WO2022218139A1 - Personalized search method and search system combined with attention mechanism - Google Patents

Abstract

Disclosed in the present invention are a personalized search method and search system combined with an attention mechanism. The search method comprises: 1, collecting and obtaining a large amount of user generation content generated by a user in an Internet information media, and performing vectorization representation; 2, constructing a dominant item group; 3, constructing and training a user preference perception model combined with the attention mechanism, the model being based on a DBN and being composed of three RBMs; 4, constructing a distribution estimation probabilistic model based on a user preference; 5, setting a population size N, and generating N new individuals by using the distribution estimation probabilistic model based on the user preference; 6, selecting N items having the highest similarity to the N new individuals in a search space to form a set of items to be recommended S^u; 7, calculating an adaptive value of each item in S^u, and 8, selecting the Top N items having the highest adaptive values in S^u as the search result, and performing personalized recommendation. In the method, different influences of different decision components on the user preference are considered, so that the user can be helped to perform personalized search more effectively.

Description

Personalized search method and search system fused with attention mechanism technical field

The invention belongs to the technical field of data mining, and particularly relates to a personalized search method and a search system.

Background technique

With the rapid development of technologies such as big data, cloud computing, and the Internet of Things, the scale of the Internet and the number of users have increased dramatically. Users have become active creators of data, gathering a large number of multi-source and heterogeneous user-generated content, and all kinds of information are intricate and presented. Explosive growth. User-generated content contains massive, multi-source, heterogeneous and dynamically evolving complex data. It has the characteristics of diverse sources and structures, sparseness, multi-modality, incompleteness, and social dissemination. It contains rich valuable information and huge Mining potential is also an important source for various Internet platforms and mobile application merchants to obtain information, improve performance and services, and become a typical big data environment. However, while these complex, multi-source and heterogeneous user-generated content bring new information to users, it also increases the difficulty for users to screen, screen and process information and finally make decisions, which brings about the problem of "information overload". As a bridge connecting users and information, personalized search and recommendation algorithms can make full use of massive multi-source heterogeneous user-generated data, predict user behavior and development trends based on users’ potential needs and cognitive preferences, and help users filter from massive amounts of information as much as possible. Content that is in line with user needs and interests and preferences can effectively alleviate "information overload" and improve user experience and the commercial interests of the website platform.

The essence of the personalized search task for user-generated content is to search for optimization goals that meet user needs and personalized preferences in the dynamic evolution space composed of multi-source heterogeneous user-generated data, which is a kind of dynamic qualitative index optimization problem. Due to this kind of complex qualitative index optimization problem, not only its objective function and performance index cannot be accurately described by mathematical functions, but even the decision variables of its optimization problem are no longer simple structured data, and often have greater subjectivity and ambiguity. , uncertainty and inconsistency, users are required to qualitatively analyze, evaluate and make decisions on the items to be searched based on their experience, knowledge and interests, so it is difficult to establish a specific and accurate mathematical model for description. In recent years, the interactive co-evolutionary computing integrated with human intelligence evaluation, which combines the user's subjective cognitive experience, intelligent evaluation decision-making and traditional evolutionary computing, is an effective way to deal with the above-mentioned complex personalized search qualitative index optimization problem.

The Chinese patent with application number CN2020102165574 discloses an interactive personalized search method driven by restricted Boltzmann machine, wherein the construction of the user interest preference model does not consider the different influences of decision variables describing different item attributes on user preference. However, the same weight is used for the decision variables of the items used, which cannot fully reflect the impact of each decision variable on user preferences, so it is difficult to build a more accurate user preference model, which further affects the effect of users' personalized search.

SUMMARY OF THE INVENTION

Purpose of the invention: In view of the problems existing in the prior art, the present invention provides a personalized search method and a search system integrating attention mechanism, wherein the search method takes into account the different influences of different decision components on user preferences, which can help users to better Personalize your search efficiently.

Technical solution: On the one hand, the present invention discloses a personalized search method integrating attention mechanism, including:

Step 1. Collect and obtain user-generated content, which includes all items that user u has evaluated, ratings and textual comments for each item, images of each item, and usefulness of other users’ evaluations of user u. Sexual evaluation score; vectorize text comments, extract features from item images, and obtain feature vectors;

Step 2. The items whose user score is greater than the preset score threshold and whose trust degree is greater than the preset trust degree threshold are formed into an advantageous project group D containing the user's preference; the items in D constitute a set S, S={(u, x _i , C _i ,T _i ,G _i )}, where x _i ∈ D, C _i is the category label vector of item x _i , T _i is the vectorized representation of the user's textual comments on item x _i , and G _i is the image of item x _i Feature vectorized representation, i=1, 2, L, |D|, |D| represents the number of items in D;

Step 3. Build a user preference perception model fused with an attention mechanism. The model is based on a deep belief network and consists of three layers of restricted Boltzmann machines, wherein the visible layer of the first layer of restricted Boltzmann machines includes the first layer of restricted Boltzmann machines. A group of visible units v ₁ , the second group of visible units v ₂ and the third group of visible units v ₃ , the hidden layer is h ₁ ; h ₁ is the visible layer, and the second layer of RBM is formed with the hidden layer h ₂ ; h ₂ is the visible layer layer, and the hidden layer h ₃ constitutes the third layer of RBM; the parameters of the user preference perception model of the fusion attention mechanism are θ={θ ₁ ,θ ₂ ,θ ₃ }={w ₁ ,a ₁ ,b ₁ , w ₂ ,a ₂ ,b ₂ ,w ₃ ,a ₃ ,b ₃ };

Using the dominant project group D, the contrast divergence learning algorithm is used to train the first-layer RBM in the user preference perception model fused with the attention mechanism, and its model parameters θ ₁ ={w ₁ ,a ₁ ,b ₁ } are obtained;

After the training of the first layer of RBM model is completed, when the state of the hidden unit is given, the activation state of each visible unit is independent, and the vector of an item x _i represents [C _i , T _i , G _i ] input to the visible layer, its first The activation probabilities of the visible units in the group, the second group, and the third group are:

where a _1,j , a _1,k and a _1,l represent the first, second and third visible cell offsets, respectively.

Calculate the information entropy of various multi-source heterogeneous data, the information entropy of the item category label is:

The information entropy of the text review vector is:

The information entropy of the item image feature vector is:

where c _ij represents the j-th element of the category label vector C _i of the item x _i , and p(c _ij ) represents the visible unit activation probability corresponding to the j-th element represented by the item category label vector in RBM1;

t _ik represents user u’s textual comments on item xi _i to represent the k-th element of T _i , p(t _ik ) represents the visible unit activation probability corresponding to the k-th element represented by the user text comment vector in RBM1;

g _il represents, p(g _il ) represents the image feature vectorization of item x _i represents the lth element of G _i , p(g _il ) represents the visible unit in RBM1 corresponding to the lth element represented by the item image feature vector activation probability;

Secondly, calculate the proportion of various types of information entropy to the total information entropy as a weight factor:

Wherein H(x _i )=H(C _i )+H(T _i )+H(G _i );

Combining the vectors C _i , T _i , and G _i to form the decision vector Ψ _i of the item x _i is input to each visible unit in v ₁ , v ₂ , and v ₃ , the activation state of each hidden unit in the hidden layer h ₁ is independent, and the first The activation probability of m ₁ hidden unit is:

Among them, m ₁ =1,2,...,M ₁ ,

is the bias of the _m1th hidden unit in _h1 ; _v1j is the state of the jth visible unit in the first group of visible units v1 of _RMB1 ; _v2k is the kth visible unit in the _second group of visible units v2 of RMB1 The state of the unit; the state of the lth visible unit in the _third group of visible units v3 in v _3l RMB1; w _{1, n, m1} are the element values in w ₁ , indicating the nth visible unit in RBM1 and the m _1th unit Connection weight between hidden units, n=1,2,...,Ф;

Represents the state of the m1th hidden unit in the hidden layer h1; σ(x)= ₁ /( ₁ +exp(-x)) is the sigmoid activation function;

After the training of RBM1 is completed, the state of each hidden unit corresponding to item x _i is obtained according to formula (9), and then the user's preference for each decision component of each item in the dominant item group D is obtained, that is, the activation probability of visible layer unit, as attention Force weight coefficient at _n (x _i ):

in

Represents Ψ _i as the state of each visible unit in the visible layer of RBM1, the state of the m ₁ hidden unit in the hidden layer h ₁ ; at _n ( _xi ) represents the attention weight of each decision component ψ _in of item x _i ;

Taking the attention weight coefficient at _n ( _xi ) as the weight coefficient of each decision component of the item _xi , the item _xi in the dominant item group D is coded based on the attention mechanism, and expressed as x _ati after coding:

x _ati =Ψ _i +at _n (x _i )×Ψ _i (12)

Input x _ati into the pre-trained RBM1 to get the visible unit activation probability V _RBM1 (x _ati ):

where x _atn' is the n'th element of x _ati ;

The self-attention mechanism operation is performed by the visible unit activation probability V _RBM1 (x _ati ) of RBM1, and the user preference attention weight vector A(x _ati ) of the dynamic learning project individual is:

A(x _ati )=softmax(a(V _RBM1 (x _ati ),w ₁ )) (14)

Among them, the softmax() function ensures that the sum of all weight coefficients is 1; the function a(V _RBM1 (x _ati ),w ₁ ) measures the attention weight coefficient of item _xi relative to user preference features, and is calculated as follows:

a(V _RBM1 (x _ati ), w ₁ )=V _RBM1 (x _ati )·(w ₁ ) ^T (15)

Combining the user preference attention weight vector A(x _ati ) and the original decision vectors C _i , T _i , G _i of the item _xi , generate the item decision vector fused with the attention mechanism:

x _i ′=A(x _ati )×Ψ _i (16)

The item decision vector x _i ′ fused with the attention mechanism is used to form the training set, and the RBM1, RBM2, and RBM3 models in the DBN are trained layer by layer. model parameter θ;

Step 4. According to the trained DBN-based user preference perception model and its model parameters that integrate the attention mechanism, establish and construct a distribution estimation probability model P(x) based on user preference:

P(x)=[P(ψ ₁ ),P(ψ ₂ ),L,P(ψ _n ),L,P(ψ _Ф )] (17)

where (ψ ₁ ,ψ ₂ ,…,ψ _n ,…,ψ _Ф ) is the original decision vector of item x, and P(ψ _n ) represents the probability of the nth decision component of the item preferred by the user;

Step 5. Set the population size N, use the distribution based on user preference to estimate the probability model P(x), and use the distribution estimation algorithm to generate N new individuals, each individual is an item; the category label vector of the vth new individual

The setting steps are as follows:

(5.1) Let v=1;

(5.2) Generate a random number z between [0, 1]; if z≤P(ψ _j =1), then the class label vector of the vth new individual

The jth element of is 1, otherwise it is 0;

(5.3) add one to v, and repeat step (5.2) until v>N;

Step 6. Select and N new individual category label vectors in the search space

The N items with the highest similarity constitute a set of items to be recommended S ^u ;

Step 7. Calculate the fitness value of each item in the item set ^Su to be recommended

in,

and

respectively represent the maximum and minimum value of the item energy function in the item set ^Su to be recommended;

is the energy function of item x ^* , x ^* ∈ S ^u , which is calculated as:

in

is the nth decision component of item x ^* ;

Step 8. Select the top N items with the highest fitness value in Su as the search result, ^TopN <N;

With the advancement of the user's interactive search process and the dynamic evolution of user behavior, according to the current user's recent evaluation data, the dominant item group D is updated, the user preference perception model fused with the attention mechanism is retrained, and the extracted user preference features are dynamically updated. , update the estimated probability model P(x) based on the distribution of user preferences.

On the other hand, the present invention also discloses a search system for realizing the above-mentioned personalized search method, including:

The user-generated content acquisition module is used to collect and acquire user-generated content, which includes all items that user u has evaluated, ratings and text comments for each item, images of each item, and user-generated content from other users. u The usefulness evaluation score of the evaluation; vectorize the text comment, extract the feature of the item image, and obtain the feature vector;

The advantageous project group building module is used to form the advantageous project group D with user preference of the projects whose user score is greater than the preset score threshold and whose trust degree is greater than the preset trust degree threshold;

The user preference perception model construction and training module is used to construct and train a user preference perception model fused with an attention mechanism; the model is based on a deep belief network and consists of three layers of restricted Boltzmann machines, of which the first layer is restricted The visible layer of the Boltzmann machine includes the first group of visible units v ₁ , the second group of visible units v ₂ and the third group of visible units v ₃ , the hidden layer is h ₁ ; h ₁ is used as the visible layer, and the hidden layer h ₂ Constitute the second-layer restricted Boltzmann machine; h ₂ as the visible layer, and the hidden layer h ₃ form the third-layer restricted Boltzmann machine; the parameter of the user preference perception model of the fusion attention mechanism is θ ={θ ₁ ,θ ₂ ,θ ₃ }={w ₁ ,a ₁ ,b ₁ ,w ₂ ,a ₂ ,b ₂ ,w ₃ ,a ₃ ,b ₃ };

The distribution estimation probability model building module based on user preference is used to build a user preference-based distribution estimation probability model P(x ):

P(x)=[P(ψ ₁ ),P(ψ ₂ ),L,P(ψ _n ),L,P(ψ _Ф )] (17)

where (ψ ₁ ,ψ ₂ ,…,ψ _n ,…,ψ _Ф ) is the original decision vector of item x, and P(ψ _n ) represents the user’s preference probability for the nth decision component of the item;

The population generation module is used to estimate the probability model P(x) based on the distribution based on user preferences, use the distribution estimation algorithm to generate N new individuals, each individual is an item, and set the category label vector of each new individual, N is preset population size;

The building block of the item set to be recommended is used to select and N new individual category label vectors in the search space

The N items with the highest similarity constitute a set of items to be recommended S ^u ;

The fitness value calculation module is used to calculate the fitness value of each item in the item set ^Su to be recommended;

The search result selection module is used to select the top ^TopN items with the highest fitness value in Su as the search result, TopN<N.

Beneficial effects: The personalized search method disclosed in the present invention makes full use of multi-source heterogeneous user-generated content, including user ratings, item category labels, user text comments, evaluation trust and item image information, and constructs user preference perception fused with attention mechanism Model, based on this user preference perception model, constructs a distribution estimation probability model based on user preference, generates new feasible solution items containing user preference, and selects multiple items with the highest fitness value as the final search result. This method can well handle the personalized search task of multi-source heterogeneous user-generated content in the big data environment, effectively guide users to conduct personalized search, help users search for satisfactory solutions as soon as possible, and improve the comprehensive performance of personalized search algorithms.

Description of drawings

Fig. 1 is the flow chart of the personalized search method disclosed by the present invention fused with attention mechanism;

FIG. 2 is a schematic structural diagram of a user preference perception model fused with an attention mechanism;

Figure 3 is a schematic diagram of the composition of a personalized search system incorporating an attention mechanism.

Detailed ways

The present invention will be further explained below in conjunction with the accompanying drawings and specific embodiments.

As shown in Figure 1, the present invention discloses a personalized search method integrating attention mechanism, including:

Step 1. Collect and obtain user-generated content, which includes all items that user u has evaluated, ratings and textual comments for each item, images of each item, and usefulness of other users’ evaluations of user u. Sexual evaluation score; vectorize text comments, extract features from item images, and obtain feature vectors;

The steps for the vectorized representation of text comments in this embodiment are: removing stop words and punctuation marks in the text comments, and performing data preprocessing; using documents: Devlin J, Chang M W, Lee K, et al.BERT: The BERT model in Pre-training of Deep Bidirectional Transformers for Language Understanding[J].arXiv:1810.04805v2[cs.CL]24 May 2019. is a vectorized representation of user text comments.

The feature extraction of the project image is to use the literature: Krizhevsky A, Sutskever I, Hinton G E. Image Net classification with deep convolutional neural networks. In: Proceedings of the 25th International Conference on Neural Information Processing Systems. Lake Tahoe, Nevada, USA: Curran Associates Inc., 2012.1097-1105. The AlexNet model in the project image is feature extraction and vectorized representation.

The usefulness judgement of other users’ evaluations on user u means that other users make useful judgements on the current user u’s evaluation information about a certain item. The user's evaluation of the current user u's evaluation information on a certain item, the total number marked as 1 is the usefulness evaluation score of the evaluation of the user u by other users. For example, the current user u has made an evaluation on item x, and user A and user B have made a usefulness judgment on the evaluation, which reflects the credibility of the current user's evaluation of item x. u Judging the usefulness of the evaluation of item x, you can filter invalid evaluations or fake reviews.

The ratio of the usefulness evaluation score of the evaluation made by other users to user u to the total number of evaluation items of user u is the trust degree of user u to the evaluation of the item.

Step 2. Construct the advantageous project group D that the user prefers;

An item whose user rating is greater than the preset rating threshold and whose trust degree is greater than the preset trust degree threshold is an item preferred by the user. Due to the characteristics of users' ambiguity, uncertainty and dynamic changes, this embodiment introduces a certain randomness into the existing user preference item groups, so as to increase the user's selection range, so that the user's selection is not too limited to the current Within the range of preference information, it adapts to the actual situation of the environment and the dynamic variability of user preferences. Thereby, the items whose scores are greater than the preset scoring threshold and whose trust degree is greater than the preset trust degree threshold, and multiple new items randomly sampled in the search space, form a dominant item group D. The new items added to the dominant item group D may or may not contain user preferences, and are random, which increases the diversity of the item group. The proportion of new projects in the advantageous project group D does not exceed 30%. In this embodiment, the new projects account for 10% of the total number of projects in the advantageous project group D.

Since new items are randomly sampled in the search space, the current user u may or may not have rated them. If the current user u has no comments on the new item, the text comments on the new item by similar users u' of the current user u are used as the evaluation of the new item by user u; if multiple similar users of user u all share the new item When making an evaluation, the evaluation of the user with the greatest similarity to user u is selected. If the similar users of the current user u do not evaluate the new item, the user u's evaluation of the new item adopts the method of random assignment.

Similar users of user u are users who have a common rating item with user u and whose similarity is greater than a preset similarity threshold. For user u' that has a common rating item with user u, u'≠u, the similarity Sim(u, u') of u and u' is:

where I _u,u' represents the set of items scored by both users u and u'; R _ux' is the user u's rating on the item x' in I _{u, u'} , and R _u'x' is the user u' to x'rating;

is the average rating of all items evaluated by user u;

is the average rating of all items evaluated for user u'.

The dominant item group D constitutes a set S, S={(u, x _i , C _i , T _i , G _i )}, where x _i ∈ D, C _i is the category label vector of item x _i , and the length is the total number of categories n ₁ ; each element c _ij in C _i is a binary variable; c _ij =1 indicates that item x _i has the j-th class label, j=1, 2, L, n ₁ ; and the labels of different classes are not mutually exclusive, Multiple category labels can exist for an item at the same time. T _i is the vectorized representation of the text comments of the user on the item _xi , the length is n ₂ ; G _i is the vectorized representation of the image feature of the item _xi , the length is n ₃ ; i=1,2,L,|D|, |D| represents the number of items in D.

The vectors C _i , T _i , and G _i are combined into a vector Ψ _i with a length of Ф, which constitutes the original decision vector of the item _xi , and each element ψ _in is the decision component of the item _xi , Ф=n ₁ +n ₂ +n ₃ , n=1, 2, . . . , Φ.

Step 3. Construct a user preference perception model fused with the attention mechanism, as shown in Figure 2, the model is based on the Deep Belief Network (DBN), and the model consists of a three-layer Restricted Boltzmann Machine (Restricted Boltzmann Machine). , RBM), wherein the visible layer of the first layer of restricted Boltzmann machine RBM1 includes the first group of visible units v ₁ , the second group of visible units v ₂ and the third group of visible units v ₃ , and the hidden layer is h ₁ ; The _first group of visible unit v1 has _n1 units, and each unit is a binary variable; the _second and _third groups of visible units v2 and _v3 have _n2 and n3 units respectively, and each unit is Real variable; h ₁ as the visible layer, and the hidden layer h ₂ form the second-layer Restricted Boltzmann Machine RBM2; h ₂ as the visible layer, and the hidden layer h ₃ form the third-layer Restricted Boltzmann Machine RBM3 . h ₁ , h ₂ , and h ₃ respectively have M ₁ , M ₂ and M ₃ hidden units, and each hidden unit is a real variable; for each RBM, the number of hidden units is selected to be 0.8-1.2 times the total number of visible units. In the example, it is set to 0.8 times. Thus, the number M ₁ of hidden units in h ₁ is:

Φ=n ₁ +n ₂ +n ₃ ,

is an upward rounding operation; the number M ₂ of hidden units in h ₂ is:

The number M ₃ of hidden units in h ₃ is:

The parameters of the user preference perception model fused with the attention mechanism are θ={θ ₁ ,θ ₂ ,θ ₃ }={w ₁ ,a ₁ ,b ₁ ,w ₂ ,a ₂ ,b ₂ ,w ₃ ,a ₃ , b ₃ }, where {w ₁ ,a ₁ ,b ₁ }, {w ₂ ,a ₂ ,b ₂ } and {w ₃ ,a ₃ ,b ₃ } represent the model parameters of RBM1, RBM2, RBM3, respectively, w _τ represents the connection weight between the visible unit and the hidden unit of the τ-th layer RBM; a _τ and b _τ represent the bias of the visible unit and the hidden unit of the τ-th layer RBM, respectively; τ ∈ {1, 2, 3}.

Using the dominant project group D, the contrastive divergence learning algorithm is used to train the first-layer restricted Boltzmann machine RBM1 in the user preference perception model fused with the attention mechanism, and its model parameters θ ₁ ={w ₁ ,a ₁ ,b ₁ }. In this step, only RBM1 is trained, which can be considered as pre-training of RBM1. In subsequent steps, RBM1, RBM2, and RBM3 will be trained layer by layer again. The decision vector Ψ _i of item x _i is composed of C _i , T _i , and G _i , and C _i , T _i , and G _i contain different user preference information. For example, the length n ₁ of the category label vector C _i is usually less than The image feature vectorization of the item represents the length n ₃ of G _i ; if each component in the decision vector of the item is treated equally, the data containing more information will flood the data containing less preference information, and this kind of preference Data with less information is a useful supplement for building user preference perception models and cannot be ignored. Therefore, the present invention considers the information entropy represented by each data type, and uses weights to adjust the components of various types of multi-source heterogeneous data input to the visible layer neural units of the user preference perception model, so as to ensure that all types of data can contribute to the construction of the user preference perception model. make an effective contribution.

After the training of the first layer of RBM model is completed, when the state of the hidden unit is given, the activation state of each visible unit is independent, and the vector of an item x _i represents [C _i , T _i , G _i ] input to the visible layer, its first The activation probabilities of the visible units in the group, the second group, and the third group are:

Among them, a _1,j , a _1,k and a _1,l represent the first, second and third group of visible unit offsets, respectively, a _1,j , a _1,k , a _1,l are combined as a ₁ , j=1,2,L,n ₁ ,k=1,2,...,n ₂ ,l=1,2,...,n ₃ .

According to the information entropy formula:

Calculate the information entropy of various multi-source heterogeneous data,

The information entropy of the item category label is:

The information entropy of the text review vector is:

The information entropy of the item image feature vector is:

where c _ij represents the j-th element of the category label vector C _i of the item x _i , and p(c _ij ) represents the visible unit activation probability corresponding to the j-th element represented by the item category label vector in RBM1;

t _ik represents user u’s textual comments on item xi _i to represent the k-th element of T _i , p(t _ik ) represents the visible unit activation probability corresponding to the k-th element represented by the user text comment vector in RBM1;

g _il represents, p(g _il ) represents the image feature vectorization of item x _i represents the lth element of G _i , p(g _il ) represents the visible unit in RBM1 corresponding to the lth element represented by the item image feature vector activation probability;

Secondly, the proportion of various information entropy to total information entropy is further calculated as a weight factor:

Wherein H(x _i )=H(C _i )+H(T _i )+H(G _i );

When the state of the visible unit is given, that is, the decision vector Ψ _i of the item x _i is formed by combining the vectors C _i , T _i , and G _i into the visible units in v ₁ , v ₂ , and v ₃ , each unit in the hidden layer h ₁ The activation states of the hidden units are conditionally independent, and the activation probability of the _m1th hidden unit is:

Among them, m ₁ =1,2,...,M ₁ ,

is the bias of the _m1th hidden unit in _h1 ; _v1j is the state of the jth visible unit in the first group of visible units v1 of _RMB1 , that is, the value of the jth element of C _i ; _v2k is the second value of RMB1 The state of the k-th visible unit in the group visible unit v ₂ , that is, the value of the k-th element of Τ _i ; the state of the l-th visible unit in the third group of visible units v ₃ of v ₃₁ RMB1, that is, the l-th element of G _i the value of;

is the element value in w ₁ , indicating the connection weight between the nth visible unit and the _m1th hidden unit in RBM1, n=1,2,...,Ф;

represents the state of the m1th hidden unit in the hidden layer h1; σ(x)= ₁ /( ₁ +exp(-x)) is the sigmoid activation function.

When the hidden unit state is given, the activation state of each visible unit is also conditionally independent, and the activation probability of the nth visible unit is:

where a _1,n represents the bias of the nth visible unit in the visible layer.

After the training of RBM1 is completed, the state of each hidden unit corresponding to item x _i can be obtained according to formula (5), and then the user's preference for each decision component of each item in the dominant project group D can be obtained, that is, the activation probability of the visible layer unit, As the attention weight coefficient at _n (x _i ):

in

Represents Ψ _i as the state of each visible unit in the visible layer of RBM1, the state of the m ₁ hidden unit in the hidden layer h ₁ ; at _n ( _xi ) represents the attention weight of each decision component ψ _in of item x _i , which reflects the self adaptive characteristics.

Taking the attention weight coefficient at _n ( _xi ) as the weight coefficient of each decision component of the item _xi , the item _xi in the dominant item group D is coded based on the attention mechanism, and expressed as x _ati after coding:

x _ati =Ψ _i +at _n (x _i )×Ψ _i (12)

where i=1,2,L,|D|;

Input x _ati into the pre-trained RBM1 to get the visible unit activation probability V _RBM1 (x _ati ):

where x _atn' is the n'th element of x _ati .

Equation (9) actually nests the activation probability of the hidden unit and the activation probability of the visible unit, namely:

Use the visible unit activation probability V _RBM1 (x _ati ) in the obtained RBM1 model and the literature: Li J, Wang Y, Mcauley J. Time Interval Aware Self-Attention for Sequential Recommendation. In: WSDM'20: The Thirteenth ACM International Conference on The self-attention mechanism proposed in Web Search and Data Mining.ACM, 2020. uses the RBM1 visible unit activation probability V _RBM1 (x _ati ) to perform the self-attention mechanism operation, and dynamically learns the individual user preference attention weight vector A ( x _ati ):

A(x _ati )=softmax(a(V _RBM1 (x _ati ),w ₁ )) (14)

Among them, the softmax() function guarantees that the sum of all weight coefficients is 1. The function a(V _RBM1 (x _ati ),w ₁ ) measures the attention weight coefficient of item _xi relative to user preference features, and is calculated as follows:

a(V _RBM1 (x _ati ), w ₁ )=V _RBM1 (x _ati )·(w ₁ ) ^T (15)

Combining the user preference attention weight vector A(x _ati ) and the original decision vectors C _i , T _i , G _i of the item _xi , generate the item decision vector fused with the attention mechanism:

x _i ′=A(x _ati )×Ψ _i (16)

Using the item decision vector x _i ' of the fusion attention mechanism to form a training set, the RBM1, RBM2, and RBM3 models in the DBN are trained layer by layer. First, the RBM1 is trained to obtain parameters {w ₁ , a ₁ , b ₁ }; ₁ pass into a ₂ in RBM2, train RBM2 on this basis, and obtain optimization parameters {w ₂ , a ₂ , b ₂ }; pass b ₂ into a ₃ in RBM3, train RBM3 on this basis, and obtain optimization parameters {w ₃ , a ₃ , b ₃ }; thus, the three-layer RBM models in the DBN network influence and correlate with each other, forming a network as a whole. After the training is completed, the DBN-based user preference perception model fused with the attention mechanism and its optimized model parameters θ are obtained.

The DBN model training method here is an improved DBN model training method based on the attention mechanism. The purpose is to better use the adaptive weight information to extract user preference features, focus on important features, and be more appropriate. It expresses the influence of different types of attribute decision components of each item on user preference characteristics in practical application scenarios, and expresses user preference characteristics more precisely.

Step 4. According to the trained DBN-based user preference perception model and its model parameters that integrate the attention mechanism, establish and construct a distribution estimation probability model P(x) based on user preference:

P(x)=[P(ψ ₁ ),P(ψ ₂ ),L,P(ψ _n ),L,P(ψ _Ф )] (17)

where (ψ ₁ ,ψ ₂ ,…,ψ _n ,…,ψФ) is the original decision vector of the item x, and P(ψ _n ) represents the user’s preference probability for the nth decision component of the item, which is calculated as follows:

First, calculate the probability distribution model p(x) based on user preference according to the dominant project group D:

p(x) is a Ф-dimensional vector, and its n-th element p(ψ _n ) is the activation probability of the n-th decision component of the user preference item; a lower bound constraint is applied to p(ψ _n ), and the constrained value is the user preference The probability P(ψ _n ) of the nth decision component of the item, namely:

ε is the preset lower bound threshold. In this embodiment, ε=0.1, that is, for the decision component whose activation probability calculated according to formula (18) is less than 0.1, the activation probability value is set to 0.1; this constraint considers the activation of the decision component When the probability is small, the decision component is randomly sampled with a certain probability value to enhance the diversity of the generated population and prevent the evolutionary optimization algorithm from prematurely converging and missing the optimal solution.

Step 5. Set the population size N, use the distribution based on user preference to estimate the probability model P(x), and use the distribution estimation algorithm (Estimation of Distribution Algorithms, EDA) to generate N new individuals, each individual is an item; vth class label vector for new individuals

The setting steps are as follows:

(5.1) Let v=1;

(5.2) Generate a random number z between [0, 1]; if z≤P(ψ _j =1), then the class label vector of the vth new individual

The jth element of is 1, otherwise it is 0;

(5.3) add one to v, and repeat step (5.2) until v>N;

Step 6. Select and N new individual category label vectors in the search space

The N items with the highest similarity constitute the item set S ^u to be recommended; in this embodiment, the Euclidean distance is used as the similarity calculation, that is, the smaller the Euclidean distance between the two vectors, the higher the similarity between the two;

Step 7. Calculate the fitness value of each item in the item set ^Su to be recommended:

In the present invention, the fitness value of the item is calculated based on the energy function, and the fitness value of the item x ^* in the recommended item set S ^u is treated.

is calculated as follows:

in,

and

respectively represent the maximum and minimum value of the item energy function in the item set ^Su to be recommended;

is the energy function (x ^* ∈ S ^u ) of item x ^* , which is calculated as:

where a _1,n represents the bias of the nth visible unit in the visible layer of RBM1,

is the nth decision component of item x ^* ,

is the bias of the _m1th hidden unit in _h1 ,

is the element value in w ₁ , indicating the connection weight between the nth visible unit and the _m1th hidden unit in RBM1.

Step 8. Select the top ^TopN items with the highest fitness value in Su as the search result, TopN<N.

Due to the dynamic evolution characteristics of multi-source heterogeneous user-generated content and the uncertainty of user interest preferences, in the early stage of the personalized evolution search process, the user preference information contained in the dominant item group D is not sufficient, so users trained based on this The user preference features extracted by the preference-aware model are relatively rough. With the advancement of the user interactive search process and the dynamic evolution of user behavior, according to the recent evaluation data of the current user, update the dominant project group D, retrain the user preference perception model integrated with the attention mechanism, and dynamically update the extracted user preference features. Track user preference changes; at the same time, update the distribution estimation probability model P(x) based on user preferences to effectively guide the direction of personalized evolutionary search, help users find user-satisfying solutions as soon as possible, and successfully complete personalized search tasks in complex environments.

This embodiment also discloses a personalized search system that realizes the above-mentioned personalized search method and integrates the attention mechanism, as shown in FIG. 3 , including:

User-generated content acquisition module 1, used to collect and acquire user-generated content, which includes all items that user u has evaluated, ratings and text comments for each item, images of each item, and comments from other users. The usefulness evaluation score of the evaluation made by the user u; vectorize the text comments, extract the feature of the item image, and obtain the feature vector;

The advantageous project group building module 2 is used to form the advantageous project group D containing the user's preference with the projects whose user score is greater than the preset score threshold and the trust degree is greater than the preset trust degree threshold;

The user preference perception model construction and training module 3 is used to construct and train the user preference perception model fused with the attention mechanism according to step 3; the model is based on a deep belief network and consists of three layers of restricted Boltzmann machines, in which the first The visible layer of a restricted Boltzmann machine includes the first group of visible units v ₁ , the second group of visible units v ₂ and the third group of visible units v ₃ , and the hidden layer is h _{1 ;} The hidden layer h ₂ constitutes the second-layer restricted Boltzmann machine; h ₂ serves as the visible layer, and the hidden layer h ₃ constitutes the third-layer restricted Boltzmann machine; the user preference perception model of the fusion attention mechanism The parameters of θ={θ ₁ ,θ ₂ ,θ ₃ }={w ₁ ,a ₁ ,b ₁ ,w ₂ ,a ₂ ,b ₂ ,w ₃ ,a ₃ ,b ₃ };

The distribution estimation probability model building module 4 based on user preference is used to build a user preference-based distribution estimation probability model P( x):

P(x)=[P(ψ ₁ ),P(ψ ₂ ),L,P(ψ _n ),L,P(ψ _Ф )] (17)

where (ψ ₁ ,ψ ₂ ,…,ψ _n ,…,ψ _Ф ) is the original decision vector of item x, and P(ψ _n ) represents the user’s preference probability for the nth decision component of the item;

The population generation module 5 is used to estimate the probability model P(x) by using the distribution based on the user preference, use the distribution estimation algorithm to generate N new individuals, each individual is an item, and set the category label vector of each new individual, N is the preset population size;

Item set to be recommended building module 6, used to select and N new individual category label vectors in the search space

The N items with the highest similarity constitute a set of items to be recommended S ^u ;

The fitness value calculation module 7 is used to calculate the fitness value of each item in the item set ^Su to be recommended according to step 7;

The search result selection module 8 is used to select the top ^TopN items with the highest fitness value in Su as the search result, TopN<N.

Claims (10)

1. The personalized search method fused with attention mechanism is characterized by including:

   Step 1. Collect and obtain the content generated by user u, which includes all items that user u has evaluated, ratings and text comments for each item, images of each item, and the evaluation of user u by other users. Usefulness evaluation score; vectorize text comments, extract features from item images, and obtain feature vectors;

   Step 2. The items whose user score is greater than the preset score threshold and whose trust degree is greater than the preset trust degree threshold are formed into an advantageous project group D containing the user's preference; the items in D constitute a set S, S={(u, x _i , C _i ,T _i ,G _i )}, where x _i ∈ D, C _i is the category label vector of item x _i , T _i is the vectorized representation of the user's textual comments on item x _i , and G _i is the image of item x _i Feature vectorized representation, i=1, 2, L, |D|, |D| represents the number of items in D;

   Step 3. Build a user preference perception model fused with an attention mechanism. The model is based on a deep belief network and consists of three layers of restricted Boltzmann machines, wherein the visible layer of the first layer of restricted Boltzmann machines includes the first layer of restricted Boltzmann machines. A group of visible units v ₁ , the second group of visible units v ₂ and the third group of visible units v ₃ , the hidden layer is h ₁ ; h ₁ as the visible layer, and the hidden layer h ₂ constitute the second layer of restricted Boltzmann machine; h ₂ as the visible layer, and the hidden layer h ₃ constitute the third-layer restricted Boltzmann machine; the parameters of the user preference perception model of the fusion attention mechanism are θ={θ ₁ , θ ₂ , θ ₃ }={w ₁ ,a ₁ ,b ₁ ,w ₂ ,a ₂ ,b ₂ ,w ₃ ,a ₃ ,b ₃ };

   Using the dominant project group D, the contrastive divergence learning algorithm is used to train the first-layer restricted Boltzmann machine in the user preference perception model fused with the attention mechanism, and its model parameters θ ₁ ={w ₁ ,a ₁ ,b ₁ };

   After the training of the first layer of RBM model is completed, when the state of the hidden unit is given, the activation state of each visible unit is independent, and the vector of an item x _i represents [C _i , T _i , G _i ] input to the visible layer, its first The activation probabilities of the visible units in the group, the second group, and the third group are:

   

   

   

   Among them, a _1,j , a _1,k and a _1,l represent the first group, the second group and the third group of visible unit offsets, respectively;

   Calculate the information entropy of various multi-source heterogeneous data, the information entropy of the item category label is:

   

   The information entropy of the text review vector is:

   

   The information entropy of the item image feature vector is:

   

   where c _ij represents the j-th element of the category label vector C _i of the item x _i , and p(c _ij ) represents the visible unit activation probability corresponding to the j-th element represented by the item category label vector in RBM1;

   t _ik represents user u’s textual comments on item xi _i to represent the k-th element of T _i , p(t _ik ) represents the visible unit activation probability corresponding to the k-th element represented by the user text comment vector in RBM1;

   g _il represents, p(g _il ) represents the image feature vectorization of item x _i represents the lth element of G _i , p(g _il ) represents the visible unit in RBM1 corresponding to the lth element represented by the item image feature vector activation probability;

   Secondly, calculate the proportion of various types of information entropy to the total information entropy as a weight factor:

   

   Wherein H(x _i )=H(C _i )+H(T _i )+H(G _i );

   Combining the vectors C _i , T _i , and G _i to form the decision vector Ψ _i of the item x _i is input to each visible unit in v ₁ , v ₂ , and v ₃ , the activation state of each hidden unit in the hidden layer h ₁ is independent, and the first The activation probability of m ₁ hidden unit is:

   

   Among them, m ₁ =1,2,...,M ₁ ,

   

   is the bias of the _m1th hidden unit in _h1 ; _v1j is the state of the jth visible unit in the first group of visible units v1 of _RMB1 ; _v2k is the kth visible unit in the _second group of visible units v2 of RMB1 The state of the unit; the state of the lth visible unit in the _third group of visible units v3 of v _3l RMB1;

   

   is the element value in w ₁ , indicating the connection weight between the nth visible unit and the m1th hidden unit in RBM1, n= ₁ , 2,...,Φ;

   

   Represents the state of the m1th hidden unit in the hidden layer h1; σ(x)= ₁ /( ₁ +exp(-x)) is the sigmoid activation function;

   After the training of RBM1 is completed, the state of each hidden unit corresponding to item x _i is obtained according to formula (9), and then the user's preference for each decision component of each item in the dominant item group D is obtained, that is, the activation probability of visible layer unit, as attention Force weight coefficient at _n (x _i ):

   

   in

   

   Represents Ψ _i as the state of each visible unit in the visible layer of RBM1, the state of the m ₁ hidden unit in the hidden layer h ₁ ; at _n ( _xi ) represents the attention weight of each decision component ψ _in of item x _i ;

   Taking the attention weight coefficient at _n ( _xi ) as the weight coefficient of each decision component of the item _xi , the item _xi in the dominant item group D is coded based on the attention mechanism, and expressed as x _ati after coding:

   x _ati =Ψ _i +at _n (x _i )×Ψ _i (12)

   Input x _ati into the pre-trained RBM1 to get the visible unit activation probability V _RBM1 (x _ati ):

   

   where x _atn' is the n'th element of x _ati ;

   The self-attention mechanism operation is performed by the visible unit activation probability V _RBM1 (x _ati ) of RBM1, and the user preference attention weight vector A(x _ati ) of the dynamic learning project individual is:

   A(x _ati )=softmax(a(V _RBM1 (x _ati ),w ₁ )) (14)

   Among them, the softmax() function ensures that the sum of all weight coefficients is 1; the function a(V _RBM1 (x _ati ),w ₁ ) measures the attention weight coefficient of item _xi relative to user preference features, and is calculated as follows:

   a(V _RBM1 (x _ati ), w ₁ )=V _RBM1 (x _ati )·(w ₁ ) ^T (15)

   Combining the user preference attention weight vector A(x _ati ) and the original decision vectors C _i , T _i , G _i of the item _xi , generate the item decision vector fused with the attention mechanism:

   x _i ′=A(x _ati )×Ψ _i (16)

   The item decision vector x _i ′ fused with the attention mechanism is used to form a training set, and the RBM1, RBM2, and RBM3 models in the DBN are trained layer by layer. Its optimized model parameter θ;

   Step 4. According to the trained user preference perception model based on the deep belief network and the model parameters of the fusion attention mechanism, establish and construct a distribution estimation probability model P(x) based on user preference:

   P(x)=[P(ψ ₁ ),P(ψ ₂ ),L,P(ψ _n ),L,P(ψ _Φ )] (17)

   where (ψ ₁ ,ψ ₂ ,…,ψ _n ,…,ψ _Φ ) is the original decision vector of item x, and P(ψ _n ) represents the user’s preference probability for the nth decision component of the item;

   Step 5. Set the population size N, use the distribution based on user preference to estimate the probability model P(x), and use the distribution estimation algorithm to generate N new individuals, each individual is an item; the category label vector of the vth new individual

   

   (v=1,2,L,N) The setting steps are as follows:

   (5.1) Let v=1;

   (5.2) Generate a random number z between [0, 1]; if z≤P(ψ _j =1), then the class label vector of the vth new individual

   

   The jth element of is 1, otherwise it is 0;

   (5.3) add one to v, and repeat step (5.2) until v>N;

   Step 6. Select and N new individual category label vectors in the search space

   

   The N items with the highest similarity constitute a set of items to be recommended S ^u ;

   Step 7. Calculate the fitness value of each item in the item set ^Su to be recommended

   

   

   in,

   

   and

   

   respectively represent the maximum and minimum value of the item energy function in the item set ^Su to be recommended;

   

   is the energy function of item x ^* , x ^* ∈ S ^u , which is calculated as:

   

   in

   

   is the nth decision component of item x ^* ;

   Step 8. Select the top N items with the highest fitness value in Su as the search result, ^TopN <N;

   With the advancement of the user's interactive search process and the dynamic evolution of user behavior, according to the current user's recent evaluation data, the dominant item group D is updated, the user preference perception model fused with the attention mechanism is retrained, and the extracted user preference features are dynamically updated. , update the estimated probability model P(x) based on the distribution of user preferences.
2. The personalized search method incorporating an attention mechanism according to claim 1, wherein the advantageous item group D further includes a new item with a proportion of n, and the new item is obtained by random sampling in the search space.
3. The personalized search method fused with attention mechanism according to claim 2, wherein if the current user u has no comments on the new item, the text comment on the new item by the similar user u' of the current user u is used as the user u's evaluation of the new item; if multiple similar users of user u have evaluated the new item, select the evaluation of the user with the greatest similarity with user u; The project is evaluated, and the user u's evaluation of the new project adopts the method of random assignment.
4. The personalized search method fused with attention mechanism according to claim 3, characterized in that, similar users of user u are users who have a common scoring item with user u, and the similarity is greater than a preset similarity threshold; User u has a user u' with a common rating item, u'≠u, the similarity Sim(u, u') of u and u' is:

   

   where I _u,u' represents the set of items scored by both users u and u'; R _ux' is the user u's rating on the item x' in I _{u, u'} , and R _u'x' is the user u' to x'rating;

   

   is the average rating of all items evaluated by user u;

   

   is the average rating of all items evaluated for user u'.
5. The personalized search method fused with the attention mechanism according to claim 1, wherein the RBM1, RBM2, and RBM3 models in the DBN are trained layer by layer, specifically:

   First train RBM1 to obtain parameters {w ₁ , a ₁ , b ₁ }; pass b ₁ into a ₂ in RBM2, train RBM2 on this basis, and obtain optimized parameters {w ₂ , a ₂ , b ₂ }; b ₂ is passed into a ₃ in RBM3, and RBM3 is trained on this basis to obtain optimized parameters {w ₃ , a ₃ , b ₃ }.
6. The personalized search method fused with attention mechanism according to claim 1, wherein the calculation of the probability P(ψ _n ) of the nth decision component of the item preferred by the user is:

   First, calculate the probability distribution model p(x) based on user preference according to the dominant project group D:

   

   p(x) is a Φ-dimensional vector, and its n-th element p(ψ _n ) is the activation probability of the n-th decision component of the user preference item; the lower bound is constrained on p(ψ _n ), and the constrained value is the user's preference The probability P(ψ _n ) of the nth decision component of the item, namely:

   

   ε is a preset lower bound threshold.
7. The personalized search method fused with attention mechanism according to claim 1, wherein, in the three-layer restricted Boltzmann machine, the number of hidden layer hidden units in each layer of restricted Boltzmann machine 0.8-1.2 times the number of visible cells in the visible layer.
8. The personalized search method fused with attention mechanism according to claim 2, wherein the proportion of new items in the dominant item group D is η<30%.
9. The personalized search method fused with attention mechanism according to claim 1, characterized in that, in step 6, Euclidean distance is used as similarity calculation, that is, the smaller the Euclidean distance between two vectors, the greater the similarity between the two vectors. higher.
10. The personalized search system integrating attention mechanism is characterized in that it includes:

    The user-generated content acquisition module is used to collect and acquire user-generated content, which includes all items that user u has evaluated, ratings and text comments for each item, images of each item, and user-generated content from other users. u The usefulness evaluation score of the evaluation; vectorize the text comment, extract the feature of the item image, and obtain the feature vector;

    The advantageous project group building module is used to form the advantageous project group D with user preference of the projects whose user score is greater than the preset score threshold and whose trust degree is greater than the preset trust degree threshold;

    The user preference perception model construction and training module is used to construct and train a user preference perception model fused with an attention mechanism; the model is based on a deep belief network and consists of three layers of restricted Boltzmann machines, of which the first layer is restricted The visible layer of the Boltzmann machine includes the first group of visible units v ₁ , the second group of visible units v ₂ and the third group of visible units v ₃ , the hidden layer is h ₁ ; h ₁ is used as the visible layer, and the hidden layer h ₂ Constitute the second-layer restricted Boltzmann machine; h ₂ as the visible layer, and the hidden layer h ₃ form the third-layer restricted Boltzmann machine; the parameter of the user preference perception model of the fusion attention mechanism is θ ={θ ₁ ,θ ₂ ,θ ₃ }={w ₁ ,a ₁ ,b ₁ ,w ₂ ,a ₂ ,b ₂ ,w ₃ ,a ₃ ,b ₃ };

    The distribution estimation probability model building module based on user preference is used to build a user preference-based distribution estimation probability model P(x ):

    P(x)=[P(ψ ₁ ),P(ψ ₂ ),L,P(ψ _n ),L,P(ψ _Φ )] (17)

    where (ψ ₁ ,ψ ₂ ,…,ψ _n ,…,ψ _Φ ) is the original decision vector of item x, and P(ψ _n ) represents the user’s preference probability for the nth decision component of the item;

    The population generation module is used to estimate the probability model P(x) based on the distribution based on user preferences, use the distribution estimation algorithm to generate N new individuals, each individual is an item, and set the category label vector of each new individual, N is preset population size;

    The building block of the item set to be recommended is used to select and N new individual category label vectors in the search space

    

    The N items with the highest similarity constitute a set of items to be recommended S ^u ;

    The fitness value calculation module is used to calculate the fitness value of each item in the item set ^Su to be recommended;

    The search result selection module is used to select the top ^TopN items with the highest fitness value in Su as the search result, TopN<N.

Images