TWI601024B - Sampling methods, systems and equipment - Google Patents

Description

Sampling analysis methods, systems and equipment

The invention relates to the field of computer network technology, and particularly relates to a sampling analysis method, system and device.

The search engine generally records the user's query record. For a large search engine, the user's query record is a large amount of data in a certain period of time, and a large percentage of the user's query keyword is a repeated query, for example, for the most recent popular event. The queries made by the users are similar or even the same. In order to provide better services, the search engine service provider processes the user's query records, and a basic processing step is to merge the same query keywords, which can greatly reduce the memory or disk space occupied by the data storage. . For example, recently there were 2,000 query keywords "Alibaba", then the combined data format is "Alibaba 2000", in which "Alibaba" represents the user query keyword and 2000 represents the query keyword for a period of time. The number of occurrences in the Query Log. However, for such statistical data that has been initially sorted out, how to perform query keyword sampling can make the sampling data close to the true distribution of the query keywords become a problem to be solved.

In the prior art, for the statistics of the "query keyword PV (Page View)" format, firstly, the proportion of each query keyword in all query keywords is calculated, wherein PV represents the query key. Statistics on the number of times a word appears on the search platform. For example, right In the query data of "Alibaba 2000", firstly, the sum of the PV values of all the query keywords in the query keyword set is counted, assuming that the total PV value is 1 million, which means that the number of keywords for all users is 1 million. Then calculate the proportion of the query keyword "Alibaba" in all query keywords. It can be seen that the ratio is 2000/1000000=0.0025. The meaning of this data is: In all query keywords, the key of "Alibaba" is the query. The probability that the word is randomly drawn is 0.0025. After the calculation result of the extraction probability of all the query keywords is obtained, the query keyword may be sampled in the set of all the query key words according to the extraction probability of a certain query keyword, thereby obtaining the final sampling data of the corresponding query keyword. Through the analysis of the sample data, the distribution of the user's query keywords is known. For example, in a query keyword set with a total PV value of 1 million, it is estimated that 10,000 query records are extracted as query keyword samples for analysis. The specific query keyword sampling process is as follows: the sampling number of the query keyword is determined according to the extraction probability of a query keyword, that is: [the number of samples of a query keyword] = [the expected number of samples] * (the key of the query Word extraction probability), wherein the number of samples of the query keyword and the expected number of samples are positive integers. For example, if the probability of the query keyword “Alibaba” is randomly selected is 0.0025, then 10000*0.0025=25 “Alibaba” query keywords are extracted as the query keyword sample in the query record “Alibaba 2000”. Similarly, the number of other query keywords that are sampled can be obtained according to the above formula; the sum of the sample numbers of all query keywords is 10,000. Compared with 1 million query records, 10,000 sample query records are analyzed and processed, and the data analyst's workload and calculation steps are greatly reduced, which improves work efficiency.

In the process of implementing the present invention, the inventors have found that at least the following problems exist in the prior art: if the number of pieces of data to be extracted is large, the sampling analysis method in the prior art can sample data to some extent analogous to the true data distribution, but When the number of extractions is medium or small, the extraction result will have a large distortion with the true distribution of the data. The reason is that in the statistics of statistics, the statistical distribution of many data has long tail characteristics. The so-called long tail is the number of entities with very low frequency of occurrence or the number of data, especially when users search for keywords through search engines. Many users query keywords only a few times, for example, the keywords of some queries only appear once or twice. Although the probability of occurrence of a keyword is very low, the total number of low-frequency query keywords is a large percentage of the total number of query keywords. For the case of such a long tail distribution, if the sampling analysis method in the prior art described above is adopted, the low frequency query keyword cannot be extracted. For example, the goal of an application is to extract 2000 query keywords, in which the total number of query PVs is 1 million. For a query keyword, the probability of being extracted is only 1 million for "e-commerce 1". One, so using the above method is not able to extract low frequency query keywords. This kind of data extracted according to the existing sampling analysis method is very different from the distribution of real data, so that it is impossible to accurately understand the user demand information and market dynamics according to the sampling analysis of the query keywords in the search engine, and thus it is not good for the user. Provide convenient and fast e-commerce online trading services.

The embodiment of the invention provides a sampling analysis method, system and device for analyzing data of a large-scale search engine query, so as to obtain real data sampling while using as little storage space as possible, and accurately understand user demand information. And market dynamics to improve service quality.

To achieve the above objective, an embodiment of the present invention provides a sampling analysis method for analyzing data of a large-scale search engine query, including the following steps: dividing a query keyword according to a query record PV value of different query keywords. Having at least one query keyword subset; calculating a sample number of the query keyword subset; extracting query data in the query keyword subset according to the sample number.

Another aspect of the present invention provides a sampling analysis device for analyzing data of a large-scale search engine query, including: a dividing module, configured to divide a query keyword according to a query record PV value of different query keywords. a sampling module for calculating a sample number of the subset of query keywords divided by the dividing module; a sampling module for obtaining according to the computing module The number of samples is used to extract query data in the subset of query keywords divided by the partitioning module.

On the other hand, the embodiment of the present invention further provides a sample analysis system for analyzing data of a large-scale search engine query, including: a search platform for providing a search service for a user query, and recording PV values of different query keywords; a sampling analysis device, configured to divide a query keyword into at least one query keyword subset according to a PV value of different query keywords recorded by the search platform, and calculate a sample number of the query keyword subset, according to the The number of samples is used to extract query data in the subset of query keywords.

Compared with the prior art, the embodiment of the present invention has the following advantages: the required query record can be randomly extracted from a large number of query keyword sets that have undergone preliminary statistical sorting, which can reduce the storage amount required for subsequent calculations, and It can solve the problem that the low-probability low-frequency query keywords in the existing sampling analysis methods are underestimated, and effectively achieve the purpose of randomly extracting the query records, so that the medium-sized or small-scale sampling can be closer to the true distribution of the data; the search engine service provider According to the sampling data, we can establish a perfect mathematical model, obtain real and effective data distribution information, accurately understand user needs and market dynamics, and appropriately adjust the service content of the search engine to better provide users with convenient and fast e-commerce online. Trading platform to improve service quality.

The technical solutions in the embodiments of the present invention are clearly and completely described in the following with reference to the accompanying drawings in the embodiments of the present invention. It is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, those of ordinary skill in the art are not doing All other embodiments obtained under the premise of creative labor are within the scope of the invention.

The first embodiment of the present invention provides a sampling analysis method for analyzing data of a large-scale search engine query, which specifically includes the following steps:

Step S101: Divide the query keyword into at least one query keyword subset according to the PV value of the different query keywords.

The PV value is specifically the number of times at least one query keyword appears on the search platform within a preset time period. Before performing the sampling analysis, first store the PV values of all the user query keywords recorded by the search platform in a time period, and sort the PV values of the different query keywords, and the sorting manner may be from small to large, or From large to small, the query keywords with the same PV value are then classified into a subset of query keywords.

Step S102: Calculate the number of samples of the query keyword in the query keyword subset.

Before sampling the query keywords for all the sets of query key phrases, first determine the number of sample analysis query keywords according to the application needs. Specifically, the SPV (Set Page View) value of the PV value in each query keyword subset is calculated first, and the SPV refers to the total PV number of a certain query keyword subset; The sum of the SPV values obtained by the query keyword subsets is obtained, and the total number of times of the web page browsing in the query keyword set is TPV (Total Page View) value, that is, all users query the key within a preset time period. The total number of times the word appears on the search platform; based on the obtained ratio of the SPV value to the TPV value, the probability that the query keyword subset is extracted can be calculated. Then, the number of samples of a certain subset of the query keywords is calculated according to the predetermined number of query keywords that need to be extracted and the extraction probability of each query keyword subset.

Step S103: Extract query data of the query keyword in the subset of query keywords according to the number of samples of the query keyword.

The query data is extracted in the query keyword subset by a random sampling method, wherein the random sampling method includes a lottery method and/or a random number method. Using the query data of the randomly selected query keywords, the distribution of different query keywords searched by the user in a period of time can be analyzed, so that the user's demand information can be understood.

Through the embodiment of the present invention, the required query data can be randomly extracted from a large number of query keyword sets that have undergone preliminary statistical sorting, which can reduce the storage amount required for subsequent calculation, and can solve many methods of small probability low frequency query. The underestimated risk effectively achieves the purpose of randomly extracting query records, so that medium-sized or small-scale sampling can be closer to the true distribution of data, thus providing search engine service providers with accurate user demand information and market dynamics. Improve service quality.

For many search engine service providers, it is necessary to provide a "current user is searching for what" feature, the purpose is to immediately output the query request sent by the user to the search engine. Especially for large search engines, many users use it for search queries, and each user usually searches for multiple query keywords on the search platform, so even for a short period of time, the search engine records User queries are a lot of information, such as receiving hundreds of millions of search requests a day. Among the recorded user query keywords, a large percentage of the keywords queried by users are duplicate queries, that is, the query keywords sent by different users are similar or even the same, for example, for the most recent hot events, Hundreds of millions of users will send out the same query request to query the event in a short period of time. The search engine service provider needs to process a large number of user query requests within a certain period of time in order to provide better services for users. One of the basic processing steps is to merge the same user query keywords, so that Greatly reduce the memory or disk space occupied by data storage. In order to understand the user's needs and provide users with better and more convenient services, it is necessary to conduct sample analysis and investigation on user query keywords within a certain period of time. Of course, the number of query keywords that determine the sample is small relative to the total number of all user query keywords. If the sampled records are sorted and analyzed directly, the probability of large-scale user query keywords being searched for during this period is higher, and those with low frequency and small scale are higher. The probability that the query keyword is drawn is very low, so the original sampling purpose cannot be achieved. In addition, due to the limited display space of the webpage, it is impossible to display all the online query keywords of all the netizens. Therefore, the mathematical model of the small-scale query record can be displayed by sampling the query keywords, and the user can be accurately reflected. The query requirements require that the sampling of this sample and the keyword distribution of a large number of user queries be realistically approximated.

Embodiment 2 of the present invention provides another sampling analysis method for the real sampling problem of large-scale search engine query with partial statistical information, and adopts a two-stage sampling method to solve the problems existing in the prior art and make the sampling data thereof. Close to the true distribution of the query. The flow of the overall method is shown in Figure 2. If the number of queries M that have been determined to be extracted is applied to a certain query, in the first stage of the present invention, the query keywords are first classified according to the PV value of each search engine query keyword, and the extraction of each category is calculated. Probability, whereby the number of query keywords extracted from each category can be calculated; in the second stage of the present invention, a random sample can be used to extract the final query in a certain category of the query key phrase. data.

In the following embodiments, the flow of the sampling methods of the first stage and the second stage will be further described in detail. In the first stage, the number of samples of each query keyword subset is calculated. The method flow is as shown in FIG. 3, and includes the following steps:

Step S301: Sort the search engine query keyword set stored in the “query keyword PV” format according to the PV value. The sorting may be a method in which the PV value is from large to small, or a method in which the PV value is from small to large, and the sorting manner does not affect the operation of the subsequent steps. For example, suppose that the total number of all user query keywords recorded by the search engine is 100,000 in a period of time. Among them, there are 2000 records for querying the "Alibaba" keyword, and the storage is "Alibaba 2000"; There are 1800 records of “e-commerce”, which are stored as “e-commerce 1800”; 500 records of “computer” are stored as “computer 500”; 500 records of “clothing” are stored, and stored as “clothing 500” ;......; In addition, there are 60 "water cups" inquiries, 60 "pencils" inquiries, 60 inquiries, and so on, all stored in the above storage format. Then, the above query sets are arranged in order of small to large, namely: "Water Cup 60", "Pencil 60", "Note Book 60", ..., "Computer 500", "Apparel 500", "E-commerce 1800" ", "Alibaba 2000", ....

Step S302, blaming the query keywords with the same PV value.

For query keywords with the same PV value, all of these query keywords can be regarded as a subset of the query keyword set QuerySet. The common attributes of these queries belonging to the QuerySet set are: the PV values of each query keyword are the same; In this way, different QuerySets can be obtained according to different PV values. Assuming that the PV value is from 1 to K (K is a natural number greater than 1), the query keyword subsets QuerySet1, QuerySet2, ... can be obtained accordingly. .., QuerySetK. Of course, in a specific case, for different user queries, the PV value of each query keyword counted by the search engine in a time period may also be discontinuous. By combining the query keywords with the same PV value in step S301, multiple query keyword subsets may be obtained in sequence, such as: QuerySet60, QuerySet500, QuerySet1800, QuerySet2000, and the like.

Step S303: Calculate the extraction probability of each query keyword subset.

For the query keyword subset QuerySetI composed of the query with PV value i, the total PV number of the query keyword subset is calculated statistically, that is, the SPV value: SPV _i = I *| QuerySetI |, where I represents the PV value I, | QuerySetI | represents the size of the sub-set of the query keyword, that is, how many query keywords with PV values belong to this sub-set, that is, for the query keyword sub-set QuerySet60 in step S302 Say, suppose there are 30 query keys of "query keyword 60", | QuerySet 60| is equal to 30, which means there are 30 different queries in the sub-set of this query keyword, then the SPV value is 60*30=1800. For the data satisfying the long tail distribution, the smaller the PV value is, the larger the number of query keywords included in the query keyword sub-set is, so although the PV value is small for a single query keyword, the SPV is The statistical information is not affected by the small PV value of a single query keyword. For example, for the "Water Cup 60" query data, its PV value is 100,000 relative to the total TPV value of all queries, even a small number relative to the PV value of 2000 of "Alibaba 2000", but the " The water cup 60" query data sub-set QuerySet60, the SPV value is 1800, and the SPV value of the query keyword subset of "Alibaba 2000" is 2000 (assuming only one "Alibaba 2000" query record )very close.

In order to calculate the extraction probability of each query keyword subset, the SPV numbers of all the query keyword subsets are summed to obtain the total PV of all the query keywords, which is called TPV; with the TPV, the sampling process can be calculated. The probability that each query key subset is extracted, for example, for a subset of query keywords with a PV value of i, the probability of being extracted is P _i = SPV _i / TPV .

Through the above steps, each query keyword subset can calculate the probability P _i to which the set is extracted. This probability is important for this sampling method because the probability of the low frequency query itself being extracted is very small for many low frequency query keywords. However, for a sub-set of query keywords consisting of the same PV value, the number of keywords of the low-frequency query is often large, so the number of SPVs of the query keyword sub-sets of the low-frequency query key words is still large. In this way, the low-frequency query keywords are sampled as a whole, and the probability of being extracted is effectively amplified, so that the final sampled data is more in line with the true distribution of the data. For example, the probability that the query keyword subset QuerySet60 is extracted is P ₆₀ =1800/100000=0.018, and the probability that QuerySet2000 is extracted is P ₂₀₀₀ =2000/100000=0.020, as can be seen from the data obtained. The probability that the query keyword subset QuerySet60 and QuerySet2000 are respectively extracted is very close.

In the first stage of the present invention, assuming that a certain application has determined the number of samples K, the number of queries that should be extracted from the set can be calculated according to the probability that each query keyword subset is extracted, for example, it is determined that K is 5000. The probability of sub-set extraction of PV=60 is 0.018, then the number of queries to be extracted from QuerySet60 is: 5000*0.018=90; assuming the set extraction probability of PV=2 is 0.010, then it needs to be extracted from QuerySet2. The number of queries is 5000*0.010=50.

In the second phase of sampling, the final query key is extracted from each subset of query keywords. Through the classification and statistics of different query data in the first stage, the number of queries that need to be extracted from a sub-set of query keywords is determined, and the second stage is randomly selected from a subset of specified query keywords. The flow of a query keyword is shown in Figure 4. Since in the first stage, it is possible to determine the number N of queries (N is a natural number) that need to be extracted in a certain subset of query keywords, when performing the sample query in the second stage, it is necessary to continuously select a subset of the query keywords. Sampling N times, each time randomly extracting a query record from the query keyword subset until N is filled. For example, for the query keyword sub-set QuerySet 60, in the first stage step S303, the number of queries that need to be extracted in the query sub-set is calculated according to the probability that the set is extracted, and therefore, the last time is performed. When querying keyword sampling, it is necessary to randomly sample 90 consecutive times from the query keyword subset to obtain 90 query records.

When extracting any search engine query record in a sub-set of query keywords, the principle that is followed when classifying all query keywords in the first stage is that the query keywords of the same PV value are classified into one class. Therefore, for each query keyword in the query keyword subset, the probability that it is extracted by a certain sampling should be an equal probability event, that is, the probability that each query keyword is extracted is the same. For example, when sampling in the query keyword subset QuerySet60, the extracted 90 records may include 2 "water cup" query records, "pencil" query records 3, and so on. In this way, when sampling the low frequency query, the obtained sampling result can be approximated to the true distribution of the data, and the purpose of the initial sampling analysis is achieved.

At this stage, a sub-set of a query keyword is sampled, and a commonly used random sampling method, such as a lottery method or a random number method, may be used for sampling. In this embodiment, the query key in the query keyword subset is sampled by a random number method, and the specific calculation process is as shown in Algorithm 1.

Algorithm 1: Extracting any query record from the query subset using the random number method

Through the embodiment of the present invention, a two-stage sampling analysis method is used to randomly extract the required query from a large number of search engine query keyword sets that have undergone preliminary statistical sorting, and the sampling result approximates the distribution of real data. This sampling method can not only reduce the storage required for subsequent calculations by retaining and utilizing preliminary statistics, but also solve the risk of underestimation of low-probability low-frequency queries in many methods, effectively achieving random sampling for sampling analysis. Target; based on the information obtained from the sampling data, the search engine service provider can accurately understand the user's needs and market dynamics, find some business opportunities, and appropriately adjust the search engine's service content, so as to better provide users with convenient and fast e-commerce. Online trading platform to improve service quality.

The third embodiment of the present invention provides a sampling analysis system for analyzing data of a large-scale search engine query. The structure thereof is as shown in FIG. 5, and includes: a search platform 1 for providing a search service for user queries, and recording different query keys. a PV value of the word; the sampling analysis device 2 is configured to divide the query keyword set into at least one query keyword subset according to the PV value of the different query keywords recorded by the search platform 1, and calculate the query keyword The number of samples of the set, and the query data is extracted in the subset of the query keywords according to the number of samples.

The structure of the sample analysis device 2 is as shown in FIG. 6 , and includes: a division module 21, configured to divide the query keyword into at least one query keyword subset according to the query record PV value of different query keywords; The group 22 is configured to calculate the number of samples of the subset of the query keywords divided by the partitioning module 21; the sampling module 23 is configured to be divided by the dividing module 21 according to the number of samples obtained by the computing module 22. The query keyword subset extracts query data.

In addition, the sampling analysis device 2 may further include: a storage module 24, configured to store the PV value, where the PV value is specifically the number of times at least one query keyword appears within a preset time period.

The partitioning module 21 may further include: a sorting sub-module 211 for sorting PV values of the different query keywords stored by the storage module 24; and a sorting sub-module 212 for sorting according to The sub-module 211 arranges the order of the query keywords with the same PV value into one query keyword subset.

The calculation module 22 may further include: a probability calculation sub-module 221 for calculating a extraction probability of the query keyword subset; and a sampling calculation sub-module 222 for calculating a sub-module according to the determined number of queries and the probability of the extraction. The extraction probability obtained by group 221 is used to calculate the number of samples of the subset of query keywords.

Through the sampling analysis system and device provided by the embodiment of the present invention, the required query data can be randomly extracted from a large number of query keyword sets that are preliminary statistically compiled, which can reduce the storage amount required for subsequent calculation, and can Solve the risk that the low-probability low-frequency query keywords are underestimated in many methods, effectively achieving the purpose of randomly extracting query records, so that medium-sized or small-scale sampling can be closer to the true distribution of data, thus providing search engine service providers with Accurate user demand information and market dynamics to improve service quality.

For the convenience of description, each part of the above system is described by function into various modules or devices. Of course, the functions of each module or device can be implemented in the same software or hardware or hardware in the practice of the present invention.

The above modules may be distributed in one device or distributed in multiple devices. The above modules can be combined into one module, or can be further divided into multiple sub-modules.

A person skilled in the art can understand that the drawings are only a schematic diagram of a preferred embodiment, and the modules or processes in the drawings are not necessarily required to implement the invention.

A person skilled in the art can understand that the modules in the apparatus in the embodiment can be distributed in the apparatus of the embodiment according to the description of the embodiment, or the corresponding changes can be located in one or more apparatuses different from the embodiment. The modules of the above embodiments may be combined into one module, or may be further split into multiple sub-modules.

The serial numbers of the embodiments of the present invention are merely for the description, and do not represent the advantages and disadvantages of the embodiments.

Through the description of the above embodiments, those skilled in the art can clearly understand that the present invention can be implemented by hardware, or by means of a software plus a necessary universal hardware platform. Based on such understanding, the technical solution of the present invention can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.), including A number of instructions are used to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to perform the methods described in various embodiments of the present invention.

The above disclosure is only a few specific embodiments of the present invention, but the present invention is not limited thereto, and any changes that can be made by those skilled in the art should fall within the protection scope of the present invention.

1. . . Search platform

2. . . Sampling analysis equipment

twenty one. . . Partition module

twenty two. . . Computing module

twenty three. . . Sampling module

twenty four. . . Storage module

211. . . Sorting submodule

212. . . Categorical submodule

221. . . Probability calculation sub-module

222. . . Sampling calculation sub-module

In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly described below. It is obvious that the drawings in the following description are only some embodiments of the present invention, Those skilled in the art can also obtain other drawings based on these drawings without paying any creative work.

1 is a flow chart of a sampling analysis method in Embodiment 1 of the present invention;

2 is a flow chart of a two-stage sampling analysis method in Embodiment 2 of the present invention;

3 is a flow chart of a first stage sampling method in Embodiment 2 of the present invention;

4 is a flow chart of a second stage sampling method in Embodiment 2 of the present invention;

5 is a schematic structural diagram of a sampling analysis system according to Embodiment 3 of the present invention;

FIG. 6 is a schematic structural diagram of a sampling analysis device according to Embodiment 3 of the present invention.

Claims (12)

A sampling analysis method for analyzing data of a large-scale search engine query, comprising: dividing a query keyword into at least one query keyword according to a PV (Page View) value of a query record of different query keywords a set, wherein the query key having the same PV value is classified into a query keyword subset; calculating a sample number of the query keyword subset, wherein, before calculating the sample number of the query keyword subset, The method includes: determining a number of query keywords to be extracted by the subsample analysis, wherein the calculating the number of samples of the query keyword subset comprises: calculating a extraction probability of the query keyword subset; and determining the number of queries extracted according to the determination The probability is extracted to calculate the number of samples of the subset of query keywords; and the query data is extracted in the subset of query keywords according to the number of samples. The sampling analysis method of claim 1, wherein before the query keyword is divided into at least one query keyword subset according to the query record PV value of different query keywords, the method further includes: storing the PV The value of the PV is specifically the number of times the query keyword appears on the search platform within a preset time period. The sampling analysis method of claim 1, wherein the dividing the query keyword into the at least one query keyword subset according to the PV value of the different query keywords comprises: sorting the PV values; The query keyword with the same PV value is classified into a query keyword. set. The sampling analysis method of claim 3, wherein the sorting the PV values comprises: sorting the PV values in a small to large manner; or sorting the PV values in a large to small manner. The sampling analysis method of claim 1, wherein calculating the extraction probability of the query keyword subset comprises: calculating a sum of SPV values of the PV values in the query keyword subset; The SPV value obtains a total number of times TPV (Total Page View) of the query record in the query keyword set; and obtains a extraction probability of the query keyword subset according to a ratio of the SPV value to the TPV value. The method of sampling analysis according to claim 1, wherein the extracting the query data in the subset of the query keywords according to the number of samples is obtained by a random sampling method. The sampling analysis method according to claim 6, wherein the random sampling method comprises: a lottery method and/or a random number method. A sampling analysis device for analyzing data of a large-scale search engine query, comprising: a dividing module, configured to divide a query keyword into a PV (Page View) value according to a query record of different query keywords At least one query keyword subset, wherein the query keyword having the same PV value is classified into a query keyword subset; a calculation module, configured to calculate a sample number of the subset of the query keywords divided by the partitioning module, wherein, before the calculating the number of samples of the subset of the query keywords, the method further comprises: determining that the sampling analysis is required The number of extracted query keywords, wherein the calculating the number of samples of the query keyword subset comprises: calculating a extraction probability of the query keyword subset; calculating the query keyword according to the determined number of extracted queries and the extraction probability The number of samples of the subset; and a sampling module for extracting query data from the subset of query keywords divided by the partitioning module according to the number of samples obtained by the computing module. The sampling analysis device of claim 8 , further comprising: a storage module, configured to store a PV value of the different query keywords, where the PV value is specifically within a preset time period, at least The number of times a query keyword appears. The sampling analysis device of claim 9, wherein the dividing module comprises: a sorting sub-module for sorting PV values of the different query keywords stored in the storage module; The class sub-module is configured to classify the query keywords having the same PV value into a query keyword subset according to the order arranged by the sorting sub-module. The sampling analysis device of claim 8, wherein the calculation module comprises: a probability calculation sub-module, configured to calculate a extraction probability of the query keyword subset; a sampling calculation sub-module, configured to calculate a sampling number of the query keyword subset according to the determined number of extracted queries and the extraction probability obtained by the probability calculation sub-module, wherein the query keyword subset is calculated Before the number of samples, it also includes: determining the number of query keywords to be extracted for the sample analysis. A sampling analysis system for analyzing data of a large-scale search engine query, comprising: a search platform for providing a search service for a user query, recording a PV (Page View) value of different query keywords; and a sampling analysis device And calculating, according to a PV value of different query keywords recorded by the search platform, a query keyword into at least one query keyword subset, and calculating a sample number of the query keyword subset, wherein the query key is calculated in the query Before the number of samples of the subset of words, the method further includes: determining a number of query keywords to be extracted by the subsample analysis, wherein the calculating the number of samples of the subset of the query keywords comprises: calculating a probability of extracting the subset of the query keywords; Calculating the number of samples of the query keyword subset according to the determined number of queries and the extraction probability; and extracting query data in the query keyword subset according to the number of samples.