CN114564639A - Course recommendation method based on deep session interest interaction model - Google Patents

Course recommendation method based on deep session interest interaction model Download PDF

Info

Publication number
CN114564639A
CN114564639A CN202210114867.4A CN202210114867A CN114564639A CN 114564639 A CN114564639 A CN 114564639A CN 202210114867 A CN202210114867 A CN 202210114867A CN 114564639 A CN114564639 A CN 114564639A
Authority
CN
China
Prior art keywords
user
interest
course
representation
dynamic
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202210114867.4A
Other languages
Chinese (zh)
Inventor
刘铁园
吴琼
古天龙
常亮
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Guilin University of Electronic Technology
Original Assignee
Guilin University of Electronic Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Guilin University of Electronic Technology filed Critical Guilin University of Electronic Technology
Priority to CN202210114867.4A priority Critical patent/CN114564639A/en
Publication of CN114564639A publication Critical patent/CN114564639A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/953—Querying, e.g. by the use of web search engines
    • G06F16/9535—Search customisation based on user profiles and personalisation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/10—Services
    • G06Q50/20—Education
    • G06Q50/205—Education administration or guidance

Landscapes

  • Engineering & Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Databases & Information Systems (AREA)
  • Theoretical Computer Science (AREA)
  • Tourism & Hospitality (AREA)
  • Physics & Mathematics (AREA)
  • Educational Administration (AREA)
  • Educational Technology (AREA)
  • Strategic Management (AREA)
  • General Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Primary Health Care (AREA)
  • Marketing (AREA)
  • General Business, Economics & Management (AREA)
  • Human Resources & Organizations (AREA)
  • Economics (AREA)
  • Health & Medical Sciences (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention discloses a course recommendation method based on a deep session interest interaction model, which is characterized by mainly comprising the following steps of: screening and preprocessing user data, sequencing behavior information of users and projects according to time, and dividing sessions by taking one day as a time interval; in order to depict the interest of the dynamic change of the user, enrich the interest representation of the user, and apply GRU to capture the dynamic preference of the user; next, inputting the recent behavior data and the dynamic interest representation of the user into a second-layer Attention network to obtain a multi-angle interest representation of the user; and finally, performing inner product on multi-angle interest representation and course vector representation of the user, and selecting each candidate item with high score to recommend the student, thereby solving the problems that the influence of noise items existing in the interaction process of the user and the items is not considered and static and low-rank vectors cannot sufficiently express the interest of the user in the current course-based recommendation method.

Description

Course recommendation method based on deep session interest interaction model
(I) the technical field
The invention relates to the technical fields of machine learning, deep learning, data mining and the like, in particular to a course recommendation method for capturing multi-angle interest preferences of a user.
(II) background of the invention
In recent years, the traditional offline education mode is stressed by the rising labor cost, meanwhile, the personalized requirements of consumers cannot be met, and the challenges of high cost, low profit and the like faced by offline education are increasingly highlighted under the background. And the rapid development of the internet and artificial intelligence is accompanied, so that the online education network environment is greatly improved and promoted. Online education refers to a method for rapid learning and content dissemination through internet technology. The existing online education platform MOOC is popular in online cloud classrooms and the like. Compared with the traditional education mode, the online education has the characteristics of high efficiency, convenience, low threshold and sufficient education resources. However, a large number of courses exist in the online learning platform, so that users are often overwhelmed, and the difficulty of decision making is increased. In order to solve the problem of information overload, course recommendation systems have come to work, aiming to provide personalized services for appropriate users at appropriate times.
The invention discloses a course recommending method and system based on dynamic weight of a graph-convolution neural network, which is published at present, and is disclosed as CN110580314A, wherein a user-course matrix is obtained by acquiring the score value of each course of a user and preprocessing the score value, the graph-convolution neural network is constructed on the basis of the user-course matrix, the user-course score matrix is predicted, and the user-course score matrix is subjected to sequence mode mining to obtain the recommended course sequence of each user. The method does not consider the noise course in the user interactive course, which can cause inaccurate recommending result and dynamic change of the user interest, but not invariable. The invention describes a course recommending method based on a deep conversation interest interaction model, which utilizes time sequence information of user and project interaction, removes noise courses of a user in course sessions by using an attention network, extracts the preference of each session, then simulates the interest of the user in dynamic change by using GRU, finally obtains the interest representation of the user, and obtains the score of each course by combining the course representation so as to recommend the course to students.
Disclosure of the invention
The invention aims to solve the problem that the influence of noise items existing in the interaction process of a user and the items is not considered in the interest extraction process of user conversation in the current course recommendation method; meanwhile, a static and low-rank vector cannot fully express the interest of the user, and the interest of the user is not constant and changes along with time, so that the generated recommendation model cannot recommend personalized courses for the user, and the like.
In order to solve the problems, the invention is realized by the following technical scheme:
and screening and preprocessing the original online learning related behavior data of the downloaded good MOOCCube data set.
And sequencing the behavior information of the user and the project according to time, and dividing the conversation by taking one day as a time interval.
Since all items in a conversation may not be the items that the user really wants to interact with, the attention network is applied to cull noise courses for each conversation and extract the interest of each conversation.
In order to characterize the dynamically changing interest of the user, the user interest expression is enriched, and the GRU is good at capturing the sequence relation, so the GRU is applied to capturing the dynamic preference of the user.
Inputting the latest behavior data of the user and the dynamic interest representation obtained in the fourth step into the second layer Attention network to obtain the multi-angle interest representation of the user.
And finally, performing inner product on the multi-angle interest representation and the course vector representation of the user, obtaining the score of each candidate item, and selecting the score with high score to recommend to the student.
Compared with the prior art, the invention has the following advantages:
in the aspect of data modeling, the time sequence behaviors of the user are divided, courses browsed by the user in one day are divided into one session, and the fact that the characteristics of the courses in the same session are highly homogeneous, and the characteristics of the courses in different sessions are greatly different.
Based on the conversation divided by sequential time modeling, the model can distribute different weights to different items by using an attention mechanism, can effectively eliminate noise courses, extracts interest point information contained in the conversation, and is beneficial to improving the accuracy of recommendation.
In the aspect of user interest modeling, interest may change due to the influence of various factors on the time interval of the conversation of a user, and a learning behavior is a continuous behavior.
After considering the dynamically changing user interest, the model also notes that items recently viewed by the user are also an important contributor to course recommendations. The invention introduces a second layer of Attention, distributes different weights for the dynamic interest representation of the user and the recently interacted items, and finally obtains the multi-angle interest representation of the user, thereby improving the performance.
Description of the drawings
FIG. 1 is a block diagram of a model of the present invention.
FIG. 2 is a flowchart of an overview of course recommendation based on a deep session interest interaction model.
FIG. 3 is a diagram illustrating extraction of session interests.
Fig. 4 is a flow chart of input data generation.
FIG. 5 is a flow chart of extracting session interests.
FIG. 6 is a flow chart of user dynamic interest generation.
FIG. 7 is a flow chart of multi-angle preference generation for a user.
(V) detailed description of the preferred embodiments
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail below with reference to the accompanying drawings in combination with specific examples.
The invention describes the specific implementation process of the method by taking course recommendation based on the deep session interest interaction model as an example. The model framework of the invention is shown in FIG. 1, and the overall process of course recommendation based on the deep session interest interaction model is shown in FIG. 2. The specific steps are explained by combining a schematic diagram:
Step 1, downloading a MOOCCube data set in an MOOCData official network, and preprocessing the data after screening.
And 2, after the screened data are obtained in the step 1, arranging the data of the students according to a time sequence, and finally dividing the sequence behavior of each user into sessions.
And 3, embedding the student course records in each session of the user into an embedding layer to obtain a user session vector matrix, and obtaining a session interest vector through an Attention layer, wherein the session interest vector is shown in fig. 3.
And 4, inputting the interest of each session into a GRU module, and simulating the interest of the dynamic change of the user so as to obtain a dynamic interest vector of the user.
And step 5, considering that the influence of each session interest item on the user is possibly different, and the items recently browsed by the user are important factors influencing the preference of the user, taking the recently browsed item set as short-term interest preference, and simultaneously inputting the short-term interest preference and the user dynamic interest vector preference into the Attention layer to obtain multi-angle user interest vector representation.
And 6, performing inner product on the multi-angle preference representation obtained in the step 5 and the candidate course vectors to obtain candidate course scores, sequencing the candidate courses, and selecting the top 10 courses to realize course recommendation.
Fig. 4 shows a flowchart of the input data generation of the present example, which includes the following specific steps:
step 1, the paper published in ACL2020 discloses an open data warehouse for large-scale online education, and this data set includes 706 real online courses, 38181 teaching videos, 114563 concepts, hundreds of thousands of courses for 199999 MOOC users, and resource libraries related to course concepts such as video viewing records. Course data and student behavior data come from the real use environment of the classroom online. The data set is subjected to primary processing, useless data are cleaned, and user.csv files in the selected data set contain the records of the learning behaviors of students. csv contains course information including the course name, the course type, the course profile and other relevant information. The student's learning behavior record is composed of different attributes, which have different meanings. For example, in the user.csv dataset, id represents the student number, name represents the student name, court _ order represents the lesson that the student watched, and enroll _ time represents the time that the student watching the corresponding lesson occurred. Csv, id denotes the course's course _ id, name denotes the course's information, course _ type denotes the type to which the corresponding course belongs, course _ info denotes the course profile, and video _ order denotes the order of the videos contained in this course.
And 2, further processing the two data sets, firstly arranging student behavior records in the user csv data set according to a time sequence, screening users with the number of watching courses larger than 10, then dividing sessions for the student behaviors by taking one day as a unit, and screening users with the number of sessions larger than 4 again. For the court.csv data set, only three types of data, namely user _ id, court _ id and roll _ time, are reserved, and other types of data are discarded. Finally, after data preprocessing, 14580 pieces of student behavior data meeting requirements, including 994 students and 632 courses, are obtained. Resulting in the final desired data set.
Through the processing, the conversation of the user at the corresponding time stamp is achieved, and due to the fact that the behavior data of the user has the following characteristics that firstly, a large number of noise items exist in each conversation of the user, secondly, the items are highly homogeneous in the same conversation, therefore, the noise course is removed for each conversation by applying the attention network, and the interest of each conversation is extracted.
FIG. 5 is a flowchart showing the steps of extracting session interests, including:
step 1, for each user U e U, dividing the session from the sequential data by day unit,
Figure BDA0003495856390000031
Figure BDA0003495856390000032
Representing the user's sequential sessions, where N represents the total number of sessions made by user u,
Figure BDA0003495856390000033
Figure BDA0003495856390000034
representing the nth session (set of items) of user u. Using the last session in the model
Figure BDA0003495856390000035
To extract the user's short-term preferences, which is also an important factor in recommending courses for the user. On the other hand, for the dynamic preference of the user, the user can select the dynamic preference
Figure BDA0003495856390000036
And (4) extracting. The invention inputs the course vector in the conversation into the Attention layer, screens out the course noise through the Attention, in form, pay Attention to the definition of the network:
Figure BDA0003495856390000037
Figure BDA0003495856390000038
wherein the content of the first and second substances,
Figure BDA0003495856390000039
are parameters of the model. We assume that the set of items in the conversation are successively marked from 1 to W, and cjA dense embedded vector representing item j. Will course vector cjObtaining a hidden representation h as input to a multilayer perceptron (MLP)1j. The function Φ (·) is the RELU nonlinear activation function.
And 2, taking the embedded u of the user as a context vector, and measuring the attention score of the user by using a softmax function. Finally, the session features represent
Figure BDA00034958563900000310
Our invention applies GRUs to model the dynamically changing interests of users, considering that the interests of users are not constant and dynamically change over time.
FIG. 6 is a flowchart showing the dynamic interest generation of a user, and the specific steps include:
Step 1, obtaining conversation characteristics IkInput into a gated neural unit (GRU) and modeling variable length sequence data can handle temporal dynamics of user item interactions and sequential patterns of user behavior, and can solve the problem of "long term dependence".
And 2, modeling the change of the session interest by applying GRU. The formula for GRU is as follows:
zt＝σ(Wz·[ht-1，It])
rt＝σ(Wr·[ht-1，Ii])
Figure BDA0003495856390000041
Figure BDA0003495856390000042
where σ (·) is the sigmoid function, z and r are the update gate and reset gate, respectively, and their vector magnitude and input ItAnd the same, wherein (·) represents dot product and (·) represents cross product.
Step 3, h obtained in step 2tIs a hidden state of the GRU, representing the user's dynamic preferences.
In addition to the dynamic interests of the user, the short-term interests of the user should be considered, firstly, ignoring the short-term interests may lead to inaccurate recommendation results, secondly, when in sparse data sets, since the data set demand of the short-term preferences is small, the short-term benefits may be a main determinant for predicting the next item and relieving the sparseness, and finally, the short-term preferences may effectively capture the interests of the user in each dimension in terms of fine granularity. The dynamic interests of the user need to be inferred in their long-term behavior, while the short-term interests of the user are reflected in their short-term behavior and captured by the embedding of the user.
FIG. 7 is a flowchart showing the multi-angle preference generation of a user, comprising the following steps:
step 1, inputting the dynamic preference ht of the user into a second layer of the Attention layer, and assigning the weight to the Attention layer.
Step 2, similar to modeling user dynamic preferences, we also turn to attention network, assigning different weights for long-term representation and item embedding in short-term item sets to capture high-level representations of users.
Figure BDA0003495856390000043
Figure BDA0003495856390000044
Wherein
Figure BDA0003495856390000045
Is a parameter of the model, and j > 0, cjIs a course of short-term preference sets
Figure BDA0003495856390000046
When j is 0, cj＝htSimilar to the user interest extraction layer, user embedding is used as a context vector to obtain a mixed representation of user interest.
The computational hybrid user representation is as follows:
Figure BDA0003495856390000047
wherein, beta0Is a weight of long-term user interest preferences. u. oftIs a final user interest representation that considers both long-term dynamic preferences and short-term preferences and assigns them different weights using an attention mechanism, which can capture non-linear interactions between the user and the item at the same time.
Step 3, in order to recommend courses to the user, after obtaining the user interest representation, calculating a preference score of the candidate item j as follows:
Rujt＝utcj
we then rank the candidate courses according to the user's preference score for the candidate purpose, the purpose of the model being to provide the user with a ranked list of courses at time t. Thus, the model utilizes a pairwise ordered objective function for our model, according to the optimization criteria of BPR. We assume that the user compares to item c, which is not observed kPreference model recommended candidate item cjTherefore, the following inequality is present:
Rujt＞Rukt
wherein R isuktIs the user not observing the item c at time tkThe preference score of (1). Thus we have a paired set of items
Figure BDA0003495856390000048
We then trained the model by maximizing the a posteriori (MAP), as follows:
Figure BDA0003495856390000049
wherein Θ is { U, C, W ═ W1，W2，b1，b2Is the parameter set of the model, σ is the logistic regression function, ΘucIs a set of embedded matrices for users and courses, Θa＝{W1，W2，Wz，WrW is a set of weight matrices for attention layer and GRU, λ ═ λuc，λaIs the regularization parameter.
It should be noted that, although the above-mentioned embodiments illustrate the present invention, the present invention is not limited thereto, and thus the present invention is not limited to the above-mentioned embodiments. Other embodiments, which can be made by those skilled in the art in light of the teachings of the present invention, are considered to be within the scope of the present invention without departing from its principles.

Claims (5)

1. A course recommendation method based on a deep session interest interaction model is characterized in that: screening and preprocessing user data, sequencing behavior information of users and projects according to time, and dividing sessions by taking one day as a time interval; in order to depict the interest of the dynamic change of the user, enrich the interest representation of the user, and apply GRU to capture the dynamic preference of the user; next, inputting the recent behavior data and the dynamic interest representation of the user into a second-layer Attention network to obtain a multi-angle interest representation of the user; and finally, performing inner product on the multi-angle interest representation and the course vector representation of the user, and selecting each candidate item with high score to recommend the student. Therefore, the problems that the influence of noise items existing in the interaction process of the user and the items is not considered and static and low-rank vectors cannot sufficiently express the interest of the user in the current course-based recommendation method are solved.
2. The course recommendation method based on the deep session interest interaction model as claimed in claim 1, wherein: in the aspect of data modeling, the invention divides the time sequence behaviors of the user, divides the courses browsed by the user in one day into one session, and divides the course characteristics in the same session into a plurality of sessions, wherein the course characteristics are highly homogeneous, and the course characteristics are greatly different among different sessions.
3. The course recommendation method based on the deep session interest interaction model as claimed in claim 1, wherein: based on the conversation divided by sequential time modeling, the model can distribute different weights to different items by using an attention mechanism, can effectively eliminate noise courses, extracts interest point information contained in the conversation, and is beneficial to improving the accuracy of recommendation.
4. The course recommendation method based on the deep session interest interaction model as claimed in claim 1, wherein: in the aspect of user interest modeling, interest may change due to the influence of various factors on the time interval of the conversation of the user, and the learning behavior is a continuous behavior, so that the experience of online learning of students is improved by simulating the change of interest between each conversation of the user by using the GRU.
5. After considering the dynamically changing user interest perspective, the model also notes that items recently viewed by the user are also an important contributor to the course recommendations. The invention introduces a second layer of Attention, distributes different weights for the dynamic interest representation of the user and the recently interacted items, and finally obtains the multi-angle interest representation of the user, thereby improving the performance.
CN202210114867.4A 2022-01-31 2022-01-31 Course recommendation method based on deep session interest interaction model Pending CN114564639A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210114867.4A CN114564639A (en) 2022-01-31 2022-01-31 Course recommendation method based on deep session interest interaction model

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210114867.4A CN114564639A (en) 2022-01-31 2022-01-31 Course recommendation method based on deep session interest interaction model

Publications (1)

Publication Number Publication Date
CN114564639A true CN114564639A (en) 2022-05-31

Family

ID=81714712

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210114867.4A Pending CN114564639A (en) 2022-01-31 2022-01-31 Course recommendation method based on deep session interest interaction model

Country Status (1)

Country Link
CN (1) CN114564639A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113435685A (en) * 2021-04-28 2021-09-24 桂林电子科技大学 Course recommendation method of hierarchical Attention deep learning model

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113435685A (en) * 2021-04-28 2021-09-24 桂林电子科技大学 Course recommendation method of hierarchical Attention deep learning model

Similar Documents

Publication Publication Date Title
Manovich AI aesthetics
WO2020228514A1 (en) Content recommendation method and apparatus, and device and storage medium
WO2020207196A1 (en) Method and apparatus for generating user tag, storage medium and computer device
CN110781321B (en) Multimedia content recommendation method and device
CN109299372B (en) Intelligent course selection recommendation method based on collaborative filtering
CN108664658B (en) Collaborative filtering video recommendation method considering dynamic change of user preference
CN111241394B (en) Data processing method, data processing device, computer readable storage medium and electronic equipment
Tang et al. Time series model for predicting dropout in massive open online courses
Wang et al. Attention-based CNN for personalized course recommendations for MOOC learners
CN111859149A (en) Information recommendation method and device, electronic equipment and storage medium
CN113435685A (en) Course recommendation method of hierarchical Attention deep learning model
Belarbi et al. User profiling in a SPOC: A method based on user video clickstream analysis
CN111563770A (en) Click rate estimation method based on feature differentiation learning
Aryal et al. MoocRec: Learning styles-oriented MOOC recommender and search engine
Sakboonyarat et al. Massive open online courses (MOOCs) recommendation modeling using deep learning
Hu et al. Learning peer recommendation using attention-driven CNN with interaction tripartite graph
CN114564639A (en) Course recommendation method based on deep session interest interaction model
Sanchez-Corcuera et al. Persuasion-based recommender system ensambling matrix factorisation and active learning models
CN111311997B (en) Interaction method based on network education resources
CN114358807A (en) User portrayal method and system based on predictable user characteristic attributes
CN116662651A (en) Heterogeneous graph neural network course recommendation method and system based on attention mechanism
WO2023087933A1 (en) Content recommendation method and apparatus, device, storage medium, and program product
Oubalahcen et al. The Use of AI in E-Learning Recommender Systems: A Comprehensive Survey
Alharbi et al. Social recommender system for predicting the needs of students/instructors: Review and proposed framework
Wang et al. MOOC resources recommendation based on heterogeneous information network

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication