Commodity recommendation system and method based on dynamic index combination
By using a product recommendation system based on dynamic indicator combinations and leveraging MPP databases and materialized view acceleration technology, the system solves the problem of real-time calculation of tens of millions of users and products on e-commerce platforms. This enables efficient and flexible recommendation logic configuration and accurate recommendations, while reducing operating costs.
Patent Information
- Application Number
- CN202511350456.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-22
- Publication Date
- 2026-02-06
AI Technical Summary
On e-commerce platforms with tens of millions of users and tens of millions of products, existing recommendation systems face challenges such as insufficient real-time computing power, inability to integrate operational strategies into recommendation logic, and a lack of interpretability and control over recommendation results.
A product recommendation system based on dynamic index combinations is adopted, including a data storage layer, a configuration management layer, and a recommendation service layer. It utilizes MPP database and materialized view acceleration technology, combined with a dynamic template engine and configuration management, to achieve efficient computing and flexible configuration.
It achieves efficient computing under large data volumes, supports flexible business needs, reduces operating costs, improves the accuracy and interpretability of recommendations, and supports multi-scenario adaptation.
Smart Images
Figure CN121481658A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of commodity recommendation, and in particular to a commodity recommendation system and method based on a dynamic index combination. BACKGROUND
[0002] With the development of e-commerce platforms, the number of commodity types and users has increased dramatically. As an important personalized service tool in the platform, the performance and recommendation effect of the recommendation system directly affect user experience and platform conversion rate. The core task of the recommendation system is to determine which commodities a user may like and to prioritize the display of these commodities in the user's browsing page, thereby improving click rate, purchase rate, and user satisfaction. Traditional recommendation methods are mostly based on a "user x commodity" scoring model, such as collaborative filtering, deep learning ranking, etc. The essence of these methods is to score each user and all commodities in the platform, and then rank and recommend them according to the score. This approach can work effectively when the number of users is small and the number of commodities is limited, but in actual e-commerce platforms, it faces several major challenges:
[0003] 1. Large number of users and commodities: The number of registered users on current mainstream e-commerce platforms can reach hundreds of millions, and the number of online commodities often exceeds tens of millions. In this case, any comprehensive calculation of "user x commodity" will involve extremely large data volume. For example, if there are 100 million users and 10 million commodities, a complete round of scoring will involve 1015 calculations. This not only requires extremely high computing resources, but also makes it difficult to output the recommendation results in real time, seriously affecting user experience.
[0004] 2. Difficulty in integrating platform operation goals into recommendation logic: The recommendation system often prioritizes user interest, while platform operation strategies (such as promoting new products, clearing inventory, and promoting activities) belong to another angle of goals. In traditional methods, operation strategies are often implemented through hard insertion of advertising slots or manual configuration of rules, which are separate from the logic of the recommendation system and cannot intelligently integrate recommendation content with operation needs.
[0005] 3. Lack of explainability and controllability of recommendation results: With the platform's emphasis on fine-tuned operations, the explainability and controllability of recommendation results have become a pain point. The platform wants to understand "why a certain user is recommended a certain commodity" and can adjust the recommendation strategy when needed, such as strengthening the exposure of certain product categories, brands, or activity commodities in a certain time period. However, traditional models are often "black box" type and cannot provide a clear recommendation logic path. SUMMARY
[0006] The application aims to provide a commodity recommendation system and method based on dynamic index combination to solve the problems in the background art: 1, insufficient real-time computing capacity in the scene of ten million users and ten million commodities; 2, operation strategy cannot be integrated into recommendation logic through configuration; 3, the recommendation result lacks explainability and control.
[0007] To achieve the above application purpose, one aspect of the application provides a commodity recommendation system based on dynamic index combination, comprising a data storage layer, a data processing and computing layer, a configuration management layer, and a recommendation service layer, wherein:
[0008] The data storage layer is used to establish an MPP (Massively Parallel Processing) database, store commodity dimension data, sales index data, person-commodity algorithm index, and bitmap index;
[0009] The configuration management layer is used to provide a user interface for configuration management and define a recommendation task;
[0010] The data processing and computing layer is used to process data and computing, and generate a commodity list;
[0011] The recommendation service layer is used to accept a recommendation request, call the configuration management platform to obtain corresponding recommendation configuration, drive the data processing and computing layer to generate a recommendation table, and return a commodity list.
[0012] Further, the person-commodity algorithm index is a user preference score index of a specific commodity calculated based on user behavior, and the user behavior includes purchase frequency, browsing time, and collection behavior.
[0013] Further, the data storage layer adopts the Materialized Views (MV) acceleration technology of the MPP database to pre-compute and store sales index data, and the sales index data includes person-category index, person-brand index, person-commodity index, category itself index, brand itself index, and commodity itself index, thereby avoiding huge overhead of real-time computing.
[0014] Further, the data storage layer adopts the Bitmap technology of the MPP database to create a bitmap index of a user, compute recommendation index computing logic configured in the system, and realize efficient computation of large data.
[0015] Further, the configuration management layer provides user modification of index selection and weight allocation, sets a time range and sorting method of data statistics for each selected index, performs normalization on the index value, defines the total number of recommended commodities and the proportion of each index in the total recommendation amount.
[0016] Further, the configuration management layer is configured with a dynamic template engine, and a user can freely configure related indexes that meet the recommended expectations, and the platform accepts the configured parameters, dynamically generates SQL query statements or data processing logic, and embeds the user-configured parameters into the predefined SQL template or data processing script.
[0017] Further, the data processing and calculation layer includes a dynamic query generator, an index aggregation and calculation module, an index normalization and weighting module, a sorting and screening module, and a drilling processing module, wherein:
[0018] The dynamic query generator is used to generate a query statement for the MPP database based on the dynamic template engine and the parameters of the configuration platform;
[0019] The index aggregation and calculation module is used to execute the dynamically generated query to retrieve data from the materialized view;
[0020] The index normalization and weighting module is used to normalize the retrieved index data, and perform weighted summation according to the user-set weight to calculate a comprehensive score;
[0021] The sorting and screening module is used to sort the goods according to the comprehensive score, and screen out the final recommended goods list according to the configured TopN data and the proportion of each index;
[0022] The drilling processing module is used to automatically associate all goods under the selected category or brand of the user according to the configuration, and perform index aggregation, normalization, weighting and sorting at the goods level.
[0023] Another aspect of the application provides a goods recommendation method based on dynamic index combination, comprising the following steps:
[0024] Step S1, a user defines a recommendation task on a configuration management platform, selects calculation indexes, sets respective time ranges, sorting methods, normalization methods and weights;
[0025] Step S2, a dynamic template engine generates a SQL query statement according to the configuration, and the statement efficiently extracts the required data from a pre-constructed MPP database materialized view;
[0026] Step S3, a data processing and calculation layer executes the query to obtain original index data;
[0027] Step S4, an index normalization and weighting module of the data processing and calculation layer processes the data to calculate a comprehensive score of each good;
[0028] Step S5, a sorting and screening module of the data processing and calculation layer determines a final recommended goods list according to the comprehensive score and the proportion requirement;
[0029] Step S6, the recommended service layer returns the generated recommendation list.
[0030] Further, the determination of the final recommended commodity list according to the comprehensive score and the proportion requirement in step S5 includes the following steps:
[0031] Step S501, descending order arrangement according to the comprehensive score;
[0032] Step S502, according to the total number of recommendations N configured and the weight proportion of each index, TopN screening is performed;
[0033] Further, for the category or brand index, through the dynamic index combination, the indexes of goods, brands and categories in each dimension are combined with each other, and normalization and sorting are performed.
[0034] Compared with the prior art, the system and method have the following advantages:
[0035] 1. High flexibility and dynamicity: the present application can realize highly customized recommendations by dynamically combining various indexes, time ranges, sorting and weights according to the changing business requirements through the dynamic template engine and the configuration platform.
[0036] 2. Excellent performance: the present application uses materialized view acceleration to pre-calculate and store complex aggregation data (such as person-category, person-goods index), avoiding the huge overhead of real-time calculation and greatly shortening the query response time. Bitmap is also used to accelerate UV calculation, which can efficiently and cost-effectively perform user deduplication and significantly improve the calculation efficiency of UV index.
[0037] 3. Strong configurability and easy iteration: the recommendation logic of the present application is encapsulated in the template, which can be adjusted by configuration without modifying the code, greatly reducing the cost of system maintenance and iteration.
[0038] 4. Precise recommendation granularity: the present application supports drilling down the category and brand indexes to the goods dimension for weighted sorting, ensuring the accuracy of the recommendations.
[0039] 5. Scientific recommendation quantity allocation: the index weight directly determines the quantity proportion of the recommended goods, making the recommendation results more reflect the business focus configured by the user. BRIEF DESCRIPTION OF DRAWINGS
[0040] Figure 1 It is a kind of category commodity recommendation system structure diagram based on dynamic index combination and materialized view acceleration.
[0041] Figure 2A kind of based on dynamic index combination and materialized view acceleration category commodity recommendation method flow chart. DETAILED DESCRIPTION
[0042] The technical solutions in the embodiments of the present application will be described clearly and completely below with the drawings in the embodiments of the present application.Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.
[0043] As Figure 1 It is the system structure diagram of an embodiment of the present application, and the present application discloses a kind of based on dynamic index combination and materialized view acceleration category commodity recommendation system. The system mainly includes data storage layer, data processing and calculation layer, configuration management layer and recommendation service layer.
[0044] Data storage layer is an MPP database, stores commodity dimension data, sales index data, person-commodity algorithm index and bitmap index. Commodity dimension data includes the hierarchical relationship of storage category (first level, second level, third level, fourth level), brand, commodity and the basic attributes of commodity (such as shelf time, etc.).
[0045] Sales index data is stored in the form of materialized view, including:
[0046] Person-category index: the order quantity, sales net, sales volume, PV (page access, Page View) of specific person (user) in specific category are counted;
[0047] Person-brand index: the order quantity, sales net, sales volume, PV of specific person (user) in specific brand are counted;
[0048] Person-commodity index: the order quantity, sales net, sales volume, PV of specific person (user) in specific commodity are counted;
[0049] Category itself index: the order quantity, sales net, sales volume, UV (independent visitor, Unique Visitor, using Bitmap optimization), shelf time from now on time length of category (each level) are counted;
[0050] Brand itself index: the order quantity, sales net, sales volume, UV (using Bitmap optimization) of brand (each level) are counted;
[0051] Commodity itself index: the order quantity, sales net, sales volume, UV (using Bitmap optimization), shelf time from now on time length of commodity (each level) are counted.
[0052] The person-goods algorithm index is used to store the algorithm index of the user's preference score for a specific good calculated based on user behavior (such as purchase frequency, browsing time, collection behavior, etc.).
[0053] The calculation of the person-goods algorithm index is based on user objective behavior data, and the core data types include but are not limited to the following categories, and each category of data needs to record the timestamp of the behavior occurrence, the associated good ID and user ID to ensure the traceability and timeliness of the data:
[0054] Purchase behavior data: including user's historical purchase records (such as purchase time, purchase quantity, purchase amount), repeat purchase situation (such as the number of repeat purchases of the same good, repeat purchase cycle), etc. This kind of data is a strong signal reflecting user preference, and usually has the highest weight contribution to the preference score.
[0055] Browsing behavior data: including user's browsing time on the good detail page (single browsing time, cumulative browsing time), browsing frequency (number of times a specific good is browsed), browsing depth (such as whether to view good reviews, specifications, related recommendations, etc.), and jump-out situation (such as whether to open the page and close it immediately), etc.
[0056] Collection and attention behavior data: including user's collection operation on goods (such as the time of adding to the collection folder, whether to keep it for a long time), attention behavior to merchants or good series, etc. This kind of data shows that the user has explicit interest in the goods, which is an important manifestation of potential purchase intention.
[0057] Interaction behavior data: including user's evaluation of goods (such as good, medium, and poor evaluations and evaluation content), question and answer interaction (such as asking questions to merchants or other users, answering other people's questions), sharing behavior (such as sharing goods to social platforms), and adding to cart behavior (such as adding goods to the shopping cart but not immediately purchasing).
[0058] The calculation of the person-goods algorithm index (with preference score as the core) follows the whole process logic of "data preprocessing - feature engineering - algorithm modeling - score output and update", which is as follows:
[0059] Data preprocessing: cleaning and standardizing the collected user behavior raw data, including handling missing values (such as filling in the null values when the user does not fill in the evaluation), removing outliers (such as high-frequency invalid browsing data generated by misoperation), data format unification (such as converting different time formats to standard timestamps), and data deduplication (such as merging repeated browsing records of the same user at the same time for the same good).
[0060] Feature engineering: Convert the pre-processed behavior data into feature variables that can be used for algorithm modeling, for example: convert "purchase frequency" into "purchase times in the past 30 days", "purchase times in the past 90 days", etc. time dimension features; convert "browsing time" into "average browsing time per session", "browsing time accounts for the proportion of the user's total browsing time during the same period", etc. relative features; convert "collection behavior" into "collection status (0 = not collected, 1 = collected)", "collection time", etc. Boolean and numerical features.
[0061] Algorithm modeling: Based on the above feature variables, a specified machine learning model (such as collaborative filtering, DSSM, DeepFM, etc.) is used to build a preference prediction model, and the target variable of the model is "user preference for a specific product". The mapping relationship between the feature variables and the target variables is learned through model training, and then the preference score is calculated.
[0062] Score output and update: After model training, the corresponding preference score is output for each "user - product" combination (usually using 0-100 points or 0-1 standardized points, the higher the score, the stronger the preference). At the same time, a dynamic updating mechanism is established, and when new behavior data (such as new purchases, browsing records) is generated by the user, the preference score is recalculated in real time or quasi-real time (such as every hour / every day), ensuring the timeliness of the index.
[0063] The machine learning model used for the calculation of the person - product algorithm index as a whole follows the core logic of "based on user behavior and product related data, the model is used to mine associated features and quantify preferences": first, based on user purchase, browsing, collection and other objective behavior data, combined with product attributes and other information, after data preprocessing and feature transformation, the model automatically captures the association between users and products, including similarity association between user groups or products, deep association between user behavior time series and product semantics, and low-order and high-order interaction association between multi-dimensional features. Finally, these association relationships are converted into quantifiable user preference scores for specific products, and can be dynamically updated with new data to support accurate preference evaluation.
[0064] Bitmap index creates a bitmap index for users in MPP database to efficiently calculate UV.
[0065] The configuration management layer is used to provide a user interface for configuration management, allowing users to configure recommendation tasks, including:
[0066] Index selection and weight allocation: users can select various sales indicators and algorithm indicators mentioned above and allocate weights to them;
[0067] Time Range Setting: Set the time range for data statistics for each selected indicator (e.g., the last 7 days, the last 30 days);
[0068] Sorting Criteria Selection: Set the sorting method for each selected indicator (ascending order, descending order, random order);
[0069] Normalization Method: Select the method for normalizing indicator values (e.g., Min-Max normalization, Z-Score normalization);
[0070] TopN Quantity and Proportion: Define the total number of recommended products (N) and the proportion of each indicator in the total number of recommendations (e.g., if the weight of indicator 1 is 0.2, then the proportion of the number of recommended products is N*0.2).
[0071] The dynamic template engine in the configuration management layer receives parameters of the configuration platform and dynamically generates SQL query statements or data processing logic. This engine is responsible for embedding user-configured indicators, time ranges, sorting, weights, and other parameters into pre-defined SQL templates or data processing scripts.
[0072] The dynamic template engine is based on the SQL generation and dynamic factor calculation system of the Thymeleaf dynamic rule, and its core components include:
[0073] Dynamic Rule Template Management Module: Stores pre-defined SQL templates containing placeholders and expressions. These templates mimic the syntax of Thymeleaf, supporting conditional statements (if), loops (each), variable fragments (define), and other functions, which are used to define the structure and logic of SQL.
[0074] Dynamic Rule Parsing Engine: Receives external incoming rule parameters (such as business conditions, time ranges, user attributes, etc.), parses and replaces placeholders and expressions in the SQL template. This engine can dynamically include or exclude certain parts of the SQL statement based on the incoming parameter values, and perform value escaping to prevent SQL injection.
[0075] SQL Execution Module: Receives the SQL statement generated by the parsing engine and submits it to the database for execution. This module is responsible for connecting to the database, executing SQL, and returning the query results.
[0076] Dynamic Factor Calculation Module: Receives data returned by the SQL execution module and performs data aggregation, conversion, and calculation based on pre-set calculation logic (which may also contain dynamic rules) to generate the final dynamic factors. This module can orchestrate multiple SQL generation and execution processes and integrate the results into a calculation flow.
[0077] Rule Parameter Interface: Provides a standardized interface for external applications to input dynamic rule parameters.
[0078] The system workflow is as follows:
[0079] 1. The external application transmits dynamic rule parameters through the interface;
[0080] 2. The dynamic rule analysis engine generates the final SQL statement according to the parameters and the preset template;
[0081] 3. The SQL execution module executes the SQL and obtains data;
[0082] 4. The dynamic factor calculation module calculates using the obtained data, and can call the SQL generation process again to complete multi-step calculation as needed;
[0083] 5. The calculation result is returned to the external application.
[0084] The data processing and calculation layer is used for processing data and calculation, and generating a commodity list. It includes:
[0085] Dynamic query generator: based on the dynamic template engine and the parameters of the configuration platform, generate a query statement for the MPP database;
[0086] Indicator aggregation and calculation module: execute the dynamically generated query to retrieve data from the materialized view;
[0087] Indicator normalization and weighting module: normalize the retrieved indicator data, and then perform weighted summation according to the user-set weights to calculate the comprehensive score;
[0088] Sorting and filtering module: sort the commodities according to the comprehensive score, and filter out the final recommended commodity list according to the configured TopN number and the proportion of each indicator;
[0089] The drilling processing module calculates the fact indicator TopM (MIN (M equals the total number of recommended commodities * weight, 10)) of the category / brand in advance according to the weight, and then calculates the corresponding commodity score topN (N = total number of recommended commodities * weight / M) according to these dimensions. When the user selects a category or brand indicator, this module will automatically associate all commodities under the category / brand according to the configuration, and perform commodity-level indicator aggregation, normalization, weighting, and sorting.
[0090] The specific process of the drilling processing module for category / brand indicators, taking category A as an example, the data is from the MPP database:
[0091] Associate commodities: associate the commodity table of the MPP database through category ID='A001', and filter out all on-shelf commodities (1200 in total, excluding 200 off-shelf commodities) under category A;
[0092] Indicator aggregation: Extract the order volume and PV of each product in the last 7 days from the'sales_mv' (sales indicator materialized view) of the MPP database, aggregate the overall indicators of category A (such as the total order volume of category A in the last 7 days = the sum of the order volumes of each product), and aggregate the overall indicators of category A (such as the total order volume of category A normalized value = 0.5).
[0093] Score calculation: First, normalize the overall indicators of category A (such as the total order volume of category A normalized value = 0.5), then calculate the overall score according to the proportion of product indicators to category indicators (such as product 1 order volume accounts for 10% of category A → product 1 indicator score = 0.5x10% = 0.05), and combine other indicator weights to calculate the comprehensive score.
[0094] Quota allocation: If the indicator weight of category A is 0.3, and the total number of recommendations N = 100, then select the top 30 (100x0.3) products from the products under category A according to the comprehensive score, and store the results in the recommendation result table of the MPP database.
[0095] The recommendation service layer receives a recommendation request from the front end or other business systems, and the parameters include user ID, recommendation scenario, configuration ID, etc., then calls the configuration management platform to obtain the corresponding recommendation configuration, drives the data processing and calculation layer to generate the recommendation list, and returns the result.
[0096] As shown in Figure 2 Another embodiment of the present application is disclosed, which is a product recommendation method based on dynamic indicator combination, comprising the following steps:
[0097] Step S1, the user uses the configuration management layer to define a recommendation task, selects indicators (such as the order volume of category A, the sales net of brand B, and the PV of product C), sets the respective time range, sorting method, normalization method and weight.
[0098] Step S2, the dynamic template engine generates a SQL query statement according to the configuration, which efficiently extracts the required data from the pre-constructed MPP database materialized view. The recommendation service layer receives the user request, finds the corresponding configuration, and passes the configuration parameters (such as indicator list, weight, time range, etc.) to the dynamic template engine. The dynamic template engine parses the configuration, generates a SQL containing indicator selection, time filtering, normalization, weighting and sorting, executes in the MPP database, and quickly returns the pre-aggregated indicator data using the materialized view.
[0099] Step S3, the data processing layer executes the query to obtain the original indicator data.
[0100] Step S4, the index normalization and weighting module processes the data and calculates the comprehensive score of each product. UV calculation uses Bitmap index for efficient deduplication. According to the configured normalization method (such as Min-Max, Z-Score), the extracted index value is processed to be in a comparable range. Multiply the normalized index value by its corresponding weight, then add it up to get the comprehensive score of each product.
[0101] Step S5, the sorting and TopN filtering module determines the final recommended product list according to the comprehensive score and the proportion requirement. It includes the following steps:
[0102] Step S501, arrange in descending order (or user-specified sorting) according to the comprehensive score.
[0103] Step S502, according to the configured total number of recommendations N, and the weight proportion of each index (for example, index 1 weight 0.4, then allocate 0.4*N vacancies to the goods mainly affected by index 1 from the total number of recommendations N), perform TopN filtering.
[0104] It will perform category or brand drilling. That is, if the user selects a category or brand index, step S5 will ensure that the aggregated index of all goods under the category / brand is calculated, and normalized and sorted. For example, if a certain index is "net sales of category A", and the weight is 0.4, then when calculating the comprehensive score, the score of product A1 (belonging to category A) will be affected by the overall net sales of category A, and when the final 100 products are selected, about 40 vacancies will be inclined to be contributed by the category.
[0105] Step S6, the recommendation service layer returns the generated recommendation list. The recommendation service layer receives the user request, finds the corresponding configuration, passes the configuration parameters (such as index list, weight, time range, etc.) to the dynamic template engine, generates SQL, executes in MPP database, and organizes the product ID, ranking, etc. Information into the final recommendation list returned to the caller.
[0106] The system reduces the human cost of operating the recommendation column in the e-commerce platform from 29 person-days to 2 person-days, reducing about 90%, realizing a qualitative leap in operational response capability. The system realizes flexible adaptation in multiple scenarios, improving the self-operation capability of the operation team.
[0107] Although embodiments of the present application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made therein without departing from the principles and spirit of the application, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A product recommendation system based on dynamic index combinations, characterized in that, It includes a data storage layer, a configuration management layer, a data processing and computing layer, and a recommendation service layer, among which: The data storage layer is used to build the MPP database, storing product dimension data, sales indicator data, person-product algorithm indicators, and bitmap indexes; The configuration management layer provides a user interface for configuration management and defines recommendation tasks using dynamic metric combinations. The data processing and computing layer is used to process data and perform calculations to generate a product list; The recommendation service layer is used to accept recommendation requests, call the configuration management platform to obtain the corresponding recommendation configuration, drive the data processing and calculation layer to generate a recommendation table, and return a list of products.
2. The product recommendation system based on dynamic index combination according to claim 1, characterized in that, The person-product algorithm metric is a user preference score for a specific product calculated based on user behavior, including purchase frequency, browsing time, and collection behavior.
3. The product recommendation system based on dynamic index combination according to claim 1, characterized in that, The data storage layer uses the materialized view acceleration technology of MPP database to pre-calculate and store sales indicator data, which includes person-category indicators, person-brand indicators, person-product indicators, category-specific indicators, brand-specific indicators, and product-specific indicators, thus avoiding the huge overhead of real-time calculation.
4. A product recommendation system based on dynamic index combination according to claim 1, characterized in that, The data storage layer uses Bitmap technology from the MPP database. By creating a bitmap index for the user, it calculates the recommendation metric calculation logic configured in the system, achieving efficient calculation of large amounts of data.
5. A product recommendation system based on dynamic index combination according to claim 1, characterized in that, The configuration management layer allows users to modify indicator selection and weight allocation, set the time range and sorting method for data statistics for each selected indicator, normalize indicator values, and define the total number of recommended products and the proportion of each indicator in the total recommendations.
6. A product recommendation system based on dynamic index combination according to claim 1, characterized in that, The configuration management layer is equipped with a dynamic template engine, which receives parameters from the configuration platform, dynamically generates SQL query statements or data processing logic, and embeds user-configured parameters into predefined SQL templates or data processing scripts.
7. A product recommendation system based on dynamic index combination according to claim 1, characterized in that, The data processing and computation layer includes a dynamic query generator, an indicator aggregation and calculation module, an indicator normalization and weighting module, a sorting and filtering module, and a drill-down processing module, among which: The dynamic query generator is used to generate query statements for MPP databases based on parameters from the dynamic template engine and configuration platform. The indicator aggregation and calculation module is used to execute dynamically generated queries and retrieve data from materialized views; The indicator normalization and weighting module is used to normalize the retrieved indicator data, perform weighted summation according to the weights set by the user, and calculate the comprehensive score. The sorting and filtering module is used to sort products based on their comprehensive scores and filter the final list of recommended products based on the configured TopN data and the proportion of each indicator. The drill-down processing module is used to automatically associate all products under the category or brand selected by the user according to the configuration, and perform product-level indicator aggregation, normalization, weighting and sorting.
8. A product recommendation method based on dynamic index combinations, characterized in that, Includes the following steps: Step S1: Users define recommendation tasks in the configuration management platform, select calculation metrics, and set their respective time ranges, sorting methods, normalization methods, and weights. Step S2: The dynamic template engine generates an SQL query statement based on the configuration. The statement will efficiently extract the required data from the pre-built MPP database materialized view. Step S3: The data processing and calculation layer executes a query to obtain the raw indicator data; Step S4: The index normalization and weighting module of the data processing and calculation layer processes the data and calculates the comprehensive score for each product. Step S5: The sorting and filtering module of the data processing and calculation layer determines the final recommended product list based on the comprehensive score and proportion requirements. In step S6, the recommendation service layer returns the generated recommendation list.
9. A product recommendation method based on dynamic index combination according to claim 8, characterized in that, Step S5, based on the overall score and proportion requirements, determines the final recommended product list, including the following steps: Step S501: Sort the results in descending order based on the overall score; Step S502: Based on the total number of recommendations N and the weight ratio of each indicator, the TopN is selected.
10. A product recommendation method based on dynamic index combination according to claim 9, characterized in that, For category or brand metrics, a dynamic metric combination method is used to combine metrics from various dimensions of products, brands, and categories, and then normalize and sort them.