JP2022090562A - Information processing device and program - Google Patents

Abstract

To provide an information processing device and a program that can use even social media where relations are sparse, by quantifying the relations between information items such as analysis targets and articles for sale.SOLUTION: An information processing device includes: a first machine learning processing unit that extracts a first analysis target group in which relations set between analysis targets are considered to have a mutually predetermined prescribed first relation, and machine-learns distributed expression information related to each of the analysis targets, based on permutations or combinations thereof; and a second machine learning processing unit that sets, as a second analysis target group, analysis targets in which relations set between the analysis targets are considered to have a mutually predetermined prescribed second relation and analysis targets in which relations set between the analysis targets are considered not to have the mutually predetermined prescribed second relation, and machine-learns distributed expression information of the analysis targets based on the relation set for each of the analysis targets included in the second analysis target group.SELECTED DRAWING: Figure 2

Description

The present invention relates to an information processing apparatus and a program that perform analysis processing related to an analysis target.

In recent years, various social media such as social network services (SNS) have become indispensable to the lives of many people. However, on the other hand, the amount of information transmitted by SNS and the like is enormous, and it is becoming difficult to provide appropriate information to the analysis target.

Patent Document 1 discloses an example in which purchase recommendation information is provided to an analysis target that has not been purchased by another analysis target among the analysis targets belonging to the analysis target group.

JP-A-2019-125359A

According to the above-mentioned conventional example, it is possible to provide information on products that are commonly purchased among users of social media such as a certain SNS. However, as the scale of social media grows, so does the number of users to be analyzed, and the interests of users diversify. The number of information such as is small compared to the number of users.

In other words, as the number of analysis targets increases, information that generally represents the relationship between the analysis targets, the products purchased with the analysis target, and the hashtag (headword) used by the analysis target becomes sparse. The reality is that it has become. It is known that the method as in the above conventional example is not effective for such a sparse relationship.

The present invention has been made in view of the above circumstances, and information that can be used by quantifying the relationship between information items such as analysis targets and products even in social media where the relationship is sparse. One of the purposes is to provide a processing device and a program.

One aspect of the present invention that solves the problems of the above-mentioned conventional example is an information processing apparatus, in which vector values set for each analysis target are set as distributed representation information for a plurality of analysis targets for which mutual relationships are set. A part of the plurality of analysis targets, which is said to have a predetermined first relationship in which the means set as the analysis target and the analysis target have a predetermined first relationship with each other, is extracted as the first analysis target group. , The first machine learning means for machine learning the distributed expression information related to each of the analysis targets included in the first analysis target group based on the sequence or combination of the extracted first analysis target group, and the analysis target. A part of the plurality of analysis targets in which the relationship set in the above is considered to be a predetermined second relationship predetermined to each other, and the relationship set between the analysis targets is a predetermined second relationship predetermined to each other. The second analysis target group is defined as a part of the plurality of analysis targets that are not related to the second analysis target group, and is based on the relationship set for each of the analysis targets included in the second analysis target group. A second machine learning means for machine learning the distributed expression information related to each of the analysis targets included in the analysis target group, and a distributed expression for each analysis target machine-learned by the first machine learning means and the second machine learning means. It includes a processing means for providing information to a predetermined processing.

According to the present invention, even in social media where the relationship is sparse, the relationship between information items such as an analysis target and a product can be quantified and used.

It is a block diagram which shows the structural example of the information processing apparatus which concerns on embodiment of this invention. It is a functional block diagram which shows the example of the information processing apparatus which concerns on embodiment of this invention. It is explanatory drawing which shows the outline example of the graph generated by the information processing apparatus which concerns on embodiment of this invention. It is a flowchart which shows the operation example of the information processing apparatus which concerns on embodiment of this invention. It is explanatory drawing which shows the example of the operation of the information processing apparatus which concerns on embodiment of this invention.

An embodiment of the present invention will be described with reference to the drawings. As illustrated in FIG. 1, the information processing apparatus 1 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, and a communication unit 15. The control unit 11 is a program control device such as a CPU, and operates based on the program stored in the storage unit 12. In the present embodiment, the control unit 11 holds the vector values set for each analysis target in the storage unit 12 as distributed expression information for a plurality of analysis targets for which the relationship with each other is set. Further, the control unit 11 extracts a part of a plurality of analysis targets determined to have a predetermined first relationship as a first analysis target group by using the relationship set between the analysis targets. Then, the control unit 11 executes the first machine learning process for machine learning the distributed expression information based on the order or combination of the extracted first analysis target group.

Further, the control unit 11 extracts an analysis target different from the first analysis target group as the second analysis target group, and is based on the relationship set for each of the analysis targets included in the extracted second analysis target group. Then, the second machine learning process for machine learning the distributed expression information related to each of the analysis targets included in the second analysis target group is executed.

In one example of the present embodiment, the control unit 11 is different from the analysis target A, in which the relationship set with a certain analysis target A is determined to be in a predetermined second relationship predetermined with each other. Other than that, it is determined that the relationship set between the analysis target (referred to as another analysis target. There may be a plurality of them) B and a certain analysis target A does not have a predetermined second relationship predetermined with each other. The second analysis target group including the analysis target B of the above is extracted from a plurality of analysis targets. Then, the control unit 11 machine-learns the distributed expression information related to each of the analysis targets included in the second analysis target group based on the relationship set for each of the analysis targets included in the extracted second analysis target group. do.

Then, the control unit 11 uses the distributed expression information for each analysis target obtained by the first machine learning process and the second machine learning process for a predetermined process, for example, a process such as determination of similarity between the analysis targets. The details of the processing of the control unit 11 will be described later.

The storage unit 12 is a memory device, a disk device, or the like, and holds a program executed by the control unit 11. The program stored in the storage unit 12 may be provided by being stored in a computer-readable and non-temporary recording medium, and may be stored in the storage unit 12. The storage unit 12 of the present embodiment also operates as a work memory of the control unit 11.

The operation unit 13 is a keyboard, a mouse, or the like, accepts an operation of the user of the information processing device 1, and outputs information representing the content of the operation to the control unit 11. The display unit 14 is a display or the like, and displays and outputs information according to an instruction input from the control unit 11.

The communication unit 15 is a network interface or the like. In an example of this embodiment, the communication unit 15 accesses the SNS server S or the like according to an instruction input from the control unit 11, and posts information to be analyzed or an analysis target from these servers S or the like. It receives information and the like and outputs it to the control unit 11. Further, the communication unit 15 may send information to the terminal U on the user side via the network according to the instruction input from the control unit 11 and have the user present the information.

Next, the contents of the operation of the control unit 11 of the present embodiment will be described. The control unit 11 of the present embodiment operates according to the program stored in the storage unit 12, and functionally, as illustrated in FIG. 2, the information acquisition unit 21, the graph generation unit 22, and the first unit. A configuration including a machine learning processing unit 23, a second machine learning processing unit 24, an integrated processing unit 25, an evaluation processing unit 26, and an output processing unit 27 is realized.

In the following explanation, as an example, an account of an SNS user such as Twitter (registered trademark) or Instagram (registered trademark) (referred to as a general account for convenience) and another account that the user follows. (In the following example, in particular, it is limited to predetermined accounts other than general accounts such as official accounts of product brands, hereinafter referred to as official accounts) and each user (general account is an official account). The information that identifies each analysis target is acquired by using each of the tags (hash tags) included in the post (regardless of) as the analysis target.

In this example, the information of the account and the tag included in the post of the account, the information specifying the general account, and the information specifying the official account followed by the general account specified by the information are in the present invention. , Corresponds to the mutual relationship between the analysis targets.

The information acquisition unit 21 includes information that identifies each of a plurality of analysis targets, including an analysis target to be evaluated, and at least one type of information item that has a predetermined relationship with at least one of the analysis targets. To get. In one example of this embodiment, one of the SNSs such as Twitter (registered trademark) and Instagram (registered trademark) is targeted, and the account of the user of the targeted SNS is analyzed.

In this example, the information acquisition unit 21 limits the list of accounts of the users of the SNS (not the entire account, but the accounts to be analyzed, from the website of the SNS or the operator of the SNS according to predetermined conditions. (Maybe), the tag (hash tag) information included in the posts from the accounts belonging to the list, and another account that the account follows for each account belonging to the list (here, it is limited to the official account). It is sufficient to acquire the information that specifies (to be).

The graph generation unit 22 sets nodes representing each of the analysis targets specified by the information acquired by the information acquisition unit 21, and generates a graph in which nodes having a predetermined relationship are connected.

FIG. 3 conceptually shows a graph generated by the graph generation unit 22 when the SNS is targeted. In FIG. 3, a layer B including a node representing each of general accounts, a layer A including a node of an official account, and a layer C including a node representing a hashtag used by each account are represented as separate layers. There is.

The graph generation unit 22 sets an edge between each layer or between nodes having a predetermined relationship with each other in the layer. Specifically, in the example of FIG. 3, the edge is set according to the following rule:
An edge is set between the layer B and the layer A between the node in the layer B and the node in the layer A representing the official account that the general account represented by the node represents.
An edge is set between the layer B and the layer C between the node in the layer B and the node in the layer C representing the hashtag included in the text posted by the general account represented by the node.
-Between the nodes in layer C, an edge is set between the nodes representing each of the plurality of hashtags included (that is, co-occurring) at one time in one post.
I will do it. In the example here, the relationship between the official account and the hashtag used by the official account is not considered.

The graph generation unit 22 sets an edge between the layer B and the layer A, between the layer B and the layer C, between the nodes in the layer B, and between the nodes in the layer C according to the above-mentioned rule. That is, the graph generation unit 22 selects the nodes of the general account belonging to the layer B one by one, and sets the edge from the selected node to the node of the official account belonging to the layer A, which the general account represented by the node follows. do.

Further, the graph generation unit 22 sets an edge from the node selected from the layer B to the node of the hash tag included in the post of the general account represented by the selected node.

The graph generation unit 22 sets this edge for each node of the general account belonging to the layer B. Further, the graph generation unit 22 may further concatenate a pair of nodes representing a pair of hashtags having a predetermined relationship and a pair of nodes representing a pair of official accounts having a predetermined relationship.

Specifically, this graph generation unit 22 is a pair of hashtags that are different from each other and are represented by a pair of nodes belonging to layer C, and a pair that appears in common (that is, co-occurs) in one post. Set an edge between a pair of nodes that represent the hashtag of. For example, when a post contains hashtags α, β, and γ, the graph generation unit 22 sets edges between hashtags α-β, β-γ, and α-γ.

Further, the graph generation unit 22 may set weights on the edges between the nodes in the layer C according to the number of times co-occurrence occurs. Further, the graph generation unit 22 describes the edge between the node of the general account (node belonging to layer B) and the node of the hashtag included in the post from the user of the general account (node belonging to layer C). However, the weight may be set according to the number of times the hashtag is used by the general account.

Here, each edge set by the graph generation unit 22 may be an edge having no direction. That is, the graph generated by the graph generation unit 22 may be an undirected graph.

The first machine learning processing unit 23 is determined to have a predetermined first relationship using the connection relationship of the nodes included in the graph generated by the graph generation unit 22, that is, the relationship set between the analysis targets. A part of a plurality of analysis targets is extracted as a first analysis target group.

As an example, the first machine learning processing unit 23 starts from an arbitrary node of the graph generated by the graph generation unit 22, and can be reached by tracing an edge a predetermined number of times from the starting point (first analysis target group). ) List (sequential column or combination) is generated. Hereinafter, a node that satisfies the condition (predetermined condition) that the edge can be reached by tracing the edge within an arbitrarily determined number of times from the starting point is referred to as a connecting node for the node that is the starting point, and these starting points are referred to. It is said that the node and the connected node are connected to each other.

Then, the first machine learning processing unit 23 machine-learns the distributed expression information based on, for example, a sequence of the extracted first analysis target group. Specifically, the first machine learning processing unit 23 performs skipgram learning by negative sampling for each node included in each generated list, and obtains vector information which is distributed representation information of each node. This method is disclosed, for example, in T. Mikolov, et al., “Distributed Representations of Words and Phrases and their Compositionality”, arXiv, arXiv: 1310.4546, so detailed description of this method is omitted here. do.

That is, the first machine learning processing unit 23 obtains vector information for each node as distributed representation information of each node of the generated graph by a so-called graph embedding method.

In the second machine learning processing unit 24, the relationship set between the analysis targets is a part of a plurality of analysis targets that are considered to have a predetermined relationship predetermined with each other, and the relationship set between the analysis targets is set. A part of the plurality of analysis targets that are not considered to have a predetermined relationship with each other as a second analysis target group, based on the relationship set for each of the analysis targets included in the second analysis target group. , Machine learning of distributed expression information related to each of the analysis targets included in the second analysis target group.

In an example of the present embodiment, when it is determined that the second machine learning processing unit 24 has a relationship of being connected to each other under a predetermined condition using graph information, it is assumed that the second machine learning processing unit 24 has the above-mentioned predetermined relationship. The second analysis target group is extracted.

Specifically, the second machine learning processing unit 24 that performs the processing of this example performs the next processing while sequentially selecting each node of the general account belonging to the layer B as the node of interest Bn. That is, the second machine learning processing unit 24 selects at least one node Cx whose edge is set between the selected node Bn and the node of interest among the nodes belonging to the layer C. Further, the second machine learning processing unit 24 selects at least one node Cy whose edge is not set between the selected node Bn and the node belonging to the layer C among the nodes belonging to the layer C. The nodes Cx and Cy selected here are the second analysis target group in the present invention.

As for the first analysis target group, the node satisfying the condition that the node can be reached by tracing the edge within a predetermined number of times from the node as the starting point is extracted here, whereas the second analysis target is extracted. Regarding the group, a node having an edge set between the node and the node as the starting point and a node having no edge set are selected, but this is an example, and is another example of the present embodiment. Then, also in the extraction of the second analysis target group, a node that satisfies the condition that the node can be reached by tracing the edge within a predetermined number of times from the node as the starting point and a node that does not satisfy the condition are selected. You may do it.

In the second machine learning processing unit 24, the distance between the distributed representation information of the node Bn of interest and the distributed representation information of the node Cx (if there are a plurality of them) is different between the distributed representation information of the node Bn of interest and the node Cy (plural). In that case, at least the distributed representation information of the node Bn of interest is determined so that the distance is closer than the distance to the distributed representation information (of each of them). The initial value of the distributed representation information of each node here may be set at random, or may be set to the vector value obtained by the first machine learning processing unit 23.

として求める。ここでΘを含む項は正則化項（regularization term）であり、

である。またここでａはゲインであり、適宜実験的に、例えばａ＝１などと定める。 After initialization, the distributed representation information of the node Bn of interest selected i-th is xi (i = 1, 2, ...), The distributed representation information of the node Cx is xj (j = 1, 2, ...), And the distribution of the node Cy. Assuming that the expression information is xj'(j'= 1, 2, ...), The second machine learning processing unit 24 sets the loss LPARIG.

Ask as. Here, the term containing Θ is a regularization term, and is a regularization term.

Is. Further, here, a is a gain, and is appropriately experimentally determined, for example, a = 1.

は、条件Ｂに係る指示関数であり、（１）式の例では所定のｍを定めたとき、

であれば「１」、そうでなければ「０」となる関数である。 Further here

Is an indicator function according to the condition B, and in the example of the equation (1), when a predetermined m is defined,

If it is, it is "1", otherwise it is "0".

That is, in this loss function LPARIG, the distance between the distributed representation information of the attention node Bn and the distributed representation information of the node Cx (if there are a plurality of them) is the variance representation information of the attention node Bn and the node Cy (if there are a plurality of them). At least the distributed representation information of the attention node Bn is determined so that the distance is closer than the distance to the distributed representation information (of each of them), but in the first place, the distributed representation information of the attention node Bn and a certain node Cx Only when the distance from the distributed expression information of the node Bn of interest is sufficiently smaller than the distance between the distributed expression information of the node Cy of interest and the distributed expression information of a certain node Cy, the difference in the distance is taken into consideration.

The second machine learning processing unit 24 has the distributed representation information xi of the node Bn of interest, the distributed representation information xj (j = 1, 2, ...) Of the node Cx, and the distributed representation of the node Cy so as to minimize this loss LPARIG. The information is machine-learned by updating xj'(j'= 1, 2, ...).

The second machine learning processing unit 24 acquires the distributed representation information xi (i = 1, 2, ...) Of each node belonging to the layer B by this processing. Since the machine learning process for updating the distributed representation information so that the loss is small in the second machine learning processing unit 24 described here can be a widely known process, more detailed processing here can be adopted. The explanation is omitted.

The integrated processing unit 25 later evaluates the node representing the analysis target by using the distributed expression information obtained by the first machine learning processing unit 23 and the distributed expression information obtained by the second machine learning processing unit 24. Generates distributed representation information about the node representing the analysis target used in the processing unit 26.

とする。なお、ここで

は、ベクトルの要素を連接することを意味する。つまりここでは分散表現情報ｙiの次元がｎy、分散表現情報ｘiの次元がｎxであるとするとき、統合分散表現情報Ｘiは、その成分の数が（ｎy＋ｎx）であり、その最初のｎy個はｙiの成分とし、残るｎx個の成分はｘiの成分としたものである。 Specifically, the integrated processing unit 25 uses the distributed representation information Xi (hereinafter referred to as integrated distributed representation information) for each node Bi (i = 1, 2, ...) Belonging to the layer B as the first for the node Bi. Using the distributed expression information yi obtained by the machine learning processing unit 23 and the distributed expression information xi obtained by the second machine learning processing unit 24,

And. In addition, here

Means concatenating the elements of a vector. That is, here, assuming that the dimension of the distributed representation information yi is ny and the dimension of the distributed representation information xi is nx, the integrated distributed representation information Xi has the number of components (ny + nx), and the first ny pieces are The components of yi are used, and the remaining nx components are the components of xi.

The evaluation processing unit 26 evaluates the similarity between individual accounts using the integrated distributed representation information for each node belonging to the layer B obtained by the integrated processing unit 25. In this example, the evaluation processing unit 26 obtains distance information between the integrated distributed representation information of the plurality of nodes belonging to the layer B. This distance information may be the Euclidean distance or the cosine similarity (the product of the norms of each vector divided by the inner product of the vectors).

Then, the evaluation processing unit 26 determines that the distance is short based on the distance information between the node and another node for each node of the personal account (if it is a Euclidean distance, a predetermined threshold value). If it is less than, or if it is a cosine similarity, it exceeds a predetermined threshold of 0 or more and less than 1) is selected as a node of a similar account. Then, the evaluation processing unit 26 outputs information for identifying a node of a similar account selected for each node of the personal account.

The output processing unit 27 displays and outputs information such as the account name and its profile of the personal account specified by the information input from the evaluation processing unit 26 to the display unit 14, or displays and outputs the information on the user side via the communication unit 15. It is presented to the user of the information processing apparatus 1 by sending it to the terminal U or the like.

[motion]
The present embodiment has the above configuration and operates as follows. In the following example, the information processing device 1 is managed by a server device managed by an SNS operator such as Twitter (registered trademark) or Instagram (registered trademark), or a contractor contracted with the SNS operator. It is assumed that it is realized as a server device and is accessible from the user of the SNS.

The information processing device 1 of this example analyzes the personal account of the user of the SNS, the hashtag included in the posted text, and the official account of a company or the like having an account in the SNS. Further, the information processing device 1 receives information for identifying the personal account from the user in advance. Specifically, the information processing apparatus 1 accepts information that identifies the user himself / herself from a user who is one of the users of the SNS as information that identifies the personal account to be evaluated.

In the example of the following behavior, the accepted information shall present another similar personal account to the account identified for evaluation.

As illustrated in FIG. 4, the information processing device 1 of this example is included in a list of personal accounts of users of the SNS and posts from personal accounts belonging to the list from the website of the target SNS or the like. Get information about tags (hash tags) and information that identifies another account that the individual account follows (here, for the sake of explanation, it is limited to the official account) for each individual account that belongs to the list. (S11: Information gathering).

The personal account acquired by the information processing apparatus 1 here does not have to be the personal account of all SNS users. However, the personal account to be evaluated shall be included.

The information processing device 1 virtually includes a layer B including a node representing each of the acquired personal accounts, a layer A including a node of an official account, and a layer C including a node representing a hashtag used by the account. Set.

Then, the information processing apparatus 1 selects the nodes of the accounts belonging to the layer B, which is the analysis target including the evaluation target, one by one in a predetermined order, and from the selected nodes, the account represented by the node follows the layer A. Set the edge to the node of the official account to which it belongs.

Further, the information processing apparatus 1 sets an edge from the node of the account belonging to the selected layer B to the node of the hash tag included in the post of the account represented by the selected node. Further, the information processing apparatus 1 sets an edge between hashtags that appear in common (that is, co-occurrence) in one post in the node representing the hashtag belonging to the layer C.

Further, the information processing apparatus 1 sets weights on the edges between the nodes in the layer C according to the number of co-occurrence of hashtags represented by the pair of nodes related to the edges. Further, the information processing apparatus 1 also has the edge between the node of the personal account (node belonging to layer B) and the node of the hashtag included in the post from the personal account (node belonging to layer C). Set the weight according to the number of times the hashtag of the personal account is used (S12: Generate graph).

Next, the information processing apparatus 1 executes a first machine learning process for machine learning the distributed representation information of the nodes based on the connection relationship of the nodes included in the graph generated in step S12 (S13). Here, specifically, the information processing apparatus 1 uses a graph embedding method to obtain distributed representation information of each node.

That is, in step S13, the information processing apparatus 1 obtains distributed representation information of each node by using a so-called DeepWalk method (Bryan Perozzi, et al., DeepWalk: Online Learning of Social Representations, arXiv: 1403.6652). v2, 27 Jun 2014). More specifically, the information processing apparatus 1 generates one random integer value γ in step S13 and resets the counter k to k = 0.

The information processing apparatus 1 selects one of the nodes V belonging to the layer B as the initial node Vj (0) (j is the number of repetitions of this process, j = 1, 2, ...). Then, the information processing apparatus 1 records this initial node Vj (0). Next, the information processing apparatus 1 randomly selects one edge connected to the initial node Vj (0), moves to the adjacent node Vj (1) via the selected edge, and causes the relevant node Vj (1). The node Vj (1) is connected to the initial node and recorded. Then, the information processing apparatus 1 increments the counter k by "1".

Hereinafter, while k <γ, one edge included in the node Vj (k) is randomly selected from the destination node Vj (k), and the node is moved to the adjacent node Vj (k + 1) to move. The process of concatenating and recording the previous node Vj (k + 1) with the recording of the node that has been moved up to that point is repeated.

Here, when randomly selecting an edge connected to the node from the current node, the information processing apparatus 1 may consider the weight of the edge. When considering the weight of the edge, the information processing apparatus 1 controls the probability of being selected by the ratio of the weight set to the edge. For example, when n edges Ei (i = 1, 2, ..., N) are connected to the current node Vj (k) and a weight wi is set for each (here, for an edge without a weight setting). Is a predetermined value, a statistical value based on the weight of the edge connected to the node and the weight of the edge (mean value, median value, maximum value, minimum value, etc.), or in the graph. The information processing device 1 sets the probability of selecting the node Ei as wi / Σwi (Σwi means the sum of wis). .. As for the method of randomly selecting according to the probability, various well-known methods can be adopted, and thus the description thereof is omitted here.

As a result, the information processing apparatus 1 has a list of γ nodes (referred to as a node list) Lj (j = 1, 2, ...) :.
Lj = [Vj (0), Vj (1), Vj (2) ..., Vj (γ)]
To get.

The information processing apparatus 1 sequentially selects each of the nodes Vi belonging to the layer B and executes the above processing to obtain a random list of nodes having different lengths from each other.

The information processing device 1 is a node list obtained by tracing the edge starting from the node Vi belonging to the layer B, and includes at least a predetermined number (for example, 1 or more) of the nodes of the personal account to be evaluated. The process of step S13 is repeatedly executed until the node list of is obtained, and a plurality of node lists Lj (j = 1, 2, ...) Are generated.

For example, the information processing apparatus 1 may sequentially extract all the nodes of the layer B when selecting the node as the starting point. Further, the information processing apparatus 1 counts the number n of nodes included in the layer B, which are different from each other, in the node list obtained by the processing, and the value of this n is a predetermined threshold value (for convenience, the first one). It may be determined whether or not it exceeds the threshold value). Here, the first threshold value may be, for example, the maximum integer value floor (αN) that does not exceed the number of a predetermined ratio α (0 <α ≦ 1) among the number N of the nodes included in the layer B. .. Then, this n is below a predetermined first threshold value, or in the obtained node list, only the number of nodes of the account to be evaluated is less than a predetermined second threshold value. If it is not included, the information processing apparatus 1 may repeat the process of selecting a node as a starting point and then repeat the process of generating another node list.

As an example, as illustrated in FIG. 5, the A layer contains nodes V11, V12, V13 ..., The B layer contains nodes V21, V22 ..., And the C layer contains nodes V31, V32, V33 ... At this time, it is assumed that an edge is set between the B-layer node V21 and the A-layer node V11, and an edge is set between the node V21 and the C-layer nodes V31 and V32. Further, it is assumed that the edge is set between the node V22 of the B layer and the node V32 of the C layer, and the edge is set between the nodes V31, 32, V32, 33, and V31, 33 of the C layer, respectively. It shall be.

At this time, for example, it is assumed that V21 is set as the initial node and a node list consisting of five nodes is generated by the Deep Walk method. In this example, it moves from V21 in the order of V21 → V11 → V21 → V31 → V33 by random walk, and the sequence L1 = [V21, V11, V21, V31, V33].
To get.

Also, when the same V21 is used as the initial node, another node list is L2 = [V21, V31, V32, V22], which is a different order.
A node list such as may be generated.

Next, the information processing apparatus 1 performs skipgram learning by, for example, negative sampling, for each node included in each generated node list, and obtains vector information which is distributed representation information of each node. That is, the information processing apparatus 1 uses the generated node list as one "sentence" and the node included in the node list as the sentence as a "word" to generate distributed expression information by a widely known method such as word2vec. ..

As a result, the information processing apparatus 1 obtains distributed representation information about nodes representing a plurality of analysis targets, including the personal account to be evaluated.

Further, in the information processing apparatus 1, a part of a plurality of analysis targets in which the relationship set between the analysis targets is considered to have a predetermined relationship predetermined with each other and the relationship set between the analysis targets are set in advance with each other. Based on the relationship set for each of the analysis targets included in the second analysis target group, the second analysis target group is a part of the plurality of analysis targets that are not considered to have a predetermined relationship. Machine learning is performed on the distributed expression information related to each of the analysis targets included in the second analysis target group (S14).

Specifically, the information processing apparatus 1 initially (unless the distributed representation information of the nodes has been calculated by the process of step S14) randomly determines the initial value of the distributed representation information of each node.

Further, the information processing apparatus 1 sequentially selects each node of the general account belonging to the layer B as the node of interest Bn. Then, the information processing apparatus 1 selects at least one node Cx (example analysis target) whose edge is set between the selected node Bn and the node belonging to the layer C. Further, the information processing apparatus 1 selects at least one node Cy (negative example analysis target) whose edge is not set between the selected node Bn and the node belonging to the layer C, in the same number as the node Cx. do.

として求める。既に説明したように、ここでΘを含む項は正則化項（regularization term）であり、

である。さらにここで

は、条件Ｂに係る指示関数であり、（１）式の例では所定のｍを定めたとき、

であれば「１」、そうでなければ「０」となる関数である。 In the information processing apparatus 1, the distributed representation information of the node Bn of interest selected i-th is xi (i = 1, 2, ...), The distributed representation information of the node Cx is xj (j = 1, 2, ...), And the node Cy. The second machine learning processing unit 24 sets the loss LPARIG as xj'(j'= 1, 2, ...).

Ask as. As already explained, the term containing Θ here is a regularization term, and it is a regularization term.

Is. Further here

Is an indicator function according to the condition B, and in the example of the equation (1), when a predetermined m is defined,

If it is, it is "1", otherwise it is "0".

Using this loss LPARIG, the information processing apparatus 1 uses the gradient method to obtain the distributed representation information xi of the node Bn of interest, the distributed representation information xj (j = 1, 2, ...) Of the node Cx, and the distributed representation information of the node Cy. The distance between the distributed expression information xi of Bn and the distributed expression information xj of the node Cx is the distance between the distributed expression information xi of Bn and the distributed expression information xj of the node Cy. Machine learn to be less than the distance between.

The information processing apparatus 1 generates the distributed representation information xi (i = 1, 2, ...) Of the node used for the evaluation process, here, each node belonging to the layer B, independently of the process in step S13. Then, the information processing apparatus 1 concatenates the distributed representation information of the node obtained in step S13 and the distributed representation information of the node obtained in step S14 for each node to obtain the integrated distributed representation information of the node (). S15).

One of the characteristics of this embodiment is that the distributed representation information of the node obtained in step S13 is obtained based on the information obtained by tracing the edge from the node (that is, the local feature of the graph). On the other hand, the distributed representation information of the node obtained in step S14 is calculated using the information of other nodes that are not connected to the node, and is obtained based on the global characteristics of the graph. It is a thing.

By using the integrated distributed representation information including the distributed representation information based on the local features and the distributed representation information based on the global features in this way, the evaluation is performed based on the information including the local features and the global features. It will be possible to do.

The information processing apparatus 1 evaluates the relationship between the nodes to be evaluated by using the integrated distributed representation information obtained for each node. In the example here, the similarity between the node of the specified personal account and the node of another personal account is evaluated (S16).

That is, the information processing apparatus 1 represents, for example, the Euclidean distance of each integrated distributed representation information as distance information between the integrated distributed representation information of the node of the personal account that is the target of evaluation and the integrated distributed representation information of the node of another personal account. Ask for.

Then, the information processing apparatus 1 is determined to have a relatively short distance between the node of the personal account to be evaluated based on the obtained distance information and the node of another personal account based on the integrated distributed expression information. Select at least one. Then, the information processing apparatus 1 outputs the information of the personal account represented by the selected node and presents it to the user (S17).

For example, the information processing apparatus 1 is not connected to the node of the personal account to be evaluated through the edge among the nodes of the official account in the layer A connected to the selected node via the edge. Obtain information that identifies the node of the official account in Layer A.

Of the official account nodes in layer A that are linked to the node selected here via the edge, the official accounts in layer A that are not linked to the node of the account that was evaluated through the edge. If there is no node, select another node from the nodes of the account that is judged to be close based on the distance information obtained earlier, and connect to the selected node via the edge. Among the nodes of the official account in the layer A, the information that identifies the node of the official account in the layer A that is not connected to the node of the account to be evaluated via the edge may be obtained.

The information processing device 1 is the information related to the official account, that is, the account name of the official account, which is specified by the node not connected to the node of the account subject to the evaluation obtained here via the edge. , The profile information and the like are presented to the user. For example, the information processing apparatus 1 presents the account name of the official account, its profile information, and the like to the user. The presentation of this information may be performed, for example, by transmitting and displaying the information to the terminal device used by the user.

According to this example of this embodiment, the relationship between the user and the official account that the user follows, the relationship between the user and the hash tag used by the user, and the hash tag that co-occurs in one post. A graph network is formed based on information such as the relationship between the graphs, and the integrated distributed representation information of the user's personal account is used for both the local and global characteristics of the graph (note that the integrated distributed representation information is the official account. , Hash tags, etc. may also be obtained), enabling quantitative evaluation of the closeness of relationships between users. Therefore, even if each user uses a common hashtag less often, that is, even if the relationship is sparse, the distance between each user is based on both local and global characteristics. Can be quantified.

[Other examples of how to set edges]
In the explanation so far, when the analysis target or the edge to be set between the nodes representing the information items is determined when the graph is generated, the analysis target and the information item are divided into layers A, B, and layers for each type. After dividing into C ..., the edge was set according to a predetermined rule.

That is, in the examples so far,
• Between layer B and layer A, an edge is set between the node in layer B and the node in layer A that represents the official account that the account represented by that node follows.
An edge is set between the layer B and the layer C between the node in the layer B and the node in the layer C representing the hashtag included in the text posted by the account represented by the node.
-Between the nodes in layer C, an edge is set between the nodes representing each of the plurality of hashtags included (that is, co-occurring) at one time in one post.
I was supposed to. However, the edge setting method is not limited to this example.

In addition to this, for example, an edge may be set between the nodes of a pair of accounts that are following each other among the nodes in the layer B.

Further, in the node in the layer A, the edges may be set for the official accounts that are judged to be similar to each other by the investigation and analysis conducted separately. For example, if the official account is an account of some brand, an edge may be set between the official accounts of the brands analyzed as having similar customers.

[Weight normalization]
In the description so far, the information processing apparatus 1 sets weights on the edges between the nodes in the layer C according to the number of times the corresponding hashtags co-occur, and also sets the nodes belonging to the layer B and the nodes belonging to the layer B. The edge between the node and the node belonging to the layer C representing the hashtag included in the post from the corresponding personal account is also set to be weighted according to the number of times the hashtag of the personal account is used.

というように正規化した重みを設定してもよい。 This weight may be normalized as follows. Specifically, in the information processing apparatus 1, the corresponding personal account (referred to as node Vi) has the corresponding hash tag (node) as the weight for the edge between layers B and C (edge between the personal account and the hash tag). Using (let's say Vj) the number of times w'ij was used,

You may set the normalized weight like this.

Further, the information processing apparatus 1 sets weights for the edges between the pair of nodes Vi and Vj in the layer C as follows. That is, in the following example, the information processing apparatus 1 sets the initial weight to w'ij = 0 when the number of co-occurrence of hashtags is equal to or less than a predetermined threshold value such as "1". This is to reduce the complexity of the graph and reduce the processing load related to machine learning. Further, the information processing apparatus 1 initially sets the weight between the hashtags co-occurring a number of times exceeding the predetermined threshold value as w'ij = Nc (where Nc is the hashtag corresponding to the nodes Vi and Vj). The number of posts that have been raised, that is, the number of co-occurrence).

と設定する。 Then, the information processing apparatus 1 uses the initial weight w'ij to determine the weight of the edge between the pair of nodes Vi and Vj in the layer C.

And set.

By normalizing the weights in this way, for example, when performing a deep walk process, each node is adjusted to be selected with a relatively equal probability.

[Example of directed graph]
Further, in the above description, the generated graph is an undirected graph with no direction. However, for example, when setting an edge, a pair of nodes in which one of them follows the other among the nodes in the layer B. During (including non-mutual follow), you may want to set a directional edge from the node of the account you are following to the node of the account you are following. In this example, between a pair of nodes Va and Vb that follow each other, an edge Eba having a direction from Va to Vb and an edge Eba having a direction from Vb to Va (that is, bidirectional edges are ) Set.

Also, in this case, even if there is no mutual follow-up between the node of the official account in layer A and the node of the account in layer B, Vc is between the node Vc in layer A and the node Vd in layer B. Even if the official account of Vd does not follow the account of Vd, if the account of Vd follows the official account of Vc, it has the edge Edd having the direction from Vc to Vd and the direction from Vd to Vc. Set the edge Edc (that is, the bidirectional edge).

Similarly, bidirectional edges are set between layers B and C, and bidirectional edges are set between a pair of nodes corresponding to hashtags co-occurring in one post even between layers C. (It may be two edges as described above) is set.

In the case of a directed graph in this way, when performing a deep walk, it is assumed that the graph moves only according to the direction of the edge. Therefore, if there is no outward edge in the destination node, the list ends there.

[Other examples of operations by the second machine learning processing unit]
Further, in the description of the present embodiment so far, the second machine learning processing unit 24 uses xi (i = 1, 2, ...) And the distribution of the node Cx for the distributed representation information of the node Bn of interest selected i-th. The expression information is xj (j = 1, 2, ...), The distributed expression information of the node Cy is xj'(j'= 1, 2, ...), And LPARIG is calculated as the loss by the equation (1).

として損失ＬBPRを求めてもよい。 However, this is only an example, and the loss LBPR may be obtained by the following equation (2) instead of the LPARIG of the equation (1). That is, the second machine learning processing unit 24

The loss LBPR may be obtained as.

In this example as well, the second machine learning processing unit 24 uses the distributed representation information xi of the node Bn of interest and the distributed representation information xj (j = 1, 2, ...) Of the node Cx so as to minimize this loss LBPR. Machine learning is performed by updating xj'(j'= 1, 2, ...) The distributed representation information of the node Cy (for example, Steffen Rendle, et.al., "BPR: Bayesian personalized ranking from implicit feedback", UAI '09: Proceedings of the Twenty-Fifth Conference on Uncertainty in Artificial Intelligence, June 2009 Pages 452-461, etc.). Since this update method can be performed by the same method as the above-mentioned example except that the loss calculation is different, the description thereof will be omitted.

The method of generating distributed representation information representing global features in the second machine learning processing unit 24 is not limited to the methods described so far, and Yehuda Koren, et al., “MATRIX FACTORIZATION TECHNIQUES FOR RECOMMENDER SYSTEMS” ”, Computer, 42 (8), pp.30-37, 2009 and the method described in Xiangunan He, et al.,“ Neural Collaborative Filtering ”, arXiv: 1708.05031v2, 26 Aug 2017, etc. You may.

[Modification example of calculation in the first machine learning processing unit]
Further, the processing of the first machine learning processing unit 23 that generates the distributed representation information representing the local feature is not limited to the example described here. The first machine learning processing unit 23 uses so-called node2vec (Aditya Grover, et al., scalable Feature Learning for Networks, DOI: http://dx.doi.org/10.1145/2939672.2939754) to distribute local features. William Hamilton, et al., "Inductive Representation Learning on Large Graphs", In Advances in neural information processing systems, pp.1024-1034 (31st Conference on Neural Information Processing Systems (NIPS 2017)) The distributed representation information may be generated by adopting the method disclosed in)).

[Example of including intermediate features]
Further, in the description so far, the second machine learning processing unit 24 has a node Cy that is not in a predetermined relationship (for example, a relationship that can be reached by tracing the edge within a predetermined number of times) from the entire graph to the node Bn of interest. However, the present embodiment is not limited to this example. For example, the second machine learning processing unit 24 generates a subgraph including the attention node Bn from the subgraphs obtained from the graph, and selects a node Cy that does not have a predetermined relationship with the attention node Bn from the subgraph. You may do it. In this example, although it is a global feature, distributed representation information representing a midrange feature is generated.

In another example, the second machine learning processing unit 24 has a relationship that can be reached by tracing the edge from the attention node Bn within α times (where α is a natural number) and the edge from the attention node Bn. Is β (β> α + 1, α, β is a natural number), and the node Cy that is not in a relationship that can be reached is selected, and the node Cx, Cy is used to represent the global characteristics of the node Bn of interest. While generating distributed representation information, using the natural number γ such that β> γ> α, select a node Cy'that is not related to be reachable by tracing the edge from the note node Bn a number of times α + 1 or more and γ times or less, and select node Cx. And this node Cy'may be used to obtain distributed representation information representing the global characteristics of the midrange of the node Bn of interest. In the latter example, the extent to which global features are included can be adjusted by controlling the range of α, β (and γ).

[Other examples]
Further, in the explanation so far of this embodiment, an example of recommending an official account to be followed is shown to a user such as SNS, but the field of use of the information processing apparatus 1 of this embodiment is limited to this example. It is not something that can be done.

For example, from the official account side, a user who is close to a user who is following himself (called a following user) but who is not following himself (called an unfollowed user) is a user's personal account. It may be analyzed by using the integrated distributed representation information of the node.

That is, the information processing apparatus 1 responds to the request from the official account side and obtains the integrated distributed representation information related to the personal account of the user who is following the official account that made the request. Then, the information processing apparatus 1 randomly selects, for example, N pieces (N is a natural number of 1 or more) from the obtained integrated distributed expression information, and the selected integrated distributed expression information and the follow-up to the official account are not performed. Calculates the distance information from the user's integrated distributed representation information. Then, when it is determined that the pair of distributed representation information represented by the distance information are similar to each other (in the case of cosine similarity, when it is greater than 0 and equal to or less than a predetermined value), the information processing apparatus 1 , Present the information of the personal account of the unfollowed user to the official account of the requesting source.

When the integrated distributed expression information related to a plurality of following users is used (when N is 2 or more), it is judged to be similar to any of them, or the integrated distributed related to a certain unfollowed user. When the number of integrated distributed expression information related to the following user judged to be similar to the expression information is n, and n / N exceeds a predetermined judgment threshold, the account of the unfollowed user Information may be presented to the requesting official account.

Further, the analysis target of the information processing apparatus 1 of the present embodiment does not need to be information related to SNS. For example, in a system that introduces a marriage partner, the registrant (user who is looking for a marriage partner) is set as a node, and the information items that the registrant is interested in (information indicating the interest group to which the registrant belongs, etc.) are different. As a node, a network may be created by connecting these nodes at an edge.

In this case, the information processing apparatus 1 obtains distributed representation information representing local features related to the network for each registrant by, for example, a network embedding method, and Bayesian personalized ranking from implicitly obtains distributed representation information representing global features. Information such as feedback is obtained for each registrant, and integrated distributed representation information that integrates these is generated for each registrant.

Then, the information processing apparatus 1 may obtain the similarity between the registrants based on the similarity between the integrated distributed expression information, and may use the information of the similarity for the process of introducing the registrant who is the recommended partner. .. Also in this example, if there are interest groups to which the same registrant belongs in common, an edge may be set between the interest groups as well.

Further, in this example, the registrant X who has already applied for dating and has been refused by another registrant Q whose distributed expression information has a high degree of similarity to the distributed expression information related to himself (referred to as registrant P) is When present, it is conceivable to present the registrant P with a low possibility of dating with respect to the registrant X.

As described above, in an example of this embodiment, the similarity with the integrated distributed representation information of P, which is one of the analysis targets, is relatively high (for example, in the case of cosine similarity, it is greater than 0 and equal to or less than a predetermined value of 1 or more. There is another analysis target Q related to the integrated distributed representation information that is judged to be (such as), and the analysis target Q should not set an edge with the other analysis target X (so to speak, with a negative example). When there is a relationship), the information processing apparatus 1 presents to the user of the information processing apparatus 1 that the relationship between the analysis target P and the information item or the analysis target X can also be a negative example. May be good. Further, as the probability, for the number N of a plurality of other analysis targets Q1, Q2 ... Related to the integrated distributed expression information judged to be similar to the analysis target P, the information item or the analysis target X in the analysis target Q is used. The number n of the analysis target Q for which the relationship between the two is a negative example may be obtained, and the ratio n / N may be shown.

[Example of distributed information for official accounts]
Further, in an example of the present embodiment, the information processing apparatus 1 may obtain the distributed representation ai of the node Vi of the official account (layer A) as follows.

として求めてもよい。 That is, the set of the index k of the node Vk of the individual account of the layer B that follows the official account Ai represented by this node Vi is written as F (i), and the total number of its elements (the individual account that follows the official account Ai). The total number) is | F (i) |, and the information processing apparatus 1 uses the distributed representation ai of the official account Ai and the integrated distributed representation information Xk of the node Vk of the layer B.

May be sought as.

In this example, the similarity (score) between the node Vj of the personal account and the node Vi of this official account Ai is determined by the cosine similarity between ai and the integrated distributed representation information Xj of the node Vj of the personal account. May be good.

1 Information processing device, 11 Control unit, 12 Storage unit, 13 Operation unit, 14 Display unit, 15 Communication unit, 21 Information acquisition unit, 22 Graph generation unit, 23 1st machine learning processing unit, 24 2nd machine learning processing unit , 25 integrated processing unit, 26 evaluation processing unit, 27 output processing unit.

Claims (6)

A means for holding vector values set for each analysis target as distributed representation information for a plurality of analysis targets for which mutual relationships are set, and
A part of the plurality of analysis targets, which are considered to have a predetermined first relationship in which the relationships set between the analysis targets are predetermined to each other, is extracted as a first analysis target group, and the extracted first. A first machine learning means for machine learning distributed expression information related to each of the analysis targets included in the first analysis target group based on the order or combination of the analysis target groups.
A part of the plurality of analysis targets in which the relationship set between the analysis targets is considered to have a predetermined second relationship predetermined to each other, and the relationship set between the analysis targets are predetermined to each other. Based on the relationship set for each of the analysis targets included in the second analysis target group, a part of the plurality of analysis targets that are not considered to have a predetermined second relationship is set as the second analysis target group. , A second machine learning means for machine learning distributed expression information related to each of the analysis targets included in the second analysis target group.
A processing means for subjecting the distributed expression information for each analysis target machine-learned by the first machine learning means and the second machine learning means to a predetermined process,
Information processing equipment including. The information processing apparatus according to claim 1.
The relationship set between the analysis targets is represented by graph information in which each of the analysis targets is a node and the nodes are connected.
The predetermined first relationship determined in advance is that it is determined that the relationship is connected to each other by a predetermined condition using the graph information.
The first machine learning means selects one of the analysis targets by a predetermined method, and a part of the plurality of analysis targets related to the node determined to be connected to the node related to the selected analysis target. Is an information processing device that extracts as a first analysis target group. The information processing apparatus according to claim 1 or 2.
The relationship set between the analysis targets is represented by graph information in which each of the analysis targets is a node and the nodes are connected.
The predetermined second relationship is that the corresponding nodes represented by the graph information are connected to each other.
The second machine learning means selects one of the analysis targets by a predetermined method, and the analysis target related to the node connected to the node related to the selected analysis target and the node related to the selected analysis target. An information processing device that extracts an analysis target related to a node that is determined not to be connected to the second analysis target group. The information processing apparatus according to claim 3.
The second machine learning means sets the analysis target related to the node determined to be connected to the node related to the selected analysis target among the analysis targets included in the second analysis target group as the normal analysis target. , The set of the selected analysis target, the positive analysis target, and the negative analysis target, with the analysis target related to the node determined not to be connected to the node related to the selected analysis target as the negative example analysis target. An information processing device that machine-learns at least the distributed expression information related to the selected analysis target using the above. The information processing apparatus according to any one of claims 1 to 4.
The processing means evaluates the degree of similarity between at least a part of the analysis target by using the distributed expression information for each analysis target machine-learned by the first machine learning means and the second machine learning means. An information processing device that executes the processing to be performed. Computer,
A means for holding vector values set for each analysis target as distributed representation information for a plurality of analysis targets for which mutual relationships are set, and
A part of the plurality of analysis targets, which are considered to have a predetermined first relationship in which the relationships set between the analysis targets are predetermined to each other, is extracted as a first analysis target group, and the extracted first. A first machine learning means for machine learning distributed expression information related to each of the analysis targets included in the first analysis target group based on the order or combination of the analysis target groups.
A part of the plurality of analysis targets in which the relationship set between the analysis targets is considered to have a predetermined second relationship predetermined to each other and the relationship set between the analysis targets are predetermined to each other. Based on the relationship set for each of the analysis targets included in the second analysis target group, a part of the plurality of analysis targets that are not considered to have a predetermined second relationship is set as the second analysis target group. , A second machine learning means for machine learning distributed expression information related to each of the analysis targets included in the second analysis target group.
A program that functions as a processing means for subjecting distributed expression information for each analysis target machine-learned by the first machine learning means and the second machine learning means to a predetermined process.

Images