Automatic decision-making method and system based on user preferences in different scenes

Abstract

The invention discloses an automatic decision-making method and system based on user preferences in different scenes, and belongs to the technical field of data processing. According to the method, multi-dimensional scene features of a user are collected, and scene types and confidence coefficients are output through a pre-training classification model; multi-source user data is directionally collected according to a scene and is converted into standardized features through cleaning and feature engineering; constructing a basic preference model library, calling a corresponding model, calculating a scene weight adjustment coefficient, and updating model parameters to generate a dynamic preference model; and outputting a commodity preference sorting result by the standardized feature input model, converting the commodity preference sorting result into decision-making instructions such as personalized recommendation and dynamic pricing, and executing the decision-making instructions. According to the method, through a scenarized dynamic model adaptation mechanism, the problems of static preference decision, poor scene adaptability and insufficient real-time performance in the prior art are solved, the user preference prediction accuracy and the personalized service conversion efficiency are improved, and the user stickiness of the e-commerce platform is enhanced.

Description

Technical Field

[0001] This invention relates to the field of e-commerce platform data processing, and in particular to a method and system for automatic decision-making based on user preferences in different scenarios. Background Technology

[0002] With the rapid development of e-commerce platforms, personalized services have become a core means to enhance user experience and commercial value, and user preference decision-making is key to achieving personalized services. Existing e-commerce platform user preference decision-making technologies mainly suffer from the following shortcomings:

[0003] Poor scenario adaptability: Most technologies rely on static historical data to build preference models without considering the differences in user scenarios (such as significant differences in user preferences between promotional scenarios and daily scenarios), resulting in a disconnect between decision-making results and actual needs. Insufficient real-time performance: The model parameters are updated over a long period of time (usually days or hours), making it impossible to respond to fluctuations in user preferences caused by real-time user interactions (such as rapid clicks on the current page or changes in dwell time). Fixed weights: Feature weights remain fixed after model training and cannot be dynamically adjusted according to scenario features (such as product popularity and user behavior frequency), which reduces the accuracy of preference prediction. Single data dimension: It relies heavily on users' historical purchase data and ignores the impact of environmental characteristics (such as device type and network environment) and scenario-related data (such as the average behavior of the group) on preferences, resulting in weak model generalization ability.

[0004] To address the aforementioned issues, there is an urgent need for an automated decision-making solution that can combine dynamic features of the scenario, update preference models in real time, and be driven by multi-dimensional data, in order to improve the accuracy and timeliness of personalized services on e-commerce platforms. Summary of the Invention

[0005] The purpose of this invention is to propose a method for automatic decision-making based on user preferences in different scenarios, aiming to solve the technical problems of poor scenario adaptability, insufficient real-time performance, fixed weights, and single data dimensions in existing e-commerce platform user preference decision-making technologies.

[0006] This invention is implemented as follows: a method for automatic decision-making based on user preferences in different scenarios, the method comprising: Collect multi-dimensional scene features of the current user, input them into a pre-trained scene classification model, and output scene type and scene confidence. Based on the output scenario type, collect user data from the corresponding data source, and perform cleaning and feature engineering on the collected data; Construct a basic preference model library; call the corresponding basic preference model from the basic preference model library according to the scenario type, and calculate the scenario weight adjustment coefficient; update the parameters of the basic preference model based on the scenario weight adjustment coefficient, and generate a dynamic preference model for the current scenario; The feature data after feature engineering is input into the dynamic preference model, and the product preference ranking result is output. According to the e-commerce platform application scenario, the preference ranking result is converted into decision instructions, and the corresponding personalized operation is executed according to the decision instructions.

[0007] Another objective of this invention is to provide a system for automatic decision-making based on user preferences in different scenarios. The system includes:

[0008] The scene recognition module is used to collect multi-dimensional scene features of the current user, input them into a pre-trained scene classification model, and output the scene type and scene confidence score. The data acquisition and feature engineering module is used to collect user data from the corresponding data source according to the output scenario type, and to clean and feature engineer the collected data. The dynamic preference model construction module is used to call the corresponding basic preference model from the basic preference model library according to the scenario type, calculate the scenario weight adjustment coefficient, update the parameters of the basic preference model based on the scenario weight adjustment coefficient, and generate the dynamic preference model under the current scenario. The preference decision and personalized execution module is used to input the feature data after feature engineering into the dynamic preference model and output the product preference ranking result; according to the e-commerce platform application scenario, the preference ranking result is converted into decision instructions, and the corresponding personalized operation is executed according to the decision instructions.

[0009] Beneficial effects of the present invention This invention discloses a method and system for automatic user preference decision-making based on different scenarios. The method collects multi-dimensional user scenario features, outputs scenario type and confidence level through a pre-trained classification model, collects multi-source user data according to scenario orientation, cleans and performs feature engineering to transform it into standardized features, constructs a basic preference model library, calls corresponding models and calculates scenario weight adjustment coefficients, updates model parameters to generate a dynamic preference model, inputs standardized features into the model to output product preference ranking results, and converts them into decision instructions such as personalized recommendations and dynamic pricing, which are then executed. This invention solves the problems of static preference decision-making, poor scenario adaptability, and insufficient real-time performance in existing technologies through a scenario-based dynamic model adaptation mechanism, improving the accuracy of user preference prediction and the efficiency of personalized service conversion, and enhancing user stickiness on e-commerce platforms. Attached Figure Description

[0010] Figure 1This is a flowchart of a preferred embodiment of the present invention for an automatic decision-making method based on user preferences in different scenarios; Figure 2 This is a system structure diagram of a preferred embodiment of the present invention, which is based on automatic decision-making of user preferences in different scenarios. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. For ease of explanation, only the parts related to the embodiments of this invention are shown. It should be understood that the specific embodiments described herein are merely for explaining this invention and are not intended to limit this invention.

[0012] This invention proposes a method and system for automatic user preference decision-making based on different scenarios, belonging to the field of data processing technology. The method collects multi-dimensional user scenario features, outputs scenario type and confidence level through a pre-trained classification model; collects multi-source user data according to scenario orientation, cleans and transforms it into standardized features; constructs a basic preference model library, calls corresponding models and calculates scenario weight adjustment coefficients, updates model parameters to generate a dynamic preference model; inputs the standardized features into the model to output product preference ranking results, transforms them into personalized recommendation, dynamic pricing, and other decision instructions, and executes them. This invention's method, through a scenario-based dynamic model adaptation mechanism, solves the problems of static preference decision-making, poor scenario adaptability, and insufficient real-time performance in existing technologies, improving the accuracy of user preference prediction and the efficiency of personalized service conversion, and enhancing user stickiness on e-commerce platforms.

[0013] Figure 1 This is a flowchart of a preferred embodiment of the present invention for an automatic decision-making method based on user preferences in different scenarios; the method includes the following steps: S1: Collect multi-dimensional scene features of the current user, input them into a pre-trained scene classification model, and output the scene type and scene confidence.

[0014] In this embodiment of the invention, the multi-dimensional scene features include time features, user behavior features, product category features, and environmental features; the time features include date type and time period; the user behavior features include the current access path and interaction frequency; the product category features include the category level of the currently viewed product; and the environmental features include the access device type and network environment. The scenario types include promotional scenarios, daily shopping scenarios, emergency demand scenarios, and category exploration scenarios; In this embodiment of the invention, the front end of an e-commerce platform (iOS / Android / PC) can be connected via an HTTP interface to collect multi-dimensional scene characteristics of users; data transmission adopts JSON format; In this embodiment of the invention, the training process of the scene classification model includes: collecting historical scene data from e-commerce platforms and corresponding manually labeled scene types to construct a training dataset; using the XGBoost algorithm as the basic classifier, taking multi-dimensional scene features as input and labeled scene types as output, and optimizing model parameters through 5-fold cross-validation to ensure that the model test set accuracy is ≥ a preset percentage (e.g., 92%).

[0015] For example, in one embodiment of the present invention, the multi-dimensional scene features are defined as follows: Time characteristics: Date type (weekday / weekend, statutory holiday / ordinary day, promotional day (such as 618, Double 11) / non-promotional day), time period (8-10 am, 12-2 pm, 8-10 pm, etc.); User behavior characteristics: Current access path (homepage → category page → product details page / homepage → search page → product details page), interaction frequency (number of clicks and page switching times within 5 minutes); Product category characteristics: The primary category to which the currently viewed product belongs (Apparel, Shoes & Bags / Food & Fresh Produce / Home Appliances & Digital Products), and the category level (Level 1 / Level 2 / Level 3); Environmental characteristics: Access device type (mobile phone / tablet / PC), network environment (4G / 5G / Wi-Fi); For example, in one embodiment of the present invention, the scene classification model is trained as follows: Dataset Construction: Six months of historical data were collected from e-commerce platforms, containing 1 million user scenario records. Each record was labeled with the scenario type (promotional scenario / daily shopping scenario / emergency demand scenario / category exploration scenario), with the sample proportions for each scenario being 30% / 40% / 15% / 15%, respectively. Model training: The XGBoost algorithm was used, with multi-dimensional scene features as input (after One-Hot encoding and normalization), and scene type probability distribution as output. The learning rate (0.1), tree depth (6), and number of leaf node samples (20) were optimized through grid search. The model achieved an accuracy of 94.2% and an F1 score of 93.8% on the test set. Scene confidence calculation: The probability value of the current scene type output by the model is taken as the scene confidence. When the confidence is ≥0.8, the scene type is determined directly. When 0.5≤confidence<0.8, it is corrected by combining the user's historical scene preferences. When the confidence is <0.5, the default scene (daily shopping scene) is triggered.

[0016] S2: Based on the output scenario type, collect user data from the corresponding data source, and clean and perform feature engineering on the collected data (i.e., convert it into standardized numerical or vector features that can be input into the preference model). The user data includes historical user behavior data, real-time interaction data, scenario-related data, and product attribute data; The historical behavior data includes historical purchase records, favorites records, and review data; the real-time interaction data includes current page dwell time, click sequence, and add-to-cart actions; the scenario-related data includes the average behavior of user groups in the scenario and the product popularity trend; the product attribute data includes product price range, specifications, and inventory status. The feature engineering includes constructing behavioral sequence features, user profile features, and product association features; The cleaning process includes removing outliers and filling in missing values. For example, in one embodiment of the present invention, user data from the corresponding data source is collected according to the output scene type as follows: For historical behavioral data, purchase records, favorite records, ratings, and return records for the past 12 months are collected from the user database; offline collection (updated daily); For real-time interactive data, collect data such as current page dwell time, click location, and add-to-cart / cancel-add-to-cart activity from the front-end interaction log library; collect data in real time (1 time / second). For scenario-related data, the average user dwell time and the top 10 clicked products in the same scenario are obtained through a real-time calculation engine; real-time calculation (5-minute window). For product attribute data, price, discount rate, specifications, inventory, and positive review rate are collected from the product database; real-time synchronization (updated every 10 minutes); In this embodiment of the invention, data cleaning uses the 3σ principle to remove outliers (such as dwell time > 30 minutes) and the KNN algorithm to fill in missing values ​​(such as missing network environment data). For example, in one embodiment of the present invention, the feature engineering method is as follows: Behavioral sequence features: The user's click sequence over the past hour is converted into a fixed-length (20) vector (using Word2Vec encoding). User profile features: Construct user tags (price sensitivity: high / medium / low, category preference: top 3 categories); Product association features: Calculate the category similarity and price similarity between the current product and the user's historically purchased products.

[0017] S3: Construct a basic preference model library; call the corresponding basic preference model from the basic preference model library according to the scenario type, and calculate the scenario weight adjustment coefficient; update the parameters of the basic preference model based on the scenario weight adjustment coefficient, and generate a dynamic preference model for the current scenario; including the following steps: S31: Construct a basic preference model library; The basic preference model library contains basic preference models corresponding to each scenario type; the basic preference model adopts an improved Transformer architecture, in which the encoder layer introduces a scenario attention mechanism and the decoder layer outputs the user's preference rating for the product. S32: Based on the scenario type, call the corresponding basic preference model from the basic preference model library and calculate the scenario weight adjustment coefficient; The scenario weight adjustment coefficient is determined based on scenario confidence, the frequency of users' historical behavior in the scenario, and the variance of the current scenario's product popularity. The formula for calculating the scene weight adjustment coefficient is as follows: α=(0.4×C + 0.3×F + 0.3×V) / 3; Where α is the scene weight adjustment coefficient (range 0.6-1.2), C is the scene confidence level (range 0-1), F is the normalized value of the user's historical behavior frequency in the current scene type (range 0-1), and V is the normalized value of the product popularity variance in the current scene (range 0-1).

[0018] S33: Based on the scene weight adjustment coefficient, dynamically update the feature weight matrix and attention weight in the basic preference model to generate a dynamic preference model for the current scene; The specific steps for updating the feature weight matrix and attention weights in the basic preference model are as follows: Multiply the feature weight matrix W of the basic preference model by the scene weight adjustment coefficient α to obtain the dynamic feature weight matrix W'=W×α; adjust the scene attention weight according to the proportion of α, amplify the attention weight according to the proportion of α when α>1, and reduce the attention weight according to the proportion of α when α<1. The triggering conditions for dynamic updates include timed triggering and event triggering. Timed triggering involves detecting changes in scenario confidence every preset time period (e.g., 30 minutes), and triggering an update when the change is greater than or equal to a preset percentage (e.g., 15%). Event triggering involves triggering an update when the accumulated user real-time interaction data reaches a preset number (e.g., 10). An incremental update mechanism is used to update only feature weights and attention parameters, with a single update cycle taking less than or equal to a preset time (e.g., 800ms), without interrupting the current decision-making service. After the update is completed, a dynamic preference model specific to the current scenario is generated.

[0019] In this embodiment of the invention, the feature weights in the feature weight matrix (W) are the weight values ​​corresponding to each element in the matrix, representing the basic influence of each input feature on preference prediction; the scene attention weights are the attention allocation weights in the attention head weight matrix of the encoder layer of the Transformer architecture, representing the attention priority between different user behavior sequences. The following section explains the updating of the feature weight matrix and attention weights in detail, using rules and examples. For updating feature weights: the feature weight matrix W of the basic preference model is multiplied by the scene weight adjustment coefficient α. Here, "weight" refers to all elements in the feature weight matrix W (i.e., the weight value corresponding to each feature). For example, the weight value of the "discount rate" feature in the basic preference model is 0.2. When α=0.81, the updated feature weight = 0.2×0.81=0.162. The weight values ​​of all features are updated synchronously according to α through matrix multiplication, forming a dynamic weight matrix W'.

[0020] Regarding the update of attention weights: Scene attention weights are increased (when α>1) or decreased (when α<1) by a ratio of α. Here, "weight" refers to the weight assigned to each attention in the attention head weight matrix. For example, in the basic preference model, the attention weight of the "click coupon" action to the "browse product details" action is 0.3. When α=0.81 (<1), the updated attention weight = 0.3×0.81=0.243. All attention weights are adjusted synchronously by a ratio of α.

[0021] For example, in one embodiment of the present invention, the basic preference model is designed as follows: Architecture: Improve the Transformer architecture by introducing a scene attention mechanism in the encoder layer (6 layers) (embedding the scene type as an attention mask) and outputting product preference scores (0-10 points) in the decoder layer (3 layers). Input features: user profile features (128 dimensions), behavior sequence features (20×64 dimensions), product attribute features (64 dimensions), and scene features (32 dimensions); Training process: Three months of historical data from an e-commerce platform (5 million user-product interaction records) were used. Actual user purchase / add-to-cart behavior was used as positive samples, and non-interactive products were randomly sampled as negative samples (positive-to-negative sample ratio 1:5). The Adam optimizer was used with a learning rate of 0.001 and 50 training epochs. The model test set MAE (mean absolute error) was ≤0.85.

[0022] For example, in one embodiment of the present invention, the method for calculating the scene weight adjustment coefficient is as follows: Scene weight adjustment coefficient α = (0.4 × C + 0.3 × F + 0.3 × V) / 3; For example, if the scenario confidence level C = 0.9 (promotional scenario), the normalized frequency value of user's historical promotional scenario behavior F = 0.8 (participation in promotional activities 8 times in the last 30 days), and the variance of product popularity in the current promotional scenario V = 0.7 (normalized variance of product clicks), then... α=(0.4×0.9+0.3×0.8+0.3×0.7) / 3=(0.36+0.24+0.21) / 3=0.81; Weight update: Multiply the feature weight matrix W of the base model by 0.81 to obtain the dynamic feature weight matrix W' = W × 0.81; adjust the scene attention weights by a ratio of 0.81. The dynamic update is triggered by timed and event-driven conditions. Timed updates are triggered every 30 minutes when the scene confidence changes and the change is ≥15%. Event-driven updates are triggered when the user's real-time interaction data reaches 10 items. An incremental update mechanism is used to update only the feature weights and attention parameters. A single update takes ≤800ms and does not interrupt the current decision service. After the update is completed, a dynamic preference model specific to the current scene is generated.

[0023] S4: Input the feature data after feature engineering into the dynamic preference model and output the product preference ranking result; according to the e-commerce platform application scenario, convert the preference ranking result into a decision instruction, and execute the corresponding personalized operation according to the decision instruction.

[0024] The application scenarios include personalized recommendations, dynamic pricing, and service adaptation. The decision instructions include: personalized recommendation instructions (such as outputting a product list sorted by preference, with a display priority of ≥80%), dynamic pricing instructions (such as setting a 5%-15% discount for the top 20% of products in terms of preference rating), and service adaptation instructions (triggering inventory warning prompts and prioritizing logistics timeliness for preferred products). For example, in one embodiment of the present invention, the preference ranking and decision instructions are as follows: Preference ranking: The dynamic preference model outputs the user's preference rating for 100-200 candidate products in the current scenario, sorts them in descending order of rating, and takes the Top 50 as the preferred product list; The decision instructions are as follows: In personalized recommendation scenarios, the personalized recommendation instructions are as follows: display the top 10 preferred products in the homepage recommendation slot, and associate the top 5 recommended products on the details page; update the recommendation list by calling the recommendation system interface (response time ≤ 500ms). In a dynamic pricing scenario, the dynamic pricing instruction is as follows: set an 8%-12% discount for the top 20% of preferred items and a 3%-5% discount for the top 21-50%; generate a limited-time discount tag (valid for 2 hours) by calling the pricing system interface. In the service adaptation scenario, the service adaptation instructions are as follows: when the top 10 preferred products trigger a "low inventory" prompt, priority is given to matching next-day delivery; by calling the inventory system and logistics system interfaces, the product tags and logistics priority are updated; Corresponding to the method for automatic decision-making based on user preferences in different scenarios described in the above embodiments, Figure 2 This diagram illustrates the structural block diagram of a system for automatic decision-making based on user preferences in different scenarios, as provided in an embodiment of this application. For ease of explanation, only the parts relevant to the embodiments of this application are shown. The system includes:

[0025] The scene recognition module is used to collect multi-dimensional scene features of the current user, input them into a pre-trained scene classification model, and output the scene type and scene confidence score. The data acquisition and feature engineering module is used to collect user data from the corresponding data source according to the output scenario type, and to clean and feature engineer the collected data. The dynamic preference model construction module is used to call the corresponding basic preference model from the basic preference model library according to the scenario type, calculate the scenario weight adjustment coefficient, update the parameters of the basic preference model based on the scenario weight adjustment coefficient, and generate the dynamic preference model under the current scenario. The preference decision and personalized execution module is used to input the feature data after feature engineering into the dynamic preference model and output the product preference ranking result; according to the e-commerce platform application scenario, the preference ranking result is converted into decision instructions, and the corresponding personalized operation is executed according to the decision instructions.

[0026] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by program instructions and related hardware. The program can be stored in a computer-readable storage medium, such as ROM, RAM, disk, optical disk, etc.

[0027] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for automatic decision making based on user preferences in different scenarios, characterized in that, The method comprises: Collecting multi-dimensional scene features of the current user, inputting them into a pre-trained scene classification model, and outputting scene types and scene confidence; According to the output scene type, collecting user data from the corresponding data source, and cleaning and feature engineering the collected data; Building a basic preference model library; calling the corresponding basic preference model from the basic preference model library according to the scene type, calculating the scene weight adjustment coefficient, updating the basic preference model parameters based on the scene weight adjustment coefficient, and generating a dynamic preference model under the current scene; Input the feature data processed by the feature engineering into the dynamic preference model, output the commodity preference ranking result; according to the e-commerce platform application scene, the preference ranking result is converted into a decision instruction, and the corresponding personalized operation is executed according to the decision instruction.

2. The method of claim 1, wherein the user preference is automatically determined based on a scenario. The training process of the scene classification model comprises: Collecting e-commerce platform historical scene data and corresponding artificial labeled scene types to build a training data set; Using XGBoost algorithm as the basic classifier, taking multi-dimensional scene features as input and labeled scene types as output, optimizing model parameters through 5-fold cross-validation to make the model test set accuracy ≥ preset percentage.

3. The method for automatic decision-making based on user preferences in different scenes according to claim 1, wherein The multi-dimensional scene features include time features, user behavior features, commodity category features and environment features; The time features include date types and time periods; the user behavior features include current access paths and interaction frequencies; the commodity category features include the category level of the currently browsed commodities; The environment features include access device types and network environments.

4. The method for automatic decision-making based on user preferences in different scenes according to claim 1, wherein The user data includes user historical behavior data, real-time interaction data, scene association data and commodity attribute data; The feature engineering includes building behavior sequence features, user portrait features and commodity association features; The cleaning includes removing outliers and filling missing values.

5. The method for automatic decision-making based on user preferences in different scenes according to claim 1, wherein The basic preference model library contains basic preference models corresponding to each scene type; the basic preference model uses an improved Transformer architecture, the encoder layer of which introduces a scene attention mechanism, and the decoder layer outputs the user's preference score for the commodity.

6. The method for automatic decision based on user preference in different scenarios as claimed in claim 1, wherein, The calculation formula of the scene weight adjustment coefficient is: α=(0.4×C + 0.3×F + 0.3×V) / 3; Wherein, α is the scene weight adjustment coefficient, C is the scene confidence, F is the historical behavior frequency normalization value of the user in the current scene type, and V is the commodity heat variance normalization value in the current scene.

7. The method for automatic decision based on user preference in different scenarios as claimed in claim 1, wherein, The updating of the feature weight matrix and the attention weight in the basic preference model is as follows: Multiply the feature weight matrix W of the basic preference model by the scene weight adjustment coefficient α to obtain the dynamic feature weight matrix W'=W×α; The scene attention weight is adjusted by a proportion of α. When α>1, the attention weight is enlarged by a proportion of α. When α<1, the attention weight is reduced by a proportion of α. 8.The method of claim 1, wherein the method further comprises: The feature weight in the feature weight matrix is a weight value corresponding to each element in the matrix, representing the basic influence of each input feature on the preference prediction. The scene attention weight is an attention distribution weight in the attention head weight matrix of the Transformer architecture encoder layer, representing the attention priority among different user behavior sequences. 9.The method of claim 1, wherein the method further comprises: The application scenarios include personalized recommendation, dynamic pricing, and service adaptation. The decision instructions include personalized recommendation instructions, dynamic pricing instructions, and service adaptation instructions.

10. A system for automatic decision making based on user preferences in different scenarios, characterized in that, The system comprises: A scene recognition module, configured to collect multi-dimensional scene features of a current user, input the multi-dimensional scene features into a pre-trained scene classification model, and output a scene type and a scene confidence; A data collection and feature engineering module, configured to collect user data from corresponding data sources according to the output scene type, and clean and perform feature engineering on the collected data; A dynamic preference model construction module, configured to call a corresponding basic preference model from the basic preference model library according to the scene type, calculate a scene weight adjustment coefficient, update parameters of the basic preference model based on the scene weight adjustment coefficient, and generate a dynamic preference model in the current scene; A preference decision and personalized execution module, configured to input the feature data processed by the feature engineering into the dynamic preference model, output a commodity preference ranking result, convert the preference ranking result into a decision instruction according to an application scenario of an e-commerce platform, and execute a corresponding personalized operation according to the decision instruction.

Images