CN114564594A - Knowledge graph user preference entity recall method based on double-tower model - Google Patents

Abstract

The invention discloses a knowledge map user preference entity recall method based on a double-tower model, and an optimization method is added to the traditional double-tower model for better learning the interaction between users and items. The model can be used to recall entities related to user preferences on the knowledge graph. First, the entities corresponding to the items in the user history in the knowledge graph are used as the starting point, and all neighbor entities are retrieved along the edges. The recalled entities are then screened through the trained and optimized twin-tower model. Finally, the above operations are repeated with the recalled entity as a new starting point. Finally, a knowledge graph that can represent user preferences and potential preferences is formed.

Description

A Recall Method for User Preference Entity in Knowledge Graph Based on Twin Tower Model

technical field

The invention relates to the technical field of knowledge graphs and deep learning, in particular to a method for recalling user preference entities of knowledge graphs based on a twin-tower model.

Background technique

A knowledge graph, a concept proposed by Google in 2012, is a knowledge base that Google uses to enhance the functionality of its search engine. Essentially, the knowledge graph aims to describe various entities or concepts and their relationships in the real world, which constitute a huge semantic network graph, where nodes represent entities or concepts, and edges are composed of attributes or relationships. Each piece of knowledge is represented in the form of a triple (h, r, t), where h represents the head entity, t represents the tail entity, and r represents the relationship between the head and tail entities. With powerful semantic processing capabilities and open organization capabilities, knowledge graphs play an important role in recommender systems, intelligent question answering, and information retrieval, laying the foundation for knowledge organization and intelligent applications in the Internet era.

Traditional recommender systems use explicit or implicit information as input to make predictions, and there are two main problems. One is the sparsity problem. In actual scenarios, the interaction information between users and items is often very sparse. Using such a small amount of observation data to predict a large amount of unknown information will greatly increase the risk of overfitting. The second is the cold start problem. For newly added users or items, there is no corresponding historical information, so it is difficult to accurately model and recommend.

Knowledge graphs contain rich semantic associations between entities, providing a potential source of auxiliary information for recommender systems. Knowledge graphs introduce more semantic relationships for items, which can deeply discover user interests. Through the different types of relational links in the knowledge graph, it is beneficial to the divergence of the recommendation results. The knowledge graph can connect the user's historical records and recommendation results, thereby improving the user's satisfaction and acceptance of the recommendation results, and enhancing the user's trust in the system.

There are mainly two types of existing methods for knowledge graph recommendation applications. One is based on the vector embedding method (embedding-based methods), through the knowledge map vector embedding algorithm, to learn the entities and relationships in the knowledge map, and obtain the vector representation of the entities and relationships, and then introduce the vectors of the entities and relationships into the recommendation. in the system framework. For example, the DKN framework (Deep Knowledge-aware Network) based on convolutional neural networks, and the CKE framework (Collaborative Knowledge base Embedding) based on collaborative knowledge base embedding. Although the knowledge graph recommendation method based on vector embedding has strong flexibility, this kind of method is usually suitable for the application of link prediction in the graph, and the recommendation scenario needs to explore the potential interests of users. The other category is path-based methods, which explore various connections between entities in the knowledge graph and provide additional guidance for recommender systems. For example, Personalized Entity Recommendation and Meta-GraphBased Recommendation. Although path-based knowledge graph recommendation methods are able to use knowledge graphs in a more natural and intuitive way, they heavily rely on manually designed meta-paths, which are difficult to optimize in practice.

SUMMARY OF THE INVENTION

In view of the problems existing in the prior art, the present invention provides a method for recalling user preference entities of knowledge graphs based on the twin tower model. An optimization method is added to the traditional two-tower model to better learn the interaction between users and items; the trained two-tower model can be used to recall entities related to user preferences on the knowledge graph; The items in the history record take the entity corresponding to the knowledge graph as the starting point, and retrieve all neighbor entities along the edge; then the recalled entities are screened through the trained optimized twin-tower model. Finally, the recalled entities are used as a new starting point, and the above operations are repeated; finally, a knowledge graph that can represent user preferences and potential preferences is formed.

The technical scheme adopted in the present invention is as follows:

A method for recalling user preference entities in knowledge graph based on two-tower model, comprising the following steps:

1. Define the user feature vector and the item feature vector as the input of the twin tower model;

2. Train the twin-tower model, and optimize the twin-tower model by combining the in-batch softmax loss function and the frequency estimation method based on the hash sequence;

3. Define the entity mapping relationship between the user history interaction matrix and the knowledge graph;

4. Through the method of preference entity propagation, input the entities and user characteristics recalled by each propagation into the optimized twin-tower model to obtain the predicted probability, screen the entities with high probability, and finally obtain the knowledge representing user preferences and potential preferences Atlas.

The process of step 1 is as follows:

1.1. User feature refers to the user's interaction behavior with items, including click records, search records, social data, personal data and sample age. User feature vector is to convert the above interaction data into vectors and concatenate them. The method of converting raw data into vectors is called vector embedding. Vector embedding is a commonly used method to represent data features in machine learning. The purpose is to extract features from raw data, that is, after mapping through neural networks. The low-dimensional vector of ;

Further, the process of 1.1 is as follows:

1.1.1. The embedding of the user click record is the weighted average of the id-type embeddings of all clicked items. The id-type embedding is a vector that maps the unique identifier of the item to the same dimension, and its weight is proportional to the browsing time of the item. The embedding calculation formula of its user click record is as follows:

表示第i个权重，v_click,i表示点击记录中第i个物品的id类embedding，n表示embedding的个数；其中，

可通过如下公式计算：Where v _click represents the embedding of the user click record,

represents the ith weight, v _{click, i} represents the id class embedding of the ith item in the click record, and n represents the number of embeddings; among them,

It can be calculated by the following formula:

表示用户对物品i浏览的时间，N表示样本总数，k表示正例总数；in

Represents the time that the user browses the item i, N represents the total number of samples, and k represents the total number of positive examples;

1.1.2. The embedding of the user's search record is the keyword of the historical search for word segmentation to obtain the entry. The process of word segmentation is to obtain the embedding of the corresponding entry through the Word2vec model, and then perform a weighted average of the embeddings of the user search records.

The word segmentation is a technology in which the search engine divides the keyword string submitted by the user into different entry tokens.

The Word2vec model converts the content vocabulary in the text and reduces it to a space vector. The value of the word vector is affected by the context, which implies the correlation between words.

The calculation formula of the embedding of its user search records is as follows:

表示第i个权重，v_search,i表示搜索记录中第i个词条的embedding，n表示embedding的个数；where v _search represents the embedding of the user search record,

Represents the ith weight, v _{search, i} represents the embedding of the ith entry in the search record, and n represents the number of embeddings;

The weight calculation of the embedding of the search record:

The validity of the search is judged as whether the user clicks the item after the search;

1.1.3. The user's social data includes the embedding weighted average of favorites, likes, and subscription data. The embedding corresponding to the collection and like data refers to the embedding of the id category of the items that the user has favorited and liked; the embedding corresponding to the subscription data refers to the embedding of the id category of the person in charge corresponding to the user subscribed items.

The calculation formula of the embedding of its user social data is as follows:

表示第i个权重，v_social,i表示搜索记录中第i个社交数据的embedding；where v _social represents the embedding of user search records,

represents the ith weight, v _social,i represents the embedding of the ith social data in the search record;

For the weight calculation of the favorite and liked embeddings:

表示用户对物品i浏览的时间，N表示样本总数，k表示正例总数；in

Represents the time that the user browses the item i, N represents the total number of samples, and k represents the total number of positive examples;

For subscription embedding weight calculation:

示被订阅者第i个物品的浏览时间，N表示样本总数，k表示正例总数；in

Indicates the browsing time of the ith item of the subscriber, N represents the total number of samples, and k represents the total number of positive examples;

1.1.4. The user's personal data includes the user's gender, age, and region; gender is a simple binary feature, and age and region are continuous features, which can be normalized to a real value in the [0,1] interval. numerical value. The embedding of the user's personal data is a vector obtained by splicing the processed values of gender, age and region;

Going a step further, the process for 1.1.4 is as follows:

1.1.4.1. Calculate the binary representation of user gender, the formula is as follows:

1.1.4.2. Calculate the normalized real value of the user's age and region. The normalization formula is as follows:

where X represents the sample value, μ is the mean of all sample data, and σ is the standard deviation of all sample data;

1.1.4.3. Perform a splicing operation on the gender binary value, age and region normalized real value described in 1.1.4 to obtain a vector. The vector splicing operation formula is as follows:

v _personal = [gender, z _age , z _region ]

where v _personal represents the user feature vector, gender represents the user’s gender, and z _age and z _region represent the normalized value of the user’s age and region, respectively;

1.1.5. The user feature vector is obtained by performing the concatenate connection operation on the embedding of the user click record, the embedding of the user search record, the embedding of the user interaction data, and the embedding of the user's personal data described in the process of 1.1. The formula is as follows:

v _user ＝concatenate(v _click ,v _search ,v _social ,v _personal )

=[v _click [1],v _click [2],…,v _search [1],v _search [2],…,v _social [1],v _social [2],…,v _personal [1], v _personal [2],…]

where v _user represents the user feature vector, v _click [i] represents the ith component of the user click embedding, v _search [i] represents the ith component of the user search record embedding, and v _social [i] represents the user's social data embedding The i-th component, v _personal [i] represents the i-th component of the user's personal data embedding;

1.2. Item features include the id of the item and its context information, and the item feature vector is formed by splicing the id class embedding of the item and the embedding of its context information;

Going a step further, the process for 1.2 is as follows:

1.2.1. The id class embedding of the item is given, which is a vector that maps the unique identifier of the item to the same dimension;

1.2.2. The context information embedding of the item is given, which is a vector obtained by passing the context information through Word2vec;

1.2.3. Perform the concatenate connection operation of the id class embedding and the context information embedding described in 1.2 to obtain the item feature vector. The formula is as follows:

v _item =concatenate(v _id ,v _context )=[v _id [1],v _id [2],…,v _context [1],v _context [2],…]

Where v _item represents the feature vector of the item, v _id represents the id class embedding of the item, v _context represents the embedding of the item context information, v _id [i] represents the i-th component of the id class embedding of the item, and v _context [i] represents the item The ith component of the embedding of context information;

In the step 2, the twin tower model is derived from DSSM (Deep Structured Semantic Models). DSSM is a deep structured semantic model. In the field of natural language processing, DSSM is often used to solve the problem of semantic similarity. The dual-tower model can be divided into input layer, presentation layer and matching layer from top to bottom, and from the horizontal direction, it is divided into two input layers, two presentation layers and one matching layer. The outputs of the two input layers are the inputs of the two presentation layers, respectively, and the outputs of the two presentation layers are aggregated to the matching layer, and the whole is in the form of a "twin tower". In the present invention, the two inputs of the input layer of the twin-tower model are the user feature vector and the item feature vector respectively, the two representation layers have the same neural network structure, and the feature vector will obtain a vector of the same dimension after being calculated by the neural network. Finally, the two vectors are L2 normalized and then inner product is performed. Further, the optimized twin-tower model of the present invention adopts a frequency estimation method based on a hash sequence to optimize the twin-tower model. This method reduces the possible sampling bias problem of negative sampling in each batch, thereby optimizing the loss function.

The process of step 2 is as follows:

其中

表示d维实数向量。2.1. Give two embedding functions with parameters:

in

Represents a d-dimensional real vector.

是通过深度神经网络提取后的用户特征向量和物品特征向量，其中

和

分别是双塔模型需要的输入值。

are the user feature vector and item feature vector extracted by the deep neural network, where

and

are the input values required by the twin-tower model, respectively.

Going a step further, the process for 2.1 is as follows:

2.1.1. The twin tower model contains two deep neural network models. The component unit of the neural network model is the perceptron. The perceptron has several inputs and one output. The output and the input are a linear relationship trained, and the output value will get the final result through a nonlinear activation function;

2.1.2. The deep neural network is extended on the neural network model. The deep neural network has one input layer, one output layer and several hidden layers. Connection, full connection means that each node is connected to all nodes in the previous layer;

其中

2.1.3. Input the user feature vector and item feature vector into the corresponding neural network. Since the model needs to be trained, the output result is represented by a function containing the parameter θ to be trained. The output results are:

in

做L2归一化处理，L2归一化公式如下：2.1.4, will output the result

To do L2 normalization, the L2 normalization formula is as follows:

其中

γ_i表示[Y]向量上第i个分量；

in

γ _i represents the i-th component on the [Y] vector;

2.2. Through the frequency estimation method based on the hash sequence, that is, using the hash sequence to record the serial number of the sampling, the sampling deviation problem that may occur in each batch of negative sampling is reduced, thereby optimizing the loss function, and looping the following steps ;

Going a step further, the process for 2.2 is as follows:

2.2.1. Take T samples randomly from the sample set, which are expressed as follows:

其中x_i表示样本T中的第i个用户特征向量，

where x _i represents the ith user feature vector in sample T,

y _i represents the i-th item feature vector in sample T,

ri represents the feedback degree of the _ith user in the sample T, and _ri ∈ [0,1].

2.2.2. Calculate the probability p _i of y _i in each sample by using the frequency estimation method based on the hash sequence;

The process described in 2.2.2 is as follows:

2.2.2.1. Let the learning rate α, the arrays A and D of size H, and the hash function h;

where the hash function h can map each _yi to an integer in the range [0, H].

对于每一物品

有：2.2.2.2. For each step t = 1, 2, ..., the set of all items in a batch of the number of training samples

for each item

Have:

2.2.2.2.1,

表示y_i上次被采样到的时刻，

表示y_i被采样一次经过的步数；where ← means assignment,

represents the last time _yi was sampled,

Indicates the number of steps that _yi is sampled once;

2.2.2.2.2,

采样概率

2.2.2.3, for each

sampling probability

并给出推导过程；2.2.3. Calculate the optimized loss function

And give the derivation process;

The process described in 2.2.3 is as follows:

2.2.3.1. Give the vector inner product formula:

的概率可以利用softmax函数计算，其公式如下：2.2.3.2. Given a vector x, get one of the M items

The probability of can be calculated using the softmax function, the formula is as follows:

where e is a natural constant;

2.2.3.3. The weighted log-likelihood loss function is:

where T represents T samples randomly taken as described in 2.2.1;

其中batch-softmax函数为：2.2.3.4. The present invention uses a negative sampling algorithm for calculation, and the negative sampling algorithm adopts a smoothing strategy, which can improve the sampling probability of low-frequency samples. We sample negative samples in the same batch. gives a minimum batch, which is

The batch-softmax function is:

2.2.3.5. In each batch, due to the power distribution phenomenon, that is, random sampling of negative samples in each batch will make popular items easy to be sampled, and these popular items will be excessively punished in the loss function, so consider The sampling is corrected by frequency, and the formula is as follows:

s ^c (x _i , y _i )=s(x _i , y _i )-log( _pi )

where s ^c ( _xi , y _i ) represents the correction value of s ( _xi , y _i ), and _pi comes from the frequency estimation algorithm based on hash sequence described in 2.2.2;

2.2.3.6. Therefore, the modified conditional probability function is:

2.2.3.7. The final loss function is:

2.2.4. Use the gradient descent algorithm to update the parameter θ to make it close to the optimal value. Among them, the gradient descent algorithm is a commonly used algorithm in machine learning, which is used to solve the model parameters;

The step 3 process is as follows:

表示第

个用户与第

个物品的交互情况，U和V分别表示用户集合和物品集合；3.1. Given a user interaction matrix

means the first

users and

The interaction of each item, U and V represent the user set and the item set, respectively;

The expression formula of the user interaction matrix Y is as follows:

的实体；3.2. Define O _i,j to represent the interaction of user i with item j, and map the items that user i interacts with to the knowledge graph

entity;

Further, the 3.2 process is as follows:

3.2.1. O _i,j is an interaction matrix consisting of row vectors of users. O _i,j represents the interaction of user i with item j, where j is the index of the item;

3.2.2. Define a HashMap to store all items. HashMap is stored in the form of key-value key-value pairs. The key here stores the index of the item, and the value stores the entity corresponding to the object;

3.2.3, from the row vector of the user interaction matrix O to represent the interaction situation of a user, this row vector is stored in an array, and this array is defined as E;

3.2.3. Define a temporary set temp_set user to store the entity, traverse the array E, if the value of the element in the array is 1, then access the item HashMap according to the index of this element, obtain the corresponding entity, and store it in the temporary set temp_set;

The step 4 process is as follows:

4.1, give the user feature vector v _{user described in the first step for a user} ;

4.2. Initialize the user's temporary set temp_set according to the third step;

4.3. Define a HashMap of user preferences, named user_map, the key value stores the number of cycles, and the value stores the triplet set;

4.4. Define a set used_set to store the recalled triples;

4.5. Let the number of loops k = 1, 2, 3..., given a maximum value K, when the number of triples in user_map is greater than K, exit the loop and execute the following steps:

4.5.1. Traverse the entities in temp_set and remove the entities in used_set

temp_set←temp_set-used_set

← means assignment, - means difference operation;

并存放到user_map，定义如下：4.5.2. Find the triplet set from the knowledge graph

And stored in user_map, defined as follows:

Where (h', r', t') represents a triple, h' represents a head entity, r' represents a relationship, and t' represents a tail entity;

4.5.3. Since there is a loop in the knowledge graph, in order to prevent the collection of recalled entities, it is necessary to add entities in temp_set to used_set;

中三元组的尾实体,存放到集合

4.5.4, take out

The tail entity of the triple, stored in the collection

中的实体，取出对应的物品特征向量v_item；4.5.5. Traversal

The entity in , take out the corresponding item feature vector v _item ;

中对应的物品的概率，并根据概率大小将实体排序；4.5.6. Through the input parameters v _user and v _item described in the second step to the twin tower model, the user and set are obtained

The probability of the corresponding item in , and the entities are sorted according to the probability;

4.5.7. Given a value τ, 0<τ≤1, it is used to determine the number of screening entities:

表示集合中的元素按概率排序，并返回一个数组，getSubVec(i,j)表示取出子数组，即原数组的第i到第j的元素，newSet()表示将数组转化为集合，

表示向上取整；in

Indicates that the elements in the set are sorted by probability and returns an array, getSubVec(i,j) means to take out the sub-array, that is, the i-th to j-th elements of the original array, newSet() means to convert the array into a set,

means rounded up;

集合中筛选出尾实体属于

的三元组：4.5.8. From

Filters out the tail entity from the collection that belongs to

triples of :

4.5.9. The method of storing triplet set Λ to user_map: user_map(k, Λ);

集合覆盖到temp_set；4.5.10, will

Set overwrites to temp_set;

4.6. After executing 4.5, the knowledge graph of user preference is finally obtained.

The beneficial effects of the present invention are as follows: the entities preferred by the user in the knowledge graph are screened through the optimized twin-tower model. The optimization method of the twin-tower model adopts the frequency estimation method based on the hash sequence, so that the items can better adapt to a variety of data distributions. The screening of knowledge graph entities can not only obtain better data, the recalled entities can be truly close to user preferences; and screening entities is conducive to the deep recall of knowledge graphs, because each recalled entity may be due to the cardinality of the previous entity. Explosive growth, if not screened, will affect computational efficiency and exploration of users’ potential preferences.

Detailed ways

The present invention will be further described below in conjunction with the embodiments.

Example:

The first step: define the user feature vector and item feature vector as the input of the twin tower model;

The process of the first step is as follows:

1.1. User feature refers to the user's interaction behavior with items, including click records, search records, social data, personal data and sample age. User feature vector is to convert the above interaction data into vectors and concatenate them. The method of converting raw data into vectors is called vector embedding. Vector embedding is a method commonly used in machine learning to represent data features. The purpose is to extract features from raw data, that is, after mapping through neural networks. The low-dimensional vector of ;

Further, the process of 1.1 is as follows

1.1.1. The embedding of the user click record is the weighted average of the id-type embeddings of all clicked items. The id-type embedding is a vector that maps the unique identifier of the item to the same dimension, and its weight is proportional to the browsing time of the item. The embedding calculation formula of its user click record is as follows:

表示第i个权重，v_click,i表示点击记录中第i个物品的id类embedding，n表示embedding的个数；其中，

可通过如下公式计算：Where v _click represents the embedding of the user click record,

represents the ith weight, v _{click, i} represents the id class embedding of the ith item in the click record, and n represents the number of embeddings; among them,

It can be calculated by the following formula:

表示用户对物品i浏览的时间，N表示样本总数，k表示正例总数；in

Represents the time that the user browses the item i, N represents the total number of samples, and k represents the total number of positive examples;

1.1.2. The embedding of the user's search record is the keyword of the historical search for word segmentation to obtain the entry. The process of word segmentation is to obtain the embedding of the corresponding entry through the Word2vec model, and then perform a weighted average of the embeddings of the user search records.

The word segmentation is a technology in which the search engine divides the keyword string submitted by the user into different entry tokens.

The Word2vec model was proposed by Mikolov et al. in 2013. This model converts the content words in the text and simplifies them into space vectors. The value of the word vector is affected by the context, which implies the correlation between words.

The calculation formula of the embedding of its user search records is as follows:

表示第i个权重，v_search,i表示搜索记录中第i个词条的embedding，n表示embedding的个数；where v _search represents the embedding of the user search record,

Represents the ith weight, v _{search, i} represents the embedding of the ith entry in the search record, and n represents the number of embeddings;

The weight calculation of the embedding of the search record:

The validity of the search is judged as whether the user clicks the item after the search;

1.1.3. The user's social data includes the embedding weighted average of favorites, likes, and subscription data. The embedding corresponding to the collection and like data refers to the embedding of the id category of the items that the user has favorited and liked; the embedding corresponding to the subscription data refers to the embedding of the id category of the person in charge corresponding to the user subscribed items.

The calculation formula of the embedding of its user social data is as follows:

表示第i个权重，v_social,i表示搜索记录中第i个社交数据的embedding；where v _social represents the embedding of user search records,

represents the ith weight, v _social,i represents the embedding of the ith social data in the search record;

For the weight calculation of the favorite and liked embeddings:

表示用户对物品i浏览的时间，N表示样本总数，k表示正例总数；in

Represents the time that the user browses the item i, N represents the total number of samples, and k represents the total number of positive examples;

For subscription embedding weight calculation:

示被订阅者第i个物品的浏览时间，N表示样本总数，k表示正例总数；in

Indicates the browsing time of the ith item of the subscriber, N represents the total number of samples, and k represents the total number of positive examples;

1.1.4. The user's personal data includes the user's gender, age and region. Among them, gender is a simple binary feature, and age and region are continuous features, which can be normalized to real values in the [0,1] interval. The embedding of the user's personal data is a vector obtained by splicing the processed values of gender, age and region;

Going a step further, the process for 1.1.4 is as follows:

1.1.4.1. Calculate the binary representation of user gender, the formula is as follows:

1.1.4.2. Calculate the normalized real value of the user's age and region. The normalization formula is as follows:

where X represents the sample value, μ is the mean of all sample data, and σ is the standard deviation of all sample data;

1.1.4.3. Perform a splicing operation on the gender binary value, age and region normalized real value described in 1.1.4 to obtain a vector. The vector splicing operation formula is as follows:

v _personal = [gender, z _age , z _region ]

where v _personal represents the user feature vector, gender represents the user’s gender, and z _age and z _region represent the normalized value of the user’s age and region, respectively;

1.1.5. The user feature vector is obtained by performing the concatenate connection operation on the embedding of the user click record, the embedding of the user search record, the embedding of the user interaction data, and the embedding of the user's personal data described in the process of 1.1. The formula is as follows:

v _user ＝concatenate(v _click ,v _search ,v _social ,v _personal )

=[v _click [1],v _click [2],…,v _search [1],v _search [2],…,v _social [1],v _social [2],…,v _personal [1], v _personal [2],…]

where v _user represents the user feature vector, v _click [i] represents the ith component of the user click embedding, v _search [i] represents the ith component of the user search record embedding, and v _social [i] represents the user's social data embedding The i-th component, v _personal [i] represents the i-th component of the user's personal data embedding;

1.2. Item features include the id of the item and its context information, and the item feature vector is formed by splicing the id class embedding of the item and the embedding of its context information;

Going a step further, the process for 1.2 is as follows:

1.2.1. The id class embedding of the item is given, which is a vector that maps the unique identifier of the item to the same dimension;

1.2.2. The context information embedding of the item is given, which is a vector obtained by passing the context information through Word2vec;

1.2.3. Perform the concatenate connection operation of the id class embedding and the context information embedding described in 1.2 to obtain the item feature vector. The formula is as follows:

v _item =concatenate(v _id ,v _context )=[v _id [1],v _id [2],…,v _context [1],v _context [2],…]

Where v _item represents the feature vector of the item, v _id represents the id class embedding of the item, v _context represents the embedding of the item context information, v _id [i] represents the i-th component of the id class embedding of the item, and v _context [i] represents the item The ith component of the embedding of context information;

The second step: train the double-tower model, and optimize the double-tower model by combining the in-batch softmax loss function and the frequency estimation method based on the hash sequence;

In the second step, the twin tower model is derived from DSSM (Deep Structured Semantic Models). DSSM is a deep structured semantic model. In the field of natural language processing, DSSM is often used to solve the problem of semantic similarity. The dual-tower model can be divided into input layer, presentation layer and matching layer from top to bottom, and from the horizontal direction, it is divided into two input layers, two presentation layers and one matching layer. The outputs of the two input layers are the inputs of the two presentation layers, respectively, and the outputs of the two presentation layers are aggregated to the matching layer, and the whole is in the form of a "twin tower". In the present invention, the two inputs of the input layer of the twin-tower model are the user feature vector and the item feature vector respectively, the two representation layers have the same neural network structure, and the feature vector will obtain a vector of the same dimension after being calculated by the neural network. Finally, the two vectors are L2 normalized and then inner product is performed. Further, the optimized twin-tower model of the present invention adopts a frequency estimation method based on a hash sequence to optimize the twin-tower model. This method reduces the possible sampling bias problem of negative sampling in each batch, thereby optimizing the loss function.

The process of the second step is as follows:

其中

表示d维实数向量。2.1. Give two embedding functions with parameters:

in

Represents a d-dimensional real vector.

是通过深度神经网络提取后的用户特征向量和物品特征向量，其中

和

分别是双塔模型需要的输入值。

are the user feature vector and item feature vector extracted by the deep neural network, where

and

are the input values required by the twin-tower model, respectively.

Further, the process of 2.1 is as follows

2.1.1. The twin tower model contains two deep neural network models. The component unit of the neural network model is the perceptron. The perceptron has several inputs and one output. The output and the input are a linear relationship trained, and the output value will get the final result through a nonlinear activation function;

2.1.2. The deep neural network is extended on the neural network model. The deep neural network has one input layer, one output layer and several hidden layers. Connection, full connection means that each node is connected to all nodes in the previous layer;

其中

2.1.3. Input the user feature vector and item feature vector into the corresponding neural network. Since the model needs to be trained, the output result is represented by a function containing the parameter θ to be trained. The output results are:

in

做L2归一化处理，L2归一化公式如下：2.1.4, will output the result

To do L2 normalization, the L2 normalization formula is as follows:

其中

γ_i表示[Y]向量上第i个分量；

in

γ _i represents the i-th component on the [Y] vector;

2.2. Through the frequency estimation method based on the hash sequence, that is, using the hash sequence to record the serial number of the sampling, the sampling deviation problem that may occur in each batch of negative sampling is reduced, thereby optimizing the loss function, and looping the following steps ;

Going a step further, the process for 2.2 is as follows:

2.2.1. Take T samples randomly from the sample set, which are expressed as follows:

其中x_i表示样本T中的第i个用户特征向量，

where x _i represents the ith user feature vector in sample T,

y _i represents the i-th item feature vector in the sample T,

ri represents the feedback degree of the _ith user in the sample T, and _ri ∈ [0,1].

2.2.2. Calculate the probability p _i of y _i in each sample by using the frequency estimation method based on the hash sequence;

The process described in 2.2.2 is as follows:

2.2.2.1. Let the learning rate α, the arrays A and D of size H, and the hash function h;

where the hash function h can map each _yi to an integer in the range [0, H].

对于每一物品

有：2.2.2.2. For each step t = 1, 2, ..., the set of all items in a batch of the number of training samples

for each item

Have:

2.2.2.2.1,

表示y_i上次被采样到的时刻，

表示y_i被采样一次经过的步数；where ← means assignment,

represents the last time _yi was sampled,

Indicates the number of steps that _yi is sampled once;

2.2.2.2.2,

采样概率

2.2.2.3, for each

sampling probability

并给出推导过程；2.2.3. Calculate the optimized loss function

And give the derivation process;

The process described in 2.2.3 is as follows:

2.2.3.1. Give the vector inner product formula:

的概率可以利用softmax函数计算，其公式如下：2.2.3.2. Given a vector x, get one of the M items

The probability of can be calculated using the softmax function, the formula is as follows:

所带的参数，e为自然常数；Among them; θ after the symbol means

With the parameters, e is a natural constant;

2.2.3.3. The weighted log-likelihood loss function is:

where T represents T samples randomly taken as described in 2.2.1;

其中batch-softmax函数为：2.2.3.4. The present invention uses a negative sampling algorithm for calculation, and the negative sampling algorithm adopts a smoothing strategy, which can improve the sampling probability of low-frequency samples. We sample negative samples in the same batch. gives a minimum batch, which is

The batch-softmax function is:

2.2.3.5. In each batch, due to the power distribution phenomenon, that is, random sampling of negative samples in each batch will make popular items easy to be sampled, and these popular items will be excessively punished in the loss function, so consider The sampling is corrected by frequency, and the formula is as follows:

s ^c (x _i , y _i )=s(x _i , y _i )-log( _pi )

where s ^c ( _xi , y _i ) represents the correction value of s ( _xi , y _i ), and _pi comes from the frequency estimation algorithm based on hash sequence described in 2.2.2;

2.2.3.6. Therefore, the modified conditional probability function is:

2.2.3.7. The final loss function is:

2.2.4. Use the gradient descent algorithm to update the parameter θ to make it close to the optimal value. The gradient descent algorithm

Algorithms commonly used in machine learning to solve model parameters;

Step 3: Define the entity mapping relationship between the user history interaction matrix and the knowledge graph;

The third step process is as follows:

表示第

个用户与第

个物品的交互情况，U和V分别表示用户集合和物品集合；3.1. Given a user interaction matrix

means the first

users and

The interaction of each item, U and V represent the user set and the item set, respectively;

The expression formula of the user interaction matrix Y is as follows:

的实体；3.2. Define O _i,j to represent the interaction of user i with item j, and map the items that user i interacts with to the knowledge graph

entity;

Further, the 3.2 process is as follows:

3.2.1. O _i,j is an interaction matrix consisting of row vectors of users. O _i,j represents the interaction of user i with item j, where j is the index of the item;

3.2.2. Define a HashMap to store all items. HashMap is stored in the form of key-value key-value pairs. The key here stores the index of the item, and the value stores the entity corresponding to the object;

3.2.3, from the row vector of the user interaction matrix O to represent the interaction situation of a user, this row vector is stored in an array, and this array is defined as E;

3.2.3. Define a temporary set temp_set user to store entities, traverse the array E, if the value of the element in the array is 1, access the item HashMap according to the index of this element, obtain the corresponding entity, and store it in the temporary set temp_set;

Step 4: Through the method of preference entity propagation, input the entities and user characteristics recalled by each propagation into the optimized twin-tower model to obtain the predicted probability, screen the entities with high probability, and finally obtain the representation of user preference and potential preference. knowledge graph;

The fourth step process is as follows:

4.1, give the user feature vector v _{user described in the first step for a user} ;

4.2. Initialize the user's temporary set temp_set according to the third step;

4.3. Define a HashMap of user preferences, named user_map, the key value stores the number of cycles, and the value stores the triplet set;

4.4. Define a set used_set to store the recalled triples;

4.5. Let the number of loops k = 1, 2, 3..., given a maximum value K, when the number of triples in user_map is greater than K, exit the loop and execute the following steps:

4.5.1. Traverse the entities in temp_set and remove the entities in used_set

temp_set←temp_set-used_set

← means assignment, - means difference operation;

并存放到user_map，定义如下：4.5.2. Find the triplet set from the knowledge graph

And stored in user_map, defined as follows:

Where (h', r', t') represents a triple, h' represents a head entity, r' represents a relationship, and t' represents a tail entity;

4.5.3. Since there is a loop in the knowledge graph, in order to prevent the collection of recalled entities, it is necessary to add entities in temp_set to used_set;

中三元组的尾实体,存放到集合

4.5.4, take out

The tail entity of the triple, stored in the collection

中的实体，取出对应的物品特征向量v_item；4.5.5. Traversal

The entity in , take out the corresponding item feature vector v _item ;

中对应的物品的概率，并根据概率大小将实体排序；4.5.6. Through the input parameters v _user and v _item described in the second step to the twin tower model, the user and set are obtained

The probability of the corresponding item in , and the entities are sorted according to the probability;

4.5.7. Given a value τ, 0<τ≤1, it is used to determine the number of screening entities:

表示集合中的元素按概率排序，并返回一个数组，getSubVec(i,j)表示取出子数组，即原数组的第i到第j的元素，newSet()表示将数组转化为集合，

表示向上取整；in

Indicates that the elements in the set are sorted by probability and returns an array, getSubVec(i,j) means to take out the sub-array, that is, the i-th to j-th elements of the original array, newSet() means to convert the array into a set,

means rounded up;

集合中筛选出尾实体属于

的三元组：4.5.8. From

Filter out the tail entity from the collection that belongs to

triples of :

集合覆盖到temp_set；4.5.9. The method of storing the triplet set Λ to user_map: user_map(k, Λ); 4.5.10. Putting the

Set overwrites to temp_set;

4.6. After executing 4.5, the knowledge graph of user preference is finally obtained.

Claims (5)

1. A knowledge graph user preference entity recalling method based on a double-tower model is characterized by comprising the following steps:

1) defining a user characteristic vector and an article characteristic vector as the input of a double-tower model;

2) training a double-tower model, and optimizing the double-tower model by combining an in-batch softmax loss function and a frequency estimation method based on a Hash sequence;

3) defining an entity mapping relation between a user historical interaction matrix and a knowledge graph;

4) and inputting the entities recalled by each transmission and the user characteristics into the optimized double-tower model user bias to obtain a prediction probability in a preference entity transmission mode, screening the entities according to the prediction probability, and finally obtaining a knowledge graph representing the user preference and the potential preference.

2. The method for recalling knowledge-graph user preference entity based on double-tower model according to claim 1, wherein the specific process of step 1) is as follows:

1.1) user characteristics refer to interactive behaviors of a user on an article, including click records, search records, social data, personal data and sample age, and user characteristic vectors are obtained by converting the interactive data into vectors and splicing the vectors; the mode of converting the original data into the vector is called vector embedding;

1.1.1) the embedding of the user click record is the weighted average of the id types of the clicked items, wherein the id types of the embedding are vectors which map the unique identifiers of the items to the same dimension, and the weight of the id types of the embedding is in direct proportion to the item browsing time; the imbedding calculation formula of the user click record is as follows:

wherein v is_clickAn embedding representing the user clicking on a record,

denotes the ith weight, v_click,iRepresenting id type embedding of the ith item in the click record, wherein n represents the number of the embedding; wherein,

can be calculated by the following formula:

wherein

Representing the time when the user browses the item i, N representing the total number of samples, and k representing the total number of positive examples;

1.1.2) the embedding of the user search record is to perform word segmentation on the keywords of the historical search to obtain entries; the Word segmentation process is to obtain embedding of a corresponding entry through a Word2vec model, and then carry out weighted average on the embedding of the user search record;

the formula for embedding the user search records is as follows:

wherein v is_searchEmbedding representing a user's search records,

denotes the ith weight, v_search,iIndicating the imbedding of the ith entry in the search record, wherein n indicates the number of the imbedding;

weight calculation of embedding of search records:

judging whether the user clicks the article after searching the article according to the searching validity;

1.1.3) the social data of the user comprises the embedding weighted average corresponding to the collection, praise and subscription data; the imbedding corresponding to the collection and approval data refers to the imbedding of the item id class collected and approved by the user; subscribing the embedding corresponding to the data refers to subscribing the embedding of the id class of a person in charge corresponding to the item by the user;

the formula for embedding the social data of the user is as follows:

wherein v is_socialEmbedding representing a user's search records,

denotes the ith weight, v_social,iEmbedding representing the ith social data in the search record;

weight calculation for favorites and praise embedding:

wherein

Representing user to itemi, browsing time, N represents the total number of samples, and k represents the total number of positive cases;

embedding weight calculation for a subscription:

wherein

Showing the browsing time of the ith item of the subscriber, wherein N represents the total number of samples, and k represents the total number of positive examples;

1.1.4) personal data of a user includes the user's gender, age, and location; the neutral characteristic is a simple binary characteristic, the age and the region belong to a continuous characteristic, and the continuous characteristic is normalized into a real numerical value in a [0,1] interval; embedding of user personal data is a vector obtained by splicing the processed values of gender, age and region;

1.1.4.1) calculates a binary representation of the user's gender, which is formulated as follows:

1.1.4.2) calculating the normalized real value of the age and the region of the user, wherein the normalized formula is as follows:

wherein X represents the sample value, μ is the mean of all sample data, and σ is the standard deviation of all sample data;

1.1.4.3) splicing the sex binary value, age and the region normalized real value in the step 1.1.4) to obtain a vector, wherein the vector splicing operation formula is as follows:

v_personal＝[gender,z_age,z_region]

wherein v is_personalRepresenting the user feature vector, gender, z_ageAnd z_regionNormalized values respectively representing the age and region of the user;

1.1.5) clicking recorded embedding by the user according to the process in the step 1.1), searching recorded embedding by the user, interacting data embedding by the user, and carrying out concatemate connection operation on the embedding of personal data of the user to obtain a user characteristic vector, wherein the formula is as follows:

v_user＝concatenate(v_click,v_search,v_social,v_personal)＝[v_click[1],v_click[2],…,v_search[1],v_search[2],…,v_social[1],v_social[2],…,v_personal[1],v_personal[2],…]

wherein v is_userRepresenting a user feature vector, v_click[i]The ith component, v, representing the user clicking on embedding_search[i]The i-th component, v, representing the user's search record embedding_social[i]The i component, v, representing the user's social data embedding_personal[i]An ith component representing user personal data embedding;

1.2) the article characteristics comprise the id of the article and the context information thereof, and the article characteristic vector is formed by splicing the id class embedding of the article and the embedding of the context information thereof;

1.2.1) giving an id class embedding of the article, wherein the id class embedding is a vector for mapping the unique identifier of the article to the same dimension;

1.2.2) giving context information embedding of the article, wherein the context information is a vector obtained by Word2 vec;

1.2.3) carrying out a containing connection operation on the id type embedding and the context information embedding in the step 1.2) to obtain an article feature vector, wherein the formula is as follows:

v_item＝concatenate(v_id,v_context)＝[v_id[1],v_id[2],…,v_context[1],v_context[2],…]

wherein v is_itemIndicating the nature of the articleEigenvectors, v_idId class embedding, v representing an item_contextEmbedding, v representing item context information_id[i]The i component, v, of the id class embedding representing the item_context[i]The ith component of embedding representing item context information.

3. The method for recalling knowledge-graph user preference entity based on double-tower model according to claim 1, wherein the specific process of the step 2) is as follows:

2.1) the embedding function gives two band parameters:

wherein

Representing a d-dimensional real number vector;

is the user feature vector and the article feature vector extracted by the deep neural network, wherein

And

respectively are input values required by the double-tower model;

2.1.1) the double tower model contains two deep neural network models; the component unit of the neural network model is a perceptron, the perceptron has a plurality of inputs and an output, wherein the output and the input are a linear relation trained, the output value can obtain the final result through a nonlinear activation function;

2.1.2) the deep neural network is expanded on the neural network model, and the deep neural network is provided with an input layer, an output layer and a plurality of hidden layers;

2.1.3) inputting the user characteristic vector and the article characteristic vector into a corresponding neural network, wherein the output result is represented by a function containing a parameter theta to be trained because the model needs to be trained, and the output results are respectively as follows:

wherein

2.1.4) will output the result

The L2 normalization processing is carried out, and the L2 normalization formula is as follows:

wherein

γ_iIs represented by [ Y]The ith component on the vector;

2.2) through a frequency estimation method based on a hash sequence, namely, the hash sequence is used for recording the serial number of the sample, the problem of sampling deviation possibly occurring in each batch of negative samples is reduced, so that a loss function is optimized, and the following steps are circulated;

2.2.1) randomly taking T samples from the sample set, expressed as follows:

wherein x_iRepresenting the ith user feature vector in sample T,

y_irepresenting the ith item feature vector in sample T,

r_irepresents the degree of feedback of the ith user in the sample T, and r_i∈[0,1]；

2.2.2) computing y in each sample by using a frequency estimation method based on a Hash sequence_iProbability p of_i；

2.2.2.1) setting arrays A and D with learning rate alpha and size H and a hash function H;

wherein the hash function h may hash each y_iMapping to a range of [0, H]An integer of (d);

2.2.2.2) for each step t ═ 1,2, …, all collections of items in one batch of the training sample number batch

For each article

Comprises the following steps:

2.2.2.2.1)

where ← represents the assignment,

denotes y_iThe time of the last time that it was sampled,

denotes y_iThe number of steps of one pass of the sampled data;

2.2.2.2.2)

2.2.2.3) for each

Probability of sampling

2.2.3) after calculation optimizationLoss function of

And giving a derivation process;

2.2.3.1) gives the vector inner product formula:

2.2.3.2) given a vector

Obtaining one of the M articles

The probability of (c) can be calculated using the softmax function, which is formulated as follows:

wherein e is a natural constant;

2.2.3.3) weighted log-likelihood loss function as:

wherein T represents 2.2.1) the randomly taken T samples;

2.2.3.4) calculating by using a negative sampling algorithm, wherein the negative sampling algorithm adopts a smoothing strategy and can improve the sampling probability of low-frequency samples; sampling negative samples in the same batch; giving a minimum batch of

Wherein the batch-softmax function is:

2.2.3.5) in each batch, the sampling is corrected by frequency, which is considered to be the following equation, because of the power distribution phenomenon, that is, the negative samples are randomly sampled in each batch, which makes the hot goods easy to sample, and the hot goods are excessively punished in the loss function:

s^c(x_i,y_i)＝s(x_i,y_i)-log(p_i)

wherein s is^c(x_i,y_i) Denotes s (x)_i,y_i) Correction value of p_iIs y_iThe probability of (d);

2.2.3.6) the conditional probability function thus modified is:

2.2.3.7) the final loss function is:

2.2.4) updating the parameter theta by adopting a gradient descent algorithm to enable the parameter theta to be close to an optimal value; the gradient descent algorithm machine learns the commonly used algorithm and is used for solving the model parameters.

4. The method for recalling knowledge-graph user preference entity based on double-tower model according to claim 1, wherein the specific process of the step 3) is as follows:

3.1) given a user interaction matrix

Is shown as

A user and the second

The interactive condition of each item, U and V respectively represent a user set and an item set;

the expression of the user interaction matrix Y is as follows:

3.2) definition of O_i,jRepresenting the interaction condition of the user i to the item j, and mapping the item with the interaction of the user i to the knowledge graph

The entity of (1);

3.2.1)O_i,jis an interactive matrix, which is composed of row vectors of users; o is_i,jRepresenting the interaction condition of a user i to an item j, wherein j is the index of the item;

3.2.2) defining a HashMap for storing all articles, wherein the HashMap is stored in a key-value key value pair mode, the key stores the index of the article, and the value stores the entity corresponding to the object;

3.2.3) expressing the interaction condition of a user from the row vector of the user interaction matrix O, and storing the row vector in an array, wherein the array is defined as Ee;

3.2.3) defining a temporary set temp _ set user storage entity, traversing the array E, and if the value of an element in the array is 1, accessing the article HashMap according to the index of the element, acquiring the corresponding entity, and storing the entity in the temporary set temp _ set.

5. The method for recalling knowledge-graph user preference entity based on double-tower model according to claim 1, wherein the specific process of the step 4) is as follows:

4.1) providing a user with the user feature vector v of the first step_user；

4.2) initializing the temporary set temp _ set of the user according to the third step;

4.3) defining a HashMap preferred by a user, namely a user _ map, storing the circulation times of a key value, and storing a triple set by a value;

4.4) defining a set used _ set for storing the recalled triples;

4.5) making the number K of loops equal to 1,2,3 …, giving a maximum value K, and when the number of triples in the user _ map is greater than K, exiting the loop, and executing the following steps:

4.5.1) traverse the entities in temp _ set, remove the entities in used _ set:

temp_set←temp_set-used_set

wherein ← represents valuation, -represents difference set operation;

4.5.2) finding triple sets from the knowledge-graph

And stored in the user _ map, defined as follows:

wherein (h ', r', t ') represents a triplet, h' represents a head entity, r 'represents a relationship, and t' represents a tail entity;

4.5.3) due to the existence of a loop of the knowledge graph, in order to prevent the collection of the recalled entities, the entities in temp _ set need to be added to used _ set;

4.5.4) taking out

Tail entities of middle triplets, deposited into collections

4.5.5) traversal

The corresponding article feature vector v is taken out from the entity in (1)_item；

4.5.6) input parameter v as described by the second step_userAnd v_itemTo the double tower model, the users and the sets are obtained

The probability of the corresponding article in (1), and ordering the entities according to the probability;

4.5.7) given a value τ,0< τ ≦ 1, for determining the number of screening entities:

wherein

Representing the elements in the set sorted by probability and returning an array, getSubVec (i, j) representing the fetching of the child array, i.e. the i to j elements of the original array, newSet () representing the conversion of the array into the set,

represents rounding up;

4.5.8) from

The screened tail entities in the set belong to

The triplet of (2):

4.5.9) method of depositing a triplet set Λ to user _ map: user _ map (k, Λ);

4.5.10) will be

Set overlay to temp _ set;

4.6) and 4.5) finally obtaining the knowledge graph preferred by the user.