Advertisement pushing method based on e-commerce big data feedback

Abstract

The invention belongs to the technical field of advertisement pushing, and particularly relates to an advertisement pushing method based on e-commerce big data feedback, which can extract a user intention snapshot in real time by constructing a multi-dimensional spatiotemporal behavior matrix and combining a sliding time window dynamic slicing technology. And intention transition events such as interest point transfer and decision intensity change are accurately identified by virtue of the trained intention transition detection model, so that user demand fluctuation can be followed up in real time, advertisement pushing is highly matched with the real-time intention of the user, the problem that pushing is disjointed with the real-time demand is effectively solved, and the accuracy of advertisement pushing is remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of advertising push technology, specifically relating to an advertising push method based on e-commerce big data feedback. Background Technology

[0002] In existing technologies, most ad pushes are based on fixed ad matching using users' historical behavior tags, which cannot capture the dynamic changes in user intent in real time. This causes the ad push system to mistakenly believe that the user's intent has not changed and continue to push the original category of products, resulting in a disconnect between the push and the user's real-time needs. To address this, an improved ad push method based on e-commerce big data feedback has been designed. Summary of the Invention

[0003] In view of the above-mentioned shortcomings in the existing technology, the present invention provides an advertising push method based on e-commerce big data feedback to solve the problems in the background technology.

[0004] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The advertising push method based on e-commerce big data feedback includes the following steps: S1: Collect real-time behavioral data streams of users on e-commerce platforms and construct a multi-dimensional spatiotemporal behavioral matrix based on these data streams. S2: The multidimensional spatiotemporal behavior matrix is ​​dynamically sliced ​​using a sliding time window. Each slice corresponds to a user intent snapshot. Intent transition events between adjacent slices are identified through an intent transition detection model. The intent transition events include interest point shift, decision intensity change, and purchase intent generation or decay. S3: Based on the user intent snapshot, generate a candidate set containing real push ads and virtual ads; S4: Push the candidate set to the user and capture their interaction feedback. By comparing the differences in user feedback to real ads and virtual ads, quantify the certainty and openness of user intent. S5: Adjust the push strategy based on the nature of the user's intent and the intent transition event.

[0005] Preferably, in step S1, the method for constructing the multidimensional spatiotemporal behavior matrix is ​​as follows: Receive real-time behavioral data streams from e-commerce platforms, analyze each behavioral event, and generate standard event tuples. The standard event tuples include: user identifier, behavior type, absolute timestamp of the behavior occurrence, page uniform resource identifier, product category code, and context information. Using seconds as the basic unit, the absolute timestamp is mapped to a consecutive second-level time slice sequence number starting from the base time, and multiple behavioral events occurring within the same second are classified into the same second-level time slice; A unified virtual shopping space coordinate system is constructed, which maps different page Uniform Resource Identifiers to first-dimensional coordinate values, maps different product category codes to second-dimensional coordinate values, and represents the user's "physical location" in a single browsing behavior as a two-dimensional coordinate point in the virtual shopping space. Each standard event tuple is transformed into a multidimensional feature vector, with the user identifier as the first dimension index, the second-level time slice sequence number as the second dimension index, and the multidimensional feature vector as the third dimension element to generate a three-dimensional spatiotemporal behavior matrix.

[0006] Preferably, the feature vector includes: a spatial coordinate vector composed of page coordinates and category coordinates, a behavior type vector that performs one-hot encoding of behavior types, and a context feature vector that performs numerical encoding of context information.

[0007] Preferably, in step S2, generating a user intent snapshot specifically involves: Within the sliding window, a fixed-length window size and a sliding step size are preset. Starting from the starting time point of the multidimensional spatiotemporal behavior matrix, continuous duration behavior data is sequentially extracted at intervals of the sliding step size to form a slice. The behavior event data within each slice forms a sub-matrix. The sub-matrix is ​​used as the input of the intent transition detection model at the corresponding time point, and the intent transition detection model generates a snapshot of the user intent at that time point.

[0008] Preferably, the training method for the intent transition detection model is as follows: From historical user behavior logs, the historical user behavior is sliced ​​using a sliding window to obtain time-series behavior slice pairs, which are then divided into training and validation sets. Based on the behavior that ultimately occurs within each time window slice, the main intent type of the slice is determined, and intent graph labels are generated; intent intensity labels are obtained by quantifying and assigning values ​​based on the density, duration, and sequence consistency of the behavior within the slice; and intent transition labels are obtained by comparing the main intent categories of adjacent slices to determine whether a behavior transition has occurred and the type of behavior transition. The parameters of the intent transition detection model are initialized. Slices and corresponding intention graph labels, intent intensity labels, and intent transition labels are selected from the training set. The intention graph classification loss, intent intensity regression loss, and intent transition detection loss are calculated for the intention graph labels, intent intensity labels, and intent transition labels, respectively. The intention graph classification loss, intent intensity regression loss, and intent transition detection loss are linearly combined to obtain the joint loss value. Through gradient backpropagation, the partial derivatives of the joint loss value with respect to the trainable parameters in the intent transition detection model are calculated in sequence to form a gradient vector. The parameters of the intent transition detection model are updated using an estimation algorithm. The model is considered to have converged when the overall performance index of the intent transition detection model on the validation set does not exceed a preset threshold within 10 consecutive cycles.

[0009] Preferably, in step S3, the real advertisement generation process is as follows: The user intent snapshot is input into the global ad pool. Multiple candidate ads with the highest semantic matching degree with the current intent snapshot are retrieved. The matching degree score between the multiple candidate ads and the user intent snapshot is calculated. The ad with the highest score is selected as the actual push ad.

[0010] Preferably, the virtual advertisement generation process is as follows: An offset is applied to the user intent snapshot, shifting it towards a neighboring intent direction to obtain an offset intent snapshot. This offset intent snapshot is used as input to the global ad pool. A pre-selected virtual ad set is retrieved from the global ad pool and output. Ads that are comparable to the real push ads in terms of category, price range, or brand but different in specific products are selected from the pre-selected virtual ad set and used as virtual ads. The real push ads and virtual ads are combined to form the final candidate set to be pushed to the user.

[0011] Preferably, in step S4, the quantification of the determinism and openness of user intent specifically refers to: The system captures user interactions with each ad in the candidate set. Based on the different scores for each interaction, the scores for each ad's interactions are summed to obtain the ad's overall feedback score. The average overall feedback score of the real-world ads and the average overall feedback score of all virtual ads are then calculated. Finally, a feedback difference metric is calculated based on these two average overall feedback scores. The calculation formula is: in The average overall feedback score for genuinely pushed advertisements, the The average overall feedback score for all virtual ads. The feedback difference measure is a very small positive number. The intention deterministic quantization value is obtained by mapping using the Sigmoid function. Calculate the intention openness quantification value based on the intention certainty quantification value. The calculation formula is: Based on the intention to determine the quantification value The user's current intent is determined by comparing it with the first threshold and the second threshold.

[0012] Preferably, the user's current intent includes: When the intention is deterministically quantified When the value exceeds the first preset threshold, it is determined that the user's current intent is highly certain, meaning they prefer genuine advertisements. When the intention is deterministically quantified When the value is below the second preset threshold, it is determined that the user's current intention is highly open, that is, they are more receptive to virtual advertisements; When the intention is deterministically quantified When the value is between the first threshold and the second threshold, the user's intent is determined to be in a state of fuzzy transition.

[0013] An advertising push system based on e-commerce big data feedback includes: The intent flow modeling module is used to collect real-time behavioral data streams of users on e-commerce platforms and construct a multi-dimensional spatiotemporal behavioral matrix based on these data streams. The dynamic slicing analysis module is used to dynamically slice the multidimensional spatiotemporal behavior matrix using a sliding time window. Each slice corresponds to a user intent snapshot. The intent transition detection model identifies intent transition events between adjacent slices. These intent transition events include interest point shifts, changes in decision intensity, and the generation or fading of purchase intent. The ad generation module is used to generate a candidate set containing real push ads and virtual ads based on user intent snapshots; An interactive push and feedback capture module is used to push the candidate set to the user and capture their interactive feedback. By comparing the differences in user feedback to real ads and virtual ads, the certainty and openness of user intent are quantified. The strategy module is used to adjust the push strategy based on the nature of the user's intent and the intent transition event.

[0014] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention constructs a multi-dimensional spatiotemporal behavior matrix and combines it with a sliding time window dynamic slicing technique to extract user intent snapshots in real time. It also uses a trained intent change detection model to accurately identify intent transition events such as interest point shifts and changes in decision intensity. This allows for real-time tracking of user demand fluctuations, ensuring that ad pushes are highly matched with users' real-time intents. This effectively solves the problem of ad pushes being out of sync with real-time needs and significantly improves the accuracy of ad pushes. 2. By setting up candidate sets for real and virtual ads and combining user interaction feedback to calculate quantitative values ​​of intent certainty and openness, different states of user intent are clearly defined. At the same time, based on the combination of intent nature and intent transition events, differentiated push strategies are formulated to realize the transformation from fixed push to personalized dynamic adaptation, thereby improving user experience and conversion efficiency. Attached Figure Description

[0015] Figure 1 A flowchart illustrating an embodiment of an invention for a method of pushing advertisements based on e-commerce big data feedback. Detailed Implementation

[0016] To enable those skilled in the art to better understand the present invention, the technical solution of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0017] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual images. They should not be construed as limiting the scope of this patent. To better illustrate the embodiments of the present invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual dimensions of the product. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0018] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "inner," and "outer" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present patent. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0019] In the description of this invention, unless otherwise explicitly specified and limited, the term "connection" or similar designation indicating a connection between components should be interpreted broadly. For example, it can refer to a fixed connection, a detachable connection, or an integral part; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can refer to the internal communication between two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0020] Example 1: like Figure 1 The advertising push method based on e-commerce big data feedback shown includes, specifically: Intent Flow Modeling Module: Collects real-time behavioral data streams of users on e-commerce platforms. These behavioral data streams include user browsing, clicking, favorites, adding to cart, placing orders, and other behavioral events. Each behavioral event is parsed to generate a standard event tuple. Each tuple contains 6 types of core information, including: user identifier, behavior type, absolute timestamp of the behavior, page unified resource identifier, product category code, and context information. Using the platform system startup time as a preset base time, and starting from the base time, the absolute timestamp of each behavioral event is converted into a continuous second-level time slice sequence number. The time slice sequence number of the first second after the base time is 1, the second second is 2, and so on. Multiple behavioral events occurring within the same second are classified into the same second-level time slice. The page uniform resource identifier is mapped to the first-dimensional coordinate value, and the product category code is mapped to the second-dimensional coordinate value. At this time, the page visited by the user in a browsing behavior and the corresponding product category together constitute a two-dimensional coordinate point in the virtual shopping space, which is used to represent the "spatial location" of the user's behavior. Each standard event tuple is transformed into a multi-dimensional feature vector, which consists of three parts: a spatial coordinate vector composed of page coordinates and category coordinates, a behavior type vector formed by one-hot encoding of behavior types, and a context feature vector. By numerically encoding the context information, mobile phones are encoded as 1 and computers as 2 in device type, and regional information is encoded by province as the corresponding value. Finally, a three-dimensional spatiotemporal behavior matrix is ​​constructed with user identifier as the first dimension index, second-level time slice sequence number as the second dimension index, and multi-dimensional feature vector as the third dimension element.

[0021] Dynamic slice analysis module: Set a fixed size and sliding step size for the sliding time window. Starting from the initial time slice of the multidimensional spatiotemporal behavior matrix, extract consecutive duration behavior data at intervals of the sliding step size to form a slice. The sub-matrix corresponding to each slice contains all the user's behavior feature vectors within that time period. Use this sub-matrix as input to the intent transition detection model to generate a user intent snapshot at the corresponding time point. The snapshot contains the user's core intent information such as behavioral preferences and interest directions within that time period. Input the user intent snapshot into the trained intent transition detection model. The model outputs the main intent type, intent intensity, and intent transition events compared to the previous snapshot for that snapshot.

[0022] The training method for the intent transition detection model is as follows: From the historical user behavior logs of the e-commerce platform, the historical behavior data is sliced ​​according to the above sliding window parameters to obtain a large number of time-series behavior slice pairs. Each slice pair is a slice of two adjacent time periods. The slice pairs are randomly divided into training set and validation set in a ratio of 8:2. For each slice, analyze the key behaviors that ultimately occur within it to determine the main intent type of that slice, which includes purchase intent, browsing intent, comparison intent, etc. The behavior density, behavior duration, and sequence consistency within a slice are quantified and assigned values. Behavior density refers to the number of behaviors per unit time; behavior duration refers to the total duration of behaviors within the slice; and sequence consistency refers to the continuity of behavior types. For example, continuous browsing of products within the same category indicates high consistency. The quantification and assignment specifically involves: Quantifying behavioral density Behavioral density The behavior density is calculated by dividing the total number of user actions within a slice by the slice duration, and then using a linear normalization method to obtain the quantified behavior density value. The calculation formula is: in This represents the minimum behavior density of similar slices in the platform's historical data. The maximum behavioral density of slices of the same type, if ,but Take 0, if ,but Take 1; Quantify the duration of the behavior Duration of the behavior That is, the cumulative time spent on all user actions within a slice from start to finish, quantified as the duration of the action. The calculation process is as follows: in This represents the total time the action takes within the slice. Quantization sequence consistency The longest continuous behavior sequence proportion method is used for calculation. First, user behavior sequences within the slice are extracted by time. Then, the behavior type and associated product category are identified to determine the longest continuous behavior subsequence. Finally, the consistency index is calculated. : The quantized value is obtained by dividing the length of the longest continuous subsequence by the total number of rows within the slice, and then optimizing the distribution using the Sigmoid function. The formula is: The coefficient 5 is used to enhance the discrimination, making When it approaches 0.5 Approaching 0.5, When close to 0 or 1 Approaching 0 or 1; The weights of the three indicators are set to calculate the final intent strength label. : in It is 0.3. It is 0.4. It is 0.3; Intent transition label: Compare the main intent labels of adjacent slices. If the intent graph type changes, it is determined that an intent transition has occurred, and the transition type is marked. The transition type includes interest point shift, decision intensity change, purchase intent generation or decay. If the intent graph type does not change, it is marked as no transition. Slices and their corresponding intention graph labels, intention intensity labels, and intention transition labels are selected from the training data and input into the model for calculation. The specific calculation formulas for each type of loss are as follows: The classification loss for the concept map labels is calculated using the cross-entropy loss function. Assume there are a total of primary intent tags. Class, model for the first The class probabilities of each sample output are: The one-hot vector corresponding to the real label is Then the single-sample classification loss The calculation formula is: The batch classification loss is the average of the classification losses of all samples.

[0023] The intention intensity label uses the mean squared error loss function to calculate the regression damage: Suppose the model is for the first The predicted value of intent intensity for each sample is The actual label value is Then the single-sample regression loss is The batch regression loss is the average of the losses of all samples.

[0024] The intention transition label uses a binary cross-entropy loss function to calculate the detection loss: The intended transition label is binary, 0 = no transition, 1 = transition exists. Let the model be for the first... The probability that a sample is predicted to be a transition is The real label is The single-sample detection loss is The batch detection loss is the average of the losses of all samples; The joint loss value is obtained by linearly combining the three types of batch losses. The formula for the joint loss function is as follows: in These are the weights for the three types of loss, all set to 0.33 by default. Then, the joint loss value is calculated using the gradient backpropagation algorithm. The partial derivatives of all trainable parameters in the model are used to form a gradient vector. The model parameters are updated using the Adam optimization algorithm. The above process is repeated, and the validation set performance index is recorded for 10 consecutive training cycles. When the performance index for 10 consecutive cycles does not exceed the threshold of 0.001, the model training is completed.

[0025] Ad generation module: Based on user intent snapshots, it generates a candidate set containing both real push ads and virtual ads. The specific process is as follows: Real ad generation: User intent snapshots are input into a global ad pool constructed via e-commerce platform servers, and a semantic matching algorithm is used to calculate the semantic matching score between the snapshot and each candidate ad. Specifically, the multidimensional information of the intent snapshot is transformed into a comprehensive feature vector. The processing method for each dimension is as follows: the intent graph type uses one-hot encoding to generate... A binary vector with intent intensity quantized as 0-1 is directly used as a one-dimensional feature. Interest categories are encoded using word embeddings to generate fixed-dimensional semantic vectors. Finally, the three types of features are concatenated to obtain the intent snapshot feature vector. And perform L2 normalization on the vector; For each candidate ad in the ad pool, its semantic description, category, price range, brand, and other attribute information are extracted, using the same encoding logic as the intent snapshot: category and brand are encoded using word embeddings, price range is mapped to numerical labels and used as one-dimensional features, and semantic description is extracted as sentence-level embedding vectors through big data analysis. All attribute-encoded vectors are concatenated to obtain the ad feature vector. And perform L2 normalization on the vector; The cosine similarity algorithm is used to calculate the cosine of the angle between the intent snapshot feature vector and the advertisement feature vector to measure their semantic similarity. The formula is as follows: in The dot product of two vectors. Let be the norms of the two vectors, respectively, and then, through a linear mapping... Convert to a matching score of 0-1; Ads with scores below 0.1 are removed, and all candidate ads are sorted by their matching scores. The ad with the highest matching score is selected as the actual push ad. Virtual ad generation: A preset offset is applied to the user's current intent snapshot. For example, if the current intent is "buy a mid-to-high-end smartphone", the offset intent becomes "buy a mid-to-high-end tablet". This causes the intent direction to shift towards the adjacent intent type, resulting in an offset intent snapshot. This offset intent snapshot is then input into the global ad pool. Similar to real push ads, virtual ads are retrieved through a semantic matching algorithm. Finally, one real push ad is combined with 2 to 3 virtual ads to form the final candidate set to be pushed to the user.

[0026] Interactive push and feedback capture module: The generated candidate set is pushed to users, and user interaction behavior with each ad in the candidate set is recorded, including clicks, browsing duration, favorites, adding to cart, and placing an order. Different scores are assigned to different interaction behaviors: 1 point for a click, 2 points for browsing duration exceeding 30 seconds, 3 points for favorites, 4 points for adding to cart, and 5 points for placing an order. The overall feedback score for each ad is accumulated based on user behavior. Feedback difference calculation: Calculate the overall feedback score of actual push ads. Calculate the average overall feedback score for all virtual ads. Calculate the feedback difference metric. The calculation formula is: in It is a very small positive number; Feedback difference measurement The intention deterministic quantization value is obtained by mapping using the Sigmoid function. Its formula is: Intent openness quantification value through formula Calculations show that the preset first threshold is 0.7, and the second threshold is 0.3. When When the user's current intent is highly certain, they tend to prefer genuine ads that match that intent; when... When the user's current intent is deemed highly open, they are more receptive to virtual ads in adjacent intent directions; when At that time, it is determined that the user's intent is in a state of ambiguity and transition.

[0027] Strategy module: Based on the nature of intent, the system is divided into three major scenarios: high determinism, high openness, and fuzzy transition. Using intent transition events as moderating factors, strategies are refined to address shifts in interest, changes in decision intensity, and the generation or fading of purchase intent, ultimately generating a targeted push notification solution. when When it is determined that the user's current intent is highly certain and the transition event is the generation of purchase intent, the pushed advertising content is as follows: the proportion of real ads increases to more than 80%, and is accompanied by 1-2 ads of complementary products in the same category. Only 1 virtual ad in the same price range is retained to verify the stability of purchase intent, without further expanding the boundaries of interest.

[0028] when When it is determined that the user's current intent is highly certain and the transition event is a change in decision intensity, that is, when the user's purchase intent intensity increases, the pushed advertising content is: real ads focusing on best-selling and highly rated products in the current interest category, supplementing conversion information such as product cost-effectiveness and inventory shortage, and increasing the ad display time.

[0029] when When it is determined that the user's current intent is highly certain and the transition event is a shift in interest, the push ad content is as follows: immediately switch real ads to the new interest point, retain 20% of the original interest point ads, observe feedback for two consecutive push cycles, and after confirming the stability of the interest shift, fully focus on the new interest point.

[0030] when When there is no jump event, the push ad slots will have 70%-80% real ads, all of which are core products in the current interest category, and 20% virtual ads, which are different models of the same category.

[0031] when When it is determined that the user's current intent is highly open and the transition event is a shift in interest, the pushed advertising content is: 40% real ads containing 30% of the original interest, and 30% virtual ads. The word embedding encoding adds cross-category association features to improve the accuracy of cross-category ad matching.

[0032] when When the user's current intent is deemed to be highly open and the transition event is the decline of purchase intent, the advertising content will be as follows: the proportion of real ads will decrease to 30%, the proportion of virtual ads will be 70%, new product and trendy product ads will be added, and user needs will be repositioned through multi-category push.

[0033] when When the user's current intent is deemed to be highly open and there is no transition event, the advertising content is as follows: the ratio of real ads to virtual ads is 1:1, the virtual ads cover 3-4 adjacent interest directions, and 2 representative products are selected for each direction. The feedback data of the virtual ads is tracked in detail, and the virtual ad direction with the highest feedback score is included in the candidate pool of real ads in the next round.

[0034] when When the user's intent is determined to be in a fuzzy transition state, and the transition event is an interest point shift, i.e. a fuzzy shift without a clear direction, the advertising content is as follows: 50% real ads represent the current fuzzy interest point, and 50% virtual ads cover 2-3 potential interest directions, with 1-2 ads for each direction. The intent certainty quantification threshold is temporarily relaxed, with the first threshold adjusted to 0.75 and the second threshold adjusted to 0.2 to avoid misjudging the nature of the intent.

[0035] when When the user's intent is determined to be in a state of ambiguity and the transition event is a fluctuation in decision intensity, the advertising content is as follows: real ads select products with multiple specifications and price points, and virtual ads select products of the same category but different price ranges to help users clarify their preferences and increase guided interactive ads to clarify intent through active interaction.

[0036] when When the user intent is determined to be in a vague transition state and there is no transition event, the advertising content is: real ads and virtual ads in a ratio of 6:4. Virtual ads focus on the current category sub-direction and do not expand across major categories. If the intent remains vague for two consecutive cycles, new product and promotional activity ads in the category will be pushed.

[0037] The above are merely embodiments of the present invention. The circuits, electronic components, and modules involved are all prior art, fully achievable by those skilled in the art, and require no further explanation. The scope of protection in this application does not involve improvements to the software and methods. Commonly known structures and characteristics in the solutions are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all prior art in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent.

Claims

1. An advertising push method based on e-commerce big data feedback, characterized in that, Includes the following steps: S1: Collect real-time behavioral data streams of users on e-commerce platforms and construct a multi-dimensional spatiotemporal behavioral matrix based on these data streams. S2: The multidimensional spatiotemporal behavior matrix is ​​dynamically sliced ​​using a sliding time window. Each slice corresponds to a user intent snapshot. Intent transition events between adjacent slices are identified through an intent transition detection model. The intent transition events include interest point shift, decision intensity change, and purchase intent generation or decay. S3: Based on the user intent snapshot, generate a candidate set containing real push ads and virtual ads; S4: Push the candidate set to the user and capture their interaction feedback. By comparing the differences in user feedback to real ads and virtual ads, quantify the certainty and openness of user intent. S5: Adjust the push strategy based on the nature of the user's intent and the intent transition event.

2. The advertising push method based on e-commerce big data feedback according to claim 1, characterized in that, In step S1, the method for constructing the multidimensional spatiotemporal behavior matrix is ​​as follows: Receive real-time behavioral data streams from e-commerce platforms, analyze each behavioral event, and generate standard event tuples. The standard event tuples include: user identifier, behavior type, absolute timestamp of the behavior occurrence, page uniform resource identifier, product category code, and context information. Using seconds as the basic unit, the absolute timestamp is mapped to a consecutive second-level time slice sequence number starting from the base time, and multiple behavioral events occurring within the same second are classified into the same second-level time slice; A unified virtual shopping space coordinate system is constructed, which maps different page Uniform Resource Identifiers to first-dimensional coordinate values, maps different product category codes to second-dimensional coordinate values, and represents the user's "physical location" in a single browsing behavior as a two-dimensional coordinate point in the virtual shopping space. Each standard event tuple is transformed into a multidimensional feature vector, with the user identifier as the first dimension index, the second-level time slice sequence number as the second dimension index, and the multidimensional feature vector as the third dimension element to generate a three-dimensional spatiotemporal behavior matrix.

3. The advertising push method based on e-commerce big data feedback as described in claim 2, characterized in that: The feature vector includes: a spatial coordinate vector composed of page coordinates and category coordinates, a behavior type vector that performs one-hot encoding of behavior types, and a context feature vector that performs numerical encoding of context information.

4. The advertising push method based on e-commerce big data feedback according to claim 3, characterized in that, In step S2, generating a user intent snapshot specifically involves: Within the sliding window, a fixed-length window size and a sliding step size are preset. Starting from the starting time point of the multidimensional spatiotemporal behavior matrix, continuous duration behavior data is sequentially extracted at intervals of the sliding step size to form a slice. The behavior event data within each slice forms a sub-matrix. The sub-matrix is ​​used as the input of the intent transition detection model at the corresponding time point, and the intent transition detection model generates a snapshot of the user intent at that time point.

5. The advertising push method based on e-commerce big data feedback according to claim 4, characterized in that, The training method for the intent transition detection model is as follows: From historical user behavior logs, the historical user behavior is sliced ​​using a sliding window to obtain time-series behavior slice pairs, which are then divided into training and validation sets. Based on the behavior that ultimately occurs within each time window slice, the main intent type of the slice is determined, and intent graph labels are generated; intent intensity labels are obtained by quantifying and assigning values ​​based on the density, duration, and sequence consistency of the behavior within the slice; and intent transition labels are obtained by comparing the main intent categories of adjacent slices to determine whether a behavior transition has occurred and the type of behavior transition. The parameters of the intent transition detection model are initialized. Slices and corresponding intention graph labels, intent intensity labels, and intent transition labels are selected from the training set. The intention graph classification loss, intent intensity regression loss, and intent transition detection loss are calculated for the intention graph labels, intent intensity labels, and intent transition labels, respectively. The intention graph classification loss, intent intensity regression loss, and intent transition detection loss are linearly combined to obtain the joint loss value. Through gradient backpropagation, the partial derivatives of the joint loss value with respect to the trainable parameters in the intent transition detection model are calculated in sequence to form a gradient vector. The parameters of the intent transition detection model are updated using an estimation algorithm. The model is considered to have converged when the overall performance index of the intent transition detection model on the validation set does not exceed a preset threshold within 10 consecutive cycles.

6. The advertising push method based on e-commerce big data feedback according to claim 5, characterized in that, In step S3, the real advertisement generation process is as follows: The user intent snapshot is input into the global ad pool. Multiple candidate ads with the highest semantic matching degree with the current intent snapshot are retrieved. The matching degree score between the multiple candidate ads and the user intent snapshot is calculated. The ad with the highest score is selected as the actual push ad.

7. The advertising push method based on e-commerce big data feedback according to claim 6, characterized in that, The virtual advertisement generation process is as follows: An offset is applied to the user intent snapshot, shifting it towards a neighboring intent direction to obtain an offset intent snapshot. This offset intent snapshot is used as input to the global ad pool. A pre-selected virtual ad set is retrieved from the global ad pool and output. Ads that are comparable to the real push ads in terms of category, price range, or brand but different in specific products are selected from the pre-selected virtual ad set and used as virtual ads. The real push ads and virtual ads are combined to form the final candidate set to be pushed to the user.

8. The advertising push method based on e-commerce big data feedback according to claim 7, characterized in that, In step S4, the quantification of the determinism and openness of user intent specifically refers to: The system captures user interactions with each ad in the candidate set. Based on the different scores for each interaction, the scores for each ad's interactions are summed to obtain the ad's overall feedback score. The average overall feedback score of the real-world ads and the average overall feedback score of all virtual ads are then calculated. Finally, a feedback difference metric is calculated based on these two average overall feedback scores. The calculation formula is: in The average overall feedback score for genuinely pushed advertisements, the The average overall feedback score for all virtual ads. The feedback difference measure is a very small positive number. The intention deterministic quantization value is obtained by mapping using the Sigmoid function. Calculate the intention openness quantification value based on the intention certainty quantification value. The calculation formula is: Based on the intention to determine the quantification value The user's current intent is determined by comparing it with the first threshold and the second threshold.

9. The advertising push method based on e-commerce big data feedback according to claim 8, characterized in that, The user's current intent includes: When the intention is deterministically quantified When the value exceeds the first preset threshold, it is determined that the user's current intent is highly certain, meaning they prefer genuine advertisements. When the intention is deterministically quantified When the value is below the second preset threshold, it is determined that the user's current intention is highly open, that is, they are more receptive to virtual advertisements; When the intention is deterministically quantified When the value is between the first threshold and the second threshold, the user's intent is determined to be in a state of fuzzy transition.

10. An advertising push system based on e-commerce big data feedback, characterized in that, include: The intent flow modeling module is used to collect real-time behavioral data streams of users on e-commerce platforms and construct a multi-dimensional spatiotemporal behavioral matrix based on these data streams. The dynamic slicing analysis module is used to dynamically slice the multidimensional spatiotemporal behavior matrix using a sliding time window. Each slice corresponds to a user intent snapshot. The intent transition detection model identifies intent transition events between adjacent slices. These intent transition events include interest point shifts, changes in decision intensity, and the generation or fading of purchase intent. The ad generation module is used to generate a candidate set containing real push ads and virtual ads based on user intent snapshots; An interactive push and feedback capture module is used to push the candidate set to the user and capture their interactive feedback. By comparing the differences in user feedback to real ads and virtual ads, the certainty and openness of user intent are quantified. The strategy module is used to adjust the push strategy based on the nature of the user's intent and the intent transition event.

Images