JPH10228487A - Automatic learning system for user taste - Google Patents

Abstract

PROBLEM TO BE SOLVED: To efficiently learn the taste of a user with respect to a content through the use of a pair of user contents given as data and the accurate liking degree. SOLUTION: When a pair of the user contents are inputted, a prediction part 5 predicts the liking degree of the pair by a present hypothesis (the value of the liking degree of past learning data stored in learning data storage part 2 and inter-user weight and inter-content weight, which are stored in storage parts 3 and 4). A learning part 6 updates the present hypothesis when the accurate liking degree of the pair is inputted. At that time, inter-user weight is updated as the function of a difference between the liking degree with respect to a pair of the user contents in past learning data and the liking degree with respect to a pair of the user contents which are newly given, and inter-content weight, and updates it as the function of the difference between the liking degree with respect to a pair of the user contents in past learning data and the liking degree with respect to a pair of the user contents which are newly given, and inter-user weight.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a user preference learning technique for providing content that matches individual user preferences in a content service such as a news article or a homepage on a computer network or the like. is there.

[0002]

2. Description of the Related Art In a learning-type navigation system which learns user preferences for contents such as news articles and homepages on a computer network or the like, and provides contents matching the preferences of individual users, a learning-type navigation system is provided. Conventionally known methods for learning the preference for a content include a “content-based method” in which prediction and learning are performed based on past preference for similar content by the same user, and prediction and learning based on similar user preference for the same content. It can be broadly divided into "user-based method" for learning.

As an example of the former content-based system, for example, NewsWeeder by Ken Lang of Carnegie Mellon University is known, and Proceedings of the
12th International Confes
rence on Machine Learning
Paper "NewsWeeder: Learnnin"
g to filtersnews ". Also, as an example of the content-based method, there is a sequential learning method by Atsuyoshi Nakamura of C & C Research Laboratories of NEC Corporation, and a paper "Bulletin Variable Real Polynomials" Learning of Preference Function Used ”.

As an example of the latter user base method, M
"Group ・ by Paul Resnick of IT
Lens (Group Lens) "is known,
The article “GroupLens: An Open Ar” published in the “Roceedings of CSCW” (1994).
feature for Collaborat
ive Filtering of Networks "
In more detail.

[0005]

In recent years, the above two types of methods have been combined, and information on the past tastes of similar users has been obtained.
The importance of a method of learning and predicting the user's preference by using both pieces of past preference information for similar contents has been strongly recognized, and an invention of an effective method for achieving this has been long-awaited. The present invention proposes a specific and effective method that exactly satisfies this condition.

[0006]

[Means for Solving the Problems]

(A) The user preference automatic learning method according to the present invention learns user preference for contents such as news articles and homepages using data consisting of a content given as data and a numerical value representing the user preference level for the content. A learning data input unit for inputting learning data consisting of a user-content pair and a numerical value representing the user's preference for the content, and a learning data storage unit for collectively storing learning data given in the past And an inter-user weight storage unit for storing inter-user weights representing similarities between users, an inter-content weight storage unit for storing inter-content weights representing similarities between contents, and the learning data input unit. Learning data stored in the learning data storage unit, A prediction unit that predicts the user's preference for the content using the weight between users stored in the weight storage unit between users and the weight between contents stored in the weight storage unit between contents, and the learning data Correct preference level of the user / content pair input from the input unit, learning data given in the past stored in the learning data storage unit, user weight stored in the user weight storage unit, and the content A learning unit that updates the inter-user weight stored in the inter-user weight storage unit and the inter-content weight stored in the inter-content weight storage unit using the inter-content weight stored in the inter-weight storage unit. , The prediction in the prediction unit,
A weighted average in which the value of the preference level of the past learning data stored in the learning data storage unit is a product of the weight between users and the weight between contents between a newly given user / content pair. Weight updating in the learning unit is performed for each user-content pair in the past learning data with respect to the weight between the user and the content between the newly-given user-content pair. And the weight between users is calculated based on the difference between the preference level for the user-content pair in the past learning data and the preference level for the newly-given user-content pair. The inter-content weight is updated as a function of the inter-content weight and the preference between the user content pair in the past learning data and the newly given user content. Has a difference between the preference degree for a pair, the configuration to be updated as a function of the user between the weights.

In the automatic user preference learning method of the present invention configured as shown in (A), when a user content pair is input from the learning data input unit, the prediction unit sets the user content pair. For the current hypothesis, the past learning data stored in the learning data storage unit, the inter-user weight stored in the inter-user weight storage unit, and the inter-content weight stored in the inter-content weight storage unit are used as current hypotheses. As the predicted value of the user's preference for the content, the product of the user-to-user weight and the content-to-content weight between the newly given user-content pair of the value of the preference of the past learning data is used as the weight, Determined by weighted average. When the correct preference level of the user / content pair is input from the learning data input unit, the learning unit sets the correct preference level, the previously given learning data stored in the learning data storage unit, and the weight between users. Using the current inter-user weight stored in the storage unit and the current inter-content weight stored in the inter-content weight storage unit, the inter-user weight and inter-content weight storage unit stored in the inter-user weight storage unit are used. The stored weight between contents, that is, the current hypothesis is updated. Specifically, for each user-content pair in the past learning data, both the inter-user weight and the inter-content weight between the newly given user-content pair are updated. The weight between users is
The difference between the preference level for the user / content pair in the past learning data and the preference level for the newly given user / content pair is updated as a function of the weight between the contents, and the weight between the contents is determined based on the past learning weight. Update is performed as a function of the difference between the preference level for the user content pair in the data and the preference level for the newly given user content pair, and the inter-user weight. By updating such weights, learning with a faster rise becomes possible as compared with a case where only similarity between users is considered or a case where only similarity between contents is considered. In the above operation example, when a user / content pair is given, prediction and learning are performed. However, this is a so-called sequential learning type operation in which learning is performed every time. During learning, it is possible to perform only learning without performing prediction, and after a certain stage of learning, a so-called batch type operation of performing only prediction can be performed.

(B) In another automatic user preference learning method according to the present invention, the learning unit may be configured to compare each user content pair in the past learning data with a newly given user content pair. Instead of updating both the inter-user weight and the inter-content weight during the period, a random number is used to randomly update the inter-user weight or the inter-content weight.

(C) Further, another automatic user preference learning method according to the present invention is a method for recognizing a user's preference for a content such as a news article or a homepage from a content given as data and a numerical value representing the user's preference for the content. A learning data input unit for inputting learning data consisting of a user / content pair and a numerical value representing the user's preference for the content, and a learning data provided in the past collectively. A learning data storage unit for storing a plurality of learning methods; a learning method storage unit for storing a plurality of learning methods; and a user content pair given to the learning data input unit. Prediction by each learning method stored in the method storage unit. A prediction unit that predicts a weighted average value representing a degree of reliability, and, for each user content pair of a question pair candidate set composed of a plurality of user content pairs, the learning method storage unit for the user's preference for the content. And a query section for inquiring the correct preference level of the user content pair having the largest variation in the predicted value and obtaining the value, and the user content input from the learning data input section. Based on the correct preference degree of the pair and the correct preference degree of the user content pair obtained by asking the question section, the hypothesis held by each learning method in the learning method storage section is updated, and the prediction section is updated. The weight representing the reliability for each learning method used by the And a learning unit for updating as a function of the difference between the good degree.

[0010] In the automatic user preference learning method of the present invention configured as shown in (C) above, when a user content pair is input from the learning data input unit, the prediction unit sets the user content pair. , The user's preference for the content is predicted by each of the learning methods stored in the learning method storage unit, and is predicted using a weighted average value of the predicted values. Is input from the learning data input unit, the learning unit updates the hypothesis held by each learning method in the learning method storage unit based on the correct preference level, and updates the learning method used by the prediction unit. The weight representing the reliability is updated as a function of the difference between the current value of each weight and the predicted value of each learning method and the correct preference given as input. As described above, by updating the hypothesis of each learning method and updating the weight representing the reliability of each learning method, the prediction value of the more reliable learning method has a greater influence as the learning progresses. As a result, highly accurate prediction can be performed with a small amount of learning data.

[0011] On the other hand, the interrogation section reduces the trouble of the user who inputs the user / content pairs to be learned by determining the user / content pairs to be learned by the system itself, and at the same time, makes the learning more efficient. to enable. That is, for each user content pair in the question pair candidate set consisting of a plurality of user content pairs, the question unit predicts the degree of preference of the user for the content by each learning method in the learning method storage unit. The correct preference level of the user / content pair having the largest variation in the predicted value is inquired to obtain the value. Then, the learning unit updates the hypothesis held by each learning method in the learning method storage unit based on the correct preference level answered by the user for the user / content pair queried by the question unit, and uses the prediction unit to The weight indicating the reliability for each learning method to be updated is updated. By performing learning based on a user-content pair having a large amount of information in this way, highly accurate prediction can be performed with a small number of questions.

The order of learning is a sequential expression in which one user / content pair is input from the learning data input unit, prediction is performed, and then a correct preference is input from the learning data input unit to learn. Alternatively, a batch type in which only learning is performed without performing prediction during learning and only prediction is performed after learning has progressed to some extent. In addition, regarding the relationship between the input from the learning data input unit and the question by the question unit, the user unit receives one user content pair from the learning data input unit, processes it, One question may be asked, or a plurality of user / content pairs may be input from the learning data input unit, and after processing them, some questions may be issued from the question unit.

(D) Another automatic user preference learning method according to the present invention is the automatic user preference learning method according to the above (C), wherein the learning method stored in the learning method storage unit includes the user preference described in the above (B). Use the automatic learning method. That is, the individual learning methods stored in the learning method storage unit are:
It includes an inter-user weight storage unit that stores inter-user weights indicating similarities between users, an inter-content weight storage unit that stores inter-content weights indicating similarities between contents, a prediction unit, and a learning unit. The prediction unit in each of the learning methods includes, for a given user / content pair, learning data stored in a learning data storage unit, an inter-user weight stored in an inter-user weight storage unit, and an inter-content weight storage unit. Is used to predict the degree of preference of the user for the content. At this time, the prediction is made by multiplying the product of the inter-user weight and the inter-content weight between the value of the preference degree of the past learning data stored in the learning data storage unit and the newly given user / content pair by the weight. , By weighted averaging. Also,
The learning unit in each learning method includes: a correct preference degree of the user / content pair; learning data given in the past stored in the learning data storage unit; a user weight stored in the user weight storage unit; The inter-user weight stored in the inter-user weight storage unit and the inter-content weight stored in the inter-content weight storage unit are updated using the inter-content weight stored in the weight storage unit. At this time, the weight between users or the weight between contents is updated at random using random numbers. That is, although the learning methods in the learning method storage unit are the same learning method, different random numbers are generally generated in updating the weight between users and the weight between contents, and different weights are updated. Specifically, there are a plurality of learning methods that perform different predictions in a specific scene.

Further, in another automatic user preference learning method according to the present invention, in the automatic user preference learning method of the above (D), the interrogation unit may include a set of a plurality of user content pairs as past question pair candidates. Of user / content pairs consisting of the user and the content appearing in the learning data of the above, all or a part of a set of pairs to which no preference is given as data in the past.

[0015]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Next, embodiments of the present invention will be described in detail with reference to the drawings.

Referring to FIG. 1, a first embodiment of an automatic user preference learning system according to the present invention is a learning data input unit 1.
A learning data storage unit 2, an inter-user weight storage unit 3,
It comprises an inter-content weight storage unit 4, a prediction unit 5, and a learning unit 6. These are, for example, a data processor (computer) having a program-controlled CPU, an input device such as a keyboard, and a storage device such as a magnetic disk.
Can be realized.

The learning data input section 1 is a section for sequentially inputting learning data. The input learning data is composed of a user / content pair and a numerical value representing the user's correct preference for the content.

The learning data storage unit 2 stores learning data given in the past collectively, and the inter-user weight storage unit 3 stores a weight matrix (a matrix of inter-user weights) representing the similarity between users. The inter-content weight storage unit 4 stores a weight matrix (a matrix of inter-content weights) representing the similarity between contents.

The prediction section 5 is a section for predicting the user's preference for the content with respect to the user / content pair input from the learning data input section 1. This prediction is based on past learning data stored in the learning data storage unit 2, user weights stored in the user weight storage unit 3, and content weights stored in the content weight storage unit 4. This is done with reference to.

When new learning data (a user-content pair and a numerical value representing the correct preference) is input from the learning data input unit 1, the learning unit 6 corrects the correct preference of the input user-content pair. Using the learning data given in the past stored in the learning data storage unit 2, the user weight stored in the user weight storage unit 3, and the content weight stored in the content weight storage unit 4, This section updates the inter-user weight stored in the inter-user weight storage unit 3 and the inter-content weight stored in the inter-content weight storage unit 4.

FIG. 2 is a flowchart showing an example of the operation of the present embodiment. In this operation example, one user / content pair is input from the learning data input unit 1, and the user's preference level for the content is predicted by the prediction unit 5. This is a case in which a learning method of repeating a process of inputting a degree of preference and updating a weight between users and a weight between contents in the learning unit 6 is employed.

Assuming that a set of users is X and a set of contents is Y, the following steps are executed in the first embodiment of the present invention. (1) A pair of a new user x @ X and a new content y @ Y is input in the learning data input unit 1 (step S)
1). (2) With respect to the input pair of the user x and the content y, the preference degree f (x,
y) is predicted by the prediction unit 5 using the current hypothesis (step S2). It is assumed that the value of f (x, y) is generally a real value. For example, a five-level evaluation (1, 2,..., 5) is assumed. (3) Correct preference degree f of user x for content y
The value of (x, y) is input by the learning data input unit 1 (step S3) and stored in the learning data storage unit 2 (step S4). (4) The current hypothesis is updated (learned) in the learning unit 6 (step S5). (5) It is determined whether to end the learning, and if not, the process returns to step S1 (step S6).

The goal of the prediction / learning algorithm is to perform highly accurate prediction with as few learning data as possible. Various conditions can be considered as conditions for terminating in step S6 (stop conditions), but there is no essential effect on the present invention, and since it largely depends on the application environment, it is not particularly specified here. In addition, the above (1)-
(4) may be repeated endlessly.

In the prediction step (2), the following hypothesis format is used. First, it is assumed that the value of the correct preference degree f (x, y) given in the past is stored in the learning data storage unit 2 as the data matrix M. At this time, for x and y given the value of f (x, y) in the past, M
(X, y) = f (x, y), and for other x, y, M (x, y) = *. Here, * indicates that no setting has been made. Next, an arbitrary user pair (x _1, x
₂ ) For ∈X × X, a weight matrix u between users describing weights u (x _1, x ₂ ) representing similarity between the users
Is provided in the inter-user weight storage unit 3. Similarly, for an arbitrary content pair (y _1, y ₂ ) ∈Y × Y, a weight matrix v between contents describing a weight v (y _1, y ₂ ) representing the similarity between the contents is given by Are provided in the weight storage unit 4 between contents. The current hypothesis is that the data matrix M
, A weight matrix u between users, and a weight matrix v between contents. Then, the newly given (x, y) ∈X
The preference of xY is predicted using a weighted average of all data stored in the data matrix M. here,
The weights for taking the weighted average are: new user, content pair x, y, user in data matrix M,
Assuming that the content pair is x ′, y ′, the product of the weight between the users and the weight between the contents, u (x, x ′) · v (y,
y ′) is used. That is, the predicted value f ＾ (x, y) of the preference of the new pair (x, y) is calculated as follows.

[0025]

(Equation 1)

The use of the product of the weight between users and the weight between contents as the weight has the following meaning. The value of f (x ', y') is close to the value of f (x, y) only when x, x 'and y, y' both have high similarity. (X, y) and (x ', y')
Is similar to the proposition that x and x 'are similar and the proposition that y and y' are similar. However, logical learning is dangerous when learning is performed from an environment containing noise and from incomplete data. Therefore, in this example, a real number product by weight representing similarity is used instead of the logical product. By doing so, prediction with robustness to noise is realized.

Next, a learning procedure, that is, a procedure for updating the current hypothesis will be described. Now, it is assumed that the correct answer f (x, y) of the preference degree for the new pair x, y is given. For each correct answer f (x ′, y ′) given in the past, if | f (x, y) −f (x ′, y ′) | takes a large value, (x, y) And the proposition that (x ', y') are similar are denied. From the observations above, this translates to the proposition that (x, x ') is similar and (y,
Since the logical product of the proposition that y ′) is similar is negated, the proposition that (x, x ′) is not similar to the proposition that (y, y ′) is not similar Either will be correct. However, it is not possible to identify which is correct from this information. Therefore, the estimated value u (x, x ′) of the similarity of (x, x ′) learned from the past data and the estimated value v (y, y ′) of the similarity of (y, y ′) The following inference is made with reference to this. That is, if the value of u (x, x ') is large (x, x')
Is strongly similar, the newly given proposition dictates the proposition that (y, y ') is dissimilar, and conversely, if the value of v (y, y') is If there is strong evidence that large (y, y ') are similar,
Indicate the proposition that (x, x ') is not similar.
The judgment of “large” or “strong” can be inferred as a logical value using a threshold or the like, but it is dangerous in learning using incomplete data from a noisy environment. | F (x, y) -f
(X ′, y ′) | and the value of u (x, x ′) determine the update width of v (y, y ′) continuously,
The magnitude of | f (x, y) −f (x ′, y ′) | and v
Depending on the magnitude of the value of (y, y ′), u (x,
x ') is updated. For example, the weight u between users
Is updated as follows,

[0028]

(Equation 2)

The weight v between contents is updated as follows.

[0030]

(Equation 3)

Note that c and d in the above equation are positive constants. c
Is called a learning rate, and can be changed depending on the number of data. In d, | f (x ′, y ′) − f (x,
y) It is appropriate to use an estimated value or the like of the average value of |. When performing the above update, it is also possible to specify the upper limit of the amount of change in the weight in one update. For example, the weight u between users is updated as follows,

[0032]

(Equation 4)

The weight v between contents may be updated as follows.

[0034]

(Equation 5)

Further, in order to avoid problems such as underflow, normalization may be performed as follows each time weights are updated.

[0036]

(Equation 6)

[0037]

(Equation 7)

By using the above-described weight updating method, it is possible to realize robust inductive inference with respect to noise, and at the same time, when only the similarity between users is considered, or when only the similarity between contents is considered. As compared with the case where is taken into account, learning with a faster start-up becomes possible.

Next, a specific example of the estimation of the preference level and the updating of the weight matrix will be described. As an example, as shown in FIG.
5 users from 1 to u5 and 7 from c1 to c7
F (u) in a state where all elements of the data matrix M indicating the past learning data in the learning data storage unit 2 are not set (indicated by? In FIG. 4)
2, c3) = 4, f (u4, c3) = 4, f (u2, c
6) It is assumed that learning data is given in the order of 1, f (u4, c6) = 1. Further, the weight between users provided in the weight storage between users 3 and the weight between users provided in the weight between contents storage unit 4 are each an initial value of 1
It is assumed that is set to Further, the constant c is -lo
It is assumed that g 0.8 is set, that is, e ^−c = 0.8, and d is 2. Also, the weights u (x, x), v
(Y, y) is not updated. At this time, the update of the predicted value and the weight of the preference degree after each learning data is given is as follows.

1. f (u2, c3) = 4 In this case, since there is no past learning data, the weight is not updated. Further, the predicted value of the preference degree is arbitrary.

2. f (u4, c3) = 4 ○ Preference value

(Equation 8) ○ Update of weights (a) (u2, c3) v. s. (U4, c3)

(Equation 9)

3. f (u2, c6) = 1 ○ Predicted value of preference

(Equation 10) Update of weight (a) (u4, c3) v. s. (U2, c6)

[Equation 11] (B) (u2, c3) v. s. (U2, c6)

(Equation 12)

4. f (u4, c6) = 1 ○ Predicted value of degree of preference

(Equation 13) ○ Update of weights (a) (u2, c3) v. s. (U4, c6)

[Equation 14] (B) (u4, c3) v. s. (U4, c6)

(Equation 15) (C) (u2, c6) v. s. (U4, c6)

(Equation 16)

Thereafter, the weight is updated each time the learning data is sequentially provided, and the accuracy of the predicted value of the preference level increases accordingly. FIG. 3 shows an example of a preference degree (an example of a user preference function). In the user preference function of this example, for example, each user u1 to content c1 and content c3
Since the preference degree of u5 is similar, the content c1
And the content c3 have a high similarity. In addition, since the preference level of the user u2 and the preference level of the user u4 for each of the contents c1 to c7 are similar, it can be seen that the user u2 and the user u4 have a high similarity degree.

Next, a second embodiment of the present invention will be described.

In the first embodiment described above, each time a correct answer to the preference for a new user / content pair is given, the weight u between users and the weight v between contents are given.
In the second embodiment, either one of them is updated at random using a random number using the update formula or the like in the first embodiment. Other than that, it is basically the same as the first embodiment. At this time, it is possible to perform each update with a probability of 1/2,
It is also possible to do this randomly with a probability depending on the associated weight. For example, the weight u (x, x ′) between the users is updated with the following probability p, and when the weight between the users is not updated (that is, the probability 1−p), the weight v between the contents is
It is conceivable to update (y, y ').

[0047]

[Equation 17]

Next, a third embodiment of the present invention will be described.

Referring to FIG. 5, a third embodiment of the automatic user preference learning method according to the present invention is
1, a learning data storage unit 12, a prediction unit 13, a learning method storage unit 14, a learning unit 15, and an interrogation unit 16. These are, for example, program-controlled C
It can be realized by a data processing device (computer) having a PU, an input device such as a keyboard, a storage device such as a magnetic disk, a display device, and the like.

The learning data input section 11 is a section for sequentially inputting learning data. The input learning data is
It consists of a user content pair and a numerical value indicating the user's correct preference for the content.

The learning data storage section 12 is a section for collectively storing learning data given in the past.

The learning mode storage section 14 is a section for storing a plurality of learning modes. Here, each learning method has a function of predicting a user's preference for the content based on a current hypothesis for a given user-content pair, and a function of predicting the user's preference for the content pair. It has a function to update (learn) the current hypothesis based on it.

The predicting section 13 is a section for predicting the user's preference for the content with respect to the user / content pair input from the learning data input section 11. In this prediction, the degree of preference of the user for the content is predicted by each learning method stored in the learning method storage unit 14, and the weight of the current reliability for each learning method is used to calculate the prediction value of the prediction value. This is performed by taking a weighted average value.

The question section 16 is a section that issues a question specifying a user / content pair by itself and receives an input of a preference level for the user / content pair. The question is displayed on a display screen, for example, and the answer is input from a keyboard or the screen.

The learning unit 15 is a unit that updates the current hypothesis of each learning method stored in the learning method storage unit 14 and the weight indicating the reliability of each learning method.

FIG. 6 is a flowchart showing an example of the operation of the present embodiment. In this operation example, one user / content pair is input from the learning data input unit 11, the user's preference for the content is predicted by the prediction unit 13, and then the user / content pair of the user / content pair is predicted from the learning data input unit 11. Input the correct preference and learner 1
In 5, the current hypothesis of each learning method and the weight indicating the reliability of each learning method are updated, and, if possible, a question is asked about the preference of one user / content pair in the question section 16, and according to the answer, The learning unit 15 performs sequential learning in which the process of updating the current hypothesis of each learning method and the weight indicating the reliability of each learning method is repeated. Less than,
This embodiment will be described in detail.

Assuming that a set of users is X and a set of contents is Y, in the present embodiment, the following steps are executed as shown in FIG. (1) A pair of a new user x @ X and a new content y @ Y is input at the learning data input unit 11 (step S1).
1). (2) With respect to the input pair of the user x and the content y, the preference degree f (x,
y) is predicted by the prediction unit 13 using the predicted values of the plurality of learning methods stored in the learning method storage unit 14 and weights representing the current reliability of each learning method (steps 12 and 1).
3). It is assumed that the value of f (x, y) is generally a real value. For example, a five-level evaluation (1, 2,..., 5) is assumed. (3) Correct preference degree f of user x for content y
The value of (x, y) is input by the learning data input unit 11 (Step S14) and stored in the learning data storage unit 12 (Step S15). (4) The learning section 6 updates (learns) the current hypotheses of the plurality of learning methods based on the correct preference input in step S14 (step S16), and assigns a weight indicating the reliability of each learning method. Update (step S17). (5) Using the current hypotheses of a plurality of learning methods, the question section 16 obtains a new pair of the user x ′ and the content y ′ for which the preference is to be asked, and inputs the correct preference for the pair from the user. (Step S18). (6) The learning data including the pair of the user x ′ and the content y ′ and the correct preference is stored in the learning data storage unit 12 (step S19). (7) The learning unit 6 updates (learns) the current hypotheses of the plurality of learning methods based on the correct preference input in step S18 (step S20), and assigns a weight indicating the reliability of each learning method. Update (step S21). (8) It is determined whether to end the learning, and if not, the process returns to step S11 (step S22).

The goal of the prediction / learning algorithm is to perform highly accurate prediction with as few learning data and the number of questions as possible. Various conditions such as the achievement of a certain level of learning accuracy can be considered as the condition for terminating in step S22 (stop condition), but there is no essential effect on the present invention. , It is not particularly specified here. Note that the above (1) to (7) may be repeated endlessly.

In the present embodiment, the learning method storage unit 14 performs high-precision prediction using a small amount of learning data.
Is achieved by predicting the degree of preference in the user-content pair given to each of the plurality of learning methods stored in the learning method, and performing prediction using a weighted average using weights representing the reliability of the learning methods. . That is, the value of f (x, y) is predicted by f ＾ (x, y) below, where A _i (x, y) is the predicted value of each learning method and w _i is its weight.

[0060]

(Equation 18)

Note that the correct preference degree f for (x, y)
When (x, y) is given as an input, each learning method A
the weights w _i of _i is updated as follows.

[0062]

[Equation 19]

By the above-described weight updating, learning progresses, and the prediction value of the learning method having higher reliability has a large influence.

Next, in order to predict with high accuracy with a small number of questions, it is necessary to ask the user / content pair having a large amount of information about its preference. For this purpose, a user / content pair in which predicted values of a plurality of learning methods stored in the learning method storage unit 14 vary are found, and a preference degree for the pair is inquired. A certain user
The variation of the predicted value with respect to the content pair can be measured by a variance of the predicted value or a weighted average of the square error of the predicted value with respect to the weighted average preference using the weight of each learning method. A method of finding pairs with a large variation in predicted values is, for example, to select a pair from among a large number of user content pairs generated using random numbers and to select a pair whose predicted value is most varied in the above sense. The method can be adopted. That is, the user content pair (x ^* , y ^* ) used for the question can be obtained as follows, where P is a set of user content pairs considered as candidates.

[0065]

(Equation 20)

In the case where the obtained user / content pair already exists in the learning data, a question of a pair candidate which does not exist in the learning data may be selected from the candidates after the next point.

Next, a fourth embodiment will be described.

The fourth embodiment is different from the third embodiment in that an algorithm for realizing the user preference automatic learning method of the second embodiment is used as a plurality of learning methods stored in the learning method storage unit 14. Use multiple copies. That is, although each uses the same learning method,
Since different random numbers are generated in updating the weight between users and between contents, and different weights are updated, there are a plurality of learning methods for performing different predictions in a specific scene. In particular, random numbers are generated from the proposition that (x, y) and (x ', y') are not similar,
The proposition that (x, x ') is not similar to (y,
This is done to resolve the ambiguity of which of the propositions that y ') is dissimilar, so that multiple copies will coexist with different conclusions about these ambiguities. By predicting the actual preference with a weighted average of those predictions by weights representing the reliability of each learning method, it is possible to perform learning with higher accuracy, and to reduce the dispersion of the predictions. By asking the user / content pair to be maximized, a question with a large amount of information can be realized.

Next, a fifth embodiment will be described.

In the fifth embodiment, as in the fourth embodiment, the user preference automatic learning method of the second embodiment is realized as a plurality of learning methods stored in the learning method storage unit 14. It uses multiple copies of the algorithm. Further, in the third and fourth embodiments, a method of generating and comparing a large number of pairs has been adopted for finding pairs with a large variation in predicted values. However, in the fifth embodiment, A specific method is used for finding a pair having a large variation in the predicted value.

That is, as a question pair candidate set including a plurality of content and user pairs, a preference level is given as data in the past in a user / content pair including a user and a content appearing in past learning data. This is a method of using all or a part of a set of unpaired pairs and inquiring about the one with the largest variation in predicted values among them.

This method uses f (x, y) and f (x ′,
y ′) is given and the ambiguity that occurs when the difference is found to be large, that is, whether the similarity of x, x ′ is low or the similarity of y, y ′ is low, is concrete. As a solution, it is based on the finding that it is effective to know the value of f (x, y ') or f (x', y) by a question. That is, since a question for such a pair has a large amount of information, a question having a large amount of information can be surely realized by selecting and asking a pair having the largest variation in predicted values among them. .

FIG. 7 shows a configuration example of the fifth embodiment.
As shown in the figure, the present embodiment has a learning data input unit 11 for sequentially inputting learning data and a learning data storage unit 12 for collectively storing learning data given in the past.
, A plurality of learning methods 14-1, 14-2,..., And a user / content pair input from the learning data input unit 11, the preference of each learning method 14-1. , 14-2,..., And takes a weighted average value of the predicted values using a weight representing the reliability of each learning method, and uses the average value as the predicted value of the user / content pair. And a question unit 16 that issues a question specifying a user / content pair by itself and receives an input of a degree of preference for the user / content pair, and each learning method 14-1, 1 stored in the learning method storage unit 14.
4-2,... And a learning unit 15 for updating weights representing the reliability of each learning method.

Each of the learning methods 14-1 and 14-2
Is composed of an inter-user weight storage unit 3, an inter-content weight storage unit 4, a prediction unit 5, and a learning unit 6 as described in FIG. Since the user preference automatic learning method according to the second embodiment is implemented, the learning unit 6 randomly updates the weight between users or the weight between contents using random numbers.

The operation example of the fifth embodiment is the same as the flowchart shown in FIG.

In FIG. 7, one user is
When a content pair is input to the learning data input unit 11 (Step S11), the learning data input unit 11 transmits the content pair to the prediction unit 13. The prediction unit 13 converts the transmitted user content pair into each of the learning methods 14-1, 14-2,.
To predict its preference.

In each of the learning methods 14-1, 14-2,..., Each prediction unit 5 stores the learning data stored in the learning data storage unit 12 and the inter-user weight storage unit 3 in the own method. The user's preference for the content is predicted using the weight between the users and the weight between the contents stored in the weight between the contents storage unit 4 (step S12). That is, based on equation (1), the weight of the preference between the past learning data stored in the learning data storage unit 12 and the weight of the inter-user weight between the newly given user-content pair and the weight of the inter-content weight are set. The weighted average with the product as the weight is used as the predicted value and returned to the prediction unit 13.

The prediction unit 13 determines each of the learning methods 14-1 and 14
−2,... And the learning methods 14-1, 14-2,
.. With the weight representing the current reliability of
Then, a predicted value is calculated according to 8 (step S13).

Next, when the correct preference level of the user / content pair is input from the learning data input unit 11 (step S14), it is stored in the learning data storage unit 12 (step S15), and the learning unit 15 is set. Is transmitted to The learning unit 15 converts the transmitted preference degree into each learning method 14
-1, 14-2,..., And update the weight between users and the weight between contents (step S1).
6). At this time, the learning unit 6 in each of the learning methods 14-1, 14-2,... Updates the inter-user weight with the probability of the above-described expression (17) using the above-described expression (2) and the like, and updates the inter-user weight. If not, the weight between contents is updated by the above-described equation (3) or the like.

The learning unit 15 includes the learning methods 14-1 and 14
Simultaneously with the learning in −2,..., The weight indicating the reliability of each learning method used by the prediction unit 13 is updated by the above-described equation (19) (step S17).

Next, the question section 16 is used as a question pair candidate set including a plurality of user / content pairs. All or a part of a set of pairs to which no preference is given is used, and each learning method 14-1, 14-2,
.. Predict the preference level, and ask the user about the preference level of the user / content pair having the largest variation in the predicted value. At this time, each learning method 14-1, 14-2,
.. Perform prediction in the same manner as when a user / content pair is given from the prediction unit 13. Next, the question unit 16
When the user inputs the correct preference level for the questioned user / content pair, the user inputs the correct preference level to the learning data storage unit 12.
(Step S19) and store it in the learning unit 1.
5

The learning unit 15 transmits the degree of preference of the transmitted learning data to the learning units 6 of the learning methods 14-1, 14-2,..., And updates the weight between users and the weight between contents ( Step S20). At this time, each learning method 14
The learning unit 6 in −1, 14-2,... Updates the inter-user weight with the above-described equation (2) at the probability of the above-described equation (17), and when the inter-user weight is not updated, the above-described equation ( 3) The weight between contents is updated by, for example,. At the same time, the learning unit 15 updates the weight indicating the reliability of each learning method used by the prediction unit 13 using the above-described equation (19) (step S21).

Next, a specific example of prediction of the degree of preference and updating of the weight matrix will be described. As an example, for simplicity, the learning methods stored in the learning method storage unit 14 are 14-1 and 1
4-2. As shown in FIG. 4, assuming five users u1 to u5 and seven contents c1 to c7, all elements of a data matrix M indicating past learning data in the learning data storage unit 12 are all It is not set (indicated by? In FIG. 4). Further, each learning method 14-
It is assumed that the inter-user weights provided in the inter-user weight storage units 3 and 14-2 and the inter-content weights provided in the inter-content weight storage unit 4 are respectively set to the initial value 1. Further, it is assumed that the constant c is set to -log 0.8, that is, e ^-c = 0.8, and d is 2.
The own weights u (x, x) and v (y, y) are not updated. At this time, as learning data, f
It is assumed that (u2, c3) = 4 is given. At this point, no question is asked because the candidate for the pair to be asked is an empty set, and the learning data f (u4, c6) = 1 is given next. For these data, the weight between users or the weight between contents in the learning methods 14-1 and 14-2 is, for example, as follows.

○ Learning method 14-1 f (u2, c3) = 4 No weight update. 2. f (u4, c6) = 1

(Equation 21)

○ Learning method 14-2 f (u2, c3) = 4 No weight update. 2. f (u4, c6) = 1

(Equation 22)

That is, for the learning data of f (u4, c6) = 1, the weight between contents is updated in the learning method 14-1, and the weight between users is updated in the learning method 14-2.

At this stage, the candidate pairs to be asked are:
(U2, c6) and (u4, c3), and the predicted values of the preference degree for each of them by the learning methods 14-1 and 14-2 are as follows. 1. f (u2, c6) (a) Learning method 14-1

(Equation 23) (B) Learning method 14-2

(Equation 24)

2. f (u4, c3) (a) Learning method 14-1

(Equation 25) (B) Learning method 14-2

(Equation 26)

In this case, both learning methods 1 for both candidate pairs
The predicted values of 4-1 and 14-2 are 2.67 and 2.33, respectively.
And 2.33 and 2.67, both pairs show the same variation. At this stage, both pairs have the same amount of information. I do. Here, it is assumed that (u2, c6) is selected and a question is made.

Next, each learning method 14-
The prediction values of the above data by 1, 14-2 and the updating of the weights representing the reliability of each learning method performed as a result are shown below. In the following example, each learning method A _i
The initial value of the weight w _i is 1.0, and is updated as follows. The prediction for the pair given first is three points which are the middle points of the degree of preference.

[0091]

[Equation 27]

In the learning method 14-1, A ₁ (u2, c3) = 3

[Equation 28] 2. A ₁ (u4, c6) = 4

(Equation 29) 3. A ₁ (u2, c6) = 2.33

[Equation 30]

In the learning method 14-2: A ₂ (u2, c3) = 3

(Equation 31) 2. A ₂ (u4, c6) = 4

(Equation 32) 3. A ₂ (u2, c6) = 2.33

[Equation 33]

In the above example, f (u2, c3) = 4
When f (u4, c6) = 1 is given as the learning data, there is an ambiguity that the similarity of either u2, u4 or c3, c6 is low, but either is unknown. Although this ambiguity is randomly resolved in the learning method 14-1 and the learning method 14-2, in the above example, in the learning method 14-1, v
(3, 6) has been updated to reduce
In -2, u (2, 4) is updated in a direction of decreasing (in FIG. 3, since u2 and u4 have a high similarity, the learning method 14-1 makes a more accurate determination. There). As a result, a difference occurs between the prediction values of the two learning methods with respect to the question point selected in the above example, and as a result, the weight of the learning method 14-1 that has performed correct ambiguity resolution is larger. (That is, w ₁ > w ₂ ),
In the prediction for the user / content pair given next, it is understood that the prediction value of the learning method 14-1 is more important.

[0095]

According to the user preference automatic learning method according to the first or second aspect, it is possible to realize inductive inference that is robust against noise, and to consider only the similarity between users, In addition, it is possible to perform learning at a rapid rise as compared with the conventional method in which only the similarity between contents is considered.

According to the user preference automatic learning method according to the third, fourth or fifth aspect, it is possible to achieve highly accurate learning with a significantly smaller number of questions than in the conventional method.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a configuration example of a first embodiment of a user preference automatic learning method according to the present invention.

FIG. 2 is a flowchart illustrating an example of an operation according to the first exemplary embodiment of the present invention.

FIG. 3 is a diagram illustrating an example of a preference level (user preference function).

FIG. 4 is a diagram showing an example of past learning data.

FIG. 5 is a block diagram showing a configuration example of a third embodiment of a user preference automatic learning method according to the present invention.

FIG. 6 is a flowchart illustrating an example of an operation according to the third exemplary embodiment of the present invention.

FIG. 7 is a block diagram illustrating a configuration example according to a fifth embodiment of the present invention.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 ... Learning data input part 2 ... Learning data storage part 3 ... User weight storage part 4 ... Content weight storage part 5 ... Prediction part 6 ... Learning part 11 ... Learning data input part 12 ... Learning data storage part 13 ... Prediction part 14 learning method storage unit 14-1, 14-2 learning method 15 learning unit 16 question unit

Claims (5)

[Claims] 1. An apparatus for learning a user's preference for a content such as a news article or a homepage by using a content provided as data and data comprising a numerical value representing the user's preference for the content. A learning data input unit for inputting learning data consisting of numerical values representing the user's preference for the content; a learning data storage unit for collectively storing learning data given in the past; An inter-user weight storage unit for storing inter-user weights, an inter-content weight storage unit for storing inter-content weights indicating similarities between contents, and a user content pair given to the learning data input unit. , Learning data stored in the learning data storage unit, and storage in the inter-user weight storage unit A prediction unit that predicts the degree of preference of the user for the content by using the obtained weight between users and the weight between contents stored in the content weight storage unit; and the user input from the learning data input unit. The correct preference degree of the content pair, the learning data given in the past stored in the learning data storage unit, the inter-user weight stored in the inter-user weight storage unit, and stored in the inter-content weight storage unit A learning unit that updates the inter-user weight stored in the inter-user weight storage unit and the inter-content weight stored in the inter-content weight storage unit, using the inter-content weight, prediction in the prediction unit, Between the value of the preference degree of the past learning data stored in the learning data storage unit and the newly given user / content pair The weight update in the learning unit is performed by using a product of the weight between the user and the weight between the contents as a weight. Is performed for both the user-to-user weight and the content-to-content weight between the user-content pair given to the user, and the user-to-user weight is determined based on the preference degree for the user-content pair in the past learning data and the new degree. Is updated as a function of the difference between the degree of preference for the user-content pair given to the user and the weight between the contents, and the weight between the contents is newly given to the degree of preference for the user-content pair in the past learning data. The difference between the preference level for the paired user content and
An automatic user preference learning method characterized by having a configuration updated as a function of the inter-user weight. 2. The learning unit according to claim 1, wherein for each user content pair in the past learning data, an inter-user weight between the newly provided user-content pair and an inter-content weight. Instead of updating both,
2. The user preference automatic learning method according to claim 1, further comprising a configuration for randomly updating a weight between users or a weight between contents using a random number. 3. An apparatus for learning a user's preference for a content such as a news article or a homepage by using a content provided as data and data comprising a numerical value representing the user's preference for the content. A learning data input unit for inputting learning data consisting of numerical values representing the user's preference for the content; a learning data storage unit for collectively storing learning data given in the past; and storing a plurality of learning methods. Learning method storage unit, for the user content pair given to the learning data input unit, the degree of preference of the user for the content is predicted by each learning method stored in the learning method storage unit, Predict with a weighted average value of the reliability of each prediction method for those prediction values For each user-content pair of a question pair candidate set composed of a plurality of user-content pairs, the user's preference for the content is predicted by each learning method in the learning method storage unit, and the most predictable A question section for asking the correct preference degree of the user-content pair having a large variation in the value to obtain the value; and a correct preference degree and the question section for the user-content pair input from the learning data input section. Updating the hypothesis held by each learning method in the learning method storage unit based on the correct preference degree of the user / content pair obtained as described above, and weighting the reliability of each learning method used by the prediction unit. As a function of the difference between the current value of each weight and the predicted value of each learning method and the correct preference given as input. User preference automatic learning method characterized by and. 4. The user preference automatic learning method according to claim 3, wherein each of the learning methods stored in the learning method storage unit is the user preference automatic learning method according to claim 2. 5. The question section as a question pair candidate set including the plurality of user / content pairs, a user / user including content and content appearing in past learning data.
5. The user preference automatic learning method according to claim 4, wherein a configuration is used in which all or a part of a set of pairs to which a preference level has not been given as data in the past is used.