Collaborative filtering recommendation method and device based on user grading
By optimizing user segmentation and similarity calculation, and combining popularity decay coefficient and cold start strategy, the problems of unmet high-value user needs and Matthew effect in collaborative filtering recommendation are solved, achieving more accurate recommendation results and platform content diversity.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SUZHOU BLUE DOLPHIN INTERACTIVE INFORMATION TECH CO LTD
- Filing Date
- 2026-02-05
- Publication Date
- 2026-05-12
AI Technical Summary
Existing collaborative filtering recommendation technologies fail to effectively incorporate user segmentation information, resulting in unmet needs of high-value users, a Matthew effect, poor cold-start recommendation performance, and limited exploration of user interests.
By employing user-level tiering, data preprocessing, and feature extraction, and adjusting similarity calculations and popularity decay coefficients based on user-level weights, combined with a cold-start strategy to optimize recommendations, the system dynamically adjusts recommendation priorities and diversity to meet the needs of high-value users and mitigate the Matthew effect.
It improved the satisfaction and retention rate of high-value users, promoted the health and diversity of the platform's content ecosystem, and improved the user experience and conversion potential during the cold start phase.
Smart Images
Figure CN122019890A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of Internet information technology, and more specifically, to a collaborative filtering recommendation method and apparatus based on user classification. Background Technology
[0002] With the rapid development of internet platforms (such as recruitment, e-commerce, and service platforms), recommendation systems have become a core bridge connecting users with goods, services, or content, and their performance directly impacts user experience and platform business efficiency. Collaborative filtering, as a classic recommendation algorithm, makes recommendations by analyzing the behavioral similarities of user groups or the relevance of items. It has advantages such as not relying on the attributes of items or services themselves and being able to discover potential user interest associations, and is widely used in various scenarios.
[0003] However, existing collaborative filtering recommendation techniques still have several significant drawbacks in practical applications:
[0004] First, traditional collaborative filtering algorithms typically treat all users as homogeneous, ignoring the inherent differences within user groups. In real-world platforms, users exhibit clear stratification characteristics, such as ordinary users, paying users, and high-value users. Current technologies do not deeply integrate this user stratification information with recommendation algorithms, resulting in an inability to prioritize the personalized needs of high-value users, and the benefits of paying users are difficult to effectively reflect through recommendation mechanisms. This impacts the retention of core users and the platform's commercial value conversion.
[0005] Secondly, traditional collaborative filtering recommendations are susceptible to the "Matthew effect," where popular content is repeatedly recommended and its exposure increases, while a large amount of high-quality long-tail content (such as niche jobs, unique products, and emerging services) lacks exposure opportunities, leading to an imbalance in the platform's content ecosystem and limited exploration of user interests.
[0006] Furthermore, in cold start scenarios where user behavior data is sparse (such as new users or users with no historical behavior), collaborative filtering algorithms struggle to accurately calculate user similarity, leading to a significant drop in recommendation accuracy and impacting the initial experience and conversion rate of new users.
[0007] Furthermore, while user classification exists in existing technologies, it is mostly limited to access control or simple tagging, and the classification results (such as level weights) are not deeply integrated into the core calculation links of recommendation algorithms (such as similarity calculation, candidate set selection, score ranking, etc.), resulting in the value of the user classification system not being fully realized in recommendation systems.
[0008] Therefore, there is an urgent need for a technical solution that can deeply integrate user segmentation with collaborative filtering recommendation, which can both prioritize the needs of high-value users and alleviate the Matthew effect and improve the cold start recommendation effect. Summary of the Invention
[0009] The purpose of this invention is to address the shortcomings of existing technologies by proposing a collaborative filtering recommendation method and apparatus based on user hierarchy.
[0010] To address the aforementioned technical problems, the present invention adopts the following technical solution:
[0011] In a first aspect, the present invention provides a collaborative filtering recommendation method based on user hierarchy, comprising the following steps:
[0012] S1. User classification processing: Obtain user data from the target platform, classify users based on preset classification dimensions, and obtain at least two user levels and the corresponding level weight for each user level.
[0013] S2. Data Preprocessing and Feature Extraction: Collect user behavior data and content data, clean the data, construct a user-content interaction matrix, and extract user features and content features;
[0014] S3. Hierarchical collaborative filtering recommendation calculation:
[0015] S31. Adjust the similarity calculation between users based on user level weights to obtain the similarity overload value;
[0016] S32. Filter the set of similar users to the target user based on the similarity overload value;
[0017] S33. Extract a candidate content set from the set of similar users, and calculate the recommendation priority score of the candidate content by combining the target user level weight and the popularity decay coefficient.
[0018] S4. Cold Start and Diversity Optimization: If the target user is a new user, an initial recommendation list is generated by combining content-based recommendation with a tiered cold start strategy.
[0019] Diversity control is applied to the candidate content set to manage the category distribution and the proportion of long-tail content in the recommendation list;
[0020] S5. Recommendation Result Generation: Sort the candidate content according to the recommendation priority score, and select the top-ranked preset number of content as the final recommendation result to push to the target user.
[0021] As a further aspect of the present invention: the preset grading dimensions include at least two of the following: commercial value dimension (payment level, consumption amount ratio), behavioral activity dimension (average daily active time, interaction frequency), and credit compliance dimension (complaint rate, default record).
[0022] The user data includes user attribute data (registration information, identity authentication status), behavioral data (browsing, clicks, favorites, transaction / application records), and commercial data (payment status, payment level, consumption amount).
[0023] As a further aspect of the present invention: the user levels include core users (high-paying / high-value), ordinary users, and new users, with different levels corresponding to different level weights (core user weight > ordinary user weight > new user weight), and the level weights are dynamically adjustable, with a weight value range of [1.0, 3.0].
[0024] As a further aspect of the present invention: in step S31, the formula for calculating the similarity overload value is:
[0025] Similarity overload value = Base similarity × (Target user level weight × 0.6 + User level weight to be matched × 0.4);
[0026] The basic similarity is calculated using the Pearson correlation coefficient or cosine similarity. This weighting ratio aims to more significantly amplify the influence of the preferences of high-value (target) users, while also taking into account the commonalities of similar user groups.
[0027] As a further aspect of the present invention: in step S33, the formula for calculating the heat attenuation coefficient is as follows:
[0028] Popularity decay coefficient = 1 / (1 + α × cumulative content exposure);
[0029] Here, α is the adjustment coefficient, with a value range of [0.001, 0.01]. By introducing this coefficient, the recommendation weight of high-exposure content is dynamically suppressed to alleviate the "Matthew effect" and promote the exposure of long-tail content.
[0030] As a further aspect of the present invention, the diversity regulation in step S4 includes at least one of the following:
[0031] The proportions of content categories, paid content, and long-tail content in the recommendation list can be adjusted. For example, the proportion of the same category of content in the recommendation list can be limited to a preset upper limit (e.g., 30%), the proportion of paid content for core users can be limited to a preset upper limit (e.g., 40%), and the proportion of long-tail content for ordinary users can be guaranteed to be no less than a preset lower limit (e.g., 20%). The specific proportions mentioned above can be dynamically configured according to the platform's operational goals, content ecosystem strategy, and user group characteristics.
[0032] Secondly, the present invention provides a collaborative filtering recommendation device based on user hierarchy, comprising:
[0033] The user classification module is used to acquire user data, classify users based on preset classification dimensions, and output user levels and corresponding level weights.
[0034] The data processing module is used to collect and clean user behavior data and content data, construct a user-content interaction matrix, and extract user features and content features.
[0035] The collaborative filtering calculation module is used to improve the collaborative filtering algorithm based on user level weights, calculate user similarity overload values, and calculate the recommendation priority score of candidate content based on the set of similar users and the target user level weights.
[0036] The optimization module is used to perform cold start processing for new users and to optimize and control the diversity of the recommendation list;
[0037] The recommendation output module is used to generate and output the final recommendation results based on the recommendation priority scores.
[0038] As a further aspect of the present invention: the user classification module includes a data acquisition unit, a classification calculation unit, and a weight configuration unit; the data acquisition unit collects user attributes, behaviors, and business data; the classification calculation unit calculates classification scores based on classification dimensions and divides user levels; the weight configuration unit assigns dynamically adjustable level weights to different levels.
[0039] The data processing module includes a data cleaning unit, a matrix construction unit, and a feature extraction unit; the data cleaning unit filters out abnormal data and removes duplicates; the matrix construction unit generates a user-content interaction matrix; and the feature extraction unit extracts level tags, interest tags, content category tags, etc.
[0040] As a further aspect of the present invention: the collaborative filtering calculation module includes a similarity calculation unit, a candidate set extraction unit, and a score calculation unit; the similarity calculation unit calculates the basic similarity and combines it with the grade weight to obtain the similarity overload value; the candidate set extraction unit extracts candidate content from the similar user set; the score calculation unit calculates the recommendation priority score by combining the grade weight and the popularity decay coefficient.
[0041] The optimization module includes a cold start unit and a diversity control unit. The cold start unit provides content-based cold start recommendations for new users; the diversity control unit limits the proportion of similar content to ensure the exposure of long-tail content.
[0042] Thirdly, the present invention provides a device for collaborative filtering recommendation based on user hierarchy, comprising a processor and a memory, wherein the memory stores at least one instruction, which is loaded and executed by the processor to implement the above method.
[0043] Compared with the prior art, the advantages of this invention are:
[0044] I. This solution deeply integrates user level weights into similarity calculation and recommendation priority ranking, which significantly amplifies the preferences of high-value users and paying users, thereby obtaining more accurate and high-quality recommendation results that better match their identity and needs, effectively improving the satisfaction, retention rate and commercial conversion efficiency of core users.
[0045] Second, this solution introduces a popularity decay coefficient to dynamically suppress the recommendation weight of highly exposed popular content, and combines it with a mandatory diversity optimization strategy to provide reasonable exposure opportunities for niche and emerging high-quality content, thereby promoting the health and diversity of the platform's content ecosystem.
[0046] Third, this solution is designed for new users with a cold start strategy that combines hierarchical weighting with content-based recommendations. In the absence of historical behavioral data, it can quickly provide relatively reasonable and diverse initial recommendations based on user registration information and platform hierarchical logic, effectively improving the user experience and subsequent conversion potential during the cold start phase.
[0047] Fourth, the key parameters of this solution, such as user classification dimensions, weights of each dimension, grade weights, popularity decay coefficient, number of similar users, number of recommended results, and specific parameters for diversity control, can be dynamically configured and adjusted according to platform characteristics and operational goals, achieving iterative optimization without reconstructing the algorithm framework. Attached Figure Description
[0048] Figure 1 This is a schematic diagram of the system implementation environment according to an embodiment of the present invention;
[0049] Figure 2 This is a schematic diagram of the overall system architecture and data flow according to an embodiment of the present invention;
[0050] Figure 3 A request processing interaction sequence diagram is recommended for this invention;
[0051] Figure 4 This is a schematic diagram of the overall process of the present invention.
[0052] Figure 5 This is a schematic diagram of the structural modules of the device of the present invention. Detailed Implementation
[0053] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0054] Example 1
[0055] (a) Description of the implementation environment
[0056] like Figure 1 As shown, the implementation environment of this embodiment of the invention includes:
[0057] Terminal device 101: including smartphones, desktop computers, tablets, etc., used to display recommendation results and collect user behavior data;
[0058] Recommendation server 102: Deploys the aforementioned user-tiered collaborative filtering recommendation device (such as...) Figure 3 As shown in the figure, it performs core logic such as data processing, hierarchical calculation, and recommendation algorithm operation;
[0059] Database 103 includes a user database (storing user attributes, grading results, and grading weights), a behavior database (storing user-content interaction records), and a content database (storing content data such as job positions, products, and services), supporting real-time data reading, writing, and querying.
[0060] The overall system architecture and data flow of this invention are as follows: Figure 2 As shown, user attributes, behaviors, and business data from multiple sources are input into the user classification and data processing system for processing and analysis, and the processing results are stored in the user database. When a user initiates a content request for goods, positions, or services through a terminal, the request carries the user's search or filtering conditions. The system integrates the user's classification information, feature vectors, and request conditions to retrieve matching content from the content database, and finally returns the recommendation results in a structured format (such as JSON), forming a complete data loop.
[0061] (II) Implementation details of core processes:
[0062] The overall flow of the recommended method of this invention is as follows: Figure 4 As shown, it includes the following steps:
[0063] S1. User classification processing: Terminal device 101 collects user data in real time, classifies users based on preset classification dimensions, and obtains at least two user levels and the corresponding level weight for each user level.
[0064] User data includes user attribute data (registration information, identity authentication status), behavioral data (browsing, clicks, favorites, transaction / application records), and commercial data (payment status, payment level, consumption amount).
[0065] The preset tiering dimensions include: commercial value (paid level, percentage of spending), behavioral activity (daily average active time, interaction frequency), and credit compliance (complaint rate, default record). As a configurable implementation, the weighting of these tiering dimensions can be set as follows: commercial value 50% (paid level 30% + spending 20%), behavioral activity 30% (daily average active time 15% + interaction frequency 15%), and credit compliance 20% (complaint rate negative scoring 10% + default record negative scoring 10%). These dimensions and weightings can be dynamically adjusted based on the platform's business priorities.
[0066] User levels include core users (high-paying / high-value users), regular users, and new users. Different levels correspond to different weights (core user weight > regular user weight > new user weight), and the weight range is [1.0, 3.0], which can be dynamically adjusted.
[0067] Tiered Score Calculation: Tiered Score = Σ (Score of each dimension × Weight of the corresponding dimension). A score ≥ 80 indicates a core user (tier weight 2.0-3.0), a score ≤ 30 < 80 indicates a regular user (tier weight 1.0-1.5), and a score < 30 indicates a new user (tier weight 1.5, exclusive to cold start). Dynamic Update Mechanism: User tiered scores are recalculated every 7 days, and user tiers and weights are updated in real time.
[0068] S2. Data Preprocessing and Feature Extraction: Collect user behavior data (user-content interaction records, including interaction type, interaction duration, and interaction frequency) and content data (such as attribute information of job positions, products, and services). After data cleaning, construct a user-content interaction matrix. The matrix elements are the interaction weights of users to content (interaction type weight × interaction duration coefficient). Extract user features (level tags, interest preference tags) and content features (category tags, popularity tags).
[0069] S3. Hierarchical collaborative filtering recommendation calculation:
[0070] S31. Adjust the similarity calculation between users based on user level weights to obtain the similarity overload value. The formula for calculating the similarity overload value is as follows:
[0071] Similarity overload value = base similarity × (target user level weight × 0.6 + user level weight to be matched × 0.4).
[0072] The basic similarity calculation uses cosine similarity: for users u and v, the basic similarity sim(u,v) = (u·v) / (||u||×||v||), where u and v are row vectors in the user-content interaction matrix;
[0073] Example of similarity overload value calculation: If the target user is a core user (weight 2.0) and the user to be matched is a regular user (weight 1.2), and the basic similarity is 0.8, then the similarity overload value = 0.8 × (2.0 × 0.6 + 1.2 × 0.4) = 0.8 × 1.68 = 1.344.
[0074] S32. Filter the set of similar users to the target user based on the similarity overload value;
[0075] S33. Extract a set of candidate content from the set of similar users, and calculate the recommendation priority score of the candidate content by combining the target user level weight and the popularity decay coefficient.
[0076] The formula for calculating the heat decay coefficient is:
[0077] Popularity decay coefficient = 1 / (1 + α × cumulative exposure of content); where α is a control coefficient, with a value range of [0.001, 0.01]. Example: If α = 0.005 and the cumulative exposure of a certain piece of content is 1000, then the popularity decay coefficient = 1 / (1 + 0.005 × 1000) = 0.667, reducing the recommendation priority of popular content.
[0078] S4, Cold Start and Diversity Optimization:
[0079] If the target users are new users, an initial recommendation list will be generated using a combination of content-based recommendation and a tiered cold start strategy. The cold start strategy for new users is as follows: During registration, user preference tags (such as job types and salary ranges on recruitment platforms) are obtained. High-quality long-tail content (quality score ≥ 0.8 and cumulative exposure < 500) in the same category is filtered from the content database. This is combined with popular high-quality content (quality score ≥ 0.9) to generate the recommendation list, with long-tail content accounting for no less than 30%.
[0080] The diversity of candidate content sets is regulated through category deduplication and content fragmentation algorithms to control the category distribution and the proportion of long-tail content in the recommendation list. In this embodiment, the following exemplary regulation strategy is adopted:
[0081] The proportion of content of the same category in the recommended list is limited to no more than 30%;
[0082] For core users, the proportion of paid content should be limited to no more than 40%.
[0083] For ordinary users, the proportion of long-tail content should be no less than 20%.
[0084] It should be noted that the specific ratios mentioned above are only examples. In actual applications, they can be dynamically configured and optimized based on platform operation goals, content ecosystem health indicators, and user feedback.
[0085] S5. Recommendation Result Generation: Sort candidate content according to recommendation priority score, select the top 20 (e.g., the first 20) as the final recommendation result and push them to the target user.
[0086] Example 2
[0087] This invention provides a user-tiered collaborative filtering recommendation device, deployed on recommendation server 102, whose module structure is as follows: Figure 5 As shown, it includes:
[0088] The user classification module is used to acquire user data, classify users based on preset classification dimensions, and output user levels and corresponding level weights. The user classification module includes a data acquisition unit, a classification calculation unit, and a weight configuration unit. The data acquisition unit collects user attributes, behaviors, and business data; the classification calculation unit calculates classification scores based on the classification dimensions and assigns user levels; the weight configuration unit assigns dynamically adjustable level weights to different levels.
[0089] The data processing module is used to collect and clean user behavior data and content data, construct a user-content interaction matrix, and extract user and content features. The data processing module includes a data cleaning unit, a matrix construction unit, and a feature extraction unit. The data cleaning unit filters out abnormal data and removes duplicates; the matrix construction unit generates the user-content interaction matrix; and the feature extraction unit extracts level tags, interest tags, content category tags, etc.
[0090] The collaborative filtering calculation module improves the collaborative filtering algorithm based on user ranking weights, calculates user similarity overload values, and calculates recommendation priority scores for candidate content based on the similar user set and the target user's ranking weights. The collaborative filtering calculation module includes a similarity calculation unit, a candidate set extraction unit, and a score calculation unit. The similarity calculation unit calculates the basic similarity and combines it with ranking weights to obtain the similarity overload value; the candidate set extraction unit extracts candidate content from the similar user set; and the score calculation unit calculates the recommendation priority score by combining ranking weights and a popularity decay coefficient.
[0091] The optimization module is used to perform cold start processing for new users and to optimize and control the diversity of the recommendation list. The optimization module includes a cold start unit and a diversity control unit. The cold start unit provides content-based cold start recommendations for new users. The diversity control unit limits the proportion of similar content to ensure the exposure of long-tail content.
[0092] The recommendation output module is used to generate and output the final recommendation results based on the recommendation priority scores.
[0093] In summary, the various modules of the device of the present invention work together. Figure 3 The interaction sequence between the terminal, the user-level data processing system, and the database during the recommendation process is further illustrated. The specific workflow is as follows:
[0094] 1. Data Collection and User Classification Stage: Users log in to terminal device 101 (Step 1). The terminal collects their basic attributes and behavioral data and sends it to recommendation server 102. The user classification module retrieves relevant data from the database, performs classification calculations and weight allocation, and stores the user feature vectors (Step 2).
[0095] 2. Recommendation Request Initiation Stage: Users initiate content requests for goods, positions, or services through their terminals, and may include additional filtering conditions (step 3).
[0096] 3. Recommendation Calculation Stage: Recommendation server 102 receives recommendation requests. The data processing module retrieves the user's grading information, grading weights, and preprocessed feature vectors from the database (steps 4-5), and simultaneously obtains all product, job, or service content data (step 6). The collaborative filtering calculation module performs calculations as described in step S3 based on the above data, and the optimization module performs cold start or diversity control (step 7, corresponding to...). Figure 4 process).
[0097] 4. Results Output Stage: The recommendation output module returns the final sorted recommendation list to the terminal device 101 for display.
[0098] The above description is merely a preferred embodiment of the present invention; however, the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and its improved concepts, should be covered within the scope of protection of the present invention.
Claims
1. A collaborative filtering recommendation method based on user hierarchy, characterized in that: Includes the following steps: S1. User classification processing: Obtain user data from the target platform, classify users based on preset classification dimensions, and obtain at least two user levels and the corresponding level weight for each user level. S2. Data Preprocessing and Feature Extraction: Collect user behavior data and content data, clean the data, construct a user-content interaction matrix, and extract user features and content features; S3. Hierarchical collaborative filtering recommendation calculation: S31. Adjust the similarity calculation between users based on user level weights to obtain the similarity overload value; S32. Filter the set of similar users to the target user based on the similarity overload value; S33. Extract a candidate content set from the set of similar users, and calculate the recommendation priority score of the candidate content by combining the target user level weight and the popularity decay coefficient. S4. Cold Start and Diversity Optimization: If the target user is a new user, an initial recommendation list is generated by combining content-based recommendation with a tiered cold start strategy. Diversity control is applied to the candidate content set to manage the category distribution and the proportion of long-tail content in the recommendation list; S5. Recommendation Result Generation: Sort the candidate content according to the recommendation priority score, and select the top-ranked preset number of content as the final recommendation result to push to the target user.
2. The collaborative filtering recommendation method based on user hierarchy according to claim 1, characterized in that: The preset grading dimensions include at least two of the following: business value dimension, behavioral activity dimension, and credit compliance dimension; The user data includes user attribute data, behavioral data, and business data.
3. The collaborative filtering recommendation method based on user hierarchy according to claim 1, characterized in that: The user levels include core users, regular users, and new users. Different user levels correspond to different level weights, and the level weights are dynamically adjustable.
4. The collaborative filtering recommendation method based on user hierarchy according to claim 1, characterized in that: In step S31, the formula for calculating the similarity overload value is: Similarity overload value = Base similarity × (Target user level weight × 0.6 + User level weight to be matched × 0.4); The basic similarity is calculated using the Pearson correlation coefficient or cosine similarity.
5. The collaborative filtering recommendation method based on user hierarchy according to claim 1, characterized in that: In step S33, the formula for calculating the heat attenuation coefficient is: Popularity decay coefficient = 1 / (1 + α × cumulative content exposure); Where α is the control coefficient, and its value ranges from [0.001, 0.01].
6. The collaborative filtering recommendation method based on user hierarchy according to claim 1, characterized in that: The diversity regulation in step S4 includes at least one of the following: Limit the proportion of content of the same category in the recommendation list to no more than a preset limit; For core users, the proportion of paid content is limited to a preset upper limit; To ensure that the proportion of long-tail content for ordinary users is not low, a preset lower limit is set.
7. An apparatus for implementing the user-level-based collaborative filtering recommendation method according to any one of claims 1-6, characterized in that, include: The user classification module is used to acquire user data, classify users based on preset classification dimensions, and output user levels and corresponding level weights. The data processing module is used to collect and clean user behavior data and content data, construct a user-content interaction matrix, and extract user features and content features. The collaborative filtering calculation module is used to improve the collaborative filtering algorithm based on user level weights, calculate user similarity overload values, and calculate the recommendation priority score of candidate content based on the set of similar users and the target user level weights. The optimization module is used to perform cold start processing for new users and to optimize and control the diversity of the recommendation list; The recommendation output module is used to generate and output the final recommendation results based on the recommendation priority scores.
8. The collaborative filtering recommendation device based on user hierarchy according to claim 7, characterized in that: The user classification module includes a data acquisition unit, a classification calculation unit, and a weight configuration unit. The data processing module includes a data cleaning unit, a matrix construction unit, and a feature extraction unit.
9. A collaborative filtering recommendation device based on user hierarchy according to claim 7, characterized in that: The collaborative filtering calculation module includes a similarity calculation unit, a candidate set extraction unit, and a score calculation unit. The optimization module includes a cold start unit and a diversity control unit.
10. A collaborative filtering recommendation device based on user hierarchy according to claim 7, characterized in that: It also includes a processor and a memory, the memory storing at least one instruction that is loaded and executed by the processor to implement the method as described in any one of claims 1 to 6.