User portrait method, device and computer-readable storage medium

Abstract

The present invention discloses a user portrait method, device and computer-readable storage medium, and relates to the field of big data technology. The user portrait method includes: constructing multiple item association sequences according to the association relationship between items, wherein the elements in each item association sequence represent vectors of items with association relationships, and the vector of each item includes an initial vector corresponding to the label values ​​of the corresponding item in multiple label dimensions; inputting the item association sequence into a word conversion model to obtain an embedding vector of each item; constructing a corresponding relationship between the label value and the embedding vector of each label dimension according to the embedding vector of each item; determining the portrait vector of the user to be portraited based on the embedding vector of the items involved in the operation of the user to be portraited; determining the label value used to describe the user to be portraited according to the similarity between the portrait vector and the embedding vector of each label value. The present invention can improve the accuracy of user portraits.

Description

Technical Field

[0001] The present invention relates to the field of big data technology, and in particular to a user profiling method, device and computer-readable storage medium. Background Art

[0002] User portraits are analyzed and mined based on the massive log data stored on the server and the large amount of data in the database, and "labels" are assigned to users. "Labels" are identifiers that can represent a certain dimension of user characteristics, such as user interest preferences, behavioral preferences, consumption characteristics, etc. User portrait results can be applied to various fields such as personalized recommendations, advertising systems, content distribution, etc.

[0003] In related technologies, user profiling is performed based on the user's own behavior data. In addition, user tags need to be constructed once for each tag dimension. Summary of the invention

[0004] After analyzing the relevant technologies, the inventors found that there is an obvious Matthew effect in the field of e-commerce, that is, more than 90% of user behaviors are sparse. If labels are built based on the user's own behavior, it is difficult to solve the problem that the user portrait is not comprehensive and accurate due to the sparsity of user behavior.

[0005] A technical problem to be solved by the embodiments of the present invention is: how to improve the accuracy of user portraits.

[0006] According to a first aspect of some embodiments of the present invention, a user portrait method is provided, comprising: constructing a plurality of item association sequences according to association relationships between items, wherein the elements in each item association sequence represent vectors of items with association relationships, and the vector of each item includes an initial vector corresponding to the label values ​​of the corresponding item in multiple label dimensions; inputting the item association sequence into a word translation model to obtain an embedding vector for each item, wherein the embedding vector of the item includes an embedding vector of the label value of each label dimension of the item; constructing a correspondence between the label value of each label dimension and the embedding vector according to the embedding vector of each item; determining a portrait vector of the user to be portraited based on the embedding vectors of the items involved in the operation of the user to be portraited; and determining the label value used to describe the user to be portraited according to the similarity between the portrait vector and the embedding vector of each label value.

[0007] In some embodiments, constructing multiple item association sequences based on the association relationship between items includes: obtaining multiple operation behavior sequences from user data, wherein the elements in each operation behavior sequence are vectors of items operated in sequence by the same user, and the vector of the item includes an initial vector corresponding to the label values ​​of multiple label dimensions of the item; constructing multiple item association sequences based on the multiple operation behavior sequences.

[0008] In some embodiments, the user data is session data, which includes items operated by the user and the operation time. The items in the operation behavior sequence are items that are continuously operated by the user and the time interval of the operation is less than a preset value.

[0009] In some embodiments, constructing multiple item association sequences based on multiple operation behavior sequences includes: constructing a graph structure based on the adjacent relationships of items in the multiple operation behavior sequences, wherein nodes in the graph structure represent items, and there are edges between nodes corresponding to two adjacent items in the operation behavior sequences; performing multiple random walks based on the graph structure to obtain multiple item association sequences, wherein the elements in each item association sequence represent vectors of items visited during the walk.

[0010] In some embodiments, in a graph structure, the weight of an edge is determined according to the number of times the items corresponding to the two nodes connected by it are adjacent in multiple operation behavior sequences, and the probability of random walk is positively correlated with the weight of the edge.

[0011] In some embodiments, the language translation model is a Word2Vec model, and the objective function of the Word2Vec model includes a positive sample calculation item, a negative sample calculation item, and a global calculation item, wherein the global calculation item is determined based on the vector of the item that has generated the ordering behavior in the item association sequence.

[0012] In some embodiments, determining the portrait vector of the user to be profiled based on the embedding vectors of the items involved in the operations of the user to be profiled includes: obtaining the operation behavior sequence of the user to be profiled, wherein the elements in each operation behavior sequence are the embedding vectors of the items operated in sequence by the user to be profiled, and the elements include the embedding vectors corresponding to the label values ​​of multiple label dimensions of the item and the product of preset weights of the corresponding label dimensions; determining the portrait vector of the user to be profiled according to the operation behavior sequence.

[0013] In some embodiments, determining the portrait vector of the user to be profiled based on the operation behavior sequence includes: obtaining an intermediate vector by weighted summing up the elements in the operation behavior sequence of the user to be profiled; and weighting the vector corresponding to each label value in the intermediate vector according to the preset weight corresponding to each label dimension to obtain the portrait vector of the user to be profiled.

[0014] In some embodiments, for each element in the operation behavior sequence of the user to be profiled, the earlier the item corresponding to the element is operated, the smaller the weight of the element.

[0015] In some embodiments, the weight of an element is determined based on a piecewise exponential decay function, the independent variable of which is the length of the operation time of the item corresponding to the element from the current time, and the independent variable is negatively correlated with the piecewise exponential decay function.

[0016] According to a second aspect of some embodiments of the present invention, there is provided a user portrait device, comprising: a sequence construction module, configured to construct a plurality of item association sequences according to association relationships between items, wherein the elements in each item association sequence represent vectors of items with association relationships, and the vector of each item includes an initial vector corresponding to the label values ​​of the corresponding item in multiple label dimensions; an embedding vector acquisition module, configured to input the item association sequence into a word conversion model to obtain an embedding vector of each item, wherein the embedding vector of the item includes an embedding vector of the label value of each label dimension of the item; a correspondence construction module, configured to construct a correspondence between the label value of each label dimension and the embedding vector according to the embedding vector of each item; a portrait vector determination module, configured to determine the portrait vector of the user to be portraited based on the embedding vectors of the items involved in the operation of the user to be portraited; a user label value determination module, configured to determine the label value used to describe the user to be portraited according to the similarity between the portrait vector and the embedding vector of each label value.

[0017] According to a third aspect of some embodiments of the present invention, there is provided a user portrait device, comprising: a memory; and a processor coupled to the memory, wherein the processor is configured to execute any one of the aforementioned user portrait methods based on instructions stored in the memory.

[0018] According to a fourth aspect of some embodiments of the present invention, there is provided a computer-readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements any one of the aforementioned user profiling methods.

[0019] Some embodiments of the above invention have the following advantages or beneficial effects. The embodiments of the present invention mine the embedding vectors of items and label values ​​based on massive user behaviors, which can learn the characteristics of user groups and alleviate the impact of the sparsity of single user behaviors on the construction of the portrait system, so that the generated embedding vectors can more accurately represent the characteristics of items and label values. On this basis, in combination with the items operated by the user, the user's portrait vector is calculated based on the generated embedding vector to describe the user's characteristics, and the label value that can better describe the user's characteristics is determined based on the similarity between the user's portrait vector and the embedding vector of the label value, thereby being able to determine a label value that is closer to the user's characteristics, thereby improving the accuracy of the user portrait.

[0020] Further features and advantages of the present invention will become apparent from the following detailed description of exemplary embodiments of the present invention with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0022] Figure 1 A schematic flow chart of a user profiling method according to some embodiments of the present invention is shown.

[0023] Figure 2 A schematic flow chart of a method for constructing an item association sequence according to some embodiments of the present invention is shown.

[0024] Figure 3 A schematic flow chart of a method for determining a portrait vector according to some embodiments of the present invention is shown.

[0025] Figure 4 A schematic diagram of the structure of a user profiling device according to some embodiments of the present invention is shown.

[0026] Figure 5 A schematic diagram of the structure of a user profiling device according to some other embodiments of the present invention is shown.

[0027] Figure 6 A schematic diagram of the structure of a user portrait device according to some further embodiments of the present invention is shown. DETAILED DESCRIPTION

[0028] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. The following description of at least one exemplary embodiment is actually only illustrative and is by no means intended to limit the present invention and its application or use. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0029] The relative arrangement of components and steps, the numerical expressions and numerical values ​​set forth in these embodiments do not limit the scope of the present invention unless specifically stated otherwise.

[0030] At the same time, it should be understood that for the convenience of description, the sizes of the various parts shown in the drawings are not drawn according to the actual proportional relationship.

[0031] Technologies, methods, and apparatus known to ordinary technicians in the relevant field may not be discussed in detail, but where appropriate, such technologies, methods, and apparatus should be considered part of the authorization specification.

[0032] In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limiting. Therefore, other examples of the exemplary embodiments may have different values.

[0033] It should be noted that like reference numerals and letters refer to similar items in the following figures, and therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.

[0034] Figure 1 FIG. 1 is a flow chart showing a user profiling method according to some embodiments of the present invention. Figure 1 As shown, the user portrait method of this embodiment includes steps S102 to S110.

[0035] In step S102, multiple item association sequences are constructed based on the association relationships between items, wherein the elements in each item association sequence represent vectors of items with an association relationship, and each item vector includes an initial vector corresponding to the label values ​​of the corresponding item in multiple label dimensions.

[0036] Label dimensions refer to label categories, such as user purchasing power, income level, etc. Label values ​​refer to all values ​​under each label dimension. For example, the label values ​​under the user purchasing power dimension include rich type, cost-effective type, etc.

[0037] By analyzing the behaviors of a large number of users, we can obtain the association relationship between items. For example, if two items appear in the item set involved in the same behavior of the same user, then it is considered that there is an association relationship between the two items. The item set is, for example, a set of clicked items within a preset time period, a set of purchased items in a single shopping behavior, a set of favorite items in a certain category of a user, etc. By counting the items involved in the behaviors of a large number of users, we can obtain multiple item association sequences.

[0038] In some embodiments, multiple operation behavior sequences are obtained from user data, wherein the elements in each operation behavior sequence are vectors of items operated in sequence by the same user, and the vector of the item includes an initial vector corresponding to the label values ​​of multiple label dimensions of the item; and multiple item association sequences are constructed based on the multiple operation behavior sequences. For example, the operation behavior sequence is directly used as the item association sequence, or the association relationship between items is further mined through the operation behavior sequence to generate the item association sequence.

[0039] In some embodiments, the item association sequence is determined based on the user's session data. For example, in the user A's session data, the user clicks on item A, item B, item C, and item D in sequence, where the time interval between the user clicking on item A and item B is very short, the time interval between the user clicking on item B and item C is longer, and the time interval between the user clicking on item C and item D is very short. Then, two sequences {item A, item B} and {item C, item D} can be generated based on the user A's session data. Furthermore, each commodity includes label values ​​of multiple dimensions, for example, {item A, item B} can be further represented as {[category of item A, brand of item A, technological attributes of item A, ...], [category of item B, brand of item B, technological attributes of item B, ...]}. By representing each object in the above sequence using a vector, an item association sequence is obtained.

[0040] In some embodiments, a vector initialization model is used to obtain a vector corresponding to each tag value, and the output of the vector initialization model is multiplied by a preset weight corresponding to the tag dimension where the tag value is located to obtain an initial vector corresponding to the tag value. i Initialize the vector information of the i-th entity (i.e., item) in the item association sequence, E i,j is the vector initialization information of the i-th entity in the j-th label dimension, w i is the initialization weight of the jth label, then U i ={w1·E i,1 ,w2·E i,2 ,w3·E i,3 ,…}. The weight corresponding to each label dimension is obtained, for example, by random initialization and updating in combination with the training process, and the initial vector corresponding to the label value is obtained, for example, by a vector initialization model such as Xavier.

[0041] In step S104, the item association sequence is input into the word conversion model to obtain an embedding vector for each item, wherein the embedding vector for the item includes an embedding vector for the label value of each label dimension of the item.

[0042] In the field of natural language processing, sentences are composed of words of smaller granularity, and sentences are composed of paragraphs, chapters, and documents. In word vector mining algorithms such as Word2Vec, a sentence is regarded as a training sample, and words are regarded as entities in the sample. The embodiment of the present invention draws on the idea of ​​text processing, and regards a sequence of item associations as a "sentence", and the information of each item operated by the user as a "word". The information of the item is composed of various label values ​​such as brand, category, price level, etc.

[0043] The present invention treats the item association sequence as text for processing, so that each item in the sequence serves as the "context" of other items. Through the word turn model used to determine the embedding vector of each word in the text according to the context of the text, the embedding vector of each item in the item association sequence can be determined. Therefore, the embedding vector of the item can be determined according to the association relationship between the items.

[0044] In some embodiments, the word turn model is a Word2Vec model. In the Word2Vec model, it mainly includes a Skip-Gram model for given input words and predicting contexts and a CBOW model for predicting input words through given contexts.

[0045] In some embodiments, the Skip-Gram model is used, and the objective function used by the Skip-Gram model is improved to introduce the user's purchase information into the calculation of the embedding vector. Formula (1) exemplarily shows the improved objective function.

[0046]

[0047] In formula (1), for an item in an item association sequence, D p represents the positive sample set, D n Denotes a negative sample set, l denotes the current item, and c denotes items other than the current item. When item c is within the preset window range of item l, that is, when the two are considered to have a contextual relationship, (l,c)∈D p ; When item c is not within the preset window range of item l, (l,c)∈D n . It indicates the parameter value θ of the Word2Vec model when the formula that follows it has the maximum value.

[0048] In the summation calculation of formula (1), in addition to the calculation based on the positive samples in the first part and the negative samples in the second part, the third part is also included in, Indicates the items that the user ordered. In each sequence, if an item operated by a user is ordered by the user, that is, a real purchase behavior is brought about, then this behavior is introduced into the objective function of Skip-Gram. Regardless of whether the item is in the sliding window of Skip-Gram, it is assumed that the behavior is related to the center point of the sliding window, that is, a global behavior is introduced into the objective function of Skip-Gram.

[0049] Therefore, the embedding vectors obtained by the model and the similarity relationship between the embedding vectors are more likely to be calculated based on the purchase transaction conversion information in the sequence. The Word2Vec algorithm itself is unsupervised. Some operations such as user clicks and favorites on items can only represent that the user is interested in them, but cannot guarantee that the user will eventually buy them. By adding a constraint of supervisory information, the learned embedding vector can be better aligned with the business goal.

[0050] In some embodiments, when the embedding vector is calculated using the improved objective function, the same item association sequence includes at most one item ordered by the user.

[0051] In step S106, based on the embedding vector of each object, a corresponding relationship between the label value and the embedding vector of each label dimension is constructed.

[0052] Since the vector of an item is composed of vectors corresponding to the label values ​​of the item, after obtaining the vector of the item, the embedding vector corresponding to each label value can be obtained through the position information of the vector of the label value in the vector of the item.

[0053] When initializing the vector of an item, the vector is initialized for each label value of the item, so that the initialization vector of the item is composed of the initialization vector of the label value. During the model training process, the vectors of the same label value in different items are updated synchronously. Therefore, the embedding vector of each label value can be obtained according to the embedding vector of the item.

[0054] In step S108, a portrait vector of the user to be profiled is determined based on the embedding vectors of the items involved in the operation of the user to be profiled, such as click, purchase, favorite, etc.

[0055] The items involved in the operations of the user to be profiled represent the items that the user to be profiled is interested in, and these items can reflect the characteristics of the user to be profiled. Therefore, after obtaining the embedding vector of the item, the user can be described based on the items involved in the operations of the user to be profiled.

[0056] In some embodiments, a weighted sum is performed on the embedding vectors of the label values ​​of the items involved in the operation of the user to be profiled to obtain a portrait vector of the user to be profiled.

[0057] In step S110, based on the similarity between the portrait vector and the embedding vector of each label value, the label value used to describe the user to be profiled is determined, thereby generating a user portrait.

[0058] For example, the tag values ​​with similarities greater than a preset threshold are used as the tag values ​​of the user to be profiled, or a preset number of tag values ​​with the highest similarities are used as the tag values ​​of the user to be profiled.

[0059] The method of the above embodiment mines the embedding vectors of items and label values ​​based on massive user behaviors, can learn the characteristics of user groups, alleviate the impact of the sparsity of single user behaviors on the construction of the portrait system, and enable the generated embedding vectors to more accurately represent the characteristics of items and label values. On this basis, in combination with the items operated by the user, the user's portrait vector is calculated based on the generated embedding vector to describe the user's characteristics, and the label value that can better describe the user's characteristics is determined based on the similarity between the user's portrait vector and the embedding vector of the label value, so that the label value that is closer to the user's characteristics can be determined, thereby improving the accuracy of the user portrait.

[0060] In some embodiments, after obtaining multiple operation behavior sequences from user data, the association relationship between items can be further mined based on the random walk model to obtain the item association sequence. Random walk means starting from a node, randomly selecting one of its adjacent points, and then starting from this adjacent point to the next node. Repeat this step and then record all the nodes passed. Through random walk, a path starting from each node can be obtained, and this path represents the structural information corresponding to this node. Figure 2 An embodiment of the method for constructing an item association sequence according to the present invention is described.

[0061] Figure 2 FIG. 1 is a flow chart showing a method for constructing an item association sequence according to some embodiments of the present invention. Figure 2 As shown, the method for constructing an item association sequence in this embodiment includes steps S202 to S206.

[0062] In step S202, multiple operation behavior sequences are obtained from user data, wherein the elements in each operation behavior sequence are vectors of items operated in sequence by the same user, and the vector of the item includes an initial vector corresponding to the label values ​​of multiple label dimensions of the item.

[0063] In some embodiments, the operation behavior sequence is a click behavior sequence.

[0064] In some embodiments, the user data is session data, which includes items operated by the user and the operation time. The items in the operation behavior sequence are items that are continuously operated by the user and the time interval of the operation is less than a preset value.

[0065] In step S204, a graph structure is constructed according to the adjacent relationship of items in the plurality of operation behavior sequences, wherein the nodes in the graph structure represent items, and there is an edge between the nodes corresponding to two adjacent items in the operation behavior sequence.

[0066] This embodiment constructs a graph structure by analyzing massive operation behavior sequence features based on the assumption that two adjacent entities (ie, two items) in a behavior sequence have a similar relationship and non-adjacent entities do not have a similar relationship.

[0067] In some embodiments, in the graph structure, the weight of an edge is determined according to the number of times the items corresponding to the two nodes connected by it are adjacent in multiple operation behavior sequences. For example, the weight of the edge between nodes in the graph structure is obtained by accumulating the similarity relationship between two entities in the operation behavior sequence.

[0068] For example, in some operation behavior sequences, item A and item B are adjacent, that is, after a user operates item A, the next item operated is item B. Therefore, there is an edge between the nodes corresponding to item A and item B. If item A and item B are adjacent in 5 operation behavior sequences, the weight of the edge between the nodes corresponding to item A and item B is determined based on 5. For example, 5 is directly used as the weight, or the number of operation behavior sequences corresponding to each edge is normalized to obtain the weight.

[0069] In step S206, multiple random walks are performed based on the graph structure to obtain multiple item association sequences, wherein the elements in each item association sequence represent vectors of items visited during the walk.

[0070] In some embodiments, the probability of random walk is positively correlated with the weight of the edge. For example, the probability of random walk is proportional to the weight of the edge.

[0071] In some embodiments, when walking from the current node to the next node, the weights of all candidate edges between the current node and the candidate next node are obtained, and the candidate nodes are nodes with edges directly connected to the current node. Then, the ratio of the weight of each edge in the candidate edges to the sum of the weights of all candidate edges is used as the probability of random walking corresponding to each edge, and then the actual selected candidate edge and its corresponding next node are determined by roulette or other methods.

[0072] Through the above embodiments, a graph structure representing the similarity between items can be constructed based on massive user data, and a large amount of data based on the similarity relationship can be generated based on a random walk as an item association sequence, so as to further mine the vector representation of items and tag values ​​based on the item association sequence. Thus, richer and more comprehensive data can be generated, and the accuracy of user portraits can be improved.

[0073] Reference below Figure 3 An embodiment of determining a portrait vector of a user to be profiled is described.

[0074] Figure 3FIG. 2 is a flow chart showing a method for determining a portrait vector according to some embodiments of the present invention. Figure 3 As shown, the portrait vector determination method of this embodiment includes steps S302 to S304.

[0075] In step S302, the operation behavior sequence of the user to be profiled is obtained. The structure of the operation behavior sequence is similar to the item association sequence. The elements in each operation behavior sequence are the embedding vectors of the items operated in sequence by the user to be profiled. The elements include the embedding vectors corresponding to the label values ​​of multiple label dimensions of the item and the product of the preset weights of the corresponding label dimensions.

[0076] In step S304, a portrait vector of the user to be profiled is determined according to the operation behavior sequence.

[0077] The operation behavior sequence can not only reflect the items that the user is interested in, but also carry the timing information of the user's operation of the items. Therefore, it can more accurately reflect the user's portrait vector.

[0078] In some embodiments, an intermediate vector is obtained by weighted summing each element in the operation behavior sequence of the user to be profiled, so that the intermediate vector eliminates the object dimension, including the vectors corresponding to each label value. Then, according to the preset weight corresponding to each label dimension, the vector corresponding to each label value in the intermediate vector is weighted to obtain the portrait vector of the user to be profiled. Therefore, the dimension of the portrait vector after weighted summing is consistent with the dimension of the label value, which facilitates the similarity calculation between the two.

[0079] In some embodiments, for each element in the operation behavior sequence of the user to be profiled, the earlier the item corresponding to the element is operated, the smaller the weight of the element. Since operations closer to the current time can better reflect the current characteristics of the user, the user's portrait vector can be determined more accurately.

[0080] In some embodiments, the weight of an element is determined according to a piecewise exponential decay function, the independent variable of which is the length of the time from the time when the item corresponding to the element is operated to the current time, and the independent variable is negatively correlated with the piecewise exponential decay function. Formula (2) exemplarily shows a decay function.

[0081]

[0082] In formula (2), k represents the distance between the time when the item is operated and the current time, and F(k) represents the weight corresponding to the item; w1, w2, ..., wi, ... represent preset weight coefficients, and the sum of these coefficients is 1; t1, t2, ..., ti, ... represent preset time coefficients, indicating the time range controlled by the 1st, 2nd, ..., i, ... segment attenuation, and when the value of ti is large, it means that the attenuation of the user behavior of the time interval around ti is weaker, and the attenuation of the time interval far exceeding ti and far less than ti is stronger. In practical applications, the influence of the attenuation function can be controlled by adjusting the number of segments, weight coefficients, and time coefficients of the function.

[0083] If the user has recently clicked on item A, item B, and item C in sequence, the embedding vectors corresponding to item A, item B, and item C are represented by a, b, and c respectively, and the time intervals between the three clicks and the current one are k1, k2, and k3 respectively, then the user portrait vector can be expressed as F(k1)a+F(k2)b+F(k3)c.

[0084] Reference below Figure 4 An embodiment of the user portrait device of the present invention is described.

[0085] Figure 4 FIG. 1 is a schematic diagram showing the structure of a user profiling device according to some embodiments of the present invention. Figure 4 As shown, the user portrait device 40 of this embodiment includes: a sequence construction module 410, which is configured to construct multiple item association sequences according to the association relationship between items, wherein the elements in each item association sequence represent vectors of items with an association relationship, and the vector of each item includes an initial vector corresponding to the label value of the corresponding item in multiple label dimensions; an embedding vector acquisition module 420, which is configured to input the item association sequence into a word conversion model to obtain an embedding vector of each item, wherein the embedding vector of the item includes the embedding vector of the label value of each label dimension of the item; a correspondence construction module 430, which is configured to construct a correspondence between the label value of each label dimension and the embedding vector according to the embedding vector of each item; a portrait vector determination module 440, which is configured to determine the portrait vector of the user to be portraited based on the embedding vector of the items involved in the operation of the user to be portraited; a user label value determination module 450, which is configured to determine the label value used to describe the user to be portraited according to the similarity between the portrait vector and the embedding vector of each label value.

[0086] In some embodiments, the sequence construction module 410, the embedding vector acquisition module 420, and the correspondence construction module 430 are deployed in an offline device, and the portrait vector determination module 440 and the user tag value determination module 450 are deployed in an online device.

[0087] In some embodiments, the sequence construction module 410 is further configured to obtain multiple operation behavior sequences from user data, wherein the elements in each operation behavior sequence are vectors of items operated in sequence by the same user, and the vector of the item includes an initial vector corresponding to the label values ​​of multiple label dimensions of the item; and multiple item association sequences are constructed based on the multiple operation behavior sequences.

[0088] In some embodiments, the user data is session data, which includes items operated by the user and the operation time. The items in the operation behavior sequence are items that are continuously operated by the user and the time interval of the operation is less than a preset value.

[0089] In some embodiments, the sequence construction module 410 is further configured to construct a graph structure based on the adjacent relationships of items in multiple operation behavior sequences, wherein the nodes in the graph structure represent items, and there are edges between the nodes corresponding to two adjacent items in the operation behavior sequence; and perform multiple random walks based on the graph structure to obtain multiple item association sequences, wherein the elements in each item association sequence represent the vector of items visited during the walk.

[0090] In some embodiments, in a graph structure, the weight of an edge is determined according to the number of times the items corresponding to the two nodes connected by it are adjacent in multiple operation behavior sequences, and the probability of random walk is positively correlated with the weight of the edge.

[0091] In some embodiments, the word diversion model is a Word2Vec model, and the objective function of the Word2Vec model includes a positive sample calculation item, a negative sample calculation item, and a global calculation item, wherein the global calculation item is determined based on the vector of the item that has generated the ordering behavior in the item association sequence.

[0092] In some embodiments, the portrait vector determination module 440 is further configured to obtain the operation behavior sequence of the user to be profiled, wherein the elements in each operation behavior sequence are embedding vectors of the items operated in sequence by the user to be profiled, and the elements include the embedding vectors corresponding to the label values ​​of multiple label dimensions of the item and the product of the preset weights of the corresponding label dimensions; determine the portrait vector of the user to be profiled according to the operation behavior sequence.

[0093] In some embodiments, the portrait vector determination module 440 is further configured to obtain an intermediate vector by weighted summing up each element in the operation behavior sequence of the user to be profiled; weighting the vector corresponding to each label value in the intermediate vector according to the preset weight corresponding to each label dimension to obtain the portrait vector of the user to be profiled.

[0094] In some embodiments, for each element in the operation behavior sequence of the user to be profiled, the earlier the item corresponding to the element is operated, the smaller the weight of the element.

[0095] In some embodiments, the weight of an element is determined based on a piecewise exponential decay function, the independent variable of which is the length of the operation time of the item corresponding to the element from the current time, and the independent variable is negatively correlated with the piecewise exponential decay function.

[0096] Figure 5 FIG. 2 shows a schematic diagram of the structure of a user profiling device according to some other embodiments of the present invention. Figure 5 As shown, the user portrait device 50 of this embodiment includes: a memory 510 and a processor 520 coupled to the memory 510, and the processor 520 is configured to execute the user portrait method in any one of the aforementioned embodiments based on the instructions stored in the memory 510.

[0097] The memory 510 may include, for example, a system memory, a fixed non-volatile storage medium, etc. The system memory may store, for example, an operating system, an application program, a boot loader, and other programs.

[0098] Figure 6 FIG. 2 shows a schematic diagram of the structure of a user portrait device according to some other embodiments of the present invention. Figure 6 As shown, the user portrait device 60 of this embodiment includes: a memory 610 and a processor 620, and may also include an input / output interface 630, a network interface 640, a storage interface 650, etc. These interfaces 630, 640, 650 and the memory 610 and the processor 620 may be connected, for example, via a bus 660. Among them, the input / output interface 630 provides a connection interface for input / output devices such as a display, a mouse, a keyboard, and a touch screen. The network interface 640 provides a connection interface for various networked devices. The storage interface 650 provides a connection interface for external storage devices such as SD cards and USB flash drives.

[0099] An embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon, wherein the program implements any of the aforementioned user profiling methods when executed by a processor.

[0100] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable non-transient storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0101] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0102] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0103] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0104] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A user profiling method, comprising: According to the association relationship between items, multiple item association sequences are constructed, wherein the elements in each item association sequence represent vectors of items with an association relationship, and each item vector includes an initial vector corresponding to the label values ​​of the corresponding item in multiple label dimensions; Inputting the item association sequence into a word conversion model to obtain an embedding vector for each item, wherein the embedding vector for the item includes an embedding vector for a label value of each label dimension of the item; According to the embedding vector of each object, a corresponding relationship between the label value of each label dimension and the embedding vector is constructed, wherein the corresponding relationship is obtained according to the positional relationship of the vector of the label value in the vector of the object; Determining a portrait vector of the user to be profiled based on the embedding vectors of the items involved in the operation of the user to be profiled; According to the similarity between the portrait vector and the embedding vector of each tag value, a tag value for describing the user to be portraited is determined.

2. The user profiling method according to claim 1, wherein: The step of constructing a plurality of item association sequences according to the association relationship between the items includes: Acquire multiple operation behavior sequences from user data, wherein the elements in each operation behavior sequence are vectors of items operated in sequence by the same user, and the vectors of the items include initial vectors corresponding to label values ​​of multiple label dimensions of the items; According to the multiple operation behavior sequences, multiple item association sequences are constructed.

3. The user profiling method according to claim 2, wherein: The user data is session data, which includes items operated by the user and the operation time. The items in the operation behavior sequence are items that are continuously operated by the user and the time interval of the operation is less than a preset value.

4. The user profiling method according to claim 2, wherein: The constructing a plurality of item association sequences according to the plurality of operation behavior sequences comprises: Constructing a graph structure according to the adjacent relationships of the items in the plurality of operation behavior sequences, wherein the nodes in the graph structure represent items, and there is an edge between the nodes corresponding to two adjacent items in the operation behavior sequences; Multiple random walks are performed based on the graph structure to obtain multiple item association sequences, wherein the elements in each item association sequence represent vectors of items visited during the walk.

5. The user profiling method according to claim 4, wherein: In the graph structure, the weight of an edge is determined according to the number of times the items corresponding to the two nodes connected by the edge are adjacent in the multiple operation behavior sequences, and the probability of random walk is positively correlated with the weight of the edge.

6. The user profiling method according to claim 1, wherein: The word diversion model is a Word2Vec model, and the objective function of the Word2Vec model includes a positive sample calculation item, a negative sample calculation item and a global calculation item, wherein the global calculation item is determined according to the vector of the item that has generated the ordering behavior in the item association sequence.

7. The user profiling method according to claim 1, wherein: The determining of the portrait vector of the user to be profiled based on the embedding vector of the item involved in the operation of the user to be profiled comprises: Obtaining an operation behavior sequence of the user to be profiled, wherein the elements in each operation behavior sequence are embedding vectors of the items sequentially operated by the user to be profiled, and the elements include the product of the embedding vectors corresponding to the label values ​​of multiple label dimensions of the items and the preset weights of the corresponding label dimensions; According to the operation behavior sequence, a portrait vector of the user to be profiled is determined.

8. The user profiling method according to claim 7, wherein: Determining the portrait vector of the user to be portraited according to the operation behavior sequence includes: Obtaining an intermediate vector by weighted summing up each element in the operation behavior sequence of the user to be profiled; According to the preset weight corresponding to each label dimension, the vector corresponding to each label value in the intermediate vector is weighted to obtain the portrait vector of the user to be profiled.

9. The user profiling method according to claim 8, wherein: For each element in the operation behavior sequence of the user to be profiled, the earlier the item corresponding to the element is operated, the smaller the weight of the element.

10. The user profiling method according to claim 9, wherein: The weight of the element is determined according to a piecewise exponential decay function, the independent variable of which is the length of the operation time of the item corresponding to the element from the current time, and the independent variable is negatively correlated with the piecewise exponential decay function.

11. A user profiling device, comprising: A sequence construction module is configured to construct a plurality of item association sequences according to association relationships between items, wherein an element in each item association sequence represents a vector of items with an association relationship, and each item vector includes an initial vector corresponding to a label value of the corresponding item in a plurality of label dimensions; An embedding vector obtaining module is configured to input the item association sequence into a word conversion model to obtain an embedding vector for each item, wherein the embedding vector for the item includes an embedding vector for a label value of each label dimension of the item; A correspondence building module is configured to build a correspondence between the label value of each label dimension and the embedding vector according to the embedding vector of each object, wherein the correspondence is obtained according to the position relationship of the vector of the label value in the vector of the object; A portrait vector determination module, configured to determine a portrait vector of the user to be portraited based on an embedding vector of an object involved in an operation of the user to be portraited; The user tag value determination module is configured to determine the tag value used to describe the user to be profiled according to the similarity between the portrait vector and the embedding vector of each tag value.

12. A user profiling device, comprising: Memory; as well as A processor coupled to the memory, the processor being configured to execute the user portrait method according to any one of claims 1-10 based on instructions stored in the memory.

13. A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the user profiling method according to any one of claims 1 to 10.

Images