Interactive clustering quality improvement method based on user intention learning

Abstract

The present invention discloses an interactive clustering quality improvement method based on user intention learning, which belongs to the field of human-computer interaction and visual analysis technology, and comprises the following steps: step 1: data collection and storage, and obtaining a data set D; step 2: extracting samples from D to obtain a sample set S; step 3: performing initial clustering on S, the system performs data analysis, and recommends a clustering adjustment scheme in a visual form; step 4: the user adjusts C S , and get C ’ S The system learns the clustering intention reflected in the user adjustment process through a semi-supervised method, and then automatically clusters a larger number of unadjusted samples. The core of the present invention is to use the grouping information of the user-adjusted samples to automatically cluster a larger number of unadjusted samples; the present invention models the association relationship of all samples by constructing a connected implicit semantic relationship graph, so that ordinary users can also use this method, which is also conducive to the cross-disciplinary and cross-data application of this method.

Description

Technical Field

[0001] The present invention belongs to the field of human-computer interaction and visual analysis technology, and in particular relates to an interactive clustering quality improvement method based on user intention learning. Background Art

[0002] Cluster analysis originated from taxonomy. In ancient taxonomy, people mainly relied on experience and professional knowledge to achieve classification, and rarely used mathematical tools for quantitative classification. With the development of human science and technology, the requirements for classification are getting higher and higher, so that sometimes it is difficult to accurately classify based on experience and professional knowledge alone. Therefore, people gradually introduced mathematical tools into taxonomy, forming numerical taxonomy, and then introduced multivariate analysis technology into numerical taxonomy to form cluster analysis.

[0003] Most existing clustering algorithms rely on the initial state setting, the quality of input data, and the user's prior knowledge. If users want to obtain high-quality output results, they need to have a deep understanding of data science and clustering principles. In addition, existing methods are prone to the following problems:

[0004] First, the initial parameters such as the number of clusters, distance measurement, and some random quantity settings will significantly affect the clustering results. However, there is a lack of clear guidance on the selection of such parameter values, and users need to determine them based on experience or repeated trial and error.

[0005] Second, clustering is an unsupervised machine learning method. The corresponding models belong to black box systems. Users without professional background knowledge find it difficult to understand its internal working mechanism, let alone modify it to optimize performance.

[0006] Third, due to the obvious differences in input data characteristics, it is difficult to obtain stable and high-precision results in cross-domain scenarios with the same set of parameter settings. Users need to spend huge manpower and computing power on different applications to find suitable settings. In other words, if existing clustering methods are to achieve ideal results, users need to have solid professional knowledge, which makes it difficult to achieve cross-disciplinary and cross-data applications. Summary of the invention

[0007] The purpose of the present invention is to propose an interactive clustering quality improvement method based on user intent learning in order to solve the problems listed in the background technology.

[0008] In order to achieve the above object, the present invention adopts the following technical solutions:

[0009] The interactive clustering quality improvement method based on user intention learning includes the following steps:

[0010] Step 1: Data collection and storage: Collect data and obtain the data set D = {d1, d2, ..., dn};

[0011] Step 2: Small batch data sampling: extract r samples from the data set D and obtain the sample set S = {s1, s2, ..., s r};

[0012] Step 3: Initial clustering of small samples: Use the clustering method kmeans to perform initial clustering on the sample set S and obtain the preliminary clustering result C S ,The system performs data analysis and recommends cluster adjustment solutions in a visual form;

[0013] Step 4: Interactive clustering adjustment: Users adjust the clustering results C based on their experience or system recommendations. S Adjust to get C ’ S , the system learns the clustering intention reflected in the user's adjustment process through a semi-supervised method, and automatically clusters the unadjusted samples according to the clustering intention; in this process, the system calculates the improvement degree of clustering quality after each step of adjustment according to the following improvement formula:

[0014]

[0015] in, Represents any two clusters C i , C j If improvement is a negative value, the system will remind the user to verify again or cancel the operation;

[0016] Step 5: Clustering intention propagation: construct a fully connected graph containing all samples to establish a connection between the current small sample and the new data sample, transform the clustering problem into a semi-supervised learning problem based on the GNN model, and automatically calibrate the unadjusted samples;

[0017] Step 6: Clustering result evaluation: The distance between two probability distributions is calculated by KL divergence. The closer the probability of samples belonging to each cluster, the less reliable the final result of the model output.

[0018] Preferably, after collecting the data in step 1, the data needs to be cleaned and integrated for preprocessing to obtain a data set D.

[0019] Preferably, the method of extracting r samples from the data set D in step 2 is uniform random sampling or blue noise sampling.

[0020] Preferably, in step 4, the user adjusts the clustering result C S When the samples are displayed on the screen, the positions of the samples are determined by the t-SNE algorithm, and samples with similar features will form a sample cluster.

[0021] Preferably, in step 4, the clustering result C S The adjustments include:

[0022] Create: If the sample does not belong to any current cluster, drag it to a blank space to form a new cluster containing only one sample;

[0023] Move: If the number of other samples with similar features to the current sample belonging to another cluster is greater than the number belonging to the current cluster, drag the current sample from the current cluster to another cluster;

[0024] Delete: If you are unsure about the ownership of the current sample, select the sample and delete it;

[0025] Merge: If the features of two clusters are very similar, drag one of them to make the corresponding display areas intersect and merge;

[0026] Segmentation: If the feature division of a cluster is too coarse, select multiple samples using the lasso tool and drag them away from the current cluster to form a new cluster.

[0027] Preferably, when the user adjusts the clustering result C S When , the system adopts the adjustment scheme from clusters to samples, which includes:

[0028] A. Use the following formula to determine the cluster to be adjusted:

[0029]

[0030] rank(C)=variation(C)+1 / (density(C)+μ)

[0031] Among them, density represents the degree of uniformity of features between samples within the cluster, and w i，j It is the cosine similarity of the feature vectors of samples i and j. The smaller the density value is, the more inconsistent the sample features within the cluster are, and the more the cluster needs to be adjusted. Variation indicates the degree of change between sample features within the cluster. The higher the variation value is, the more dissimilar the sample features are, and the more the cluster needs to be adjusted. The rank value indicates the degree to which the cluster needs to be adjusted. The larger the rank value is, the more the cluster needs to be adjusted.

[0032] B. Determine the samples to be adjusted: According to the clusters determined in A, find the k nearest neighbors for each sample it contains. The k value is the same as the number of clusters. At the same time, calculate the distribution of these k nearest neighbors in each cluster {p1,p2,...,p k}, the system adjusts the recommended samples to the cluster with the largest distribution probability.

[0033] Preferably, the propagation of clustering intention is realized by training a neural network model, which includes two graph convolution layers, each layer uses a ReLU activation function, and finally passes through a softmax function to obtain the probability that each sample belongs to each cluster; the definition of the loss function L of the model uses information of both labeled and unlabeled data, where the proportion of unlabeled data information ɑ(.) increases with the increase in the number of training rounds;

[0034] The calculation formula of the loss function L is as follows:

[0035]

[0036] The calculation formula of ɑ(.) is as follows:

[0037]

[0038] Preferably, in the clustering result evaluation, the cluster to which the new sample belongs is determined by the probability distribution P output by the semi-supervised model. c ={c1,c2,...,c k}, where k represents the number of clusters, c i Indicates the possibility that the current sample belongs to the i-th cluster, 1<=i<=k, and the sample will eventually be assigned to the cluster with the greatest possibility.

[0039] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0040] 1. The core of the present invention is to use the grouping information of samples adjusted by the user to automatically cluster a larger number of unadjusted samples; specifically, it is to establish a relationship between adjusted and unadjusted samples so that the information of the adjusted samples is transmitted only to samples with a strong association with them as much as possible; the present invention models the association relationship of all samples by constructing a connected implicit semantic relationship graph, and samples with similar features are interconnected. The user's intention can be transmitted from the labeled small samples to the unlabeled new samples through the path on the graph, so that the user does not need to have a high level of professional background knowledge to use the method, which is also conducive to the cross-disciplinary and cross-data application of the method; at the same time, the unsupervised clustering problem is converted into a semi-supervised learning problem, and the labeled nodes used for graph neural network training represent small batches of samples manually adjusted by the user. These nodes reflect specific domain knowledge and clustering requirements, and play an important guiding role in the clustering of unlabeled data.

[0041] 2. In the present invention, interactive clustering adjustment is adopted, that is, the data and its clustering results are mapped to the visual space through methods such as scatter plots, force-directed layout and t-SNE, which is convenient for users to understand and perceive and improve the efficiency of outlier identification; combined with domain knowledge or expert experience, users can use devices such as mouse and keyboard to interact with the visual elements corresponding to the data to achieve "what you see is what you get" clustering adjustment, which is much more efficient than directly operating on abstract data.

[0042] 3. In the present invention, the clustering quality is analyzed by multi-view linkage. During the adjustment process, the user may perform some unreasonable operations due to negligence or forgetfulness, which will greatly reduce the input quality of the GNN clustering model. Therefore, this method analyzes the statistical relationship between the data before and after each step of adjustment. The results of the analysis are displayed in multiple linkage views of the visual interface, which can not only recommend adjustment plans to users, but also quantify the impact of adjustment operations on clustering quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 A flowchart of the interactive clustering quality improvement method based on user intent learning proposed by the present invention;

[0044] Figure 2 A neural network model diagram of the interactive clustering quality improvement method based on user intent learning proposed by the present invention;

[0045] Figure 3 A diagram showing changes in clustering quality improvement evaluation indicators corresponding to four interactive clustering adjustment operations of the user in the interactive clustering quality improvement method based on user intention learning proposed by the present invention and the system;

[0046] Figure 4 A comparison chart of clustering results of the interactive clustering quality improvement method based on user intent learning proposed in the present invention;

[0047] Figure 5 A diagram showing the types of adjustment operations of the interactive clustering quality improvement method based on user intent learning proposed in the present invention. DETAILED DESCRIPTION

[0048] The present invention is described below in conjunction with specific embodiments.

[0049] like Figure 1 As shown, the interactive clustering quality improvement method based on user intention learning in this embodiment includes the following steps:

[0050] Step 1: Data collection and storage: This method can be flexibly applied to clustering tasks of various types of data such as images and texts. After the collected data is cleaned, integrated and other preprocessing operations, the data set D = {d1, d2, ..., d n}. Use appropriate feature extraction methods for different types of data. For example, extract the feature vector of the text through the word2vec model to obtain the corresponding data feature vector set V = {v1, v2, ..., v n};

[0051] Step 2: Small batch data sampling: Use uniform random sampling or blue noise sampling to extract r samples S = {s1, s2, ..., s r The purpose of sampling is to reduce labor costs. Since humans have limited ability in searching and comparing tasks, reaction time and error rate will increase with the increase of sample size. Therefore, it is necessary to control the number of samples for manual inspection. Generally, the value of r is between 20-30, which can be obtained through repeated user experiments.

[0052] Step 3: Initial clustering of small samples: For the small samples collected, simple and efficient traditional clustering methods such as kmeans are used to obtain the initial clustering result C of S S This process does not require detailed parameter adjustment and allows cluster assignment errors to exist. Its main function is to reduce the difficulty of manual adjustment to a certain extent. At the same time, the system performs data analysis and recommends cluster adjustment plans in a visual form. If the user has relevant expert experience and can provide more reasonable parameter values, the number of cluster adjustment steps will be greatly reduced;

[0053] Step 4: Interactive Cluster Adjustment: C S The distribution of samples in the graph is displayed on the main view of the front-end interface. Overall, the position of sample points in the screen space is determined by the t-SNE algorithm, but this will make the boundaries between different clusters unclear. Therefore, each cluster is regarded as a single node, and they are connected two by two to form a complete graph structure, and then the center point position of each cluster is calculated by the force-directed layout algorithm. The size of the space occupied by a cluster is proportional to the number of samples it contains, and the position of samples within the cluster is still determined by the t-SNE algorithm.

[0054] The user makes adjustments based on experience or system recommendations to obtain C ’ S ,like Figure 5 As shown, adjustment operations include the following types:

[0055] Create: If the sample does not belong to any current cluster, drag it to a blank space to form a new cluster containing only one sample;

[0056] Move: If most samples with similar features to the current sample belong to another cluster, drag the sample from the current cluster to the target cluster;

[0057] Delete: If you are unsure about the ownership of the current sample, you can select the sample and delete it;

[0058] Merge: If the features of two clusters are very similar, you can drag one of them to make the corresponding display areas intersect and merge;

[0059] Segmentation: If the feature division of a cluster is too rough, you can use the lasso tool to select multiple samples and drag them away from the current cluster to form a new cluster.

[0060] In order to improve the efficiency of user adjustment and avoid duplication, omission and conflict, the system recommends the adjustment plan of "from clusters to samples". Specifically, the following indicators are used to measure clusters and samples:

[0061] A. Determine the cluster to be adjusted: density reflects the degree of semantic unity between samples in the cluster, where w i，j It is the cosine similarity of the feature vectors of samples i and j. The smaller the density value, the less semantically consistent the samples in the cluster are, and the more the cluster needs to be adjusted. Variation represents the degree of change between sample features in the cluster. The higher the variation, the less similar the sample features are, and the more the cluster needs to be adjusted. These two indicators are combined to calculate the rank of each cluster (μ is a minimum value to prevent the denominator from being 0, and can be set to 0.001 in practical applications), and sorted in descending order. The cluster with the largest rank is recommended as the one that needs the most adjustment, where:

[0062]

[0063] rank(C)=variation(C)+1 / (density(C)+μ).

[0064] B. Determine the samples to be adjusted: According to the clusters determined in A, find the k nearest neighbors for each sample it contains. The k value is the same as the number of clusters. At the same time, calculate the distribution of these k nearest neighbors in each cluster {p1,p2,...,p k}, the system adjusts the recommended samples to the cluster with the largest distribution probability.

[0065] During this process, there is a certain probability that the user will reject the recommended solution and make adjustments based only on their own domain knowledge. In order to prevent conflicts between the user's operation and the distribution characteristics of the data itself, the system calculates the improvement degree of clustering quality after each step of adjustment according to the following improvement formula:

[0066]

[0067] in, Represents any two clusters C i , C j If improvement is a negative value, it means that the overall clustering quality has declined, and the system will remind the user to verify again or cancel the operation.

[0068] like Figure 3 The records show the user's four interactive clustering adjustment operations and the changes in the system's corresponding clustering quality improvement evaluation indicators.

[0069] Step 5: Clustering intention propagation, which is divided into two steps:

[0070] S1. Implicit relationship graph construction. In order to establish a connection between the current small sample and the new data sample, a fully connected graph containing all samples is constructed. The weight of the edge in the graph is equal to the semantic similarity between the two end nodes. When the number of nodes in the graph is n, the undirected fully connected graph has n(n-1) / 2 edges. When n is large, the computational complexity is high. Therefore, it can be based on the similarity threshold θ s Filter out unimportant edges. s The value is determined based on the statistical distribution of all edge weights, making the number of edges as small as possible while satisfying the connectivity of the relationship graph G;

[0071] S2. Based on G constructed in the previous step, the clustering task of all new samples is transformed into a semi-supervised learning process, and the pseudo labels of small samples are used to identify different categories. The propagation of clustering intention is achieved through the training of the neural network model. The structure of the model is as follows: Figure 2 As shown in the figure, it contains two layers of graph convolutional layers, each layer uses the ReLU activation function, and finally passes through the softmax function to obtain the probability of each node belonging to each category. The definition of the model's loss function L uses both labeled and unlabeled data information, where the proportion of unlabeled data information ɑ(.) increases with the number of training rounds;

[0072] Among them, the calculation formula of the loss function L is as follows:

[0073]

[0074] The calculation formula of ɑ(.) is as follows:

[0075]

[0076] Step 6: Clustering result evaluation: The probability distribution P of the cluster to which the new sample belongs is output by the semi-supervised model c ={c1,c2,...,c k}, where k represents the number of clusters, c iIndicates the possibility that the current sample belongs to the i-th cluster, 1<=i<=k, and the sample will eventually be assigned to the cluster with the greatest possibility. In order to evaluate the reliability of the clustering results, P c With the average probability distribution P u For comparison, the distance between two probability distributions is calculated by Kullback–Leibler (KL) divergence. In theory, the closer the probability of samples belonging to each cluster, the less reliable the final result of the model output. Therefore, the larger the KL distance, the more reliable the model's prediction of the sample cluster.

[0077] like Figure 4 The comparison chart of clustering results is recorded. In the chart, (a) represents the benchmark data, and different color depths represent samples belonging to different clusters; (b) is the result obtained by the method proposed in this patent; (c) is the result obtained by the kmeans method. From the area indicated by the dotted box, it can be found that the method proposed in this patent can significantly improve the clustering quality of the traditional method.

Claims

1. An interactive clustering quality improvement method based on user intention learning, characterized in that: The following steps are involved: Step 1: Data collection and storage: Collect data and obtain the data set D = {d1, d2, ..., d n }; Step 2: Small batch data sampling: extract r samples from the data set D and obtain the sample set S = {s1, s2, ..., s r }; Step 3: Initial clustering of small samples: Use the clustering method kmeans to perform initial clustering on the sample set S and obtain the preliminary clustering result C S ,The system performs data analysis and recommends cluster adjustment solutions in a visual form; Step 4: Interactive clustering adjustment: Users adjust the clustering results C based on their experience or system recommendations. S Adjust to get C ’ S , the system learns the clustering intention reflected in the user's adjustment process through a semi-supervised method, and automatically clusters the unadjusted samples according to the clustering intention; in this process, the system calculates the improvement degree of clustering quality after each step of adjustment according to the following improvement formula: in, Represents any two clusters C i , C j If improvement is a negative value, the system will remind the user to verify again or cancel the operation; Step 5: Clustering intention propagation: construct a fully connected graph containing all samples to establish a connection between the current small sample and the new data sample, transform the clustering problem into a semi-supervised learning problem based on the GNN model, and automatically calibrate the unadjusted samples; Step 6: Clustering result evaluation: The distance between two probability distributions is calculated by KL divergence. The closer the probability of samples belonging to each cluster, the less reliable the final result of the model output.

2. The interactive clustering quality improvement method based on user intention learning according to claim 1 is characterized in that: After collecting the data in step 1, the data needs to be cleaned, integrated and preprocessed to obtain data set D.

3. The interactive clustering quality improvement method based on user intention learning according to claim 1 is characterized in that: In step 2, the method of extracting r samples from the data set D is uniform random sampling or blue noise sampling.

4. The interactive clustering quality improvement method based on user intention learning according to claim 1 is characterized in that: In step 4, the user adjusts the clustering result C S When the samples are displayed on the screen, the positions of the samples are determined by the t-SNE algorithm, and samples with similar features will form a sample cluster.

5. The interactive clustering quality improvement method based on user intention learning according to claim 4 is characterized in that: In step 4, the clustering result C S The adjustments include: Create: If the sample does not belong to any current cluster, drag it to a blank space to form a new cluster containing only one sample; Move: If the number of other samples with similar features to the current sample belonging to another cluster is greater than the number belonging to the current cluster, drag the current sample from the current cluster to another cluster; Delete: If you are unsure about the ownership of the current sample, select the sample and delete it; Merge: If the features of two clusters are very similar, drag one of them to make the corresponding display areas intersect and merge; Segmentation: If the feature division of a cluster is too coarse, select multiple samples using the lasso tool and drag them away from the current cluster to form a new cluster.

6. The interactive clustering quality improvement method based on user intention learning according to claim 5 is characterized in that: When the user adjusts the clustering result C S When , the system adopts the adjustment scheme from clusters to samples, which includes: A. Use the following formula to determine the cluster to be adjusted: rank(C)=variation(C)+1 / (density(C)+μ) Among them, density represents the degree of uniformity of features between samples within the cluster, and w i，j It is the cosine similarity of the feature vectors of samples i and j. The smaller the density value is, the more inconsistent the sample features within the cluster are, and the more the cluster needs to be adjusted. Variation indicates the degree of change between sample features within the cluster. The higher the variation value is, the more dissimilar the sample features are, and the more the cluster needs to be adjusted. The rank value indicates the degree to which the cluster needs to be adjusted. The larger the rank value is, the more the cluster needs to be adjusted. B. Determine the samples to be adjusted: According to the clusters determined in A, find the k nearest neighbors for each sample it contains. The k value is the same as the number of clusters. At the same time, calculate the distribution of these k nearest neighbors in each cluster {p1,p2,...,p k }, the system adjusts the recommended samples to the cluster with the largest distribution probability.

7. The interactive clustering quality improvement method based on user intention learning according to claim 1 is characterized in that: The propagation of clustering intention is realized through the training of the neural network model. The model contains two layers of graph convolutional layers. Each layer uses the ReLU activation function. Finally, the softmax function is used to obtain the probability that each sample belongs to each cluster. The definition of the model's loss function L uses the information of both labeled and unlabeled data. The proportion of unlabeled data information ɑ(.) increases with the increase of the number of training rounds. The calculation formula of the loss function L is as follows: The calculation formula of ɑ(.) is as follows:

8. The interactive clustering quality improvement method based on user intention learning according to claim 1, characterized in that: In the clustering result evaluation, the cluster to which the new sample belongs is determined by the probability distribution P output by the semi-supervised model. c ={c1,c2,...,c k }, where k represents the number of clusters, c i Indicates the possibility that the current sample belongs to the i-th cluster, 1<=i<=k, and the sample will eventually be assigned to the cluster with the greatest possibility.

Images