Learning auxiliary system based on artificial intelligence

Abstract

The invention provides a learning assistance system based on artificial intelligence. The learning assistance system comprises a client and a server, the client is used for inputting personalized data; the data input mode comprises voice input, video shooting and manual typing; the server comprises an artificial intelligence analysis unit and a learning suggestion and plan output unit; analyzing the collected data by using a machine learning algorithm, cleaning the data, describing the data, checking the distribution of the data, comparing the relationship between the data, cultivating the intuition of the data, summarizing the data, and deducing the learning habit and the understanding degree of some courses; on the basis of the knowledge graph, personalized learning resources such as courses, videos and articles are intelligently recommended according to learning characteristics and requirements of students; the system is superior to manual tutoring, and does not have the problems of level difference, emotion difference, standing prejudice and the like like manual tutoring, and does not have a single function or a basic function like existing common learning tools; the system can deeply supervise the learning process, formulate a personalized learning plan and the like, and comprehensively manage the final learning result.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a learning assistance system based on artificial intelligence. Background Technology

[0002] Currently, most tutoring is still conducted manually, that is, with the assistance and guidance of teachers, parents or high-achieving classmates, in order to improve students' academic performance.

[0003] The biggest problem with human assistance is that teaching or tutoring levels vary from person to person. Students taught by teachers and parents of different levels will have different outcomes. Secondly, because people have emotions (they may intentionally or unintentionally pass on their negative emotions to students when they are unhappy) and biases (they may dislike student A or favor student B, etc.), these subjective factors can easily affect the effectiveness of teaching and tutoring, ultimately leading to the assistance and guidance failing to achieve the desired effect of improving students' grades.

[0004] In the existing technology, there are also some auxiliary teaching tools such as reading pens, dictionary pens, and learning machines. The common feature of these products is that they can only play a certain or limited role.

[0005] For example, a reading pen may simply read aloud and translate words, and its effectiveness in learning is hard to guarantee; a learning machine may just be a question bank or video library, and still cannot provide individualized instruction for different students, so the learning outcome is not guaranteed; these tool products are at best icing on the cake, and cannot achieve the same effect as an excellent tutor. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, this invention provides an artificial intelligence-based learning assistance system that can create different learning plans tailored to each student's individual circumstances. This includes the collection and analysis of learning data, as well as the supervision and assessment of the learning process and the formulation of learning plans. The system architecture is scientifically designed and highly efficient, effectively and intelligently assisting students in completing learning tasks and improving their academic performance.

[0007] The technical solution adopted by the present invention to solve its technical problem is: an artificial intelligence-based learning assistance system, comprising: a client and a server;

[0008] Furthermore, the client is used for personalized data input; the data input methods include: voice input, video recording, and manual typing;

[0009] As an example, the client is used to interact with learners, such as for data collection and information feedback.

[0010] As an example, the client can be a mobile phone, tablet, computer APP, or a highly customized integrated hardware and software learning device.

[0011] As an example, to ensure that the system can proactively interact with learners from time to time (simple APP applications can only passively wait for people to open and use them), it is preferable that the client is a highly customized integrated hardware and software learning device. This device can keep the client program in an active state at all times, so as to proactively remind learners of learning plans and other tasks.

[0012] Furthermore, the server includes: an artificial intelligence analysis unit and a learning suggestion and plan output unit;

[0013] Furthermore, the overall architecture design of the artificial intelligence analysis unit includes:

[0014] The collected data is analyzed using machine learning algorithms, the data is cleaned, described, its distribution is examined, relationships between data are compared, data intuition is developed, and the data is summarized to infer learning habits and the level of understanding of certain courses.

[0015] Specifically, the machine learning algorithm architecture is designed as follows:

[0016] Algorithm ① Normalization;

[0017] Data smoothing is achieved by offfitting the data;

[0018]

[0019] Where: x Min Let x be the minimum value of the variable x. Max This represents the maximum value of the variable x.

[0020] Algorithm ② Standardization;

[0021] Data consistency is ensured through standardization.

[0022]

[0023] μ is the mean of the variable x, and σ is the standard deviation of the variable x.

[0024] Algorithm ③ Approximate Inference;

[0025] Based on the Double Robust method; infer habit type, X and W are both confounding variables, but only X is a relevant variable in habit type; t is treatment, i.e. the action applied, and the value of t is 1 or 0, representing different states;

[0026] Step 1: Use X, W, T to predict Y, and define the predicted value as g. t (X,W);

[0027] The second step is to use a classification model to predict the propensity score of T using X and W, defined as p. t (X,W);

[0028] Where: T is a discrete variable, i represents the index of the smallest physical unit of the model, T i The action applied to the i-th unit (i.e., T) i =t), and restrict g t (X,W) is a regression-based model;

[0029] After the first and second steps described above, calculate an adjusted result variable:

[0030]

[0031] Then adjust Y i,t DR Calculate the difference Y between the experimental group and the control group. i1 DR -Y i0 DR Regression X yields the habit type;

[0032] Algorithm 4: Building a knowledge graph: Build a knowledge graph to classify subject knowledge, record the relationships between knowledge points, and provide students with a systematic learning path;

[0033] Furthermore, constructing the knowledge graph is an iterative update process. According to the logic of knowledge acquisition, each iteration includes four stages: knowledge storage, information extraction, knowledge fusion, and knowledge computation.

[0034] I. Knowledge Storage: A graph database is used to store various types of knowledge, including: basic attribute knowledge, relational knowledge, event knowledge, time-series knowledge, and resource knowledge, etc.

[0035] II. Information Extraction: Extract entities, attributes, and relationships between entities from various types of data sources, and form ontological knowledge representation based on this.

[0036] Based on the copy mechanism, relational facts are extracted through an end-to-end neural model. The Seq2Seq (input sequence, output sequence) framework is used to extract relations, head entities, and tail entities sequentially.

[0037] coding:

[0038] decoding:

[0039] in: This is the output of the encoded part at time m. Let x be the hidden layer state at time m of the encoded part. m This is the input at time m of the encoded part. This is the output of the decoding part at time m. For the hidden layer state at time m in the decoding part, u m This is the input at time m for the decoding part;

[0040] Relation extraction: The calculation can be performed using the Softmax function;

[0041] Definition of the Softmax function:

[0042]

[0043] Among them: Z i Let be the output value of the i-th node, C be the number of output nodes, i.e. the number of categories; e represent the prediction result; the Softmax function can be used to convert the output values ​​of multi-class classification into a probability distribution ranging from [0,1] to 1.

[0044] Header entity extraction: Select one entity from n tokens during the current decoding process.

[0045] For each token's encoding, add the output of the current decoding; w is the weight, and the meanings of the other parameters are the same as described above;

[0046] According to p e =Softmax([q e ;q NA From n tokens, select the one with the highest probability as the entity; NA represents N prediction results;

[0047] Tail entity extraction: Similar to head entity extraction, only the predicted head entity from the previous step needs to be masked.

[0048] III. Knowledge Integration: After acquiring new knowledge, integrate the new knowledge to eliminate contradictions and ambiguities. For example, some entities may have multiple expressions, and a particular term may correspond to multiple different entities.

[0049] Similarity calculation:

[0050] SIM(E1,E2)=ω1SIM(P1,P2)+ω2SIM(C1,C2)+ω3SIM(S1,S2)

[0051] Where: SIM() represents similarity, ω represents the weight of each feature, SIM(P1,P2), SIM(C1,C2), SIM(S1,S2) represent semantic similarity based on attribute labels, semantic similarity based on category label matching, and semantic similarity based on keywords in unstructured text, respectively. The entity with the highest similarity and greater than a certain threshold is taken as the alignment result output, and is regarded as having the same semantic orientation.

[0052] IV. Knowledge Computation: New knowledge that has undergone knowledge fusion needs to be evaluated for quality before it can be added to the knowledge base to ensure the quality of the knowledge base.

[0053] As an example, the quality assessment can be conducted with partial human intervention in the screening process.

[0054] The Continuous Bag-Of-Words (CBOW) model is used to convert entity description text information into entity representations, which are then used for objective function learning. A set of keywords containing the main concepts of the entities is extracted from the description text. The first n keywords are selected as input, and the encoded word vectors are simply added together to obtain the text representation: e = x1 + x2 + ... + x k

[0055] Where: x k This represents the embedding of the k-th word in the set of keywords belonging to entity e;

[0056] Furthermore, the overall architecture design of the learning suggestion and plan output unit includes:

[0057] Based on the knowledge graph, personalized learning resources, such as courses, videos, and articles, are intelligently recommended according to students' learning characteristics and needs.

[0058] The intelligent recommendation refers to achieving the best possible recommendation effect through recommendation algorithms.

[0059] As an example, the recommendation algorithm includes: content-based recommendation algorithm, collaborative filtering algorithm, association rule-based recommendation algorithm, and / or knowledge-based recommendation algorithm.

[0060] As an example, the collaborative filtering algorithm used for book similarity calculation is designed with the following formula:

[0061] Where: the denominator |N(i)| is the number of users who like book i, and the numerator is the number of users who like both book i and book j; therefore, the above formula can be understood as what percentage of users who like book i also like book j, and thus recommends the corresponding books to learners.

[0062] The beneficial effects of this invention are:

[0063] This system is superior to human tutoring because it is not subject to issues such as differences in skill level, emotions, and biases, nor does it have the limited or basic functions of common learning tools.

[0064] This system can deeply monitor the learning process, create personalized learning plans, and comprehensively manage the final learning outcomes, making it an excellent learning support system. With this system, parents can use it as a powerful tool for family learning guidance; not only will children not be unable to receive effective guidance due to parents' limited knowledge, but it will also free up parents' spare time.

[0065] This system provides teachers with a highly capable assistant; they no longer have to worry about not being able to take care of all students due to limited energy, and they can be freed from the heavy daily work of evaluating learning outcomes (such as grading test papers) to focus their energy on teaching research, which can be described as a win-win situation.

[0066] This system uses artificial intelligence technology to effectively assist students' learning, including but not limited to the collection and analysis of learning data, as well as the supervision of the learning process, the evaluation of learning outcomes, the provision of learning suggestions, and the creation of personalized learning plans for each student. Attached Figure Description

[0067] Figure 1 This is a schematic diagram of the structure of an artificial intelligence-based learning assistance system according to the present invention.

[0068] Figure 2 This is a schematic diagram illustrating an example of how a machine learning algorithm in an artificial intelligence-based learning assistance system describes the collected data.

[0069] Figure 3 This is a schematic diagram illustrating an example of a graph database in an artificial intelligence-based learning assistance system according to the present invention.

[0070] Figure 4 This is a schematic diagram illustrating an example of a collaborative filtering algorithm in an artificial intelligence-based learning assistance system according to the present invention. Detailed Implementation

[0071] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0072] Reference Figures 1 to 4 As shown, an artificial intelligence-based learning assistance system includes: a client 102 and a server 101;

[0073] Furthermore, the client 102 is used for personalized data input; the data input methods include: voice input, video recording, and manual typing;

[0074] As an example, the client 102 is used to interact with learners, such as data collection and information feedback.

[0075] As an example, the client 102 can be a mobile phone, tablet, computer APP application, or a highly customized integrated hardware and software learning device.

[0076] As an example, to ensure that the system can actively interact with learners from time to time (simple APP applications can only passively wait for people to open and use them), the client 102 is preferably a highly customized integrated hardware and software learning device. This device can keep the client program in an active state at all times, so as to actively remind learners of learning plans and other tasks.

[0077] Furthermore, the server 101 includes: an artificial intelligence analysis unit and a learning suggestion and plan output unit;

[0078] Furthermore, the overall architecture design of the artificial intelligence analysis unit includes:

[0079] The collected data is analyzed using machine learning algorithms, the data is cleaned, and the data is described (using descriptive statistics, charts, etc.). Figure 2 (As shown), examine the distribution of data, compare the relationships between data, cultivate intuition for data, summarize the data, and infer learning habits and the level of understanding of certain courses.

[0080] Specifically, the machine learning algorithm architecture is designed as follows:

[0081] Algorithm ① Normalization;

[0082] Data smoothing is achieved by offfitting the data;

[0083]

[0084] Where: x Min Let x be the minimum value of the variable x. Max This represents the maximum value of the variable x.

[0085] Algorithm ② Standardization;

[0086] Data consistency is ensured through standardization.

[0087]

[0088] μ is the mean of the variable x, and σ is the standard deviation of the variable x.

[0089] Algorithm ③ Approximate Inference;

[0090] Based on the Double Robust method; infer habit type, X and W are both confounding variables, but only X is a relevant variable in habit type; t is treatment, i.e. the action applied, and the value of t is 1 or 0, representing different states;

[0091] Step 1: Use X, W, T to predict Y, and define the predicted value as g. t (X,W);

[0092] The second step is to use a classification model to predict the propensity score of T using X and W, defined as p. t (X,W);

[0093] Where: T is a discrete variable, i represents the index of the smallest physical unit of the model, T i The action applied to the i-th unit (i.e., T) i =t), and restrict g t (X,W) is a regression-based model;

[0094] After the first and second steps described above, calculate an adjusted result variable:

[0095]

[0096] Then adjust Y i,t DR Calculate the difference Y between the experimental group and the control group. i1 DR -Y i0 DR Regression X yields the habit type;

[0097] Algorithm 4: Building a knowledge graph: Build a knowledge graph to classify subject knowledge, record the relationships between knowledge points, and provide students with a systematic learning path;

[0098] Furthermore, constructing the knowledge graph is an iterative update process. According to the logic of knowledge acquisition, each iteration includes four stages: knowledge storage, information extraction, knowledge fusion, and knowledge computation.

[0099] I. Knowledge Storage: Graph database is used (see appendix) Figure 3 The storage method shown can be used to store various types of knowledge, including: basic attribute knowledge, related knowledge, event knowledge, time sequence knowledge, and resource knowledge.

[0100] II. Information Extraction: Extract entities, attributes, and relationships between entities from various types of data sources, and form ontological knowledge representation based on this.

[0101] Based on the copy mechanism, relational facts are extracted through an end-to-end neural model. The Seq2Seq (input sequence, output sequence) framework is used to extract relations, head entities, and tail entities sequentially.

[0102] coding:

[0103] decoding:

[0104] in: This is the output of the encoded part at time m. Let x be the hidden layer state at time m of the encoded part. m This is the input at time m of the encoded part. This is the output of the decoding part at time m. For the hidden layer state at time m in the decoding part, u m This is the input at time m for the decoding part;

[0105] Relation extraction: The calculation can be performed using the Softmax function;

[0106] Definition of the Softmax function:

[0107]

[0108] Among them: Z i Let be the output value of the i-th node, C be the number of output nodes, i.e. the number of categories; e represent the prediction result; the Softmax function can be used to convert the output values ​​of multi-class classification into a probability distribution ranging from [0,1] to 1.

[0109] Header entity extraction: Select one entity from n tokens during the current decoding process.

[0110] For each token's encoding, add the output of the current decoding; w is the weight, and the meanings of the other parameters are the same as described above;

[0111] According to p e =Softmax([q e ;q NA From n tokens, select the one with the highest probability as the entity; NA represents N prediction results;

[0112] Tail entity extraction: Similar to head entity extraction, only the predicted head entity from the previous step needs to be masked.

[0113] III. Knowledge Integration: After acquiring new knowledge, integrate the new knowledge to eliminate contradictions and ambiguities. For example, some entities may have multiple expressions, and a particular term may correspond to multiple different entities.

[0114] Similarity calculation:

[0115] SIM(E1,E2)=ω1SIM(P1,P2)+ω2SIM(C1,C2)+ω3SIM(S1,S2)

[0116] Where: SIM() represents similarity, ω represents the weight of each feature, SIM(P1,P2), SIM(C1,C2), SIM(S1,S2) represent semantic similarity based on attribute labels, semantic similarity based on category label matching, and semantic similarity based on keywords in unstructured text, respectively. The entity with the highest similarity and greater than a certain threshold is taken as the alignment result output, and is regarded as having the same semantic orientation.

[0117] IV. Knowledge Computation: New knowledge that has undergone knowledge fusion needs to be evaluated for quality before it can be added to the knowledge base to ensure the quality of the knowledge base.

[0118] As an example, the quality assessment can be conducted with partial human intervention in the screening process.

[0119] The Continuous Bag-Of-Words (CBOW) model is used to convert entity description text information into entity representations, which are then used for objective function learning. A set of keywords containing the main concepts of the entities is extracted from the description text. The first n keywords are selected as input, and the encoded word vectors are simply added together to obtain the text representation: e = x1 + x2 + ... + x k

[0120] Where: x k This represents the embedding of the k-th word in the set of keywords belonging to entity e;

[0121] Furthermore, the overall architecture design of the learning suggestion and plan output unit includes:

[0122] Based on the knowledge graph, personalized learning resources, such as courses, videos, and articles, are intelligently recommended according to students' learning characteristics and needs.

[0123] The intelligent recommendation refers to achieving the best possible recommendation effect through recommendation algorithms.

[0124] As an example, the recommendation algorithm includes: content-based recommendation algorithm, collaborative filtering algorithm, association rule-based recommendation algorithm, and / or knowledge-based recommendation algorithm.

[0125] As an example, the collaborative filtering algorithm used for book similarity calculation is designed with the following formula:

[0126] Where: the denominator |N(i)| is the number of users who like book i, and the numerator is the number of users who like both book i and book j; therefore, the above formula can be understood as what percentage of users who like book i also like book j, and thus recommends the corresponding books to learners.

[0127] This system surpasses human tutoring in many ways. Unlike human tutors, it avoids issues such as varying skill levels, emotional differences, and biases. It also differs from common learning tools that offer only single or basic functions. This system can deeply monitor the learning process, create personalized learning plans, and comprehensively manage final learning outcomes, making it an excellent learning support system. With this system, parents can utilize it as a powerful tool for home learning guidance. Parents' limited knowledge levels will prevent children from receiving effective guidance, freeing up their spare time. This system provides teachers with a highly capable assistant, freeing them from the burden of assessing learning outcomes (such as grading tests) and allowing them to focus on teaching research – a win-win situation. Based on artificial intelligence technology, this system effectively assists student learning, including but not limited to the collection and analysis of learning data, monitoring the learning process, evaluating learning outcomes, providing learning suggestions, and creating personalized learning plans for each student.

[0128] The above description is only a preferred embodiment of the present invention. It should be understood that the above description of the embodiments is only for the purpose of helping to understand the method and core idea of ​​the present invention, and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, etc. made within the idea and principle of the present invention should be included within the scope of protection of the present invention.

Claims

1. A learning assistance system based on artificial intelligence, characterized in that, include: Client and server; The client is used for entering personalized data; Data entry methods include: voice input, video recording, and manual typing; The server includes: an artificial intelligence analysis unit and a learning suggestion and plan output unit; The overall architecture design of the artificial intelligence analysis unit includes: The collected data is analyzed using machine learning algorithms, the data is cleaned, described, its distribution is examined, relationships between data are compared, data intuition is developed, and the data is summarized to infer learning habits and the level of understanding of certain courses. The overall architecture design of the learning suggestion and plan output unit includes: Based on the knowledge graph, and considering students' learning characteristics and needs, intelligent recommendations provide personalized learning resources. These intelligent recommendations refer to achieving the best possible recommendation results through recommendation algorithms.

2. The artificial intelligence-based learning assistance system according to claim 1, characterized in that, The client is used to interact with learners, including data collection and information feedback.

3. The artificial intelligence-based learning assistance system according to claim 1, characterized in that, The client is a mobile phone, tablet, computer APP application or a highly customized integrated hardware and software learning device.

4. The artificial intelligence-based learning assistance system according to claim 1, characterized in that, To ensure that the system can proactively interact with learners at irregular intervals, the client is a highly customized integrated hardware and software learning device that keeps the client program always active, thus proactively reminding learners of their learning plans.

5. A learning assistance system based on artificial intelligence according to any one of claims 1-4, characterized in that, The machine learning algorithm architecture is designed as follows: Algorithm ① Normalization; Data smoothing is achieved by offfitting the data; Where: χ Min χ is the minimum value of the variable x. Max This represents the maximum value of the variable x. Algorithm ② standardization; Data consistency is ensured through standardization. μ is the mean of the variable x, and σ is the standard deviation of the variable x. Algorithm ③ Approximate Inference; Based on the Double Robust method; infer habit type, X and W are both confounding variables, but only X is a relevant variable in habit type; t is treatment, i.e. the action applied, and the value of t is 1 or 0, representing different states; Step 1: Use X, W, T to predict Y, and define the predicted value as g. t (X,W); The second step is to use a classification model to predict the propensity score of T using X and W, defined as p. t (X,W); Where: T is a discrete variable, i represents the index of the smallest physical unit of the model, T i The action applied to the i-th unit (i.e., T) i =t), and restrict g t (X,W) is a regression-based model; After the first and second steps described above, calculate an adjusted result variable: Then adjust Y i,t DR Calculate the difference Y between the experimental group and the control group. i1 DR -Y i0 DR Regression X yields the habit type; Algorithm 4: Building a knowledge graph: A knowledge graph is built to classify subject knowledge, record the relationships between knowledge points, and provide students with a systematic learning path.

6. The artificial intelligence-based learning assistance system according to claim 5, characterized in that, Constructing the knowledge graph is an iterative process. According to the logic of knowledge acquisition, each iteration includes four stages: knowledge storage, information extraction, knowledge fusion, and knowledge computation.

7. The artificial intelligence-based learning assistance system according to claim 6, characterized in that, The four stages specifically include: I. Knowledge Storage: A graph database is used to store various types of knowledge, including: basic attribute knowledge, relational knowledge, event knowledge, time-series knowledge, and resource knowledge; II. Information Extraction: Extract entities, attributes, and relationships between entities from various types of data sources, and form ontological knowledge representation based on this. Based on the copy mechanism, relational facts are extracted through an end-to-end neural model. The Seq2Seq (input sequence, output sequence) framework is used to extract relations, head entities, and tail entities sequentially. coding: decoding: in: This is the output of the encoded part at time m. Let x be the hidden layer state at time m of the encoded part. m This is the input at time m of the encoded part. This is the output of the decoding part at time m. For the hidden layer state at time m in the decoding part, u m This is the input at time m for the decoding part; Relation extraction: The calculation can be performed using the Softmax function; Definition of the Softmax function: Among them: Z i Let be the output value of the i-th node, C be the number of output nodes, i.e. the number of categories; e represent the prediction result; the Softmax function can be used to convert the output values ​​of multi-class classification into a probability distribution ranging from [0,1] to 1. Header entity extraction: Select one entity from n tokens during the current decoding process. For each token's encoding, add the output of the current decoding; w is the weight; According to p e =Softmax([q e ;q NA From n tokens, select the one with the highest probability as the entity; NA represents N prediction results; Tail entity extraction: Similar to head entity extraction, only the predicted head entity from the previous step needs to be masked. III. Knowledge Integration: After acquiring new knowledge, integrate the new knowledge to eliminate contradictions and ambiguities. For example, some entities may have multiple expressions, and a particular name may correspond to multiple different entities. Similarity calculation: SIM(E1,E2)=ω1SIM(P1,P2)+ω2SIM(C1,C2)+ω3SIM(S1,S2) Where: SIM() represents similarity, ω represents the weight of each feature, SIM(P1,P2), SIM(C1,C2), SIM(S1,S2) represent semantic similarity based on attribute labels, semantic similarity based on category label matching, and semantic similarity based on keywords in unstructured text, respectively. The entity with the highest similarity and greater than a certain threshold is taken as the alignment result output, and is regarded as having the same semantic orientation. IV. Knowledge Computation: New knowledge that has undergone knowledge fusion needs to be evaluated for quality before it can be added to the knowledge base to ensure the quality of the knowledge base. The continuous bag-of-words model is used to convert entity description text information into entity representations, and then the entity representations are used for objective function learning. A set of keywords containing the main concepts of the entities is extracted from the description text, and the first n keywords are selected as input. The encoded word vectors are then simply added together to obtain the text representation: e = x1 + x2 + ... + x k Where: x k This represents the embedding of the k-th word in the set of keywords belonging to entity e.

8. The artificial intelligence-based learning assistance system according to claim 7, characterized in that, The recommendation algorithms include: content-based recommendation algorithms, collaborative filtering algorithms, association rule-based recommendation algorithms, and / or knowledge-based recommendation algorithms.

9. A learning assistance system based on artificial intelligence according to claim 8, characterized in that, The collaborative filtering algorithm is used for book similarity calculation, and the algorithm formula is designed as follows: Where: the denominator |N(i)| is the number of users who like book i, and the numerator is the number of users who like both book i and book j; therefore, the above formula can be understood as what percentage of users who like book i also like book j, and thus recommends the corresponding books to learners.

Images