B2B steel resource recommendation method, electronic equipment and storage medium

Abstract

The invention relates to a B2B steel resource recommendation method based on short-term behavior intensity of buying parents, electronic equipment and a storage medium, and the method comprises the following steps: 1) collecting long-term and short-term behavior data of a user, and extracting long-term and short-term behavior characteristics; 2) constructing a buyer behavior intensity sequence model by using a deep learning and time sequence analysis method, and screening out resources highly matched with user interests; 3) filtering results according to industry characteristics and business rules, removing resources which do not meet business requirements, and dynamically adjusting filtering rules according to the business requirements which change in real time; and 4) optimizing the arrangement sequence of the resources based on a high-level sorting algorithm, and realizing accurate recommendation of the B2B steel resources. Compared with the prior art, the B2B steel resource recommendation method has the advantages that the recommendation accuracy is high, the service adaptability is high, and a recommendation system has real-time performance and dynamic performance.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of consumption recommendation, in particular to a B2B steel resource recommendation method based on long-term and short-term behavior intensity of buyers, an electronic device and a storage medium. BACKGROUND

[0002] In the B2B steel industry, the transaction relationship between buyers and sellers usually has the characteristics of complexity and long-term, and the decision-making process often involves multiple stakeholders, in addition, the transaction volume involving the steel industry is usually large, the resource types are various, and the industry has strong speciality. Compared with B2C recommendation system, B2B recommendation system not only needs to consider the short-term consumption demand of buyers, but also needs to consider the long-term purchasing behavior and the development direction of enterprises. In steel transactions, the behavior intensity of buyers in different time periods reflects their different purchasing demand types, and the strength of these purchasing demands puts higher requirements on accurate recommendation methods and recommendation strategies.

[0003] Existing B2B recommendation technologies mostly rely on traditional collaborative filtering algorithms, content recommendation and deep learning-based recommendation models. Specifically, collaborative filtering algorithms usually analyze user behavior data to predict products that users may be interested in, content recommendation relies more on product attribute information for matching, and deep learning-based recommendation models automatically learn complex user behavior patterns for prediction.

[0004] These methods in the B2B scenario, because the buyer demand cycle is long and unstable, the traditional recommendation model often faces the long tail problem, resulting in less expected recommendation effect. Face the following limitations: (1) Lack of long-term behavior modeling: Many existing technologies focus too much on users' short-term behavior (such as browsing, clicking, etc.), and ignore long-term purchasing behavior. This leads to the possibility that the recommended results may not fully reflect the potential strategic needs of buyers.

[0005] (2) Unable to adapt to complex decision-making process: B2B purchasing decisions are usually made by multiple people (in addition to company operators in steel transactions, there are also company employees participating), and existing models perform poorly in capturing the behavior characteristics of multiple participants, and cannot effectively handle the dynamic interaction between different stakeholders.

[0006] (3) Lack of real-time performance: Although many recommendation systems can make real-time recommendations, they are often based on static models and lack sensitivity to fluctuations in buyer demand. SUMMARY

[0007] The present application aims to overcome the defects of the prior art and provide a B2B steel resource recommendation method based on the long-term and short-term behavior intensity of buyers, an electronic device and a storage medium, to solve or partially solve the problems of lack of long-term behavior modeling in the recommendation model, single recommendation content, and lack of real-time demand of buyers.

[0008] The object of the present application can be achieved by the following technical solutions: The present application provides a B2B steel resource recommendation method based on the long-term and short-term behavior intensity of buyers, the core of which includes three modules: recall layer, business filtering layer and sorting layer, and the specific working principle is as follows: Recall layer: based on the modeling results of the long-term and short-term behavior of buyers, resources with high matching degree of interest are recalled. Through the modeling of buyer behavior based on deep learning model, resources with consistent potential demand can be accurately identified. The recall layer mainly plays the role of "screening" and filters out resources irrelevant to user interest.

[0009] Business filtering layer: in the candidate resources recalled, secondary screening is carried out in combination with industry, product category and business demand, and resources that do not match the current demand of buyers or do not match the actual business scenario are removed. This layer further ensures the practicality and accuracy of the recommendation.

[0010] Sorting layer: the final sorting layer sorts the candidate resources by considering multiple factors such as buyer interest intensity, resource relevance, purchase history, etc. Advanced sorting algorithm (such as XGBoost) is used for optimization to ensure the maximum conversion rate of the recommendation results.

[0011] Specifically, the following steps are included: S1. Collect long-term and short-term behavior data of users, including purchase frequency, procurement amount, resource category preference, supplier procurement preference, procurement cycle, browsing behavior, clicking behavior and potential purchase behavior, and clean, normalize and handle missing values of the data, and build and extract long-term and short-term behavior characteristics of users; S2. In the recall stage, use deep learning and time series analysis method to build buyer behavior intensity sequence model to capture user long-term and short-term interest preference and potential demand, and select resources highly matched with user interest; S3. Filter the results according to industry characteristics and business rules to remove resources that do not meet business demand, and dynamically adjust the filtering rules according to real-time changing business demand; S4. In the sorting stage, based on advanced sorting algorithm, combined with the long-term and short-term behavior intensity of users and the diversity of recommended resources, the arrangement order of resources is optimized to realize accurate recommendation of B2B steel resources.

[0012] As a preferred technical solution, step S2, specifically includes the buyer behavior intensity sequence model: S21. Long-term behavioral modeling: For numerical features, normalization is performed, mapping the data to [0, 1] or a preset interval to eliminate the influence of different magnitudes, using the minimum-maximum normalization formula: In the formula, X is the original feature value, and Xmin and Xmax are the minimum and maximum values ​​of the feature column, respectively; For categorical features, one-hot encoding is used to convert each category into a binary vector. Rules are set according to different feature combinations to form a rule set, and an initial interest model is constructed. The K-Means clustering algorithm is used to cluster buyers into different groups using the processed buyer data as samples. S22. Short-term behavior modeling: Short-term behavioral features are assigned dynamic weights based on a time decay factor. The shorter the behavior duration, the larger the weight coefficient; the longer the behavior duration, the smaller the weight coefficient. The weights are calculated using an exponential decay function. In the formula, e is the natural constant, λ is the decay rate, and t is the time variable.

[0013] As a preferred technical solution, the filtering rules in step S3 are dynamically adjusted using regression or random forest models. The filtering model is trained based on historical filtering data, and the model adapts to business changes through periodic updates.

[0014] As a preferred technical solution, the advanced ranking algorithm in step S4 is XGBoost, which is used to rank candidate resources based on their combined features.

[0015] As a preferred technical solution, the XGBoost sorting algorithm determines the maximum depth of the tree, the learning rate, and the number of iterations, and iteratively constructs a decision tree in a gradient boosting manner. In each iteration, it fits the prediction residual of the previous round, optimizes the model by minimizing the objective function, and uses cross-validation to evaluate the model performance and adjust the parameters.

[0016] As a preferred technical solution, the candidate resources are sorted by inputting the candidate resource feature data into the trained XGBoost model, and the model outputs the prediction score of each candidate resource. The candidate resources are then arranged in descending or ascending order based on the prediction scores.

[0017] As a preferred technical solution, the candidate resources are sorted according to business objectives.

[0018] As a preferred technical solution, the following steps are also included: conducting A / B testing regularly to evaluate the effectiveness of the recommendation algorithm, adjusting the recommendation strategy based on user feedback, and optimizing the recommendation model by collecting feedback data in real time.

[0019] Another aspect of the present invention is to provide an electronic device including one or more processors and a memory, wherein the memory stores one or more programs, the one or more programs including instructions for executing the above-described B2B steel resource recommendation method based on the intensity of buyers' long-term and short-term behaviors.

[0020] Another aspect of the present invention provides a computer-readable storage medium including one or more programs executable by one or more processors of an electronic device, said one or more programs including instructions for performing the above-described B2B steel resource recommendation method based on the intensity of buyers' long-term and short-term behaviors.

[0021] Compared with the prior art, the present invention has at least one of the following beneficial effects: 1. To address the problem that recommendation models lack long-term behavior modeling and rely on short-term behavior modeling, this invention adopts a technical approach that combines long-term and short-term behavior modeling for comprehensive analysis and weights short-term behavior processing, thereby enabling the recommendation system to more accurately reflect the real needs of buyers.

[0022] 2. To address the issues of recommendation systems being unable to adapt to complex decision-making processes and having limited content recommendations, this invention introduces a business filtering layer, utilizes machine learning models, and combines business rules and objectives with historical data analysis to achieve a technical effect that makes the recommendation system more suitable for personalized recommendation services in the B2B steel industry and improves business adaptability.

[0023] 3. To address the issue of the lack of real-time response to buyer demands, this invention employs a technical means of dynamically adjusting filtering rules based on real-time changes in business needs. This enables the recommendation system to adjust recommendation results according to real-time changes in buyer behavior, making the recommendation system more flexible and adaptable to rapidly changing market demands, thus enhancing its real-time performance and dynamism.

[0024] 4. To address the issues of low conversion rates of candidate resources and inaccurate recommendation results, this invention employs an advanced ranking algorithm that combines the intensity of users' long-term and short-term behaviors with the diversity of recommended resources to optimize the order of resource arrangement, thereby achieving the technical effect of accurate recommendation of B2B steel resources. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0027] Example 1 To address the problems existing in the aforementioned prior art, this embodiment provides a B2B steel resource recommendation method based on the intensity of buyers' long-term and short-term behaviors, the implementation steps of which include: 1. Data Collection and Preprocessing: Collect long-term and short-term user behavior data, including browsing history, purchase history, query keywords, search frequency, etc., and perform preliminary processing.

[0028] 2. Recall Layer: In the recall phase, deep learning and time series analysis methods are used to construct a buyer behavior intensity sequence model. This model can capture changes in short-term behavior and uncover potential patterns in long-term behavior. Preliminary resource recall is performed by comparing the historical behavior of buyers with that of other similar buyers.

[0029] 3. Business Filtering Layer: This layer filters the recall results based on industry characteristics and business rules, removing resources that do not meet business needs. For example, a specific product may only be applicable to a particular industry or may not align with the company's long-term strategic needs.

[0030] 4. Ranking Layer: In the ranking stage, the order of resources is optimized by using a ranking algorithm, taking into account the intensity of users' long-term and short-term behaviors and the diversity of recommended resources, in order to ensure the accuracy of the recommendation results.

[0031] Figure 1 The diagram illustrates the specific process of the recommendation method of this invention. Feature engineering is performed on the collected and preprocessed data to extract behavioral features. Based on these behavioral features, long-term and short-term behavioral models are constructed. The filtering rules are then dynamically adjusted according to business needs. Candidate resources are ranked according to an advanced ranking algorithm. Recommended resources are displayed based on the ranking results. The method also includes A / B testing to optimize the recommendation model, thus achieving accurate recommendation of B2B steel resources.

[0032] like Figure 1 As shown in step 01, the data is collected and preprocessed: Collect user behavior data (such as clicks, browsing, and purchase records) and product data (such as categories, prices, and sales volume), and perform preprocessing such as cleaning, normalization, and missing value handling on the collected raw data to ensure data quality and consistency; construct and extract long-term and short-term user behavior features for use in subsequent recall layers.

[0033] In B2B steel trading, a well-designed and constructed feature set can significantly improve the effectiveness of accurate recommendations. This involves comprehensively considering long-term buyer behavior characteristics (such as purchase frequency, purchase amount, resource category preferences, buyer preferences for certain suppliers, and purchase cycles over a past period), short-term buyer behavior (such as search behavior, browsing behavior, click behavior, potential purchase behavior after adding items to the cart, and participation in promotional activities), cross-features between users and resources (such as resource-click conversion rate, the intensity of interest in resources after clicks / adding items to the cart, and historical resource similarity), and industry characteristics (social inventory, seasonal factors in the steel industry, etc.).

[0034] like Figure 1 As shown in step 02, feature engineering is used to extract multi-dimensional features that accurately reflect the long-term and short-term behavior of buyers. A specific example is provided below: Buyer long-term behavioral characteristics: (1) Purchase frequency Retrieve all purchase order records from the transaction system database for buyers within a specified time period (e.g., the past 1 year, 3 years, etc.). Calculate the number of orders within this time period and divide by the number of days in the period to obtain the average daily purchase frequency; alternatively, purchase frequency can be calculated by weekly or monthly time windows to analyze periodic changes in purchase frequency.

[0035] (2) Purchase amount In the transaction system, the transaction amount field for each order is extracted. The total purchase amount is obtained by summing the amounts of all orders within a specified time period; further calculations can be made to determine the average purchase amount per order (total purchase amount ÷ number of orders), which can be used to analyze the scale of purchases and purchasing power.

[0036] (3) Resource category preference From the order details, obtain information on the specific categories of steel purchased (such as rebar, wire rod, plate, etc.). Calculate the percentage of different resource categories in the total purchase volume; categories with higher percentages represent the buyer's preferred resource categories. Alternatively, calculate the proportion of each resource category's purchase amount to the total purchase amount, which more directly reflects value-based preferences.

[0037] (4) Purchasing preferences for certain suppliers Retrieve the supplier name or supplier ID field from the order records. Calculate the number of orders or the total purchase amount from different suppliers, and determine the percentage of each supplier in the total purchases. Suppliers with high percentages are those preferred by the buyer; alternatively, the dynamic changes in the buyer-supplier relationship can be analyzed by calculating trends in purchasing preferences over a period of time.

[0038] (5) Procurement cycle Extract the order placement time for each order. Analyze the time intervals between adjacent orders, and determine the buyer's purchasing cycle by statistically analyzing the mean and median of order intervals over a period of time. Time series analysis can also be used to observe whether there are seasonal or regular changes in the purchasing cycle.

[0039] Buyer short-term behavior characteristics: (1) Search behavior Obtain information such as buyer search keywords, search time, and search frequency from website or platform search logs. Analyze the search frequency for different keywords to understand buyer needs regarding steel product attributes (such as specifications and materials) and price ranges. Furthermore, time series analysis can be used to observe changes in search behavior frequency over short periods (such as the past week or month).

[0040] (2) Browsing behavior By utilizing website or platform page view logs, we can record the steel product pages visited by buyers, the duration of their visits, and the order in which they browse. By analyzing the number of views and total viewing time for different product pages, we can assess buyers' interest in different steel products. By analyzing the browsing order, we can infer buyers' purchasing decision-making process and product comparison behavior. Furthermore, we can calculate browsing depth (e.g., number of pages viewed ÷ number of visits) to measure the activity level of buyers' browsing behavior.

[0041] (3) Click behavior Record buyer clicks and their timing on steel product images, detail page links, and inquiry buttons in the website or platform's interaction logs. Analyze the click counts for different elements to understand buyer preferences regarding product display formats and information retrieval needs; compare the click-through rates (clicks ÷ impressions) of different products to assess product attractiveness and display effectiveness.

[0042] (4) Potential purchase behavior of adding vehicles From the shopping cart system, obtain records of when buyers add steel products to their carts and the timestamps. Analyze the quantity and total amount of products added to the cart to assess the potential scale of buyer purchases; calculate the cart conversion rate (order amount ÷ amount added to cart) to evaluate the efficiency of converting cart purchases into actual orders, and determine the conversion rate of buyer purchase intentions and behavior.

[0043] like Figure 1 As shown in step 03, the specific content of the recall layer is as follows: Based on long-term user behavior characteristics (such as the frequency of purchases, purchase amount, resource category preference, purchase preference for certain suppliers, and purchase cycle over a period of time), a long-term user interest model is constructed to identify users' interest preferences for certain resources.

[0044] For numerical features such as purchase frequency and purchase amount, normalization is performed to map the data to [0,1] or a specific interval, eliminating the influence of different magnitudes and facilitating subsequent calculations. The min-max normalization formula is used: , where X is the original feature value, and Xmin and Xmax are the minimum and maximum values ​​of the feature column, respectively.

[0045] For categorical features such as resource categories and suppliers, one-hot encoding is used to convert each category into a binary vector. For example, if there are three resource categories: rebar, wire rod, and plate, rebar is encoded as [1,0, 0], wire rod is encoded as [0, 1, 0], and plate is encoded as [0, 0, 1].

[0046] The rule-based model uses simple rules to build an initial model. For example, if a buyer has purchased rebar more than 50% of their total purchases in the past 12 months, and the purchase amount accounts for more than 40% of their total purchases, they are considered to have a high interest in rebar. Similar rules are set based on different combinations of characteristics to form a rule set, thus initially constructing an interest model.

[0047] The K-Means clustering algorithm was used, with the processed buyer data as samples, each containing the processed feature vector. The clustering algorithm divided buyers into different groups, where buyers within the same group exhibited similar behavioral characteristics, representing similar interests and preferences.

[0048] Analyze recent user behavior patterns (such as search behavior, browsing behavior, click behavior, potential purchase behavior via add-to-cart, and participation in promotional activities) to capture short-term changes in user interests. Assign dynamic weights to short-term behavioral characteristics (such as searches, clicks, add-to-cart, etc.) based on a time decay factor: the weight coefficient for the past 3 days is set to 0.8, for the past 7 days to 0.5, and for the past 15 days to 0.3, using an exponential decay function. (λ is the decay rate, which can be optimized based on business data) Calculate the weights to highlight the impact of recent behavior on interest preferences.

[0049] By modeling user behavior using deep neural networks and other models, we can identify potential needs that users are interested in but have not explicitly expressed.

[0050] By combining long-term and short-term interest modeling results, resources that highly match user interests are selected. At this point, the number of recalled resources may be large, requiring further refined filtering.

[0051] like Figure 1 As shown in step 04, the specific content of the business filtering layer is as follows: Based on actual business needs, such as resource type, steel grade and application, and downstream user groups, a series of filtering rules are defined. These rules may include product attribute filtering, inventory limits, and sales strategies.

[0052] Based on the resources and business needs identified in the recall results, a secondary screening process is performed to remove resources that do not fit the business scenario or current requirements. For example, if a certain type of resource is in short supply, or if a user is currently focused on resources from a specific supplier, the business filtering layer will automatically remove irrelevant resources.

[0053] Based on real-time changes in business needs (such as promotional activities and seasonal trends), the filtering rules are dynamically adjusted to ensure the timeliness and accuracy of recommended resources. Logistic regression or random forest models are introduced, and the intelligent filtering model is trained based on historical filtering data. For example, inputting features such as resource category, inventory level, and buyer industry tags, the output is a filtering probability (0-1), automatically removing resources with a filtering probability greater than 0.7. The model adapts to business changes through periodic updates (daily incremental training), improving the intelligence level of the filtering rules.

[0054] like Figure 1 As shown in step 05, the specific content of the sorting layer is as follows: Extract various features from the recalled resources, such as the intensity of buyer interest (output by a deep learning model), resource relevance (e.g., resource similarity, relevance), buyer purchase history, click-through rate of similar resources, conversion rate, etc.

[0055] Advanced ranking algorithms (such as XGBoost, LightGBM, etc.) are used to rank candidate resources by combining multiple features to ensure that the conversion rate of each candidate resource can be accurately evaluated.

[0056] The specific process of sorting based on sorting algorithms such as XGBoost is as follows: Various features are extracted from the recalled resources, including: buyer interest intensity (which can be represented by a score output by a deep learning model), resource relevance (calculated through resource similarity, such as cosine similarity to measure the similarity between steel product attributes), buyer purchase history (purchase frequency, amount, category preference, etc.), and click-through rate and conversion rate of similar resources.

[0057] For numerical features, min-max normalization is used to map the feature values ​​to the [0, 1] interval to eliminate the influence of dimensions; for categorical features, one-hot encoding is used to convert them into numerical vectors.

[0058] In the collected historical recommendation data, the actual conversion status of candidate resources (whether they were converted into purchases, which can be marked as 0 or 1) is used as a label, which together with the above features constitutes the training dataset.

[0059] Determine parameters such as the maximum depth of the tree (e.g., set to 6 to limit the complexity of the tree), the learning rate (e.g., 0.1 to control the step size of each iteration), and the number of iterations (e.g., 100 times).

[0060] The decision tree is built iteratively using gradient boosting. In each iteration, the model is fitted with the prediction residuals from the previous round. The model is optimized by minimizing the objective function (e.g., mean squared error for regression problems, log loss function for classification problems). During training, cross-validation (e.g., 5-fold cross-validation) is used to evaluate model performance, and parameters are adjusted to prevent overfitting.

[0061] The feature data of the candidate resources to be ranked are input into the trained XGBoost model, and the model outputs a prediction score for each candidate resource, which reflects the likelihood of resource conversion or the degree of matching with buyer demand.

[0062] Candidate resources are sorted in descending or ascending order based on their predicted scores (descending order is generally used for conversion rate prediction problems). Resources with higher scores are ranked first and are given priority for recommendation to buyers, thereby ensuring that the conversion rate of each candidate resource can be accurately assessed and precise recommendations can be achieved.

[0063] The algorithm model predicts the conversion rate of each candidate resource and ranks them according to the conversion rate. During the ranking process, other factors are taken into account, such as product display frequency, user's historical transaction limit, and marketing strategy, to further improve the recommendation effect.

[0064] Besides conversion rate, the ranking layer can further incorporate optimization of other objectives, such as increasing user purchase frequency and overall user spending. The objective function of the ranking algorithm should be adjusted according to business goals. This adjustment requires consideration of both business objectives and actual data feedback, as detailed below. 1. Add new target items To increase user purchase frequency, a term related to purchase frequency can be added to the original objective function (such as a function centered on conversion rate). For example, a metric measuring the interval between user purchases can be introduced. The shorter the interval between purchases, the higher the weight or contribution of this term in the objective function, thus guiding the model to recommend steel resources that encourage frequent purchases.

[0065] To increase users' overall purchasing volume, items related to users' historical or projected transaction amounts can be added. For example, based on users' past purchasing amounts, corresponding consumption growth expectations can be set for different users. In the objective function, this is reflected in the better ranking of recommended resources for users with high consumption expectations, thereby promoting their consumption growth.

[0066] 2. Adjust the weights of existing target items: When the business focus shifts from simply pursuing conversion rates to considering long-term user value (such as continued user purchases and platform loyalty), the weight of conversion rate in the objective function can be reduced, while the weight of metrics related to long-term user value (such as user retention rate and repurchase rate) can be increased. This way, the model will focus more on recommending resources that bring long-term benefits during optimization.

[0067] Conversely, if the platform prioritizes rapidly increasing order volume in the short term, it can increase the weight of conversion rate, allowing the model to prioritize recommending steel resources that are most likely to facilitate transactions.

[0068] Ensure that the recommendation results maximize overall benefits, including both short-term and long-term benefits; Short-term benefits: These mainly include steel sales revenue, gross profit, and order volume. They are quantified by statistically analyzing the actual transaction volume generated by recommended resources and calculating the profit per order (considering procurement costs, transportation costs, marketing costs, etc.). For example, if a user places an order after a certain batch of steel is recommended, the resulting increase in transaction volume represents the short-term benefit of that recommendation.

[0069] Long-term benefits include user retention rate (the percentage of users who continue to use the platform to purchase steel within a certain period), user repurchase rate (the frequency with which users make repeat purchases), user recommendation rate of the platform, and the platform's increased market share in the steel market. For example, by recommending high-quality steel resources that meet users' long-term needs, users are more willing to continue purchasing from the platform, thereby increasing user retention and repurchase rates, which constitutes long-term benefits.

[0070] like Figure 1 As shown in step 06, the specific content of the recommendation results is as follows: Based on the ranking results, recommended resources are selected and displayed to the user. Display formats may include personalized recommendations, homepage recommendations, and search result recommendations.

[0071] Conduct regular A / B testing to evaluate the effectiveness of the recommendation algorithm and adjust the recommendation strategy based on user feedback. Optimize the recommendation model by collecting real-time feedback data (such as clicks, purchases, retention rates, etc.) to improve user experience and conversion rates.

[0072] Example 2 Based on the foregoing embodiments, this embodiment provides an electronic device, including a processor, an internal bus, a network interface, memory, and non-volatile memory, and may also include other hardware required for various services. The processor reads the corresponding computer program from the non-volatile memory into memory and then runs it to achieve the above-mentioned functions. Figure 1The method described above is a B2B steel resource recommendation method based on the intensity of buyers' long-term and short-term behaviors. Of course, besides software implementation, this invention does not exclude other implementation methods, such as logic devices or a combination of hardware and software. In other words, the execution entity of the processing flow is not limited to individual logic units, but can also be hardware or logic devices.

[0073] Example 3 Based on the foregoing embodiments, this embodiment provides a computer-readable storage medium storing a computer program that can be used to execute the above-described... Figure 1 This document provides a B2B steel resource recommendation method based on the intensity of buyers' short-term and long-term behaviors. Computer-readable media, including both permanent and non-permanent, removable and non-removable media, can store information using any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0074] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A B2B steel resource recommendation method based on the intensity of buyers' long-term and short-term behavior, characterized in that, Includes the following steps: S1. Collect long-term and short-term user behavior data, including purchase frequency, purchase amount, resource category preference, supplier purchase preference, purchase cycle, browsing behavior, click behavior and potential purchase behavior of adding items to cart, and clean, normalize and handle missing values ​​of the data to construct and extract long-term and short-term user behavior characteristics; S2. During the recall phase, deep learning and time series analysis methods are used to construct a buyer behavior intensity sequence model to capture users' long-term and short-term interests and potential needs, and to screen out resources that are highly matched with users' interests. S3. Filter the recall results based on industry characteristics and business rules, remove resources that do not meet business needs, and dynamically adjust the filtering rules according to real-time changes in business needs; S4. In the ranking stage, based on the advanced ranking algorithm, combined with the long-term and short-term behavioral intensity of users and the diversity of recommended resources, the order of resource arrangement is optimized to realize the recommendation of B2B steel resources.

2. The B2B steel resource recommendation method based on the intensity of buyers' long-term and short-term behaviors as described in claim 1, characterized in that, In step S2, the buyer behavior intensity sequence model specifically includes: S21. Long-term behavioral modeling: For numerical features, normalization is performed, mapping the data to [0, 1] or a preset interval to eliminate the influence of different magnitudes, using the minimum-maximum normalization formula: In the formula, X is the original feature value, and Xmin and Xmax are the minimum and maximum values ​​of the feature column, respectively; For categorical features, one-hot encoding is used to convert each category into a binary vector. Rules are set according to different feature combinations to form a rule set, and an initial interest model is constructed. The K-Means clustering algorithm is used to cluster buyers into different groups using the processed buyer data as samples. S22. Short-term behavior modeling: Short-term behavioral features are assigned dynamic weights based on a time decay factor. The shorter the behavior duration, the larger the weight coefficient; the longer the behavior duration, the smaller the weight coefficient. The weights are calculated using an exponential decay function. In the formula, e is the natural constant, λ is the decay rate, and t is the time variable.

3. The B2B steel resource recommendation method based on the intensity of buyers' long-term and short-term behaviors as described in claim 1, characterized in that, For the filtering rules in step S3, dynamic adjustment is achieved using logistic regression or random forest models. The filtering model is trained based on historical filtering data and adapted to business changes through periodic updates.

4. The B2B steel resource recommendation method based on the intensity of buyers' long-term and short-term behaviors as described in claim 1, characterized in that, The advanced ranking algorithm in step S4 is XGBoost, which is used to rank candidate resources based on their combined features.

5. The B2B steel resource recommendation method based on the intensity of buyers' long-term and short-term behaviors according to claim 4, characterized in that, The XGBoost ranking algorithm determines the maximum depth of the tree, the learning rate, and the number of iterations. It iteratively constructs a decision tree using gradient boosting, fits the prediction residual of the previous round in each iteration, optimizes the model by minimizing the objective function, and uses cross-validation to evaluate model performance and adjust parameters.

6. The B2B steel resource recommendation method based on the intensity of buyers' long-term and short-term behaviors according to claim 4, characterized in that, The candidate resources are sorted by inputting their feature data into a trained XGBoost model. The model outputs a prediction score for each candidate resource, and the candidate resources are sorted in descending or ascending order based on the prediction scores.

7. The B2B steel resource recommendation method based on the intensity of buyers' long-term and short-term behaviors according to claim 4, characterized in that, The candidate resources are sorted according to business objectives.

8. The B2B steel resource recommendation method based on the intensity of buyers' long-term and short-term behaviors according to claim 1, characterized in that, It also includes the following steps: Regularly conduct A / B testing to evaluate the effectiveness of the recommendation algorithm, adjust the recommendation strategy based on user feedback, and optimize the recommendation model by collecting feedback data in real time.

9. An electronic device, characterized in that, include: One or more processors and a memory, wherein the memory stores one or more programs, the one or more programs including instructions for executing the B2B steel resource recommendation method based on the intensity of buyer long-term and short-term behavior as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, Includes one or more programs executed by one or more processors of an electronic device, said one or more programs including instructions for executing the B2B steel resource recommendation method based on the intensity of buyer long-term and short-term behavior as described in any one of claims 1-8.

Images