Personalized user portrait group discovery method based on multi-view clustering and drift compensation

Abstract

The invention provides a multi-view clustering method combining multi-kernel learning and matrix decomposition. The multi-view clustering method is applied to personalized user portrait group discovery in a recommendation system. For each user data of the initial multiple views, a plurality of kernel matrixes are obtained through kernel function mapping, a common low-dimensional embedding is learned through a matrix decomposition form, a clustering algorithm is implemented on the consensus embedding, and an initial user group discovery result is obtained. Along with time lapse, part of users in the model may have interest offset, or new users are added, for this purpose, the model is endowed with clustering feature (CF) vectors based on an original user group, a small number of feature vectors can represent a large number of users, and then the feature vectors are updated in real time when model drift occurs so as to obtain a user portrait group after drift compensation. The whole model can realize real-time user portrait group discovery, and accurate pushing of a recommendation system is realized.

Description

Technical Field

[0001] The present invention relates to the fields of data mining and machine learning, and in particular to a personalized user portrait group discovery method based on multi-view clustering and drift compensation. Background Art

[0002] Recommendation systems are an important technical support for the modern digital economy and are widely used in e-commerce, social networks, online education, news recommendations and other fields. User portraits are a set of multidimensional features built by recommendation systems for users, which are used to characterize user interests, preferences, behavioral characteristics and demographic attributes. Through user portraits, recommendation systems can divide users into multiple groups, locate their needs, and optimize recommendation effects. Since user behavior may change over time, the established model may drift, and the user's data is no longer a single view. Multi-view information can capture group information more comprehensively and help to accurately locate needs.

[0003] Clustering is one of the most important tasks in data mining, which aims to divide a set of objects into multiple clusters based on the similarities between objects, so that objects within a cluster are similar and objects between clusters are different. Multi-view clustering aims to integrate information from multiple different feature sets or data sources. In practical applications, data is often multidimensional and multi-source. For example, in image recognition tasks, the same image can be described by different features, such as color, texture, and shape. In bioinformatics, the same biological entity can be described in different ways, such as gene expression, protein sequence, and cell function. These multiple descriptions of the same object represent observations from different perspectives and may contain complementary information, allowing for a more comprehensive characterization of the object and thus more accurate clustering. However, in the existing multi-view clustering branches, multi-kernel learning uses a linear combination of base kernels to obtain the optimal kernel, and finally decomposes the optimal kernel into eigenvectors, which has high time complexity, large memory consumption, and poor scalability. Multi-view clustering based on matrix decomposition is another important branch. Matrix decomposition has lower time complexity. Matrix decomposition-based methods generally decompose the data matrix into two parts to find a low-dimensional embedding that can better reflect the data structure. In this process, different constraints are imposed on each part of the matrix. However, the meaning of these constraints and the interpretation of each part of the matrix are still lacking, especially when extended to the multi-view setting, not every constraint is reasonable. Summary of the invention

[0004] In order to solve the above-mentioned defects existing in the prior art, the present invention proposes a real-time clustering algorithm based on multi-view clustering and drift compensation, and uses the algorithm to cluster user groups in the recommendation system. As user interests shift and new users join, the algorithm updates the model based on the drift compensation mechanism, thereby being able to discover personalized user groups in real time.

[0005] According to a first aspect of the present invention, a personalized user portrait group discovery method based on multi-view clustering and drift compensation is provided, comprising:

[0006] Step 1: Collect data from multiple views and construct a multi-view data matrix X = {X1, X2, …, X V}, each matrix X i is the feature matrix of a single view. Using the kernel function φ i Map it to a high-dimensional space under multiple views to obtain the kernel matrix of multiple views

[0007] Where X={X1,X2,…,X V} consists of V views, a single view matrix d i is the feature dimension under the i-th view, and n is the number of samples.

[0008] When selecting kernel functions for different views, you can use different kernel functions for different views, such as the Gaussian kernel function. The kernel matrices of different views are obtained by adjusting σ.

[0009] Step 2: Initial user group division based on multi-view clustering, by alternating updates ω and H to maximize the objective function

[0010] in represents the low-dimensional embedding under a single view, H is the learned consensus embedding, and HH is required T =I, the second term of the objective function forces is close enough to H, ω is the weight vector for the view, and the sum of the components of ω is required to be 1.

[0011] The objective function is derived from First, for each view mapped v (X v ) to perform matrix decomposition, is the basis matrix under v views, k is the desired embedding dimension, generally chosen as the number of clusters, and different basis matrices represent different information of each view to a certain extent. k×n is the consensus embedding, and finally sent to k-means to obtain the clustering result. Further, by assuming that the basis matrix is ​​a linear combination of high-dimensional space data points, the original formula becomes Finally, the formula is expanded, adding αtr(G v H) item, the objective function is transformed into a concise α is the parameter that controls the second term.

[0012] Step 3: Divide the initial user groups and perform k-means clustering on the consensus embedding H to obtain the clustering result C = {C1, C2, ..., C k}, represents k clusters, each cluster corresponds to a group of users with similar interests.

[0013] Step 4: Record the CF (clustering feature) vector of each cluster. For cluster C i , whose CF vector includes N i LS i and SS i . N i is the number of samples in the cluster, LS i is the sum of samples in the cluster, SS i is the sum of squares of samples within the cluster. With these three features, the cluster center can be naturally expressed radius

[0014] Step 5: Update the model for drift. As time goes by, the interests of some users shift or new users join, which makes the original model no longer applicable. This is called drift. In order to maintain the efficiency of the model, the drift phenomenon is no longer re-clustered, but updated on the basis of the original model to adapt to the new data distribution. This is called drift compensation mechanism in the present invention. Drift occurs in the following two situations:

[0015] 1) There is a change in the interest behavior of existing users. Suppose user x i When a behavior change occurs, the new data of V views are spliced ​​together to obtain Calculated by the distance function d(·,·) The distance from the original cluster center like Then user x i If the user has not left the group, then update the CF vector of the cluster: like Then user x i No longer belongs to the current interest group, calculate the new cluster Continue to judge if Userx i Join Cluster C l , and update cluster C l CF vector: N l ←N l +1, If there is still Then x i Create a new cluster for the center.

[0016] 2) A new user joins the system. Suppose user x tNewly added system, calculate its cluster If d(u t ,μ l )<2r l , x t Join Cluster C l In the update cluster C l CF vector: N l ←N l +1,LS l ←LS l +u t , SS l ←SS l +u t 2 ; If d(u t ,μ l )>2r l , then x t Create a new cluster for the center.

[0017] Step 6: Check whether clusters need to be merged. For any two clusters C i and C j , if d(μ l ,μ l )∈[2r i ,2r j ], without loss of generality, assuming r i <r j , then merge the two clusters into C l , and calculate C l The CF vector is: N l ←N i +N j ,LS l ←LS i +LS j , SS l ←SS i +SS j .

[0018] According to a second aspect of the present invention, there is provided a computer device, characterized in that it comprises:

[0019] a memory for storing instructions; and

[0020] A processor is used to call the instructions stored in the memory to execute the personalized user portrait group discovery method based on multi-view clustering and drift compensation of the first aspect.

[0021] According to a third aspect of the present invention, there is provided a computer-readable storage medium, characterized in that instructions are stored therein, and when the instructions are executed by a processor, the personalized user portrait group discovery method based on multi-view clustering and drift compensation of the first aspect is executed.

[0022] Compared with the prior art, the above technical solution conceived by the present invention has at least the following beneficial effects:

[0023] (1) It can process multi-view data and obtain richer information in the field of user group discovery.

[0024] (2) Compared with traditional kernel-based multi-view clustering, the present invention does not involve the learning of the optimal kernel, so there is no eigenvalue decomposition step, which can reduce the computational complexity at the theoretical level.

[0025] (3) In the initial clustering stage of multi-view, the present invention removes the orthogonality constraint for each view and instead adds an orthogonality constraint to the consensus embedding. This is more in line with the assumption that single-view information in multi-view data is inaccurate and can achieve higher clustering accuracy on general multi-view datasets.

[0026] (4) In response to the phenomenon of user interest changes, it is possible to monitor in real time and adjust the existing clustering structure without the need for re-clustering, so feedback is obtained faster.

[0027] (5) In response to the phenomenon of new users joining, the present invention uses CF vectors to represent each user group. CF vectors are additive, so the user grouping results can be quickly updated without visiting individual users.

[0028] (6) In the process of gradually updating the clustering results, it is possible to capture the development trend of user groups and provide more accurate personalized services.

[0029] It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0031] Figure 1 It is a flowchart of a personalized user portrait group discovery method based on multi-view clustering and drift compensation according to an exemplary embodiment. DETAILED DESCRIPTION

[0032] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0033] Taking user data in a movie recommendation system as an example, the personalized user portrait group discovery method based on multi-view clustering and drift compensation of the present invention is used for analysis. Figure 1 As shown, user portrait groups are discovered by performing cluster analysis on user data from multiple views. That is, through cluster analysis, we can know which user portrait group each user belongs to, and then accurately push movies.

[0034] The data contains 1484 user samples, each user has 2 views of behavior data, and the number of initial user groups is given as 10. The multi-view user data is shown in the following table:

[0035] User view:

[0036] User ID age gender Year of registration Number of film readings 1 34 0 2022 122 2 25 0 2020 51 3 19 1 2018 6 … … … … … 1484 32 1 2020 180

[0037] Movie View:

[0038]

[0039] The specific implementation steps of the embodiment are as follows:

[0040] Step 1: Collect data from multiple views and construct a multi-view data matrix X = {X1, X2, …, X V}, each matrix X i is the feature matrix of a single view. Using different kernel functions φ i Map it to a high-dimensional space under multiple views to obtain the kernel matrix of multiple views In this embodiment, V=2.

[0041] Where X={X1,X2,…,X V} consists of V views, a single view matrix d i is the feature dimension under the i-th view, and n is the number of samples. In this embodiment, d1=4, d2=5, n=1684, X1 is the user view, including age, gender, etc.; X2 is the movie view, including the number of clicks on the movie type, and all view data are converted into numerical data.

[0042] When selecting kernel functions for different views, you can use different kernel functions for different views, such as the Gaussian kernel function. The kernel matrices of different views are obtained by adjusting σ. In this embodiment, σ of the two views is 1.

[0043] Step 2: Update by alternation ω and H to maximize the objective function Get the consensus embedding H.

[0044] Step 3: Perform k-means clustering on the consensus embedding H and obtain the clustering result C = {C1, C2, ..., C k}, represents k clusters, each cluster corresponds to a group of users with similar interests.

[0045] Step 4: Record the CF (clustering feature) vector of each cluster. For cluster C i , whose CF vector includes N i LS i and SS i . N i is the number of samples in the cluster, LS i is the sum of samples in the cluster, SS i is the sum of squares of samples in the cluster. At the same time, the cluster center is recorded radius

[0046] Step 5: Check whether model drift occurs. Drift occurs in the following two situations:

[0047] 1) There is a change in the interest behavior of existing users. Suppose user x i When a behavior change occurs, the new data of V views are spliced ​​together to obtain Calculated by the distance function d(·,·) The distance from the original cluster center like Then user x i If the user has not left the group, then update the CF vector of the cluster: like Then user x i No longer belongs to the current interest group, calculate the new cluster Continue to judge if Userx i Join Cluster C l , and update cluster C l CF vector: N l ←N l +1, If there is still Then x i Create a new cluster for the center.

[0048] 2) A new user joins the system. Suppose user x t Newly added system, calculate its cluster If d(u t ,μ l )<2r l , x t Join Cluster C l In the update cluster C l CF vector: N l ←N l +1,LS l ←LS l +u t , SS l ←SS l +u t 2 ; If d(u t ,μ l )>2r l , then x t A new cluster is established for the center. In this embodiment, the distance function d(·,·) adopts the cosine distance function:

[0049] Step 6: Check whether clusters need to be merged. For any two clusters C i and C j , if d(μ l ,μ l )∈[2r i ,2r j ], without loss of generality, assuming r i <r j , then merge the two clusters into C l , and calculate C l The CF vector is: N l ←N i +N j ,LS l ←LS i +LS j , SS l ←SS i +SS j .

[0050] Finally, through the updated user group discovery results, the movies pointed to by the central user of each user group are pushed to each user group, and the present invention can realize real-time push.

[0051] Other embodiments of the present invention will be readily apparent to those skilled in the art after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses or adaptations of the present invention that follow the general principles of the present invention and include common knowledge or customary techniques in the art that are not disclosed in this disclosure. The specification and examples are to be considered exemplary only, and the true scope and spirit of the present invention are indicated by the following claims.

[0052] It should be understood that the present invention is not limited to the exact construction that has been described above and shown in the drawings and that various modifications and changes may be made without departing from the scope thereof. The scope of the present invention is limited only by the appended claims.

Claims

1. A personalized user portrait group discovery method based on multi-view clustering and drift compensation, characterized in that: include: Step 1: Collect data from multiple views and construct a multi-view data matrix X = {X1, X2, …, X V }, each matrix X i is the feature matrix of a single view; using the kernel function φ i Map it to a high-dimensional space under multiple views to obtain the kernel matrix of multiple views Where X={X1,X2,…,X V } consists of V views, a single view matrix d i is the feature dimension under the i-th view, and n is the number of samples; When selecting kernel functions for different views, you can use different kernel functions for different views, such as the Gaussian kernel function. The kernel matrix of different views is obtained by adjusting σ; Step 2: Initial user group division based on multi-view clustering, by alternating updates ω and H to maximize the objective function in represents the low-dimensional embedding under a single view, H is the learned consensus embedding, and HH is required T =I, the second term of the objective function forces is close enough to H, ω is the weight vector for the view, and the sum of the components of ω is required to be 1; The objective function is derived from First, for each view mapped v (X v ) to perform matrix decomposition, is the basis matrix under v views, k is the desired embedding dimension, generally chosen as the number of clusters, and different basis matrices represent different information of each view to a certain extent; H∈R k×n is the consensus embedding, which is finally sent to k-means to obtain the clustering result; further, by assuming that the basis matrix is ​​a linear combination of high-dimensional space data points, the original formula becomes Finally, the formula is expanded, adding αtr(G v H) item, the objective function is transformed into a concise α is the parameter that controls the second term; Step 3: Divide the initial user groups and perform k-means clustering on the consensus embedding H to obtain the clustering result C = {C1, C2, ..., C k }, representing k clusters, each cluster corresponds to a group of users with similar interests. Step 4: Record the CF (clustering feature) vector of each cluster. For cluster C i , whose CF vector includes N i LS i and SS i ; N i is the number of samples in the cluster, LS i is the sum of samples in the cluster, SS i is the sum of squares of samples within the cluster; with these three features, the cluster center can be naturally represented radius Step 5: Update the model for drift phenomenon. As time changes, the interests of some users shift, or new users join, which makes the original model no longer applicable, which is called drift phenomenon. In order to maintain the efficiency of the model, the drift phenomenon is no longer re-clustered, but updated on the basis of the original model to adapt to the new data distribution. This is called drift compensation mechanism in the present invention. Drift occurs in the following two situations: 1) There is a change in the interest behavior of existing users. Suppose user x i When a behavior change occurs, the new data of V views are spliced ​​together to obtain Calculated by the distance function d(·,·) The distance from the original cluster center like Then user x i If the user has not left the group, then update the CF vector of the cluster: like Then user x i No longer belongs to the current interest group, calculate the new cluster Continue to judge if Userx i Join Cluster C l , and update cluster C l CF vector: N l ←N l +1, If there is still Then x i Create a new cluster for the center; 2) A new user joins the system. Suppose user x t Newly added system, calculate its cluster If d(u t ,μ l )<2r l , x t Join Cluster C l In the update cluster C l CF vector: N l ←N l +1, LS l ←LS l +u t , SS l ←SS l +u t 2 ; If d(u t ,μ l )>2r l , then x t Create a new cluster for the center; Step 6: Check whether clusters need to be merged. For any two clusters C i and C j , if d(μ l ,μ l )∈[2r i ,2r j ], without loss of generality, assuming r i <r j , then merge the two clusters into C l , and calculate C l The CF vector is: N l ←N i +N j ,LS l ←LS i +LS j , SS l ←SS i +SS j .

2. The personalized user portrait group discovery method based on multi-view clustering and drift compensation according to claim 1, characterized in that: Also includes: The parameter α in the objective function has a value range of {10 0 ,10 1 ,...,10 9 }.

3. A computer device, characterized in that: include: A memory for storing instructions; as well as A processor, configured to call the instructions stored in the memory to execute the personalized user portrait group discovery method based on multi-view clustering and drift compensation as described in any one of claims 1-2.

4. A computer-readable storage medium, characterized in that: Instructions are stored, and when the instructions are executed by the processor, the personalized user portrait group discovery method based on multi-view clustering and drift compensation as described in any one of claims 1-2 is executed.

Images