CN111797321B

CN111797321B - Personalized knowledge recommendation method and system for different scenes

Info

Publication number: CN111797321B
Application number: CN202010646490.8A
Authority: CN
Inventors: 史玉良; 李琳; 程林; 张坤; 孔凡玉
Original assignee: Shandong University
Current assignee: Shandong University
Priority date: 2020-07-07
Filing date: 2020-07-07
Publication date: 2021-04-27
Anticipated expiration: 2040-07-07
Also published as: CN111797321A

Abstract

The disclosure provides a personalized knowledge recommendation method and system for different scenes, which are used for acquiring user data information and judging the using scene of the user data information; if the short-term conversation type scene is detected, establishing a conversation learning model, extracting user interaction items in the conversation as input, modeling each interaction behavior through an attention mechanism, and outputting the preference of conversation content; establishing an edge information learning module, and predicting the current user requirement by mining the edge information of the project by using the graph structure information; if the scene is a long-term continuous scene, a classification-layering attention mechanism is adopted for coding, preference learning is carried out on two levels of knowledge items and categories, and user requirements are predicted; and recommending related information according to the predicted user requirements, wherein the recommended information is more accurate.

Description

Personalized knowledge recommendation method and system for different scenes

Technical Field

The disclosure belongs to a data information processing method, and relates to a personalized knowledge recommendation method and system oriented to different scenes.

Background

The statements in this section merely provide background information related to the present disclosure and may not necessarily constitute prior art.

With the innovation and rapid development of internet technology, internet information resources are increased rapidly, and it is very difficult for network users to accurately acquire required contents from massive information resources; for the internet platform, it is also a very challenging matter how to accurately provide appropriate information services for users, improve the efficiency of user access, and optimize the user experience. The recommendation system plays a key role in relieving information overload, and is widely used by a plurality of internet platforms such as e-commerce and online learning websites. With the increasing of internet information resources, the actual value of the recommendation system is higher and higher. The key of the recommendation system is to model the preference of the user according to historical access information (such as scores and clicks) of the user, and predict the demand or interest preference of the user, so that the user is helped to obtain interested contents or services from mass data more quickly.

For a software knowledge learning platform, the data volume of the platform is large, and the knowledge system and the types are various, such as video-class courses and document-type contents. Meanwhile, the knowledge data structure is different, and the organization form is different. In order to provide personalized services, improve user experience, and enhance interactivity between a user and a platform, personalized knowledge recommendation needs to be better performed by combining the platform and the user. Specifically, on one hand, the organization and processing of heterogeneous data on a platform needs to be considered, such as organizing heterogeneous knowledge items using knowledge graphs; on the other hand, the knowledge level difference between different users needs to be considered, and the current requirements and preferences of the users can be better learned by modeling the current knowledge level information of the users.

To the knowledge of the inventors, in the current recommendation method research, the RNN-based sequence modeling method does not consider the potential relationship between the interactive item and other items, and is difficult to capture deeper user requirements. In the method, by combining the knowledge recommendation background of the software knowledge learning platform, scene-based and hierarchical modeling is adopted to perform better knowledge recommendation, and short-term conversation type scene recommendation and long-term continuous type scene recommendation are mainly included. The short-term conversation type scene means that a user may intensively retrieve and browse the associated knowledge items in order to learn a certain knowledge in some time period, that is, the user has a strong purpose of accessing the platform in a subject manner in the time period. The long-term continuous scene means that the access of the user is general and continuous, and the user can continue to learn corresponding knowledge according to the previous learning progress, so that a continuous process of learning the knowledge by the user is embodied. For short-term conversation type recommendation scenes, the early recommendation method mostly only considers the relationship between a user and an item (user-item), which may cause poor recommendation effect; for personalized recommendation based on user knowledge level, the current sequence recommendation model is difficult to capture across categories.

Disclosure of Invention

The invention provides an individualized knowledge recommendation method and system for different scenes to solve the problems, and the method and system are used for carrying out targeted processing on two scenes, namely a short-term conversation scene and a long-term continuous scene, so that the knowledge level representation and the current knowledge requirement of a user can be studied more deeply, and more accurate knowledge recommendation is carried out.

According to some embodiments, the following technical scheme is adopted in the disclosure:

a personalized knowledge recommendation method oriented to different scenes comprises the following steps:

acquiring user data information and judging a use scene of the user data information;

if the short-term conversation type scene is detected, establishing a conversation learning model, extracting user interaction items in the conversation as input, modeling each interaction behavior through an attention mechanism, and outputting the preference of conversation content; establishing an edge information learning module, and predicting the current user requirement by mining the edge information of the project by using the graph structure information;

if the scene is a long-term continuous scene, a classification-layering attention mechanism is adopted for coding, preference learning is carried out on two levels of knowledge items and categories, and user requirements are predicted;

and recommending the relevant information according to the predicted user requirements.

As an alternative embodiment, user data information is acquired, and if the user has historical data exceeding a set value, the user is considered to be in a long-term continuous type scene, otherwise, the user is considered to be in a short-term conversation type scene.

As an alternative implementation, in a short-term conversational scenario, interactive items are mapped to corresponding entities of a knowledge graph based on a series of continuous interactive behaviors of a user in a conversation, and corresponding edge information vector representations are obtained by aggregating attributes and edge-related entities of the entities, so that edge information learning is realized.

As an alternative implementation, in a short-term conversation type scene, a cooperative attention mechanism is introduced based on a user conversation behavior vector and an edge information vector, the importance of each interaction is calculated, the correlation between conversation items and items in a knowledge graph spectrum is quantified, and a comprehensive vector representation of user preference is obtained;

constructing a softmax function by utilizing the comprehensive vector of the user preference; and calculating a loss function of the output value of the softmax function, and training the learning parameters of the gating cycle unit by adopting a back propagation algorithm to finish the training of the model.

As an alternative implementation, in a long-term continuous scene, the information sequence of the learning platform is accessed based on the user history, and the information sequence is mapped into a corresponding vector representation through an embedding operation, so that the user sequence information is encoded.

As an alternative embodiment, based on information encoding of a user sequence, a hidden state of each time step is acquired through a gating cycle unit to obtain information transmission conditions in a history sequence, so that an implicit state vector set is obtained.

As an alternative implementation, based on the obtained implicit state vector set, classifying according to the original label to obtain the implicit vectors corresponding to the knowledge of the same category; converting the user embedded vector into a category query vector, and introducing a knowledge sensitive attention mechanism to calculate the attention weight of the user to each category knowledge;

performing inner product operation on the obtained attention weight of the user to each category of knowledge and a known candidate item set to construct a softmax function; and calculating a loss function of the output value of the softmax function, and training the learning parameters of the gating cycle unit by adopting a back propagation algorithm to finish the training of the model.

A personalized knowledge recommendation system facing different scenes comprises the following steps:

the scene judging module is configured to acquire user data information and judge a use scene of the user data information;

the short-term conversation type scene processing module is configured to establish a conversation learning model, extract user interaction items in a conversation as input, model each interaction behavior through an attention mechanism, and output the preference of conversation contents; establishing an edge information learning module, and predicting the current user requirement by mining the edge information of the project by using the graph structure information;

the long-term continuous scene processing module is configured to encode by adopting a classification-layering attention mechanism, perform preference learning on two levels of a knowledge item and a category and predict user requirements;

and the information recommending module is configured to recommend the relevant information according to the predicted user requirement.

As an alternative embodiment, the short-term conversational scene processing module includes:

the session information learning module is used for modeling the sequential interaction of the users in the session by utilizing the context-associated GRU network, modifying the operation in the traditional GRU unit, adding the interaction behavior of the users into each gate function, and obtaining the vector representation of the user session behavior;

the edge information learning module is used for mapping the interactive items to corresponding entities of a Knowledge Graph (KG) based on a series of continuous interactive behaviors of the user in the session along with time change, and acquiring corresponding edge information vector representation by aggregating attributes and edge related entities of the entities;

the comprehensive vector acquisition module is used for introducing a cooperative attention mechanism based on the user session behavior vector and the edge information vector, calculating the importance of each interaction, quantifying the correlation between the session item and the items in the knowledge graph spectrum, and acquiring comprehensive vector representation of user preference;

the recommendation model training module is used for constructing a softmax function by utilizing the comprehensive vector of the user preference; calculating a loss function of the output value of the softmax function, and training the learning parameters of the GRU by adopting a back propagation algorithm to finish the training of the model;

and the feedback updating module is used for outputting the recommendation result of the experimental sample set after the model training is finished, comparing the recommendation result with the actual user behavior, feeding back and updating the bottom data information, continuously optimizing the weight value of the data and perfecting the user behavior recommendation result.

As an alternative embodiment, the long-term continuous scene processing module includes:

the user sequence information coding module is used for mapping the user sequence information into corresponding vector representation by embedding operation by utilizing the information sequence of the user history access learning platform;

the hidden state vector learning module is used for obtaining information transmission conditions in the historical sequence through a GRU (generalized regression unit) method based on the coding vector of the user sequence information;

the user knowledge level learning module is used for classifying according to the original labels based on the obtained hidden state vector set to obtain hidden vectors corresponding to knowledge of the same category; converting the user embedded vector into a category query vector, and introducing a knowledge sensitive attention mechanism to calculate the attention weight of the user to each category knowledge level;

the recommendation model training module is used for carrying out inner product operation on the attention weight of each category of knowledge and a known candidate item set by a user to construct a softmax function; calculating a loss function of the output value of the softmax function, and training the learning parameters of the GRU by adopting a back propagation algorithm to finish the training of the model;

A computer-readable storage medium, wherein a plurality of instructions are stored, and the instructions are adapted to be loaded by a processor of a terminal device and execute the personalized knowledge recommendation method oriented to different scenes.

A terminal device comprising a processor and a computer readable storage medium, the processor being configured to implement instructions; the computer readable storage medium is used for storing a plurality of instructions, and the instructions are suitable for being loaded by a processor and executing the personalized knowledge recommendation method facing different scenes.

Compared with the prior art, the beneficial effect of this disclosure is:

(1) the method and the system respectively carry out recommendation research from two angles of a short-term conversation type scene and a long-term continuous type scene, so that the purpose of personalized recommendation is achieved;

(2) for a short-term conversation type scene, the method is based on a series of continuous interaction behavior data (such as browsing, clicking and the like) of a user in a historical conversation, and considers two aspects of user conversation information learning and edge information learning to better mine potential preference information of the user. For user session information learning, based on a sequence of historical access sessions of users, modeling sequential interaction of the users in the sessions by using a context-associated GRU network to obtain user session behavior vector representation; for the learning of the edge information, the inventor maps interactive items in the session to corresponding entities of a knowledge graph, and corresponding edge information vector representation is obtained by aggregating the attributes of the entities and edge related entities;

(3) aiming at a short-term conversation type scene, the method introduces a cooperative attention mechanism based on a user conversation behavior vector and an edge information vector in consideration of the fact that the items in a user conversation and the information of the items in a knowledge graph have relevance, calculates the importance of each interaction, quantifies the relevance between the conversation items and the items in the knowledge graph, and obtains comprehensive vector representation of user preference, thereby improving the recommendation effect;

(4) aiming at a long-term continuous scene, the method is based on an information sequence of a user history access knowledge platform, in order to learn the current knowledge level change condition of the user, a user embedded vector is converted into a category query vector, a knowledge sensitive attention mechanism is introduced, and the attention weight of the user to each category of knowledge is calculated, so that more suitable recommendation is completed.

Drawings

The accompanying drawings, which are included to provide a further understanding of the disclosure, illustrate embodiments of the disclosure and together with the description serve to explain the disclosure and are not to limit the disclosure.

Fig. 1 is an overall flowchart of a personalized recommendation method for different scenes according to an embodiment of the present disclosure;

FIG. 2 is a flowchart of a process for collaborative attention based user behavior recommendation provided by an embodiment of the present disclosure;

FIG. 3 is a flow chart of a process for attention network recommendation based on knowledge level awareness provided by an embodiment of the present disclosure;

FIG. 4 is a diagram illustrating an attention mechanism recommendation provided by an embodiment of the present disclosure;

FIG. 5 is a flowchart of a recommendation embodiment for different scenarios provided by an embodiment of the present disclosure;

fig. 6 is a schematic structural diagram of a recommendation device facing different scenes according to an embodiment of the present disclosure.

The specific implementation mode is as follows:

the present disclosure is further described with reference to the following drawings and examples.

It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the disclosure. Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure belongs.

It is noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of example embodiments according to the present disclosure. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, and it should be understood that when the terms "comprises" and/or "comprising" are used in this specification, they specify the presence of stated features, steps, operations, devices, components, and/or combinations thereof, unless the context clearly indicates otherwise.

As shown in fig. 1 to 5, a personalized recommendation method for different scenes in this embodiment includes:

for a short-term conversational type scenario,

A. the method comprises the steps of collecting user session data, including historical data and current interactive data, and conducting data preprocessing on the acquired massive user session interactive data, including data cleaning, missing data completion, data definition and storage.

Specifically, interactive data of a user are acquired as application examples based on an online purchasing platform, music interactive data and E-commerce interactive data are acquired respectively, and a data set is divided into training data and testing data. As shown in table 1, a data set of interactions in a historical session is for a user.

Table 1 basic statistical information of data sets

B. Based on a series of continuous interactive behaviors (such as browsing, clicking and the like) of a user in a session, the sequential interaction of the user in the session is modeled by using a context-associated GRU network, the operation in a traditional GRU unit is modified, the interactive behaviors of the user are added into various gate functions, and a user session behavior vector representation is obtained, so that session information learning is realized.

Specifically, the general generation process of the user session information learning in step B is as follows:

B1. session Behavior Learning (SBL) is the modeling of short-term session behavior of a given user u, more specifically, for a given user access session sequence S_u＝{(x₁,b₁),(x₂,b₂),…,(x_i,b_i)}，x_iThe ith item within the session is represented,_irepresenting the way the user interacts with item i (e.g., click, purchase, etc.). Firstly, modeling is carried out according to continuous interactive behaviors (such as browsing, clicking and the like) of a user in a session, and preference information in the current session of the user is obtained.

B2. Modeling sequential interactions of users in a session using context-dependent GRU networks and modifying operations in conventional GRU units, b_iThe interaction behavior of (2) is added to each gate function separately. Previous studies have demonstrated that GRUs perform better than LSTM in session-based recommendations. Hidden state h in GRU_tMay be the previous hidden state h_t-1And candidate hidden states

Linear interpolation between:

z_t＝σ(W_zx_t+V_zb_t+U_zh_t-1) (2)

r_t＝σ(W_rx_t+V_rb_t+U_rh_t-1) (4)

wherein z is_tUpdating the gate function;

is a candidate activation function; r is_tIs a reset gate function; w_z,V_z,U_z,W_r,V_r,U_rRespectively, are weight coefficients.

As described above, implicit state h is used_tSave the current meetingThe intention information of the user in the conversation can be obtained to obtain a user conversation preference set I_u:

I_u＝{h_s,1,h_s,2,…,h_s,t} (5)

Wherein I_u∈R^D×tD is the dimension of the hidden state and R represents the set of real numbers.

C. In the learning of the edge information, the interactive item is mapped to the corresponding entity in the Knowledge Graph (KG), and the attribute of the entity and the edge related entity are aggregated to obtain the corresponding neighborhood information representation N (x)_n) And performing relevant potential interest learning. But in real-world knowledge maps, the number and relationship of the neighbors of an entity may vary greatly. To keep the computation of each batch stable and efficient, a fixed-size neighbor set is randomly selected without using its full neighbors.

Specifically, the general generation process of the user edge information learning in step C is as follows:

C1. mapping the interactive items to corresponding entities in a Knowledge Graph (KG) based on a user historical conversation sequence, and aggregating attributes and edge related entities of the entities to obtain corresponding neighborhood information representation N (x)_n). Computing a neighborhood representation v of an entity vⁱN (v) is:

wherein

Representing a relationship weight score between entities.

C2. For each interactive item i_nCalculating i layer by layer_nAnd represents neighborhood information of the entity item in the form of neighbor aggregation:

agg_neighbor＝σ(W·v_N(v)+b) (7)

aggregation is a key step for obtaining neighborhood information representation, and the implementation aggregates each entity in a session to obtain a neighborhood information set N_uWatch, watchEdge information collection of interactive items within a session:

N_u＝{agg_neighbor,1,agg_neighbor,2,…,agg_neighbor,t} (8)

D. introducing a cooperative attention mechanism based on the user session behavior vector and the edge information vector, calculating the importance of each interaction, quantifying the correlation between the session items and the items in the knowledge graph spectrum, and obtaining a comprehensive vector representation of the user preference:

the conversation behavior vector I obtained in the step B is used_uAnd the edge information vector N obtained in step C_uAs input to the cooperative attention network, a correlation matrix c is calculated:

wherein, W_cF is a tanh function as a weight coefficient.

Then, taking the correlation matrix c as a feature, the attention for calculating the preference and edge information in the session is:

P^s＝f₁(W_NN_u+(W_iI_u+W_ih_s,t)C^T) (10)

in the same way as above, the first and second,

P^u＝f₁(W_iI_u+(W_NN_u+W_tagg_neighbor,t)C^T) (12)

wherein f is₁,f₂Tan h function and softmax function, W, respectively_s,W_t,W_tAre weight coefficients.

Depending on the weight of attention, the interdependent characterization of the user's long-term and short-term preferences can be computed as a weighted sum of their interactive characterizations:

and

to emphasize the effect of the current term on user preferences, this implementation uses P_u＝[I_co-u；h_s,t],P_s＝[P_co-u；agg_neighbor,t]The final intra-session preference and edge information potential preference are represented separately.

Finally, the two are combined with the candidate item set, and comprehensive vector representation of user preference is obtained

E. D, recommending the user behavior of the sample to be predicted, and outputting the result in the step D

Inputting softmax layer results in output values as follows:

where y' represents the output vector of the model, representing the probability distribution of the candidate items, each probability representing the probability that the item is the next visit.

F. And (3) adopting cross entropy as a loss function, and if y is a real class distribution, defining the loss function as follows:

then, optimization was performed using a random gradient descent optimizer.

User behavior recommendation is performed on a sample to be tested, a recommendation result is pushed and compared with an actual user behavior, and table 2 describes performance comparison of methods in the user behavior recommendation:

TABLE 2 Performance comparison of user behavior recommendations

Based on the results in table 2, the performance of the proposed user behavior recommendation model of this embodiment is superior to other methods.

For a long-term continuous-type scene recommendation,

A. the method comprises the steps of collecting user access knowledge platform data, including historical data and current data, preprocessing the acquired massive user access data, including data cleaning, missing data completion, data definition and storage.

Specifically, historical access data of a user are obtained as an application example based on a knowledge software knowledge learning platform, and the obtained access data of the user mainly comprise a user ID, a user access course ID, time T for the user to access the course, a user access progress P, difficulty level H for the user to access the course, a category C to which the user accesses the course, and the like. Firstly, the acquired data is processed, and too few and too long user data samples are screened to obtain user access data with uniform length. The specific user data description is shown in table 3.

TABLE 3 basic statistics of the data set

Recording	Number of	Number after screening
			User recording	114827	109696
Course(s)	837	445
			User' s	12600	10398
Chapters and sections	3227	2164
			Class of course	26	26

B. Based on the historical information sequence of the user accessing the software knowledge learning platform, the current knowledge level of the user can be analyzed hierarchically and the current required knowledge of the user can be presumed.

Let u denote the user, the route that the user has historically learned: the initial full sequence is:

S(u)＝{k₁,k₂,…,k_n} (1)

further, an internal (implicit) sequence is obtained according to the classification of the class label to which the knowledge belongs:

c_i＝{k₁,k₂,…,k_mwhere m < n (2)

It should be noted that k is₁,k_iNot necessarily consecutive items in the initial full sequence.

Therefore, the user history learning complete sequence can be expressed as:

C. historical access sequence S for a given user_k＝{k₁,k₂,…,k_nN is the number of items accessed by the user history. The sequence K is then converted by the embedding layer into a space vector representation E ═ E₁,e₂,…,e_n}，e_iFor a D-dimensional vector, D is the number of all terms, and UserID is also converted into a spatial vector representation by embedding.

D. Based on the user sequence information code obtained in the step C, the hidden state h of each time step is obtained by using a GRU-based method_t，h_tMay be the previous hidden state h_t-1And candidate hidden states

Linear interpolation between:

wherein the content of the first and second substances,

z_t＝σ(W_zx_t+U_zh_t-1) (5)

at the same time, candidate activation functions

Comprises the following steps:

wherein the gate function r is reset_tThe calculation is as follows:

r_t＝σ(W_rx_t+U_rh_t-1) (7)

W,U,W_z,U_z,W_r,U_rare weight coefficients.

By obtaining the implicit state h for each time step, as described above_iTo obtain the information transmission condition in the history sequence, and obtain the implicit state vector set H_S＝{h_s,1,h_s,2,…,h_s,t}，I_u∈R^D×tWhere R represents a natural dataset and D is a dimension of the hidden state.

E. Classifying the hidden state vector set obtained in the step D according to the original labels to obtain hidden vectors corresponding to knowledge of the same category, wherein one category is represented as:

C_n＝{h_s,i,h_s,j,…,h_s,k} (8)

wherein n, i, j, k are constants.

Then, the implicit vectors within the same class are concatenated in series:

C_n＝[h_s,ih_s,j,…,h_s,k] (9)

wherein n, i, j, k are constants.

To obtain a current knowledge level representation of a user, a user embedded vector is first converted into a category query vector q_cThe expression is as follows:

q_c＝RelU(V_c×e_u+b_c) (10)

then, the category-level knowledge sensitive attention score is calculated as follows:

α_p(c_i)＝c^Tσ(W_p×q_c+b_p) (11)

the end-user knowledge level representation u_pIs the sum of the attention representations of the classes, weighted according to their attention weights:

to verify the improvement of the attention mechanism to the recommendation effect, the performance difference of LSTM + attention and LSTM can be compared, as shown in fig. 4. From the fact that the LSTM + attention result is superior to the LSTM, the attention mechanism has certain learning capacity in sequence modeling, and the effect of the model can be improved in sequence recommendation.

F. Herein, a negative sampling technique is adopted to jointly predict click scores of K +1 candidate sets in model training. The K +1 candidate sets consist of one positive sample of one user and K randomly selected negative samples.

Is formed by a candidate set k_iAnd a user current knowledge level representation u_pThen calculated by the softmax function:

based on the above description, an attention network personalized recommendation algorithm based on knowledge level perception is given.

The attention network personalized recommendation algorithm process based on knowledge level perception is as follows:

inputting: user access record sequence K, user embedding vector u, candidate set V

And (3) outputting: score for each candidate knowledge item

01. Model initialization

02.for k in sequence do

03.v＝Embedding{k}

04.Dense{v},then GRU-based{v_i}

05. Acquiring hidden state in GRU network

06.end for

07.for c in C:

08.for h in c:

09.Concat{h_i}

10.Attention layer{category-level}

11. Calculating the contribution score of each knowledge to the current knowledge level on the category level through the formula (12)

12. By equation (13), a knowledge level representation of the current user is obtained

13. Candidate knowledge item k by equations (14) (15)_c-iWith user representation u_pCalculating recommendation score by dot product

The core idea of the algorithm is that modeling is carried out based on user sequence access information and characteristic information of classification of the user sequence access information, and the current knowledge level state of a user is learned, so that knowledge recommendation is carried out more reasonably. Specifically, sequence modeling is carried out through a GRU-based method, context information accessed by a user sequence is learned, reclassification processing is carried out according to class characteristics of knowledge items, and the hidden states of the same class are connected in series (lines 2-6); then, at the category level, the contribution score of each type of knowledge to the current user knowledge level is calculated using an attention mechanism (lines 7-10); finally, a knowledge level representation of the current user is obtained and a recommendation score for each candidate is calculated (lines 11-13).

G. And (3) adopting cross entropy as a loss function, and if y is a real class distribution, defining the loss function as follows:

then, optimization was performed using a random gradient descent optimizer.

This implementation adopts TensorFlow and Keras to realize the framework of the model and the baseline model that propose, in order to prevent overfitting, used two dropout layers: the first dropout layer is used between the knowledge item embedding layer and the GRU layer, and the discarding rate is set to be 20%; the second dropout layer has a drop rate set to 30% before the user knowledge level is expressed. In order to accurately classify hidden states in the GRU, the text acquires category feature labels corresponding to time steps by position. During training, the learning rate is set to 0.001 and the batch size is set to 128. In the optimization process, parameter optimization is performed by using an Adam optimizer.

User behavior recommendation is performed on the sample to be tested, the recommendation result is pushed and compared with the actual user behavior, and table 4 describes the performance comparison of the method in the user behavior recommendation:

TABLE 4 comparison of Properties

Based on the results in table 4, the proposed recommendation model (KLAN) of this embodiment performs better than other methods.

The short-term conversational recommendation scene proposed by the embodiment is generally that a user frequently accesses the content or knowledge of the platform within a certain time, and the purpose is strong. The long-term conversation type recommendation scene is usually directed at a knowledge learning scene with an advanced level and has long-term continuity. Such as the learning history of shallow and deep for each system of knowledge.

As shown in fig. 6, the present embodiment provides a personalized recommendation device facing different scenes, including:

for short-term conversational-type scene recommendations,

(1) the session information learning module is used for modeling the sequential interaction of the users in the session by utilizing the context-associated GRU network, modifying the operation in the traditional GRU unit, adding the interaction behavior of the users into each gate function, and obtaining the vector representation of the session behavior of the users, thereby realizing the session information learning;

(2) the edge information learning module is used for mapping the interactive items to corresponding entities of a Knowledge Graph (KG) based on a series of continuous interactive behaviors (such as browsing, clicking and the like) which change along with time in a conversation of a user, and obtaining corresponding edge information vector representation by aggregating attributes and edge related entities of the entities, so that the edge information learning is realized.

(3) The comprehensive vector acquisition module is used for introducing a cooperative attention mechanism based on the user session behavior vector and the edge information vector, calculating the importance of each interaction, quantifying the correlation between the session item and the items in the knowledge graph spectrum, and acquiring comprehensive vector representation of user preference;

(4) the recommendation model training module is used for constructing a softmax function by utilizing the comprehensive vector of the user preference; calculating a loss function of the output value of the softmax function, and training the learning parameters of the GRU by adopting a back propagation algorithm to finish the training of the model;

specifically, in the recommendation model training module, the cross entropy is used as a loss function of the recommendation model to realize the training of the model.

(5) And the feedback updating module is used for outputting the recommendation result of the experimental sample set after the model training is finished, comparing the recommendation result with the actual user behavior, feeding back and updating the bottom data information, and continuously optimizing the weight value of the data, thereby continuously improving the user behavior recommendation.

For a long-term continuous-type scene recommendation,

(1) the user sequence information coding module is used for mapping the user sequence information into corresponding vector representation by utilizing the information sequence of the user history access learning platform through embedding operation, so that the user sequence information coding is realized;

(2) an implicit state vector learning module, which is used for obtaining information transmission conditions in a history sequence through a GRU method based on the coding vector of the user sequence information so as to learn the implicit state vector set,

(3) the user knowledge level learning module is used for classifying according to the original labels based on the obtained hidden state vector set to obtain hidden vectors corresponding to knowledge of the same category; meanwhile, in order to obtain the current knowledge level representation of the user, converting the user embedded vector into a category query vector, and introducing a knowledge sensitive attention mechanism to calculate the attention weight of the user to each category knowledge level;

(4) the recommendation model training module is used for carrying out inner product operation on the attention weight of each category of knowledge and a known candidate item set by a user to construct a softmax function; calculating a loss function of the output value of the softmax function, and training the learning parameters of the GRU by adopting a back propagation algorithm to finish the training of the model;

In another embodiment, a personalized recommendation device facing different scenes is characterized by further comprising:

and the preprocessing module is used for preprocessing data of mass historical user data, and comprises data cleaning, missing data completion, data definition and normalization processing.

In another embodiment, a computer-readable storage medium is provided, on which a computer program is stored, which program, when being executed by a processor, carries out the steps of the recommendation method as shown in fig. 1.

In another embodiment, a computer device is provided, comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the steps of the proposed method as shown in fig. 1 when executing the program.

As will be appreciated by one skilled in the art, embodiments of the present disclosure may be provided as a method, system, or computer program product. Accordingly, the present disclosure may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present disclosure may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and so forth) having computer-usable program code embodied therein.

The present disclosure is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the disclosure. It will be understood that each flow and/or block of the flow diagrams and/or block diagrams, and combinations of flows and/or blocks in the flow diagrams and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

The above description is only a preferred embodiment of the present disclosure and is not intended to limit the present disclosure, and various modifications and changes may be made to the present disclosure by those skilled in the art. Any modification, equivalent replacement, improvement and the like made within the spirit and principle of the present disclosure should be included in the protection scope of the present disclosure.

Although the present disclosure has been described with reference to specific embodiments, it should be understood that the scope of the present disclosure is not limited thereto, and those skilled in the art will appreciate that various modifications and changes can be made without departing from the spirit and scope of the present disclosure.

Claims

1. A personalized knowledge recommendation method oriented to different scenes is characterized by comprising the following steps: the method comprises the following steps:

if the scene is a long-term continuous scene, a classification-layering attention mechanism is adopted for coding, preference learning is carried out on two levels of knowledge items and categories, and user requirements are predicted; the method specifically comprises the following steps: based on the information sequence of the user history access learning platform, mapping the information sequence into corresponding vector representation through embedding operation to realize the information coding of the user sequence; based on information coding of a user sequence, acquiring a hidden state of each time step through a gating circulation unit to obtain an information transmission condition in a historical sequence, so as to obtain a hidden state vector set; recommending related information according to the predicted user requirements;

classifying according to the original labels based on the obtained hidden state vector set to obtain hidden vectors corresponding to knowledge of the same category; converting the user embedded vector into a category query vector, and introducing a knowledge sensitive attention mechanism to calculate the attention weight of the user to each category knowledge;

2. The method as claimed in claim 1, wherein the method comprises: and acquiring user data information, and if the user has historical data exceeding a set value, considering the user to be in a long-term continuous scene, or else, considering the user to be in a short-term session scene.

3. The method as claimed in claim 1, wherein the method comprises: during a short-term conversation type scene, based on a series of continuous interactive behaviors of a user in a conversation, mapping interactive items to corresponding entities of a knowledge map, and aggregating attributes and edge related entities of the entities to obtain corresponding edge information vector representation so as to realize edge information learning;

or, when in a short-term conversation type scene, a cooperative attention mechanism is introduced based on a user conversation behavior vector and an edge information vector, the importance of each interaction is calculated, the correlation between conversation items and items in a knowledge graph spectrum is quantized, and comprehensive vector representation of user preference is obtained;

4. A personalized knowledge recommendation system oriented to different scenes is characterized in that: the method comprises the following steps:

the hidden state vector learning module is used for obtaining information transmission conditions in the historical sequence through a GRU (generalized regression unit) method based on the coding vector of the user sequence information; classifying according to the original label to obtain an implicit vector corresponding to the knowledge of the same category; converting the user embedded vector into a category query vector, and introducing a knowledge sensitive attention mechanism to calculate the attention weight of the user to each category knowledge;

5. The system of claim 4, wherein the system comprises: the short-term session type scene processing module comprises:

the edge information learning module is used for mapping the interactive items to corresponding entities of the knowledge graph based on a series of continuous interactive behaviors of the user in the session along with time change, and acquiring corresponding edge information vector representation by aggregating attributes and edge related entities of the entities;

6. The system of claim 4, wherein the system comprises: the long-term continuous scene processing module comprises:

7. A computer-readable storage medium characterized by: a plurality of instructions are stored, the instructions are suitable for being loaded by a processor of the terminal equipment and executing the personalized knowledge recommendation method facing different scenes, wherein the personalized knowledge recommendation method is as claimed in any one of claims 1 to 3.

8. A terminal device is characterized in that: the system comprises a processor and a computer readable storage medium, wherein the processor is used for realizing instructions; the computer readable storage medium is used for storing a plurality of instructions, and the instructions are suitable for being loaded by a processor and executing the personalized knowledge recommendation method facing different scenes, wherein the personalized knowledge recommendation method is as claimed in any one of claims 1 to 3.