A cross-scale clustering method to meet the needs of personalized recommendation

Abstract

The present invention belongs to the field of personalized recommendation technology, and discloses a cross-scale clustering method that meets personalized recommendation needs. The method comprehensively considers the scale characteristics of the data itself and the personalized needs of users. First, through single-link hierarchical clustering based on multi-scale factors, the potential distribution law of data attribute characteristics is excavated, the scale division of each attribute is realized, and a multi-scale information table is constructed; then, the transformation of the data attribute scale is combined to reflect the user's attention to different attributes, and on the basis of non-uniform grid division, cross-scale clustering combining micro and macro is realized, and a personalized recommendation scheme that is closer to user needs is obtained based on the clustering results. The present invention not only considers the impact of the scale characteristics of the data itself on personalized recommendation, but also takes user needs as one of the factors of personalized recommendation. It has a strict theoretical basis, strong versatility, and is easy to use flexibly.

Description

Technical Field

[0001] The present invention belongs to the field of personalized recommendation technology, and specifically relates to a cross-scale clustering method that meets the needs of personalized recommendation. Background Art

[0002] In the era of big data, accurate personalized recommendations are of great significance for meeting the personalized needs of users. At present, with the diversified development of the recommendation field, personalized recommendation methods emerge in an endless stream, mainly including: collaborative filtering-based recommendation, content-based recommendation, hybrid recommendation, knowledge-based recommendation, etc. Among them, the recommendation method based on collaborative filtering finds other users with similar interests to the current user and recommends the content that similar users like to the current user; content-based recommendation finds similar items and recommends them to users based on the user's ratings of different items; hybrid recommendation combines multiple recommendation algorithms (such as: collaborative filtering combined with content-based recommendation) to improve the accuracy and diversity of recommendations; and the knowledge-based recommendation method recommends results that meet the user's own needs based on user needs and the use of relevant domain knowledge. Among these existing recommendation methods, the main one is based on correlation analysis, that is, to infer the content that the user may be interested in based on the user's historical behavior data. However, correlation is not equivalent to causality. Combining causality and providing recommendations that better meet the real needs of users is the development trend of explainable personalized recommendation.

[0003] As actual data becomes more diverse and the scale of data increases, the same set of data can have multiple values ​​on each attribute at different measurement scales, that is, the data can be presented at different scale levels, and there is a transformation of information granules from fine to coarse in the scale space. Considering the different importance of different attributes of the same data under different user needs, the scales of key attributes and non-key attributes can be different as needed. When making recommendations, if a single macro scale is used to analyze different users, the details of the key attributes cannot be captured, and the accuracy of the final recommendation results decreases; on the contrary, if a single micro scale is used for analysis, non-essential attributes are over-refined, the complexity of mining information is high, and unnecessary resources are wasted. Therefore, the scale of data attributes is one of the important factors affecting recommendations. If the transformation of the scales of various data attributes can be combined in the recommendation, it can not only reflect causality, but also improve the interpretability of personalized on-demand recommendations. However, none of the existing recommendation technologies has effectively considered the impact of the scale characteristics of the data itself on personalized recommendations.

[0004] As an important data mining technology, cluster analysis divides unlabeled data into several non-overlapping clusters in an unsupervised learning manner, so that the difference between data in different classes is as large as possible and the difference within data in different classes is as small as possible. Cluster analysis technology has gradually become a recommendation method. For example, patent application CN102254028A discloses a personalized commodity recommendation method and system integrating attributes and structural similarity, which establishes a correlation information network diagram based on the purchase relationship between customers and commodities, and uses clustering technology to improve the efficiency of generating the nearest neighbor in the recommendation process; however, the existing clustering at a single scale often fails to meet different practical needs and cannot provide analysis at different scales in a reasonable way. With the research of multi-scale theory in different disciplines, some scholars have proposed some valuable multi-scale clustering ideas and methods. For example, patent application CN114037175A proposes a large-scale bus network hierarchical optimization method based on multi-scale clustering. However, when performing multi-scale clustering, the existing methods usually focus on the information at different scales of the same attribute, and do not pay attention to the synergy between different scales of different attributes. Considering that users attach different importance to different attributes under different needs, cross-scale clustering combining micro and macro will be closer to the expectations of different individual needs. However, there is no cross-scale clustering method in the prior art that meets the needs of personalized recommendation. Summary of the invention

[0005] To solve the above technical problems, the present invention provides a cross-scale clustering method that meets the needs of personalized recommendation. It takes into account the different degrees of importance that research subjects attach to each attribute under different needs, comprehensively describes the scale division of each attribute of the data at different levels or granularities, and uses cross-scale clustering that combines micro and macro, considering the impact of the scale characteristics of the data itself on personalized recommendation, to obtain an explainable personalized recommendation solution that meets the needs of different individuals.

[0006] The cross-scale clustering method for meeting personalized recommendation requirements of the present invention comprises the following steps:

[0007] Step 1: Build industry data samples The property set , , for any attribute , extract all the values ​​it contains, recorded as , ;in, Indicates the number of data samples; Indicates the data dimension, that is, the number of attributes; Indicates attribute value; Representation attributes The number of values ​​of ;

[0008] Step 2: Construct attributes through the similarity difference of neighboring elements Multi-scale factor set , ;in Indicates the number of multi-scale factors; based on the multi-scale factors, each attribute set corresponding to the data is subjected to single-link hierarchical clustering, and the result is converted into a multi-scale information table corresponding to the industry field data;

[0009] Step 3: Based on user needs and multi-scale information tables, select cross-scale combinations of attributes of industry data, calculate the grid step size, and perform grid division according to the values ​​of each attribute at different scales. The division method changes as needed to form a non-uniform grid structure with different scale combinations.

[0010] Step 4: Map the data into non-uniform grid cells, use cross-scale clustering methods to form grid clusters, and mine the information of each grid cluster to form personalized recommendation solutions for users.

[0011] Furthermore, step 2 is specifically as follows:

[0012] Step 2-1: For the attribute value list , use formula (1) to calculate the similarity difference of each neighbor element , the similarity difference of the nearest neighbor elements after deduplication Considered as a multi-scale factor , and finally obtain the multi-scale factor set ;

[0013] (1)

[0014] Step 2-2, traverse the multi-scale factor set, use formula (2) to calculate the similarity of adjacent attribute values ​​according to the single-link hierarchical clustering rule, and compare the result with the current multi-scale factor, and merge the attribute values ​​that are smaller than the current multi-scale factor at one time:

[0015] (2)

[0016] in, Represents different clusters, Respectively represent the clusters data points, Represents a cluster Repeat step 2-2 until the traversal is completed and the silhouette coefficient of the attribute value cluster under different multi-scale factors is calculated. If it is greater than the given silhouette coefficient value , then output the result, and finally get the multi-level cluster result ,in, Representation attributes In the Layer Clusters of classes;

[0017] Step 2-3: For each level cluster result , analyze each cluster The interval range and meaning of the attribute value in the middle, giving each interval a unique identifier , Indicates the number of values ​​under the current scale; maps the attribute value to the corresponding interval identifier, and uses the interval identifier as the value of the current data object under this scale to obtain the attribute The corresponding multi-scale information is obtained; similarly, the multi-scale information of all other attributes is calculated; finally, the multi-scale information of all attributes is integrated to form a multi-scale information table of the original data in the industry field.

[0018] Furthermore, step 3 is specifically as follows:

[0019] Step 3-1: Select attribute scale combination according to user needs ,in is the number of attributes, For attributes Selected scale; for any attribute , find the triples corresponding to different identifiers at the current scale in the multi-scale information table , where any two adjacent triplets and have , ; and They refer to the minimum value, maximum value and number of data in the two adjacent identification intervals. is the number of values ​​under the current scale;

[0020] Step 3-2: If the boundary values ​​between different identifiers are discontinuous due to the characteristics of the cluster, the interval density is introduced to divide the discontinuous boundaries. According to formulas (3) and (4), the adjacent interval triples are calculated. and The interval density and transition point :

[0021] (3)

[0022] (4);

[0023] Step 3-3: Calculate the attributes according to formula (5) In scale The corresponding non-uniform grid step size is :

[0024] (5)

[0025] Step 3-4: According to formula (6), get the scale combination Non-uniform grid space under :

[0026] (6).

[0027] Furthermore, step 4 is specifically as follows:

[0028] Step 4-1: Map data samples to non-uniform grid space In the,dataset,, the data will be presented in different grids under different,scale combinations of different user requirements;

[0029] Step 4-2: Calculate the non-uniform grid space according to formulas (7) and (8) Each grid Density and the average density of the entire grid space :

[0030] (7)

[0031] (8)

[0032] in, and Respectively represent the grid The number and size of the internal data. is the number of grids;

[0033] Step 4-3, compare the density of each grid with the average density. If it is greater than the average density, it is a dense grid, otherwise, it is a sparse grid;

[0034] Step 4-4: According to the density connectivity principle, dense grids are merged to form grid clusters;

[0035] Step 4-5: According to the cluster core points in each grid cluster, the features of the core points of each cluster are extracted to form a personalized recommendation plan.

[0036] The beneficial effects of the present invention are as follows: the method of the present invention not only considers the influence of the scale characteristics of the data itself on personalized recommendation, but also takes user demand as one of the factors of personalized recommendation. On the one hand, starting from the attributes of the data in the industry field, the potential distribution law of each attribute feature of the data is excavated in an unsupervised learning manner, and each attribute is multi-scaled by single-link hierarchical clustering based on multi-scale factors, and a multi-scale information table reflecting the different levels or granularity patterns of each attribute of the original data of the industry is constructed, so as to comprehensively characterize the scale division of each attribute of the data at different levels or granularity; on the other hand, considering the different emphasis of the research object on each attribute under different needs, the transformation of the scale of each attribute of the data is combined when recommending, and the data dimension space is gridded with reference to the multi-scale information table, and the transformation of the attribute scale is controlled by the grid step size, and the cross-scale clustering combining micro and macro is realized by non-uniform grid division, which effectively considers the influence of the scale characteristics of the data itself on personalized recommendation, obtains clustering results that are closer to the needs of different individuals, and improves the interpretability of personalized recommendation. The present invention has a strict theoretical basis, strong versatility, easy to understand and flexible application. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 is a flow chart of the method of the present invention;

[0038] Figure 2 This is a workflow diagram of personalized recommendation based on cross-scale clustering in an embodiment of the present invention. DETAILED DESCRIPTION

[0039] In order to make the contents of the present invention more clearly understood, the present invention is further described in detail below based on specific embodiments in conjunction with the accompanying drawings.

[0040] like Figure 1 and Figure 2 As shown, a cross-scale clustering method for meeting personalized recommendation needs described in the present invention includes the following steps:

[0041] Step 1: Build industry data samples The property set , , extract attributes All the values ​​contained are recorded as , ;in Indicates the number of data; Indicates the attribute dimension, that is, the number of attributes; Indicates attribute values, Indicates the number of values ​​under the current scale;

[0042] Step 2: Construct attributes through the similarity difference of neighboring elements Multi-scale factor set , ;in, Indicates the number of multi-scale factors. Based on the multi-scale factors, each attribute set corresponding to the data is subjected to single-link hierarchical clustering, and the result is converted into a multi-scale information table corresponding to the original data.

[0043] Step 3: Based on user needs and the multi-scale information table, select cross-scale combinations of data attributes, calculate the grid step size, and perform grid division according to the values ​​of each attribute at different scales. The division method changes as needed to form a non-uniform grid structure with different scale combinations.

[0044] Step 4: Map the data into non-uniform grid cells, use cross-scale clustering methods to form grid clusters, and mine the information of each grid cluster to form a personalized recommendation solution that better meets user needs.

[0045] The present invention is further described below by way of embodiments in conjunction with the accompanying drawings.

[0046] Implementation Cases:

[0047] Taking the personalized recommendation of healthy food as an implementation case, the application and implementation of the cross-scale clustering method that meets the needs of personalized recommendation is explained. The nutrients in food are closely related to human health, including: fat content that affects blood lipids; carbon water content that affects blood sugar; sodium salt content that affects blood pressure, etc. Different users attach different importance to different nutrients in food. For example, people with high blood lipids need to control the fat content in food; people with high blood lipids and a family history of high blood sugar pay attention to both the fat content and carbon water content of food.

[0048] First, we collected 30,000 sets of nutritional data for various foods. Table 1 shows the information description of the food nutritional data set. A complete data set includes fat content, carbon water content, and sodium content.

[0049] Table 1 Description of food nutrition data set

[0050]

[0051] (1) Construct a multi-scale information table of food nutritional composition data;

[0052] First, according to step 2 of the present invention, the silhouette coefficient is set to 0.5, and the attributes of the data are divided into multiple scales to construct a corresponding multi-scale information table, as shown in Table 2. Among them, each identifier represents: extremely low content , lower content , slightly lower content , Slightly low content , slightly high content , slightly higher content , higher content , extremely high content , No. The second scale identification corresponding to the attribute in the second layer is interpreted as: basic standard , General Standards , Normal Standard , Abnormal Standard , No. The third scale corresponding to the attribute in the third layer is interpreted as: health , non-health .

[0053] Table 2 Multi-scale information table of food nutritional composition data

[0054]

[0055] (2) Combine user needs and multi-scale information tables to perform cross-scale clustering;

[0056] Assume that there are four types of users with different needs: 1. Patients with three highs; 2. Healthy people; 3. Patients with high blood lipids and critical blood sugar; 4. Patients with high blood lipids. Considering the different needs of the four groups of people, when cross-scale clustering for personalized recommendation, the scale combinations selected for the four types of users are , , and , using different scale combinations to reflect the different levels of importance that different users place on each attribute. Small scales show that details are important, while large scales show the opposite.

[0057] According to step 3-1 to step 3-4, the grid step size of each attribute is calculated to form a non-uniform grid structure.

[0058] According to step 4-1 to step 4-4, the data is mapped to the grid space, and the cross-scale clustering method is used to obtain the final clusters. The specific results are shown in Tables 3 to 6.

[0059] Table 3 Cross-scale clustering results under scale combination

[0060]

[0061] Table 4 Clustering results under scale combination

[0062]

[0063] Table 5 Clustering results under scale combination

[0064]

[0065] Table 6 Clustering results under scale combination

[0066]

[0067] Tables 3 and 4 show the clustering results at a single microscopic and a single macroscopic scale. The clustering under the scale is for user group type 1, and a total of 15 clusters are obtained. The data characteristics between the clusters are slightly different and the information is rich, which meets the precise and detailed needs of the three high people. The subscale clustering is for user group type 2. After clustering, only 4 clusters are obtained. The data of each cluster have no obvious characteristics in most attribute dimensions. Food is only roughly divided into four types: healthy carbohydrates, unhealthy lipids, healthy sodium, and uncertain, which meets the rough needs of healthy people.

[0068] Tables 5 and 6 show the cross-scale clustering results under the combination of macro and micro scales. The number of clusters obtained by clustering is a compromise. Each cluster takes into account both the feature differences and the similarity of the data within the cluster. Cross-scale clustering under scale combination shows that after clustering, the characteristics of the data in each cluster on attribute 1 are obvious, the characteristics of attribute 2 are relatively obvious, and the characteristics of attribute 3 are very random, which reflects the importance and relative attention of user group type 3 to attribute characteristics 1 and 2. Cross-scale clustering under scale combination shows that after clustering, each cluster has obvious characteristics only in attribute 1 dimension, and attributes 2 and 3 are very random, which reflects the importance of user group type 4 on attribute 3, that is, the concern of hyperlipidemia patients on the fat content of food.

[0069] (2) Formation of personalized recommendation schemes for user group type 3;

[0070] According to steps 4-5, a personalized recommendation plan for user group type 3 (hyperlipidemia; blood sugar prevention; other no special requirements) is formed. Based on the clustering results in Table 5 above, since the cluster center can fully reflect the typical characteristics of each cluster, the central sample point of each cluster is used as the representative food, and the results shown in Table 7 can be obtained.

[0071] Table 7 Cross-scale clustering results and representative foods for user group type 3

[0072]

[0073] Further analysis of the clustering results in Table 7 shows that when the actual demand of this type of user is to control the fat content in the first scale And below range, carbon water content in When the range is below , among the 10 clusters obtained by clustering, since the samples in clusters 1, 2, and 3 meet the user's needs in three-dimensional attributes, cabbage, crucian carp, and kale can be used as strong recommended foods; the samples in clusters 4, 5, 7, and 8 do not meet the rules because of their attribute 1 or attribute 2, and roasted gluten, snow pear, and pecan are not recommended foods; although the samples in clusters 6, 9, and 10 are uncertain or unhealthy in attribute 3, they meet the rules in attributes 1 and 2 that users care about, and poached eggs, celery, and crab meat can also be recommended foods. Therefore, the recommended representative foods that meet the needs of this type of users include cabbage, crucian carp, kale, poached eggs, celery, and crab meat, while roasted gluten, snow pear, and pecan are not recommended foods.

[0074] In summary, the method of the present invention uses cross-scale clustering to combine macro and micro scales to capture the intrinsic characteristics of data on demand, and the clustering results can perfectly meet the needs of users. In practical applications, combined with the personalized needs of different users, cross-scale grid clustering can provide more targeted recommendation solutions for each user.

[0075] The above is only a preferred solution of the present invention and is not intended to be a further limitation of the present invention. All equivalent changes made using the contents of the present invention's specification and drawings are within the protection scope of the present invention.

Claims

1. A cross-scale clustering method that meets the needs of personalized recommendation, characterized in that: The following steps are involved: Step 1: Construct food nutrition data sample The property set , , for any attribute , extract all the values ​​it contains, recorded as , ;in, Indicates the number of food nutrition data samples; Indicates the data dimension, that is, the number of attributes; Indicates attribute value; Representation attributes The number of values ​​of ; Step 2: Construct attributes through the similarity difference of neighboring elements Multi-scale factor set , ;in Indicates the number of multi-scale factors; based on the multi-scale factors, each attribute set corresponding to the data is subjected to single-link hierarchical clustering, and the result is converted into a multi-scale information table corresponding to the food nutritional composition data; Step 3: Based on user needs and the multi-scale information table, select cross-scale combinations of attributes of food nutritional data, calculate the grid step size, and perform grid division according to the values ​​of each attribute at different scales. The division method changes as needed to form a non-uniform grid structure with different scale combinations; Step 4: Map the data into non-uniform grid cells, use cross-scale clustering methods to form grid clusters, and mine the information of each grid cluster to form personalized recommendation solutions for users.

2. A cross-scale clustering method for meeting personalized recommendation needs according to claim 1, characterized in that: Step 2 is as follows: Step 2-1: For the attribute value list , use formula (1) to calculate the similarity difference of each neighbor element , the similarity difference of the nearest neighbor elements after deduplication Considered as a multi-scale factor , and finally obtain the multi-scale factor set ; (1)， Step 2-2, traverse the multi-scale factor set, use formula (2) to calculate the similarity of adjacent attribute values ​​according to the single-link hierarchical clustering rule, and compare the result with the current multi-scale factor, and merge the attribute values ​​that are smaller than the current multi-scale factor at one time: (2)， in, Represents different clusters, Respectively represent the clusters data points, Represents a cluster Repeat step 2-2 until the traversal is completed and the silhouette coefficient of the attribute value cluster under different multi-scale factors is calculated. If it is greater than the given silhouette coefficient value , then output the result, and finally get the multi-level cluster result ,in, Representation attributes In the Layer Clusters of classes; Step 2-3: For each level cluster result , analyze each cluster The interval range and meaning of the attribute value in the middle, giving each interval a unique identifier , map the attribute value to the corresponding interval identifier, and use the interval identifier as the value of the current data object at this scale to obtain the attribute The corresponding multi-scale information is obtained; similarly, the multi-scale information of all other attributes is calculated; finally, the multi-scale information of all attributes is integrated to form a multi-scale information table of the original data in the industry field.

3. A cross-scale clustering method for meeting personalized recommendation needs according to claim 1, characterized in that: Step 3 is as follows: Step 3-1: Select attribute scale combination according to user needs ,in is the number of attributes, For attributes Selected scale; for any attribute , find the triples corresponding to different identifiers at the current scale in the multi-scale information table , where any two adjacent triplets and have , ; and They refer to the minimum value, maximum value and number of data in the two adjacent identification intervals. is the number of values ​​under the current scale; Step 3-2: If the boundary values ​​between different identifiers are discontinuous due to the characteristics of the cluster, the interval density is introduced to divide the discontinuous boundaries. According to formulas (3) and (4), the adjacent interval triples are calculated. and The interval density and transition point : (3)， (4)； Step 3-3: Calculate the attributes according to formula (5) In scale The corresponding non-uniform grid step size is : (5)， Step 3-4: According to formula (6), get the scale combination Non-uniform grid space under : (6)。 4. A cross-scale clustering method for meeting personalized recommendation needs according to claim 1, characterized in that: Step 4 is as follows: Step 4-1: Map data samples to non-uniform grid space In the,dataset,, the data will be presented in different grids under different,scale combinations of different user requirements; Step 4-2: Calculate the non-uniform grid space according to formulas (7) and (8) Each grid Density and the average density of the entire grid space : (7)， (8)， in, and Respectively represent the grid The number and size of the internal data. is the number of grids; Step 4-3, compare the density of each grid with the average density. If it is greater than the average density, it is a dense grid, otherwise, it is a sparse grid; Step 4-4: According to the density connectivity principle, dense grids are merged to form grid clusters; Step 4-5: According to the cluster core points in each grid cluster, the features of the core points of each cluster are extracted to form a personalized recommendation plan.

Images