An education resource recommendation method based on a user preference vector model

Abstract

The application discloses an education resource recommendation method based on a user preference vector model, and comprises the following steps: S1, constructing a user preference vector model based on the preference relation of a user to an education resource attribute; the user preference vector model comprises a user preference vector model based on statistical entropy reduction maximization constructed for a new user and a user preference vector model based on the inner product of a user-resource preference score matrix and a resource-attribute matrix constructed for an old user; S2, finding a substitute resource in the case that a certain attribute value corresponding resource is missing in an education resource library based on the user preference vector model; and S3, combining the learning expectation in the education process and the personalized characteristics of knowledge points, and recommending the education resource based on the user preference vector model. The application can construct the user preference model for the new user and the old user respectively, and can select the substitute resource in the case that the resource is missing, and can realize the education resource recommendation of different learning expectations through the constructed user preference vector model.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of user model construction and education resource recommendation, and particularly relates to an education resource recommendation method based on a user preference vector model. BACKGROUND

[0002] In recent years, with the successful application of recommendation systems, many scholars have applied personalized recommendation technology to the education field to push education resources according to the personalized characteristics of users. Unlike movie or commodity recommendation, education resources contain not only educational content but also the attributes of education resources. Therefore, education resource recommendation not only involves content recommendation but also the selection of the attributes of education resources. The present application is directed to the latter, that is, in the process of resource recommendation, after the recommended content (knowledge points) are determined, the attributes of the education resources corresponding to the knowledge points, such as media types and content backgrounds, need to be further determined. Users have different preferences for different attributes, and the learning effects brought by this are also different. How to obtain user preferences and establish a preference model is one of the key problems of education resource recommendation.

[0003] A user model is the basis of education resource recommendation. Common education resource recommendation includes three kinds of recommendation algorithms, namely, hybrid filtering, collaborative filtering and content-based recommendation. Their essence is to obtain user preferences so as to recommend learning content that meets the user preferences from a large number of education resources. Therefore, obtaining user preferences and establishing a user model is an important content in education resource recommendation.

[0004] Research on how to obtain user preferences and establish a user model has been thoroughly studied. Some scholars use tags to obtain user preferences, for example and the like proposed an enhanced recommendation with tags, regarding tags as the basis for obtaining user preferences, constructing a user model and generating a learning resource recommendation sequence based on user preferences, but this method requires users to spend a lot of time tagging resources. Che et al. proposed a recommendation algorithm based on tag clustering and user preferences, quantified user preferences by obtaining tags, divided user preferences using a clustering algorithm and established a preference model of similar users. However, this method does not consider the characteristic attributes of tags.

[0005] Some scholars consider establishing a user model from the long-term and short-term preferences of users. For example, Yu et al. introduced the concept of user preference fluctuation to measure the degree of short-term preference change of users and established a short-term user preference model. However, there are still some defects in accurately quantifying user preference fluctuation. Liu et al. proposed a long-term and short-term user preference model, integrated the long-term and short-term preferences of users together and generated accurate recommendations. However, how to determine the representation of the user preference model was not considered.

[0006] In addition, there are also researches on establishing user preference model by other methods. For example, Wang et al. establish user preference model from the aspect of user visual preference, but the complexity of obtaining user preference is high. Jawaheer et al. propose to use explicit and implicit feedback of users as the key to user preference modeling. However, this method does not combine explicit and implicit feedback to construct a unified user model. Higuchi et al. propose two online learning algorithms to construct user preference, which sequentially learn through all prediction settings and limited prediction settings to obtain user preference. However, this algorithm needs certain historical data to establish user preference model, and therefore has a cold start problem for new users.

[0007] In addition, the user model constructed by the above various methods cannot be used for correlation operation; most recommendation systems rarely consider the missing of items and the selection method of substitute items. SUMMARY

[0008] The present application aims to provide an education resource recommendation method based on user preference vector model, which constructs user preference model for new users and old users respectively, and can select substitute resources in the case of missing resources, and can realize education resource recommendation with different learning expectations through the constructed user preference vector model.

[0009] To achieve the above-mentioned purpose, the present application provides an education resource recommendation method based on user preference vector model, comprising the following steps:

[0010] S1, constructing a user preference vector model based on the preference relationship of users to the attributes of education resources:

[0011] The user preference vector model includes a user preference vector model based on statistical entropy reduction maximization constructed for new users and a user preference vector model based on the inner product of user-resource preference score matrix and resource-attribute matrix constructed for old users;

[0012] S2, finding substitute resources in the case of missing resources corresponding to a certain attribute value in the education resource library based on the user preference vector model;

[0013] S3, combining the learning expectation in the education process and the personalized characteristics of knowledge points, and performing education resource recommendation based on the user preference vector model.

[0014] As a further scheme of the present application, the content of the education resources has multiple different attributes in step S1, and users have different tendency selection for different attribute values, which is represented as the preference relationship of users to the attribute value of the education resources;

[0015] The attribute space of the education resources is denoted as A, which is composed of n attributes, and is denoted as:

[0016] A = {S1,S2,…,S} n}

[0017] Among them, S i Let represent the i-th attribute, and n be the number of attributes. The set of values ​​for each attribute is denoted as:

[0018] S i ={t i1 ,t i2 ,…,t ik}

[0019] Among them, t ik It is the kth value, and t is defined as follows: ij With t ik (j≠k) are mutually exclusive;

[0020] The space A, composed of the composite attributes of educational resources, is the Cartesian product of these attributes, and its size is:

[0021] A = |S1| × |S2| × … × |S n |

[0022] Among them, |S i | represents S i Number of elements in the middle;

[0023] For two elements t in the attribute set ij and t ik If the user has t ij The preference is greater than that for t ik If the preference is t, then it is called t ij Preference dominance t ik ;

[0024] Elements in the educational resource attributes are used as vertices of a directed graph, and their dominance relations are used as edges of the directed graph, forming a dominance relation graph. The adjacency matrix of the dominance relation graph is the dominance relation matrix. The ranking of attribute value preferences is obtained based on the dominance relation matrix. Then, the attribute value preference ranking vector is used to represent the user preference vector model.

[0025] As a further aspect of the present invention: In step S1, the principle of maximizing entropy reduction in the preference vector model for new users is as follows: for missing values ​​in the dominant relationship matrix, the missing values ​​are filled using historical statistical data of all existing users. By comparing the information entropy change before and after replacing a certain filling value, the key attribute that makes the matrix entropy value change the most is selected and submitted to the user for evaluation.

[0026] As a further aspect of the present invention: the two preference dominance matrices before and after the selection of the key attribute submitted to the user to obtain their preferences satisfy a first-order difference matrix relationship;

[0027] The first-order difference matrix relation is expressed as: matrix R1, R2 is a dominant relation matrix, R3 is the difference between matrix R1 and R2: R3=R1-R2, if R3 has and only has one non-zero element, then R1 is called the first-order difference matrix of R2;

[0028] For the dominant relation matrix R=(r ij ), the information entropy H(R) calculation formula is:

[0029]

[0030] Where r ij represents the attribute value t i The dominant situation of t j , p(r ij ) represents the dominant probability of t i To t j , 1-p(r ij ) represents the dominant probability of media type t j To t i ;

[0031] The entropy reduction calculation formula is:

[0032] ΔH(R1, R2)=|H(R1)-H(R2)|

[0033] Where, ΔH(R1, R2) is the difference between the entropy values of matrix R1 and R2, and the element position formula corresponding to the maximum entropy reduction is:

[0034] (i, j)=argmax(ΔH(R1, R2))

[0035] The attribute corresponding to the position of the maximum entropy reduction element is submitted to the user for evaluation, the user makes a preference selection on the given preference attribute, fills in the elements in the user preference dominant relation matrix, and constructs a user preference vector model according to the user preference dominant relation matrix.

[0036] As a further scheme of the application: in step S1, facing old users, the user-resource preference score matrix G contains the corresponding score of the user u to the resource e i The resource-attribute matrix F contains the value condition whether the resource e i Has attribute t ij ; the user-resource preference score matrix G and the resource-attribute matrix F are subjected to inner product operation, the preference degree of the user to the resource attribute is calculated according to the score of the user to the resource; and the formula for calculating the preference degree of the user to the media attribute based on statistical average is:

[0037]

[0038] In the formula, I u,tr represents the set of resources that have attribute t. u,e This represents the actual rating that user u gave to resource e;

[0039] For the user-resource rating matrix G and the resource-attribute matrix F, the user preference vector model M is:

[0040] M = GF * Q

[0041] Where GF represents the inner product of G and F, and * represents the Hadamard product, Q is the number of resources with attribute t in the user rating matrix R |I u，t The reciprocal of | (1 / |I) u,t |); By combining the historical rating information of users and resources with the attribute features contained in the resources, the user's preference for the attribute features of the resources is calculated, thereby constructing a user preference vector model.

[0042] As a further aspect of the present invention: the method for finding substitute resources in step S2 is as follows: the educational resource attribute vector V is encoded in binary. When an educational resource has n attributes, it is represented by a binary string of length n.

[0043]

[0044] Each bit corresponds to the existence of a specific value for a certain attribute; that is, a bit being 1 indicates that the educational resource attribute contains the corresponding value, and 0 indicates that it does not contain the corresponding value. It is easy to see that the educational resource can be represented in at most 100 different ways. kind;

[0045] The substitute resource e is calculated by using the dot product of the user preference vector model M and the educational resource attribute vector V. r * When selecting alternative resources, the corresponding educational resource attribute vectors must satisfy the condition of maximizing the dot product, i.e., satisfying the following formula:

[0046]

[0047] Wherein d(u,e) r ) represents the vector distance between a user and an educational resource in the educational resource attribute space. If 1 / (M·V) is used to represent the distance, then the smaller the distance, the larger the dot product. The principle is the same for both.

[0048] As a further aspect of the present invention: the educational resource recommendation method in step S3 includes:

[0049] When faced with new knowledge points, the system recommends educational resources that closely match the user's preference vector model and satisfy their preferences. The recommended educational resource e... r * Satisfying the maximum dot product:

[0050]

[0051] Facing the challenging educational resources which are far away from the user preference vector model of old knowledge points, the recommended educational resources e r * Satisfy the minimum dot product:

[0052]

[0053] Compared with the prior art, the beneficial effects of the present application are as follows:

[0054] The construction method of the user preference vector model of the two types of users, new users and old users, is provided, which helps to more accurately recommend educational resources;

[0055] The substitute resource selection method based on preference ranking in the case of resource missing is proposed, and the user preference vector model is applied to realize the educational resource recommendation method with different learning expectations, which can accurately recommend educational resources, avoid the case of no educational resource recommendation caused by the missing of certain resources, and effectively improve the user satisfaction. DETAILED DESCRIPTION

[0056] The present application will be further described below through examples.

[0057] A kind of educational resource recommendation method based on user preference vector model, comprising the following steps:

[0058] S1, based on the preference relationship of user to educational resource attribute, construct user preference vector model:

[0059] The user preference vector model includes the user preference vector model based on statistical entropy reduction maximization constructed for new users and the user preference vector model based on the inner product of user-resource preference score matrix and resource-attribute matrix constructed for old users;

[0060] S2, based on the user preference vector model, find substitute resources in the case of missing resources corresponding to certain attribute values in the educational resource library;

[0061] S3, based on the user preference vector model, combine the learning expectations in the educational process and the personalized features of knowledge point mastery to recommend educational resources.

[0062] Specifically, in step S1, the educational resource content has multiple different attributes, such as "media type", "content carrier", "background information", etc. Users have different tendencies in different attribute value conditions, which is manifested as the preference relationship of users to attribute value conditions;

[0063] The attribute space of educational resources is denoted as A, which is composed of n attributes, denoted as:

[0064] A = {S1, S2, …, S n}

[0065] where S i represents the i-th attribute, n is the number of attributes, and the value set of each attribute is denoted as:

[0066] S i = {t i1 , t i2 , …, t ik}

[0067] where t ik is the k-th value, and it is defined that t ij and t ik (j≠k) are mutually exclusive;

[0068] The space A composed of the composite attributes of the educational resources is the Cartesian product of the attributes, and its size is:

[0069] A = |S1| × |S2| × … × |S n |

[0070] where |S i | represents the number of elements in S i ;

[0071] For example, A = {S1 = media type, S2 = content carrier, S3 = background information}. S1 = {t 11 = audio, t 12 = video, t 13 = text, t 14 = image}, S2 = {t 21 = theory, t 22 = explanation, t 23 = test question, t 24 = example, t 25 = mind map}, and S3 = {t 31 = entertainment, t 32 = life, t 33 = science, t 34 = history}. At this time, |A| = 80.

[0072] The preference relationship of the user for the attributes of the educational resources is as follows: for two elements t ij and t ik in the attribute set, if the user prefers t ij more than t ik , then t ij is said to be preferred to t ik ;

[0073] The elements in the education resource attribute are taken as vertices of a directed graph, and the dominant relations are taken as edges of the directed graph to form a dominant relation graph. An adjacency matrix of the dominant relation graph is a dominant relation matrix. The ordering of the attribute value preference degree is calculated based on the dominant relation matrix. Then, the user preference vector model is represented by using the attribute value preference degree ordering vector.

[0074] The user preference vector model construction method based on statistical entropy reduction maximization is used for new users. Shannon introduced entropy from thermodynamics to information theory and proposed the concept of information entropy to measure the degree of information disorder and represent the uncertainty of events. The information entropy calculation formula of a random event X is:

[0075]

[0076] wherein p(x i ) represents the probability of X=x i .

[0077] The principle of entropy reduction maximization in the preference vector model construction for new users in step S1 is that the missing values in the dominant relation matrix are filled by using the historical statistical data of all existing users. By comparing the information entropy changes before and after replacing a certain filled value, the key attribute with the highest matrix entropy value change is selected and submitted to the user for evaluation.

[0078] The two preference dominant matrices corresponding to the key attribute selection before and after being submitted to the user for obtaining the preference satisfy the first-order difference matrix relationship. The higher the entropy value change is, the more the uncertainty is eliminated, and the lower the uncertainty degree of the matrix is. Thus, the user preference vector model is constructed.

[0079] The first-order difference matrix relationship is represented as follows: the matrices R1 and R2 are dominant relation matrices, and R3 is the difference between the matrices R1 and R2: R3=R1-R2. If R3 has and only has one non-zero element, R1 is called the first-order difference matrix of R2.

[0080] The key attribute selection is proposed according to the statistical entropy reduction maximization.

[0081] For the dominant relation matrix R=(r ij ), the information entropy H(R) calculation formula is:

[0082]

[0083] wherein r ij represents the dominant relation of the attribute value t i to t j , p(r ij ) represents the dominant probability of t i to t j , and 1-p(r ij ) represents the media type tj The dominant probability of t i ;

[0084] The entropy reduction formula is:

[0085] ΔH(R1, R2) = |H(R1) - H(R2)|

[0086] Wherein, ΔH(R1, R2) is the difference of the entropy value of the matrix R1 and R2, and the element position formula corresponding to the maximum entropy reduction is:

[0087] (i, j) = argmax(ΔH(R1, R2))

[0088] The attribute corresponding to the maximum entropy reduction element position is submitted to the user for evaluation, the user makes a preference selection on the given preference attribute, fills in the elements in the user preference dominant relation matrix, and constructs a user preference vector model according to the user preference dominant relation matrix.

[0089] For old users, the user preference vector model construction method based on the inner product of the user-resource preference score matrix and the resource-attribute matrix. In step S1, for old users, the user-resource preference score matrix G contains the corresponding score of the user u to the resource e i ; the resource-attribute matrix F contains the value condition of whether the resource e i has the attribute t ij ; the user-resource preference score matrix G and the resource-attribute matrix F are operated by inner product, the preference degree of the user to the resource attribute is calculated according to the score of the user to the resource; and the formula for calculating the preference degree of the user to the media attribute based on statistical average is:

[0090]

[0091] Wherein, I u,t represents the set of resources with t attribute characteristics, and r u,e represents the actual score (absolute score) of the user u to the resource e

[0092] For the matrix user-resource score matrix G and the resource-attribute matrix F, the user preference vector model M is:

[0093] M = GF*Q

[0094] Wherein, GF represents the inner product of G and F, * represents Hadamard product, and Q is the reciprocal (1 / |I u，t |) of the number of resources with attribute t in the user score matrix R |I u,t |); the historical score information of the user and the resource and the attribute characteristics contained by the resource are combined to calculate the preference of the user to the resource attribute characteristics, so as to construct the user preference vector model.

[0095] For the problem of resource missing, a substitute resource selection method based on user preference vector model is adopted.

[0096] Specifically, the way of finding substitute resources in step S2 is: the education resource attribute vector V is encoded in binary, when the education resource has n attributes, a binary string with length n is used to represent:

[0097]

[0098] Each bit corresponds to whether a specific value of an attribute exists, that is, a bit of 1 indicates that the corresponding value is included in the education resource attribute, and 0 indicates that the corresponding value is not included. It is easy to know that the education resource representation has at most kinds;

[0099] The dot product of the user preference vector model M and the education resource attribute vector V is used to calculate the substitute resource e r * Taking M = ((2, 4, 3.5, 1), (2, 3, 1, 4, 5), (3, 4, 2, 1)) and V0 = ((0, 1, 0, 0), (0, 0, 0, 0, 1), (0, 1, 0, 0)) as an example, it is easy to get M·V = 13, and V0 is the education resource e r 0 corresponding to the vector. However, when e r 0 does not exist in the candidate resource library E R , a substitute resource needs to be selected, and the corresponding education resource satisfies the maximum dot product, that is, the following formula:

[0100]

[0101] where d(u, e r ) represents the vector distance between the user and the education resource in the education resource attribute space. If the distance is represented by 1 / (M·V), the smaller the distance, the larger the dot product, and the principles are the same.

[0102] Combined with the learning expectations in the education process and the personalized characteristics of knowledge points, two types of education resource recommendation methods based on user preference vector model are provided, which are the recommendation methods for new knowledge points and old knowledge points.

[0103] Specifically, the education resource recommendation method in step S3 includes:

[0104] For new knowledge points, the education resource that meets the user's preference vector model is recommended, and the recommended education resource e r * satisfies the maximum dot product:

[0105]

[0106] Facing the challenging educational resources of old knowledge point recommendation away from user preference vector model, to improve the learning ability of learners, the recommended educational resources e r * Satisfy the minimum dot product:

[0107]

[0108] Through the learning of different presentation methods of old knowledge points, more association and memory support about old knowledge points are provided to learners, so that learners can have a deeper understanding and consolidation of previous knowledge points.

Claims

1. An education resource recommendation method based on a user preference vector model, characterized in that, The method comprises the following steps: S1, constructing a user preference vector model based on the preference relationship of the user to the attribute of the educational resource; The user preference vector model includes a user preference vector model based on statistical entropy reduction maximization constructed for new users and a user preference vector model based on the inner product of the user-resource preference score matrix and the resource-attribute matrix constructed for old users; The educational resource content has multiple different attributes, and the user has different tendency choices for different attribute value conditions, which is manifested as the preference relationship of the user to the attribute value condition; The attribute space of the education resource is denoted as A , which is composed of n an attribute, denoted as: A ={ S 1, S 2, … , S n} wherein, S i denotes the i th attribute, n is the number of attributes, and the set of values for each attribute is denoted by S i = { t i1 , t i2 , … , t ik} wherein t ik is the k th value, and t ij have mutual exclusivity, wherein t ik j ≠ k ;​ The space of the composite attributes of the educational resources A is the Cartesian product of the attributes, whose size is: A =| S 1|×| S 2|×…×| S n | wherein, S i | represents S i number of elements in the middle For two elements in a set of attributes t ij and t ik , the preference for t ij is said to dominate the preference for t ik if the user prefers t ij more than t ik ; The elements in the educational resource attribute are taken as the vertices of a directed graph, and the dominant relationship is taken as the edge of the directed graph to form a dominant relationship graph. The adjacency matrix of the dominant relationship graph is a dominant relationship matrix. The attribute value preference degree is sorted based on the dominant relationship matrix, and then the attribute value preference degree sorting vector is used to represent the user preference vector model; The principle of entropy reduction maximization for constructing the preference vector model for new users is that the missing values in the dominant relationship matrix are filled by calculating the historical statistical data of all existing users. By comparing the information entropy changes before and after replacing a certain filling value, the key attribute with the highest matrix entropy change is selected and submitted to the user for evaluation; S2, based on the user preference vector model, finding a substitute resource when the corresponding resource of a certain attribute value in the educational resource library is missing; S3, combining the learning expectation in the educational process and the personalized characteristics of knowledge points, and recommending the educational resource based on the user preference vector model. 2.The education resource recommendation method based on a user preference vector model according to claim 1, characterized in that, The two preference dominant matrices corresponding to the key attribute selection before and after the user is submitted to obtain the preference satisfy the first-order difference matrix relationship; The first-order difference matrix relation is expressed as: matrix R 1, R 2 is a dominant relation matrix, R 3 is the difference between matrix R 1 and R 2: R 3= R 1- R 2, if R 3 has and only has one element that is not 0, then R 1 is called the first-order difference matrix of R 2; For the dominant relation matrix R =( r ij ), the information entropy H ( R ) calculation formula is: ； wherein r ij denotes the attribute value t i to t j the dominant case, p ( r ij ) denotes t i to t j the dominant probability; The entropy reduction calculation formula is: Δ H ( R 1, R 2)=| H ( R 1)- H ( R 2)|; wherein Δ H ( R 1, R 2) is the difference between the entropy values of the matrices R 1 and R 2, and the element position corresponding to the maximum entropy reduction is given by the formula: i , j )=argmax(Δ H ( R 1, R 2)) The attribute corresponding to the position of the maximum entropy reduction element is submitted to the user for evaluation. The user makes a preference selection for the given preference attribute, fills in the elements in the user preference dominant relationship matrix, and constructs the user preference vector model according to the user preference dominant relationship matrix. 3.The education resource recommendation method based on a user preference vector model according to claim 1, characterized in that, The user-resource preference score matrix G in step S1, which is oriented to the old user, contains the user u e i The corresponding score of the resource; the resource-attribute matrix F contains the resource e i t ij Whether the resource has the value of the attribute; the inner product operation is performed on the user-resource preference score matrix G and the resource-attribute matrix F, the preference degree of the user to the resource attribute is calculated according to the score of the user to the resource; and the formula for calculating the preference degree of the user to the media attribute based on statistical average is:​​ ； wherein I u,t indicates that the resource has t a set of attribute features, r u,e indicates that the user u has an actual rating for the resource e ; For Matrix user-resource score matrix G and resource-attribute matrix F then user preference vector model M is: M = GF * Q； wherein GF denotes G the inner product, F while * denotes the Hadamard product, Q is a user rating matrix R having attributes t | the number of resources in the set I u，t | the inverse (1 / | I u,t |); the user preference vector model is constructed by combining the historical rating information of the user and the resources and the attribute features contained in the resources to calculate the user's preference for the attribute features of the resources. 4.The education resource recommendation method based on a user preference vector model according to claim 1, characterized in that, The way of looking for substitute resource in step S2: the education resource attribute vector V Adopt binary code, when the education resource has n attributes, use binary string with length n to express: V =(( v 11 , v 12 ,...,),...,( v n1 , v n2 ,...,)) Each bit corresponds to whether a specific value of an attribute exists, that is, a bit of 1 indicates that the corresponding value is included in the educational resource attribute, and 0 indicates that the corresponding value is not included. It is easy to know that the educational resource has at most kinds of forms; Adopting user preference vector M Dot product calculation with education resource attribute vector V Substitute resource e r * Select substitute resource, which corresponding education resource attribute vector satisfies dot product maximum, namely satisfy following formula: ； wherein, E R represents an educational resource, M · V represents a dot product, d(u, e r represents a vector distance between a user and an educational resource in an educational resource attribute space, with 1 / M · V represents a distance, the smaller the distance, the greater the dot product.

5. The educational resource recommendation method based on the user preference vector model according to claim 4, characterized in that: The educational resource recommendation method in step S3 includes: The recommended educational resources meet the preferences of the user preference vector model for new knowledge points, and the recommended educational resources meet the preferences e r * Dot product maximum: ； In the face of old knowledge point recommendation away from the user preference vector model of challenging educational resources, the recommended educational resources e r * Dot product minimum: 。

Images