Cold start user interest prediction and diversion method based on cross-platform mapping

Abstract

This invention discloses a method for cold-start user interest prediction and traffic acquisition based on cross-platform mapping, including the following steps: Step 1: Obtaining the original cross-platform e-commerce data of the target cold-start users; Step 2: Determining the cross-platform mapping relationship and unifying fields of the original cross-platform e-commerce data using weighted Hamming distance; Step 3: Constructing field representations and selectively embedding mappings; Step 4: Predicting cold-start user interests using an improved AutoInt network; Step 5: Performing matching and filtering based on Jensen-Shannon divergence to generate a target traffic acquisition candidate set; Step 6: Determining the target traffic acquisition object and the target traffic acquisition reach path and generating traffic acquisition execution results; Step 7: Updating the parameters of the improved AutoInt network based on feedback behavior. This invention improves the accuracy of cold-start user interest prediction and the stability of traffic acquisition conversion through the improved AutoInt network.

Description

Technical Field

[0001] This invention relates to the field of e-commerce data processing technology, and in particular to a method for predicting and driving traffic to cold-start users based on cross-platform mapping. Background Technology

[0002] With the continuous integration and development of e-commerce platforms, content platforms, social platforms, and search platforms, the demand for interest identification and precise traffic redirection for cold-start users is constantly growing. Intelligent processing technologies for cross-platform user behavior mapping, interest prediction, and traffic redirection decision control have received widespread attention. Existing e-commerce recommendation or traffic distribution systems mainly rely on single-platform behavioral data, static user tags, or simple collaborative filtering methods for cold-start user interest identification. However, these systems commonly suffer from the following problems in practical applications: User behavior records, content interaction records, and product-related features differ significantly across platforms in terms of field structure, semantic expression, and record granularity. Existing cross-platform data association methods often rely solely on account binding or simple field matching for mapping, making it difficult to achieve stable cross-platform mapping under conditions of missing user identifiers, sparse behavior, or strong platform heterogeneity. This results in incomplete basic interest information for cold-start users and insufficient reliability of field-uniform results. Cold-start users on e-commerce platforms exhibit few explicit transaction behaviors and have weak historical preferences that can be directly utilized. Existing interest prediction methods often employ conventional embedding networks or ordinary attention mechanisms to uniformly model input fields, which easily leads to unclear field pairing relationships, weak interest signals being easily interfered with by irrelevant fields, and repeated accumulation of high-order interactions. This results in large fluctuations and insufficient discriminativeness in the prediction results of product category interests, brand interests, price range interests, and activity object interests. During the traffic acquisition execution phase, existing methods typically select candidate products or activity pages directly based on a single rating result. This lacks fine-grained measurement of the differences between the interest prediction distribution and the candidate object tag distribution. It is also difficult to combine platform reach conditions, terminal display conditions, and traffic acquisition acceptance conditions to collaboratively determine the traffic acquisition target and traffic acquisition path, resulting in low accuracy of traffic acquisition target matching, poor adaptability of reach paths, and unstable traffic acquisition conversion effects.

[0003] Therefore, how to provide a method for predicting and driving user interest in cold starts based on cross-platform mapping is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a method for predicting and driving traffic to cold-start users based on cross-platform mapping. This invention fully utilizes weighted Hamming distance, improved AutoInt network, and Jensen-Shannon divergence, and describes in detail the method for predicting and driving traffic to cold-start users in e-commerce. It has the advantages of strong cross-platform interest recognition capability, high accuracy of cold-start prediction, high accuracy of traffic target matching, and stable traffic conversion effect.

[0005] A cold-start user interest prediction and traffic redirection method based on cross-platform mapping according to an embodiment of the present invention includes the following steps: Step 1: Obtain the target cold start users' cross-platform e-commerce raw data; Step 2: Perform cross-platform mapping relationship determination and field unification processing on the original cross-platform e-commerce data using weighted Hamming distance to obtain a standard field feature set; Step 3: Perform field representation construction and selective embedding mapping on the standard field feature set to generate the target embedding vector sequence; Step 4: Input the target embedding vector sequence into the improved AutoInt network, perform cold start user interest prediction, and generate a target user interest prediction probability vector; the improved AutoInt network includes a field pairing generation module, a displacement difference interaction module, an interaction aggregation module, and a prediction output module; the displacement difference interaction module performs pre-displacement difference attention interaction and post-displacement difference attention interaction, and introduces duplicate interaction deletion logic; Step 5: Based on Jensen-Shannon divergence, match and filter the target user interest prediction probability vector with the candidate e-commerce object set to generate a target traffic candidate set that matches the interests and preferences of the target cold start users; Step 6: Based on the target traffic candidate set, determine the target traffic targets and target traffic reach paths, execute traffic processing, and generate traffic execution results; Step 7: Based on the traffic redirection execution results, obtain the feedback behavior of the target cold start users to the target traffic redirection object, and update the parameters of the improved AutoInt network based on the feedback behavior.

[0006] Optionally, the cross-platform e-commerce raw data includes the target cold start user's account identification information, device association information, product browsing information, product click information, product collection information, search history information, product-related content interaction information, promotional activity participation information, and time information on e-commerce platforms, content platforms, social platforms, and search platforms.

[0007] Optionally, step two specifically includes: For each data record in e-commerce platforms, content platforms, social platforms, and search platforms, fields are extracted from the original cross-platform e-commerce data to obtain the cross-platform field encoding sequence of the current data record; The i-th data record from the first platform and the j-th data record from the second platform are selected as candidate record pairs. The first platform and the second platform are two different platforms among e-commerce platform, content platform, social platform and search platform; If the fields of the corresponding dimensions of the two cross-platform field encoding sequences of the candidate record pair are the same, the field difference judgment value is assigned to 0; otherwise, the field difference judgment value is assigned to 1. Statistically analyze the value distribution of each field in all candidate record pairs, and calculate the information entropy of each field based on the value distribution of each field to obtain the field weight of each field; The field difference judgment values ​​and corresponding field weights are weighted and fused to obtain the weighted Hamming distance of candidate record pairs; Calculate the average weighted Hamming distance of all candidate record pairs to obtain the mapping threshold; If the weighted Hamming distance of the current candidate record pair is less than or equal to the mapping threshold, then it is determined that the current candidate record pair has a cross-platform mapping relationship, and a mapping record pair is generated. Based on the mapping record pairs, the corresponding cross-platform e-commerce raw data is uniformly mapped and merged into fields, and then organized into a standard field feature set according to the data record dimension × field dimension structure. The fields include user behavior fields, content preference fields, product attribute fields, brand preference fields, price sensitivity fields, and activity response fields.

[0008] Optionally, step three specifically includes: Field representations are constructed for the user behavior field, content preference field, product attribute field, brand preference field, price sensitivity field, and activity response field respectively, resulting in vectors for user behavior field, content preference field, product attribute field, brand preference field, price sensitivity field, and activity response field. Perform a field retention determination on each field vector to determine which field vectors are retained and which are not. The field retention determination method is as follows: count the number of non-zero code bits in the current field vector. If the number of non-zero code bits is greater than or equal to the average number of non-zero code bits in all field vectors, then the current field vector is determined to be a retained field vector; otherwise, the current field vector is determined to be a non-retained field vector. The retained field vectors are linearly mapped through a trainable field embedding matrix to generate field embedding vectors; For the vector of unretained fields, a missing identifier vector is constructed by one-hot encoding, and the missing identifier vector is indexed and mapped by a trainable placeholder embedding matrix to generate a preset placeholder embedding vector. Arrange the field embedding vector and the preset placeholder embedding vector according to the field order to generate the target embedding vector sequence.

[0009] Optionally, step four specifically includes: In the field pairing generation module, pairwise pairing processing is performed on the embedding vectors in the target embedding vector sequence. If the m-th embedding vector and the n-th embedding vector are not both preset placeholder embedding vectors, then the m-th embedding vector and the n-th embedding vector are concatenated to obtain the paired input vector. The paired input vectors are passed through a linear mapping layer and a GELU activation layer to generate field-paired representation vectors. Arrange all field pairing representation vectors in ascending order of the embedding vector index of the first position and the embedding vector index of the last position to generate a sequence of field pairing representation vectors. The displacement differential interaction module adopts a T-layer stacked displacement differential interaction unit. Each displacement differential interaction unit performs pre-displacement differential attention interaction and post-displacement differential attention interaction on the input field pairing representation vector sequence to generate a displacement differential interaction vector sequence. Between displacement differential interaction units, the displacement differential interaction vector sequence is updated by repeating interaction deletion logic to obtain the field pairing representation vector sequence of the input of the next layer displacement differential interaction unit; The displacement difference interaction vector sequence after repeated interaction deletion and update in layer T is input into the interaction aggregation module, and mapped through the linear mapping layer and GELU activation layer to generate interest prediction feature vector; In the prediction output module, the interest prediction feature vector is passed through a linear mapping layer and the Softmax function to generate the target user interest prediction probability vector; Each dimension of the target user interest prediction probability vector represents the predicted probability value of the target cold-start user's interest in the current interest tag; the interest tags include product category interest tags, brand interest tags, price range interest tags, store interest tags, and activity target interest tags.

[0010] Optionally, the preceding displacement differential attention interaction and the following subsequent displacement differential attention interaction specifically include: Based on the pairing representation vector of the i-th field, select the pairing representation vector of the (i-1)-th field and the pairing representation vector of the (i+1)-th field, and construct the forward displacement difference vector and the backward displacement difference vector through vector difference; In the forward displacement differential attention interaction, the current field pairing representation vector is mapped to the query vector through a trainable query mapping matrix; The forward shift difference vector is mapped to the forward difference key vector and the forward difference value vector respectively through the trainable forward difference key mapping matrix and forward difference value mapping matrix; Based on the query vector, the previous difference key vector, and the previous difference value vector, the attention mechanism is used to generate the previous shift difference interaction vector. In the post-displacement differential attention interaction, the pre-displacement differential interaction vector is added to the current field pairing representation vector to generate an intermediate query vector; The shift difference vector is mapped to the shift key vector and the shift value vector respectively through the trainable shift key mapping matrix and the shift value mapping matrix. Based on the intermediate query vector, the subsequent difference key vector, and the subsequent difference value vector, the attention mechanism is used to generate the subsequent shift difference interaction vector. The front displacement differential interaction vector and the back displacement differential interaction vector are concatenated by features, and the displacement differential interaction vector is generated through a linear mapping layer. The displacement difference interaction vectors are arranged in order to form a displacement difference interaction vector sequence.

[0011] Optionally, the duplicate interaction deletion logic specifically includes: The similarity between the j-th displacement difference interaction vector in the t-th layer displacement difference interaction vector sequence and each displacement difference interaction vector in the (t-1)-th layer displacement difference interaction vector sequence is calculated to obtain the interaction similarity. The maximum interaction similarity is obtained by taking the maximum value of all interaction similarities corresponding to the j-th displacement difference interaction vector. Calculate the mean and standard deviation of all maximum interaction similarities in the displacement difference interaction vector sequence of layer t, and use the sum of the mean and the standard deviation as the interaction similarity threshold; If the maximum interaction similarity of the j-th displacement difference interaction vector is greater than or equal to the interaction similarity threshold, then the j-th displacement difference interaction vector is determined to be a duplicate interaction item and deleted; otherwise, the j-th displacement difference interaction vector is retained. The retained displacement differential interaction vectors are arranged in the original sequence order to generate the field pairing representation vector sequence of the input to the next layer displacement differential interaction unit.

[0012] Optionally, step five specifically includes: The candidate e-commerce objects in the candidate e-commerce object set include candidate products, candidate stores, candidate activity pages, candidate live streaming rooms, and candidate special topic pages, and each candidate e-commerce object corresponds to a candidate object tag response vector. The candidate object label response vector is vector normalized to obtain the candidate object distribution vector; The average vector of the target user interest prediction probability vector and the candidate object distribution vector is used as the intermediate distribution vector; Calculate the KL divergence of the target user interest prediction probability vector relative to the intermediate distribution vector to obtain the first divergence value; Calculate the KL divergence of the candidate object distribution vector relative to the intermediate distribution vector to obtain the second divergence value; Calculate the mean of the first divergence value and the second divergence value to obtain the Jensen-Shannon divergence of the candidate e-commerce object; The average Jensen-Shannon divergence of all candidate e-commerce objects is calculated to obtain the screening statistics. Initialize the target traffic candidate set, which represents a set of candidate e-commerce objects that match the interests and preferences of the target cold start users; If the Jensen-Shannon divergence of a candidate e-commerce object is less than or equal to the screening statistic, then the candidate e-commerce object will be added to the target traffic candidate set.

[0013] Optionally, step six specifically includes: Obtain the platform deployment information, terminal display information, and traffic acquisition information of each candidate e-commerce object in the target traffic acquisition candidate set. Combine the platform access conditions, terminal display conditions, and traffic acquisition conditions of the target cold start users to determine the traffic acquisition object suitability of each candidate e-commerce object and obtain the candidate traffic acquisition object set. Based on the set of candidate traffic acquisition targets, the candidate e-commerce object with the smallest Jensen-Shannon divergence is identified as the target traffic acquisition target; Obtain the path condition information of the target traffic reach path, which includes the site recommendation display path, message push path, search result insertion path, activity page jump path, and live broadcast entrance push path; Based on the platform reach conditions, terminal display conditions, and traffic acquisition conditions of the target cold start users, the path condition information of each traffic acquisition path is used to determine the path adaptation and obtain the target traffic acquisition path of the target traffic acquisition object. Following the target user acquisition path, push the target acquisition object to the target cold start users and generate acquisition execution results.

[0014] Optionally, the feedback behaviors include click behavior, dwell behavior, favorite behavior, add to cart behavior, coupon behavior, order behavior, and repeat purchase behavior; Based on the feedback behavior, construct the feedback supervision result of the target user interest prediction probability vector, and construct parameters to update the loss value according to the difference between the feedback supervision result and the target user interest prediction probability vector. The model parameters of the improved AutoInt network are updated based on the parameter update loss value. The model parameters include: linear mapping layer parameters of the field pairing generation module; query mapping matrix, front difference key mapping matrix, front difference value mapping matrix, back difference key mapping matrix, back difference value mapping matrix and linear mapping layer parameters of the displacement difference interaction module; linear mapping layer parameters of the interaction aggregation module; and linear mapping layer parameters of the prediction output module.

[0015] The beneficial effects of this invention are: This invention performs cross-platform mapping relationship determination and field unification processing on the original cross-platform e-commerce data of target cold-start users. This allows heterogeneous behavioral data from e-commerce platforms, content platforms, social platforms, and search platforms to form a unified standard field feature set, reducing the impact of insufficient data on a single platform and semantic inconsistencies in cross-platform fields on cold-start interest identification. In the interest prediction stage, through field representation construction, selective embedding mapping, and the field pairing generation module, displacement differential interaction module, interaction aggregation module, and prediction output module in the improved AutoInt network, the target embedded vector sequence can explicitly establish field pairing relationships when entering interest prediction. Furthermore, the differential attention interaction between adjacent field pairings is strengthened through pre-displacement differential attention interaction and post-displacement differential attention interaction. At the same time, the repeated interaction deletion logic is combined to suppress the accumulation of repeated interaction items in the multi-layer interaction process, thereby improving the prediction accuracy and discriminativeness of product category interest, brand interest, price range interest, store interest, and activity object interest. In the candidate selection phase, the Jensen-Shannon divergence is used to measure the difference between the predicted probability vector of target user interests and the tag response distribution of candidate e-commerce objects, thereby enhancing the distribution consistency discrimination ability of the matching and selection of candidate products, candidate stores, candidate activity pages, candidate live broadcast rooms, and candidate special pages. In the traffic acquisition execution phase, the target traffic acquisition objects and target traffic acquisition paths are collaboratively determined by combining platform reach conditions, terminal display conditions, and traffic acquisition acceptance conditions, improving the accuracy of traffic acquisition object selection and the rationality of traffic acquisition path adaptation. In the feedback update phase, the model parameters of the improved AutoInt network are updated based on click behavior, dwell behavior, favorite behavior, add-to-cart behavior, coupon redemption behavior, order placement behavior, and repeat purchase behavior, forming a continuous optimization closed loop for cold-start user interest prediction and traffic acquisition. Therefore, this invention is of great significance for improving cross-platform cold-start user interest recognition capabilities, increasing e-commerce traffic acquisition matching accuracy, enhancing traffic acquisition conversion stability, and reducing the cost of manual rule configuration. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of a cold start user interest prediction and traffic redirection method based on cross-platform mapping proposed in this invention; Figure 2 This is a flowchart of the cross-platform mapping relationship determination process in a cold start user interest prediction and traffic generation method based on cross-platform mapping proposed in this invention. Figure 3 This is a flowchart of the improved AutoInt network structure in a cold start user interest prediction and traffic redirection method based on cross-platform mapping proposed in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figures 1-3 A cold-start user interest prediction and traffic generation method based on cross-platform mapping includes the following steps: Step 1: Obtain the target cold start users' cross-platform e-commerce raw data; Step 2: Perform cross-platform mapping relationship determination and field unification processing on the original cross-platform e-commerce data using weighted Hamming distance to obtain a standard field feature set; Step 3: Perform field representation construction and selective embedding mapping on the standard field feature set to generate the target embedding vector sequence; Step 4: Input the target embedding vector sequence into the improved AutoInt network, perform cold start user interest prediction, and generate a target user interest prediction probability vector; the improved AutoInt network includes a field pairing generation module, a displacement difference interaction module, an interaction aggregation module, and a prediction output module; the displacement difference interaction module performs pre-displacement difference attention interaction and post-displacement difference attention interaction, and introduces duplicate interaction deletion logic; Step 5: Based on Jensen-Shannon divergence, match and filter the target user interest prediction probability vector with the candidate e-commerce object set to generate a target traffic candidate set that matches the interests and preferences of the target cold start users; Step 6: Based on the target traffic candidate set, determine the target traffic targets and target traffic reach paths, execute traffic processing, and generate traffic execution results; Step 7: Based on the traffic redirection execution results, obtain the feedback behavior of the target cold start users to the target traffic redirection object, and update the parameters of the improved AutoInt network based on the feedback behavior.

[0019] In this embodiment, the cross-platform e-commerce raw data includes the target cold start user's account identification information, device association information, product browsing information, product click information, product collection information, search history information, product-related content interaction information, promotional activity participation information, and time information on e-commerce platforms, content platforms, social platforms, and search platforms.

[0020] In this embodiment, step two specifically includes: For each data record in e-commerce platforms, content platforms, social platforms, and search platforms, fields are extracted from the original cross-platform e-commerce data to obtain the cross-platform field encoding sequence of the current data record; The i-th data record from the first platform and the j-th data record from the second platform are selected as candidate record pairs. The first platform and the second platform are two different platforms among e-commerce platform, content platform, social platform and search platform; If the fields of the corresponding dimensions of the two cross-platform field encoding sequences of the candidate record pair are the same, the field difference judgment value is assigned to 0; otherwise, the field difference judgment value is assigned to 1. Statistically analyze the value distribution of each field in all candidate record pairs, and calculate the information entropy of each field based on the value distribution of each field to obtain the field weight of each field; The field difference judgment values ​​and corresponding field weights are weighted and fused to obtain the weighted Hamming distance of candidate record pairs; Calculate the average weighted Hamming distance of all candidate record pairs to obtain the mapping threshold; If the weighted Hamming distance of the current candidate record pair is less than or equal to the mapping threshold, then it is determined that the current candidate record pair has a cross-platform mapping relationship, and a mapping record pair is generated. Based on the mapping record pairs, the corresponding cross-platform e-commerce raw data is uniformly mapped and merged into fields, and then organized into a standard field feature set according to the data record dimension × field dimension structure. The fields include user behavior fields, content preference fields, product attribute fields, brand preference fields, price sensitivity fields, and activity response fields.

[0021] In this embodiment, step three specifically includes: Field representations are constructed for the user behavior field, content preference field, product attribute field, brand preference field, price sensitivity field, and activity response field respectively, resulting in vectors for user behavior field, content preference field, product attribute field, brand preference field, price sensitivity field, and activity response field. The field representation construction specifically involves: performing dictionary-based encoding on categorical field values ​​and range-based bucketing encoding on numeric field values; then concatenating the corresponding encoding results according to the internal attribute order of the fields to obtain user behavior field vectors, content preference field vectors, product attribute field vectors, brand preference field vectors, price sensitivity field vectors, and activity response field vectors. Categorical field values ​​include behavior type identifiers, source platform identifiers, content theme category identifiers, product category identifiers, brand identifiers, price range level identifiers, and activity type identifiers. Numerical field values ​​include pageview counts, click counts, favorite counts, search counts, content dwell time, content interaction counts, product price value, product discount value, historical average transaction price, price fluctuation range, activity participation counts, coupon redemption counts, add-to-cart counts, and order counts.

[0022] Perform a field retention determination on each field vector to determine which field vectors are retained and which are not. The field retention determination method is as follows: count the number of non-zero code bits in the current field vector. If the number of non-zero code bits is greater than or equal to the average number of non-zero code bits in all field vectors, then the current field vector is determined to be a retained field vector; otherwise, the current field vector is determined to be a non-retained field vector. In this invention, the purpose of performing field retention determination is to first distinguish between field vectors with relatively sufficient information expression and field vectors with insufficient information expression when cold-start user cross-platform data has sparsity and differences in field completeness, thereby avoiding the indiscriminate entry of all fields into the subsequent embedding mapping process. Since different fields may have missing, weak responses, or insufficient effective coding bits during actual data collection, directly performing conventional embedding mapping on such information-deficient field vectors can easily introduce noisy or invalid coding into the subsequent interest prediction process, reducing the expression quality of the target embedding vector sequence. Through field retention determination, field vectors with an average number of non-zero coding bits are identified as retained field vectors and subjected to normal embedding mapping, while information-deficient field vectors are identified as unretained field vectors and replaced with preset placeholder embedding vectors. This maintains the integrity of field positions while reducing the interference of low-quality fields on subsequent field pairing generation, displacement difference interaction, and interest prediction results, thereby improving the stability and effectiveness of the entire model input representation.

[0023] The retained field vectors are linearly mapped through a trainable field embedding matrix to generate field embedding vectors; For the vector of unretained fields, a missing identifier vector is constructed by one-hot encoding, and the missing identifier vector is indexed and mapped by a trainable placeholder embedding matrix to generate a preset placeholder embedding vector. Arrange the field embedding vector and the preset placeholder embedding vector according to the field order to generate the target embedding vector sequence.

[0024] In this embodiment, step four specifically includes: In the field pairing generation module, pairwise pairing processing is performed on the embedding vectors in the target embedding vector sequence. If the m-th embedding vector and the n-th embedding vector are not both preset placeholder embedding vectors, then the m-th embedding vector and the n-th embedding vector are concatenated to obtain the paired input vector; where m and n represent the indices of the target embedding vector sequence. The paired input vectors are passed through a linear mapping layer and a GELU activation layer to generate field-paired representation vectors. Arrange all field pairing representation vectors in ascending order of the embedding vector index of the first position and the embedding vector index of the last position to generate a sequence of field pairing representation vectors. The displacement differential interaction module adopts a T-layer stacked displacement differential interaction unit. Each displacement differential interaction unit performs pre-displacement differential attention interaction and post-displacement differential attention interaction on the input field pairing representation vector sequence to generate a displacement differential interaction vector sequence. Between displacement differential interaction units, the displacement differential interaction vector sequence is updated by repeating interaction deletion logic to obtain the field pairing representation vector sequence of the input of the next layer displacement differential interaction unit; The displacement difference interaction vector sequence after repeated interaction deletion and update in layer T is input into the interaction aggregation module, and mapped through the linear mapping layer and GELU activation layer to generate interest prediction feature vector; In the prediction output module, the interest prediction feature vector is passed through a linear mapping layer and the Softmax function to generate the target user interest prediction probability vector; Each dimension of the target user interest prediction probability vector represents the predicted probability value of the target cold-start user's interest in the current interest tag; the interest tags include product category interest tags, brand interest tags, price range interest tags, store interest tags, and activity target interest tags.

[0025] In this embodiment, the forward displacement differential attention interaction and the backward displacement differential attention interaction specifically include: Based on the pairing representation vector of the i-th field, select the pairing representation vectors of the (i-1)-th and (i+1)-th fields, and construct a forward displacement difference vector and a backward displacement difference vector through vector difference; where the forward displacement difference vector is the vector difference between the pairing representation vectors of the i-th and (i-1)-th fields, and the backward displacement difference vector is the vector difference between the pairing representation vectors of the (i+1)-th and i-th fields; when the pairing representation vector of the i-th field is the first vector, the pairing representation vector of the (i-1)-th field is set as the zero vector; when the pairing representation vector of the i-th field is the last vector, the pairing representation vector of the (i+1)-th field is set as the zero vector; where i represents the index of the field pairing representation vector sequence. In the forward displacement differential attention interaction, the current field pairing representation vector is mapped to the query vector through a trainable query mapping matrix; The forward shift difference vector is mapped to the forward difference key vector and the forward difference value vector respectively through the trainable forward difference key mapping matrix and forward difference value mapping matrix; Based on the query vector, the previous difference key vector, and the previous difference value vector, the attention mechanism is used to generate the previous shift difference interaction vector. In the post-displacement differential attention interaction, the pre-displacement differential interaction vector is added to the current field pairing representation vector to generate an intermediate query vector; The shift difference vector is mapped to the shift key vector and the shift value vector respectively through the trainable shift key mapping matrix and the shift value mapping matrix. Based on the intermediate query vector, the subsequent difference key vector, and the subsequent difference value vector, the attention mechanism is used to generate the subsequent shift difference interaction vector. The front displacement differential interaction vector and the back displacement differential interaction vector are concatenated by features, and the displacement differential interaction vector is generated through a linear mapping layer. The displacement difference interaction vectors are arranged in order to form a displacement difference interaction vector sequence.

[0026] In this embodiment, the duplicate interaction deletion logic specifically includes: The similarity between the j-th displacement difference interaction vector in the t-th layer displacement difference interaction vector sequence and each displacement difference interaction vector in the (t-1)-th layer displacement difference interaction vector sequence is calculated to obtain the interaction similarity; where t ranges from 2 to T; j is the index of the displacement difference interaction vector sequence. The maximum interaction similarity is obtained by taking the maximum value of all interaction similarities corresponding to the j-th displacement difference interaction vector. Calculate the mean and standard deviation of all maximum interaction similarities in the displacement difference interaction vector sequence of layer t, and use the sum of the mean and the standard deviation as the interaction similarity threshold; If the maximum interaction similarity of the j-th displacement difference interaction vector is greater than or equal to the interaction similarity threshold, then the j-th displacement difference interaction vector is determined to be a duplicate interaction item and deleted; otherwise, the j-th displacement difference interaction vector is retained. The retained displacement differential interaction vectors are arranged in the original sequence order to generate the field pairing representation vector sequence of the input to the next layer displacement differential interaction unit.

[0027] In this invention, the improved AutoInt network maintains consistency with the standard AutoInt network in terms of overall task localization and basic structural logic. Both belong to interest prediction networks oriented towards modeling the interaction of structured field features, and both are based on the embedded representation of input features, obtaining user interest prediction results through interaction modeling between features. The standard AutoInt network typically includes an input layer, an embedding layer, a feature interaction layer, and an output layer. The input layer receives features from various fields, the embedding layer maps discrete or numerical fields into embedding vectors of a uniform dimension, the feature interaction layer models the correlation between embedding vectors of different fields through a self-attention mechanism to obtain a high-order feature interaction representation, and the output layer generates the target prediction result based on the interactive feature representation.

[0028] The improved AutoInt network does not directly adopt the structure of the standard AutoInt network where field embedding vectors directly enter the self-attention interaction layer. Instead, it reconstructs the feature interaction path based on the standard AutoInt network. Specifically, the improved AutoInt network adds a field pairing generation module. It first performs pairwise pairing processing on the embedding vectors in the target embedding vector sequence, and then generates field pairing representation vectors through feature concatenation, a linear mapping layer, and a GELU activation layer. These vectors are then arranged in ascending order of the previous and subsequent position indices to form a sequence of field pairing representation vectors. In the interaction modeling stage, the improved AutoInt network no longer uses the conventional self-attention interaction method of the standard AutoInt network. Instead, it employs a displacement difference interaction module. Through pre-displacement difference attention interaction and post-displacement difference attention interaction, it constructs pre-displacement difference vectors and post-displacement difference vectors using the vector differences between adjacent field pairing representations, and completes the difference attention interaction separately. Simultaneously, a duplicate interaction deletion logic is introduced between multiple layers of displacement difference interaction units. This identifies and deletes highly repetitive interaction items between the current layer and the previous layer, and then inputs the retained displacement difference interaction vectors into the next layer.

[0029] Through the aforementioned improvements, the enhanced AutoInt network is better suited for cold-start user interest prediction in e-commerce scenarios. The field pairing generation module embeds previously independent fields into explicit field pairing representations, which helps to more directly characterize the combined relationships between user behavior fields, content preference fields, product attribute fields, brand preference fields, price sensitivity fields, and activity response fields. The displacement difference interaction module utilizes the difference information between adjacent field pairing representations to perform attention interactions. Compared to the standard AutoInt network's conventional global field interaction method, it can better highlight the differences and changing trends between adjacent pairings, thereby enhancing the ability to characterize local interest evolution features and weak interest association features. The duplicate interaction deletion logic can suppress the continuous accumulation of highly repetitive interaction items during multi-layer interaction, avoiding redundant interaction information from interfering with subsequent interest prediction. In summary, the enhanced AutoInt network improves the ability to model field combination relationships, extract differential interactions, and enhance the effectiveness of multi-layer interaction results compared to the standard AutoInt network, thus improving the accuracy, stability, and discriminativeness of cold-start user interest prediction.

[0030] In this embodiment, step five specifically includes: The candidate e-commerce objects in the candidate e-commerce object set include candidate products, candidate stores, candidate activity pages, candidate live streaming rooms, and candidate special topic pages, and each candidate e-commerce object corresponds to a candidate object tag response vector. The candidate object label response vector is vector normalized to obtain the candidate object distribution vector; The average vector of the target user interest prediction probability vector and the candidate object distribution vector is used as the intermediate distribution vector; Calculate the KL divergence of the target user interest prediction probability vector relative to the intermediate distribution vector to obtain the first divergence value; Calculate the KL divergence of the candidate object distribution vector relative to the intermediate distribution vector to obtain the second divergence value; Calculate the mean of the first divergence value and the second divergence value to obtain the Jensen-Shannon divergence of the candidate e-commerce object; The average Jensen-Shannon divergence of all candidate e-commerce objects is calculated to obtain the screening statistics. Initialize the target traffic candidate set, which represents a set of candidate e-commerce objects that match the interests and preferences of the target cold start users; If the Jensen-Shannon divergence of a candidate e-commerce object is less than or equal to the screening statistic, then the candidate e-commerce object will be added to the target traffic candidate set.

[0031] In this invention, Jensen-Shannon divergence is more suitable for measuring the distributional consistency between the predicted probability vector of target user interests and the distribution vector of candidate e-commerce objects. Compared to directly using Euclidean distance, cosine similarity, or one-way KL divergence, Jensen-Shannon divergence simultaneously considers the deviations of both the target user interest distribution and the candidate object tag distribution from the intermediate distribution vector, thus more stably reflecting the similarity between the two in the overall probability distribution shape. Furthermore, Jensen-Shannon divergence is a symmetric measure, and will not produce different results due to different comparison directions, making it more suitable for bidirectional matching determination between user interest distribution and candidate object distribution. At the same time, Jensen-Shannon divergence is insensitive to local probability fluctuations, reducing the interference caused by abnormally large or small probability values ​​of single interest tags on the screening results, thereby making the screening process for candidate products, candidate stores, candidate activity pages, candidate live broadcast rooms, and candidate special pages smoother. Based on this, using Jensen-Shannon divergence can improve the matching accuracy and screening stability of candidate e-commerce objects and target cold-start user interests.

[0032] In this embodiment, step six specifically includes: Obtain the platform deployment information, terminal display information, and traffic acquisition information of each candidate e-commerce object in the target traffic acquisition candidate set. Combine the platform access conditions, terminal display conditions, and traffic acquisition conditions of the target cold start users to determine the traffic acquisition object suitability of each candidate e-commerce object and obtain the candidate traffic acquisition object set. Among them, the platform reach conditions are determined based on the target cold start user's account identification information, platform interaction record information, and time information. The platform reach conditions include platform activity indicators and platform login status indicators. The terminal display conditions are determined based on the target cold start user's device association information. The terminal display conditions include terminal type indicators and terminal online status indicators. The traffic acquisition conditions are determined based on the platform deployment information of each candidate e-commerce object, page accessibility information, and the target cold start user's current platform access status. The traffic acquisition conditions include page accessibility indicators and object entry availability indicators. Based on the set of candidate traffic acquisition targets, the candidate e-commerce object with the smallest Jensen-Shannon divergence is identified as the target traffic acquisition target; Obtain the path condition information of the target traffic reach path, which includes the site recommendation display path, message push path, search result insertion path, activity page jump path, and live broadcast entrance push path; Based on the platform reach conditions, terminal display conditions, and traffic acquisition conditions of the target cold start users, the path condition information of each traffic acquisition path is used to determine the path adaptation and obtain the target traffic acquisition path of the target traffic acquisition object. Following the target user acquisition path, push the target acquisition object to the target cold start users and generate acquisition execution results.

[0033] In this embodiment, feedback behaviors include click behavior, dwell behavior, favorite behavior, add to cart behavior, coupon behavior, order behavior, and repeat purchase behavior; Based on the feedback behavior, construct the feedback supervision result of the target user interest prediction probability vector, and construct parameters to update the loss value according to the difference between the feedback supervision result and the target user interest prediction probability vector. The model parameters of the improved AutoInt network are updated based on the parameter update loss value. The model parameters include: linear mapping layer parameters of the field pairing generation module; query mapping matrix, front difference key mapping matrix, front difference value mapping matrix, back difference key mapping matrix, back difference value mapping matrix and linear mapping layer parameters of the displacement difference interaction module; linear mapping layer parameters of the interaction aggregation module; and linear mapping layer parameters of the prediction output module.

[0034] Example 1: To verify the feasibility of this invention in practice, the method of this invention was applied to the cold start user acquisition scenario of a large comprehensive e-commerce platform. This platform simultaneously operates a product mall, short video content channels, brand event venues, and live-streaming shopping modules, with an average of approximately 86,000 new registered users per day. A significant portion of these new users only engage in limited browsing, searching, or content dwelling, and have not yet established a stable ordering pattern. Traditional single-platform recommendation methods struggle to accurately identify their interests and preferences, often resulting in issues such as misaligned recommended product categories, inaccurate matching of event pages, ineffective access to live-streaming entry points, and low conversion rates for adding users to cart after acquisition. Especially during major platform promotions, the sources of new users are complex, including users imported through search ads, users redirected from content communities, and users acquired through social sharing. The explicit behaviors and dispersed interest expressions of users from different sources within the platform are sparse, leading to problems such as missing interest tags, distorted tag distribution, and unstable selection of acquisition targets when simply constructing interest tags based on in-site click records.

[0035] In this invention, the platform selects 128,000 cold-start users who have newly registered within 45 consecutive days and have zero historical order counts as the target group, and simultaneously accesses cross-platform behavioral data from e-commerce platforms, content platforms, social platforms, and search platforms. The collected data includes account identification information, device association information, product browsing information, product click information, product collection information, search history information, product-related content interaction information, promotional activity participation information, and time information. First, these cross-platform e-commerce raw data undergo field extraction and encoding processing. Then, according to the method described in the claims, weighted Hamming distance is used to determine the cross-platform mapping relationship. Data records with confirmed mapping relationships are merged into unified user records and then organized into a standard field feature set. Further, field representations are constructed and selectively embedded for user behavior fields, content preference fields, product attribute fields, brand preference fields, price sensitivity fields, and activity response fields to generate target embedding vector sequences, which are then input into an improved AutoInt network.

[0036] In practical applications, the field pairing generation module first pairs the embedded valid fields to avoid unclear pairing relationships caused by using only a single field for interest judgment. The displacement differential interaction module strengthens the interest difference relationship between adjacent field pairings through pre-displacement differential attention interaction and post-displacement differential attention interaction. For example, weak interest combinations such as "browsing the maternal and infant category - searching for strollers" and "staying over beauty content - clicking on brand pages" can be further amplified. The duplicate interaction deletion logic removes duplicate interaction items between layers, reducing the problem of information redundancy accumulation after multiple layers of interaction. After the model outputs the probability distribution of product category interest tags, brand interest tags, price range interest tags, store interest tags, and activity object interest tags, it then uses Jensen-Shannon divergence to match and filter the tag response distribution of candidate products, candidate stores, candidate activity pages, candidate live broadcast rooms, and candidate special pages. Finally, it combines platform reach conditions, terminal display conditions, and traffic acquisition conditions to determine the actual traffic acquisition targets and traffic acquisition paths. The platform will distribute the selected target objects to homepage recommendation slots, message push slots, search insertion slots, and activity page jump entrances, and continuously record feedback behaviors such as clicks, dwell time, favorites, add to cart, coupon redemption, and order placement for subsequent model parameter updates.

[0037] To demonstrate how this invention addresses the issues of unclear user interest during cold starts, unstable cross-platform mapping, and low conversion rates, this embodiment sets up three comparison schemes. Comparison scheme one uses a single-platform collaborative filtering method, employing only browsing and click data from within the e-commerce site; comparison scheme two uses a standard embedded network plus fully connected prediction method, using cross-platform data but without field pairing generation or shift differential interaction; comparison scheme three uses the standard AutoInt method, employing unified embedding and conventional self-attention interaction; the experimental group uses the method of this invention. The comparison results are shown in Table 1.

[0038] Table 1. Comparison of the effectiveness of different methods in cold start user interest prediction and traffic acquisition scenarios.

[0039] As shown in Table 1, the method of this invention outperforms the comparative scheme in all evaluation indicators, indicating that the present invention has better overall performance in cold start user interest prediction and traffic generation applications. Compared with the first comparative scheme, the cross-platform mapping accuracy of the method of this invention has increased from 71.8% to 92.7%, indicating that by performing cross-platform mapping relationship determination and field unification processing through weighted Hamming distance, the association and merging of data between heterogeneous platforms can be completed more effectively; the interest tag prediction accuracy has increased from 63.4% to 84.9%, and the interest tag prediction F1 value has increased from 0.612 to 0.836, indicating that by constructing field representations, selectively embedding mapping, and using the field pairing generation module, displacement difference interaction module, and duplicate interaction deletion logic in the improved AutoInt network, the potential interest features of cold start users can be extracted more accurately and the interest tag discrimination ability can be improved.

[0040] Compared to Comparative Scheme 2 and Comparative Scheme 3, the method of the present invention improves the candidate object matching hit rate by 13.7 percentage points and 7.8 percentage points, respectively; the click-through rate of traffic generation by 3.2 percentage points and 2.1 percentage points, respectively; the add-to-cart conversion rate by 2.2 percentage points and 1.5 percentage points, respectively; the coupon conversion rate by 2.5 percentage points and 1.8 percentage points, respectively; and the order conversion rate by 1.32 percentage points and 0.94 percentage points, respectively. This indicates that the present invention can not only improve the accuracy of interest prediction results, but also further improve the precision of candidate e-commerce object screening and the actual conversion effect after traffic generation.

[0041] Therefore, this invention can effectively solve the problems of difficulty in uniformly mapping cross-platform data, inaccurate characterization of cold start user interests, and low matching accuracy of traffic redirection targets in the prior art, thereby significantly improving the accuracy, stability, and practical application effect of cold start user interest prediction and traffic redirection processing.

[0042] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for cold-start user interest prediction and traffic generation based on cross-platform mapping, characterized in that, Includes the following steps: Step 1: Obtain the target cold start users' cross-platform e-commerce raw data; Step 2: Perform cross-platform mapping relationship determination and field unification processing on the original cross-platform e-commerce data using weighted Hamming distance to obtain a standard field feature set; Step 3: Perform field representation construction and selective embedding mapping on the standard field feature set to generate the target embedding vector sequence; Step 4: Input the target embedding vector sequence into the improved AutoInt network, perform cold start user interest prediction, and generate a target user interest prediction probability vector; the improved AutoInt network includes a field pairing generation module, a displacement difference interaction module, an interaction aggregation module, and a prediction output module; the displacement difference interaction module performs pre-displacement difference attention interaction and post-displacement difference attention interaction, and introduces duplicate interaction deletion logic; Step 5: Based on Jensen-Shannon divergence, match and filter the target user interest prediction probability vector with the candidate e-commerce object set to generate a target traffic candidate set that matches the interests and preferences of the target cold start users; Step 6: Based on the target traffic candidate set, determine the target traffic targets and target traffic reach paths, execute traffic processing, and generate traffic execution results; Step 7: Based on the traffic redirection execution results, obtain the feedback behavior of the target cold start users to the target traffic redirection object, and update the parameters of the improved AutoInt network based on the feedback behavior.

2. The method for cold-start user interest prediction and traffic redirection based on cross-platform mapping according to claim 1, characterized in that, The cross-platform e-commerce raw data includes the target cold start user's account identification information, device association information, product browsing information, product click information, product collection information, search history information, product-related content interaction information, promotional activity participation information, and time information on e-commerce platforms, content platforms, social platforms, and search platforms.

3. The method for cold start user interest prediction and traffic redirection based on cross-platform mapping according to claim 1, characterized in that, Step two specifically includes: For each data record in e-commerce platforms, content platforms, social platforms, and search platforms, fields are extracted from the original cross-platform e-commerce data to obtain the cross-platform field encoding sequence of the current data record; The i-th data record from the first platform and the j-th data record from the second platform are selected as candidate record pairs. The first platform and the second platform are two different platforms among e-commerce platform, content platform, social platform and search platform; If the fields of the corresponding dimensions of the two cross-platform field encoding sequences of the candidate record pair are the same, the field difference judgment value is assigned to 0; otherwise, the field difference judgment value is assigned to 1. Statistically analyze the value distribution of each field in all candidate record pairs, and calculate the information entropy of each field based on the value distribution of each field to obtain the field weight of each field; The field difference judgment values ​​and corresponding field weights are weighted and fused to obtain the weighted Hamming distance of candidate record pairs; Calculate the average weighted Hamming distance of all candidate record pairs to obtain the mapping threshold; If the weighted Hamming distance of the current candidate record pair is less than or equal to the mapping threshold, then it is determined that the current candidate record pair has a cross-platform mapping relationship, and a mapping record pair is generated. Based on the mapping record pairs, the corresponding cross-platform e-commerce raw data is uniformly mapped and merged into fields, and then organized into a standard field feature set according to the data record dimension × field dimension structure. The fields include user behavior fields, content preference fields, product attribute fields, brand preference fields, price sensitivity fields, and activity response fields.

4. The method for cold start user interest prediction and traffic redirection based on cross-platform mapping according to claim 1, characterized in that, Step three specifically includes: Field representations are constructed for the user behavior field, content preference field, product attribute field, brand preference field, price sensitivity field, and activity response field respectively, resulting in vectors for user behavior field, content preference field, product attribute field, brand preference field, price sensitivity field, and activity response field. Perform a field retention determination on each field vector to determine which field vectors are retained and which are not. The field retention determination method is as follows: count the number of non-zero code bits in the current field vector. If the number of non-zero code bits is greater than or equal to the average number of non-zero code bits in all field vectors, then the current field vector is determined to be a retained field vector; otherwise, the current field vector is determined to be a non-retained field vector. The retained field vectors are linearly mapped through a trainable field embedding matrix to generate field embedding vectors; For the vector of unretained fields, a missing identifier vector is constructed by one-hot encoding, and the missing identifier vector is indexed and mapped by a trainable placeholder embedding matrix to generate a preset placeholder embedding vector. Arrange the field embedding vector and the preset placeholder embedding vector according to the field order to generate the target embedding vector sequence.

5. The method for cold-start user interest prediction and traffic redirection based on cross-platform mapping according to claim 1, characterized in that, Step four specifically includes: In the field pairing generation module, pairwise pairing processing is performed on the embedding vectors in the target embedding vector sequence. If the m-th embedding vector and the n-th embedding vector are not both preset placeholder embedding vectors, then the m-th embedding vector and the n-th embedding vector are concatenated to obtain the paired input vector. The paired input vectors are passed through a linear mapping layer and a GELU activation layer to generate field-paired representation vectors. Arrange all field pairing representation vectors in ascending order of the embedding vector index of the first position and the embedding vector index of the last position to generate a sequence of field pairing representation vectors. The displacement differential interaction module adopts a T-layer stacked displacement differential interaction unit. Each displacement differential interaction unit performs pre-displacement differential attention interaction and post-displacement differential attention interaction on the input field pairing representation vector sequence to generate a displacement differential interaction vector sequence. Between displacement differential interaction units, the displacement differential interaction vector sequence is updated by repeating interaction deletion logic to obtain the field pairing representation vector sequence of the input of the next layer displacement differential interaction unit; The displacement difference interaction vector sequence after repeated interaction deletion and update in layer T is input into the interaction aggregation module, and mapped through the linear mapping layer and GELU activation layer to generate interest prediction feature vector; In the prediction output module, the interest prediction feature vector is passed through a linear mapping layer and the Softmax function to generate the target user interest prediction probability vector; Each dimension of the target user interest prediction probability vector represents the predicted probability value of the target cold-start user's interest in the current interest tag; the interest tags include product category interest tags, brand interest tags, price range interest tags, store interest tags, and activity target interest tags.

6. The method for cold start user interest prediction and traffic redirection based on cross-platform mapping according to claim 1, characterized in that, The aforementioned forward displacement differential attention interaction and backward displacement differential attention interaction specifically include: Based on the pairing representation vector of the i-th field, select the pairing representation vector of the (i-1)-th field and the pairing representation vector of the (i+1)-th field, and construct the forward displacement difference vector and the backward displacement difference vector through vector difference; In the forward displacement differential attention interaction, the current field pairing representation vector is mapped to the query vector through a trainable query mapping matrix; The forward shift difference vector is mapped to the forward difference key vector and the forward difference value vector respectively through the trainable forward difference key mapping matrix and forward difference value mapping matrix; Based on the query vector, the previous difference key vector, and the previous difference value vector, the attention mechanism is used to generate the previous shift difference interaction vector. In the post-displacement differential attention interaction, the pre-displacement differential interaction vector is added to the current field pairing representation vector to generate an intermediate query vector; The shift difference vector is mapped to the shift key vector and the shift value vector respectively through the trainable shift key mapping matrix and the shift value mapping matrix. Based on the intermediate query vector, the subsequent difference key vector, and the subsequent difference value vector, the attention mechanism is used to generate the subsequent shift difference interaction vector. The front displacement differential interaction vector and the back displacement differential interaction vector are concatenated by features, and the displacement differential interaction vector is generated through a linear mapping layer. The displacement difference interaction vectors are arranged in order to form a displacement difference interaction vector sequence.

7. The method for cold start user interest prediction and traffic redirection based on cross-platform mapping according to claim 1, characterized in that, The duplicate interaction deletion logic specifically includes: The similarity between the j-th displacement difference interaction vector in the t-th layer displacement difference interaction vector sequence and each displacement difference interaction vector in the (t-1)-th layer displacement difference interaction vector sequence is calculated to obtain the interaction similarity. The maximum interaction similarity is obtained by taking the maximum value of all interaction similarities corresponding to the j-th displacement difference interaction vector. Calculate the mean and standard deviation of all maximum interaction similarities in the displacement difference interaction vector sequence of layer t, and use the sum of the mean and the standard deviation as the interaction similarity threshold; If the maximum interaction similarity of the j-th displacement difference interaction vector is greater than or equal to the interaction similarity threshold, then the j-th displacement difference interaction vector is determined to be a duplicate interaction item and deleted; otherwise, the j-th displacement difference interaction vector is retained. The retained displacement differential interaction vectors are arranged in the original sequence order to generate the field pairing representation vector sequence of the input to the next layer displacement differential interaction unit.

8. The method for cold start user interest prediction and traffic redirection based on cross-platform mapping according to claim 1, characterized in that, Step five specifically includes: The candidate e-commerce objects in the candidate e-commerce object set include candidate products, candidate stores, candidate activity pages, candidate live streaming rooms, and candidate special topic pages, and each candidate e-commerce object corresponds to a candidate object tag response vector. The candidate object label response vector is vector normalized to obtain the candidate object distribution vector; The average vector of the target user interest prediction probability vector and the candidate object distribution vector is used as the intermediate distribution vector; Calculate the KL divergence of the target user interest prediction probability vector relative to the intermediate distribution vector to obtain the first divergence value; Calculate the KL divergence of the candidate object distribution vector relative to the intermediate distribution vector to obtain the second divergence value; Calculate the mean of the first divergence value and the second divergence value to obtain the Jensen-Shannon divergence of the candidate e-commerce object; The average Jensen-Shannon divergence of all candidate e-commerce objects is calculated to obtain the screening statistics. Initialize the target traffic candidate set, which represents a set of candidate e-commerce objects that match the interests and preferences of the target cold start users; If the Jensen-Shannon divergence of a candidate e-commerce object is less than or equal to the screening statistic, then the candidate e-commerce object will be added to the target traffic candidate set.

9. A method for cold-start user interest prediction and traffic redirection based on cross-platform mapping according to claim 1, characterized in that, Step six specifically includes: Obtain the platform deployment information, terminal display information, and traffic acquisition information of each candidate e-commerce object in the target traffic acquisition candidate set. Combine the platform access conditions, terminal display conditions, and traffic acquisition conditions of the target cold start users to determine the traffic acquisition object suitability of each candidate e-commerce object and obtain the candidate traffic acquisition object set. Based on the set of candidate traffic acquisition targets, the candidate e-commerce object with the smallest Jensen-Shannon divergence is identified as the target traffic acquisition target; Obtain the path condition information of the target traffic reach path, which includes the site recommendation display path, message push path, search result insertion path, activity page jump path, and live broadcast entrance push path; Based on the platform reach conditions, terminal display conditions, and traffic acquisition conditions of the target cold start users, the path condition information of each traffic acquisition path is used to determine the path adaptation and obtain the target traffic acquisition path of the target traffic acquisition object. Following the target user acquisition path, push the target acquisition object to the target cold start users and generate acquisition execution results.

10. A method for predicting and directing cold-start user interests based on cross-platform mapping according to claim 1, characterized in that, The feedback behaviors include click behavior, dwell behavior, favorite behavior, add to cart behavior, coupon behavior, order behavior, and repeat purchase behavior; Based on the feedback behavior, construct the feedback supervision result of the target user interest prediction probability vector, and construct parameters to update the loss value according to the difference between the feedback supervision result and the target user interest prediction probability vector. The model parameters of the improved AutoInt network are updated based on the parameter update loss value. The model parameters include: linear mapping layer parameters of the field pairing generation module; query mapping matrix, front difference key mapping matrix, front difference value mapping matrix, back difference key mapping matrix, back difference value mapping matrix and linear mapping layer parameters of the displacement difference interaction module; linear mapping layer parameters of the interaction aggregation module; and linear mapping layer parameters of the prediction output module.

Images