1.0% to 0.7%, since 0.7% is less than 0.8% minus 0.2% equals 0.6%, which is not true, the conversion rate remains above 0.6%, so the operation to narrow the selection range is not triggered. When the sliding window conversion rate further drops to 0.5%, since 0.5% is less than 0.6%, the operation to narrow the selection range is triggered. When the conversion rate rebounds from a low point, it needs to rise above 0.6% before it stops narrowing. This hysteresis mechanism effectively avoids frequent adjustments when the conversion rate fluctuates slightly around the threshold.

[0078] Based on the adjusted ranking threshold Re-execute the similar audience recall operation, which involves searching the vector retrieval index for top-ranked cosine similarity for each representation vector in the enhanced query representation set. The candidate users are then processed sequentially, including deduplication, frequency control, cold start characterization compensation, category bias correction, and regional inventory constraint processing, resulting in a new target audience set. This closed-loop process continues, with the advertising system constantly collecting conversion status data. Based on this data, the system calculates the sliding window conversion rate and dynamically adjusts the ranking threshold, achieving adaptive adjustment of the target audience set. Through this feedback mechanism, the selection system can automatically optimize the recall strategy based on actual campaign performance. When conversion results are poor, the selection range is tightened to improve accuracy; when conversion results are good, the selection range is expanded to increase coverage, thus achieving a dynamic balance between advertising effectiveness and scale.

[0079] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for expanding and selecting e-commerce audiences based on similar demographics, characterized in that: Includes the following steps: Step 1: Select users who meet the preset conversion conditions from the user database of the e-commerce platform as seed users, obtain the behavioral event sequence of each seed user, extract features from the behavioral event sequence and input it into the sequence encoder to obtain the seed user representation vector corresponding to the seed user, and construct the seed representation set by combining all the seed user representation vectors. Step 2: Construct a conditional diffusion generation model. The conditional diffusion generation model includes a forward noise injection module, a reverse denoising generation module, a generation validity gating module, and a diversity constraint module. The forward noise injection module gradually adds noise to the seed user representation vector to obtain a full noise latent vector. The reverse denoising generation module gradually denoises the initial noise vector with the seed user representation vector as a condition to obtain candidate virtual representation vectors. The generation validity gating module and the diversity constraint module screen the candidate virtual representation vectors to obtain virtual similar user representation vectors. All virtual similar user representation vectors constitute a virtual user representation set. Step 3: Merge the seed representation set and the virtual user representation set to obtain the enhanced query representation set. Based on the enhanced query representation set, perform an approximate nearest neighbor search in the vector retrieval index to obtain the initial candidate user set. Perform filtering and constraint processing on the initial candidate user set in sequence to obtain the target audience user set. Adaptively adjust the selection range according to the conversion status data returned by the campaign.

2. The method according to claim 1, characterized in that, In step 1, the preset conversion condition is that the purchase of goods is completed within the current statistical period and the purchase amount reaches the preset amount threshold. The method for generating the behavioral event sequence is as follows: obtain the behavioral event records generated by the seed user within the current statistical period. The behavioral event records include product browsing events, product adding to cart events, product favorite events, order submission events, payment completion events, and product review events. The records are arranged in ascending order according to the timestamp of each behavioral event record to form a behavioral event sequence.

3. The method according to claim 2, characterized in that, In step 1, the feature extraction method for the behavioral event sequence is as follows: for each behavioral event record in the behavioral event sequence, extract the product category identifier, product price range identifier, and product brand identifier of the product associated with the behavioral event record. Concatenate the product category identifier, product price range identifier, and product brand identifier into an event feature vector. Arrange all event feature vectors in chronological order to form an event feature vector sequence. Input the event feature vector sequence into the sequence encoder.

4. The method according to claim 1, characterized in that, In step 2, the execution process of the forward noise injection module is as follows: the seed user representation vector is used as the original input vector, the total number of diffusion steps is set to a positive integer T, a noise intensity sequence containing T elements and monotonically increasing from step 1 to step T is pre-constructed, and the noise superposition operation is performed sequentially from step 1 to step T. In the noise superposition operation of step k, the sampled noise vector of step k is obtained from the standard normal distribution, and the sampled noise vector of step k is multiplied by the kth noise intensity value in the noise intensity sequence to obtain the scaled noise vector of step k. The intermediate vector output of step k-1 is added to the scaled noise vector of step k to obtain the intermediate vector output of step k. The intermediate vector output of step 0 is the original input vector, and the intermediate vector output of step T is used as the full noise latent vector.

5. The method according to claim 4, characterized in that, In step 2, the inverse denoising generation module includes a denoising neural network. The denoising neural network adopts a multilayer perceptron structure, which includes an input concatenation layer, a first hidden layer, a second hidden layer, a third hidden layer, and an output mapping layer connected in series. The input concatenation layer is used to concatenate the vector to be denoised, the conditional vector, and the step number encoding vector into a single input vector. The first, second, and third hidden layers all contain linear transformation units and nonlinear activation units. The output mapping layer maps the output of the third hidden layer into a noise estimation vector.

6. The method according to claim 5, characterized in that, In step 2, the execution process of the reverse denoising generation module is as follows: select one seed user representation vector from the seed representation set as the condition vector, sample the initial noise vector from the standard normal distribution and use it as the denoising vector to be denoised in step T, perform denoising operations sequentially from step T to step 1, in the denoising operation of step j, perform sinusoidal position encoding on the current step number j to obtain the step number encoding vector, input the denoising vector to be denoised in step j, the condition vector and the step number encoding vector into the denoising neural network to output the noise estimation vector of step j, multiply the noise estimation vector of step j with the jth noise intensity value in the noise intensity sequence to obtain the scaled noise estimation vector of step j, subtract the scaled noise estimation vector of step j from the denoising vector to be denoised in step j-1 to obtain the denoising vector to be denoised in step j-1, and use the denoising vector to be denoised in step 0 as the candidate virtual representation vector.

7. The method according to claim 6, characterized in that, In step 2, the execution process of the validity gating module is as follows: calculate the cosine similarity between the candidate virtual representation vector and the condition vector. If the cosine similarity between the candidate virtual representation vector and the condition vector is less than the preset lower limit threshold for validity similarity, discard the candidate virtual representation vector. If the cosine similarity between the candidate virtual representation vector and the condition vector is greater than or equal to the preset lower limit threshold for validity similarity, further calculate the average cosine similarity between the candidate virtual representation vector and all seed user representation vectors in the seed representation set. If the average cosine similarity is less than the preset lower limit threshold for distribution consistency, discard the candidate virtual representation vector. If the average cosine similarity is greater than or equal to the preset lower limit threshold for distribution consistency, determine that the candidate virtual representation vector passes the validity gating.

8. The method according to claim 7, characterized in that, In step 2, the execution process of the diversity constraint module is as follows: For each seed user representation vector, the reverse denoising generation module and the generation validity gating module are repeatedly executed until a preset number of M candidate virtual representation vectors that pass the validity gating are obtained. The M candidate virtual representation vectors that pass the validity gating are used to form a vector group to be screened. The cosine similarity between any two candidate virtual representation vectors in the vector group to be screened is calculated. If the cosine similarity between any two candidate virtual representation vectors is greater than the preset diversity similarity upper limit threshold, one of the two candidate virtual representation vectors is removed, and the remaining candidate virtual representation vectors in the vector group to be screened are retained as virtual similar user representation vectors.

9. The method according to claim 1, characterized in that, In step 3, the filtering process includes deduplication, frequency control, and cold start representation compensation. Deduplication involves retaining only one record for each candidate user that appears repeatedly. Frequency control involves removing candidate users whose cumulative number of ad exposures within the current statistical period has reached a preset frequency limit. Cold start representation compensation involves identifying cold start users whose number of behavioral event records is less than a preset minimum record threshold, querying the user profile database for the preferred product category set of active user groups with the same attributes based on the cold start user's registered region and age range, performing average pooling on the category embedding vectors of each product category in the preferred product category set to obtain a compensation category vector, and concatenating the original representation vector of the cold start user with the compensation category vector and compressing it through a linear mapping layer to obtain the compensated representation vector.

10. The method according to claim 9, characterized in that, In step 3, the constraint processing includes category bias correction processing and regional inventory constraint processing; Category bias correction is performed as follows: The percentage of each product category in the candidate user set is calculated. For product categories whose percentage exceeds the preset category percentage limit, the excess ratio is calculated. The excess ratio is the difference between the actual percentage of the product category and the preset category percentage limit, divided by the actual percentage of the product category. 1 minus the excess ratio is used as the retention probability of candidate users associated with product categories whose percentage exceeds the preset category percentage limit. Bernoulli sampling is then performed based on the retention probability to determine whether to retain candidate users. Regional inventory constraint processing involves removing candidate users whose delivery address is not in the list of deliverable regions or whose corresponding region has zero inventory. The adaptive adjustment method is as follows: calculate the sliding window conversion rate based on the conversion status data, set the hysteresis interval width value, if the sliding window conversion rate is lower than the difference between the lower limit of the conversion rate threshold and the hysteresis interval width value, increase the ranking threshold N of the near nearest neighbor search to narrow the selection range, if the sliding window conversion rate is higher than the sum of the upper limit of the conversion rate threshold and the hysteresis interval width value, decrease the ranking threshold N to expand the selection range, if the sliding window conversion rate is between the difference between the lower limit of the conversion rate threshold and the hysteresis interval width value and the sum of the upper limit of the conversion rate threshold and the hysteresis interval width value, keep the ranking threshold N unchanged.

Images