CN105138653A - Exercise recommendation method and device based on typical degree and difficulty - Google Patents

Abstract

The invention discloses an exercise recommendation method based on the typical degree and difficulty. The method comprises the steps that exercise doing situations comprising the exercise number, difficulty and the passing rate of each target user on each type of exercises are calculated; according to feature vectors of the target users, the similarities between every two feature vectors of the target users are calculated, and many users with the highest similarity are selected for each target user as nearest neighbours; according to exercise doing situation of the nearest neighbour users, the scores of the target users for exercise not finished are predicted; a plurality of target exercises with the highest scores are selected as recommendation results of the target users. An exercise recommendation device comprises a calculating module, a first selecting module, a scoring module and a second selecting module. A traditional recommendation method is improved through the feature that exercises have the difficulty, and a good effect is obtained. The exercise recommendation method and device can be applied in a learning website in the future to help users to select learning content and formulate individualized learning schemes.

Description

A kind of exercise question recommend method based on typical degree and difficulty and recommendation apparatus thereof

Technical field

The present invention relates to data mining, machine learning and information retrieval field, relate to the recommendation field of collaborative filtering, particularly relate to a kind of exercise question recommend method based on typical degree and difficulty and recommendation apparatus thereof.

Background technology

At present, there are commending system and the proposed algorithm of comparative maturity, the proposed algorithm of current main-stream is mainly divided into collaborative filtering recommending (collaborativefiltering, CF), content-based recommendation (contentbased, CB) and mixing recommend method (hybridmethods).Mixing proposed algorithm namely comprehensive the above two, reach better effect.

In content-based recommendation system, article can be described to the vector of a series of property value, and the property value describing article characteristics is called as " content ".Content-based recommendation system is exactly the history scoring behavior according to user, finds user preference, and recommends and the akin article of its preference.This method is mainly used in the article recommending to describe with text, such as: documents and materials, news etc.

Collaborative filtering predicts the scoring of user to unknown article by the similarity between the similarity between user or article.Between Main Basis user, the arest neighbors be called as based on user of similarity is recommended; Between Main Basis article, the arest neighbors be called as based on article of similarity is recommended.The prerequisite that collaborative filtering recommending method is set up is that the hobby of hypothesis user remains unchanged for a long period of time.

No matter be content-based recommendation system, or traditional Collaborative Filtering Recommendation System, there is it not enough when applying it to knowledge recommendation field.Content-based recommendation system wants to describe exactly the feature of recommended article, and it is mapped with user preference, when facing the comparatively fuzzy data of the such feature of similar exercise question, is just difficult to accurate description, to cause recommendation inaccurate.Tradition collaborative filtering method does not need accurate description article, but needs a large amount of score data, when applying it to recommendation knowledge, this one side of recommendation exercise question, is difficult to enough score data.

Summary of the invention

The invention provides a kind of exercise question recommend method based on typical degree and difficulty and recommendation apparatus thereof, the present invention effectively can overcome conventional recommendation technology when being applied in knowledge recommendation, article characteristics is difficult to describe, score information is few and do not take into full account the technical matters of this key character of item difficulty, described below:

Based on an exercise question recommend method for typical degree and difficulty, described exercise question recommend method comprises the following steps:

Calculate each targeted customer and do topic situation on each type exercise question, comprising: the quantity of exercise question, difficulty and the percent of pass of targeted customer on this type exercise question;

According to targeted customer's proper vector, calculate the similarity between arbitrary target user characteristics vector, and to each targeted customer, select user that some similarities are the highest as arest neighbors;

Foundation arest neighbors user does topic situation, and target of prediction user is to the scoring not doing exercise question;

To targeted customer, the target exercise question selecting some scorings the highest is as the recommendation results of targeted customer.

Wherein, described exercise question recommend method also comprises:

Debug or invalid data, then sort from small to large by submission time;

The user that statistics relates to and exercise question, form user's set and an exercise question set.

Wherein, described targeted customer's proper vector is specially:

＜type ₁:typicality ₁,type ₂:typicality ₂,…,type _i:typicality _i…,type _n:typicality _n＞

Wherein, type _irepresent topic types; Typicality _ithat targeted customer is at type _itypical degree on type exercise question; I is topic types numbering; N is topic types sum.

Based on an exercise question recommendation apparatus for typical degree and difficulty, described exercise question recommendation apparatus comprises:

Computing module, doing topic situation for calculating each targeted customer, comprising: the quantity of exercise question, difficulty and the percent of pass of targeted customer on this type exercise question on each type exercise question;

First selects module, for according to targeted customer's proper vector, calculates the similarity between arbitrary target user characteristics vector, and to each targeted customer, selects user that some similarities are the highest as arest neighbors;

Grading module, for doing topic situation according to arest neighbors user, target of prediction user is to the scoring not doing exercise question;

Second selects module, and for targeted customer, the target exercise question selecting some scorings the highest is as the recommendation results of targeted customer.

Wherein, described exercise question recommendation apparatus also comprises:

Pretreatment module, for debug or invalid data, then sorts by submission time from small to large; The user that statistics relates to and exercise question, form user's set and an exercise question set.

The beneficial effect of technical scheme provided by the invention is:

1, the present invention is that the application of commending system in user learning process provides new approaches, introduces and do topic difficulty to represent its ability level in traditional user characteristics represents, improves the effect of conventional recommendation method on exercise question is recommended to a certain extent.

2, follow the coordination filter method based on user to compare, the present invention can better describe user characteristics, obtains exercise question recommendation results more accurately, user can be helped to select learning content, improve learning efficiency.

Accompanying drawing explanation

Fig. 1 is a kind of process flow diagram of the exercise question recommend method based on typical degree and difficulty;

Fig. 2 be degree-of-difficulty factor introduce ratio affect schematic diagram;

Fig. 3 is the comparison schematic diagram of Distance conformability degree;

Fig. 4 is the comparison schematic diagram of cosine similarity;

Fig. 5 is a kind of structural representation of the exercise question recommendation apparatus based on typical degree and difficulty;

Fig. 6 is another structural representation of a kind of exercise question recommendation apparatus based on typical degree and difficulty.

In accompanying drawing, the list of parts representated by each label is as follows:

1: computing module; Select module at 2: the first;

3: grading module; Select module at 4: the second;

5: pretreatment module.

Embodiment

For making the object, technical solutions and advantages of the present invention clearly, below embodiment of the present invention is described further in detail.For convenience of description, claim in the embodiment of the present invention to want the user of recommended exercise question to be targeted customer, claim to want recommended one exercise question to be target exercise question.

Embodiment 1

Embodiments provide a kind of exercise question recommend method based on typical degree and difficulty, see Fig. 1, this exercise question recommend method comprises the following steps:

101: preprocessed data;

The primary data of embodiment of the present invention process is that user does the submission record inscribed.First the data of debug or invalid (such as: delete critical data item), then sort from small to large by submission time.After completing above-mentioned steps, the user that statistics relates to and exercise question, form user's set and an exercise question set.

102: classify to exercise question set, calculation question object difficulty, to generate problem description;

In embodiments of the present invention, use exercise question type label (as " dynamic programming ", " computational geometry ", " combinatorics " ...) exercise question set is classified.In addition, the difficulty of per pass exercise question will also be calculated.In embodiments of the present invention, the difficulty of exercise question is described to the function of submitted number of times and percent of pass two factors, and these two factor data statisticss all by obtaining step 101 draw.When the difficulty of exercise question and type known after, per pass problem description is become two tuples: < type, difficulty >.

103: calculate each targeted customer and do topic situation, i.e. the quantity of done this type exercise question, difficulty and the percent of pass of this user on this type exercise question on each type exercise question;

Do topic situation according to above-mentioned, calculate the typical degree of this targeted customer on this type exercise question; The typical degree of targeted customer on each type exercise question forms the feature interpretation (i.e. proper vector) of this targeted customer:

＜type ₁:typicality ₁,type ₂:typicality ₂,…,type _i:typicality _i…,type _n:typicality _n＞

Wherein, type _irepresent topic types, typicality _ithat this targeted customer is at type _itypical degree on type exercise question, i is topic types numbering; N is topic types sum.

104: according to targeted customer's proper vector, calculate the similarity between arbitrary target user characteristics vector, and to each targeted customer, select user that some similarities are the highest as its " arest neighbors ";

First the user characteristics vector utilizing step 103 to be formed, calculates the similarity between any two targeted customer's proper vectors, in this, as the similarity between two targeted customers; Then to each targeted customer, select user that some similarities are the highest as its " arest neighbors "." arest neighbors " herein, the arest neighbors namely described in traditional collaborative filtering method, using the Main Basis of marking to target exercise question as target of prediction user.

105: foundation arest neighbors user does topic situation, and target of prediction user is to the scoring not doing exercise question.

In embodiments of the present invention, for target exercise question and targeted customer, calculate the arest neighbors number of users having done this target exercise question, as user, the prediction of this target exercise question is marked after being normalized to [0,1].

106: to targeted customer, the target exercise question selecting some scorings the highest is as the recommendation results to this targeted customer.

This step 106 is according to step 105 and step 103.It is higher that targeted customer does the frequency inscribed, and recommend the quantity of exercise question more to it, user does the frequency inscribed and obtained by step 103.Determine after recommending number, according to the prediction scoring that step 105 produces, select the exercise question of respective numbers best result as recommendation results.

The present invention with the last submission time of user in experimental data for boundary, after acquisition a period of time internal object user do topic situation, doing topic situation by comparing recommendation results with actual, evaluating the accuracy of recommending.

In sum, the embodiment of the present invention has taken into full account this key character of item difficulty by step 101-step 106, improves conventional recommendation method, obtains good effect.

Embodiment 2

Below in conjunction with concrete computing formula, example, technical scheme in embodiment 1 is described in detail, refers to hereafter:

201: preprocessed data;

That is, some illegal data are got rid of, such as: some have lacked the data of critical data item; In addition, statistics forms initial user's set and exercise question set, with the data area of clear and definite embodiment of the present invention process.

202: exercise question set is classified, calculation question object difficulty;

Usually, the exercise question done by more people is simpler, and the topic that percent of pass is higher is simpler, compares both this, and the former is more important, because some exercise question percent of pass is very high, but only has one or two people to submit to, and this exercise question often and remarkable.In addition, variant between different classes of exercise question, such as: the exercise question of computational geometry class, because the size of code of solving a problem is large, detailed problem is many, causes the percent of pass of this type of exercise question entirety not high.Comprehensive these actual conditions above, the degree-of-difficulty factor of exercise question is calculated by formula (1).

${\mathrm{df}}_i=1.0-(\frac{3\&CenterDot;{\mathrm{subCnt}}_i}{4\&CenterDot;{\mathrm{mxSub}}_j}+\frac{{\mathrm{acRate}}_i}{4\&CenterDot;{\mathrm{mxAc}}_j})---\left(1\right)$

Wherein, j is exercise question p _iaffiliated exercise question class number; MxSub _jin j class exercise question, the submission number of times of the exercise question that submitted number of times is maximum; MxAc _jin j class exercise question, the percent of pass of the exercise question that percent of pass is the highest.SubCnt _iand acRate _irepresent exercise question p respectively _isubmitted number of times and percent of pass.The item difficulty coefficient calculated by above formula belongs to [0,1].

203: calculate user characteristics vector;

User characteristics vector and the typical degree of user on every class exercise question.

According to the computing method (Cai of TyCo, Yi, Leung, Ho-fung, LiQ, etal.TyCo:TowardsTypicality-basedCollaborativeFilteringR ecommendation [C] // 201022ndInternationalConferenceonToolswithArtificialInte lligenceIEEEComputerSociety, 2010:97-104), the typical degree of user in a class exercise question respective user group, depend on two factors, one is the average score of user to this type exercise question, and another is the frequent degree that user does this kind of exercise question.User is higher, then interested in this kind of exercise question to the scoring of this type of exercise question; It is more frequent that user does this type of exercise question, then interested in this type of exercise question.Two values are calculated respectively, the mean value that the typical degree of user in this type of exercise question respective user group will be these two values according to above two factors.User u _itypical degree v in jth class exercise question respective user group _{i, j}computing formula as shown in Equation 2.

$v_{i,j}=\frac{1}{2}({\stackrel{\&OverBar;}{R}}_{i,j}+\frac{S_{i,j}}{{\mathrm{\&Sigma;}}_{k=1}^n{S}_{i,k}})---\left(2\right)$

Wherein, user u _ito the average score of done jth class exercise question; S _i,juser u _ido the quantity of j class exercise question; S _i,kfor user u _ido the quantity of k class exercise question; N is topic types sum.

In fact, in this application scenarios, user to per pass exercise question only have by with not by two kinds of results, being 1 by scoring, otherwise is 0; Therefore, user is to the average score R of every class exercise question _i,jall 1, can not the preference of representative of consumer or ability level.

The application scenarios of TyCo mainly predicts that film is marked, and (DF_TyCo) method that the embodiment of the present invention proposes is mainly used in exercise question recommendation.Difficulty is a key character specific to exercise question, user do the difficulty of exercise question, represent the ability level of user to a certain extent.In the embodiment of the present invention, user typical case's degree computing formula as shown in Equation 3.

$v_{i,j}=\mathrm{\&beta;}\frac{S_{i,j}}{{\mathrm{\&Sigma;}}_{k=0}^{n-1}{S}_{i,k}}+(1-\mathrm{\&beta;})\frac{D_{i,j}}{{\mathrm{\&Sigma;}}_{k=0}^{n-1}{D}_{i,k}}---\left(3\right)$

Wherein, n is exercise question classification sum; D _i,jfor user u _ido the item difficulty summation of jth class exercise question; S _i,jfor user u _ido the quantity of jth class exercise question; β is undetermined coefficient, is determined by experiment optimal value; S _i,kfor user u _ido the quantity of kth class exercise question; D _i,kfor user u _ido the item difficulty summation of kth class exercise question.

Formula 2, compared with formula 3, eliminates average score factor, adds item difficulty factor, is that DF_TyCo is relative to the improvement of TyCo in exercise question exemplary application.

204: calculate user's similarity;

This step utilizes the vectorial similarity calculated between any two users of user characteristics.The embodiment of the present invention have employed cosine similarity formula and Distance conformability degree formula respectively, and contrasts its effect by experiment.Two input vectors in calculating formula of similarity are the proper vector of two users respectively, that is: with wherein, v _i,kfor user u _itypical degree on kth class exercise question; v _j,kfor user u _jtypical degree on kth class exercise question.After input vector is determined, the concrete computation process of similarity is conventionally known to one of skill in the art, and the embodiment of the present invention does not repeat this.

205: select arest neighbors;

After having obtained the similarity between user, for each user chooses " arest neighbors " that the some users the most similar to it form this user.The embodiment of the present invention is that each user chooses the similar users of fixed qty as its arest neighbors.Too much or the very few effect that all can affect recommendation of arest neighbors user, select an optimal value by repeatedly contrast experiment in the embodiment of the present invention, the number of the embodiment of the present invention to arest neighbors does not limit.

206: prediction user marks to exercise question;

Targeted customer calculates according to its arest neighbors the scoring of target exercise question.The exercise question that the arest neighbors user of targeted customer did, targeted customer also may do, and more similar user, it does topic situation more has reference value.The embodiment of the present invention is thought, any user it is submitted to and the exercise question passed through scoring be 1, remaining exercise question scoring be 0, then with the similarity between arest neighbors user and targeted customer for weights, do weighted mean, the result obtained is that targeted customer marks to the prediction of target exercise question.

207: select to recommend exercise question.

Have two kinds of methods herein, one is selection threshold value, and the exercise question that targeted customer's scoring is exceeded threshold value is as recommendation results, and another kind is for each targeted customer recommends the exercise question of fixed qty.In the embodiment of the present invention, be adopted as each targeted customer and recommend the exercise question of fixed qty to be that example is described.

By experimental data according to the time, be divided into test set and training set, utilize the data in training set to produce recommendation results, with the Data Comparison in test set, weigh the accuracy of recommending.

In sum, the embodiment of the present invention improves conventional recommendation method by the feature that above-mentioned steps 201-step 207 utilizes exercise question to have " difficulty ", obtains good effect.The embodiment of the present invention can be applied in study website in the future, helps user to select learning content, formulates personalized Learning Scheme.

Embodiment 3

Below in conjunction with concrete example, computing formula, accompanying drawing, feasibility checking is carried out to the technical scheme in embodiment 1,2, described below:

In experiment, the value of the number K of arest neighbors user is allowed to be respectively 5,10 ..., 70, observation experiment result, to probe into the impact of K value on experimental result; Under certain K value, change user and do the weight factor of topic difficulty when describing user characteristics, observation experiment result also calculates its accuracy rate.

The present invention adopts F value (F-measure) and user's accuracy rate (PU) to carry out evaluation experimental result.For convenience, claim " to user u _xrecommend one exercise question p _y" for once to recommend.

If the recommendation produced adds up to recomNum, the number of accurate recommendation is accuate, in test set, all users do topic and add up to realSum, then accuracy pred and recall rate recall computing method are respectively as shown in formula (4), formula (5).

$pred=\frac{accuate}{recomNum}---\left(4\right)$

$recall=\frac{accuate}{realSum}---\left(5\right)$

Had pred and recall just can calculate F value, the computing method of F value are as shown in formula (6).

$F=2\&times;\frac{pred\&times;recall}{pred+recall}---\left(6\right)$

Wherein, pred and recall reflects the quality of result respectively from two aspects, and F value is the two comprehensive embodiment, and in experiment, F value is larger, and result is better.

User's accuracy rate refers to the number percent that can be accounted for all targeted customers by the targeted customer accurately recommended.Fix recommendation 10 road exercise question to each targeted customer, if for certain targeted customer, have at least one exercise question to be recommend accurately, then claim this targeted customer accurately to be recommended.User's accuracy rate computing formula is as shown in formula (7):

$PU=\frac{predUser}{allUser}---\left(7\right)$

Wherein, predUser is by the number of users accurately recommended, and allUser is the total number of users of recommended mistake.

As shown in Figure 2, user characteristics is described as be in the typical degree on each class exercise question.Introduce degree-of-difficulty factor, namely with user do the part of difficulty for user characteristics of exercise question.As can be seen from Figure 2, introducing difficulty can improve experimental result, illustrates in recommendation, and introducing item difficulty, is rational.Experimental result shows, along with weight factor increases to 1.0 gradually from 0, user's accuracy rate presents the trend of falling after rising, especially when weight factor is about 0.85, namely user do topic difficulty account for user characteristics 85%, all the other factors when accounting for 15% effect best, thus prove under this application scenarios, introduce item difficulty, the effect of conventional recommendation can be improved.

TyCo is the collaborative filtering method based on typical degree, is introduced in commending system, can improve the experimental result of conventional recommendation method.In the embodiment of the present invention, on this basis, introduce item difficulty, further improve experimental result when being applied in exercise question recommendation by commending system.

In fact, compare with UBCF method, TyCo introduces the concept of typical degree, improve traditional collaborative filtering method, and the embodiment of the present invention is on the basis of TyCo, introduces item difficulty, improves result again.Fig. 3 and Fig. 4 respectively show under different calculating formula of similarity, the comparison of the embodiment of the present invention and the collaborative filtering method (TyCo) based on typical degree, the collaborative filtering method (UBCF) based on user, can find out from figure, when K gets each value, this method () comparatively can obtain better result by additive method.Prove this method be better than in accuracy other both.

Embodiment 4

Based on an exercise question recommendation apparatus for typical degree and difficulty, see Fig. 5, this exercise question recommendation apparatus comprises:

Computing module 1, doing topic situation for calculating each targeted customer, comprising: the quantity of exercise question, difficulty and the percent of pass of targeted customer on this type exercise question on each type exercise question;

First selects module 2, for according to targeted customer's proper vector, calculates the similarity between arbitrary target user characteristics vector, and to each targeted customer, selects user that some similarities are the highest as arest neighbors;

Grading module 3, for doing topic situation according to arest neighbors user, target of prediction user is to the scoring not doing exercise question;

Second selects module 4, and for targeted customer, the target exercise question selecting some scorings the highest is as the recommendation results of targeted customer.

Wherein, see Fig. 6, this exercise question recommendation apparatus also comprises:

Pretreatment module 5, for debug or invalid data, then sorts by submission time from small to large; The user that statistics relates to and exercise question, form user's set and an exercise question set.

The embodiment of the present invention achieves by above-mentioned module the feature utilizing exercise question to have " difficulty " and improves conventional recommendation method, obtains good effect.The embodiment of the present invention can be applied in study website in the future, helps user to select learning content, formulates personalized Learning Scheme.

The embodiment of the present invention is to the model of each device except doing specified otherwise, and the model of other devices does not limit, as long as can complete the device of above-mentioned functions.

It will be appreciated by those skilled in the art that accompanying drawing is the schematic diagram of a preferred embodiment, the invention described above embodiment sequence number, just to describing, does not represent the quality of embodiment.

The foregoing is only preferred embodiment of the present invention, not in order to limit the present invention, within the spirit and principles in the present invention all, any amendment done, equivalent replacement, improvement etc., all should be included within protection scope of the present invention.

Claims (5)

1., based on an exercise question recommend method for typical degree and difficulty, it is characterized in that, described exercise question recommend method comprises the following steps:

Calculate each targeted customer and do topic situation on each type exercise question, comprising: the quantity of exercise question, difficulty and the percent of pass of targeted customer on this type exercise question;

According to targeted customer's proper vector, calculate the similarity between arbitrary target user characteristics vector, and to each targeted customer, select user that some similarities are the highest as arest neighbors;

Foundation arest neighbors user does topic situation, and target of prediction user is to the scoring not doing exercise question;

To targeted customer, the target exercise question selecting some scorings the highest is as the recommendation results of targeted customer.

2. a kind of exercise question recommend method based on typical degree and difficulty according to claim 1, it is characterized in that, described exercise question recommend method also comprises:

Debug or invalid data, then sort from small to large by submission time;

The user that statistics relates to and exercise question, form user's set and an exercise question set.

3. a kind of exercise question recommend method based on typical degree and difficulty according to claim 1, it is characterized in that, described targeted customer's proper vector is specially:

＜type ₁:typicality ₁,type ₂:typicality ₂,…,type _i:typicality _i…,type _n:typicality _n＞

Wherein, type _irepresent topic types; Typicality _ithat targeted customer is at type _itypical degree on type exercise question; I is topic types numbering; N is topic types sum.

4., based on an exercise question recommendation apparatus for typical degree and difficulty, it is characterized in that, described exercise question recommendation apparatus comprises:

Computing module, doing topic situation for calculating each targeted customer, comprising: the quantity of exercise question, difficulty and the percent of pass of targeted customer on this type exercise question on each type exercise question;

First selects module, for according to targeted customer's proper vector, calculates the similarity between arbitrary target user characteristics vector, and to each targeted customer, selects user that some similarities are the highest as arest neighbors;

Grading module, for doing topic situation according to arest neighbors user, target of prediction user is to the scoring not doing exercise question;

Second selects module, and for targeted customer, the target exercise question selecting some scorings the highest is as the recommendation results of targeted customer.

5. a kind of exercise question recommendation apparatus based on typical degree and difficulty according to claim 4, it is characterized in that, described exercise question recommendation apparatus also comprises:

Pretreatment module, for debug or invalid data, then sorts by submission time from small to large; The user that statistics relates to and exercise question, form user's set and an exercise question set.