Advertisement delivery condition inference device, method, program, and advertisement delivery condition learning device

Abstract

The present invention involves: acquiring a purpose for delivering an advertisement; acquiring, as inference data, an action history sequence which includes first actions relating to recognition and utilization by customers regarding a service, second actions relating to reception, by the customers, of the advertisement having been delivered, and dates and times when actions were performed and which has actions in future dates and times being masked; predicting a probability value for an action in each date and time by using an action model; identifying and replacing an inducing action on the basis of conditions according to the purpose and the probability value predicted for the acquired inference data; selecting, on the basis of the conditions according to the purpose and a probability value predicted for inference data on the inducing action having undergone the replacement, intervention actions from among the second actions; and outputting, as delivery conditions, an intervention action which causes the increase amount of the probability value for the inducing action before and after the intervention to be equal to or greater than a prescribed value and the corresponding date and time, and outputting, as an effect, the increase amount.

Description

[Technical Field]

[0001] The disclosed technology relates to an advertisement delivery condition inference device, an advertisement delivery condition inference method, an advertisement delivery condition inference program, and an advertisement delivery condition learning device. [Background technology]

[0002] Many companies are conducting marketing activities using ad delivery. There are two main ways to deliver ads.

[0003] One is ad delivery driven by corporate objectives. Generally, when a company delivers ads, the objectives are to increase awareness of the service, encourage new use, encourage continued use (prevent cancellations), encourage cross-service use, and encourage frequent use. In line with these ad delivery objectives, marketers manually create hypotheses about the customer profile, and then determine conditions for ad delivery such as target audience, content, timing, and delivery channel based on those hypotheses, and deliver the ads according to those conditions. Ad delivery driven by corporate objectives often involves delivering ads specialized for a specific service to a large number of customers at once.

[0004] The other is customer behavior-driven ad delivery, which delivers personalized ads based on the status and behavior of each individual customer. For example, collaborative filtering and recommendation systems (see, for example, Non-Patent Document 1) predict the next action a customer may take based on their behavioral patterns and deliver ads related to that action. Customer behavior-driven ad delivery often targets multiple services rather than a specific service. [Prior art documents] [Non-patent literature]

[0005] [Non-Patent Document 1] Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, Peng Jiang, "BERT4Rec: Sequential Recommendation with Bidirectional Encoder Representations from Transformer", arXiv:1904.06690v2 [cs.IR], 21 Aug 2019. Summary of the Invention [Problem to be solved by the invention]

[0006] While ad delivery driven by corporate objectives is effective in promoting specific services in line with their objectives, it is difficult to match with the state and behavior of each individual customer, which can lead to a decrease in the percentage of ads that are successful in appealing to the advertiser (CVR = number of successful ads / number of ads delivered).On the other hand, ad delivery driven by customer behavior is effective in increasing customer satisfaction by appealing to content that is closely related to the customer, but the ads do not necessarily meet the objectives that the company is aiming for, and there is a problem that the number of ads delivered may decrease even if the CVR is increased.

[0007] Conventional technology has made it difficult to balance such ad delivery objectives with customer-driven ad delivery. Ad delivery driven by corporate objectives can result in the inclusion of ad delivery that is inappropriate for customer behavior, lowering CVR and reducing advertising effectiveness. Ad delivery driven by customer behavior can result in the inclusion of ad delivery that is inappropriate for corporate objectives, reducing the number of ads delivered that are in line with corporate objectives and reducing advertising effectiveness.

[0008] The disclosed technology has been developed in consideration of the above points, and aims to personalize advertising delivery based on customer behavior in accordance with the purpose of the advertising in an environment where behavioral history can be tracked on a customer-by-customer basis. [Means for solving the problem]

[0009] A first aspect of the present disclosure is an advertisement delivery condition inference device, comprising: a purpose acquisition unit that acquires a purpose for delivering an advertisement; a data acquisition unit that acquires, as inference data, a series of behavioral histories at each date and time, including a first behavior related to at least one of a customer's awareness of and use of a service and a second behavior related to the customer's receipt of a delivered advertisement, and a date and time at which each of the first behavior and the second behavior was performed, in which behaviors at each date and time in the future than the present are masked; a prediction unit that predicts a probability value that the masked behaviors in the inference data are each of a plurality of predetermined behaviors using a behavior model that predicts behaviors at each future date and time from behaviors at each past date and time; and a prediction unit that predicts a probability value that the masked behaviors in the inference data are each of a plurality of predetermined behaviors according to the purpose acquired by the purpose acquisition unit. and a replacement unit that identifies a triggering action from the first action based on conditions and a probability value predicted by the prediction unit for the inference data acquired by the data acquisition unit, and replaces the masked action at the date and time when the triggering action was identified with the triggering action; and an output unit that selects an intervention action from the second action based on conditions according to the purpose acquired by the purpose acquisition unit and a probability value predicted by the prediction unit for the inference data after replacement with the triggering action, and outputs the intervention action and the corresponding date and time as distribution conditions, where the increase in the probability value of the triggering action predicted by the prediction unit for the inference data from which the intervention action was selected is greater than or equal to a predetermined value compared to the probability value of the triggering action before the intervention action was selected.

[0010] A second aspect of the present disclosure is an advertisement delivery condition inference method, in which an intention acquisition unit acquires an intention to deliver an advertisement, a data acquisition unit acquires, as inference data, a series of behavioral histories at each date and time including a first behavior related to at least one of a customer's awareness of and use of a service and a second behavior related to the customer's receipt of a delivered advertisement, and a date and time when each of the first behavior and the second behavior was performed, in which behaviors at each date and time in the future than the present are masked, a prediction unit predicts a probability value that the masked behavior in the inference data is each of a plurality of predetermined behaviors using a behavior model that predicts behaviors at each future date and time from behaviors at each past date and time, and a replacement unit replaces the masked behaviors in the inference data acquired by the intention acquisition unit with a probability value that the masked behavior in the inference data is each of a plurality of predetermined behaviors, and a replacement unit replaces the masked behaviors in the inference data with a probability value that the masked behavior in the inference data is each of a plurality of predetermined behaviors using a behavior model that predicts behaviors at each future date and time from behaviors at each past date and time. A triggering action is identified from the first action based on conditions according to the purpose and the probability value predicted by the prediction unit for the inference data acquired by the data acquisition unit, and the masked action at the date and time when the triggering action was identified is replaced with the triggering action. An output unit selects an intervention action from the second action based on conditions according to the purpose acquired by the purpose acquisition unit and the probability value predicted by the prediction unit for the inference data after replacement with the triggering action, and outputs the intervention action and the corresponding date and time as delivery conditions, where the increase in the probability value of the triggering action predicted by the prediction unit for the inference data from which the intervention action was selected is greater than or equal to a predetermined value compared to the probability value of the triggering action before the intervention action was selected.

[0011] A third aspect of the present disclosure is an advertisement delivery condition inference program that causes a computer to function as each unit of the advertisement delivery condition inference device.

[0012] A fourth aspect of the present disclosure is an advertising delivery condition learning device including: a data acquisition unit that acquires, as learning data, a series of behavioral histories, including a first behavior related to at least one of a customer's awareness of and use of a service, and a second behavior related to the customer's receipt of a delivered advertisement, and behavioral histories at each date and time, the first behavior and the second behavior being the date and time when each of the first behavior and the second behavior was performed, with some of the behavioral history being masked; and a learning unit that learns the behavioral model by inputting all of the learning data acquired by the data acquisition unit into the key and the query and inputting only the first behavior and the second behavior of the learning data into the value, of the keys, queries, and values ​​that constitute an attention mechanism provided in the behavioral model that predicts behavior at each future date and time from behavior at each past date and time, and updating parameters of the behavioral model so that a prediction result predicted by the behavioral model for the masked behavior in the learning data matches a correct behavior. [Effects of the Invention]

[0013] According to the disclosed technology, in an environment where behavioral history can be tracked on a customer-by-customer basis, it is possible to personalize advertisement delivery based on customer behavior in accordance with the purpose of the advertisement. [Brief explanation of the drawings]

[0014] [Figure 1] FIG. 2 is a block diagram showing a hardware configuration of the advertisement delivery condition inference device. [Figure 2] FIG. 2 is a block diagram illustrating an example of a functional configuration of an advertisement delivery condition inference device. [Figure 3] FIG. 10 is a diagram illustrating an example of a sequence of behavioral histories. [Figure 4] FIG. 10 is a diagram illustrating an example of a behavior ID dictionary. [Figure 5] FIG. 10 is a diagram showing an example of a behavior history sequence in which the sequence has been adjusted to a specified length. [Figure 6] FIG. 10 is a diagram illustrating an example of training data after masking processing. [Figure 7] FIG. 10 is a diagram illustrating an example of training data to which sequence position identifiers are assigned. [Figure 8] FIG. 10 is a diagram illustrating the structure of a model for a comparative example that receives training data as input and outputs hidden layer values ​​of keys, queries, and values. [Figure 9] FIG. 10 is a diagram illustrating the structure of a model for a comparative example, which outputs hidden layer values ​​of a key, a query, and a value in an intermediate layer. [Figure 10] FIG. 10 is a diagram illustrating the structure of a model that outputs a probability value indicating the likelihood of each behavior occurring, in a comparative example. [Figure 11] FIG. 10 is a diagram showing the structure of a model that inputs learning data and outputs values ​​of a hidden layer of values ​​in the behavior model of this embodiment. [Figure 12] FIG. 10 is a diagram showing the structure of a model that receives learning data as input and outputs hidden layer values ​​of keys and queries for the behavior model of this embodiment. [Figure 13] FIG. 10 is a diagram showing the structure of a model that outputs values ​​of hidden layers of values ​​in an intermediate layer in the behavioral model of this embodiment. [Figure 14] FIG. 10 is a diagram showing the structure of a model that outputs hidden layer values ​​of keys and queries in the intermediate layer of the behavioral model of this embodiment. [Figure 15] FIG. 10 is a diagram showing the structure of a behavior model according to the present embodiment, which outputs a probability value indicating the likelihood of each behavior occurring. [Figure 16] 10 is a flowchart showing the flow of a learning process. [Figure 17] 10 is a flowchart showing the flow of an inference process. [Figure 18] FIG. 10 is a diagram for explaining a specific example of an inference process. DETAILED DESCRIPTION OF THE INVENTION

[0015] An example of an embodiment of the disclosed technology will be described below with reference to the drawings. Note that the same reference numerals are used to designate identical or equivalent components and parts in each drawing. Also, the dimensional proportions in the drawings are exaggerated for the sake of explanation and may differ from the actual proportions.

[0016] Fig. 1 is a block diagram showing the hardware configuration of an advertisement delivery condition inference device. As shown in Fig. 1, an advertisement delivery condition inference device 10 includes a CPU (Central Processing Unit) 11, a ROM (Read Only Memory) 12, a RAM (Random Access Memory) 13, a storage 14, an input unit 15, a display unit 16, and a communication I / F (Interface) 17. Each component is connected to each other via a bus 19 so as to be able to communicate with each other.

[0017] The CPU 11 is a central processing unit that executes various programs and controls each component. That is, the CPU 11 reads programs from the ROM 12 or the storage 14 and executes the programs using the RAM 13 as a work area. The CPU 11 controls the above components and performs various arithmetic processing in accordance with the programs stored in the ROM 12 or the storage 14. In this embodiment, the ROM 12 or the storage 14 stores an advertisement delivery condition inference program for executing the learning process and inference process described below.

[0018] The ROM 12 stores various programs and various data. The RAM 13 temporarily stores programs or data as a working area. The storage 14 is composed of storage devices such as an HDD (Hard Disk Drive) or SSD (Solid State Drive), and stores various programs including the operating system and various data.

[0019] The input unit 15 includes a pointing device such as a mouse and a keyboard, and is used to perform various inputs. The display unit 16 is, for example, a liquid crystal display, and displays various information. The display unit 16 may employ a touch panel system and function as the input unit 15. The communication I / F 17 is an interface for communicating with other devices. For this communication, for example, a wired communication standard such as Ethernet (registered trademark) or FDDI, or a wireless communication standard such as 4G, 5G, or Wi-Fi (registered trademark) is used.

[0020] Next, a description will be given of the functional configuration of the advertisement delivery condition inference device 10. Fig. 2 is a block diagram showing an example of the functional configuration of the advertisement delivery condition inference device 10.

[0021] 2, the advertisement delivery condition inference device 10 has, as functional components, a data acquisition unit 22, a learning unit 24, a behavioral model 26, a goal acquisition unit 28, a prediction unit 30, a substitution unit 32, and an output unit 34. Each functional component is realized by the CPU 11 reading out an advertisement delivery condition inference program stored in the ROM 12 or the storage 14, expanding it in the RAM 13, and executing it.

[0022] During learning, the data acquisition unit 22 acquires learning data, which is a series of behavioral histories for each customer. The behavioral history includes a first behavior related to at least one of the customer's awareness and use of the service, a second behavior related to the customer's receipt of a distributed advertisement, and the date and time when each of the first behavior and the second behavior was performed. The behavioral history may also include other supplementary information. The series of behavioral histories is a behavioral history for each date and time from a past date and time to the present.

[0023] FIG. 3 shows an example of a behavioral history sequence. In the example of FIG. 3, each behavioral history includes five items: "customer identifier," "date and time," "behavioral identifier," "supplementary information 1," and "supplementary information 2." The "date and time" is the date and time when the customer identified by the "customer identifier" performed the behavior identified by the "behavioral identifier." The "behavioral identifier" is an identifier for each of the first behavior and the second behavior. A different identifier is assigned to the first behavior for each service recognized or used or for each type of service, and a different identifier is assigned to the second behavior for each advertisement received or for each type of advertisement. For example, the behavioral identifier for recognition of service 1 may be A, the behavioral identifier for recognition of service 2 may be B, the behavioral identifier for use of service 1 may be C, the behavioral identifier for use of service 2 may be D, etc. The behavioral identifier for receipt of advertisement 1 may be X, the behavioral identifier for receipt of advertisement 2 may be Y, etc. The supplementary information may be, for example, customer attributes such as age, or information related to the service such as points held.

[0024] The data acquiring unit 22 generates an activity ID dictionary in which activity identifiers included in the learning data are paired with activity IDs linked to the activity identifiers. FIG. 4 shows an example of the activity ID dictionary. The data acquiring unit 22 includes three special activity identifiers, [MASK], [UNK], and [PAD], in advance in the activity ID dictionary. [MASK] is an identifier used in the masking process described below, [UNK] is an identifier used when the activity is unknown, and [PAD] is an identifier used as a filler when the sequence does not satisfy a specified length. The data acquiring unit 22 converts the activity identifiers of the learning data into activity IDs based on the generated activity ID dictionary.

[0025] In addition, the data acquisition unit 22 rearranges each behavior history included in the learning data by customer identifier in the order of "date and time" and adjusts the behavior history sequence to a specified length ("8" in this embodiment). The length of the sequence is the number of data (number of records, number of lines) of behavior history included in the learning data. If the behavior history sequence does not meet the specified length, the data acquisition unit 22 creates a behavior history sequence in which the behavior ID is set to the behavior ID corresponding to the behavior identifier [PAD] and the value of the numerical field is set to 0, thereby filling in the sequence. Figure 5 shows an example of a behavior history sequence of a customer identified by customer identifier "001", in which the behavior identifiers are converted to behavior IDs, rearranged in the order of date and time, and adjusted to the specified length. In the example of Figure 5, the "date and time" field is converted to Unix time.

[0026] The data acquisition unit 22 performs a masking process to mask some of the behavioral histories included in the learning data. For example, the data acquisition unit 22 randomly selects 20% of the behavioral histories of the entire sequence, and converts the behavior IDs of the selected behavioral histories into behavior IDs corresponding to the behavior identifier [MASK]. The data acquisition unit 22 also replaces the values ​​of the numerical items of the selected behavioral histories with 0. FIG. 6 shows an example of the learning data after the masking process. FIG. 6 shows an example in which the behavioral history in the second row is masked.

[0027] Note that setting the masked behavior history percentage to 20% is just an example and is not limited to this, and the masking percentage may be selected from probability values ​​less than 100%. Furthermore, the behavior history to be masked is not limited to being selected randomly, and a certain number of recent behavior histories may be selected to improve the prediction accuracy of future behavior histories.

[0028] The data acquiring unit 22 assigns a sequence position identifier to the learning data after the masking process, which indicates the position of each behavior history in the sequence, i.e., the ordinal number of each behavior history in the sequence. FIG. 7 shows an example of learning data to which sequence position identifiers have been assigned. In the example of FIG. 7, sequence position identifiers ranging from 0 to 7 (a value 1 less than the specified length of the sequence) are assigned in order from the oldest behavior history to the newest behavior history in terms of "date and time". The data acquiring unit 22 passes the learning data after the masking process and after the assignment of sequence position identifiers to the learning unit 24.

[0029] Furthermore, during inference, the data acquisition unit 22 acquires inference data, which is a series of behavioral histories for each customer. The data configuration of the inference data is the same as the data configuration of the learning data shown in FIG. 3. As with the learning data, the data acquisition unit 22 also converts behavior identifiers into behavior IDs using the generated behavior ID dictionary, sorts the behavior histories in chronological order, aligns the series to a specified length, and assigns a series position identifier to the inference data. At this time, the data acquisition unit 22 creates a behavior history for future dates and times in which the behavior ID is the behavior ID corresponding to the behavior identifier [MASK] and other items are set to 0. The data acquisition unit 22 passes the inference data after the series position identifier has been assigned to the prediction unit 30.

[0030] The learning unit 24 uses the training data after the masking process and after the assignment of sequence position identifiers to train a behavior model 26 that predicts behavior at each future date and time from behavior at each past date and time. The behavior model 26 is a model equipped with an attention mechanism including a key, a query, and a value, and receives the training data after the masking process and after the assignment of sequence position identifiers as input, and outputs a prediction result of the behavior ID of the masked behavior history. The behavior model 26 may be, for example, an MLM (Masked Language Model).

[0031] 8 to 10 show the structure of a model of BERT (Bidirectional Encoder Representations from Transformers) described in Non-Patent Document 1 above as a comparative example of this embodiment. FIG. 8 is a diagram showing the structure of a model that receives training data as input and outputs hidden layer values ​​of keys, queries, and values ​​in a comparative example. FIG. 9 is a diagram showing the structure of a model that outputs hidden layer values ​​of keys, queries, and values ​​in an intermediate layer in a comparative example. FIG. 10 is a diagram showing the structure of a model that outputs a probability value indicating the likelihood of each behavior occurring in a comparative example.

[0032] The comparative example includes an attention mechanism in the model and uses a technique called self-attention. That is, as shown in FIGS. 8 to 10, in the comparative example, the behavior ID and sequence position identifier are provided as input to each of the key, query, and value constituting the attention mechanism, and the same input is provided to each of the key, query, and value in the intermediate layer. More specifically, as shown in FIG. 8, in the comparative example, the behavior ID and sequence position identifier are converted into embedding vectors (item embedding and position embedding), respectively. Then, the item embedding and position embedding are input to each of the key, query, and value, respectively. The similarity between the query and key is then calculated in the attention layer, which includes matmul and softmax. Next, the matrix product of the similarity and the value is calculated in matmul, and linearly transformed in Linear to obtain the output of the first hidden layer (hidden_1(value, query, key)).

[0033] As shown in Figure 9, in the (n+1)th hidden layer, the output of the nth hidden layer (hidden_n(value, query, key)), which is the same value, is input to each of the key, query, and value. Subsequent processing in the (n+1)th hidden layer is the same as that in the first hidden layer shown in Figure 8. Then, as shown in Figure 10, in the output layer, matmul calculates the matrix product of the final hidden layer output (hidden_n(value, query, key)) and the item embedding, and the vector value of item bias is added to the calculation result. The addition result is then converted using a softmax function so that the sum = 1, and the probability value of each behavior ID is calculated and output.

[0034] On the other hand, in this embodiment, all items of the learning data are input as the key and query, and only the behavior ID item of the learning data is input as the value. A more detailed description will be given with reference to the structure of the model in this embodiment shown in Figures 11 to 15. Note that in Figures 11 to 15, detailed description of the same configuration as the structure of the model in the comparative example shown in Figures 8 to 10 will be omitted.

[0035] FIG. 11 is a diagram showing the structure of a model for the behavioral model 26 of this embodiment that outputs hidden layer values ​​for values ​​from input of learning data. FIG. 12 is a diagram showing the structure of a model for the behavioral model 26 of this embodiment that outputs hidden layer values ​​for keys and queries from input of learning data. FIG. 13 is a diagram showing the structure of a model for the behavioral model 26 of this embodiment that outputs hidden layer values ​​for values ​​in the intermediate layer. FIG. 14 is a diagram showing the structure of a model for the behavioral model 26 of this embodiment that outputs hidden layer values ​​for keys and queries in the intermediate layer. FIG. 15 is a diagram showing the structure of a model for the behavioral model 26 of this embodiment that outputs probability values ​​indicating the likelihood of each behavior occurring.

[0036] As shown in FIG. 11, in the first hidden layer, values ​​of items other than the behavior ID and sequence position identifier of the training data are converted into embedding vectors (feature embedding). Then, only the item embedding is input as value, and all embedding vectors are input as query and key. This model configuration results in an output (hidden_1(value)) to be input as value in the second hidden layer. Furthermore, as shown in FIG. 12, all of the item embedding, position embedding, and feature embedding are input as key, query, and value, respectively. This model configuration results in an output (hidden_1(query,key)) to be input as query and key in the second hidden layer.

[0037] 13, in the n+1th hidden layer that outputs a value to be input to the value in the n+2th layer, hidden_n(value) is input to the value, and hidden_n(query, key) is input to each of the query and key. Also, as shown in FIG. 14, in the n+1th hidden layer that outputs a value to be input to the key and query in the n+2th layer, hidden_n(query, key) is input to all of the key, query, and value.

[0038] Then, as shown in FIG. 15, in the output layer, the value of Value (hidden_n(value)), which is the output of the final hidden layer, is used to calculate and output the probability value of each behavior ID. Note that the design of the layer that calculates the probability value is not limited to this. For example, it may be designed so that the item embedding and hidden_n(value) each have a norm of 1, the matrix product is replaced with the inner product value of a unit matrix, and the probability value of each behavior ID is derived from the angle between the vectors.

[0039] In this way, the behavioral model 26 of this embodiment imposes the above-mentioned input / output constraints and adopts an input format different from the self-attention of the comparative example. In this embodiment, by using only the time series of the behavioral ID as the value, overlearning is suppressed, the embedding vectors of the date and time and supplemental information are prevented from affecting as noise, and the quality of the embedding vector (semantic vector) of the behavioral ID is improved. Note that, although n=3 in this embodiment, n may be any number equal to or greater than 0, and is not limited to this.

[0040] The learning unit 24 inputs the masked learning data and the assigned serial position identifiers to the behavioral model 26 in accordance with the input / output constraints of the behavioral model 26, thereby obtaining a prediction result of the masked behavioral ID predicted by the behavioral model 26. The learning unit 24 then repeatedly updates the parameters of the behavioral model 26 so that the prediction result matches the behavioral ID before masking, i.e., the correct behavioral ID, until a learning termination condition is met. The termination condition may be, for example, when the number of parameter updates reaches a predetermined number, when the value of the loss function calculated during the parameter update process becomes equal to or less than a predetermined value, or when the difference between the value of the loss function calculated last time and the value of the loss function calculated this time becomes equal to or less than a predetermined value. The learning unit 24 stores the behavioral model 26, in which the parameters are set when the termination condition is met, in a predetermined storage area of ​​the advertisement delivery condition inference device 10.

[0041] The purpose acquisition unit 28 acquires the purpose for distributing the advertisement. The purpose may be, for example, to increase awareness of the service, promote new use of the service, promote continued use of the service, promote cross-service use, or promote frequent use.

[0042] The prediction unit 30 uses the behavior model 26 learned by the learning unit 24 to predict the probability that the masked behavior in the inference data is each of a plurality of predetermined behaviors. Specifically, the prediction unit 30 inputs the inference data after the serial position identifiers have been assigned, which has been passed from the data acquisition unit 22, into the behavior model 26, and predicts the likelihood of each behavior occurring at each future date and time in the inference data, i.e., the probability value for each behavior ID. The prediction result may be in a format such as behavior ID=3:0.4, behavior ID=4:0.4, behavior ID=5:0.15, and behavior ID=6:0.05, for example.

[0043] Based on the prediction result by the prediction unit 30 for the inference data acquired by the data acquisition unit 22, the replacement unit 32 identifies as the triggering behavior a first behavior that satisfies the conditions according to the purpose acquired by the purpose acquisition unit 28 and has a probability value predicted by the prediction unit 30 equal to or greater than a predetermined value. The replacement unit 32 replaces the masked behavior at the date and time at which the triggering behavior was identified with the triggering behavior. That is, the replacement unit 32 replaces the behavior ID corresponding to [MASK] at the relevant date and time with the behavior ID corresponding to the triggering behavior. The replacement unit 32 predicts the probability value of each behavior ID at each date and time for the inference data after replacement with the behavior ID corresponding to the triggering behavior using the behavior model 26.

[0044] The conditions according to the purpose may be determined based on the purpose acquired by the purpose acquisition unit 28 and the corresponding relationship between the purpose and the conditions defined in advance. Alternatively, the purpose acquisition unit 28 may acquire a written version of the purpose as a condition. If the purpose is to increase awareness of the service, the conditions may include a condition limiting the behavior related to customer awareness of the service. If the purpose is to promote new use of the service, the conditions may include a condition limiting the behavior related to customer use of a service that the customer has not used before. If the purpose is to promote continued use of the service, the conditions may include a condition limiting the behavior related to customer use of a service that the customer has used before. If the purpose is to promote cross-service use, the conditions may include a condition limiting the behavior related to the customer using two or more services on the same date and time. If the purpose is to promote frequent use, the conditions may include a condition limiting the behavior related to the customer's use of a service that the customer uses frequently. Furthermore, if the purpose is to promote frequent use, the conditions may include a narrowing condition related to the track record or frequency of past behavior related to the customer's use of a service that the customer uses frequently. The conditions according to the purpose of advertisement distribution are not limited to the above examples, and may be any conditions that satisfy the advertisement purpose.

[0045] The output unit 34 selects one of a plurality of predetermined second actions as the intervention action based on the purpose of the advertisement and the prediction result for the inference data after replacement by the replacement unit 32. Then, the output unit 34 causes the prediction unit 30 to predict the probability value of the triggering action for the inference data for which the intervention action has been selected. As the intervention action, the action with the highest probability value indicated by the prediction result may be selected from among the actions according to the purpose acquired by the purpose acquisition unit 28, or all of the actions according to the purpose may be selected. Then, the output unit 34 outputs, as the delivery condition, the intervention action for which the increase in the probability value of the triggering action before and after selecting the intervention action is equal to or greater than a predetermined value, and the corresponding date and time. Note that, if multiple intervention actions are selected, the increase in the probability value of the triggering action before and after the intervention for each intervention action may be calculated, and the intervention action with the largest increase may be set as the delivery condition. Furthermore, the output unit 34 outputs the increase as an index indicating the effectiveness of delivering an advertisement based on the delivery condition.

[0046] Next, the operation of the advertisement delivery condition inference device 10 will be described. Fig. 16 is a flowchart showing the flow of the learning process by the advertisement delivery condition inference device 10. Fig. 17 is a flowchart showing the flow of the inference process by the advertisement delivery condition inference device 10. The CPU 11 reads out the advertisement delivery condition inference program from the ROM 12 or the storage 14, expands it in the RAM 13, and executes it, thereby performing the learning process and the inference process. The learning process and the inference process are examples of an advertisement delivery condition inference method of the disclosed technology.

[0047] First, the learning process shown in FIG. 16 will be described.

[0048] In step S10, the CPU 11, functioning as the data acquisition unit 22, acquires learning data, which is a series of behavioral histories for each customer. Next, in step S12, the CPU 11, functioning as the data acquisition unit 22, generates a behavior ID dictionary in which behavior identifiers included in the learning data are paired with behavior IDs linked to the behavior identifiers, and converts the behavior identifiers of the learning data into behavior IDs based on the behavior ID dictionary.

[0049] Next, in step S14, the CPU 11, as the data acquisition unit 22, sorts each behavior history included in the learning data in order of "date and time" for each customer identifier, and aligns the behavior history series to a specified length. Next, in step S16, the CPU 11, as the data acquisition unit 22, masks some behavior history included in the learning data. Next, in step S18, the CPU 11, as the data acquisition unit 22, assigns a series position identifier indicating the position of each behavior history in the series to the learning data after the masking process.

[0050] Next, in step S20, the CPU 11, as the learning unit 24, learns a behavioral model 26 equipped with an attention mechanism using the training data after the masking process and after the assignment of the sequence position identifier. At this time, the CPU 11, as the learning unit 24, updates parameters so that the prediction result of the behavioral model 26 matches the correct answer, under the constraint that only the behavioral ID of the training data is input as the value constituting the attention mechanism, and all of the training data is input as the key and query. The CPU 11, as the learning unit 24, repeats updating the parameters until a learning termination condition is satisfied. The CPU 11, as the learning unit 24, stores the behavioral model 26, in which the parameters set when the termination condition is satisfied, in a predetermined storage area of ​​the advertisement delivery condition inference device 10, and the learning process ends.

[0051] Next, the inference process shown in Fig. 17 will be described. Here, the inference process will be described with reference to a specific example shown in Fig. 18. In Fig. 18, each block corresponds to one behavior history, and the symbol in each block indicates a behavior identifier corresponding to the behavior ID of that behavior history. In Fig. 18, the notation related to the behavior ID is expressed using the behavior identifier corresponding to that behavior ID.

[0052] In step S30, the CPU 11, functioning as the data acquisition unit 22, acquires inference data, which is a series of behavioral histories for each customer. Next, in step S32, the CPU 11, functioning as the data acquisition unit 22, converts the behavior identifiers in the inference data into behavior IDs based on the behavior ID dictionary generated in the learning process.

[0053] Next, in step S34, the CPU 11, as the data acquisition unit 22, sorts each behavior history included in the learning data in order of "date and time" for each customer identifier, and aligns the behavior history series to a specified length. Next, in step S36, the CPU 11, as the data acquisition unit 22, creates a behavior history for future date and time behavior history with the behavior ID corresponding to the behavior identifier [MASK] as the behavior ID and other items set to 0. In the example of FIG. 18, as shown in (1), for inference data with a series length of 8, the series length of the behavior history series up to the current date and time is set to 4, and the series length of the behavior history series for future date and time is set to 4. Next, in step S38, the CPU 11, as the data acquisition unit 22, assigns a series position identifier to the inference data.

[0054] Next, in step S40, the CPU 11, functioning as the prediction unit 30, inputs the inference data after the sequence position identifiers have been assigned to the behavior model 26, and predicts the probability value for each behavior ID at each date and time, as shown in FIG. 18(2).

[0055] Next, in step S42, the CPU 11, functioning as the goal acquisition unit 28, acquires a goal for distributing the advertisement. Next, in step S44, the CPU 11, functioning as the substitution unit 32, identifies as the triggering behavior a first behavior that satisfies the condition corresponding to the goal acquired in step S42 and has a probability value predicted in step S40 equal to or greater than a predetermined value. For example, assume that triggering behaviors are limited to behaviors identified by behavior identifiers {B, C, D, E} according to the goal. In this case, the CPU 11, functioning as the substitution unit 32, identifies, as the triggering behavior, behavior IDs for which the predicted probability value for each behavior ID corresponding to the behavior identifiers {B, C, D, E} is equal to or greater than a predetermined value at each future date and time. Note that if the goal of the advertisement is to promote frequent use, the search may be further narrowed down to behavior IDs indicating behaviors that have been performed with a frequency equal to or greater than a predetermined value in past and current performance. 18 shows a case where, based on the probability value shown in (2), the behavior ID corresponding to the behavior identifier "MASK" with the sequence position identifier=6 is replaced with the behavior ID corresponding to the behavior identifier "B" as shown in (3). Then, the CPU 11, as the replacement unit 32, causes the behavior model 26 to predict the probability value of each behavior ID at each date and time for the inference data after replacement with the behavior ID corresponding to the triggered behavior as shown in (4) of FIG.

[0056] Next, in step S46, the CPU 11, as the output unit 34, selects one of a plurality of predetermined second actions as the intervention action based on the purpose of the advertisement and the prediction result for the inference data after replacement by the replacement unit 32. Then, the CPU 11, as the output unit 34, causes the prediction unit 30 to predict the probability value of the triggering action for the inference data for which the intervention action has been selected. For example, assume that the intervention action is limited to actions identified by action identifiers {X, Y, Z} depending on the purpose. In this case, the CPU 11, as the output unit 34, selects, as the action ID of the intervention action, action IDs for which the predicted probability value for each of the action IDs corresponding to the action identifiers {X, Y, Z} is equal to or greater than a predetermined value at each future date and time before the triggering action. The example in FIG. 18 illustrates a case where, based on the probability value shown in (4), the action ID corresponding to the action identifier "X" is selected as the action ID with the sequence position identifier = 4, as shown in (5). 18(6), the CPU 11 as the output unit 34 returns the behavior ID replaced with the behavior ID corresponding to the behavior identifier "B" in (3) to the behavior ID corresponding to the behavior identifier "MASK." Then, the CPU 11 as the output unit 34 predicts the probability value of each behavior ID at each date and time using the behavior model 26 for the inference data in which the behavior ID corresponding to the intervention behavior has been selected.

[0057] Next, in step S48, the CPU 11, as the output unit 34, calculates the increase in the probability value of the triggering behavior before and after selecting the intervention behavior. In the example of FIG. 18, (2) is the probability value before the intervention, and (6) is the probability value after the intervention. That is, (2) is the probability value of each behavior ID "if the intervention behavior is not present," and (4) is the probability value of each behavior ID "if the intervention behavior is present." The CPU 11, as the output unit 34, uses the intervention behavior for which the increase in the probability value of the triggering behavior before and after the intervention is equal to or greater than a predetermined value and the corresponding date and time as delivery conditions, and outputs the increase as an index showing the effectiveness of delivering an advertisement based on the delivery conditions, and the inference process ends.

[0058] As described above, the advertising condition deduction device according to this embodiment acquires a purpose for delivering an advertisement. The advertising condition deduction device also acquires a series of behavioral histories for each customer as inference data. The behavioral history includes a first behavior related to at least one of the customer's awareness and use of a service, a second behavior related to the customer's receipt of a delivered advertisement, and the date and time at which each of the first and second behaviors occurred. The inference data is a series of behavioral histories from past dates and times to future dates and times, in which behavior at each future date and time is masked. The advertising condition deduction device also predicts a probability value that the masked behavior in the inference data is each of a plurality of predetermined behaviors using a behavior model that predicts behavior at each future date and time from behavior at each past date and time. The advertising condition deduction device then identifies, as a triggering behavior, a first behavior that satisfies a condition according to the acquired purpose and whose predicted probability value for the acquired inference data is equal to or greater than a predetermined value, and replaces the masked behavior at the date and time at which the triggering behavior was identified with the triggering behavior. The advertising condition deduction device also selects one of a plurality of predetermined second actions as an intervention action based on the advertising objective and the probability value of each action predicted for the replaced inference data. The advertising condition deduction device then outputs the intervention action for which the probability value of the triggering action predicted based on the inference data from which the intervention action was selected increases by a predetermined value or more relative to the probability value of the triggering action before the intervention action was selected, along with the corresponding date and time as the delivery condition and the increase amount as its effect. This makes it possible to personalize customer behavior-driven ad delivery according to the advertising objective in an environment where behavioral history can be tracked on a customer-by-customer basis. In other words, it is possible to combine the characteristics of both corporate objective-driven and customer behavior-driven ad delivery, increase CVR and delivery numbers, and improve advertising effectiveness.

[0059] More specifically, the advertisement delivery condition inference device according to this embodiment predicts the customer's future behavior "if there is no advertisement delivery" from a customer behavior starting point, and then classifies the behavior that satisfies the conditions in line with the company's objectives and has a high probability value as the triggering behavior. This makes it possible to determine the triggering behavior of an advertising goal that balances both customer behavior and company objectives. Furthermore, it is possible to predict the customer's future behavior "if there is a triggering behavior" from a customer behavior starting point, and to determine behavior related to advertisement receipt from a customer behavior starting point. Furthermore, it is possible to predict the customer's future behavior "if there is an behavior related to advertisement receipt" from a customer behavior starting point, and to predict advertisement delivery conditions and effects from a company's objective starting point.

[0060] Furthermore, the ad delivery condition inference device according to this embodiment imposes constraints on the input and output of the attention mechanism in the structure of the behavior model that predicts the probability value indicating the likelihood of each behavior. Specifically, only items that identify behavior are input to the value that constitutes the attention mechanism, and all items are input to the key and value. This prevents overlearning of supplementary information about the date and time and behavior, and further improves the accuracy of customer behavior-driven predictions.

[0061] In the above embodiment, the distribution condition inference device is configured to include a learning unit, but the learning device and the inference device may be implemented as separate computers. In this case, the learning device may include a part of the data acquisition unit that functions during the learning process and the learning unit, and the inference device may read the behavior model learned by the learning device and execute the inference process.

[0062] In addition, in the above embodiment, the specified length of a behavioral history sequence is 8, the length of the sequence up to the present is 4, and the length of the future sequence is 4; however, this is not limited to this. The interval for acquiring behavioral history, the length of the sequence, etc. may be determined according to the period of the ad distribution campaign, etc. For example, the length of the inference data sequence may be 512, of which 422 sequences may be the sequence up to the present and 90 sequences may be the future sequence. In this case, if the acquisition interval for behavioral history is 6 slots per day (4-hour intervals), 70.3 days of behavioral history may be acquired and used as the sequence up to the present, and 15 days of blank behavior history (where future behavior is masked) may be used as the future sequence.

[0063] Furthermore, the learning process and inference process executed by the CPU after reading the software (program) in the above embodiment may be executed by various processors other than the CPU. Examples of such processors include programmable logic devices (PLDs) such as field-programmable gate arrays (FPGAs), whose circuit configuration can be changed after fabrication, and dedicated electrical circuits such as application-specific integrated circuits (ASICs), which are processors with circuit configurations specifically designed to execute specific processes. The learning process and inference process may be executed by one of these various processors, or by a combination of two or more processors of the same or different types (e.g., multiple FPGAs, or a combination of a CPU and an FPGA). The hardware structure of these various processors is, more specifically, an electrical circuit that combines circuit elements such as semiconductor devices.

[0064] In the above embodiment, the advertisement delivery condition inference processing program is pre-stored (installed) in the ROM 12 or the storage 14, but the present invention is not limited to this. The program may be provided in a form stored in a non-transitory storage medium such as a CD-ROM (Compact Disk Read Only Memory), a DVD-ROM (Digital Versatile Disk Read Only Memory), or a USB (Universal Serial Bus) memory. The program may also be downloaded from an external device via a network.

[0065] The following additional notes are provided regarding the above-described embodiments.

[0066] (Additional note 1) Memory and at least one processor coupled to said memory; Including, The processor: Obtain the purpose of delivering advertisements, acquire, as inference data, a series of behavioral histories at each date and time including a first behavior related to at least one of a customer's awareness of and use of a service, and a second behavior related to the customer's receipt of a distributed advertisement, and the dates and times at which the first behavior and the second behavior were performed, in which behaviors at each date and time in the future than the present are masked; predicting a probability value that the masked behavior in the inference data is each of a plurality of predetermined behaviors using a behavior model that predicts behavior at each future date and time from behavior at each past date and time; Identifying a triggering behavior from the first behavior based on the conditions according to the acquired purpose and the probability value predicted for the acquired inference data, and replacing the masked behavior at the date and time when the triggering behavior was identified with the triggering behavior; An intervention action is selected from the second actions based on the conditions according to the acquired purpose and the probability value predicted for the inference data after replacement with the triggering action, and the intervention action for which the increase in the probability value of the triggering action predicted for the inference data from which the intervention action was selected is equal to or greater than a predetermined value relative to the probability value of the triggering action before selecting the intervention action, and the corresponding date and time are output as distribution conditions. The advertisement delivery condition inference device is configured as follows.

[0067] (Additional note 2) A non-transitory recording medium storing a program executable by a computer to execute an advertisement delivery condition inference process, The advertisement delivery condition inference process includes: Obtain the purpose of delivering advertisements, acquire, as inference data, a series of behavioral histories at each date and time including a first behavior related to at least one of a customer's awareness of and use of a service, and a second behavior related to the customer's receipt of a distributed advertisement, and the dates and times at which the first behavior and the second behavior were performed, in which behaviors at each date and time in the future than the present are masked; predicting a probability value that the masked behavior in the inference data is each of a plurality of predetermined behaviors using a behavior model that predicts behavior at each future date and time from behavior at each past date and time; Identifying a triggering behavior from the first behavior based on the conditions according to the acquired purpose and the probability value predicted for the acquired inference data, and replacing the masked behavior at the date and time when the triggering behavior was identified with the triggering behavior; An intervention action is selected from the second actions based on the conditions according to the acquired purpose and the probability value predicted for the inference data after replacement with the triggering action, and the intervention action for which the increase in the probability value of the triggering action predicted for the inference data from which the intervention action was selected is equal to or greater than a predetermined value relative to the probability value of the triggering action before selecting the intervention action, and the corresponding date and time are output as distribution conditions. Non-transitory recording media, including

[0068] (Additional note 3) Memory and at least one processor coupled to said memory; Including, The processor: acquiring, as learning data, a series of behavioral histories, in which a portion of the behavioral history is masked, the series including a first behavior related to at least one of a customer's awareness of and use of a service, a second behavior related to the customer's receipt of a distributed advertisement, and the date and time at which each of the first behavior and the second behavior was performed; Among the keys, queries, and values ​​constituting an attention mechanism provided in a behavioral model that predicts behavior at each future date and time from behavior at each past date and time, all of the learning data is input to the keys and queries, and only the first and second actions of the learning data are input to the values, and the behavioral model is trained by updating parameters of the behavioral model so that prediction results predicted by the behavioral model for masked actions in the learning data match correct actions. The advertisement delivery condition learning device is configured as follows.

[0069] (Additional note 4) A non-transitory recording medium storing a program executable by a computer to execute an advertisement delivery condition learning process, The advertisement delivery condition learning process includes: acquiring, as learning data, a series of behavioral histories, in which a portion of the behavioral history is masked, the series including a first behavior related to at least one of a customer's awareness of and use of a service, a second behavior related to the customer's receipt of a distributed advertisement, and the date and time at which each of the first behavior and the second behavior was performed; Among the keys, queries, and values ​​constituting an attention mechanism provided in a behavioral model that predicts behavior at each future date and time from behavior at each past date and time, all of the learning data is input to the keys and queries, and only the first and second actions of the learning data are input to the values, and the behavioral model is trained by updating parameters of the behavioral model so that prediction results predicted by the behavioral model for masked actions in the learning data match correct actions. Non-transitory recording media, including [Explanation of symbols]

[0070] 10. Advertisement delivery condition inference device 14. Storage 15 Input section 16 Display section 17 Communication I / F 19 Bus 22 Data Acquisition Section 24 Learning Department 26 Behavioral Model 28 Objective acquisition part 30 Prediction Department 32 Replacement part 34 Output section

Claims

1. a purpose acquisition unit that acquires a purpose for delivering an advertisement; a data acquisition unit that acquires, as inference data, a series of behavioral histories at each date and time, including a first behavior related to at least one of a customer's awareness of and use of a service, a second behavior related to the customer's receipt of a distributed advertisement, and the dates and times at which the first behavior and the second behavior were performed, in which behaviors at each date and time in the future than the present are masked; a prediction unit that predicts a probability value that the masked behavior in the inference data is each of a plurality of predetermined behaviors using a behavior model that predicts behavior at each future date and time from behavior at each past date and time; a replacement unit that identifies a triggering behavior from the first behavior based on a condition according to the purpose acquired by the purpose acquisition unit and a probability value predicted by the prediction unit for the inference data acquired by the data acquisition unit, and replaces the masked behavior at the date and time when the triggering behavior was identified with the triggering behavior; and an output unit that selects an intervention action from the second actions based on a condition according to the purpose acquired by the purpose acquisition unit and a probability value predicted by the prediction unit for the inference data after replacement with the triggering action, and outputs, as distribution conditions, the intervention action and a corresponding date and time, for which an increase in the probability value of the triggering action predicted by the prediction unit for the inference data from which the intervention action has been selected is equal to or greater than a predetermined value with respect to the probability value of the triggering action before the selection of the intervention action; An advertisement delivery condition inference device including:

2. The advertisement distribution condition inference device according to claim 1 , wherein the output unit outputs the increase amount as an index indicating an effect when the advertisement is distributed based on the distribution condition.

3. If the purpose is to increase awareness of the service, the conditions include conditions that are limited to behavior related to customer awareness of the service; If the purpose is to promote new use of a service, the conditions include a condition that limits the customer's behavior regarding use of a service that the customer has not used before, If the purpose is to continue promoting a service, the conditions include conditions that are limited to the customer's behavior regarding the use of a service that the customer has used, If the purpose is to promote cross-service usage, the condition includes a condition limiting the customer's behavior to using two or more services on the same date and time; If the purpose is to promote frequent use, the conditions include a condition that limits the behavior of the customer to the use of a service that is frequently used by the customer.

3. The advertisement delivery condition inference device according to claim 1.

4. The advertisement delivery condition inference device according to claim 3 , wherein when the purpose is to promote frequent use, the conditions include a narrowing condition relating to the track record or frequency of past use of a service that is frequently used by the customer.

5. further comprising a learning unit configured to learn the behavioral model having an attention mechanism; the data acquisition unit acquires, as learning data, a series of behavioral histories, which are behavioral histories at each date and time including the first behavior, the second behavior, and dates and times when each of the first behavior and the second behavior was performed, and in which a part of the behavioral histories is masked; the learning unit inputs all of the learning data acquired by the data acquisition unit into the key and the query among the key, query, and value constituting the attention mechanism of the behavioral model, and inputs only the first behavior and the second behavior of the learning data into the value, and updates parameters of the behavioral model so that a prediction result predicted by the behavioral model for the behavior masked in the learning data matches a correct behavior.

3. The advertisement delivery condition inference device according to claim 1.

6. The purpose acquisition unit acquires a purpose for delivering the advertisement, a data acquisition unit acquires, as inference data, a series of behavioral histories at each date and time, including a first behavior related to at least one of a customer's awareness of and use of a service, a second behavior related to the customer's receipt of a distributed advertisement, and a date and time when each of the first behavior and the second behavior was performed, in which behaviors at each date and time in the future than the present are masked; a prediction unit predicts a probability value that the masked behavior in the inference data is each of a plurality of predetermined behaviors using a behavior model that predicts behavior at each future date and time from behavior at each past date and time; a replacement unit, based on a condition according to the purpose acquired by the purpose acquisition unit and a probability value predicted by the prediction unit for the inference data acquired by the data acquisition unit, identifies a triggering behavior from the first behavior, and replaces the masked behavior at the date and time when the triggering behavior was identified with the triggering behavior; an output unit selects an intervention action from the second actions based on the conditions according to the purpose acquired by the purpose acquisition unit and the probability value predicted by the prediction unit for the inference data after replacement with the triggering action, and outputs, as distribution conditions, the intervention action for which the increase in the probability value of the triggering action predicted by the prediction unit for the inference data from which the intervention action has been selected is equal to or greater than a predetermined value with respect to the probability value of the triggering action before the intervention action was selected, and the corresponding date and time. A method for inferring ad delivery conditions.

7. 3. An advertisement delivery condition inference program for causing a computer to function as each unit of the advertisement delivery condition inference device according to claim 1.

8. a data acquisition unit that acquires, as learning data, a series of behavioral histories, including a first behavior related to at least one of a customer's awareness of and use of a service, a second behavior related to the customer's receipt of a distributed advertisement, and the dates and times when each of the first behavior and the second behavior was performed, with a portion of the behavioral history masked; a learning unit that learns the behavioral model by inputting all of the learning data acquired by the data acquisition unit into the key and the query, and inputting only the first behavior and the second behavior of the learning data into the value, of the keys, queries, and values ​​that constitute an attention mechanism included in the behavioral model that predicts behavior at each future date and time from behavior at each past date and time, and updating parameters of the behavioral model so that a prediction result predicted by the behavioral model for the masked behavior in the learning data matches a correct behavior; An advertisement delivery condition learning device including:

Images