Method, apparatus, and computer device for predicting user group to which a user belongs based on search-by-shooting behavior

Abstract

The present invention belongs to the field of data processing technology and provides a method, apparatus, and computer device for predicting a user's group based on photo search behavior. The method comprises: dividing users into different sets so that users in the same set have the same group information, wherein the group information is related to the user's photo search behavior; for users in the set, based on the similarity level of the users' photo search behavior, screening out seed users whose group information has a confidence level greater than a first preset value; for non-seed users, calculating the similarity between the non-seed user's photo search behavior and that of various sub-users, and predicting the group to which the non-seed user belongs based on the similarity. The non-seed users include users who lack the group information and users in the set whose group information has a confidence level no greater than a first preset value. The present invention can more accurately predict the group to which a non-seed user belongs, further improving prediction accuracy, and further optimizing the method.

Description

Technical Field

[0001] The present invention belongs to the field of data processing technology, and is particularly applicable to the field of education. More specifically, it relates to a method, apparatus, and computer equipment for predicting a user's group based on photo and search behavior. Background Art

[0002] With the advent of the big data era, data has exploded. To address information overload, recommendation systems have been widely used in online services such as e-commerce, content sharing, social networks, and forums. Consequently, recommendation systems need to provide recommendations for different groups of people. In addition to recommendation systems for individuals, recommendation systems for user groups are also needed.

[0003] Currently, many electronic education products have a photo search function. Users use the terminal's camera to take a photo of a question on paper. After the terminal takes the photo and displays the captured image, the user can search for the answer to the question by manipulating the selection box displayed on the terminal interface, thereby completing the photo search function. However, in the existing technology, due to the inaccurate group classification of users of the above-mentioned electronic education product apps, it is impossible to implement customized recommendation services for each user group. In addition, there are still many issues worth studying in terms of the application and data prediction of the large amount of data generated by photo search behavior. In addition, there are technical problems such as low accuracy in existing user group predictions and slow data processing speeds caused by the large amount of user data.

[0004] Therefore, it is necessary to provide a method for predicting the group to which a user belongs based on photo-search behavior to solve the above problems. Summary of the Invention

[0005] (1) Technical issues to be resolved

[0006] The present invention aims to solve technical problems in existing educational product APPs, such as inaccurate group classification of users, low prediction accuracy of existing methods, and slow data processing speed caused by large amounts of user data.

[0007] (2) Technical solution

[0008] To solve the above technical problems, one aspect of the present invention proposes a method for predicting the group to which a user belongs based on photo search behavior, wherein the photo search behavior refers to the behavior of initiating a photo search request and obtaining search results, and the method comprises the following steps: dividing users into different sets so that users in the same set have the same group information, and the group information is related to the user's photo search behavior; based on the similarity level of the user's photo search behavior, screening out seed users in each set whose confidence level of the group information is greater than a first preset value; for non-seed users, calculating the similarity of the photo search behavior of the non-seed user with various sub-users, and predicting the group to which the non-seed user belongs based on the similarity, the non-seed users include users who lack the group information and users in the set whose confidence level of the group information is not greater than a first preset value.

[0009] According to a preferred embodiment of the present invention, before screening the seed users, the method further includes: labeling the search results of the photo search behavior, and converting the labeled search results into a feature vector to characterize the user's photo search behavior; and subsequently calculating the similarity of the user's photo search behavior based on the feature vector.

[0010] According to a preferred embodiment of the present invention, labeling the search results of the photo search behavior and converting the labeled search results into a feature vector includes: obtaining the user's historical photo search behavior, and representing the historical photo search behavior as a label sequence based on the corresponding search results, each label in the label sequence represents at least one feature of the search result; converting the label sequence into a vector sequence; and normalizing the vector sequence into the feature vector.

[0011] Optionally, the method further includes: characterizing the user as a feature vector of his / her photo-searching behavior: defining the name of the feature vector as the user identifier, and defining the length of the feature vector as a feature representation of the user's photo-searching behavior.

[0012] Optionally, the search results corresponding to the historical search behavior are represented as a label sequence, including: labeling the search results; and deduplicating the search result data after labeling.

[0013] Optionally, the photo search behavior refers to the behavior of initiating a photo-based search request to obtain a search result.

[0014] Optionally, the photo is a full-page photo; and the tag sequence includes: test questions and book pages.

[0015] Optionally, the group information includes at least one of the following: the user's region, school, grade, class, or group.

[0016] Optionally, the tags generated in the step of tagging the search results of the photo search behavior include at least one of the following tags: textbook, test paper, book, exercise book, page number, test question.

[0017] According to a preferred embodiment of the present invention, the step of screening out seed users includes: performing clustering processing on the feature vectors; and selecting users in the largest cluster obtained after clustering in the set as the seed users.

[0018] Optionally, a community discovery algorithm is used within each set to perform the clustering process on the users in the set.

[0019] According to a preferred embodiment of the present invention, after calculating the similarity between the shooting and searching behaviors of non-seed users and various sub-users, the method further includes: screening out seed users whose similarity with the non-seed users is within a predetermined range; when predicting the group to which the non-seed user belongs, predicting the group to which the non-seed user belongs or the group characteristics based on the group information of the screened seed users.

[0020] According to a preferred embodiment of the present invention, a user space is defined, in which each user is a vertex, the similarity relationship between adjacent users is an edge, and the similarity of the feature vectors of the adjacent users' photo search behavior is the weight of the edge. When seed users whose confidence in the group information is greater than a first preset value are screened out, the users in each set are clustered using the Louvain community discovery algorithm to obtain the seed users; the calculation of the similarity of the photo search behavior of non-seed users and various sub-users includes: in the user space, calculating the distance between the non-seed user and the various sub-users as the similarity; the screening of seed users whose similarity with the non-seed user is within a predetermined range includes: calculating the average distance between any two seed users in the largest class in each set, and screening out seed users whose distance between the non-seed user and the seed users of each set is less than the average distance.

[0021] According to a preferred embodiment of the present invention, predicting the group to which the non-seed user belongs based on the original group information of the screened seed users includes: counting the number of screened seed users according to their original group information, and taking the original group with the largest number of corresponding seed users as the predicted group to which the non-seed user belongs; when the number of users included in the cluster to which the non-seed user belongs is within a predetermined range and the cluster contains a seed user, predicting the group to which the seed user belongs as the group to which the non-seed user belongs.

[0022] According to a preferred embodiment of the present invention, before counting the number of the screened seed users according to their original group information, the method further includes: screening the screened seed users according to their regional information to eliminate seed users that are not in the same region as the non-seed users.

[0023] The second aspect of the present invention proposes a prediction device for predicting the group to which a user belongs based on photo search behavior, wherein the photo search behavior refers to the behavior of initiating a photo search request and obtaining search results. The prediction device includes: a grouping module, which divides users into different sets so that users in the same set have the same group information, and the group information is related to the user's photo search behavior; a screening module, which screens out seed users in each set whose confidence level of the group information is greater than a first preset value based on the similarity level of the user's photo search behavior; a prediction module, which calculates the similarity of the photo search behavior of the non-seed user with that of various sub-users, and predicts the group to which the non-seed user belongs based on the similarity, and the non-seed user includes users who lack the group information and users in the set whose confidence level of the group information is not greater than a first preset value.

[0024] A third aspect of the present invention proposes a computer device comprising a processor and a memory, wherein the memory is used to store a computer executable program. When the computer program is executed by the processor, the processor executes the method of predicting the group to which a user belongs based on photo-search behavior.

[0025] A fourth aspect of the present invention provides a computer program product storing a computer executable program, which, when executed, implements the method of predicting the group to which a user belongs based on photo-search behavior.

[0026] (3) Beneficial effects

[0027] Compared to existing technologies, this method, based on the similarity of users' photo-search behavior, selects seed users from each set whose group information has a confidence level greater than a first preset value. For non-seed users in the group to be predicted, the method first calculates the photo-search behavior similarity between the non-seed user and various sub-users. Then, based on this similarity, it predicts the group to which the non-seed user belongs. These non-seed users include users with missing group information and users whose group information in the set has a confidence level no greater than a first preset value. This allows for more accurate prediction of the group to which non-seed users belong, further improving prediction accuracy and achieving a simple and efficient algorithm.

[0028] Furthermore, the present invention uses the search results of the photo search behavior to perform vector conversion and use it to characterize the user vector, so as to obtain more accurate photo search behavior data and more accurately characterize the user's group information; by removing non-seed users from each original group, a more accurate seed user set (i.e., user group category) can be obtained, and the prediction accuracy can be improved; and by using a feature search engine to search the seed user database, the user feature vectors of the non-seed users are similar to the user feature vectors of the seed users in the seed user database, so as to screen out the seed users whose similarity is within a predetermined range, and use the original group with the largest number of corresponding seed users as the predicted group to which the non-seed user belongs, which can effectively avoid the problem of slow data processing speed caused by the large amount of user data, can more accurately predict the group to which the non-seed user belongs, can further improve the prediction accuracy, and can further optimize the method. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 This is a flowchart of an example of a method for predicting a user's group based on photo-search behavior according to embodiment 1 of the present invention;

[0030] Figure 2 is a flowchart of another example of the method for predicting the group to which a user belongs based on photo-search behavior according to embodiment 1 of the present invention;

[0031] Figure 3 is a schematic diagram showing an example of obtaining a seed user set by clustering processing in the method of embodiment 1 of the present invention;

[0032] Figure 4 This is a flowchart of another example of the method for predicting the group to which a user belongs based on photo-search behavior according to embodiment 1 of the present invention.

[0033] Figure 5 is a schematic diagram of an example of a prediction device for predicting a group to which a user belongs based on photo-search behavior according to embodiment 2 of the present invention;

[0034] Figure 6 is a schematic diagram of another example of a prediction device for predicting a group to which a user belongs based on photo-search behavior according to embodiment 2 of the present invention;

[0035] Figure 7 is a schematic diagram of another example of a prediction device for predicting a group to which a user belongs based on photo-search behavior according to embodiment 2 of the present invention;

[0036] Figure 8 is a schematic structural diagram of a computer device according to an embodiment of the present invention;

[0037] Figure 9 is a schematic diagram of a computer program product according to an embodiment of the present invention. DETAILED DESCRIPTION

[0038] In the introduction of specific embodiments, the detailed description of the structure, performance, effect or other features is intended to enable those skilled in the art to fully understand the embodiments. However, this does not preclude those skilled in the art from implementing the present invention with a technical solution that does not include the aforementioned structure, performance, effect or other features under specific circumstances.

[0039] The flowcharts in the accompanying drawings are merely illustrative of the process flow and do not necessarily include all of the content, operations, and steps in the flowcharts, nor do they necessarily imply that all of the steps in the flowcharts must be executed in the order shown. For example, some of the steps in the flowcharts may be separated, some may be combined or partially combined, and so on. The execution order shown in the flowcharts may be changed according to actual circumstances without departing from the spirit of the present invention.

[0040] Frames in the accompanying drawings Figure 1 The term "functional entity" generally refers to a functional entity and does not necessarily correspond to a physically independent entity. That is, these functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different networks and / or processing unit devices and / or microcontroller devices.

[0041] The same reference numerals in the accompanying drawings represent the same or similar elements, components or parts, and thus repeated descriptions of the same or similar elements, components or parts may be omitted below. It should also be understood that although the first, second, third and other numbered adjectives may be used herein to describe various devices, elements, components or parts, these devices, elements, components or parts should not be limited by these adjectives. In other words, these adjectives are only used to distinguish one from another. For example, the first device may also be called the second device, but this does not deviate from the essential technical solution of the present invention. In addition, the terms "and / or" and "and / or" refer to all combinations including any one or more of the listed items.

[0042] The present invention provides a method for predicting a user's group based on photo search behavior. This method uses a feature search engine to build a database of all seed users, calculates similarity between the user feature vectors of the user to be predicted and the user feature vectors of the seed users, and screens out seed users whose similarity is within a predetermined range. The group with the largest number of corresponding seed users is then used as the predicted group for the user to be predicted. This method effectively avoids the slow data processing speed caused by large amounts of user data, more accurately predicts the groups to which non-seed users belong, further improves prediction accuracy, and further optimizes the method. The user to be predicted is a user who lacks group information or whose group information has low confidence.

[0043] It should be noted that the above-mentioned feature search engine refers to searching for information from the user database and feeding it back to the user using a specific strategy based on user needs and specific algorithms. For example, the faiss tool is used as a feature search engine. When establishing a user database, the input is the ID of each database element and its feature vector. When the feature search engine searches, information can be returned by setting parameters, such as setting the element ID that is less than a specific similarity and the similarity with the element ID. In the present invention, the database building elements include at least one of the following elements: a label, an identification code corresponding to each label (i.e., an element ID), a vector corresponding to each identification code, and user IDs of various sub-users (i.e., a user account or user identification code), etc., wherein labels include textbooks, test papers, books, exercise books, page numbers, test questions, etc.; the identification code is a label code or label ID (i.e., an element ID) representing each label.

[0044] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings.

[0045] Figure 1 This is a flowchart of an example of a method for predicting a user's group based on photo-search behavior according to embodiment 1 of the present invention. Figure 1 As shown, the present invention provides a method for predicting the group to which a user belongs based on photo-search behavior, the method comprising:

[0046] Step S101 : dividing users into different groups so that users in the same group have the same group information, where the group information is related to the user's photo-search behavior.

[0047] The technical solution of this invention is primarily based on the fact that users with the same or similar photo-searching behaviors generally belong to the same group or have the same group characteristics. Seed users with highly confident group information can be used to predict the groups or group characteristics of other users (e.g., users with no group information or with low-confidence group information).

[0048] It should be noted that, in the present invention, the group characteristics refer to the common characteristics that characterize a certain group (ie, a certain set of seed users). For example, the group characteristics of the senior high school students in Beijing at least include the region Beijing and the grade senior high school.

[0049] Based on the above principle, this step divides users into different sets so that users in the same set have the same group information, thereby establishing a candidate set for selecting seed users, so as to facilitate subsequent screening of seed users with higher confidence.

[0050] Step S102 : Screening out seed users whose confidence level of the group information is greater than a first preset value based on the similarity level of the users' photo-searching behaviors.

[0051] For each set, this step selects users whose group information has a high confidence level as seed users based on the similarity level of their photo-search behavior. Specifically, users whose group information has a confidence level greater than a first preset value can be selected. The specific value of the first preset value can be determined by those skilled in the art based on practical circumstances, for example, based on accuracy and recall requirements, to ensure that the user's group information is credible.

[0052] Optionally, multiple clusters can be obtained through clustering processing, and the users in the largest cluster of the set have higher confidence and are called seed users.

[0053] Step S103: For non-seed users, calculate the similarity between the photo-shooting and search behaviors of the non-seed user and various sub-users, and predict the group to which the non-seed user belongs based on the similarity. The non-seed users include users who lack the group information and users whose confidence in the group information in the set is not greater than a first preset value.

[0054] In this step, the similarity between the photo-search behavior of the non-seed user and the seed users in each set is calculated. If the photo-search behavior similarity between the non-seed user and the seed users in a certain set is high, the group information of the seed users in the set can be used to predict the group of the non-seed user.

[0055] The method of the present invention for predicting a user's group based on photo search behavior can more accurately predict the group to which a user with unknown or low confidence group information belongs based on the similarity of the user's photo search behavior, further helping to provide users with more accurate services or service push.

[0056] It should be noted that in the present invention, the "photo search" behavior refers to the act of initiating a photo search request and obtaining search results. There are many usage scenarios. The following uses the application scenario of photo search in an educational service product that includes the "photo search" function as an example to explain it in detail.

[0057] Figure 2 4 is a flowchart illustrating another example of the method for predicting a group to which a user belongs based on photo-search behavior according to the first embodiment of the present invention.

[0058] like Figure 2 As shown, in step S201, users are divided into different groups so that users in the same group have the same group information, and the group information is related to the user's shooting and searching behavior.

[0059] In this example, for example, the user's APP registration information for the educational service product can be called to identify the school information or grade information in the APP registration information, and the school or grade group information in the APP registration information can be used to divide the users into multiple sets (i.e., seed user candidate sets, used to select seed users with higher confidence).

[0060] It should be noted that the set of this embodiment (i.e., the seed user candidate set) is a temporary user set established for selecting seed users based on predicted needs. For example, if you want to predict whether the user group is using the People's Education Edition or the Beijing Normal University Edition, when dividing the users into multiple sets, the group information based on may include at least one of region, school, grade, and class. A user can have multiple group information at the same time. For example, according to actual needs, the user's grade can also be predicted, but it is not limited to this. It can also predict the region, school, and the same type of users who may have common needs (such as tutoring needs), etc.

[0061] Specifically, the APP registration information may also include regional information, school stage information, user account, etc. Preferably, the APP registration information is, for example, the user's registration information for an education service product APP, wherein the education service product APP has a photo search function.

[0062] It should be noted that because app registration information can contain false information, group information can be unreliable. Therefore, this is referred to as "original group information." In other words, this step divides users with original group information into different groups, each with some common group information. Original group information refers to existing group information that contains unreliable content.

[0063] More specifically, users in the same group have the same original group information. Specifically, for example, the original group information may include at least one of a school, a grade, a class, and a group. The original group information may include, but is not limited to, a school and a grade. In other examples, the original group information may also include a region, a school period, and the like.

[0064] It should be noted that the above description is only given as a preferred example and should not be understood as a limitation of the present invention.

[0065] In the above application scenario, the historical search behavior data of all users is obtained from the education service product APP, and the historical search behavior data is labeled and processed. The historical search behavior data may include search request data and / or search result data, wherein the full set of users includes multiple sets in step S201 (i.e., alternative sets of seed users).

[0066] Optionally, the search request data includes a photo search request, which is a photo search request for a full-page photo. Users in the same group are likely to use the same textbooks and supplementary teaching materials. Based on the full-page photos, the group to which the predicted user belongs can be quickly calculated. Based on this assumption, the more content the photo contains, the higher the prediction speed and accuracy. In some examples, a trained machine learning model can be used to detect and identify whether the search image in each search request data is a full-page test paper or book, or whether it contains more than a specific number of questions.

[0067] Specifically, when the captured image is a whole page of a test paper or a whole page of a book or an image containing more than a specific number of questions, the captured image of the corresponding client can be screened out and normalized.

[0068] In some embodiments, the above-mentioned request normalization is performed on the search image to determine whether the search behavior data of different users are the same request. In particular, if the search behavior data of two or more users are the same request, the number of times each user made the same request within a specific time period is determined, and the probability that each user belongs to the same set (user group) is calculated, which is used for subsequent deduplication of the search result data after the annotation processing.

[0069] For example, if two people's image search records hit the same requirement once, then there is a probability p that the two people are in the same class. When the two people's image search records hit the same requirement n times, when calculated according to probability, the probability that the two people are in the same class is 1-(1-p)^n, so the larger n is, the greater the probability of being in the same class.

[0070] Therefore, a photo search image of a whole page of a test paper or a whole page of a book is selected as a search request, and the photo search image (i.e., the search request) or the photo search result is labeled with textbooks, test papers, books, exercise books, page numbers, test questions, etc. The labeled search results are then converted into vectors and used to characterize the photo search behavior of each user. This allows for more accurate photo search behavior feature data to be obtained, so that the user's group information can be more accurately predicted based on the similarity of the photo search behavior.

[0071] It should be noted that because the same question appears in many different textbooks, exercise books, and test papers, that is, the same needs of different users (searching for the same question) cannot particularly well express the similarity between the two users, so in the search demand scenario screening, this embodiment gives priority to using full-page search images rather than targeting a single question, thereby being able to more accurately determine the group information of each user.

[0072] In addition, for the selection of full page test paper or full page book, the above is only described as a preferred example and cannot be understood as limiting the present invention. In other examples, half page images, or images with more than a certain number of questions can also be used to search.

[0073] It should be noted that the tagging process is to label (or identify) the search request data and / or search result data. Figure 2 The annotation process (ie, labeling process or identification process) is described in detail.

[0074] like Figure 2 As shown, the method of the present invention further includes: labeling the search results of the photo search behavior and converting the labeled search results into feature vectors in step S202. Subsequently, in steps 203 and 204, the similarity of the user's photo search behavior is calculated based on the feature vectors.

[0075] It should be noted that step S202 of tagging search results only needs to be completed before calculating the similarity of the photo-search behavior, and the specific implementation node is not limited.

[0076] In some embodiments, step S202 includes:

[0077] Step 1: Label the search results of the photo search behavior to form a label sequence corresponding to the photo search behavior.

[0078] This step obtains the user's photo-search behavior history, and the data representation of the photo-search behavior history can be in the form of a digital tag sequence. Exemplarily, this step labels the search results corresponding to the historical photo-search behavior obtained (i.e., the processing process of identifying tags) to form a tag sequence (i.e., represented as a tag sequence), and each tag in the tag sequence represents at least one feature of the search result. The photo-search behavior refers to the behavior of initiating a photo-based search request to obtain a search result.

[0079] Step 2: Convert the label sequence into a vector sequence. This step vectorizes the label sequence into a vector sequence.

[0080] Step 3: Normalize the vector sequence into a feature vector (in this example, a photo search behavior feature vector) to characterize each user's photo search behavior. This step calculates the vector sequence to obtain a unique vector (subsequently used as the user's photo search behavior feature vector) with the same dimension as the vector in the previous vector sequence.

[0081] Converting the tag sequence into a vector sequence refers to vectorizing (or converting) each tag in the tag sequence into a vector, forming multiple vectors corresponding to each tag, i.e., forming a vector sequence; normalization refers to converting the vector sequence after the vectorization of each tag into a single vector, i.e., the aforementioned photo search behavior feature vector. This feature vector (e.g., a feature vector or feature vector matrix of a specific dimension) is used to subsequently calculate the photo search behavior similarity between two users.

[0082] To facilitate understanding, some specific examples are given below. For example, if a user has searched 10 times, there will be 10 corresponding search results. Each search result can be represented by tid or pid. Tid is the question id, and pid is the picture id of the page. Then, the tid and pid representing the photo search behavior are vectorized. This user has searched 10 times, so it is equivalent to a tid / pid sequence of length 10. Because the tid and pid have been vectorized, the vectorized sequence can also be subjected to a series of calculations (i.e. normalization) to become a vector. At this point, a user is represented by a vector. In summary, in the photo search scenario, the overall idea of step S202 is to vectorize the user or the user's photo search behavior. The method is to vectorize each user's photo search behavior, and then perform some normalization calculation on the behavior vectorization sequence to become a vector, so that the behavior sequence = user, so that the user becomes vectorized, which is convenient for the subsequent prediction of user groups based on the similarity of photo search behavior.

[0083] The tags generated by the step of labeling the search results of the photo search may include at least one of the following: textbook, test paper, book, exercise book, page number, and test question. Furthermore, the tags of textbook, test paper, book, exercise book, page number, and test question are annotated (or marked) to form a material mark, a test paper mark, a book mark, an exercise book mark, a page number mark, and a test question mark.

[0084] Optionally, the search results corresponding to the historical photo search behaviors of each user within a specific time period are labeled (labeled) to form a label sequence corresponding to the historical search results of each user, and the labeled photo search behavior result feature data (i.e., a label sequence including at least one label identifier among the above identifiers) is obtained, and the label sequence is vector-converted to generate a photo search behavior feature vector to characterize the photo search behavior of each user, and used for subsequent calculation of the photo search behavior similarity between two users.

[0085] For example, the tag sequence includes multiple information strips corresponding to the photo search image (i.e., photo search request data), each information strip is represented by an identification code, wherein the information strip is, for example, a question without an answer to be retrieved, a question containing an answer, etc.

[0086] For another example, the label sequence includes multiple information pieces corresponding to the question search result data, wherein the question search result data includes a test question identifier (represented by TID) and / or a teaching material identifier (represented by PID). Specifically, within one month, the question search results of user 1 include A1:1+2=3, p2, A2:15-2=13, A4, p5…An=10-2=8. After labeling the question search results of user 1, the generated label sequence of user 1 is (A1, p2, A2, A4, P5…An). In this example, it is a row of label sequence, but it is not limited to this in specific implementation. The label sequence may change according to the acquisition time period or the user's search behavior.

[0087] The labeled search results are converted into feature vectors, the information items in the label sequence of each user are vectorized, and the vectors of each information item in the label sequence are converted into vectors of specific dimensions.

[0088] For vector conversion (or vectorization), in this example, the fastText method is used, and the skip-gram algorithm is adopted to establish a vector conversion model, and the vector conversion model is used to vectorize the user's tag sequence.

[0089] Specifically, an embedded learning method is used, and a training data set is used to train the vector conversion model. The hidden layer of the neural network can output a vector of fixed dimension for each annotation (identification code or label identifier), wherein the training data set includes a label sequence of historical users (a label sequence formed by annotating the photo search results), a vector corresponding to each identification code in the label sequence, and a feature vector of historical photo search behavior.

[0090] In this example, vectorized encoding can be performed using the average pooling method, that is, the vectors of a specific dimension of all annotations (identification codes or label identifiers) are averaged to obtain a search behavior feature vector used to characterize the user's search behavior.

[0091] Optionally, the user can also be represented as a feature vector of the user's photo-search behavior (within a specific time period), where the name of the feature vector is defined as the user identifier, and the length of the feature vector (i.e., the specific vector dimension) is defined as the characteristic description of the user's photo-search behavior.

[0092] It should be noted that in this example, the search results are tagged, but this is not limited to this. In other examples, the search behavior data or the search result data and the search behavior data are also tagged. The above is only an optional example and should not be understood as a limitation of the present invention.

[0093] Preferably, the search result data (and / or search behavior data) after annotation processing is deduplicated, and then the label sequence corresponding to the search result data (and / or search behavior data) after deduplication processing is vector-converted to generate a feature vector such as the above-mentioned search behavior feature vector.

[0094] Therefore, through step S202, a label sequence corresponding to the photo search behavior is obtained, and vector conversion is performed to further generate feature vectors such as photo search behavior feature vectors, which can more accurately characterize the user's photo search behavior, obtain more accurate photo search behavior data, and can also be used for subsequent similarity calculations between users.

[0095] It should be noted that the above description is only provided as an example and should not be construed as limiting the present invention.

[0096] Next, in step S203, seed users whose confidence level of the group information is greater than a first preset value are screened out based on the similarity level of the users' photo-searching behaviors.

[0097] Optionally, clustering is performed using the user's photo search behavior feature vector obtained in step S202 to form multiple classes (also called clusters), each of which represents a subset of users with similar photo search behavior. The users included in the candidate seed user set in step S201 are all users, or in other words, all users. Based on the similar photo search behavior of users in the same real group, it can be determined that the confidence level of users in the largest class of the same set (e.g., users registered in the same class) is higher.

[0098] Specifically, we cluster the feature vectors of all users' photo and search behaviors to determine the confidence level of each user's group information. This allows us to identify seed users, defined as users whose original group information has a confidence level greater than a first preset value. For users with a confidence level greater than the first preset value, we consider the original group information to be credible.

[0099] Furthermore, the similarity level of the photo search behavior between two users among all users is calculated, that is, the similarity of the photo search behavior feature vectors of any two users is calculated, and clustering is performed based on the calculated photo search behavior similarity to obtain multiple sets (that is, the sets corresponding to the original groups).

[0100] Optionally, a community discovery algorithm is used to perform clustering processing on each of the multiple sets obtained by the above method (or each set obtained in step S201 ), specifically, clustering processing is performed on the users in each set.

[0101] Specifically, the community discovery algorithm includes defining a user space, in which each user is a vertex (e.g. Figure 3 1 to 10 in the user list), the similarity relationship between adjacent users is an edge (e.g. Figure 3 d1~d10), the similarity of the feature vectors of adjacent users’ photo-search behavior is the weight of the edge to form a user relationship network, which can be seen in Figure 3 .

[0102] Figure 3 It is a schematic diagram showing an example of obtaining a seed user set by clustering processing in the method of embodiment 1 of the present invention.

[0103] like Figure 3 As shown, the Louvain community discovery algorithm is used to cluster users in a set (e.g., the set represented by circles in the figure) to obtain the largest cluster (the set formed by the relationship networks of users 1, 2, 3, 5, 7, 9, and 10 in the figure). The users in the largest cluster obtained after clustering are used as seed users, and the largest cluster (i.e., users 1, 2, 3, 5, 7, 9, and 10 in the figure) is used as the seed user set. Other users (e.g., users 4, 8, and 6) have original group information that is accurate or has low confidence, and are therefore referred to as non-seed users.

[0104] Furthermore, all the sets are clustered in sequence to screen seed users until all the sets have completed the screening of seed users.

[0105] In this way, the seed user set and non-seed users can be determined more accurately.

[0106] Preferably, a seed user database is established using the obtained seed users, seed user sets and user IDs (ie user accounts or user identification codes) of various sub-users, and the user database is queried using a feature search engine to predict the groups or group features to which non-seed users belong.

[0107] Optionally, the seed user database also includes user IDs of various sub-users and feature vectors represented by test paper identification information and / or textbook identification information.

[0108] It should be noted that the above description is only provided as an example and should not be construed as limiting the present invention.

[0109] Next, in step S204, for non-seed users, the similarity between the shooting and searching behaviors of the non-seed user and various sub-users is calculated, and the group (or group characteristics) to which the non-seed user belongs is predicted based on the similarity. The non-seed users include users who lack the group information and users whose confidence in the group information in the set is not greater than a first preset value.

[0110] Specifically, the question search request data and / or question search result data of the user to be predicted are obtained, and the question search result data includes a test question identifier and / or a teaching material identifier.

[0111] Furthermore, the non-seed user is characterized as a photo-search behavior feature vector through the test question identifier and / or teaching material identifier (ie, photo-search behavior), and the user to be predicted is determined to be a non-seed user.

[0112] In this example, the similarity between the photo-search behavior of the user to be predicted (i.e., the non-seed user) and various sub-users is calculated respectively. Specifically, the distance between the photo-search behavior feature vector of the user to be predicted (i.e., the non-seed user) and the photo-search behavior feature vector of the seed user in the seed user database is calculated respectively (as the similarity). For example, in the user space, the distance between the non-seed user and various sub-users is calculated as the similarity to predict the user group to which the non-seed user belongs.

[0113] Specifically, a feature search engine is used to search the seed user database, and seed users whose similarity with the user to be predicted (i.e., the non-seed user) is within a predetermined range are screened out. When predicting the group to which the non-seed user belongs, the group to which the non-seed user belongs or the group characteristics are predicted based on the group information of the screened seed users.

[0114] Furthermore, screening out seed users whose similarity with the non-seed users is within a predetermined range includes: calculating an average distance between any two seed users in the largest class in each set, and screening out seed users whose distance between the non-seed user and the seed users in each set is less than the average distance.

[0115] Therefore, the feature-based search engine searches the seed user database and returns a search result, which is a list of seed users whose similarity is less than a predetermined range, and the list of seed users includes a specific number of seed users.

[0116] Next, the number of screened seed users (i.e., the seed user list) is counted according to the sets corresponding to their original group information, and the set with the largest number of corresponding seed users is used as the predicted group or group feature to which the non-seed user belongs, or the group feature of the non-seed user is predicted as the group information to which the seed user belongs.

[0117] Optionally, when the number of users included in the cluster (corresponding set) to which the non-seed user belongs is within a predetermined range and the cluster (corresponding set) contains a seed user, the group to which the non-seed user belongs is predicted as the group to which the seed user belongs, or the group characteristics of the non-seed user are predicted as the group information to which the seed user belongs.

[0118] Preferably, when the regional information is strongly correlated with the group information, for example, when predicting a user's class, the relationship between the user's region and the user's class is relatively sticky. Before counting the number of seed users after screening according to their original group information, the seed users can be screened according to their regional information to eliminate seed users that are not in the same region as the non-seed users. Subsequently, the seed users (i.e., the seed user list) after eliminating the seed users that are not in the same region are counted according to the sets corresponding to their original group information, and the set with the largest number of corresponding seed users is used as the predicted group or group feature to which the non-seed user belongs.

[0119] In addition, regional information-based screening can also be performed before similarity calculations. First, users in the collection who are not in the same region are eliminated. Then, a feature search engine is used to search the seed user database. Similarity is calculated between the feature vectors of non-seed users and the feature vectors of seed users in the seed user database to screen out seed users whose similarity is within a predetermined range. The original group with the largest number of corresponding seed users is used as the predicted group to which the non-seed user belongs. This effectively avoids the problem of slow data processing speed caused by large amounts of user data, enables more accurate prediction of the group to which the non-seed user belongs, further improves prediction accuracy, and further optimizes the method.

[0120] It should be noted that the above description is only provided as an example and should not be construed as limiting the present invention.

[0121] Figure 4 3 is a flowchart showing another example of the method for predicting the group to which a user belongs based on photo-search behavior according to embodiment 1 of the present invention.

[0122] like Figure 4 As shown, the step S403 of denoising the seed users is also included, wherein, since the steps S401, S402 and S404 are respectively Figure 1 Step S201, step S203 and step S204 are the same, so the description of step S401, step S402 and step S404 is omitted. Step S403 will be described in detail below.

[0123] In step S403, denoising is performed on the seed users to generate a more accurate set of seed users.

[0124] Specifically, the user relationship network graph includes user vertices and edge weights. The user relationship network graph has users as vertices, similar relationships between users as edges, and the similarity of the feature vectors of two users (in this example, the photo search behavior feature vectors) as edge weights.

[0125] If a user fills in their information, they may be placed in a cluster based on the information they filled in, making them a candidate for a seed user. However, this information may be unreliable, and we need to filter out noise points such as random information and expired information. Therefore, we use a community discovery algorithm to filter out noise points. A seed user is a node in the largest cluster in the community discovery algorithm. If a user does not fall into this largest cluster, then even though they have filled in their information, they are not a seed user.

[0126] Based on this, for users in the same original group, most user vertices will be divided into the same community, and noise points will be divided into other small communities. Thus, the noise points (i.e., non-seed users) are determined and removed to obtain the final seed user set.

[0127] In another example, after clustering users in the same original group, seed users in the largest cluster whose distance from the center point is greater than a specific distance are removed. In other words, outliers (i.e., non-seed users) whose distance from the center point is greater than a specific distance are removed to generate the final seed user set.

[0128] In another example, a judgment is made based on the regional information of each user in the largest cluster set, and seed users in the area that does not belong to the largest cluster set are removed to generate a final seed user set.

[0129] Specifically, for example, non-seed users include users whose class information or grade information is not true, users whose school address and area information are inconsistent, and so on.

[0130] Therefore, by removing non-seed users from each original group, a more accurate set of seed users (ie, user group categories) can be obtained, and the prediction accuracy can be improved.

[0131] The above method for predicting user groups based on photo and search behavior is intended only to illustrate the present invention. The order and number of steps are not particularly limited. Furthermore, the steps in the above method can be split into two or three steps, or some steps can be combined into a single step, depending on the actual example.

[0132] Compared to existing technologies, this method, based on the similarity of users' photo-search behavior, selects seed users from each set whose group information has a confidence level greater than a first preset value. For non-seed users in the group to be predicted, the method first calculates the photo-search behavior similarity between the non-seed user and various sub-users. Then, based on this similarity, it predicts the group to which the non-seed user belongs. These non-seed users include users with missing group information and users whose group information in the set has a confidence level no greater than a first preset value. This allows for more accurate prediction of the group to which non-seed users belong, further improving prediction accuracy and achieving a simple and efficient algorithm.

[0133] Furthermore, the present invention uses the search results of the photo search behavior to perform vector conversion and use it to characterize the user vector, so as to obtain more accurate photo search behavior data and more accurately characterize the user's group information; by removing non-seed users from each original group, a more accurate seed user set (i.e., user group category) can be obtained, and the prediction accuracy can be improved; and by using a feature search engine to search the seed user database, the user feature vectors of the non-seed users are similar to the user feature vectors of the seed users in the seed user database, so as to screen out the seed users whose similarity is within a predetermined range, and use the original group with the largest number of corresponding seed users as the predicted group to which the non-seed user belongs, which can effectively avoid the problem of slow data processing speed caused by the large amount of user data, can more accurately predict the group to which the non-seed user belongs, can further improve the prediction accuracy, and can further optimize the method.

[0134] Example 2

[0135] The following describes an apparatus embodiment of the present invention, which can be used to perform the method embodiment of the present invention. Details described in the apparatus embodiment of the present invention should be considered supplementary to the above-described method embodiment; details not disclosed in the apparatus embodiment of the present invention can be implemented with reference to the above-described method embodiment.

[0136] Reference Figures 5 to 7 , a prediction device 500 for predicting a group to which a user belongs based on photo-searching behavior according to embodiment 2 of the present invention will be described.

[0137] According to a second aspect of the present invention, the present invention further provides a prediction device 500 for predicting a group to which a user belongs based on photo search behavior, wherein the photo search behavior refers to the behavior of initiating a photo search request and obtaining search results.

[0138] Specifically, the prediction device 500 includes: a grouping module 501, which is used to divide users into different groups so that users in the same group have the same group information, and the group information is related to the user's photo search behavior; a screening module 502, which screens out seed users in each group whose confidence in the group information is greater than a first preset value based on the similarity level of the user's photo search behavior; a prediction module 503, which, for non-seed users, calculates the similarity between the photo search behavior of the non-seed user and various sub-users, and predicts the group to which the non-seed user belongs based on the similarity, and the non-seed user includes users who lack the group information and users in the group whose confidence in the group information is not greater than the first preset value.

[0139] like Figure 6As shown, it also includes a clustering module 601, which is used to label the search results of the photo search behavior before screening the seed users, and convert the labeled search results into a feature vector to characterize the user's photo search behavior; subsequently, the similarity of the user's photo search behavior is calculated based on the feature vector.

[0140] Specifically, the search results of the photo search behavior are labeled, and the labeled search results are converted into feature vectors, including: obtaining the user's historical photo search behavior, and representing the historical photo search behavior as a label sequence based on the corresponding search results, each label in the label sequence represents at least one feature of the search result; converting the label sequence into a vector sequence; and normalizing the vector sequence into the feature vector.

[0141] Optionally, the method further includes: characterizing the user as a feature vector of his / her photo-searching behavior: defining the name of the feature vector as the user identifier, and defining the length of the feature vector as a feature representation of the user's photo-searching behavior.

[0142] Optionally, the historical search behavior is characterized as a tag sequence according to the corresponding search results, including: labeling the search results; and deduplicating the search result data after labeling.

[0143] Optionally, the photo search behavior refers to the behavior of initiating a photo-based search request to obtain a search result.

[0144] Optionally, the photo is a full-page photo; and the tag sequence includes: test questions and book pages.

[0145] Optionally, the group information includes at least one of the following: the user's region, school, grade, class, or group.

[0146] Furthermore, the tags generated in the step of tagging the search results of the photo search behavior include at least one of the following tags: textbook, test paper, book, exercise book, page number, and test question.

[0147] Furthermore, the clustering module 601 is further configured to perform clustering processing on the feature vectors; and use the user in the largest cluster obtained after the clustering in the set as the seed user.

[0148] Optionally, a community discovery algorithm is used within each set to perform the clustering process on the users in the set.

[0149] Optionally, a community discovery algorithm is used within each set to cluster the users in the set, and the users in the largest cluster in the obtained set are the seed users.

[0150] Specifically, the use of the community discovery algorithm includes defining a user space, in which each user is a vertex, the similarity relationship between adjacent users is an edge, and the similarity of the feature vectors of the adjacent users' shooting and search behavior is the weight of the edge. When the seed users whose confidence level of the group information is greater than a first preset value are screened out, the Louvain community discovery algorithm is used to cluster the users in each set to obtain the seed users.

[0151] Preferably, the obtained seed users, seed user sets and user IDs (i.e. user accounts or user identification codes) of various sub-users are used to establish a seed user database, and the user database is queried using a feature search engine to predict the groups or group features to which non-seed users belong.

[0152] Each set of seed users corresponds to a user group, the seed users include school labels and grade labels, and the seed users are users whose original group information has a confidence level greater than a first preset value.

[0153] like Figure 7 As shown, it also includes a calculation module 701, which is used to calculate the similarity between the shooting and searching behaviors of non-seed users and various sub-users in, for example, a seed user database, and screen out seed users whose similarity with the non-seed users is within a predetermined range. When predicting the group to which the non-seed user belongs, the group to which the non-seed user belongs is predicted based on the group information of the screened seed users.

[0154] Furthermore, the calculation module 701 further includes: calculating the distance between the photographing and searching behavior feature vector of the non-seed user and the photographing and searching behavior feature vectors of the various sub-users as the similarity.

[0155] Specifically, seed users whose similarity with the non-seed users is within a predetermined range are screened out from the seed user database; when predicting the group to which the non-seed user belongs, the group to which the non-seed user belongs or the group characteristics are predicted based on the group information of the screened seed users.

[0156] Furthermore, screening out seed users whose similarity with the non-seed users is within a predetermined range includes: calculating an average distance between any two seed users in the largest class in each set, and screening out seed users whose distance between the non-seed user and the seed users in each set is less than the average distance.

[0157] Next, the number of screened seed users is counted according to their original group information, and the original group with the largest number of corresponding seed users is used as the predicted group to which the non-seed user belongs; when the number of users included in the cluster to which the non-seed user belongs is within a predetermined range and the cluster contains a seed user, the group to which the seed user belongs is predicted to be the group to which the non-seed user belongs.

[0158] Preferably, before counting the number of the screened seed users according to their original group information, the method further includes: screening the screened seed users according to their region information, and eliminating seed users that are not in the same region as the non-seed users.

[0159] It should be noted that, in Example 2, descriptions of the same parts as in Example 1 are omitted.

[0160] Those skilled in the art will appreciate that the modules in the above device embodiments may be distributed in the device as described, or may be modified accordingly and distributed in one or more devices different from the above embodiments. The modules in the above embodiments may be combined into one module or further split into multiple submodules.

[0161] Compared to existing technologies, this method, based on the similarity of users' photo-search behavior, selects seed users from each set whose group information has a confidence level greater than a first preset value. For non-seed users in the group to be predicted, the method first calculates the photo-search behavior similarity between the non-seed user and various sub-users. Then, based on this similarity, it predicts the group to which the non-seed user belongs. These non-seed users include users with missing group information and users whose group information in the set has a confidence level no greater than a first preset value. This allows for more accurate prediction of the group to which non-seed users belong, further improving prediction accuracy and achieving a simple and efficient algorithm.

[0162] Furthermore, the present invention uses the search results of the photo search behavior to perform vector conversion and use it to characterize the user vector, so as to obtain more accurate photo search behavior data and more accurately characterize the user's group information; by removing non-seed users from each original group, a more accurate seed user set (i.e., user group category) can be obtained, and the prediction accuracy can be improved; and by using a feature search engine to search the seed user database, the user feature vectors of the non-seed users are similar to the user feature vectors of the seed users in the seed user database, so as to screen out the seed users whose similarity is within a predetermined range, and use the original group with the largest number of corresponding seed users as the predicted group to which the non-seed user belongs, which can effectively avoid the problem of slow data processing speed caused by the large amount of user data, can more accurately predict the group to which the non-seed user belongs, can further improve the prediction accuracy, and can further optimize the method.

[0163] Example 3

[0164] This embodiment takes the class prediction of an online education app user as an example and provides a method for predicting a user's class based on the above-mentioned method based on the photo search behavior. The method mainly includes the following steps:

[0165] Step 1. Unify the annotation of user's photo search behavior. The annotation is defined as the normalization of user photo search needs, that is, the user's search behavior for the same question is regarded as the same need (different users will upload different photo search needs for retrieval); the annotation method is that because the same question will appear in many different textbooks, exercise books, and test papers, that is, the same needs of different users (searching for the same question) cannot particularly well express the similarity between the two users. Therefore, the whole page photo search need is used for demand scenario screening, because the whole page photo search user takes a photo to search for a whole page of test questions, not just a single question. If different users upload the same whole page of test questions, then other users will not be able to express the similarity between the two users. The probability of using the same textbook, workbook, or test paper can reach over 80%. If multiple photos and records all hit the same full page of test questions, then this probability will increase very quickly when calculated as independent events, and the confidence level will be very high. Most of the resources hit by full-page photo search requirements come from textbooks, workbooks, and test papers. We uniformly record the TID (test question) and PID (book page) of these resources as keys. After full normalization, we can obtain a finite set of keys consisting of TIDs and PIDs, which is the tag set. Then, user search behavior can be represented as a tag sequence, where each element in the sequence represents a specific tag key.

[0166] Step 2: Digitally encode the KEY. The finite set of all KEY codes can be regarded as a complete dictionary. The elements in the dictionary can be regarded as words. The label sequence is an article composed of words arranged in different orders. Therefore, each user can be regarded as an independent article. The digital codes in the dictionary can be vectorized through the embedded learning method. The embedded learning method uses the fasttext tool and the skip-gram algorithm. The input is a multi-line annotation sequence, each line is an article, and each article is composed of words arranged in different orders. The model itself is a shallow neural network. After training, the hidden layer of the neural network can output a fixed-dimensional vector for each annotation (word).

[0167] Step 3: Through encoding vectorization, the article (tag sequence) can also be vectorized and encoded using the average pooling method. Ultimately, each user can be represented by a vector, and the vector cosine similarity of users with similar search behaviors is also large. In actual calculations, we use one minus the vector cosine similarity to characterize the similarity of the photo search behavior between any two users. The more similar the photo search behavior, the smaller the similarity metric. For users in the same class, the similarity metric between any two users is likely to be very small (because users in the same class searched for very similar textbooks, workbooks, and test papers during the same time period).

[0168] Step 4: A considerable number of users have filled in school and grade information on the Zuoyebang app, defining the class as the same grade in the same school. This means that the user has filled in the class information. However, different users may fill in different information for the same school name. We sorted out the names and regional information of more than 300,000 primary schools, junior high schools, and high schools across the country, normalized the school names, and mapped them to the normalized school names based on the province, city, and school information filled in by the user.

[0169] Step 5. Define users as fixed points, user similarity relationships as edges, and vector similarity as edge weights. Any user can calculate the similarity of their photo search behavior with all other users. Zuoyebang has hundreds of millions of monthly active users. If the similarity of user photo search behavior is calculated pairwise, the calculation scale is on the order of billions. By placing users who have filled in their class information into corresponding class sets, we can obtain millions of class sets. Then, within each set, we use the Louvain community discovery algorithm to cluster users. The largest cluster in the set is considered reliable, and users not in the largest cluster are removed from the set. The input of the Louvain algorithm is the vertices and edge weights in the graph, and the output is the community set composed of vertices. In our scenario, most vertices will be divided into one community, and noise points will be divided into other small communities.

[0170] Step 6: Users in the largest cluster of a class are called seed users. Untrusted users who have not filled in the school label or have been excluded are called unknown class users. A feature search engine is used to build a database for all seed users. The average distance (similarity measure) between any two users in each largest cluster is calculated. The feature vector engine can be used to calculate the set of seed users with a distance less than the average distance for each unknown class user. The feature engine uses the Faiss tool. When building the database, the input is the ID and feature vector of each database element. When searching the feature engine, parameters can be set to return the ID of the element in the database that is less than the specified distance, as well as the specific distance (similarity) to the element.

[0171] Step 7: For each unknown class user, a set of similar seed users can be calculated. Based on the geographical information of the unknown class user, similar seed users from non-local schools in the set are filtered out as noise points. Then, votes are counted according to the seed user's class label, and the class with the most recalled votes is regarded as the label of the unknown class user. In actual operation, a certain proportion of unknown class users will have an empty set of similar seed users (too few photo search behaviors or new users), and the prediction accuracy for these users is low.

[0172] The class prediction method based on the similarity of users' photo-search behaviors requires relatively little computation and has high accuracy, making it easier to provide high-quality services to users in the future.

[0173] Example 4

[0174] The following describes a computer device embodiment of the present invention. This computer device can be considered a specific physical implementation of the method and apparatus embodiments of the present invention described above. Details described in the computer device embodiment of the present invention should be considered supplementary to the above-described method or apparatus embodiments; details not disclosed in the computer device embodiment of the present invention can be implemented with reference to the above-described method or apparatus embodiments.

[0175] Figure 8 The computer device of one embodiment of the present invention includes a processor and a memory, wherein the memory is used to store a computer executable program. When the computer program is executed by the processor, the processor executes Figure 1 method.

[0176] like Figure 8 As shown, the computer device is implemented as a general-purpose computing device. The processor may be one or multiple processors working in concert. The present invention also does not exclude distributed processing, meaning that the processors may be dispersed across different physical devices. The computer device of the present invention is not limited to a single entity but may also be the sum of multiple physical devices.

[0177] The memory stores a computer executable program, generally a machine-readable code, which can be executed by the processor to enable the computer device to perform the method of the present invention, or at least some steps in the method.

[0178] The memory includes a volatile memory, such as a random access memory unit (RAM) and / or a cache memory unit, and may also be a non-volatile memory, such as a read-only memory unit (ROM).

[0179] Optionally, in this embodiment, the computer device further includes an I / O interface for exchanging data with external devices. The I / O interface can represent one or more of several types of bus structures, including a storage unit bus or storage unit controller, a peripheral bus, an accelerated graphics port, a processing unit, or a local bus using any of a variety of bus structures.

[0180] It should be understood that Figure 8 The computer device shown is merely an example of the present invention. The computer device of the present invention may also include elements or components not shown in the above examples. For example, some computer devices also include display units such as screens, while others include human-computer interaction elements such as buttons and keyboards. As long as the computer device is capable of executing a computer-readable program stored in its memory to implement the method of the present invention or at least some of the steps of the method, it is considered a computer device within the scope of the present invention.

[0181] Figure 9 Schematic diagram of a computer program product according to an embodiment of the present invention. Figure 9 As shown, a computer program product stores a computer executable program, and when the computer executable program is executed, the above-mentioned method of the present invention is implemented. The computer-readable storage medium may include a data signal propagated in baseband or as part of a carrier wave, which carries a readable program code. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. The readable storage medium may also be any readable medium other than a readable storage medium, which may send, propagate, or transmit a program for use by or in combination with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium may be transmitted using any appropriate medium, including but not limited to wireless, wired, optical cable, RF, etc., or any suitable combination of the above.

[0182] The program code for performing the operations of the present invention may be written in any combination of one or more programming languages, including object-oriented programming languages such as Java, C++, and the like, as well as conventional procedural programming languages such as "C" or similar programming languages. The program code may be executed entirely on the user computing device, partially on the user device, as a stand-alone software package, partially on the user computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving a remote computing device, the remote computing device may be connected to the user computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0183] Through the above description of the implementation mode, it is easy for those skilled in the art to understand that the present invention can be implemented by hardware capable of executing a specific computer program, such as the system of the present invention, and the electronic processing unit, server, client, mobile phone, control unit, processor, etc. contained in the system. The present invention can also be implemented by computer software that executes the method of the present invention, such as control software executed by a microprocessor, an electronic control unit, a client, a server, etc. However, it should be noted that the computer software that executes the method of the present invention is not limited to being executed by one or a specific hardware entity, and it can also be implemented in a distributed manner by unspecified specific hardware. For computer software, the software product can be stored in a computer-readable storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.), or it can be distributed and stored on a network, as long as it enables a computer device to execute the method according to the present invention.

[0184] The specific embodiments described above further illustrate the objectives, technical solutions, and beneficial effects of the present invention. It should be understood that the present invention is not inherently related to any specific computer, virtual device, or computer equipment, and various general-purpose devices can also implement the present invention. The above description is merely a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included in the scope of protection of the present invention.

Claims

1. A method for predicting a user's group based on photo search behavior, wherein the photo search behavior refers to the act of initiating a photo search request and obtaining search results, characterized in that: The method comprises the following steps: Based on the fact that users with the same or similar photo-searching behaviors have the same group characteristics, users are divided into different sets, so that users in the same set have the same group information, and an alternative set of seed users is established. The group information is related to the user's photo-searching behavior, and the group characteristics represent the common characteristics of a sub-set of users. The photo-searching image of the search request for the photo-searching behavior is detected and identified, and after normalization, it is determined whether the user's photo-searching behavior is the same request. When two or more users have the same request, the number of times each user has made the same request within a specific time period is determined, and the probability that each user belongs to the same set is calculated. The search results of the photo search behavior are labeled according to the number of times the photo search behavior is performed to form a label sequence corresponding to the photo search behavior, and the labeled search results are converted into a feature vector to represent the user's photo search behavior, specifically including: obtaining the user's historical photo search behavior, and representing the historical photo search behavior as a label sequence based on the corresponding search results, each label in the label sequence represents at least one feature of the search result, converting the label sequence into a vector sequence, and normalizing the vector sequence into the feature vector; calculating the similarity of the user's photo search behavior based on the feature vector; and Search behavior similarity level, defining a user space, wherein each user in the user space is a vertex, similarity relationships between adjacent users are edges, and similarity of feature vectors of adjacent users' search behaviors is an edge weight, and a community discovery algorithm is used within each set to cluster the users in each set to screen out seed users in each set whose confidence level of the group information is greater than a first preset value, wherein screening out seed users in each set whose confidence level of the group information is greater than the first preset value further comprises: clustering the feature vectors and selecting the user in the largest cluster obtained in the set as the seed user; For non-seed users, the similarity between the photo-shooting and search behaviors of the non-seed user and various sub-users is calculated, seed users whose similarity with the non-seed user is within a predetermined range are screened out, and the group to which the non-seed user belongs is predicted based on the similarity, wherein the non-seed users include users who lack the group information and users whose confidence in the group information in each set is not greater than a first preset value; wherein, predicting the group to which the non-seed user belongs based on the similarity includes predicting the group to which the non-seed user belongs or group characteristics based on the group information of the screened seed users.

2. The method according to claim 1, characterized in that Also includes: Represent the user as a feature vector of his / her photo-search behavior: define the name of the feature vector as the user identifier, and define the length of the feature vector as the characteristic description of the user's photo-search behavior.

3. The method according to claim 1 or 2, characterized in that Also includes: The historical search behavior is represented as a label sequence according to the corresponding search results, including: labeling the search results; deduplicating the search result data after labeling; and / or, The photo search behavior refers to the behavior of initiating a search request based on a photo to obtain search results; and / or, The photo is a full-page photo; the label sequence includes: test questions and book pages; and / or, The group information includes at least one of the following: the user's region, school, grade, class, or group; and / or, The tags generated in the step of tagging the search results of the photo search behavior include at least one of the following tags: textbook, test paper, book, exercise book, page number, test question; and / or, Using the community discovery algorithm includes: using the Louvain community discovery algorithm to cluster users in each set to obtain seed users.

4. The method according to claim 1, wherein Calculating the similarity between the photo-searching behavior of the non-seed user and various sub-users includes: In the user space, the distance between the non-seed user and various sub-users is calculated as the similarity.

5. The method according to claim 4, characterized in that The screening out of seed users whose similarity with the non-seed users is within a predetermined range includes: calculating an average distance between any two seed users in the largest class in each set, and screening out seed users whose distance between the non-seed user and the seed users in each set is less than the average distance.

6. The method according to claim 4, characterized in that Predicting the group to which the non-seed user belongs based on the filtered seed user's group information includes: Count the number of seed users after screening according to their group information, and take the group with the largest number of seed users as the predicted group to which the non-seed user belongs; When the number of users included in the cluster to which the non-seed user belongs is within a predetermined range and the cluster includes a seed user, the group to which the seed user belongs is predicted as the group to which the non-seed user belongs.

7. The method according to claim 6, characterized in that Before counting the number of filtered seed users according to their group information, the following steps are also included: Screening is performed based on the region information of the screened seed users, and seed users that are not in the same region as the non-seed users are eliminated.

8. A prediction device for predicting a user's group based on photo search behavior, wherein the photo search behavior refers to the act of initiating a photo search request and obtaining search results, characterized in that: The prediction device comprises: A grouping module divides users into different sets based on the fact that users with the same or similar photo-searching behaviors have the same group characteristics, so that users in the same set have the same group information, and establishes an alternative set of seed users for selection. The group information is related to the user's photo-searching behavior, and the group characteristics represent the common characteristics of a sub-set of users. The photo-searching image of the search request for identifying the photo-searching behavior is detected and normalized to determine whether the user's photo-searching behavior is the same request. If two or more users have the same request, the number of times each user has made the same request within a specific time period is determined, and the probability that each user belongs to the same set is calculated. The screening module labels the search results of the photo-search behavior according to the number of times the photo-search behavior is performed, forming a label sequence corresponding to the photo-search behavior, and converting the labeled search results into a feature vector to represent the user's photo-search behavior. Specifically, the module includes: obtaining the user's historical photo-search behavior, and representing the historical photo-search behavior as a label sequence based on the corresponding search results, each label in the label sequence represents at least one feature of the search result, converting the label sequence into a vector sequence, and normalizing the vector sequence into the feature vector; calculating the similarity of the user's photo-search behavior based on the feature vector, Furthermore, based on the similarity level of the users' photo-searching behaviors, a user space is defined, wherein each user is a vertex, the similarity relationship between adjacent users is an edge, and the similarity of the feature vectors of the photo-searching behaviors of adjacent users is the weight of the edge. The users in each set are clustered using a community discovery algorithm within each set to screen out seed users whose confidence level of the group information is greater than a first preset value, wherein screening out seed users whose confidence level of the group information in each set is greater than the first preset value further includes: clustering the feature vectors; and selecting the user in the largest cluster obtained in the set after the clustering as the seed user; The prediction module calculates the similarity between the photo-shooting and search behaviors of the non-seed user and various sub-users, screens out seed users whose similarity with the non-seed user is within a predetermined range, and predicts the group to which the non-seed user belongs based on the similarity, wherein the non-seed user includes users who lack the group information and users whose confidence in the group information in each set is not greater than a first preset value; wherein, predicting the group to which the non-seed user belongs based on the similarity includes predicting the group to which the non-seed user belongs or group characteristics based on the group information of the screened seed users.

9. A computer device comprising a processor and a memory, wherein the memory is used to store a computer executable program, wherein: When the computer executable program is executed by the processor, the processor executes the method for predicting the group to which a user belongs based on photo-search behavior as described in any one of claims 1 to 7.

Images