JP4967320B2 - Service providing equipment - Google Patents

Description

  The present invention relates to an apparatus that provides services such as presentation of information to a user and automatic adjustment of devices used by the user.

  In recent years, with the explosive spread of computer networks, the amount of information that individuals can contact every day is increasing. As a specific example, the amount of e-mail received by an individual in a day is significantly increased compared to the past. Against this background, the following problems have arisen in recent years. In other words, such a large amount of information usually includes information that is beneficial to computer users and information that is useless. On the other hand, since the amount of information is increasing as described above, the effort for the user to select useful information from these large amounts of information also tends to increase.

  Therefore, in order to selectively present meaningful information from such information or to provide useful information in a timely manner, a network model called a Bayesian network is used to determine the importance and presentation of information. In recent years, a number of information processing techniques for probabilistically estimating timing have been developed. An example is shown in Patent Document 1.

Patent Document 2 estimates the importance of information to be presented before presentation, checks whether the user treats the information as important after presentation, and learns the network model used for the importance estimation processing. An example of executing processing is disclosed.
Japanese translation of publication No. 2004-506961 Japanese translation of PCT publication No. 2002-529820 JP 2003-248647 A

  In the above-described conventional information processing technology using, for example, a Bayesian network, the learning process for estimating the importance or priority of information is performed using explicit training data or implicit training. It has become. However, creating explicit training data is generally time consuming and burdensome to the user. On the other hand, in the case of performing implicit training, for example, as disclosed in Patent Document 3, there is a method in which the user determines from the operation of deleting the corresponding mail. It is clear that e-mail may not be applicable depending on the user's behavior, such as a user who immediately deletes an e-mail regardless of its priority and importance.

  In addition, although there is a description of monitoring the user's activity using the background monitor for the implicit training even in the one disclosed in Patent Document 1, since there is no disclosure about the specific content, It is assumed that this is due to a conventional technique (for example, the simple method described above).

  In the technique disclosed in Patent Document 2, specific information such as a sender's e-mail viewing time is used, and priority and importance are determined based on whether the time exceeds a set threshold. It is supposed to be. In this case, because the amount of documents is extremely small, the system may determine that the email is not important because the browsing time is short even though it is an important email, and incorrect information is fed back. Resulting in.

  As described above, factors that are effective for learning when feedback is provided in the information related to the reaction of the user after presentation vary depending on each user. In other words, if information related to various reactions is not combined, learning with sufficient accuracy corresponding to each user cannot be performed, and as a result, compared to learning from explicit feedback information from users. Therefore, the learning effect is lowered.

  Further, in the conventional information processing technology using, for example, a Bayesian network, for example, it is not possible to cope with newly received e-mails unless the learning process for estimating the importance of information is continuously performed. On the other hand, by continuously performing the learning process, for example, if the user is busy and accidentally handled emails that are considered important under normal conditions, this is reflected in the learning. , And similar emails can be mistakenly determined to be insignificant thereafter.

  As described above, in the conventional technique, it is difficult to improve the accuracy of learning because no consideration is given to whether or not the learning should be performed.

  Such a problem is expected not only in the presentation of information such as e-mails but also in the case of learning control in order to automatically adjust devices that provide various services such as room brightness and air conditioning temperature. When such devices are automatically adjusted and learned by sensing the environment and the user, if the user reacts differently than usual, it will be reflected in the learning, and the subsequent control will be inappropriate. May end up.

  The present invention has been made in view of the above circumstances, and an object thereof is to provide a service providing apparatus that can improve the accuracy of learning.

  The present invention for solving the problems of the conventional example is a service providing apparatus, a providing means for providing a predetermined service to a user, and an occurrence probability of a predetermined event related to the provided predetermined service A first estimating means for calculating determination information related to a result of the calculation, using a result learned based on information related to the service provided in the past and a past occurrence history, The plurality of types of information relating to the reaction of the user who has received the service is acquired, and the provided information is used using the acquired information and the learning result based on the past occurrence history of the predetermined event. A second estimation unit that calculates a probability of occurrence of a predetermined event related to the service and generates second determination information related to a result of the calculation, wherein the first estimation unit is based on the second determination information Provided It is characterized by performing learning using the occurrence history of bis the predetermined event.

  In addition, the present invention for solving the problems of the conventional example is a service providing apparatus, which provides a predetermined service for a user and a predetermined event related to the provided predetermined service. The first estimation that calculates the occurrence probability using the learning result based on the information related to the service provided in the past and the past occurrence history and presents the determination information related to the result of the calculation to the user Using the means and a plurality of types of information related to the reaction of the user who received the service, and using the acquired information and a result learned based on the past occurrence history of the predetermined event, Based on the second estimation means for calculating the occurrence probability of a predetermined event relating to the provided service, and generating second determination information related to the result of the calculation, the determination information and the second determination information In charge of service Judgment means for judging whether or not learning of the first estimation means is performed using the information and the occurrence history of the predetermined event, and the first estimation means is provided according to the result of the judgment The learning is performed using the service and the occurrence history of the predetermined event.

  The present invention for solving the problems of the conventional example is a service providing apparatus, a providing means for providing a predetermined service to a user, and an occurrence probability of a predetermined event related to the provided predetermined service A first estimating means for calculating determination information related to a result of the calculation, using a result learned based on information related to the service provided in the past and a past occurrence history, A classification that represents the classification result by acquiring information related to the user's environment, classifying the environment represented by the acquired information into one of the predetermined environmental types according to a predetermined rule. Environment determination means for outputting result information, information related to the reaction of the user who has been provided with the service, the acquired information, and a result learned based on a past occurrence history of the predetermined event for Calculating a probability of occurrence of a predetermined event relating to the provided service, and generating second determination information related to a result of the calculation, the determination information, the classification result information, and the second determination information And determining means for determining whether or not to perform learning of the first estimating means using information relating to the provided service and the occurrence history of the predetermined event, the first estimating means Is characterized by performing learning using the provided service and the occurrence history of the predetermined event according to the result of the determination.

  Here, the second estimating means may acquire a plurality of types of information as information relating to a reaction of a user who has received the service.

  Further, the first estimation means performs a process of extracting a plurality of predetermined learning element information from a service provided in the past and learning and acquiring a relationship between the plurality of learning element information and the occurrence history. And the determination means provides a service provided based on information output from each of the first estimation means, the environment determination means, and the second estimation means, and time variations of the plurality of learning element information. It is also possible to determine whether or not learning of the first estimating means is performed using the occurrence history of the predetermined event.

  Further, the second estimating means extracts a plurality of predetermined learning element information from the information relating to the user's reaction acquired in the past, and determines the relationship between the plurality of learning element information and the occurrence history. A process for acquiring learning is performed, and the determination unit is configured to detect time variation of information output from each of the first estimation unit, the environment determination unit, and the second estimation unit, and the plurality of learning element information. On the basis of this, it may be determined whether or not the second estimation means is to be learned, and the second estimation means may perform learning according to the result of the determination by the determination means.

  Further, whether or not the determination means can learn at least one of the first estimation means and the second estimation means based on the time variation of the information related to the user's environment acquired by the environment determination means. It is also possible to judge.

  Further, the provided service may be an information presentation service, and the information related to the service provided in the past may be information presented in the past.

  Also, a service providing method according to an aspect of the present invention uses a service providing apparatus that provides a predetermined service to a user, and a service that has previously provided an occurrence probability of a predetermined event related to the provided service. And calculating the result based on the learning result based on the past occurrence history and outputting the first determination information related to the occurrence probability, and acquiring information on the user's environment, Classifying the environment represented by the acquired information according to a predetermined rule into one of predetermined environmental types, outputting classification result information representing the classification result, and providing the service Obtaining information related to the reaction of the received user, and using the obtained information and the learning result based on the past occurrence history of the predetermined event, the predetermined event related to the provided service. And calculating the occurrence probability of the second and generating second determination information related to the result of the calculation, and based on the first determination information, the classification result information, and the second determination information, It is determined whether or not learning related to the step of generating the first determination information is performed, and learning related to the step of generating the first determination information is performed based on a result of the determination.

  In addition, a program according to another aspect of the present invention provides a computer that provides a predetermined service to a user, the occurrence probability of a predetermined event related to the provided service, and information related to the service provided in the past. The calculation is performed using the learning result based on the past occurrence history, the procedure for outputting the first determination information related to the occurrence probability, and the information related to the user's environment are acquired, and the predetermined rule According to the procedure for classifying the environment represented by the acquired information into one of predetermined environmental types and outputting the classification result information representing the classification result, and the user who received the service The occurrence probability of the predetermined event related to the provided service using the acquired information and the learning result based on the past occurrence history of the predetermined event. A first determination information based on the first determination information, the classification result information, and the second determination information, and executing a procedure for generating second determination information related to a result of the calculation. It is determined whether or not a learning process related to a procedure for generating the first determination information is performed, and a learning process related to a procedure for generating the first determination information is performed based on a result of the determination.

  Embodiments of the present invention will be described with reference to the drawings. As shown in FIG. 1, the service providing apparatus according to the embodiment of the present invention includes a control unit 11, a storage unit 12, an operation unit 13, a display unit 14, a state sensor group 15, and a communication unit 16. ing. Here, a case where a service for presenting information such as e-mail is provided as an example of the provided service will be described. That is, this service providing apparatus operates as an information presentation apparatus.

  Here, the control unit 11 is, for example, a CPU or the like, and operates according to a program stored in the storage unit 12. In the present embodiment, the control unit 11 executes processing as first estimation means, environment determination means, second estimation means, and determination means of the present invention. The specific processing contents of the control unit 11 will be described in detail later.

  The storage unit 12 includes a memory element such as a RAM and a ROM, and / or a disk device. The storage unit 12 stores a program executed by the control unit 11. The storage unit 12 also operates as a work memory for the control unit 11. As will be described later, the storage unit 12 of the present embodiment holds a network acquired by learning.

  The operation unit 13 is a keyboard, a mouse, or the like. The operation unit 13 accepts a user's instruction operation and outputs the content of the instruction operation to the control unit 11. The display unit 14 is a display or the like, and displays information such as an e-mail according to an instruction input from the control unit 11 to present information.

  The state sensor group 15 includes, for example, records (logs) on various actions (actions) on the user's e-mail performed via the operation unit 13, user behavior such as a camera and an infrared sensor, as well as the user's behavior. Includes devices that obtain information about responses to services and sensors that measure. The devices and sensors included in the state sensor group 15 output the acquired or measured information to the control unit 11 as information related to the user's reaction.

  The communication unit 16 is a network card, a wireless LAN, or the like, and sends information via the network in accordance with an instruction input from the control unit 11. The communication unit 16 receives information that arrives via the network and outputs the information to the control unit 11.

  Here, the content of the process performed by the control part 11 is demonstrated. The processing executed by the control unit 11 is functionally configured to include a first estimator 21, an information presentation unit 22, a second estimator 24, and a learning control unit 25, as shown in FIG. Yes. In the following, for ease of explanation, it is assumed that the information to be presented is e-mail information, and the determination information generated by the first estimator 21 and the like is e-mail importance information.

  When receiving the electronic mail, the control unit 11 stores the electronic mail in the storage unit 12. And the 1st estimator 21 produces | generates the determination information showing the importance of the said received email. In the present embodiment, it is assumed that the first estimator 21 determines importance using a Bayesian network. Here, since the method of determining the importance of information using the Bayesian network is well known, its details are omitted, but here as a specific example, a plurality of predetermined nodes ( It is assumed that some of the learning element information) candidates are used as nodes of the Bayesian network.

  That is, the first estimator 21 sends the sender name, the sender address, and the sender classification (for example, referring to address book information of the received e-mail, when the user classifies the sender, the classification Node information), a title or a character string included in the text (such as “X month X day”, whether or not there is an expression indicating a date and time, whether or not there is a predetermined keyword), a transmission date and time, and the like. The first estimator 21 uses a plurality of nodes selected from these node candidates to form a Bayesian network (a network related to the causal relationship of occurrence of each node). In the present embodiment, each parameter related to the Bayesian network, which node is selected from the node candidates, and the network structure between the nodes are targets of learning processing.

  In other words, the first estimator 21 forms a Bayesian network based on the occurrence relationship of each node with respect to the email received in the past and the validity of the past determination result (occurrence history). Then, the first estimator 21 obtains an event occurrence probability that the electronic mail received by the control unit 11 is determined to be an important mail by the user, and when it is estimated that the electronic mail is determined to be an important mail. The determination information that the mail is important is output to the information presentation unit 22. However, this determination information may be a display of the occurrence probabilities themselves relating to the importance levels of all received e-mails, or all the received e-mails may be arranged in descending order of occurrence probabilities relating to importance levels.

  The information presentation unit 22 outputs the received e-mail and the determination information input from the first estimator 21 to the display unit 14 and presents it to the user. When the determination information is not input from the first estimator 21, the received e-mail may be presented as it is. When the determination information that the mail is important is not input, the e-mail is presented. You may complete | finish a process, without it.

  The second estimator 24 performs an explicit feedback information input operation regarding the importance of the email input by the user input from the operation unit 13 or the user input to the email input from the state sensor group 15. Based on the information related to the reaction or the like, it is estimated whether or not the user has determined that the presented information is important, and second determination information representing the result of the estimation is output.

  In the present embodiment, the second estimator 24 also uses the Bayesian network to determine the importance. Further, in the present embodiment, it is assumed that the second estimator 24 constitutes a part of the same Bayesian network as the first estimator 21. The second estimator 24 also uses a part of a plurality of predetermined node candidates as nodes of the Bayesian network.

  The second estimator 24 sends the sender name, sender address, sender category of the received e-mail (for example, if the user classifies the sender with reference to the address book information, information on the category) ), Character strings included in the title and body text (such as “X month X day”, whether or not there is an expression representing the date and time, whether or not there is a predetermined keyword, etc.), information on the date and time of transmission, etc. Corresponding node candidate, user instruction input from operation unit 13, viewing time of received e-mail, document amount of e-mail, deleted, printed, attached document opened A plurality of nodes are selected from the node candidates corresponding to the user's reaction input from the state sensor group 15 regarding various information such as whether a link related to the WWW described in the e-mail is opened, Forming a Ann network (network according to causality of the occurrence of each node). As the state sensor group 15, it is effective to use information on a plurality of types of reactions so that more effective factors can be reflected in the learning of diverse users. The user's instruction operation input from the unit 13, the browsing time of the received e-mail, the document amount of the e-mail, whether it was deleted, printed, attached document opened, or written in the e-mail Inferring the importance of e-mail by incorporating factors that are effective for the majority of users when estimating the importance of e-mail based on information about user responses such as whether a link to a web page has been opened be able to. In the present embodiment, each parameter related to the Bayesian network, which node is selected from the node candidates, and the network structure between the nodes are targets of learning processing.

  Further, whether or not the e-mail is important for the node for the information extracted from the e-mail and the node related to the state sensor group 15 relating to the user's instruction operation and various user reactions input from the operation unit 13 This is particularly effective for improving the accuracy of the second estimator 24 because it expresses various conditional dependencies.

  The second estimator 24 forms a Bayesian network on the basis of the occurrence relationship of each node with respect to the email received in the past and the past determination result (occurrence history). Then, the second estimator 24 obtains the probability of occurrence of an event that the user has determined that the electronic mail received by the control unit 11 is an important mail, and outputs the determination result related to the event as second determination information. .

  In the present embodiment, the first estimator 21 and the second estimator 24 operate using one Bayesian network. In this case, when the first estimator 21 performs estimation, Information indicating that the status of the node called “evidence” has been confirmed is stored in the Bayesian network in the node candidate corresponding to the information extracted from the presented e-mail, and the user corresponding to the user's reaction is stored in the user Since the response is still unknown at this time, the first estimation is performed without storing the information that the state of the node called “evidence” is confirmed in the Bayesian network. In the case of performing the second estimation, it is possible to store information that the state of the node called “evidence” has been confirmed in the node corresponding to the information extracted from the e-mail, and not to store it intentionally. It is also possible. On the other hand, information about the user's reaction is estimated by storing it in the Bayesian network as information that the state of the node called “evidence” has been confirmed.

  However, even when information that the state of the node called “evidence” has been confirmed as described above is required, if this is not obtained in all nodes, or “evidence” called “hidden node” is not obtained at all. Nodes may exist, in which case they remain indeterminate, probabilities learned from past history, or subjective probabilities set by the designer, and other "evidence" It is also possible to perform estimation using

  Note that the control unit 11 responds to information that is the basis of learning in the first estimator 21 and the second estimator 24, that is, information on e-mails, information on user instruction operations, and user responses. Related information or the like may be stored in the storage unit 12 as past learning data in association with the received e-mail.

  The learning control unit 25 uses the information output from each of the second estimators 24 as a teacher signal, and performs learning for improving the performance of the first estimator 21 without a teacher signal from an explicit user. The relationship between the nodes involved in the various responses of the user in the second estimator 24 and the nodes that determine whether or not the e-mail is important is that the designer does not learn by giving an appropriate subjective probability. Selection is also possible. Alternatively, starting from the above-described subjective probability or uniform distribution and other probability values excluding subjectivity as much as possible, we obtain explicit feedback on the importance of the received email from the user. It is possible to simultaneously learn both the first estimator and the second estimator, which is particularly desirable in the initial stage of learning.

  In the case of the present embodiment, the teacher signal output from the second estimator 24 is uniformly important as long as it is a numerical value higher than a certain threshold (for example, an important occurrence probability of 0.5 or more) Using the concept of probability propagation in a Bayesian network, it is also possible to propagate the occurrence probability that is important and to learn using the expected value and the certainty of each state regarding whether or not the email is important. In such a case, for example, a steepest descent method or an EM algorithm may be used.

  The learning control unit 25 learns the first estimator 21 and the second estimator 24 using the information presented this time based on information output from the first estimator 21 and the second estimator 24. It can also be determined whether or not to do so.

  In the present embodiment, the first estimator 21 determines that “A: the email to be presented is important” or “B: the email to be presented is not important”. Further, as the determination of the second estimator 24, determination is made that “A: the presented email is treated as important” or “B: the presented email is not treated as important”.

  As described above, according to the present invention, information such as “A: the electronic mail to be presented is important” can be presented as the information presented to the user as the determination of the first estimator 21, and further the second estimation. The device 24 can make an accurate determination that “B: the presented e-mail was not treated as important” by making an integrated determination from the reactions of various users. As a result, learning for reducing the burden on the user and reducing the burden on the user and improving the determination accuracy of the first estimator 21 as much as possible by explicitly forcing the provision of information on the importance of the e-mail from the user is performed. Can be done.

  In this case, the first estimator 21 and the second estimator 24 and the like use the Bayesian network to generate e-mail importance level information, but the first estimator and the second estimator. Can be composed of separate Bayesian networks, but also a naive Bayes classifier, which is a special case of Bayesian networks, or support vector machines, co-occurrence patterns, decision trees, etc. An inference, classification process, or a combination thereof may be used.

  Although the presentation of information has been described so far, the service provided to the user is not limited to such information provision, but the brightness of the lighting, the height of the chair, the temperature of the air conditioner, the blowing power, the direction, etc. It can also be used to control other devices. For lighting, there is voltage control, and for the height of the chair, there is a vertical driving of the seating surface by a stepping motor. The present invention can also be applied to these, and by learning the first estimation means using the determination information, the classification result information, and the second determination information when the control is performed. Learning accuracy can be improved.

  In addition, it is not necessary for the service providing device and the user to directly act on the user such as in the same room. For example, it is applied to a device for remotely monitoring a device that is automatically controlled in a separate room with a television monitor. May be. In this case, the monitor's reaction and the monitor's reaction are acquired as information.

  Further, in another aspect of the embodiment of the present invention, the processing of the control unit 11 is slightly different. That is, the processing executed by the control unit 11 in this other aspect is functionally the first estimator 21, the information presenting unit 22, the second estimator 24, and the learning control unit 25 as shown in FIG. It is comprised including. In the following, for ease of explanation, it is assumed that the information to be presented is e-mail information, and the determination information generated by the first estimator 21 and the like is e-mail importance information.

  Note that the state sensor group 15 in this case includes sensors that measure user behavior, body temperature, heart rate, and other environmental information, such as a camera and an infrared sensor. The sensors included in the state sensor group 15 output the measured information to the control unit 11 as environment information. The state sensor group 15 is not always necessary.

  When receiving the electronic mail, the control unit 11 stores the electronic mail in the storage unit 12. And the 1st estimator 21 produces | generates the determination information showing the importance of the said received email. In the present embodiment, it is assumed that the first estimator 21 determines importance using a Bayesian network. Here, since the method of determining the importance of information using the Bayesian network is well known, its details are omitted, but here as a specific example, a plurality of predetermined nodes ( It is assumed that some of the learning element information) candidates are used as nodes of the Bayesian network.

  That is, the first estimator 21 sends the sender name, the sender address, and the sender classification (for example, referring to address book information of the received e-mail, when the user classifies the sender, the classification Node information), a title or a character string included in the text (such as “X month X day”, whether or not there is an expression indicating a date and time, whether or not there is a predetermined keyword), a transmission date and time, and the like. The first estimator 21 uses a plurality of nodes selected from these node candidates to form a Bayesian network (a network related to the causal relationship of occurrence of each node). In the present embodiment, each parameter related to the Bayesian network, which node is selected from the node candidates, and the network structure between the nodes are targets of learning processing.

  In other words, the first estimator 21 forms a Bayesian network based on the occurrence relationship of each node with respect to the email received in the past and the validity of the past determination result (occurrence history). Then, the first estimator 21 obtains an event occurrence probability that the electronic mail received by the control unit 11 is determined to be an important mail by the user, and when it is estimated that the electronic mail is determined to be an important mail. The determination information that the mail is important is output to the information presentation unit 22.

  The information presentation unit 22 outputs the received e-mail and the determination information input from the first estimator 21 to the display unit 14 and presents it to the user. When the determination information is not input from the first estimator 21, the received e-mail may be presented as it is. When the determination information that the mail is important is not input, the e-mail is presented. You may complete | finish a process, without it.

  The environment determination unit 23 determines the user's environment as one of predetermined environmental types based on the user's instruction operation input from the operation unit 13, the environment information input from the state sensor group 15, or the like. And classification result information representing the classification result is generated. As a specific example, here, the environmental type may be classified into a type in which the user is “in a normal state” and a type in which the user is “in a normal state”. For example, the environment determination unit 23 determines whether a user is seated and performing an operation based on information input from the state sensor group 15 when an instruction operation related to an e-mail is input from the operation unit 13. Check out. If the user is seated and not performing an operation, it is determined that the user is not in a normal state, and if the user is seated and performing an operation, it is determined that the user is in a normal state. Whether or not the user is seated may be determined by a sensor attached to the chair, or may be determined by performing image processing for estimating the posture of the user from the video of the user captured by the camera.

  Note that a Bayesian network may also be used for this determination by the environment determination unit 23. That is, the operation speed related to the operation input, the number of operation errors (number of times the delete key was pressed, etc.), whether or not seated, whether or not utterances were made when opening an e-mail, and whether or not eyes were moved (captured with a camera) It is also possible to determine whether or not the user is “in a normal state” by using each occurrence relationship such as can be realized by image processing for recognizing an eye part from an image.

  Furthermore, the environment determination unit 23 may use the information stored in the storage unit 12 to determine whether or not the user is “normal”. For example, when the user's schedule information is stored in the storage unit 12, the user is “before going out” by referring to the schedule information and information on a clock (not shown) (clock that measures the current date and time). The occurrence relationship such as whether or not, “I just returned from the destination” may be determined.

  Similar to the environment determination unit 23, the second estimator 24 uses the presented information based on the user's instruction operation input from the operation unit 13, the environment information input from the state sensor group 15, and the like. It is estimated whether the person determined that the person is important, and second determination information representing the result of the estimation is output.

  In the present embodiment, the second estimator 24 also uses the Bayesian network to determine the importance. The second estimator 24 also uses a part of a plurality of predetermined node candidates as nodes of the Bayesian network.

  The second estimator 24 sends the sender name, sender address, sender category of the received e-mail (for example, if the user classifies the sender with reference to the address book information, information on the category) ), Character strings included in the title and body text (such as “X month X day”, whether or not there is an expression representing the date and time, whether or not there is a predetermined keyword, etc.), information on the date and time of transmission, etc. A plurality of nodes are selected from a corresponding node candidate and a node candidate corresponding to a user's reaction, such as a user instruction operation input from the operation unit 13 or environment information input from the state sensor group 15. To form a Bayesian network (a network related to the causal relationship of occurrence of each node). In the present embodiment, each parameter related to the Bayesian network, which node is selected from the node candidates, and the network structure between the nodes are targets of learning processing.

  In other words, the second estimator 24 forms a Bayesian network based on the occurrence relationship of each of the above-described e-mails and past determination results (occurrence history). Then, the second estimator 24 obtains the probability of occurrence of an event that the user has determined that the electronic mail received by the control unit 11 is an important mail, and outputs the determination result related to the event as second determination information. .

  In addition, the control part 11 is the information used as the basis of learning in these 1st estimators 21, the 2nd estimator 24, and the environment judgment part 23, ie, the information of an email, the information of a user's instruction operation, Environmental information or the like may be stored in the storage unit 12 as past learning data in association with the received e-mail.

  The learning control unit 25 uses the information presented this time based on the information output from each of the first estimator 21, the environment determination unit 23, and the second estimator 24, and uses the information presented this time. It is determined whether or not the estimator 24 learns.

  In the present embodiment, the first estimator 21 determines that “A: the email to be presented is important” or “B: the email to be presented is not important”. Further, as the determination of the second estimator 24, determination is made that “A: the presented email is treated as important” or “B: the presented email is not treated as important”.

  First, the learning control unit 25 checks whether or not the determination result of the first estimator 21 matches the determination result of the second estimator 24. If they do not match, the parameters of the Bayesian network of the second estimator 24 are changed with reference to the contents of the user's instruction operation, environment information, and the like. For example, when it is determined from the environmental information that the user is “before going out”, among the nodes of the Bayesian network of the second estimator 24 that is related in advance to the occurrence of the determination, “the number of operation errors” The occurrence probability of an operation error is increased by controlling the occurrence probability. Alternatively, a Bayesian network composed of other nodes excluding the node with the number of operation mistakes is reconfigured, and the second determination information is regenerated using the reconfigured Bayesian network.

  In this way, the learning control unit 25 adjusts the second estimator 24 to check whether or not it matches the determination result of the first estimator 21. If they match, the second determination information after regeneration (after adjustment) is fed back to the first estimator 21 to cause the first estimator 21 to learn.

  Even when the user is busy, such as before he / she goes out, even if he / she accidentally behaves as if it is not important, the learning control unit 25 Since the adjustment of the second estimator 24 for determining the behavior is performed and the determination result is controlled, the chances that the temporary circumstances based on the accidental factor are reflected in the learning are reduced. Accuracy can be improved.

  Further, when the regenerated second determination information is still different from the determination information output from the first estimator 21, the learning control unit 25 determines the first estimator 21, the second estimator 24, and the environment determination unit 23. The information used for the e-mail presented this time is compared with each past learning result.

  For example, the first estimator 21 compares the occurrence probability distribution of each node as a past learning result with the occurrence relationship corresponding to each node related to the electronic mail presented this time. If the product of the probabilities corresponding to each occurrence relationship is greater than or equal to a predetermined threshold value, it is determined that the information related to the current e-mail is within the past learning range. For example, it is assumed that the probability that the sender of the e-mail is A is pA, and the probability that the e-mail transmission time is from 8 am to 10 am is p 1. In this case, if the e-mail received this time is a sender A and sent at 6 am, pA × (1−pB) is calculated and does this exceed a predetermined threshold value? Check for no. According to this, when there is a high probability that the sender of the e-mail is not A from past cases, pA becomes small. That is, an event that an e-mail is received from the sender A has a high probability of being outside the range of past learning cases. In this way, it is determined whether or not the electronic mail presented this time is within the range of past learning. The same determination is performed for the second estimator 24 and the environment determination unit 23.

  The learning control unit 25 inputs the result of comparing the information used by the first estimator 21, the second estimator 24, and the environment determination unit 23 with respect to the past learning result with respect to the e-mail presented this time. receive. This result is either (I) that is within the learning range or (O) that is outside the learning range for each, and is one of eight combinations as shown in FIG.

  The learning control unit 25 executes a process determined for each combination depending on which combination is selected. As a specific example, first, when everything is within the range of learning as in the first case, the first estimator 21 is made to learn using the second determination information output from the second estimator 24. . In the following description, the combination of the first estimator 21, the second estimator 24, and the environment determination unit 23 is arranged in this order and expressed as (III) or the like.

  As in the second case, when (IIO), it can be determined that an unexpected situation has occurred in the environmental information or the like. In order to make a determination according to such a case, the learning control unit 25 responds to the user's reaction such as the user's instruction operation input from the operation unit 13 or the environment information input from the state sensor group 15. Based on such conditions, it is determined whether or not learning is performed according to a predetermined rule. For example, when the user's instruction operation is “many errors” and the environmental information is “the user is not busy”, a rule such as learning is determined in advance and stored in the storage unit 12. What is necessary is just to determine whether learning is performed according to the said rule.

  Further, when the situation is not defined in this rule, it is a node having a high occurrence probability from a node included in the Bayesian network of the environment determination unit 23, and a node that is occurring, or has a low occurrence probability, and A node that has not occurred (hereinafter referred to as an occurrence matching node) is searched. This node has a discrepancy with the past learning data. The learning control unit 25 checks whether or not the node is a node related to information extracted from the e-mail. If it is such a node, the node is added to at least one of the Bayesian networks of the first estimator 21 and the second estimator 24. If it is not, the node is added to the Bayesian network of the second estimator 24. In this case, if past learning data is stored in the storage unit 12, the learning may be performed again based on the past learning data.

  Further, as in the third, fourth, and fifth of FIG. 4, the first estimator 21 and the second estimator 24, although the environment determination unit 23 determines that it is within the learning range. If any one of is determined to be outside the scope of learning, that is, if any of (OII), (IOI), or (OOI), there is an unforeseen situation in the environmental information etc. It can be determined that it has not occurred. In this case, the learning of the first estimator 21 or the second estimator 24 is suspicious. An occurrence matching node is searched from node candidates that are not selected (not used) for the Bayesian network of the first estimator 21 and the second estimator 24. This node has a discrepancy with the past learning data. If there is an occurrence matching node, the learning control unit 25 adds it to at least one of the Bayesian networks of the first estimator 21 and the second estimator 24. In this case, if past learning data is stored in the storage unit 12, the learning may be performed again based on the past learning data.

  When learning is performed again in this way, the Bayesian network before learning again is saved and stored in the storage unit 12, and the first estimator 21 and the second estimator 24 after learning again are used. On the other hand, determination information and second determination information related to the e-mail received this time are generated, and it is checked whether or not they match. When the accuracy deteriorates, such as when there is a mismatch, or when the difference in event occurrence probability that the user deems important is greater than before learning again, the saved Bayesian network is read out. Return to the original. In the case where the determination information and the second determination information match, the Bayesian network that has been learned again is used thereafter.

  Further, when the evacuated Bayesian network is read out and restored, it is a node included in the Bayesian network of the first estimator 21 and the second estimator 24 even though the occurrence probability is low. A node that has occurred or a node that has not occurred despite its high occurrence probability (hereinafter referred to as an occurrence mismatched node) is searched. This node is likely to be a defective node. Therefore, a Bayesian network excluding the node is generated, or the actual occurrence relationship of the node is ignored, and the occurrence relationship of the node is estimated from other nodes (or nodes). Instead, the determination information or the second determination information is generated according to the estimated occurrence relationship. In other words, the node is deleted and the network structure is changed. Also in this case, if the past learning data is stored in the storage unit 12, the learning may be performed again based on the past learning data.

  When learning is performed again in this way, the Bayesian network before the node deletion or the network structure is changed is also saved in the storage unit 12 and retained, and the first estimator 21 after the learning is performed again. And the second estimator 24 generate determination information and second determination information related to the currently received e-mail, and check whether they match. When the accuracy deteriorates, such as when there is a mismatch, or when the difference in event occurrence probability that the user deems important is greater than before learning again, the saved Bayesian network is read out. Return to the original. In the case where the determination information and the second determination information match, the Bayesian network that has been learned again is used thereafter.

  If the determination information and the second determination information do not match even if the occurrence matching node is added or the occurrence mismatch node is deleted as described above, the environment determination unit 23 does not match the Bayesian network. A part is selected from the included nodes, and the selected node is added to at least one of the Bayesian networks of the first estimator 21 and the second estimator 24 to replace the defective node. Let The node to be selected here may be registered in advance. Further, a network in which all the nodes included in the Bayesian networks of the first and second estimators 21 and 24 and the environment determination unit 23 are extracted is generated, and the past learning data stored in the storage unit 12 is generated. A node may be selected by performing learning based on the result and comparing the learning result with the networks of the first and second estimators 21 and 24. The node to be specifically selected may be a node or a node group surrounding a defective node (or node group). Then, a Bayesian network in which the defective node is replaced by the selected node is generated.

  Note that if a defective node cannot be identified, a new node may be obtained from the user, or a new node may be added based on information that can be acquired via the network. Even when a new node is added here, if the determination information and the second determination information do not match, the same processing as when the environment determination unit 23 determines that it is out of the learning range (that is, (Process when any of (IOO), (OIO), and (OOO)) is performed. These processes will be described later.

  Further, if the determination information and the second determination information do not coincide with each other even if the processing is performed, the first estimator 21 is made to learn based on the estimation result of the second estimator 24 on a trial basis. . Then, how much the accuracy of the estimation related to the email received in the past is affected by the learning is calculated. That is, based on the past e-mail and past learning data stored in the storage unit 12, the determination information and the second determination output from each of the first estimator 21 and the second estimator 24 before and after trial learning. Compare information. Then, the probability that the determination information does not match is checked, and the probability of the mismatch of the determination information of the first estimator 21 after the trial learning is higher than the probability of the mismatch of the determination information of the first estimator 21 before the trial learning. When it is determined that there is a negative influence on the estimation accuracy, such as an increase, the learning control unit 25 controls the first estimator 21 not to perform learning. In this case, the learning control unit 25 counts the number of times such control is performed, and when the count value becomes a certain value or more within a predetermined period, the learning unit 25 presents information to notify that effect on the display unit 14.

  Furthermore, as in the sixth and seventh in FIG. 4, when the environment determination unit 23 determines that it is out of the learning range (when it becomes (IOO), (OIO)), the environment is unexpected. There is a high possibility that one of the first estimator 21 and the second estimator 24 has determined that it is out of the learning range. That is, in this case, it is considered that one of the first estimator 21 and the second estimator 24 cannot cope with an unexpected situation.

  In this case, the occurrence matching node is searched from the nodes included in the Bayesian network of the environment determination unit 23. Then, the learning control unit 25 checks whether or not the node is a node related to information extracted from the e-mail. If it is such a node, the node is added to at least one of the Bayesian networks of the first estimator 21 and the second estimator 24. If it is not, the node is added to the Bayesian network of the second estimator 24. In this case, if past learning data is stored in the storage unit 12, the learning may be performed again based on the past learning data.

  Further, a network in which all the nodes included in the Bayesian networks of the first and second estimators 21 and 24 and the environment determination unit 23 are extracted is generated, and the past learning data stored in the storage unit 12 is generated. The node to be added may be selected by performing learning based on the result and comparing the learning result with the networks of the first and second estimators 21 and 24.

  Further, as in the eighth example of FIG. 4, when it is (OOO), this corresponds to the system itself encountering an unexpected situation. In this case, the output of the first estimator 21 may be stopped.

  As described above, the control unit 11 uses the first estimator 21 that presents information to the user and the second estimator 24 that evaluates the validity of the information presented based on the actual reaction of the user. In addition to comparing the results of these judgments, if each judgment result is inconsistent, the learning mode (nodes used for learning, network structure, etc.) changes depending on whether the cause is an environmental cause, etc. Let Or suppress learning itself. As a result, it is possible to prevent an unexpected situation based on an accidental occurrence from being learned as it is.

  Here, the first estimator 21, the second estimator 24, and the like use the Bayesian network to generate e-mail importance information and the like, but not only the Bayesian network, for example, Bayesian A naive Bayes classifier, which is a special case of a network, or inference such as support vector machine, co-occurrence pattern, decision tree, classification processing, or a combination thereof may be used.

  In addition, the learning control unit 25 performs temporal variations of the determination information and the second determination information output by the first estimator 21, the second estimator 24, and the environment determination unit 23, and the temporal changes of the nodes used by each. Variations are analyzed by a well-known time series analysis process. Then, based on the temporal variation, it is determined whether or not the first estimator 21 and the second estimator 24 are to be learned, and when it is determined that the learning is to be performed, the first estimator 21 and the second estimator 24 learnings may be performed.

  Here, the temporal fluctuation includes information such as the magnitude of the fluctuation, an increasing / decreasing tendency of the occurrence probability, and a fluctuation cycle. As a specific example, when the reception frequency of the email from the sender A tends to increase (the occurrence probability that the sender of the email is A) tends to increase, the learning control unit 25 stores the storage unit 12 is extracted only for a recent predetermined period from the past learning data stored in 12, and the network of the first estimator 21 and the second estimator 24 is learned again based on the extracted recent past learning data. . As a result, a network more suitable for recent trends is generated.

  In addition, when the fluctuation cycle is detected, the past learning data for a period exceeding the fluctuation cycle is extracted, and the network of the first estimator 21 and the second estimator 24 is again connected with the extracted past learning data. You may learn. In this way, learning accuracy can be improved.

  Although the presentation of information has been described so far, the service provided to the user is not limited to such information provision, but the brightness of the lighting, the height of the chair, the temperature of the air conditioner, the blowing power, the direction, etc. It can also be used to control other devices. For lighting, there is voltage control, and for the height of the chair, there is a vertical driving of the seating surface by a stepping motor. The present invention can also be applied to these, and whether or not the first estimation means is learned using the determination information, the classification result information, and the second determination information when the control is performed. It is possible to improve learning accuracy by determining whether or not to perform learning.

  In addition, it is not necessary for the service providing device and the user to directly act on the user such as in the same room. For example, it is applied to a device for remotely monitoring a device that is automatically controlled in a separate room with a television monitor. May be. In this case, the monitor's reaction and the monitor's reaction are acquired as information.

  Furthermore, also in this example, the state sensor group 15 includes user behaviors such as records (logs) regarding various actions (actions) on the user's e-mail performed via the operation unit 13, etc. It may include a device that obtains information about the response to the service, and a sensor that measures it. The devices and sensors included in this state sensor group 15 output the acquired or measured information to the control unit 11 as environmental information including information related to the user's reaction.

  In addition, as the state sensor group 15, it is effective to use information on a plurality of types of reactions so that more effective factors can be reflected in learning in diverse user reactions. The user's instruction operation input from the operation unit 13, the browsing time of the received e-mail, the document amount of the e-mail, whether it has been deleted, printed, attached document opened, or written in the e-mail Information about the user's response, such as whether a link to the web page was opened.

  And the control part 11 can take in a factor effective for many users when judging the importance of an e-mail by estimating using several types of information regarding these user's reactions. For example, whether a node for information extracted from an e-mail and a node related to a state sensor group 15 related to a user's instruction operation and various user responses input from the operation unit 13 are important. In some cases, it is connected without involving a node relating to the determination, and this expresses various conditional dependencies, which is particularly effective for improving the accuracy of the second estimator 24.

It is a block diagram showing an example of a service providing apparatus according to an embodiment of the present invention. It is a functional block diagram showing an example of the service provision apparatus which concerns on an aspect with embodiment of this invention. It is a functional block diagram showing an example of the service provision apparatus which concerns on another side surface of embodiment of this invention. It is explanatory drawing showing the example of information which concerns on the processing content of a learning control part.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 11 Control part, 12 Storage part, 13 Operation part, 14 Display part, 15 State sensor group, 16 Communication part, 21 1st estimator, 22 Information presentation part, 23 Environment judgment part, 24 2nd estimator, 25 Learning control Department.

Claims (8)

Providing means for providing a predetermined service to the user;
The probability of occurrence of multiple events predetermined in relation to the service, and information relating to services provided in the past, calculated using the historical occurrence history, the result of learning based on, have been prespecified First estimation means for presenting to the user determination information indicating whether or not a specific event occurs ;
A plurality of types of information related to the reaction of the user who received the service is acquired, and the occurrence probability of a plurality of events predetermined in relation to the service is determined, and the occurrence relationship of each event in the service provided in the past And second learning information indicating whether or not the specific event specified in advance occurs by calculating using the learning result based on the occurrence history and the event related to the acquired information. Second estimating means for
Including
The first estimation means performs learning using the provided service and the occurrence histories of the plurality of events predetermined in relation to the service based on the second determination information. Providing device. Providing means for providing a predetermined service to the user;
The probability of occurrence of multiple events predetermined in relation to the service, and information relating to services provided in the past, calculated using the historical occurrence history, the result of learning based on, have been prespecified First estimation means for presenting to the user determination information indicating whether or not a specific event occurs ;
A plurality of types of information related to the reaction of the user who received the service is acquired, and the occurrence probability of a plurality of events predetermined in relation to the service is determined, and the occurrence relationship of each event in the service provided in the past And second learning information indicating whether or not the previously specified specific event occurs is calculated using the learning result based on the past occurrence history and the event related to the acquired information. A second estimating means;
Based on the determination information and the second determination information, learning of the first estimation unit is performed using information related to the provided service and occurrence history of a plurality of events predetermined in relation to the service. Judging means for judging whether or not
Including
The first estimation means performs learning using the provided service and a plurality of event occurrence histories predetermined in relation to the service based on the result of the determination. Providing means for providing a predetermined service to the user;
The probability of occurrence of multiple events predetermined in relation to the service, and information relating to services provided in the past, calculated using the historical occurrence history, the result of learning based on, have been prespecified First estimation means for presenting to the user determination information indicating whether or not a specific event occurs ;
A classification result that obtains information related to the user's environment, classifies the environment represented by the acquired information according to a predetermined rule into one of the predetermined environmental types, and represents the classification result Environmental judgment means for outputting information;
A plurality of types of information related to the reaction of the user who received the service is acquired, and the occurrence probability of a plurality of events predetermined in relation to the service is determined, and the occurrence relationship of each event in the service provided in the past And second learning information indicating whether or not the specific event specified in advance occurs by calculating using the learning result based on the occurrence history and the event related to the acquired information. Second estimating means for
Based on the determination information, the classification result information, and the second determination information, the information on the provided service and the occurrence history of a plurality of events predetermined in relation to the service are used. A determination means for determining whether or not to perform learning;
Including
The first estimation means performs learning using the provided service and a plurality of event occurrence histories predetermined in relation to the service based on the result of the determination. The service providing apparatus according to claim 3,
The first estimation means performs a process of extracting a plurality of predetermined learning element information from a service provided in the past, and learning and acquiring a relationship between the plurality of learning element information and the occurrence history. ,
Said determining means, said first estimation means, and the environment judgment unit, information that each output of the second estimation means, and on the basis of the time variation of the plurality of learning element information, services and the services provided A service providing apparatus for determining whether or not to perform learning of the first estimating means using occurrence history of a plurality of events determined in advance in relation to . The service providing device according to any one of claims 3 to 4 ,
The second estimating means extracts a plurality of pieces of predetermined learning element information from information relating to user responses acquired in the past, and learns the relationship between the plurality of learning element information and the occurrence history. Process
The determination means is based on information output from each of the first estimation means, the environment determination means, and the second estimation means, and time variation of the plurality of learning element information. Decide whether or not to learn,
The service providing apparatus, wherein the second estimation unit performs learning according to a determination result of the determination unit. The service providing apparatus according to any one of claims 3 to 5 ,
The determination means determines whether or not to learn at least one of the first estimation means and the second estimation means based on a time variation of information related to the user's environment acquired by the environment determination means. A service providing apparatus characterized by that. The service providing apparatus according to any one of claims 3 to 6 ,
The provided service is an information presentation service, and the information related to a service provided in the past is information presented in the past. A computer that provides predetermined services to users ,
The probability of occurrence of multiple events predetermined in relation to the service, and information relating to services provided in the past, calculated using the historical occurrence history, the result of learning based on, have been prespecified First estimation means for presenting to the user determination information indicating whether or not a specific event occurs ;
A plurality of types of information related to the reaction of the user who received the service is acquired, and the occurrence probability of a plurality of events predetermined in relation to the service is determined, and the occurrence relationship of each event in the service provided in the past And second learning information indicating whether or not the specific event specified in advance occurs by calculating using the learning result based on the occurrence history and the event related to the acquired information. Second estimating means for
The first estimating means performs learning using the provided service and the occurrence histories of the plurality of events predetermined in relation to the service based on the second determination information ;
A program characterized by functioning as

Images