JP2013117921A - Ranking function learning device, method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To suppress deterioration of ranking accuracy to learn a ranking function at a high speed.SOLUTION: A set of training data stored in a training data storage part 21 is divided into N subsets by a division part 22 to be allocated to N calculation nodes 23. Parameters regarding a ranking function are learned by each of the N calculation nodes 23. In the respective calculation nodes 23, update of Lagrange undetermined multipliers and local parameters by a local update part 32, notification of the updated Lagrange undetermined multipliers and the local parameters to other calculation nodes 23 and acquisition of the updated Lagrange undetermined multipliers and the local parameters from other calculation nodes 23 by a synchronization part 33, and update of global constraint parameters by global update means are repeated until values of the global constraint parameters are converged.

Description

  The present invention relates to a ranking function learning apparatus, method, and program, and in particular, a ranking for learning a ranking function for determining a presentation order for a search result (a set of documents) obtained by searching based on a search query. The present invention relates to a function learning apparatus, method, and program.

  The web search system can be regarded as a problem of extracting a document set that matches an input search request word set (query) from all document sets to be searched. However, recent web search systems construct a system that regards a document having a higher degree of matching with a query as a ranking problem that is displayed higher than extracting a document set that matches the query. That is, when a query is given to the search system, as shown in FIG. 12, a “search score” corresponding to the matching degree between the query and the document is obtained for all documents to be searched or all documents including the query. The search results are presented by calculating and displaying the documents in descending order according to the search score.

  The suitability of each document for a query is calculated using a number of factors such as a score based on query frequency such as TF-IDF and a score based on link analysis such as PageRank. Here, these factors are called “ranking factors”, and the process of determining the presentation order from a number of ranking factors is called “ranking process”.

A number of techniques for constructing a ranking processing apparatus using training data created manually are proposed as methods for realizing ranking processing of search documents suitable for human intuition (for example, Non-Patent Document 1). ). The training data includes a pair of an assumed query and a document set that matches the query, and each document is manually labeled with a degree of matching with the query. At this time, if the number of ranking factors used when calculating the fitness between the query and each document is D, each document is expressed as a D-dimensional vector in which D scores based on each ranking factor are arranged. be able to. Accordingly, each document is represented by a vector of z = (z ₁ ,..., Z _D ) and is called a “feature vector” of the document. Where z _i is the score of the i-th ranking factor. Let z _{i, j} be the feature vector of the document whose query ID is i and whose document number in the document set that matches the query is j. An example of training data is shown in FIG.

  In FIG. 13, each row represents the feature expression and the degree of matching of the search result document for a certain query. A higher matching score indicates a more appropriate result for the query. As a point of caution, since the matching level is given to a pair of a query and a document, different matching levels are given to different queries even for the same document. As the degree of fitness, for example, a multi-stage value (for example, five levels) determined and given by the subject is used.

A conventional technique for constructing a ranking processing device from given training data will be described. The goal here is to build a function that gives a ranking that matches the fitness of the given training data. Here, a function that gives ranking is called a “ranking function”. In other words, when a query in the training data is given, the function is a function that gives the entire order relation of the document set related to the query according to the training data. In practice, however, constructing a function that determines the total order relationship is expensive in terms of computational complexity. Therefore, a combination of two documents with different fitness levels assigned to training data for each document set in each query is created with all possible combinations, and from the viewpoint of which of the two documents has a higher fitness level. A method of building a ranking function is often used (see Non-Patent Document 1 above). Therefore, when actually using the training data, the training data is used in the form of (y _m , x _m ). However, as shown in FIG. 14, for a pair of documents (z _{i, j} , z _{i, k} ) having a query ID of i-th and different relevance, the relevance of the left document in the paired document is convenient. Let y _m be a variable that becomes 1 when the right document is high and −1 when the right document is high, and let x _m be the difference vector of each document represented as a feature vector (ie, x _m = z _{i, j} −z _{i , k} ).

  Here, w represents a parameter representing the weight (reliability) for each ranking factor. Also, let M be the total number of document pairs in the training data. At this time, in the RankSVM method proposed in the above-mentioned Non-Patent Document 1, a ranking function is constructed using a solution to an optimization problem expressed by the following equation (1).

However, w · x _m represents the inner product of w and x _m .

  At this time, the generated ranking function is expressed by a parameter vector ^ w. This ^ w is a value corresponding to the reliability of each factor that determines the ranking, obtained based on training data. Therefore, the ranking function f () is expressed by the following equation (2).

  The ranking process itself is performed by arranging the documents in descending order based on the output value of the ranking function of the above equation (2).

  Since a search system has a wide variety of search requests, it is difficult to generate a ranking function that accurately presents a ranking for any query. In order to generate a ranking function that is as accurate as possible, it is desirable to construct a ranking function using more training data. Therefore, the ranking function generation process requires a framework for efficiently handling a lot of training data.

  Several methods have been proposed to cope with the large scale of training data. Basically, a method that supports distributed parallel processing is currently mainstream. Divided partial training data is allocated to a plurality of calculation nodes, and each calculation node independently builds a ranking function using only the assigned data. After that, combining a plurality of completed ranking functions by each method is the main method currently used (for example, Non-Patent Documents 2, 3, and 4).

Thorsten Joachims. Optimizing search engines using clickthrough data.In Proc. Of the eighth ACM international conference on Knowledge Discovery and Data mining (KDD '02), pages 133-142, 2002. Krysta M. Svore and Christopher JC Burges.Large-scale learning to rank using boosted decision trees.In Ron Bekkerman, Misha Bilenko, and John Langford, editors, Scaling Up Machine Learning: Parallel and Distributed Approaches.Cambridge Univ.Press, May 2011 . Stephen Tyree, Kilian Wienberger, Kunal Agrawal, and Jennifer Paykin.Parallel boosted regression trees for web search ranking.In WWW, 2011. Jerry Ye, Jyh-Herng Chow, Jiang Chen, and Zhaohui Zheng. Stochastic gradient boosted distributed decision trees. In CIKM, 2009.

  However, the methods used in the above Non-Patent Documents 2 to 4 are simple combining methods, and thus the accuracy is almost lower than the ranking function constructed using the entire training data. In other words, the speed can be increased by distributed processing, but there is an aspect that sacrifices accuracy.

  The present invention has been made in view of the above circumstances, and an object of the present invention is to provide a ranking function learning device, method, and program capable of learning a ranking function at high speed while suppressing a decrease in ranking accuracy. To do.

  In order to achieve the above object, a ranking function learning device according to a first invention is a device for learning parameters related to a ranking function for ranking a search result obtained by searching a set of documents based on a search query. Training data storage means for storing a set of training data including the feature value obtained for each document of the search result for the query and the fitness for the search query, and N pieces (N is 2 or more) for learning parameters related to the ranking function And a dividing unit that divides the set of training data stored in the training data storage unit into N subsets and assigns it to the N computing nodes, Each of the N computational nodes is based on a subset of the training data assigned by the dividing means. A local update means for updating a local parameter which is a parameter related to the ranking function, and notifies the other calculation node of the local parameter updated by the local update means, and the other calculation Based on synchronization means for acquiring the local parameter notified from the node, the local parameter of the other calculation node acquired by the synchronization means, and the local parameter updated by the local update means The global update means for updating the global parameter for matching the local parameter of each computation node, and whether the value of the global parameter has converged. Until the global parameter value has converged. Updating by means, notification and acquisition by the synchronization means, and is configured to include a convergence judging means for repeating the updating by the global update means.

  The ranking function learning method according to the second invention includes training data storage means for storing a set of training data including a feature value obtained for each document of a search result for a search query and a fitness for the search query; Ranking function learning in an apparatus for learning parameters related to a ranking function for ranking a search result obtained by searching a set of documents based on a search query, including a calculation node (N is a natural number of 2 or more) and a dividing unit Dividing the set of training data stored in the training data storage means into N subsets by the dividing means and assigning the N calculation nodes to the N computing nodes; Learning a parameter related to the ranking function by each of the computation nodes, The step of learning by the step of updating the local parameter, which is a parameter related to the ranking function, based on the subset of the training data assigned by the dividing unit by the local updating unit, Means for notifying the local parameter updated by the local update means to another computation node, and obtaining the local parameter notified from the other computation node; and global update means The parameter relating to the ranking function based on the local parameter of the other calculation node acquired by the synchronization unit and the local parameter updated by the local update unit, and for each calculation To match the local parameters of the node A step of updating a global parameter; and by means of convergence determination means, whether or not the value of the global parameter has converged, and the local update until it is determined that the value of the global parameter has converged Updating by the means, notification and acquisition by the synchronizing means, and repeating the updating by the global updating means.

  According to the first and second aspects of the invention, the dividing means divides the set of training data stored in the training data storage means into N subsets to be the N calculation nodes. assign. A parameter related to the ranking function is learned by each of the N calculation nodes.

  At this time, in each calculation node, the local update unit updates the parameter related to the ranking function and the local parameter based on the subset of the training data assigned by the dividing unit. The synchronization unit notifies the other calculation node of the local parameter updated by the local update unit, and acquires the local parameter notified from the other calculation node. A parameter related to the ranking function based on the local parameter of the other computation node acquired by the synchronization unit and the local parameter updated by the local update unit by a global updating unit; and , Update global parameters to match the local parameters of each compute node. The convergence determining means determines whether or not the global parameter value has converged, and until it is determined that the global parameter value has converged, the updating by the local updating means, the notification by the synchronizing means, and The acquisition and the update by the global update means are repeated.

  In this way, in each computation node, the local parameters are updated based on the assigned subset of training data, and the global parameters are updated using the local parameters obtained from other computation nodes. By repeating this until convergence, it is possible to suppress a decrease in ranking accuracy and learn the ranking function at high speed.

  A ranking function learning device according to a third invention is a device for learning parameters relating to a ranking function for ranking a search result obtained by searching a set of documents based on a search query, and for each document of the search result for the search query. Training data storage means storing a set of training data including the obtained feature value and the degree of fitness for the search query, and N calculation nodes (N is a natural number of 2 or more) for learning parameters related to the ranking function And a dividing means for dividing the set of training data stored in the training data storage means into N subsets and assigning the set to the N calculation nodes, a global updating means, and a convergence determining means. , Each of the N computational nodes is based on a subset of the training data assigned by the dividing means, A parameter that is a parameter related to the ranking function and includes a local update unit that updates a local parameter, wherein the global update unit is configured to update the local parameter updated by the local update unit of all the computation nodes. Based on the ranking function and a means for updating a global parameter for matching the local parameter of each computation node, and the convergence determining means includes the global parameter Until the value of the global parameter has converged and until it is determined that the value of the global parameter has converged, the update by the local update unit and the update by the global update unit of each calculation node are repeated. It is characterized by.

  A ranking function learning method according to a fourth aspect of the present invention includes a training data storage unit that stores a set of training data including a feature value obtained for each document of a search result for a search query and a fitness for the search query; A ranking function for ranking a search result obtained by searching a set of documents based on a search query, including a calculation node (N is a natural number of 2 or more), a dividing unit, a global updating unit, and a convergence determining unit. A ranking function learning method in an apparatus for learning a parameter relating to the calculation method, wherein the dividing unit divides the set of training data stored in the training data storage unit into N subsets and performs the N calculations. Assigning parameters to the ranking function by each of the N calculation nodes and assigning to the nodes A step of updating by the global updating unit and a step of determining by the convergence determining unit, wherein the step of learning by the calculation node includes the training data allocated by the dividing unit by a local updating unit. Updating a parameter that is a parameter related to the ranking function based on a subset of the local function, and updated by the global updating unit is performed by the local updating unit of all the computation nodes. Updating a global parameter based on the updated local parameter for the ranking function and for matching the local parameter of each computation node, the convergence The step of determining by the determining means includes It is determined whether or not the global parameter value has converged, and the update by the local update unit and the update by the global update unit of each calculation node are performed until it is determined that the global parameter value has converged. It is a repeating step.

  A program according to a fifth invention is a program for causing a computer to function as each means of the ranking function learning device.

  As described above, according to the ranking function learning device, method, and program of the present invention, each calculation node updates local parameters based on a subset of assigned training data, and other calculation nodes. Using the local parameters obtained from the above, it is possible to learn the ranking function at high speed by suppressing the decrease in ranking accuracy by repeating updating the global parameters until convergence. can get.

It is a figure which shows the flow of the ranking function learning process which concerns on the 1st Embodiment of this invention. It is the schematic which shows the structure of the ranking function learning apparatus which concerns on the 1st Embodiment of this invention. It is a flowchart which shows the content of the ranking function learning process routine in the calculation node of the ranking function learning apparatus which concerns on the 1st Embodiment of this invention. It is the schematic which shows the other example of a structure of the ranking function learning apparatus which concerns on the 1st Embodiment of this invention. It is a figure which shows the flow of the ranking function learning process which concerns on the 2nd Embodiment of this invention. It is the schematic which shows the structure of the ranking function learning apparatus which concerns on the 2nd Embodiment of this invention. It is the schematic which shows the structure of the ranking function learning system which concerns on the 3rd Embodiment of this invention. It is the schematic which shows the structure of the learning control apparatus which concerns on the 3rd Embodiment of this invention. It is the schematic which shows the structure of the learning apparatus which concerns on the 3rd Embodiment of this invention. It is a graph which shows a performance comparison result. It is a figure which shows the experimental result at the time of performing distributed parallel processing. It is a figure for demonstrating the process by a ranking processing apparatus. It is a figure which shows training data. It is a figure which shows training data.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

[First Embodiment]
<Outline of the invention>
The processing framework of the present invention basically assumes a distributed computing environment. Here, it is assumed that there are N computation nodes as shown in FIG. Next, M pieces of training data are divided into N pieces. At this time, the n-th subset of the divided training data is written as M _n . Also assume that each subset has no overlapping data. That is, ∪ _n M _n = (1, ..., M), ∩ _n M _n = empty set.

After the set of training data is divided and assigned to each computation node, each computation node uses only the training data assigned to itself, and uses the conventional method shown in the above equation (1) to calculate the ranking function. learn. At this time, the parameter used in the ranking function obtained at the nth computation node is _denoted by vn. This v _n corresponds to w in the above equation (1). In the framework of the present invention, each calculation node can independently learn the parameters used in the ranking function, so that there is a computational advantage that the learning time can be reduced to 1 / N by simple calculation. However, since v _n is generated from only partial training data, it can be used only for partial information, so that the ranking accuracy is basically higher than the ranking function using w trained using the entire training data. Is generally low. Further, a method of selecting the best v _n from v _n obtained independently, or a method of binding effectively v _n is not known until now.

In the present invention, a constraint parameter w is introduced that v _n obtained at each computation node must match. By including this restriction, v _n generated from only partial data must match throughout, so it is possible to learn a function that indirectly includes information on the entire training data. . In other words, by dividing the training data set, a function with high ranking accuracy as in the conventional ranking function learned using the entire training data is maintained while retaining the property of shortening the learning time of the ranking function. It becomes possible to learn. In addition, using a distributed computing environment makes it easy to handle training data even if it increases, so it becomes possible to construct a ranking function using a quantity of training data that could not be used in the past. It becomes possible to learn the ranking function. In the framework of the present invention, the parameter ^ w used for the ranking function can be obtained by solving the optimization equation shown by the following equation (3).

In the present invention, Reference 1 (Daniel Gabay and Bertrand Mercier. A dual algorithm for the solution of nonlinear variational problems via finite element approximation. Computers and Mathematics with Applications, 2 (1): 17-40, 1976.) and reference 2 (alternating direction method of multipliers described in Stephen Boyd, Neal Parikh, Eric Chu, Borja Peleato, and Jonathan Eckstein. Distributed optimization and statistical learning via the alternating direction method of multipliers. Foundations and Trends in Machine Learning, 2011.) The optimization problem is solved using an optimization framework called (ADMM). First, the optimization equation of the above equation (3) is modified in accordance with the ADMM framework to obtain an objective function L _ρ shown in the following equation (4).

Here, μ _n corresponds to the Lagrange multiplier when the constraint of the above equation (4) is expressed by the Lagrange undetermined multiplier method, and ρ / 2 || v _n −w || ² is an efficient optimization solution. Is an additional term for ρ is a parameter determined manually, and ρ> 0. According to the above reference 2, the above equation is replaced by u _n = μ _n / ρ. As a result, the above equation (4) is transformed into the following equation (5).

By using the equation (5), the actual optimization problem, v _n, w, optimal solution by finding repeated sequentially u _n is guaranteed to be obtained. Also, as in the second embodiment, which will be described later, the final optimal solution ^ w can be obtained even if the training data increases sequentially, that is, the number of computation nodes increases. Guaranteed by reference 3 (Pedro Forero, Alfonso Cano, and Georgios Giannakis. Consensus-based distributed support vector machines. JMLR, 2010.). By utilizing these properties, it is possible to perform a ranking function learning process by distributed processing and sequential input of data.

<System configuration>
The ranking function learning device 100 according to the first exemplary embodiment of the present invention receives training data given for learning as an input, and outputs parameters related to the ranking function. The ranking function learning device 100 is configured by a computer including a CPU, a RAM, and a ROM storing a program for executing a ranking function learning processing routine to be described later, and functionally configured as follows. Has been. As shown in FIG. 2, the ranking function learning device 100 includes an input unit 10, a calculation unit 20, and an output unit 30.

  The input unit 10 accepts a plurality of training data based on the document pairs shown in FIG. 14 as the input training data. Each training data includes a difference vector composed of differences between a plurality of feature values of a search result document pair for a certain search query, and a variable indicating the magnitude relationship of the degree of fitness for the search query. The training data that is input includes training data based on various search queries.

  Further, the input unit 10 receives the number of calculation nodes N and the parameter ρ input manually.

The calculation unit 20 includes a training data storage unit 21, a division unit 22, and N calculation nodes 23 ₁ to 23 _N. Note that, when an arbitrary calculation node among the calculation nodes 23 ₁ to 23 _N is indicated, it is referred to as a calculation node 23.

  The training data storage unit 21 stores a training data set including a large amount of training data received by the input unit 10. The data structure of the training data storage unit 21 is shown in FIG.

As shown in FIG. 14, each row is a document ID of a search result document pair for a certain search query, and the right document is high when the matching of the left document is high for convenience. It represents a variable y _m that sometimes becomes −1 and a difference vector x _{m of} each document represented by each feature vector.

The dividing unit 22 divides the training data set stored in the training data storage unit 21 into N subsets and assigns them to _N computing nodes 23 ₁ to 23 _N. Further, the dividing unit 22 notifies the input parameter ρ to each of the _N calculation nodes 23 ₁ to 23 _N.

Each of the _N calculation nodes 23 ₁ to 23 _N includes a divided data storage unit 31, a local update unit 32, a synchronization unit 33, a global update unit 34, and a convergence determination unit 35. Each of the N divided data storage units 31, the local update unit 32, the synchronization unit 33, the global update unit 34, and the convergence determination unit 35 exists, but the processing units having similar functions are represented by the same numbers. ing.

  The divided data storage unit 31 stores a subset of the training data set assigned to the calculation node 23.

The local update unit 32 updates the Lagrange undetermined multiplier u _n and the local parameter v _n using a subset of the training data set stored in the divided data storage unit 31 as described below.

(Update of the Lagrange multiplier u _n)
As a first process, the local updating section 32 updates the Lagrange multipliers u _n in the compute node 23n.

When w and v _n are fixed, the direction of the optimum value of u _n is u of the objective function L _ρ (w, v _n , u _n ) shown in the above equation (5) as shown in the following equation (6). _The gradient direction with respect to _n .

  From the relationship of the above equation (6), the update equation shown in the following equation (7) is obtained.

Local update unit 32, according to the above (7), updates the Lagrange multipliers u _n. Since the update formula (7) can be calculated independently by each calculation node 23, it is not necessary to communicate with other calculation nodes 23.

(Local parameter v _n update)
Next, the local update unit 32 updates the local parameter v _n in the calculation node 23 _n .

At the time of the iterative calculation k (k is a variable for managing the number of repetitions), when u _n and w are fixed, the optimal solution for each v _n is as shown in the following equation (8), The problem is to find v _n that minimizes the objective function L _ρ (w, v _n , u _n ) shown.

If only the term related to the local parameter v _n is extracted according to the definition, the following equation (9) is obtained.

Equation (9) can be regarded as RankSVM adding the bias term (-w + u _n) to the model shown in a conventional manner. Further, the bias term (-w + u _n), since the constant can now be solved in the same way as RankSVM conventional methods. If (−w + u _n ) = 0, this is consistent with the RankSVM equation. That is, in the individual compute nodes is equivalent to generating a parameter v _n with RankSVM conventional methods independently using a subset of the training data given.

The local update unit 32 updates the local parameter v _n according to the above equation (9). Also, the parameter v _n can be updated independently at each computation node 23, similarly to u _n .

The synchronization unit 33 notifies u _n and v _n updated this time at the calculation node 23 to all the calculation nodes 23 _i other than itself. Further, the synchronization unit 33 receives u _i and v _i updated this time notified from all the other computation nodes 23 _i . By this processing, each computation node 23 can obtain the values of _un and v _n possessed by all the computation nodes 23 _n .

The global updating unit 34 updates the global constraint parameter w as follows using u _n and v _n received from all the other computation nodes 23 _n .

When u _n and v _n are fixed, the optimal solution of w is a point where the gradient of w of the objective function L _ρ (w, v _n , u _n ) shown in the above equation (5) becomes a zero vector. From the relationship, the following relational expressions (10) to (13) are obtained.

_{^{However, ¯v = Σ N n = 1}} v n / N, is _{^{¯u = Σ N n = 1 u}} n / N.

  From the relational expression shown in the above expression (13), w at the time of the iterative calculation k is obtained by the following expression (14).

All v _n and u _n are required to update the global constraint parameter w.

Therefore, the global updating unit 34 updates the global constraint parameter w according to the above equation (14) using the values of _un and v _n possessed by all the obtained computation nodes 23 _n .

  Each calculation node 23 obtains w independently. As a point to be noted here, each computation node 23 independently obtains w, but the obtained w is identical in all computation nodes 23. As a processing method, a process may be performed in which w is calculated by any one calculation node 23 and then notified to each calculation node 23. However, in that case, it is necessary to wait for the other calculation nodes until the calculation of the selected calculation node 23 ends and the result is notified. In the present embodiment, a case will be described as an example in which the same calculation is performed in each calculation node 23.

  The convergence determination unit 35 determines whether the obtained global constraint parameter w has converged to an optimum value.

Two small positive real numbers ε ₁ and ε ₂ are given, and it is determined that convergence has occurred when the following equations (15) and (16) are satisfied (see Reference 2).

  When the convergence is not determined in the convergence determination, k = k + 1 is set, and the process returns to the local updating unit 32. If it is determined that the process has converged, the iterative process is terminated.

  The convergence determination process may be performed by any one computation node 23 as in the case of w, and the result may be notified to the whole. However, since synchronous processing is required, in this embodiment, a case will be described as an example where convergence is also individually determined in all the computation nodes 23, and the processing is terminated if it is determined that convergence has occurred. The result of this convergence determination is always the same at all the computation nodes, so the result is the same even if the determination is made individually.

  The convergence determination unit 35 outputs the global constraint parameter w obtained when determined to have converged by the output unit 30 as a parameter constituting the ranking function. In this embodiment, the case where the global constraint parameter w is output from any one computation node 23 has been described as an example. However, the global constraint parameter w is output from all computation nodes 23. Also good. When generating the actual ranking, the search score is calculated using the above equation (2) as in the conventional method.

<Operation of ranking function learning device>
Next, the operation of the ranking function learning apparatus 100 according to the present embodiment will be described. First, when a training data set including a large amount of training data is input to the ranking function learning device 100, the input training data set is stored in the training data storage unit 21 by the ranking function learning device 100. Then, in the ranking function learning device 100, the dividing unit 22 divides the training data set in the training data storage unit 21 into N subsets, and assigns them to _N calculation nodes 23 ₁ to 23 _N. Each subset of the training data set is stored in each divided data storage unit 31 of each of the computation nodes 23 ₁ to 23 _N.

Then, the ranking function learning process routine shown in FIG. 3 is executed by each calculation node 23 of the ranking function learning device 100. Hereinafter, a case where the calculation is executed by the calculation node 23 _n will be described.

First, in step S101, each of the Lagrange undetermined multiplier u _n , the local parameter v _n , and the global constraint parameter w is initialized by giving an appropriate value (for example, 0).

In step S102, the local update unit 32 initializes the Lagrange undetermined multiplier u _n , the local parameter v _n , and the global constraint parameter w initialized in step S101, or the last updated Lagrange undetermined multiplier u _n. Based on the local parameter v _n and the global constraint parameter w, the Lagrange undetermined multiplier u _n is updated according to the above equation (7).

In the next step S103, the local by the update unit 32, and the Lagrange multiplier u _n updated in step 102, the initialized global constraint parameters w in step S101, or the last updated global constraint parameters w Then, based on the subset of the training data set stored in the divided data storage unit 31, the local parameter v _n is updated according to the above equation (9).

In step S104, the synchronization unit 33 notifies the Lagrange undetermined multiplier u _n updated in step S102 and the local parameter v _n updated in step 103 to the other calculation nodes 23, and other The updated Lagrange undetermined multiplier u _i and the local parameter v _i updated in step 103 are acquired from all the computation nodes 23 _i (i = 1,..., N−1, n + 1,...). , N).

In step S105, the Lagrangian undetermined multiplier u _n updated in step S102 and the local parameter v _n updated in step 103 and the other calculation nodes 23 _{i in} step S104 are all updated by the global updating unit 34. Based on the acquired Lagrange undetermined multiplier u _{i and the} local parameter v _i , the global constraint parameter w is updated according to the above equation (14).

In the next step S106, based on the local parameter v _n updated in step 103, the global constraint parameter w updated in step 105, and the global constraint parameter w updated last time, It is determined whether the global constraint parameter w has converged according to the equations (15) and (16). If the above equations (15) and (16) are not satisfied, it is determined that they have not converged, and the process returns to step S102. On the other hand, if the above equations (15) and (16) are satisfied, it is determined that convergence has occurred, and the process proceeds to step S107.

In step S107, the global constraint parameter w ^k finally updated in step S105 is output by the output unit 30 as the obtained optimal solution ^ w, and the ranking function learning processing routine is terminated.

  As described above, according to the ranking function learning device according to the first embodiment, the Lagrange undetermined multiplier and local parameters are updated based on a subset of the assigned training data in each calculation node, Using the Lagrange undetermined multiplier and local parameters obtained from other computing nodes, it is possible to learn the ranking function at high speed by suppressing the decrease in ranking accuracy by repeating updating the global constraint parameters until convergence. can do.

  In addition, it is possible to automatically construct a ranking processing apparatus that is used in a web search engine or the like and that determines in what order the relevant documents are presented when displaying the document search results.

In the above-described embodiment, the case where each calculation node 23 includes the global update unit 34 and the convergence determination unit 35 has been described as an example. However, the present invention is not limited to this. For example, as illustrated in FIG. 4, each calculation node 23 includes a divided data storage unit 31 and a local update unit 32, and the calculation unit 20 includes one global update unit 34 and one convergence determination unit 35. It may be configured. In this case, the global update unit 34 uses the u _n and v _n obtained in all the computing nodes 23 _n, and updates the global constraint parameters w, the convergence determination unit 35, resulting global It is determined whether the constraint parameter w has converged to an optimum value. If the convergence is not determined in the convergence determination, the global constraint parameter w obtained for each computation node 23 is notified, and the process returns to the processing by the local updating unit 32.

[Second Embodiment]
Next, a second embodiment will be described. In addition, about the part which becomes the structure similar to 1st Embodiment, the same code | symbol is attached | subjected and description is abbreviate | omitted.

  The second embodiment is different from the first embodiment in that processing is performed when new training data is added to the training data set.

<Outline of the invention>
One thing to consider when generating the ranking function is that the set of documents to be searched by the web search system changes from moment to moment. As one method corresponding to this phenomenon, for example, a method of sequentially increasing training data suitable for each occasion and updating the ranking function to a more suitable one for the current state can be considered. In addition, since a pair of a query obtained when a human uses a search system and a document selected at that time can be regarded as training data, it may be possible to reflect these feedbacks in real time.

  In the said nonpatent literatures 2-4, since the point of how a new training data is taken in and a ranking function is constructed | assembled when training data increases sequentially, training data is sequentially used. There is a problem that it is not always efficient when adding.

  Here, the processing when the training data according to the present invention is sequentially added will be described with reference to FIG.

  Even when training data is added sequentially, the framework of the present invention can easily cope. Essentially, the learning method using the static training data set described in the first embodiment can be used as it is. In the processing when the training data increases every time, it is assumed that the training data exists infinitely. However, it is assumed that, at a certain time t, it happens that there are n computing nodes operating. At that time, in the framework of the present invention, it is possible to continue the processing partially by using only n computing nodes that are operating. The training data is assigned in such a way that a set of training data increased at each time is allocated to one newly added computing node so that n + 1 computing nodes can be used at time t + 1. The learning process can be continued even when the number increases sequentially.

  This is because the method of the present invention can perform almost all computations independently at each computation node. In addition, in the update of w, which is the only information sharing among all compute nodes, the solution ^ w obtained by running N compute nodes without stopping from the beginning and the compute nodes that can be used by adding sequentially are available. The solution ^ w obtained by using all the computation nodes including it has the same property. This property enables efficient calculation even in an environment where training data increases sequentially.

<System configuration>
As illustrated in FIG. 6, the calculation unit 220 of the ranking function learning device 200 according to the second embodiment includes a training data storage unit 21, a division unit 22, N calculation nodes 23 ₁ to 23 _N , and data addition. Part 222 is provided. When new training data is added to the training data storage unit 21, the calculation unit 220 adds at least one (two in the example of FIG. 6) calculation nodes 223 _{N +} corresponding to the addition of training data. ₁ and 223 _{N + 2} are provided. It should be noted that when any calculation node of the calculation nodes 223 _{N + 1} and 223 _{N + 2} is indicated, it is referred to as a calculation node 223.

When a training data set has been added up to a certain point in time, the data adding unit 222 assigns the additional training data set to the computation node 223 _{N + 1} . When the training data set is further added up to the next time point, the data adding unit 222 assigns a further addition of the training data set to the calculation node 223 _{N + 2} .

Each of the computation nodes 223 _{N + 1 to} 223 _{N + 2} includes an additional data storage unit 231, a local update unit 32, a synchronization unit 33, a global update unit 34, and a convergence determination unit 35.

  The additional data storage unit 231 stores an additional training data set assigned to the calculation node 223.

  The local update unit 32 and the convergence determination unit 35 of the calculation node 223 are the same as the calculation node 23.

The synchronization unit 33 of the calculation node 223 notifies all the calculation nodes 23 and 223 other than itself of the u _n and v _n updated this time at the calculation node 223 _n . Further, the synchronization unit 33 of the calculation node 23 also notifies all the calculation nodes 23 and 223 other than itself of the u _i and v _i updated this time in the calculation node 23 _i . Each synchronization unit 33 of the computing node 23,223 were reported from other computing nodes 23,223 all receive this updated u _n and v _n. This process, individual compute nodes 23,223 can obtain the value of u _n and v _n with the all compute nodes 23,223.

Each global update unit 34 of the computing node 23,223, using the u _n and v _n received from other computing nodes 23,223 all, according to the above (14), it updates the global constraint parameters w.

<Operation of ranking function learning device>
Next, the operation of the ranking function learning device 200 according to the present embodiment will be described. First, when a training data set composed of a large amount of training data is input to the ranking function learning device 200, the input training data set is stored in the training data storage unit 21 by the ranking function learning device 200. In the ranking function learning device 200, the dividing unit 22 divides the training data set in the training data storage unit 21 into N subsets and assigns them to N computing nodes 231 to 23N.

  Then, the ranking function learning processing routine shown in FIG. 3 is executed by each calculation node 23 of the ranking function learning device 200.

When a training data set is additionally input to the ranking function learning device 200 during execution of the ranking function learning processing routine by each calculation node 23, the training data set additionally input by the ranking function learning device 200 is stored in the training data storage. Stored in the unit 21. In the ranking function learning apparatus 200, the data adding unit 222 assigns the additional portion of the training data set in the training data storage unit 21 to the calculation node 223 _{N + 1} . The ranking function learning processing routine is executed by the calculation node 223 _{N + 1} of the ranking function learning device 200 in the same manner as each calculation node 23.

At this time, in step S104 of the ranking function learning processing routine executed in each calculation node 23, 223, the updated Lagrange undetermined multiplier u _i and the local parameter v _i are notified to the other calculation nodes 23, 223. The updated Lagrange undetermined multiplier u _i and the local parameter v _i updated in step 103 are obtained from all the other computation nodes 23 and 223.

Further, when a training data set is additionally input to the ranking function learning device 200 during execution of the ranking function learning processing routine by each of the calculation nodes 23 and 223 _{N + 1} , the training additionally input by the ranking function learning device 200 is input. The data set is stored in the training data storage unit 21. The data addition unit 222 assigns a further addition of the training data set in the training data storage unit 21 to the calculation node 223 _{N + 2} . The ranking function learning processing routine is executed by the calculation node 223 _{N + 2} of the ranking function learning device 200 in the same manner as each calculation node 23.

Further, each time a training data set is additionally input, calculation nodes 223 _{N + 3} , 223 _{N + 4} ,... And a calculation node 223 are added, and a further addition of the training data set is assigned, and ranking function learning is performed. The processing routine is executed similarly.

  As described above, according to the ranking function learning device according to the second embodiment, when learning the ranking function used in the ranking processing device, it is possible to simultaneously increase the scale of training data and sequentially input the training data. To do. Thereby, it becomes possible to construct a ranking function with higher accuracy than before by using a large amount of training data that could not be handled conventionally. In addition, since the ranking function can be constructed while increasing the training data sequentially, it becomes easy to adapt the ranking function in accordance with the data that changes every moment.

  In the search ranking function learning process, the search query actually input by the user and the Web page clicked by the user can be used as pseudo training data. It is possible to improve the performance of the ranking function by sequentially incorporating these so-called user feedbacks.

  In the above embodiment, the ranking function learning process routine executed before the addition of training data has been described as an example of the case where training data is added while the calculation node is executing. It is not limited to this. The training data may be added after the execution of the ranking function learning processing routine is completed by N calculation nodes. In this case, the values of the various variables u, v, and w finally obtained by the ranking function learning process routine that has been executed before the addition of the training data are stored, and the stored values are used as initial values of the various variables. The ranking function learning processing routine may be executed by each of a plurality of calculation nodes including the added calculation node. Thereby, it is possible to effectively use the result learned before the training data is added.

[Third Embodiment]
Next, a third embodiment will be described. In addition, about the part which becomes the structure similar to 1st Embodiment, the same code | symbol is attached | subjected and description is abbreviate | omitted.

  In the third embodiment, in the distributed parallel computing environment provided with a plurality of learning devices connected by a network, parameter updating is performed by distributed parallel computation by a plurality of learning devices. It is different from the form.

<System configuration>
As shown in FIG. 7, the ranking function learning system 300 according to the third embodiment includes a learning control device 301 and N learning devices 302 _{1 to} 302 _N. The learning control device 301 and the N learning devices 302 _{1 to} 302 _N are connected via a network 303. Note that when any learning device among the learning devices 302 _{1 to} 302 _N is shown, it is referred to as a learning device 302.

  As illustrated in FIG. 8, the learning control device 301 includes an input unit 10, a calculation unit 320, and an output unit 330.

  The calculation unit 320 includes a training data storage unit 21 and a division unit 22.

Dividing unit 22, a training data set stored in the training data storage unit 21, it is divided into N subsets, via the network 303 to the N learning apparatus 302 ₁ to 302 _N. The dividing unit 22 transmits the input parameter ρ to each of the _N learning apparatuses 302 _{1 to} 302 N via the network 303.

Each of the _N learning devices 302 _{1 to} 302 _N includes an input unit 340, a calculation unit 350, and an output unit 360, as shown in FIG.

  The input unit 340 receives a subset of the training data set transmitted from the learning control device 301. Further, the input unit 340 receives information transmitted from the other learning device 302 via the network 303.

  The calculation unit 350 includes a divided data storage unit 31, a local update unit 32, a synchronization unit 33, a global update unit 34, and a convergence determination unit 35.

  The divided data storage unit 31 stores a subset of the training data set transmitted to the learning device 302.

The synchronization unit 33 transmits u _n and v _n updated this time by the learning device 302 _n to all the learning devices 302 other than the learning device 302 _n via the network 303. The synchronization unit 33 also receives the currently updated u _i and v _i transmitted from all the other learning devices 302 _i . By this processing, each learning device 302 can acquire the values of u _n and v _n possessed by all the learning devices 302 _n .

The global updating unit 34 uses the u _i and v _i received from all the other learning devices 302 _i to update the global constraint parameter w according to the above equation (14).

  The convergence determination unit 35 determines whether the obtained global constraint parameter w has converged to an optimum value, and configures the global constraint parameter w obtained when it is determined to have converged to form a ranking function. The parameter is transmitted to the learning control apparatus 301 by the output unit 360 as a parameter.

<Operation of ranking function learning system>
Next, the operation of the ranking function learning system 300 according to the present embodiment will be described. First, when a training data set including a large amount of training data is input to the learning control device 301, the input training data set is stored in the training data storage unit 21 by the learning control device 301. Then, in the learning control device 301, the dividing unit 22 divides the training data set in the training data storage unit 21 into N subsets, and transmits them to the N learning devices 302 via the network 303. Assign to each learning device 302. A subset of the training data set is stored in the divided data storage unit 31 of the learning devices 302 _{1 to} 302 _N.

  Then, the ranking function learning process routine shown in FIG. 3 is executed by each learning device 302.

  The global constraint parameter w finally updated by at least one learning device 302 is transmitted to the learning control device 301 via the network 303. The learning control device 301 outputs the global constraint parameter w received by the learning device 302 by the output unit 330.

  As described above, according to the ranking function learning system according to the third embodiment, since the ranking function learning process by the distributed parallel processing is performed by a plurality of learning devices connected via the network, the processing is performed at high speed. And can cope with further enlargement of training data.

  In the above embodiment, a technique corresponding to the sequential addition of training data in the above-described second embodiment may be applied. In this case, every time training data is added, the number of learning devices that perform ranking learning processing may be increased.

<Experimental example>
Next, the results of experiments using the ranking function learning method proposed in the embodiment of the present invention will be described.

  Experiments were conducted for each case where the training data set was static and where training data was added sequentially. As a comparative method, the method described in Reference 3 (Ryan McDonald, Keith Hall, and Gideon Mann. Distributed training strategies for the structured perceptron. In NAACL HLT, 2010.) was used.

  FIG. 10 shows the experimental results. The vertical axis of the graph of FIG. 10 is the accuracy, and the horizontal axis is the number of repetitions during the learning process. “Batch” represents an experimental result when all training data is used from the beginning, and “Stream” represents a field blue experimental result in which training data is sequentially input. “Batch ADMM” and “Stream ADMM” represent the experimental results of the proposed ranking function learning method proposed in the embodiment of the present invention, and “Batch Iter. Param Mix” and “Stream Iter. Param Mix” The experimental results when the comparison method is used are shown. It was found that when the ranking function learning method proposed in the embodiment of the present invention is used, the accuracy is higher than that of the conventional method. Moreover, even if it inputs sequentially, it turned out that almost the same solution as when using all training data from the beginning is obtained.

  In addition, an experiment using a method of distributed parallel processing was performed. Experiments were conducted when the number of computers for learning processing was 1, 12, 36, and 72, respectively. FIG. 11 shows the experimental results. The vertical axis of the graph in FIG. 11 is the error rate, and the horizontal axis is the number of repetitions during the learning process. It was found that the error can be reduced by increasing the number of computers that perform distributed parallel processing.

  Note that the present invention is not limited to the above-described embodiment, and various modifications and applications are possible without departing from the gist of the present invention.

  For example, although the ranking function learning device, the learning control device, and the learning device described above have a computer system inside, if the “computer system” uses a WWW system, a homepage providing environment ( Or a display environment).

  In the present specification, the embodiment has been described in which the program is installed in advance. However, the program can be provided by being stored in a computer-readable recording medium.

10, 340 Input unit 20, 220, 320, 350 Calculation unit 21 Training data storage unit 22 Division unit 23, 223 Calculation node 31 Division data storage unit 32 Local update unit 33 Synchronization unit 34 Global update unit 35 Convergence determination unit 100, 200 Ranking function learning device 231 Additional data storage unit 300 Ranking function learning system 301 Learning control device 302 Learning device

Claims (9)

A device for learning parameters related to a ranking function for ranking a search result obtained by searching a set of documents based on a search query,
Training data storage means for storing a set of training data including a feature value obtained for each document of a search result for a search query and a fitness for the search query;
N calculation nodes (N is a natural number of 2 or more) for learning parameters related to the ranking function;
Dividing means for dividing the set of training data stored in the training data storage means into N subsets and assigning it to the N computation nodes;
Each of the N compute nodes is
Local updating means for updating a local parameter that is a parameter related to the ranking function based on a subset of the training data assigned by the dividing means;
Synchronizing means for notifying other local nodes of the local parameters updated by the local updating means, and acquiring the local parameters notified from the other computational nodes;
Based on the local parameters of the other computation nodes acquired by the synchronization means and the local parameters updated by the local update means, parameters relating to the ranking function, and Global update means for updating global parameters for matching the local parameters;
It is determined whether the value of the global parameter has converged, and until it is determined that the value of the global parameter has converged, the update by the local update unit, the notification and acquisition by the synchronization unit, and the global A ranking function learning device comprising: a convergence determining unit that repeats updating by the updating unit. The local updating means is determined in advance using a subset of the training data assigned by the dividing means, a Lagrange undetermined multiplier updated last time, the local parameters, and the global parameters. Updating the Lagrange undetermined multiplier and the local parameter to optimize the value of the objective function;
The synchronization means notifies the Lagrange undetermined multiplier and the local parameter updated by the local update means to another calculation node, and also notifies the Lagrange undetermined multiplier and the local parameter notified from the other calculation node. Specific parameters,
The global update means is based on the Lagrange undetermined multiplier and the local parameter of the other computation node acquired by the synchronization means, and the Lagrange undetermined multiplier and the local parameter updated by the local update means. The ranking function learning device according to claim 1, wherein the global parameter is updated so as to optimize the value of the objective function. An additional computation node that learns parameters related to the ranking function when the set of training data is added to the training data storage means;
The additional compute node is
Local update means for updating the local parameters based on the set of the added training data;
Synchronizing means for notifying other local nodes of the local parameters updated by the local updating means, and acquiring the local parameters from the other computational nodes;
Global update means for updating the global parameters based on the local parameters of the other computation nodes acquired by the synchronization means;
It is determined whether the value of the global parameter has converged, and until it is determined that the value of the global parameter has converged, the update by the local update unit, the notification and acquisition by the synchronization unit, and the global A convergence determination unit that repeats updating by the updating unit,
The synchronization means of each of the computation nodes notifies the other computation nodes and the additional computation nodes of the local parameters updated by the local update means, and the other computation nodes and the additional computations. The ranking function learning device according to claim 1, wherein the local parameter notified from the node is acquired. A device for learning parameters related to a ranking function for ranking a search result obtained by searching a set of documents based on a search query,
Training data storage means for storing a set of training data including a feature value obtained for each document of a search result for a search query and a fitness for the search query;
N calculation nodes (N is a natural number of 2 or more) for learning parameters related to the ranking function;
Dividing means for dividing the set of training data stored in the training data storage means into N subsets and assigning the set to the N computing nodes;
Including global update means and convergence determination means,
Each of the N compute nodes is
Local updating means for updating a local parameter that is a parameter relating to the ranking function based on a subset of the training data assigned by the dividing means;
The global update means is a parameter related to the ranking function based on the local parameters updated by the local update means of all the calculation nodes, and the local parameters of each calculation node A means of updating global parameters to match,
The convergence determination means determines whether or not the value of the global parameter has converged, and updates the calculation nodes by the local update means until it is determined that the value of the global parameter has converged and A ranking function learning device, characterized in that it is means for repeating the update by the global update means. Training data storage means for storing a set of training data including the feature value obtained for each document of the search result for the search query and the fitness for the search query, and N calculations (N is a natural number of 2 or more) A ranking function learning method in an apparatus for learning parameters related to a ranking function for ranking a search result obtained by searching a set of documents based on a search query, including nodes and dividing means,
Dividing the set of training data stored in the training data storage means by the dividing means into N subsets and assigning it to the N computing nodes;
Learning parameters related to the ranking function by each of the N computational nodes;
The step of learning by the computation node comprises:
Updating a parameter related to the ranking function and a local parameter based on a subset of the training data assigned by the dividing unit by a local updating unit;
Notifying other local nodes of the local parameters updated by the local updater by means of synchronization, and obtaining the local parameters notified from the other local nodes;
A parameter related to the ranking function based on the local parameter of the other computation node acquired by the synchronization unit and the local parameter updated by the local update unit by a global updating unit; and Updating global parameters to match the local parameters of each compute node;
The convergence determining means determines whether or not the global parameter value has converged, and until it is determined that the global parameter value has converged, the updating by the local updating means, the notification by the synchronizing means, and A ranking function learning method including: obtaining and repeating updating by the global updating means. The step of updating by the local updating means uses a subset of the training data allocated by the dividing means, a Lagrange undetermined multiplier updated last time, the local parameter, and the global parameter, Updating the Lagrange undetermined multiplier and the local parameter to optimize the value of a predetermined objective function;
The step of notifying and acquiring by the synchronization means notifies the Lagrangian undetermined multiplier and the local parameter updated by the local update means to other calculation nodes and the notification from the other calculation nodes. Obtaining a Lagrange undetermined multiplier and the local parameters;
The step of updating by the global updating means includes the Lagrange undetermined multiplier and the local parameter of the other calculation node acquired by the synchronizing means, and the Lagrange undetermined multiplier and the local parameter updated by the local updating means. The ranking function learning method according to claim 5, wherein the global parameter is updated so as to optimize the value of the objective function based on a correct parameter. Learning a parameter related to the ranking function when the set of training data is added to the training data storage means by an additional computation node;
Learning with the additional compute node comprises:
Updating the local parameters based on the set of the added training data by a local updating means;
Notifying other local computing nodes of the local parameters updated by the local updating means by means of synchronization, and obtaining the local parameters from the other computational nodes;
Updating the global parameter based on the local parameter of the other computation node acquired by the synchronization unit by a global update unit;
The convergence determining means determines whether or not the global parameter value has converged, and until it is determined that the global parameter value has converged, the updating by the local updating means, the notification by the synchronizing means, and And repeating the acquisition and updating by the global updating means,
The step of notifying and obtaining by each synchronization means of the computation node notifies the other computation node and the additional computation node of the local parameter updated by the local update means, and The ranking function learning method according to claim 5 or 6, wherein the local parameter notified from a node and the additional calculation node is acquired. Training data storage means for storing a set of training data including the feature value obtained for each document of the search result for the search query and the fitness for the search query, and N calculations (N is a natural number of 2 or more) A ranking function learning method in an apparatus for learning parameters related to a ranking function for ranking a search result obtained by searching a set of documents based on a search query, including a node, a dividing unit, a global updating unit, and a convergence determining unit There,
Dividing the set of training data stored in the training data storage means by the dividing means into N subsets and assigning it to the N computing nodes;
Learning a parameter related to the ranking function by each of the N computational nodes;
Updating by the global updating means;
Determining by the convergence determination means,
The step of learning by the computation node comprises:
Updating a parameter related to the ranking function and a local parameter based on a subset of the training data assigned by the dividing unit by a local updating unit;
The step of updating by the global updating means is a parameter relating to the ranking function based on the local parameter updated by the local updating means of all the computation nodes, and the local of each computation node Updating global parameters to match the global parameters,
The step of determining by the convergence determination means determines whether or not the value of the global parameter has converged, and the local update means of each calculation node until it is determined that the value of the global parameter has converged The ranking function learning method is characterized in that it is a step of repeating the update by the update by the global update means.   The program for functioning a computer as each means of the ranking function learning apparatus of any one of Claims 1-4.