An appliance maintenance cold start recommendation method based on an improved clustering algorithm and a Catboost model
By using an improved K-means clustering algorithm and Catboost model, combined with user and community information, the cold start problem of the home repair recommendation system was solved, achieving more accurate recommendation results.
Patent Information
- Application Number
- CN202310071190.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-01-17
- Publication Date
- 2025-12-23
- Estimated Expiration
- 2043-01-17
AI Technical Summary
Home repair recommendation systems suffer from a cold start problem. Existing methods, such as those based on user participation, degrade user experience; those based on demographic features have unsatisfactory recommendation results; those based on associated data struggle to match user information; and those based on social network data raise privacy concerns.
An improved K-means clustering algorithm and Catboost model are used to form user profiles by integrating user information and community information, dividing users into different user groups, and making recommendations using similarity calculation and classification prediction.
It improves the accuracy and precision of recommendations, solves the outlier problem, expands the user dimension, enhances the accuracy of classification prediction, and achieves better recommendation results.
Smart Images
Figure CN116361675B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of home maintenance recommendation system cold start, and particularly relates to a household appliance maintenance cold start recommendation method based on an improved clustering algorithm and a Catboost model. BACKGROUND
[0002] With the in-depth development of the Internet, a large amount of information is spread on the Internet, and a serious information overload occurs. Users are difficult to find valuable information from the massive information flow. Personalized recommendation is a good method to solve the information overload. However, the recommendation system often faces the problem of cold start of the recommendation system. The cold start problem refers to that when a new user or item enters the system, it is difficult for the system to calculate the new user and predict the user's preferred item due to the lack of user and item relationship data.
[0003] The new user cold start problem can be divided into complete cold start and partial cold start. The complete cold start refers to that there is no data in the recommendation system that can be used to build the relationship between the user and the item. The partial cold start refers to that there is a small amount of data in the system, for example, a new user has rated a few items.
[0004] The solution to the complete cold start problem of the new user can be roughly divided into two categories: a method based on auxiliary data and a method based on user participation. The method based on user participation is to conduct a survey on the user, such as asking the new user to select the type of preferred item when the new user enters the system. This type of method is to obtain the user preference or item rating data through explicit user feedback. The method based on auxiliary data is to expand the user's dimension through the user's other information data to obtain the user's preference. Common data includes demographic data, social network data, and context information. Demographic statistics generally divide users into different groups through head statistics characteristics, and makes recommendations according to the preferences of the groups. Social network data generally refers to the data left by users on various social networks, which often reflects the characteristics of the user. According to these characteristics, the preferred item of the user is calculated. Context information generally refers to the time, space, and device information of the user. These characteristics often directly divide the user into different groups, and different groups have different preferred items. Recommendations can be made according to these information.
[0005] The method based on associated data associates LOD (Linked Open Data) with the user to obtain more data of the user, covering knowledge in many fields.
[0006] The home repair recommendation is different from other industries and has certain particularity. For the user of the home appliance repair industry, the home appliance of the user will only need the home appliance repair service when the home appliance has a problem, and various repair services cannot be recommended at the time of recommendation, and the user needs to be calculated to speculate which services the user needs. The user's browsing information is less, the demand is single, and the like, resulting in cold start of the recommendation system.
[0007] Among them, the method based on user participation makes the user do more operations and needs the user to participate. However, the user is often unwilling to leave too much information, and this method reduces the user experience and leads to the loss of the user. The method based on demographic characteristics uses high-heat products, and it is difficult to recommend general-heat or heat products, and the recommendation effect is often not ideal. The method based on associated data needs to successfully match the associated data for each user in the field of home repair recommendation, which is a difficult thing. The method based on social network data has difficulty in obtaining the social information of each user, and some social information may be the privacy of the user, and there are certain legal security problems.
[0008] In summary, the prior art problems are:
[0009] The existing recommendation only needs the home appliance repair service when a problem occurs, and various repair services cannot be recommended at the time of recommendation, and the user needs to be calculated to speculate which services the user needs. The user's browsing information is less, the demand is single, and the like, resulting in cold start of the recommendation system. The method based on user participation makes the user do more operations and needs the user to participate. However, the user is often unwilling to leave too much information, and this method reduces the user experience and leads to the loss of the user. The method based on demographic characteristics uses high-heat products, and it is difficult to recommend general-heat or heat products, and the recommendation effect is often not ideal. The method based on associated data needs to successfully match the associated data for each user in the field of home repair recommendation, which is a difficult thing. The method based on social network data has difficulty in obtaining the social information of each user, and some social information may be the privacy of the user. SUMMARY
[0010] To solve the above technical problems, the present application provides a home appliance repair cold start recommendation method based on an improved clustering algorithm and a Catboost model, comprising:
[0011] S1: Obtain original user information and community information in the city, and fuse the obtained community information and the user information to form a new user portrait;
[0012] S2: According to the order number of service data of the historical home repair history of the new user portrait, the users in the cell are counted, and the users with the order number less than N are divided into user group 1, and the users with the order number greater than N are divided into user group 2;
[0013] S3: For user group 1, an improved K-means clustering algorithm is used for clustering to obtain a similar user group, and the top 1 service with the most preferred product quantity in the similar user group is recommended to the user;
[0014] S4: For user group 2, a Catboost model is used to classify and predict the repair service of the user according to the community information, and the recommended service is obtained.
[0015] The beneficial effects of the present application are:
[0016] 1. The present application solves the cold start problem of home repair recommendation by using a double recommendation module, which is a recommendation based on similarity calculation and a classification model recommendation based on community data. These two methods are used together to accurately recommend users and improve the accuracy of the recommendation.
[0017] 2. The present application uses an improved clustering algorithm to make recommendations for users. A reasonable number of K clustering centers are selected according to the community location, which solves the problem of isolated points. The relationship between data points and clusters is obtained according to the calculation of user similarity, which can better perform clustering. According to the recommendation of each clustering group, the service for users is better recommended.
[0018] 3. The present application uses easily accessible community information as auxiliary information to expand the user dimension. More information can make the classification model more easily learn the relationship between features and classification prediction, and improve the accuracy of classification prediction, so as to accurately recommend products for users. BRIEF DESCRIPTION OF DRAWINGS
[0019] Figure 1 The flowchart of the home appliance repair cold start intelligent recommendation method of the present application is shown in the figure. DETAILED DESCRIPTION
[0020] The technical solutions in the embodiments of the present application will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. 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.
[0021] A home appliance repair cold start recommendation method based on an improved clustering algorithm and a Catboost model, as shown in the figure, comprises: Figure 1
[0022] S1: Obtain original user information and community information in the city, fuse the obtained community information and user information to form a new user portrait;
[0023] S2: According to the order number of the service data of the historical family maintenance history of the new user portrait, the users in the community are counted, the users with order number less than N are divided into user group 1, and the users with order number greater than N are divided into user group 2;
[0024] S3: For user group 1, an improved K-means clustering algorithm is used for clustering to obtain similar user groups, and the top 1 service with the most preferred products in the similar user group is recommended to the user;
[0025] S4: For user group 2, a Catboost model is used to classify and predict the maintenance service of the user according to the community information, and the recommended service is obtained.
[0026] According to the service data of the historical family maintenance history, the community information in the city is obtained, the community information refers to the publicly available information on the Internet, such as community construction age, community level, community id, community greening ratio, community attention ratio and various features; Then match the community id with the user's community id in the historical order, fuse the obtained community information and user information to form a new user portrait.
[0027] In the historical order data, the order number is counted with community category as the classification condition, the order number of a community less than N is divided into user group 1, and the order number greater than N is divided into user group 2. Because when the data volume is too small, it is difficult for a classification model to find the relationship between the features and the results. If a classification model is used, it may produce poor classification results due to insufficient data volume.
[0028] The improved K-means clustering algorithm comprises:
[0029] The community location is selected as the clustering center instead of the initial clustering center randomly selected by the K-means algorithm; In the K-means algorithm, when the iteration number exceeds, the similarity between the user far away from the cluster center and the cluster center user is calculated, and the user of the point is assigned to the cluster with the highest similarity.
[0030] For user group 1, an improved clustering algorithm is used for clustering, comprising:
[0031] S31: Select K community locations as the initial center of clustering of each cluster, and set the iteration number;
[0032] S32: Calculate the distance between the users in the user group 1 and the selected K cluster centers, and through multiple iterations, add the users in the user group 1 along the cluster center with the smallest distance to their cluster group, for the isolated users, calculate the similarity between the isolated users and the users in each cluster, and assign the users of the point to the cluster with the highest similarity;
[0033] S33: Calculate the center of the cluster, and update the center point of the cluster;
[0034] S34: Repeat steps S32-S33 until the samples in the cluster no longer change, and obtain the similar user group.
[0035] Some data objects may be iterated multiple times without being added to a certain cluster, which are single users, and the local center points in the cluster after clustering exceed three times the tie distance, which are called isolated users;
[0036] The local center points in the cluster after clustering exceed three times the tie distance, and the similarity of the users in the cluster group is very low.
[0037] Therefore, some clusters with a relatively close distance to the data object are selected, the similarity between the data object and the cluster object is calculated according to the user similarity, and the data object is added to the cluster with the highest similarity; the similarity calculation includes:
[0038] According to the new portrait of the user, the community information of the user is divided into category type features and numerical type features, the similarity of the category type features between the isolated user and the users in the cluster is calculated first, and then the similarity of the numerical type features between the isolated user and the users in the cluster is calculated, the similarity of the category type features and the similarity of the numerical type features are multiplied to obtain the final similarity between the isolated user and the users in the cluster.
[0039] The similarity of the category type features includes:
[0040]
[0041] Wherein, S f (u1, u2) represents the category type feature similarity between the isolated user u1 and the user u2 in the cluster, f 1i represents the feature of the isolated user u1, f 2i represents the feature of the user u2 in the cluster.
[0042] The similarity of the numerical type features includes:
[0043]
[0044] Wherein, S n (u1, u2) represents the numerical type feature similarity between the isolated user u1 and the user u2 in the cluster, W iweight of the i-th categorical feature, n imax maximum value of the i-th numerical feature, n imin minimum value of the i-th numerical feature, n i the i-th numerical feature.
[0045] calculating the final similarity, including:
[0046] S nf (u1,u2)=S n (u1,u2)*S f (u1,u2)
[0047] where S nf (u1,u2) represents the final similarity between the isolated user u1 and the user u2 in the cluster, S f (u1,u2) represents the category feature similarity between the isolated user u1 and the user u2 in the cluster, S n (u1,u2) represents the numerical feature similarity between the isolated user u1 and the user u2 in the cluster.
[0048] For the current features, there are original features such as time, space, and community features. In general, the user currently has categorical features and numerical features. Numerical features such as community housing prices, green rate, etc. And category features such as month, day, district level, etc. For user group 1, both types of features will be calculated for similarity, and the most similar user group will be recommended. For users in user group 1, first calculate the similarity according to the category feature to get the corresponding temporary similar user group, and then get the final similar user group from the temporary similar user group.
[0049] For user group 2, the user features and community features have been obtained. For a small amount of missing data, use the mode of these data to fill in; then use the home repair services purchased by the user as the model to predict the label; then use the machine learning model gradient boosting tree model Catboost for classification prediction; At the same time, when training the model, simulate the real situation, such as: the first 25 days of a month as the training set, and the last 5 days as the test set. Dividing by time can better simulate the real situation; After that, the model will automatically classify the user, and the label of the classification is the service that the system will recommend.
[0050] Classification prediction using Catboost model, including:
[0051] S41: missing value processing of community information, using the mode of community information to fill in missing values;
[0052] S42: The service data of the user's historical home repair history is taken as a prediction label of the Catboost model, the Catboost model is trained, the users in the user group 2 are classified and predicted by the trained Catboost model, and the recommended service is obtained.
[0053] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.
Claims
1. A cold start recommendation method for home appliance repair based on an improved clustering algorithm and the Catboost model, characterized in that, include: S1: Obtain raw user information and community information of the city where the user is located, and integrate the obtained community information with the user information to form a new user profile; S2: Based on the number of orders in the historical home repair service data of the new user profile, the community users are statistically analyzed. Users with fewer than N orders are divided into user group 1, and users with more than N orders are divided into user group 2. S3: For user group 1, use the improved K-means clustering algorithm to cluster and obtain similar user groups. Recommend the top 1 service with the most preferred products among the similar user groups to the users. For user group 1, an improved clustering algorithm is used for clustering, including: S31: Select K community locations as the initial cluster centers for each cluster, and set the number of iterations; S32: Calculate the distance between users in user group 1 and the selected K cluster centers. Through multiple iterations, add users in user group 1 to their clusters along the cluster center with the smallest distance. For isolated users, calculate the similarity between the isolated user and users in each cluster, and assign the user at that point to the cluster with the highest similarity. Isolated users: users who have not yet joined a cluster after clustering, and users within a cluster whose distance from the center point exceeds three times the average distance. S33: Calculate the center of the cluster and update the center point of the cluster; S34: Repeat steps S32-S33 until the samples in the cluster no longer change, thus obtaining a similar user group; The similarity calculation includes: Based on the new user profile, the user's community information is divided into categorical features and numerical features. First, the similarity of the categorical features between isolated users and users in the cluster is calculated. Then, the similarity of the numerical features between isolated users and users in the cluster is calculated. The similarity of the categorical features and the similarity of the numerical features are multiplied to obtain the final similarity between isolated users and users in the cluster. Calculating the similarity of numerical features includes: Among them, S n (u1, u2) represents the numerical feature similarity between isolated user u1 and user u2 in the cluster, W i This represents the weight of the i-th type of feature. This represents the maximum value of the i-th type of numerical feature. Let n represent the minimum value of the i-th type of numerical feature. i Represents the i-th type of numerical feature; S4: For user group 2, the Catboost model is used to classify and predict the user's repair services based on community information, and the recommended services are obtained.
2. The cold start recommendation method for home appliance repair based on an improved clustering algorithm and Catboost model according to claim 1, characterized in that, The improved K-means clustering algorithm includes: Instead of randomly selecting initial cluster centers, community locations are chosen as cluster centers in the K-means algorithm. For isolated points in the K-means algorithm, when the number of iterations exceeds a certain threshold, users at that point are assigned to the cluster with the highest similarity by calculating the similarity between users far from the cluster center and users at the cluster center.
3. The cold start recommendation method for home appliance repair based on an improved clustering algorithm and Catboost model according to claim 1, characterized in that, Calculating the similarity of categorical features includes: Among them, S f (u1, u2) represents the categorical feature similarity between isolated user u1 and user u2 in the cluster, f 1i f represents the characteristics of an isolated user u1. 2i This represents the characteristics of user u2 in the cluster.
4. The cold start recommendation method for home appliance repair based on an improved clustering algorithm and Catboost model according to claim 1, characterized in that, The final similarity is calculated, including: S nf (u1,u2)=S n (u1,u2)*S f (u1,u2) Among them, S nf (u1, u2) represents the final similarity between isolated user u1 and user u2 in the cluster, S f (u1, u2) represents the categorical feature similarity between isolated user u1 and user u2 in the cluster, S n (u1,u2) represents the numerical feature similarity between isolated user u1 and user u2 in the cluster.
5. The cold start recommendation method for home appliance repair based on an improved clustering algorithm and Catboost model according to claim 1, characterized in that, Using the Catboost model for classification prediction includes: S41: The missing values in the new user profile formed by integrating community information and user information are handled, and the missing values in the new user profile are filled with the mode of the missing values. S42: Use the user's historical home repair service data as the prediction label for the Catboost model, train the Catboost model, and use the trained Catboost model to classify and predict repair services for users in user group 2 to obtain recommended services.