Information processing device, information processing method, and program

Abstract

To estimate a position bias in consideration of the influence of a bias even when the bias occurs in an item arrangement position.SOLUTION: An information processing device acquires an arrangement provability representing the provability that each of n items (n is a natural number of 2 or higher) is arranged in k positions (k is a natural number of 2 or higher), converts the n items to m embedded vectors (m is a natural number of 2 or higher) representing feature abstraction representations of the n items, calculates an allocation provability representing the provability of allocation to the m embedded vectors from the n items, and derives a provability representation where each of the m embedded vectors is arranged in each of the k positions using the arrangement probability distribution and the allocation provability distribution.SELECTED DRAWING: Figure 5

Description

[Technical field]

[0001] The present invention relates to an information processing device, an information processing method, and a program. [Background technology]

[0002] In recent years, rankings of various items such as products and search results are widely provided in web services used by users. Click data, such as a user's click history, is used to construct such rankings. Click data can be called implicit feedback because it is generated by user actions, not by users' explicit statements or ratings. That is, click data can reflect how users actually behave when interacting with a system or website, rather than what they like or dislike. In this respect, click data provides rich feedback implicitly, and therefore is used to improve personalized rankings.

[0003] When a user selects and clicks an item from a number of items displayed on a screen, the user checks the item before clicking, and the position of the item may affect the checking and clicking. Such a bias in the user's checking depending on the item's position is called position bias. Non-Patent Document 1 discloses a technology for estimating position bias according to a regression-type EM (Expectation-Maximization) algorithm based on the position where the item is placed and the click history of the item. [Prior art documents] [Non-patent literature]

[0004] [Non-Patent Document 1] Xuanhui Wang, et al., “Position Bias Estimation for Unbiased Learning to Rank in Personal Search”, Proceedings of the 11th ACM International Conference on Web Search and Data Mining (WSDM), ACM (2018), pp. 610-618 Summary of the Invention [Problem to be solved by the invention]

[0005] In a screen where a plurality of items are displayed, each of the plurality of items is often placed in a fixed position. For example, in a carousel advertisement, the placement and order of each item is predetermined by the creator of the advertisement. In this manner, in a situation where each of the plurality of items is placed in a fixed position, a bias occurs in the positions where the items are placed. In other words, a bias occurs in the positions where the items are actually placed, relative to the plurality of positions where the items can be placed. The above document does not disclose a mechanism for estimating a position bias by taking into account the influence of the bias in item placement positions when such a bias occurs in the placement positions of items.

[0006] In view of the above-mentioned problems, the present disclosure aims to establish an algorithm for estimating a positional bias while taking into account the influence of bias in the placement positions of items, even when such bias occurs. [Means for solving the problem]

[0007] An information processing device according to one embodiment of the present disclosure includes an acquisition unit that acquires placement probabilities representing the probability that each of n items (n is a natural number of 2 or greater) will be placed at k positions (k is a natural number of 2 or greater); a conversion unit that converts the n items into m (m is a natural number of 2 or greater) embedding vectors that represent abstract representations of features of the n items; a calculation unit that calculates allocation probabilities representing the probability of allocation from the n items to the m embedding vectors; and a derivation unit that derives a probability representation that each of the m embedding vectors will be placed at each of the k positions using a distribution of the placement probabilities and a distribution of the allocation probabilities.

[0008] An information processing method according to one embodiment of the present disclosure includes obtaining placement probabilities representing the probability that each of n items (n is a natural number greater than or equal to 2) is placed at k positions (k is a natural number greater than or equal to 2), converting the n items into m embedding vectors (d is a natural number greater than or equal to 2) representing abstract representations of features of the n items, calculating allocation probabilities representing the probability of allocation from the n items to the m embedding vectors, and deriving a probability representation that each of the m embedding vectors is placed at each of the k positions using a distribution of the placement probabilities and a distribution of the allocation probabilities.

[0009] A program according to one aspect of the present disclosure is a program for causing a computer to execute an information processing method, the information processing method including: obtaining placement probabilities representing the probability that each of n items (n is a natural number greater than or equal to 2) is placed at k positions (k is a natural number greater than or equal to 2); converting the n items into m embedding vectors (m is a natural number greater than or equal to 2) representing abstract representations of features of the n items; calculating allocation probabilities representing the probability of allocation from the n items to the m embedding vectors; and deriving a probability representation that each of the m embedding vectors is placed at each of the k positions using a distribution of the placement probabilities and a distribution of the allocation probabilities. Effect of the Invention

[0010] According to the present invention, an algorithm is provided for estimating a position bias by taking into account the influence of bias in the placement positions of items. [Brief description of the drawings]

[0011] [Figure 1] FIG. 1 shows an example of the configuration of an information processing system. [Figure 2A] FIG. 2A illustrates an example distribution of placement probabilities π(i,k) according to one embodiment. [Figure 2B] FIG. 2B illustrates an example distribution of allocation probabilities P(e|i) according to one embodiment. [Figure 2C] FIG. 2C illustrates an example distribution of placement probabilities π(e,k) according to one embodiment. [Diagram 3] FIG. 3 is a diagram for explaining generation of an embedding vector. [Figure 4] FIG. 4 shows the algorithm according to one embodiment. [Diagram 5] FIG. 5 illustrates an example of a functional configuration of an information processing device. [Figure 6] FIG. 6 shows an example of content with multiple items arranged in multiple locations. [Figure 7] FIG. 7 illustrates an example of a hardware configuration of an information processing device according to an embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0012] Hereinafter, with reference to the attached drawings, an embodiment for carrying out the present invention will be described in detail. Among the components disclosed below, those having the same functions are given the same reference numerals, and their description will be omitted. Note that the embodiment disclosed below is an example of a means for realizing the present invention, and should be appropriately modified or changed depending on the configuration of the device to which the present invention is applied and various conditions, and the present invention is not limited to the following embodiment. Furthermore, not all of the combinations of features described in the present embodiment are necessarily essential to the solution of the present invention.

[0013] In this embodiment, a probabilistic model is used that is an improvement of the position-based click model that represents a probabilistic model of the probability that a user clicks on an item, as disclosed in Non-Patent Document 1. First, the conventional position-based click model disclosed in the document is described, and then a configuration for estimating a position bias according to a regression-type EM algorithm based on a model that is an improvement of the conventional position-based click model is described. Note that in this disclosure, the term "probability" may be understood as the rate at which a target event occurs, rather than referring to an unknown possibility.

[0014] [Traditional location-based click model] We will now explain the location-based click model disclosed in Non-Patent Document 1. Here, we assume that a user clicks on any item from one or more recommended items (e.g., advertisements or products) displayed (placed) on a display screen in association with a web service. Item i is the recommended item. Variable C is the reward variable. If clicks are the subject of reward, then variable C = "1" indicates that the displayed item was clicked, and variable C = "0" indicates that the displayed item was not clicked. User u is a user having a user context including one or more user attributes (information about the user device or the user himself / herself) specific to user u. Position k is the position at which the item is displayed among multiple displayable (placeable) positions.

[0015] In the location-based click model, the click probability P(C=1|i,u,k) conditional on item i, user u, and location k is expressed as the multiplication of two latent probabilities as shown in Equation (1).

number

[0016] In the following explanation, the two probabilities on the right side of equation (1) are the item-user relevance μ(i,u) and the location bias θ k That is, μ(i,u)=P(R=1|i,u) and θ k = P(E = 1 | k). In Non-Patent Document 1, the relevance μ(i, u) and the position bias θ k is estimated using the regression EM algorithm. In the regression EM algorithm, the probability included in the probability model (i.e., the relevance μ(i,u) and the position bias θ k ) is estimated using maximum likelihood.

[0017] Although there may be a plurality of positions where an item can be placed on a display screen, generally, items are often placed at fixed positions based on the experience and expertise of advertisement creators and marketers. That is, referring to formula (1), among the multiple positions that position k can take, the positions at which item i is placed are often not all of the multiple positions. In such a case where a bias occurs in the actual placement position of an item with respect to the multiple positions at which the item can be placed, the conventional position-based click model cannot accurately express the position bias, and an accurate position bias cannot be estimated using the regression-type EM algorithm.

[0018] In this embodiment, a plurality of items are converted into a plurality of embedding vectors, and an improved location-based click model is defined based on the plurality of embedding vectors. Then, the relevance between items and users and the location bias expressed by the improved location-based click model are estimated using a regression EM algorithm.

[0019] [Information processing system configuration] Fig. 1 shows an example of the configuration of an information processing system 1 in this embodiment. As an example, the information processing system 1 includes an information processing device 10 and a user device 11 used by a user, as shown in Fig. 1. The information processing device 10 and the user device 11 are configured to be able to communicate with each other. Note that in the following description, although one user device 11 is shown in Fig. 1, multiple user devices may be configured to be able to communicate with the information processing device 10. Also, the term user device 11 may be understood to mean any one of the multiple user devices.

[0020] The user device 11 is a device having a display unit, such as a smartphone, a tablet terminal, or a smart TV. The user device 11 is configured to be able to communicate with the information processing device 10 via a public network such as 5G (fifth generation mobile communication system) or a wireless communication network such as a wireless LAN (Local Area Network). When the user device 11 is a device having a GUI (Graphic User Interface) equipped on a display unit such as a liquid crystal display, such as a smartphone or a tablet terminal, the user can perform various operations using the GUI. The operations include various operations on content such as images displayed on a display screen, such as a tap operation, a slide operation, and a scroll operation using a finger or a stylus.

[0021] Note that user device 11 is not limited to the device shown in Fig. 1, and may be a device such as a desktop PC (Personal Computer) or a notebook PC. In this case, each user may perform operations using input devices such as a mouse and a keyboard. User device 11 may also be provided with a separate display unit.

[0022] The information processing device 10 may be a server device that provides an electronic commerce platform such as a marketplace, and the user device 11 can use web services (internet-related services) provided by the information processing device 10. Note that the information processing device 10 is not limited to being the above-mentioned server device, and the user device 11 may be configured to use web services provided by a server device (not shown) separate from the information processing device 10 via the information processing device 10.

[0023] The user can click (select) any item among a plurality of items (e.g., advertisements or products) provided by the web service and displayed on the display screen of user device 11. If the clicked item is a product, after the click, a description of the product and information for the purchase procedure may be displayed on the display unit of user device 11. Also, if the clicked item is an advertisement, specific contents of the advertisement may be displayed on the display unit of user device 11.

[0024] The information processing device 10 is configured to provide a web service to the user device 11, observe the user's behavior in the web service, and receive a report thereof. For example, the information processing device 10 can obtain observation data reflecting the user's behavior history in the web service by enabling reception of a report indicating the user's behavior history. In this embodiment, the information processing device 10 observes a user's behavior history including a click action by the user on any of a plurality of items displayed on the display unit of the user device 11. The observation data acquired by the information processing device 10 includes at least one of a user context, information on the location of the item, item information of the item, and click information.

[0025] The user context includes one or more user attributes (information about the user device 11 or the user himself / herself) specific to the user. The user attributes include, for example, the user's name, the user's address, product delivery destination information, credit card information held by the user, and user demographic information. The demographic information is information indicating demographic user attributes such as gender, age, residential area, occupation, and family structure. The user attributes can be registered, for example, by the user in order to use a web service. In addition to this, or instead of this, the information processing device 10 can acquire the user attributes by analyzing the web pages viewed by the user, the places clicked in the past, and the like.

[0026] The information on the location of the item includes information on the location of the item clicked by the user on the display unit of the user device 11 . The item information includes information for identifying the item and may also include one or more item characteristics, such as color or size. The click information is information indicating whether or not a click has been made.

[0027] The information processing device 10 generates a data set based on such observation data and predetermined configuration information related to the display of items. The configuration information includes information on a plurality of items (including item features) displayed on the display screen of the user device and information on all positions (a plurality of positions) where the items can be placed. Then, the information processing device 10 converts a plurality of items included in the data set into a plurality of embedding vectors, and defines a position bias according to an improved position-based click model using the embedding vectors. The improved position-based click model corresponds to a model obtained by improving the conventional position-based click model of the above formula (1). Then, the information processing device 10 estimates the position bias using a regression type EM algorithm. The regression type EM algorithm in this embodiment corresponds to an improvement of the regression type EM algorithm disclosed in Non-Patent Document 1. In the following description, the former is also referred to as an improved regression type EM algorithm, and the latter is also referred to as a conventional regression type EM algorithm. The improved position-based click model and the improved regression EM algorithm will be described below along with the operation of the information processing device 10.

[0028] [Operation of information processing device] The information processing device 10 acquires observation data n times (n is a natural number equal to or greater than 2) for a web service provided to the user device 11, and generates a data set D based on the observation data and predetermined configuration information. One observation may be an observation during a certain period of time. Note that in this embodiment, it is assumed that the information processing device 10 generates the data set D based on the observation data and predetermined configuration information. However, the original data for generating the data set D is not limited to these, and other data may be used as long as the data set D can be generated by observing a user's clicking operation.

[0029] The generated data set D is expressed as the following equation (2).

number

[0030] Let A={(i,k)} be the set of actions including all possible (i,k) pairs, and π be the function that maps user u to the distribution of the actions, in other words, the policy (rule) for placing item i at position k. In this embodiment, the probability of placing (assigning) item i to position k (k∈K) (hereinafter also referred to as placement probability) is represented as π(i,k). In many cases, the policy is determined by a marketing person based on experience and expertise. As a result, the distribution of the placement probability π(i, k) is mostly deterministic and static. In this embodiment, the data set D is a data set that follows such a conventional policy, and the types of (i, k) pairs in the data set D are limited for all possible (i, k) pairs.

[0031] When the information processing device 10 acquires the data set D, it calculates the placement probability π(i, k) for the data set D. The placement probability π(i, k) for the data set D represents the probability (proportion) with which an item i included in the data set D is placed at each position K among positions at which a plurality of items can be placed. As described above, in this embodiment, the types of pairs (i, k) in the data set D are limited. In order to quantify the degree of sparseness of such pairs (i, k), the information processing device 10 calculates an index (hereinafter also referred to as an arrangement distribution index) indicating the bias of the distribution of the arrangement probability π(i, k).

[0032] As a first example, if the placement distribution index is expressed as the sparsity ratio J, the index can be expressed as in equation (3). The sparsity ratio represents the proportion of non-missing values ​​in the placement distribution (for example, a missing value refers to (i, k) that is missing in the data set D).

number

[0033] As a second example, the placement distribution index can be expressed as the similarity of the distribution of placement probability to a uniform distribution in which n items i are uniformly placed at all possible positions k. One example of the similarity is the Kullback-Leibler divergence D KL and the index can be expressed as in equation (4).

number

[0034] The sparse ratio J in equation (3) indicates the percentage of (i, k) that is actually observed. Therefore, the smaller the sparse ratio J, the lower the percentage of (i, k) that is actually observed, and the more sparse the item placement is with respect to all possible placement positions. On the other hand, the Kullback-Leibler divergence D in equation (4) KL The larger the distribution distribution index, the greater the deviation (dissimilarity) from the uniform distribution, indicating that there is a bias in the arrangement positions of the items. By calculating such an arrangement distribution index, the information processing device 10 can grasp the degree of bias in the arrangement positions of the items in the data set D.

[0035] FIG. 2A shows an example of the distribution of placement probability π(i, k) for dataset D. Specifically, FIG. 2A shows placement probability π(i, k) which indicates the probability (proportion) with which an item is placed at each position when all items are items i0, i1, and i2, and all positions where the items can be placed are k0, k1, and k2. b (i, k0), π b (i, k1), π b (i, k2), as an example. In the table of FIG. 2A, π biased (i, k) with placement probability π b It is represented as (i,k). As shown in the table, in Dataset D, item i0 is fixedly placed at position k0, item i1 is fixedly placed at position k1, and item i2 is fixedly placed at position k2. In other words, in Dataset D, it can be seen that the pairs of items and placement positions are biased toward (i0, k0), (i1, k1), and (i2, k2).

[0036] In the case of the distribution of the arrangement probability π(i, k) for the data set D shown in FIG. 2A, the sparse ratio J for the data set D can be calculated as shown in equation (5) according to equation (3).

number

[0037] In addition, according to equation (4), the Kullback-Leibler divergence D for the data set D is KL can be calculated as shown in equation (6).

number

[0038] As described above, in the data set D, bias occurs for all possible (i, k) pairs for the (i, k). That is, the (i, k) pairs in the data set D have bias and sparseness problems. When estimating the position bias using such a data set D and formula (1) according to the conventional technology, the estimation may not be performed correctly due to the bias and sparseness problems of the (i, k) pairs included in the data set D.

[0039] To address this issue, the information processing device 10 generates multiple embedding vectors from multiple items in the dataset D. Specifically, the information processing device 10 generates embedding vectors from n items i j From (j = 1 to n), we generate m embedding vectors e (e ∈ E, where E is a set of embedding vectors). The embedding vector e corresponds to a vector that represents latent contexts that abstract multiple item features.

[0040] The embedding vector e will be specifically described with reference to Fig. 3. Fig. 3 is a diagram for explaining generation of the embedding vector according to this embodiment. In this embodiment, there are n items i, and each item is associated with a item features (a is a natural number equal to or greater than 2). The item features are features for identifying an item, such as color or size. First, the information processing device 10 prepares an n×a matrix (item×item feature) from the n items i included in the dataset D. Each column of the n×a matrix corresponds to a feature vector, and each feature vector represents an item feature of each of the n items.

[0041] The information processing device 10 converts (maps) an n×a matrix (item×item feature) into an n×m matrix (item×latent context). This conversion can be performed by a known feature expression extraction technique such as LSI (Latest Semantic Indexing) or VAE (Variational Auto-Encoder). Each column of the converted n×m matrix corresponds to an embedding vector, and each embedding vector represents the latent context of each of the n items. This generates m embedding vectors. The embedding vector corresponds to a vector that abstracts the a item features, and can also be called an abstracted feature vector.

[0042] By using LSI or VAE to convert an n×a matrix into an n×m matrix, the dimension is reduced, i.e., m becomes smaller than a. For example, in LSI, the dimension can be reduced by grouping items with similar meanings in the item features of multiple items. Also, in VAE, the dimension can be reduced by compression processing. In this way, the size of the n×m matrix becomes smaller than the size of n×a, and the sparse (i,k) pairs in dataset D are converted into denser (e,k) pairs.

[0043] The probability of allocation from item i to embedding vector e (probability of embedding vector e conditioned on item i) is represented by allocation probability P(e|i). FIG. 2B shows an example of the distribution of the allocation probability P(e|i) for dataset D. Specifically, FIG. 2B shows an example of the distribution of the allocation probability P(e|i) for all items: items i0, i1, and i2, and the number of embedding vectors m=2. In this embodiment, P(e|i) is calculated for each item i so that the total of the allocation probabilities P(e|i) (the sum of the conditional probabilities of allocating item i to embedding vector e) is 1. That is, Σ e∈E(e|i)=1. In FIG. 2B, for example, the allocation probability P(e0|i0) from position i0 to embedding vector e0 is 1 / 2, and the allocation probability P(e1|i0) from position i0 to embedding vector e1 is 1 / 2, so P(e0|i0)+P(e1|i0)=1. For positions i1 and i2, P(e0|i1)+P(e1|i1)=1 / 3+2 / 3=1 and P(e0|i2)+P(e1|i2)=1 / 4+3 / 4=1, respectively.

[0044] The association between items and users shown on the right side of equation (1) can be expressed as equation (7) using the allocation probability P(e|i) and the embedding vector e.

number

number

[0045] As described above, the information processing device 10 generates the embedding vector e and samples the reward w, and from the data set D, a data set D including the embedding vector e and the reward w as shown in the following formula (9) is obtained. e can be generated.

number

[0046] Moreover, the conventional position-based click model shown in equation (1) can be expressed as an improved position-based click model with an embedding vector, as shown in equation (10), by using the embedding vector e.

number

[0047] In the following explanation, the two probabilities on the right side of equation (10) are the relevance μ(e,u) between the embedding vector and the user, and the position bias θ ek That is, μ(e,u)=P(R=1|e,u) and θ ek =P(E=1|k). The position bias θ shown in Eq. (1) k In order to distinguish from the position bias θ ek represents the probability that user u recognizes each of the k positions where the m embedding vectors are placed.

[0048] As described above, in this embodiment, for each item i, allocation probability P(e|i) is calculated so that the sum of P(e|i) is 1, and P(e|i) is defined as shown in the following equation (11).

number

[0049] If two items i and i' are similar, their corresponding embedding spatial representation e i,l and e i’,l We assume that the distributions of the allocation probabilities P(e|i) and P(e|i') are similar. Based on this assumption, even when observing a specific (item, position) pair such as (i0, k0) and (i1, k1) in dataset D, we can obtain the assignment probabilities P(e0|i0), P(e1|i0), P(e0|i1), and P(e1|i1). eBy using this method, data for both positions of embedding vectors e0 and e1 (i.e., (e0, k0) and (e0, k1) in the case of embedding vector e0) can be used, which solves the problems of bias and sparseness described above.

[0050] The information processing device 10 derives an arrangement probability π(e, k) representing a probabilistic expression at which an embedding vector e is arranged, using the distribution of the arrangement probability π(i, k) and the distribution of the allocation probability P(e|i) for each position k in the data set D. In this embodiment, the arrangement probability π(e, k) is derived (calculated) using the sum of the multiplication of the allocation probability P(e|i) and the arrangement probability π(i, k). FIG. 2C shows an example of the placement probability π(e, k) in the case of the allocation probability P(e|i) shown in FIG. 2B. The placement probability π(e, k) is calculated by dividing the allocation probability P(e|i) shown in FIG. 2B by the placement probability π b Dataset D based on (i,k) e We show the placement probability π(e,k) for

[0051] The placement probability π(e p ,k q ) is the dataset D e For all items i in p |i) and placement probability π b (i,k q ) and the sum of the multiplications. For example, in FIG. 2C, the allocation probability π(e0, k0) can be calculated as in the following equation (12).

number

[0052] As shown in FIG. 2C, the arrangement probability π(e, k) is calculated by multiplying the arrangement probability π b Compared to (i, k), we can see that the problems of data bias and sparseness have been resolved. Therefore, the dataset D containing the embedding vector e and position k ise It is expected that the position bias can be estimated more accurately by using

[0053] Next, the information processing device 10 calculates the data set D e The position bias is estimated using the regression EM algorithm described in Non-Patent Document 1. Conventionally, the position bias shown in the conventional position-based click model shown in Equation (1) is estimated using the regression EM algorithm described in Non-Patent Document 1. In other words, the recursive RM algorithm is applied to repeat the expectation step and the maximization step to estimate the relevance μ(i,u) and the position bias θ k was optimized.

[0054] In this embodiment, as an improved recursive RM algorithm, first, in the expectation step, in the t+1 iteration for a certain time t, the distribution of the hidden variables E and R is calculated by θ ek (t) and μ (t) Estimate from (e,u). θ ek (t) and μ (t) (e, u) are the position bias θ at time t, as explained with reference to Equation (10). ek and the relevance μ(e,u) between the embedding vector and the user.

number

[0055] From equation (13), the data set D e For all data points, we can calculate the probability P(E=1|u,e,w,k) and probability P(R=1|u,e,w,k). The probability P(E=1|u,e,w,k) represents the probability that an association exists (=1) between user u and embedding vector e, given user u, embedding vector e, reward w, and position k. The probability P(E=1|u,e,w,k) represents the probability that position k is recognized by user u, given user u, embedding vector e, reward w, and position k.

[0056] In the maximization step, we use the probabilities from the expectation step to find ek (t+1) and μ (t+1) Calculate (e,k).

number

[0057] The improved regression EM algorithm according to this embodiment, based on equations (13) and (14), is shown in Fig. 4. The processing steps of this algorithm will be described below in order. Action 1 As input, we take a dataset D that contains a user (user context) u, an item i, a click c, and a position k, and a position bias θ ek , receive the embedding vector and the user association: μ(e,u), and the assignment probability P(e|i). θ ek may be a predetermined initial value, and μ(e,u) may be an empty regression model. Process 2 to 4 For every user u, item i, click c, and location k in the dataset D, we sample reward w from click c with allocation probability P(e|i). Specifically, we sample reward w for w∈{0, 1}, i.e., 0 or 1, according to (8). Action 5 From the set of rewards w sampled in process 4 and dataset D, a dataset D containing user (user context) u, item i, reward w, and position k is created. e Prepare (generate) the following. Process 6 to 14 Steps 7 to 13 are repeated until the condition of step 14 is met (repeated from time t to time t+1). Action 7 Let set S be the empty set. Steps 8 to 11 Dataset D e For all users u, embedding vectors e, clicks c, and positions k in, we sample a relationship r that takes r ∈ {0, 1}, i.e., 0 or 1, from the probability P(R = 1 |u, e, w, k) based on equation (13). Next, we generate a union S of the users u, embedding vectors e, relationship r, and set S. Action 12 Taking μ(e,u) and S as input, update μ(e,u) according to the Gradient Boosted Decision Tree (GBDT). Because the relevance between items and users can be nonlinear, we use the GBDT method to learn μ(e,u). Action 13 According to equation (14), θ ek Update. Action 14 Updated θ at time t and time t+1 ek If the difference between the values ​​is less than a predetermined value, it is determined that the convergence condition is satisfied and the process ends. -3 Here, the updated θ at time t and time t+1 ek In addition to the value of μ(e,u), if the difference between the updated μ(e,u) at time t and time t+1 is equal to or less than a predetermined value, it may be determined that the convergence condition is satisfied and the process may be terminated. Action 15 θ k and returns μ(e,u).

[0058] In this way, in this embodiment, even if the data set D has a limited number of pairs (i, k) indicating the arrangement positions of items, the item i is converted into an embedding vector e, and a data set D including the (e, k) pair is created. eThen, the position bias θ is calculated according to the improved recursive EM algorithm shown in FIG. ek is estimated. This allows the position bias θ k The position bias θ ek It becomes possible to estimate the following.

[0059] The information processing device 10 calculates the position bias θ according to the improved recursive EM algorithm shown in FIG. ek When the estimated position bias θ ek Based on this, the arrangement position of the item for each user may be adjusted. Specifically, first, the information processing device 10 estimates the relevance between the item and the user μ(i, u)=P(R=1|i, u) in the position-based click model shown in formula (1) using a conventional regression EM algorithm, and identifies one or more items highly related to the user u. In addition to this, or instead of this, the information processing device 10 may identify one or more items in order of relevance to the user u.

[0060] Next, the information processing device 10 calculates the position bias θ in the improved position-based click model shown in equation (10). ek =P(E=1|k) is estimated using the improved regression EM algorithm shown in FIG. ek Based on this, a position that is likely to be recognized by the user u is identified, and the identified one or more items are arranged (matched) at the identified position. This increases the possibility that the user u will click on the one or more items, i.e., the CTR (click-through rate) is improved, and the advertising effect can be increased.

[0061] [Functional configuration of information processing device] 5 shows an example of the configuration of an information processing device 10 according to this embodiment. The information processing device 10 includes a first data set generation unit 101, a bias calculation unit 102, an embedding vector generation unit 103, an allocation probability calculation unit 104, a second data set generation unit 105, a probability expression derivation unit 106, a relevance estimation unit 107, a positional bias estimation unit 108, a content creation unit 109, and a content provision unit 110.

[0062] The first data set generation unit 101 generates a data set D reflecting a user's behavior history in a web service. For example, the first data set generation unit 101 generates the data set D based on observation data of the user's behavior history in a web service and predetermined configuration information related to the display of items. As described above, the data set D includes a user u, an item i, a click c, and a position k in each of the 1st to nth observations. The item i is associated with a number of item features. Moreover, the first data set generation unit 101 calculates and acquires the arrangement probability π(i, k) for the data set D. An example of the distribution of the arrangement probability π(i, k) for the data set D is shown in FIG. 2A.

[0063] The bias calculation unit 102 calculates a layout distribution index indicating an index showing the bias of the distribution of the layout probability π(i, k) calculated by the first data set generation unit 101. In this embodiment, as described above, the sparse ratio J defined by the formula (3) and the Kullback-Leibler divergence D defined by the formula (4) are calculated. KL The smaller the sparse ratio J, the more biased the item placement. On the other hand, the Kullback-Leibler divergence D KL The larger the value, the more bias there is in the placement of the items.

[0064] The embedding vector generation unit 103 converts n items i included in the dataset D into m embedding vectors e that represent abstract expressions of item features of the n items i. In this embodiment, as described with reference to FIG. 3, the embedding vector generation unit 103 prepares an n×a matrix (item×item feature) from the n items i included in the dataset D. Then, the embedding vector generation unit 103 converts (maps) the matrix into an n×m (item×latent context) matrix. The embedding vector generation unit 103 generates each column of the converted n×m matrix as an embedding vector e.

[0065] The allocation probability calculation unit 104 calculates an allocation probability P(e|i) that represents the probability of allocation from n items i included in the dataset D to m embedding vectors e (the probability of embedding vector e conditional on item i). An example of the distribution of the allocation probability P(e|i) for the dataset D is shown in FIG. 2B. The allocation probability calculation unit 104 can calculate P(e|i) for each item i such that the sum of the allocation probabilities P(e|i) is 1.

[0066] The second data set generator 105 generates the data set D e Specifically, as described above, the second dataset generation unit 105 samples a reward w that takes the value of 0 or 1 according to equation (8) from a click c with an allocation probability P(e|i) calculated by the allocation probability calculation unit 104. Next, the second dataset generation unit 105 generates a dataset D e Generate.

[0067] The probability expression derivation unit 106 derives an arrangement probability π(e, k) representing a probability expression at which the embedding vector e is arranged, using the distribution of the arrangement probability π(i, k) and the distribution of the allocation probability P(e|i) for each position k in the data set D. The probability expression derivation unit 106 can derive (calculate) the arrangement probability π(e, k) using the sum of multiplications of the allocation probability P(e|i) and the arrangement probability π(i, k). An example of the distribution of the probability expression is shown in FIG. 2C.

[0068] The relevance estimation unit 107 estimates the relevance μ(i, u) between an item i and a user u in the data set D. Specifically, the relevance estimation unit 107 estimates the relevance μ(i, u) shown in equation (1) by using a conventional regression type EM algorithm.

[0069] The position bias estimation unit 108 estimates the data set D e Position bias θ ek Specifically, the position bias estimation unit 108 estimates the position bias θ ek is estimated using the improved recursive EM algorithm shown in FIG. Alternatively, or in addition, the position bias estimator 108 may estimate a position bias θ k Specifically, the position bias estimation unit 108 may estimate the position bias θ k is estimated using the conventional regression-type EM algorithm.

[0070] The position bias estimation unit 108 may switch between estimating the position bias using the conventional regression EM algorithm and the improved regression EM algorithm, based on the placement distribution index calculated by the bias calculation unit 102. For example, when the sparse ratio J is used as the placement distribution index, the position bias estimation unit 108 estimates the position bias θ ekcan be estimated by using the improved regression EM algorithm shown in Fig. 4. On the other hand, when the sparseness ratio J is greater than the predetermined value (the number of missing values ​​in the placement distribution is less than a predetermined level), the position bias θ k can be estimated using a conventional recursive EM algorithm. As another example, the Kullback-Leibler divergence D KL When using the position bias estimation unit 108, the amount of information D KL When the position bias θ shown in Eq. (10) is equal to or greater than a predetermined value (when the bias in the layout distribution is equal to or greater than a predetermined level), ek can be estimated by using the improved recursive EM algorithm shown in FIG. 4. On the other hand, the position bias estimation unit 108 estimates the amount of information D KL is smaller than the predetermined value (the bias in the arrangement distribution is less than a predetermined level), the position bias θ k can be estimated using a conventional recursive EM algorithm. In this way, if there is no bias in the placement of items, the dataset D e Without generating a position bias θ k By estimating, it is possible to reduce the processing load.

[0071] The content creating unit 109 calculates the relationship μ(i, u) between the item i and the user u in the data set D estimated by the relationship estimation unit 107 and the location bias θ estimated by the location bias estimation unit 108. k Or position bias θ ek In the case where the content is an advertisement content including a plurality of advertisements (i.e., items), the content generating unit 109 ranks the plurality of advertisements in order of relevance to the user u based on the relevance μ(i,u). Then, the content generating unit 109 ranks the plurality of advertisements in order of relevance to the user u based on the position bias θ k Or position bias θ ek Based on the sequence, the ordered advertisements are assigned to each position to create advertisement content. The content providing unit 110 provides the user u with the content created by the content creating unit 109. For example, the content providing unit 110 causes the created content to be displayed on a display unit of a user device used by the user u.

[0072] In this manner, in the information processing device 10 according to this embodiment, first, the first dataset generation unit 101 acquires the arrangement probability π(i, k) representing the probability that each of the multiple items is arranged at multiple positions in the dataset D. After that, the embedding vector generation unit 103 converts the multiple items into multiple embedding vectors. Then, the allocation probability calculation unit 104 calculates the allocation probability P(e|i) representing the probability of allocation from the multiple items to the multiple embedding vectors. Next, the probability expression derivation unit 106 uses the distribution of the arrangement probability and the distribution of the allocation probability to derive, for each of the multiple positions, the arrangement probability π(e, k) indicating the probability expression at which each of the multiple embedding vectors is arranged. The position bias estimation unit 108 estimates the position bias θ ek is estimated using the improved recursive EM algorithm shown in FIG.

[0073] Furthermore, in the information processing device 10, the content creation unit 109 creates content so as to place items highly relevant to the user in positions that the user is likely to recognize, and the content providing unit 110 can provide the generated content to the user.

[0074] An example of content in which multiple items are arranged at multiple positions is shown in Fig. 6. In Fig. 6, the upper part shows advertising content 60 in which advertising items 611, 612, and 613 are fixedly arranged at positions 601, 602, and 603, respectively. Here, it is assumed that, among the relevance between user u and each advertising item estimated by the relevance estimation unit 107, the relevance between user u and advertising item 612 is the highest, followed by the relevance between user u and advertising items 613 and 611 in that order. Furthermore, the position bias θ for each position estimated by the position bias estimation unit 108 is ek(or position bias θ k ), it is assumed that the position bias for position 601 is the highest, followed by the position bias for positions 602 and 603. In this case, the content creating unit 109 creates advertisement content 61 in which advertisement items 612, 613, and 611 are assigned to positions 601, 602, and 603, respectively, as shown in the lower part of FIG. 6. Then, the content providing unit 110 provides the created advertisement content 61 to the user u. For example, the content providing unit 110 controls the advertisement content 61 to be displayed on the display unit of the user device 11 of the user u.

[0075] On the display unit of the user device 11 of the user u, in the advertising content 61, the advertising item 612 that is of greater interest to the user u is placed at the position 601 with a higher position bias. This not only results in a more personalized display form for the user u, but also increases the probability that the user u will click on the advertising item 612, improving the CVR (conversion rate) and realizing effective marketing.

[0076] [Hardware configuration of information processing device 10] Next, a description will be given of an example of a hardware configuration of the information processing device 10. The user device 11 may also have a similar hardware configuration. FIG. 7 is a block diagram showing an example of a hardware configuration of the information processing device 10 according to the present embodiment. The information processing apparatus 10 according to the present embodiment can be implemented on any single or multiple computers, mobile devices, or any other processing platform. 7, the information processing device 10 is illustrated as being implemented in a single computer, but the information processing device 10 according to the present embodiment may be implemented in a computer system including multiple computers. The multiple computers may be connected to each other via a wired or wireless network so as to be able to communicate with each other.

[0077] 7, the information processing device 10 may include a CPU (Central Processing Unit) 701, a ROM (Read Only Memory) 702, a RAM (Random Access Memory) 703, a HDD (Hard Disk Drive) 704, an input unit 705, a display unit 706, a communication I / F 707, a GPU (Graphics Processing Unit) 708, and a system bus 709. The information processing device 10 may also include an external memory. The CPU 701 generally controls the operation of the information processing device 10, and controls each of the components (702 to 708) via a system bus 709, which is a data transmission path.

[0078] The ROM 702 is a non-volatile memory that stores a control program and the like necessary for the CPU 701 to execute processing. Note that the program may be stored in a non-volatile memory such as the HDD 704 or an SSD (Solid State Drive) or an external memory such as a removable storage medium (not shown). The RAM 703 is a volatile memory and functions as a main memory, a work area, etc., of the CPU 701. That is, when executing processing, the CPU 701 loads necessary programs, etc. from the ROM 702 into the RAM 703 and executes the programs, etc., to realize various functional operations.

[0079] The HDD 704 stores, for example, various data and various information required when the CPU 701 performs processing using a program. The HDD 704 also stores, for example, various data and various information obtained when the CPU 701 performs processing using a program. The input unit 705 is composed of a keyboard and a pointing device such as a mouse. The display unit 706 is configured with a monitor such as a liquid crystal display (LCD), etc. The display unit 706 may function as a GUI (Graphical User Interface) by being configured in combination with the input unit 705.

[0080] The communication I / F 707 is an interface that controls communication between the information processing device 10 and an external device. The communication I / F 707 provides an interface with a network and executes communication with an external device via the network. Various data, various parameters, etc. are transmitted and received between the external device and the communication I / F 707. In this embodiment, the communication I / F 707 may execute communication via a wired LAN (Local Area Network) or a dedicated line conforming to a communication standard such as Ethernet (registered trademark). However, the network available in this embodiment is not limited to this, and may be configured as a wireless network. This wireless network includes wireless PANs (Personal Area Networks) such as Bluetooth (registered trademark), ZigBee (registered trademark), and UWB (Ultra Wide Band). It also includes wireless LANs (Local Area Networks) such as Wi-Fi (Wireless Fidelity) (registered trademark), and wireless MANs (Metropolitan Area Networks) such as WiMAX (registered trademark). It also includes wireless WANs (Wide Area Networks) such as 4G and 5G. Note that the network only needs to connect each device so that they can communicate with each other and be capable of communication, and the communication standard, scale, and configuration are not limited to those described above. The GPU 708 is a processor specialized for image processing, and can cooperate with the CPU 701 to perform predetermined processes.

[0081] At least some of the functions of each element of the information processing device 10 shown in Fig. 5 can be realized by the CPU 701 executing a program. However, at least some of the functions of each element of the information processing device 10 shown in Fig. 5 may be operated as dedicated hardware. In this case, the dedicated hardware operates under the control of the CPU 701.

[0082] [User Attributes] As mentioned above, in this embodiment, the user context comprises one or more user attributes associated with the user. Examples of user attributes are mentioned here. The user attributes include factual characteristics (factual information) about a device owned by a user (user device) or a user. The factual characteristics may be factual characteristics (information) that are actually or objectively obtained from the user device or the user. The user attributes may also include user attributes (estimated user attributes) estimated by applying the factual features to a trained machine learning model. The machine learning model is configured to, for example, input the factual features of a target user and output the probability (correspondence probability) that each of a plurality of user attributes corresponds (matches) the target user. The estimated user attributes can be determined from the correlation probability.

[0083] [Item Features] As mentioned above, in this embodiment, an item is associated with a number of item features. Examples of item features are now mentioned. The item characteristics may include information identifying the item (item ID), information identifying the genre (higher classification) of the item, information identifying the shop where the item is sold (shop ID), etc. The item characteristics may also include transaction information (number of transactions, etc.) between the item ID and genre ID, and between the item ID and shop ID, depending on the transaction history.

[0084] Although specific embodiments have been described above, the embodiments are merely illustrative and are not intended to limit the scope of the present invention. The apparatus and method described in this specification can be embodied in forms other than those described above. Furthermore, the above-described embodiments can be appropriately omitted, substituted, and modified without departing from the scope of the present invention. Such omitted, substituted, and modified forms are included in the scope of the claims and their equivalents, and belong to the technical scope of the present invention.

[0085] The disclosure of this embodiment includes the following configuration. [1] An acquisition unit that acquires a placement probability representing the probability that each of n items (n is a natural number equal to or greater than 2) is placed in k positions (k is a natural number equal to or greater than 2); A conversion unit that converts the n items into m embedding vectors (m is a natural number equal to or greater than 2) that represent abstract expressions of features of the n items; A calculation unit for calculating an allocation probability representing a probability of allocation from the n items to the m embedding vectors; a derivation unit that derives a probability representation in which each of the m embedding vectors is placed at each of the k positions using the distribution of the placement probability and the distribution of the allocation probability; An information processing device having the above configuration.

[0086] [2] The information processing device described in [1], wherein, in the assignment probability, for each of the n items, the sum of conditional probabilities of assigning the item to the embedding vector is 1.

[0087] [3] Each of the n items is associated with a number of characteristics (a being a natural number equal to or greater than 2); The information processing device of [1], wherein the conversion unit converts the n items associated with the a features into the m embedding vectors, where m is less than a.

[0088] [4] The information processing device according to any one of [1] to [3], further comprising an estimation unit that estimates a position bias representing the probability that a user will recognize each of the k positions at which the m embedding vectors are placed, based on the allocation probability.

[0089] [5] a first estimation unit that estimates a first position bias representing a probability that a user recognizes each of the k positions in which the m embedding vectors are placed based on the assignment probability; and and a second estimation unit that estimates a second position bias representing the probability that a user will recognize each of the k positions at which the n items are placed, based on the placement probability.

[0090] [6] Further comprising a bias calculation unit that calculates a bias in the distribution of the configuration probability; when the deviation of the distribution of the arrangement probability is equal to or greater than a predetermined level, the first estimation unit estimates the first position bias; when the deviation of the distribution of the arrangement probabilities is less than the predetermined level, the second estimation unit estimates the second position bias. [5] An information processing device according to the present invention.

[0091] [7] The information processing device described in [6], wherein the bias calculation unit calculates the proportion of the n items placed among the k positions in the distribution of the placement probability as the bias of the distribution of the placement probability.

[0092] [8] The information processing device described in [6], wherein the bias calculation unit calculates the similarity of the distribution of the placement probability to a uniform distribution of the n items at the k positions as the bias of the distribution of the placement probability.

[0093] [9] The information processing device according to [8], wherein the bias calculation unit calculates the similarity using Kullback-Leibler divergence. [Explanation of symbols]

[0094] 1: information processing system, 10: information processing device 10, 11: user device, 101: first data set generation unit, 102: bias calculation unit, 103: embedding vector generation unit, 104: allocation probability calculation unit, 105: second data set generation unit, 106: probability expression derivation unit, 107: relevance estimation unit, 108: position bias estimation unit, 109: content creation unit, 110: content provision unit

Claims

1. an acquisition unit that acquires a placement probability representing a probability that each of n items (n is a natural number of 2 or more) is placed at k positions (k is a natural number of 2 or more); a conversion unit that converts the n items into m embedding vectors (m is a natural number equal to or greater than 2) that represent abstract representations of features of the n items; A calculation unit for calculating an allocation probability representing a probability of allocation from the n items to the m embedding vectors; an estimation unit that estimates a position bias representing a probability that a user recognizes each of the k positions at which the m embedding vectors are arranged based on the assignment probability; 13. An information processing device comprising:

2. 2 . The information processing apparatus according to claim 1 , further comprising a creating unit that creates advertising content in which a plurality of advertising items are arranged at a plurality of positions based on the position bias.

3. an acquisition unit that acquires a placement probability representing a probability that each of n items (n is a natural number of 2 or more) is placed at k positions (k is a natural number of 2 or more); a conversion unit that converts the n items into m embedding vectors (m is a natural number equal to or greater than 2) that represent abstract representations of features of the n items; A calculation unit for calculating an allocation probability representing a probability of allocation from the n items to the m embedding vectors; a first estimation unit that estimates a first position bias representing a probability that a user recognizes each of the k positions at which the m embedding vectors are arranged, based on the assignment probability; a second estimation unit that estimates a second position bias representing a probability that a user recognizes each of the k positions at which the n items are arranged, based on the arrangement probability; 13. An information processing device comprising:

4. 4. The information processing device according to claim 3, further comprising a creating unit that creates advertisement content in which a plurality of advertisement items are arranged at a plurality of positions based on the first position bias or the second position bias.

5. 4. The information processing apparatus according to claim 1, wherein in the assignment probability, a sum of conditional probabilities of assigning an item to the embedding vector for each of the n items is 1.

6. Each of the n items is associated with a number of features (a being a natural number equal to or greater than 2); The information processing device according to claim 1 or 3, wherein the conversion unit converts the n items associated with the a features into the m embedding vectors, where m is smaller than a.

7. A bias calculation unit that calculates a bias in the distribution of the arrangement probability, when the deviation of the distribution of the arrangement probability is equal to or greater than a predetermined level, the first estimation unit estimates the first position bias; when the deviation of the distribution of the arrangement probabilities is less than the predetermined level, the second estimation unit estimates the second position bias.

4. The information processing apparatus according to claim 3.

8. 8. The information processing device according to claim 7, wherein the bias calculation unit calculates a ratio of the n items to the k positions in the distribution of the placement probability as a bias of the distribution of the placement probability.

9. The information processing device according to claim 7 , wherein the bias calculation unit calculates a similarity of the distribution of the arrangement probabilities to a uniform distribution of the n items at the k positions as the bias of the distribution of the arrangement probabilities.

10. 10. The information processing apparatus according to claim 9, wherein the bias calculation unit calculates the similarity using Kullback-Leibler divergence.

11. An information processing method executed by an information processing device, Obtaining placement probabilities that represent the probability that each of n items (n is a natural number equal to or greater than 2) is placed at k positions (k is a natural number equal to or greater than 2); converting the n items into m embedding vectors (where m is a natural number equal to or greater than 1) that represent an abstract representation of features of the n items; calculating assignment probabilities representing probabilities of assignment from the n items to the m embedding vectors; estimating a position bias representing the probability that a user recognizes each of the k positions in which the m embedding vectors are located based on the assigned probabilities; 13. An information processing method comprising:

12. An information processing method executed by an information processing device, Obtaining placement probabilities that represent the probability that each of n items (n is a natural number equal to or greater than 2) is placed at k positions (k is a natural number equal to or greater than 2); converting the n items into m embedding vectors (where m is a natural number equal to or greater than 1) that represent an abstract representation of features of the n items; calculating assignment probabilities representing probabilities of assignment from the n items to the m embedding vectors; estimating a first position bias representing a probability that a user recognizes each of the k positions in which the m embedding vectors are located based on the assigned probabilities; estimating a second position bias representing a probability that a user will recognize each of the k positions in which the n items are placed based on the position probabilities; 13. An information processing method comprising:

13. A program for causing a computer to execute an information processing method, the information processing method comprising: Obtaining placement probabilities that represent the probability that each of n items (n is a natural number equal to or greater than 2) is placed at k positions (k is a natural number equal to or greater than 2); converting the n items into m embedding vectors (where m is a natural number equal to or greater than 1) that represent an abstract representation of features of the n items; calculating assignment probabilities representing probabilities of assignment from the n items to the m embedding vectors; estimating a position bias representing the probability that a user recognizes each of the k positions in which the m embedding vectors are located based on the assigned probabilities; Including, the program.

14. A program for causing a computer to execute an information processing method, the information processing method comprising: Obtaining placement probabilities that represent the probability that each of n items (n is a natural number equal to or greater than 2) is placed at k positions (k is a natural number equal to or greater than 2); converting the n items into m embedding vectors (where m is a natural number equal to or greater than 1) that represent an abstract representation of features of the n items; calculating assignment probabilities representing probabilities of assignment from the n items to the m embedding vectors; estimating a first position bias representing a probability that a user recognizes each of the k positions in which the m embedding vectors are located based on the assigned probabilities; estimating a second position bias representing a probability that a user will recognize each of the k positions in which the n items are placed based on the position probabilities; Including, the program.

Images