Method and device for crowd diffusion based on seed crowd

Abstract

The embodiment of the invention provides a crowd diffusion method based on seed crowds, and the method comprises the steps: carrying out the feature compression of a representation vector of each user sample in a crowd data set, and obtaining a compression code corresponding to each user sample, and a seed user set composed of seed user samples belonging to the target population. And dividing the crowd data set into a plurality of user groups according to the compressed codes. And determining a distinction degree index of each user group about the seed user set. And determining a plurality of user groups according to the discrimination index of each user group, and classifying the user samples in the user groups into the target crowd.

Description

Technical Field

[0001] One or more embodiments of the present specification relate to the field of user big data mining and analysis, and in particular, to a method and device for population diffusion based on a seed population. Background Art

[0002] Population diffusion is a commonly used user mining and group segmentation technology. Its purpose is to discover and screen out a larger target population similar to the seed population from a population data set containing a large number of user samples based on the existing seed population. This target population can be a population that shows high conversion potential in marketing activities, or a population with certain characteristics that requires special business processing. Accurately and quickly identifying the target population can facilitate better provision of corresponding services or targeted business processing for the target population, thereby improving user experience and business results.

[0003] Therefore, we hope to have a solution that can use technical means to efficiently analyze the seed population in a population data set with massive user samples, find similar target populations, and improve the efficiency of population diffusion. Summary of the invention

[0004] One or more embodiments of this specification describe a method and device for population diffusion based on a seed population, which uses compressed coding matching with low computational complexity to group massive user samples and retrieve target populations in groups, thereby reducing time complexity and solving the technical problems mentioned above.

[0005] According to a first aspect, a method for population diffusion based on a seed population is provided, comprising:

[0006] Feature compression is performed on the representation vector of each user sample in the crowd data set to obtain compression codes corresponding to each user sample. The crowd data set includes a seed user set composed of seed user samples belonging to the target population.

[0007] According to the compression coding, the crowd data set is divided into a plurality of user groups.

[0008] Determine the discrimination index of each user group with respect to the seed user set.

[0009] According to the discrimination index of each user group, several user groups are determined, and the user samples therein are classified into the target population.

[0010] According to one implementation, the compression coding is M-bit coding; the feature compression includes:

[0011] Obtain M hyperplane vectors; perform an encoding operation on a first representation vector of any first user sample to determine a corresponding first compression code, wherein the j-th element is used to represent a first position relationship between the first representation vector and the N-th hyperplane vector.

[0012] In a scenario of the above implementation, the first position relationship is a directed distance, the compression code is a binary code, and the encoding operation includes:

[0013] If the directed distance between the first representation vector and the Nth hyperplane vector is less than 0, the jth element is encoded as 0.

[0014] If the directed distance between the first representation vector and the Nth hyperplane vector is not less than 0, the jth element is encoded as 1.

[0015] According to an implementation, for any two user samples, a first distance between their compression codes is negatively correlated with a first similarity between their representation vectors.

[0016] In one scenario of the above implementation, the first distance is the Hamming distance, and the first similarity is the cosine similarity.

[0017] According to an implementation manner, dividing the crowd data set into a plurality of user groups includes: dividing user samples having the same compression code into the same group.

[0018] According to an implementation manner, determining the distinguishability index of each user group with respect to the seed user set includes:

[0019] For any user group, determine a first quantity ratio of the seed user samples contained therein in the seed user set, and a second quantity ratio of the user samples contained therein in the population data set.

[0020] A discrimination index of the user group is determined, which is positively correlated with the first quantity ratio and negatively correlated with the second quantity ratio.

[0021] According to one implementation, determining a number of user groups and classifying the user samples therein into the target population includes:

[0022] The multiple user groups are sorted according to their respective discrimination indexes to obtain a first sequence.

[0023] Along the first sequence, user samples in the current user group are classified into the target population in sequence until a preset condition is met.

[0024] In one scenario of the above implementation, the preset condition includes: the number of user samples in the target population reaches a preset first number; or the number of processed user groups reaches a preset second number.

[0025] According to one implementation, determining a number of user groups and classifying the user samples therein into the target population includes:

[0026] Several user groups whose differentiation index is greater than a preset threshold are determined from the multiple user groups.

[0027] The user samples in the several user groups are classified into the target population.

[0028] According to a second aspect, a device for population diffusion based on a seed population is provided, comprising:

[0029] The compression module is configured to perform feature compression on the representation vector of each user sample in the population data set to obtain the compression code corresponding to each user sample. The population data set includes a seed user set composed of seed user samples belonging to the target population.

[0030] The grouping module is configured to divide the crowd data set into a plurality of user groups according to the compression coding.

[0031] The calculation module is configured to determine a distinguishability index of each user group with respect to the seed user set.

[0032] The determination module is configured to determine a number of user groups according to the discrimination index of each user group, and classify the user samples therein into the target population.

[0033] According to a third aspect, a computer program product is provided, comprising a computer program / instruction, which, when executed by a processor, implements the steps of the method described in the first aspect.

[0034] According to a fourth aspect, a computing device is provided, comprising a memory and a processor, wherein an executable code is stored in the memory, and when the processor executes the executable code, the method described in the first aspect is implemented.

[0035] In the method and device provided in the embodiments of this specification, the storage space of user feature data can be effectively reduced by compressing the representation vectors of massive user samples. By using compression coding, the representation similarity calculation between user samples is converted into a code matching with lower computational complexity, and similar user samples are divided into groups. Then, the crowd search is performed in groups, avoiding brute force search of one-to-one comparison on massive user samples. This significantly improves the efficiency of crowd diffusion and expands its application scope. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0037] Figure 1 This is a diagram of the architecture of a method for population diffusion based on a seed population disclosed in an embodiment of this specification;

[0038] Figure 2 A flow chart of a method for population diffusion based on a seed population provided according to an embodiment of this specification;

[0039] Figure 3 The present invention is a schematic diagram of a device for spreading a population based on a seed population according to an embodiment of the present specification. DETAILED DESCRIPTION

[0040] The solution provided by the embodiments of this specification is described below in conjunction with the accompanying drawings.

[0041] As mentioned above, crowd diffusion is an important part of user analysis. As a commonly used user mining and group segmentation technology, its purpose is to retrieve a larger target population that is similar or related to the seed population from a large number of users based on the known seed population.

[0042] In real business scenarios (e.g., social network operations, e-commerce recommendations, advertising, etc.), it is often necessary to find more target groups with similar characteristics based on a seed population (e.g., a user group with a low degree of social relationship with a certain user, a user group that has purchased a certain product, a user group that has clicked on a certain advertisement, etc.), thereby expanding the seed population with a small sample size into a target population with more user samples. Precision marketing for this type of target population can usually achieve the same marketing effect as the seed population. This is of great significance for expanding marketing coverage and improving user conversion rates.

[0043] The features of each user sample can be expressed using a representation vector. The user's representation vector is a compact and representative vector generated by processing the original features of each user sample. The original features of the user sample may include demographic information, user behavior data, user interaction records and other information, which are usually high-dimensional and sparse. Usually, based on the original features, common representation learning methods such as word embedding and factor decomposition machines can be used to convert these original features into a multi-dimensional real-valued vector as the representation vector of the user sample. In this way, the similarity between different user samples can be determined by comparing the representation vectors that can express the characteristics of the user samples.

[0044] In order to achieve crowd diffusion, in a related technology, the user samples are characterized by similarity calculation and sorting, and the Top-N user samples that are most similar to each seed user sample are obtained as the target population. In this process, in order to comprehensively evaluate the full user samples in the crowd data set, it is inevitable to calculate the representation similarity between each seed user sample and the full user sample, and then perform a brute force search on the full user samples. This method can achieve crowd diffusion on a small data set. Once the number of full user samples in the crowd data set reaches a certain scale (in Internet business systems, the number of user samples in the crowd data set can usually reach hundreds of millions), this one-to-one calculation and brute force retrieval method will become slower and slower and cannot meet real-time requirements.

[0045] On the one hand, the calculation of the similarity between representation vectors is often measured by the distance or angle between the two representation vectors. For example, for two k-dimensional representation vectors A and B:

[0046] The Euclidean distance can be used to measure the similarity between the representations:

[0047]

[0048] Expands to:

[0049] Sqrt((A 1 -B 1 ) 2 +(A 2 -B 2 ) 2 +…+(A k -B k ) 2 ) (2)

[0050] That is, in this manner, to calculate the representation similarity between a pair of representation vectors, k representation vector element difference calculations, k square (product) calculations, and 1 square root calculation are required.

[0051] You can also use cosine similarity to measure the similarity between representations:

[0052]

[0053] in:

[0054] A·B=A 1 *B 1 +A 2 *B 2 +…+A k *B k (4)

[0055] |A|=Sqrt((A 1 ) 2 +(A 2 ) 2 +…+(A k ) 2 ) (5)

[0056] |B|=Sqrt((B 1 ) 2 +(B 2 ) 2 +…+(B k ) 2 ) (6)

[0057] That is, in this way, to calculate the representation similarity between a pair of representation vectors, k representation vector element multiplication calculations and 2k square (product) calculations are required to obtain the modulus of the representation vectors.

[0058] It can be seen that in order to calculate the similarity between representation vectors, multiple product operations that are positively correlated with the dimensions of the representation vectors are required, which has a high computational complexity.

[0059] On the other hand, when facing a population dataset with a massive number of user samples, it is necessary to calculate the representation similarity between each seed user sample in the seed user set and each user sample in the population dataset. This calculation method is equivalent to performing a Cartesian product expansion on the seed user set and the population dataset, which will bring huge computational overhead and a time complexity of O(n 2 ).

[0060] Through the above analysis, it can be seen that in a population data set with massive user samples, the method of calculating the representation similarity one-to-one and performing brute force retrieval on the entire data to achieve population diffusion will consume a lot of computing resources, have low retrieval efficiency, and cannot meet real-time requirements.

[0061] In order to solve the above problems, a method and device for population diffusion based on seed population are proposed in the embodiments of this specification. On the one hand, by compressing the characteristics of the representation vector of the user sample, a compression code that occupies a smaller storage space and can correctly express the feature similarity relationship between different user samples is obtained. On the other hand, by using matching calculations between compression codes, instead of the representation similarity calculation that consumes a lot of computing resources, a large number of user samples are grouped according to similarity. On the other hand, by calculating and comparing the discrimination index of different groups, the group close to the seed population is determined, thereby obtaining the target population and avoiding a brute force search of all user samples.

[0062] exist Figure 1 In the figure, a method architecture diagram for population diffusion based on seed population is shown. In a population data set with a large number of user samples, according to business rules, a part of user samples can be circled as seed user samples to form a seed population. The goal of population diffusion is to retrieve user samples similar to the seed user samples in the population data set based on the seed population as the target population. First, feature compression is performed on each user sample in the population data set. Feature compression is to compress the representation vector of the user sample into a compression code that occupies less data space. The compression code can correctly reflect the similarity relationship between different user samples. Then, the compression code is matched, and user samples with the same or similar (depending on the accuracy requirements of population diffusion) compression codes are divided into the same group. According to the number of seed user samples and non-seed user samples contained in the group, the sensitivity of each group to the seed population can be determined as the discrimination index of the group. Finally, according to the discrimination index of each group, the groups are sorted, and user samples are taken out from the sorted groups in turn and classified into the target population, so that the target population that can meet the business needs can be obtained.

[0063] Following the above technical ideas. Figure 2 In FIG. 1 , a flow chart of a method for population diffusion based on a seed population according to an embodiment of the present specification is shown. It can be understood that the method can be executed by any device, equipment, platform, or device cluster with computing and processing capabilities. Figure 2In one embodiment, the method includes at least the following steps. S201: Perform feature compression on the representation vector of each user sample in the population data set to obtain the compression code corresponding to each user sample, wherein the population data set includes a seed user set composed of seed user samples belonging to the target population. S203: According to the compression code, the population data set is divided into multiple user groups. S205: Determine the discrimination index of each user group with respect to the seed user set. S207: Determine several user groups according to the discrimination index of each user group, and classify the user samples therein into the target population.

[0064] The specific implementation methods of the above steps will be described in detail below with reference to the accompanying drawings.

[0065] The crowd data set contains a large number of user samples, some of which are seed user samples, which constitute the seed user set. It should be noted that in some embodiments, the seed user samples may also come from other user data sets other than the crowd data set, and the embodiments of this specification do not limit the source of the seed user samples.

[0066] Based on the crowd data set, in step S201, feature compression is performed on the representation vector of each user sample in the crowd data set to obtain the compression code corresponding to each user sample. As mentioned above, according to the original features of the user sample, the representation vector of the user sample can be obtained through sampling processing. In this step, feature compression is performed on the representation vector of each user sample in the crowd data set, in order to obtain a smaller data volume and a compression code that can correctly reflect the similarity between different user samples.

[0067] In the specific implementation, a variety of feature compression methods can be used. For example, dimensionality reduction can be achieved by projecting the representation vector into a low-dimensional space, or by performing principal component analysis on the representation vector and retaining some of the main feature information therein to achieve dimensionality reduction of the representation vector, and so on.

[0068] In a specific implementation, a hash algorithm may be used to create a corresponding digital "fingerprint" based on the characterization vector. The hash function may compress the characterization vector data to reduce the amount of data.

[0069] Assume that the representation vector of a user sample is a d-dimensional real-valued vector, the representation vector of user sample i is The goal of feature compression is to compress the representation vector into an M-bit compression code. In some typical practices, the value of M is 20.

[0070] First, randomly obtain M d-dimensional hyperplane vectors

[0071] Then, for any user sample i, the representation vector u i An encoding operation is performed to determine the corresponding compression code, wherein the j-th element is used to represent the positional relationship between the characterization vector and the j-th hyperplane vector.

[0072] Through the above feature compression operation, the d-dimensional representation vector can be compressed into a compressed code with M bits of data. Further, in a specific scenario, the representation vector can be compressed into a binary code string, and the above position relationship can be represented by the representation vector u i With hyperplane r j The directed distance s between ij Indicates that, according to s ij The positive and negative signs of the bit are used to generate the binary code e of user sample i ij , as its compressed encoding:

[0073]

[0074] e ij =1(s ij ≥0), j=1,2,...,M (8)

[0075] Among them, 1(·) is the indicative function. The above formula shows that if the representation vector u of user sample i i , the directed distance s to the jth hyperplane ij If it is less than 0, the j-th element of the compression code is encoded as 0, otherwise, the j-th element of the compression code is encoded as 1. Thus, the compression code e of user sample i can be obtained i =(e i1 ,e i2 ,…,e iM )∈{0,1} M .

[0076] It should be understood that in the above implementation, the compression-coded element coding bit (j) corresponds one-to-one with the position (j) of the hyperplane for comparing the representation vector, but this does not mean that the corresponding relationship is a limitation on the compression-coded element coding bit. In other implementations, the compression-coded element coding bit may correspond to the position of the comparison hyperplane in a misaligned manner, but at least there is a certain comparison relationship between the compression-coded element and the position of the comparison hyperplane. For example, a check information bit is inserted before each element representing the position relationship, then the comparison relationship between the compression-coded element and the position relationship is: the 2j-th element is used to represent the position relationship between the representation vector and the j-th hyperplane vector.

[0077] In this scenario, the representation vector of a high-dimensional real-valued vector is mapped to a low-dimensional binary coding space. Moreover, for any two user samples, the higher the similarity of their representations in the original d-dimensional space, the closer the distance between their corresponding compressed codes in the coding space, and vice versa. In other words, for any two user samples, the distance between their compressed codes is negatively correlated with the similarity between their representation vectors.

[0078] In one example, for any two user samples, the Hamming distance can be used to represent the distance between their compressed codes, and the cosine similarity can be used to represent the similarity between their representation vectors. That is, the Hamming distance between the compressed codes obtained after feature compression can be used to measure the angle between the original representation vectors. In other examples, other metrics for representing the similarity between vectors and other metrics for representing the similarity between codes can also be used to measure the relationship between the compressed code and the representation vector. This specification does not list them one by one.

[0079] After obtaining the compression codes corresponding to the user samples in the crowd data set, the user samples can be divided into different groups according to the compression codes. Therefore, in step S203, the crowd data set is divided into multiple user groups according to the compression codes. Since the compression codes can correctly reflect the similarity between different user samples, user samples with similar features can be identified according to the compression codes, and these user samples can be classified into the same group.

[0080] According to one implementation, user samples with the same compression code can be divided into the same group. Assume that the set of binary compression codes is According to the rule that one coding combination is divided into one group, there are 2 M different groups, and the kth group is recorded as B k , which contains the compressed code Sample of all users:

[0081] B k ={i|e i =c k}, k = 1, 2, ..., 2 M (9)

[0082] Through the feature compression operation as described above, user samples with similar original feature vectors can obtain the same compression code in the low-dimensional space; and then through the grouping operation, similar user samples can be grouped into the same group. In this way, in the subsequent target population search, according to the business requirements for the target population search accuracy, it is only necessary to select user samples from one or several groups with similar features. In the above calculation process, the representation similarity calculation between all user samples is avoided, and only one feature compression is performed on each user sample, which reduces the time complexity of the representation similarity calculation from O(n 2 ) is reduced to the constant level O(n).

[0083] Next, in step S205, for each group, the discrimination index of the group for the seed user set is calculated. In each group, since several user samples with similar characteristics are included, the similarity between the user samples included in the group and the seed user samples can be determined by evaluating the expression sensitivity of the group to the seed population characteristics. Generally, when the user samples included in a group appear more frequently in the seed user set, but appear less frequently in the population data set containing a large number of non-seed user samples, it can be indicated that the compression coding corresponding to the group has significant category differentiation ability for the seed user set. Accordingly, the user samples included in the group can be classified as part of the target population.

[0084] According to one implementation, for any user group, first determine the first quantity ratio of the seed user samples contained therein in the seed user set and the second quantity ratio of the user samples contained therein in the population data set. Then determine the discrimination index of the user group, where the discrimination index is positively correlated with the first quantity ratio and negatively correlated with the second quantity ratio.

[0085] Assume that the seed user set is The crowd dataset is Then the kth group B k The proportions of the numbers in the seed user set and the crowd data set are:

[0086] The first quantity ratio:

[0087] The second quantity share:

[0088] Where k = 1, 2, ..., 2 M , |·| represents the cardinality of the set, that is, the number of elements.

[0089] For any group B k , its discrimination degree D for the seed user set k, which can be defined as:

[0090]

[0091] D k The larger the value, the more likely it is that group B k The higher the frequency of the user samples included in the seed user set, the lower the frequency in the crowd data set, which means that group B k The higher the discrimination of seed user set.

[0092] According to another implementation, the contribution of the group to the target population, that is, the information value (IV) of the group, can be used as the distinguishing index of the group for the seed user set. This implementation can be summarized as follows:

[0093] Assume that the seed user set is The crowd dataset is Then the kth group B k The proportion of the number of seed users and the number of people in the data set is the first proportion shown in formula (10) and formula (11) respectively. And the second number of

[0094] Referring to the idea of ​​information value, for any group B k , its discrimination degree D for the seed user set k , which can be defined as:

[0095]

[0096] D k The larger the value, the more likely it is that group B k The greater the contribution to the target population, the greater the contribution of group B. k The higher the discrimination of seed user set.

[0097] After the above steps are processed, the discrimination index of each group is obtained through evaluation. Next, in step S207: according to the discrimination index of each user group, several user groups are determined, and the user samples therein are classified into the target population. In this step, according to the discrimination index of the group, combined with the business requirements for the similarity of the target population to the seed population in the population diffusion, several groups that meet the business needs can be determined, and the user samples contained in these groups are selected as the target population, so that the population diffusion task based on the seed population can be completed.

[0098] According to one implementation, the groups may be sorted according to their discrimination index, and then the sorted groups or user samples in the groups may be conditionally selected to meet the service requirements. The implementation may be summarized as follows: the plurality of user groups are sorted according to their respective discrimination indexes to obtain a first sequence. Assume that in the obtained first sequence, the index of each group may be expressed as:

[0099]

[0100] Then, we can know that the relationship between the discrimination indicators of different groups is:

[0101]

[0102] Next, along the first sequence, user samples in the current user group are classified into the target population in sequence until the preset condition is met. In a specific scenario, the preset condition can be flexibly set according to business needs.

[0103] In one example, the preset condition may be that the number of user samples in the target population reaches a preset first number. Under this preset condition, the target population obtained by population diffusion It can be expressed as:

[0104]

[0105] Wherein, m is the first number set in the preset condition, that is, the minimum number of user samples required in the target population. t is the minimum number of groups required to meet the preset condition:

[0106]

[0107] In another example, the preset condition may also be that the number of processed user groups reaches a preset second number. Under this preset condition, the expected number of groups can be set, and then the user samples in the groups are sorted according to the first sequence, and the process of successively classifying the user samples in the groups into the target population is stopped when the number of processed groups reaches the expected number, and the population diffusion task can be completed.

[0108] The above is a description of an implementation method for selecting user samples to enter the target population. In some embodiments, according to another implementation method, according to business requirements, the qualified groups are first selected, and then the user samples in the selected groups are classified as the target population. This implementation method can be summarized as follows: determining several user groups whose discrimination index is greater than a preset threshold from the multiple user groups; and classifying the user samples in the several user groups into the target population.

[0109] Under this preset condition, the target population obtained through population diffusion It can be expressed as:

[0110]

[0111] Among them, n is the preset threshold value set according to business needs, that is, the minimum value that the discrimination index corresponding to the selected group should meet. In this way, user samples whose similarity with the seed user samples does not exceed the preset threshold can be included in the target population category, thereby achieving the population diffusion task based on the seed population.

[0112] According to one or more embodiments, a method for population diffusion based on a seed population is described in detail above. By adopting the above method provided in the embodiments of this specification, in the process of implementing population diffusion, the representation vector of the user sample can be compressed to obtain a compressed code with a smaller storage space, and the user sample can be assigned to different groups based on the compressed code, and the target population can be retrieved in groups. The original high computational complexity of the representation similarity calculation can be effectively converted into a simple and efficient coding matching calculation. At the same time, the use of group retrieval can reduce the time complexity of population diffusion to a constant level, thereby improving the computational efficiency of population diffusion.

[0113] In this specification, the word "first" in terms such as the first user sample and the first quantitative percentage, as well as the corresponding "second", "third" (if any), etc. in the text, are merely for the convenience of distinction and description and do not have any limiting meaning.

[0114] The foregoing describes certain embodiments of the present specification, and other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims may be performed in an order different from that in the embodiments, and the desired results may still be achieved. In addition, the processes depicted in the accompanying drawings do not necessarily have to be performed in the specific order or sequential order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0115] Figure 3 Schematic diagram of a device for population diffusion based on seed population according to an embodiment of this specification. The device 300 is deployed in a computing device, which can be implemented by any device, equipment, platform, device cluster, etc. with computing and processing capabilities. Figure 2 The device 300 includes:

[0116] The compression module 301 is configured to perform feature compression on the representation vector of each user sample in the population data set to obtain compression codes corresponding to each user sample. The population data set includes a seed user set consisting of seed user samples belonging to the target population.

[0117] The grouping module 302 is configured to divide the crowd data set into a plurality of user groups according to the compression coding.

[0118] The calculation module 303 is configured to determine a discrimination index of each user group with respect to the seed user set.

[0119] The determination module 304 is configured to determine a number of user groups according to the discrimination index of each user group, and classify the user samples therein into the target population.

[0120] According to another embodiment of the present specification, there is also provided a computer program product, including a computer program / instruction, which implements the aforementioned combination when executed by a processor. Figure 2 The steps of the method.

[0121] According to another embodiment, the present specification further provides a computing device, including a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the aforementioned combination is implemented. Figure 2 The steps of the method.

[0122] Those skilled in the art should be aware that in one or more of the above examples, the functions described in the embodiments of the present invention may be implemented using hardware, software, firmware, or any combination thereof. When implemented using software, these functions may be stored in a computer-readable medium or transmitted as one or more instructions or codes on a computer-readable medium.

[0123] The specific implementation methods described above further describe the purpose, technical solutions and beneficial effects of the embodiments of the present invention in detail. It should be understood that the above description is only a specific implementation method of the embodiments of the present invention and is not intended to limit the scope of protection of the present invention. Any modification, equivalent replacement, improvement, etc. made on the basis of the technical solution of the present invention shall be included in the scope of protection of the present invention.

Claims

1. A method for population diffusion based on a seed population, comprising: Performing feature compression on the representation vector of each user sample in the crowd data set to obtain compression codes corresponding to each user sample, wherein the crowd data set includes a seed user set consisting of seed user samples belonging to the target population; According to the compression coding, the crowd data set is divided into a plurality of user groups; Determine the distinguishability index of each user group with respect to the seed user set; According to the discrimination index of each user group, several user groups are determined, and the user samples therein are classified into the target population.

2. The method according to claim 1, wherein: The compression code is an M-bit code; The feature compression includes: Get M hyperplane vectors; An encoding operation is performed on a first characterization vector of any first user sample to determine a corresponding first compression code, wherein the j-th element is used to represent a first position relationship between the first characterization vector and the N-th hyperplane vector.

3. The method according to claim 2, wherein: The first positional relationship is a directed distance, and the compression code is a binary code; The encoding operation includes: If the directed distance between the first representation vector and the Nth hyperplane vector is less than 0, the jth element is encoded as 0; If the directed distance between the first representation vector and the Nth hyperplane vector is not less than 0, the jth element is encoded as 1.

4. The method according to claim 1, wherein: For any two user samples, the first distance between their compressed codes is negatively correlated with the first similarity between their representation vectors.

5. The method according to claim 4, wherein: The first distance is the Hamming distance, and the first similarity is the cosine similarity.

6. The method according to claim 1, wherein: The dividing the crowd data set into a plurality of user groups includes: dividing user samples having the same compression code into the same group.

7. The method according to claim 1, wherein: The step of determining the distinguishability index of each user group with respect to the seed user set includes: For any user group, determine a first quantity ratio of the seed user samples contained therein in the seed user set, and a second quantity ratio of the user samples contained therein in the population data set; A discrimination index of the user group is determined, which is positively correlated with the first quantity ratio and negatively correlated with the second quantity ratio.

8. The method according to claim 1, wherein: The determining of a number of user groups and classifying the user samples therein into the target population includes: Sorting the multiple user groups according to their respective discrimination indexes to obtain a first sequence; Along the first sequence, user samples in the current user group are classified into the target population in sequence until a preset condition is met.

9. The method according to claim 8, wherein the preset condition comprises: The number of user samples in the target population reaches a preset first number; Alternatively, the number of processed user groups reaches a preset second number.

10. The method according to claim 1, wherein: The determining of a number of user groups and classifying the user samples therein into the target population includes: Determine, from the multiple user groups, a number of user groups whose discrimination index is greater than a preset threshold; The user samples in the several user groups are classified into the target population.

11. A device for population diffusion based on seed population, comprising: A compression module is configured to perform feature compression on a representation vector of each user sample in a population data set to obtain a compression code corresponding to each user sample, wherein the population data set includes a seed user set consisting of seed user samples belonging to a target population; a grouping module configured to divide the crowd data set into a plurality of user groups according to the compression coding; A calculation module configured to determine a distinguishability index of each user group with respect to a seed user set; The determination module is configured to determine a number of user groups according to the discrimination index of each user group, and classify the user samples therein into the target population.

12. A computer program product, comprising a computer program / instruction, which, when executed by a processor, implements the steps of the method according to any one of claims 1 to 10.

13. A computing device comprising a memory and a processor, characterized in that: The memory stores executable codes, and when the processor executes the executable codes, the method according to any one of claims 1 to 10 is implemented.

Images