TWI639092B - Music recommendation method and computer program product thereof - Google Patents

Abstract

A music recommendation method, which includes using a computer's processor to convert a project user broadcast matrix, and merging with a project label frequency matrix into a project user label matrix, based on a non-negative matrix factorization method to generate a Decomposition matrix for calculating a similarity matrix based on social content optimization, generating an artist tag-driven similarity matrix based on an artist tag frequency matrix; and generating a number of unpurchased items based on a user through the processor Target items, use the similarity matrix optimized based on social content and the similarity matrix driven by the artist tags to determine several related items, and based on these related items, predict a rating of each unpurchased item with a scoring prediction algorithm. Each unpurchased item is ranked according to the rating to generate a ranking list.

Description

Music recommendation method and its computer program product

The invention relates to a music recommendation method, in particular to a music recommendation method based on social information.

The increasing advancement of multimedia technology has made available music data grow at an explosive rate, making it difficult to efficiently and effectively obtain preferred music pieces (ie, items) from a large amount of music data. In order to solve this problem, music recommendation (prediction) content can be appropriately added for the user to choose. The so-called music recommendation can refer to a set of predictive algorithms for learning music preferences from the user's usage log. This is a challenging problem, because the user's preferences are usually expressed by explicit votes and implicit expressions. votes). The former is, for example, the user ’s favorite music rating, while the latter includes favorite music play counts, tags, comments, listening history, and other types of usage information. )Wait.

For example, the conventional Collaborative Filtering (CF) method can be used to predict a user's music preferences. The execution process of the method can be divided into a rating prediction stage and an item selection stage. Although the selection phase of the project is easy to complete, there are still issues such as Rating Diversity, Rating Sparsity, and Lack of Rating Information in the prediction phase of the rating. Taking the diversity of ratings as an example, Table 1 shows the rating matrix of four users for four items. Because the results of different users' ratings of the same item are inconsistent, it often leads to a huge gap between user preferences and predicted results.

In order to solve the above problems, another improved CF prediction method has been developed, which uses implicit user preferences, such as social media hits, and more and more researches use social media messages to confirm users. 'S music preferences, instead of using traditional ratings, also uses user-related information about music clips. It is worth noting that although social media information has proven to be relevant to users ’preferences, the method of predicting users’ relevance to music clips is still not robust enough to be used to determine users ’true preferences for music.

In view of this, it is necessary to provide a music recommendation method to solve the problems of the conventional technology.

The main purpose of the present invention is to provide a music recommendation method and its computer program product, which can integrate label information and rating information as a basis for recommendation, so as to improve the recommendation effect due to insufficient information and improve the computational effectiveness of music recommendation.

To achieve the above purpose, the present invention provides a music recommendation method that can be executed by a computer that recommends music. The method may include the steps of: storing an item user broadcast matrix and an item tag frequency in a memory of the computer Matrix and an artist ’s label frequency matrix; using one of the computer ’s processors to perform calculations, convert the project user broadcast matrix into a project user rating matrix, and merge the project user rating matrix with the project label frequency matrix into a Project user label matrix, the project user label matrix is calculated according to a non-negative matrix factorization method to generate a non-negative decomposition matrix, and the non-negative decomposition matrix is calculated according to a similarity algorithm to generate a similarity based on social content optimization Sex matrix, the artist tag frequency matrix is calculated according to the similarity algorithm to generate an artist tag-driven similarity matrix; the processor responds to a user ’s visit, and generates several items based on the user ’s unpurchased items Target items, using the similarity matrix optimized based on social content and the similarity matrix driven by the artist ’s tags to determine several related items, based on these related items, a rating prediction algorithm is used to predict a rating of each unpurchased item, and Each unpurchased item is ranked according to the rating to generate a ranking list; and the ranking list is output by a monitor of the computer to recommend several music items in the ranking list.

In one embodiment of the present invention, the related items are determined by the top k% related phases of the similarity matrix optimized based on social content and the similarity matrix driven by the artist tag, where k is a positive number.

In an embodiment of the present invention, the processor may convert the elements of the project user broadcast matrix to the elements of the project user rating matrix according to a threshold. The threshold is defined as follows: T = μ-τ * σ, where T represents the threshold, μ represents an average value of the user's play times, σ represents a standard deviation of the user's play times, and τ is a weight parameter.

In one embodiment of the present invention, the range of the threshold is divided into two equivalent sub-ranges, and the two equivalent sub-ranges respectively include a set of rated values, and the number of rated values contained in the set of second rated values may be Different; the weight parameter can be 0.5.

On the other hand, in order to achieve the above purpose, the present invention also provides a computer program product. After the computer program product is loaded and executed by a computer, the computer can perform the aforementioned music recommendation method.

S1‧‧‧Offline preprocessing steps

S2‧‧‧ Online prediction steps

T‧‧‧ Threshold for converting the number of playbacks into ratings

μ‧‧‧Average number of playback times

σ‧‧‧ standard deviation

τ‧‧‧ weight parameter

Fig. 1: Flow chart of the music recommendation method according to an embodiment of the invention.

Fig. 2: A schematic diagram of an example of converting the playback times of the music recommendation method into ratings according to an embodiment of the present invention.

Fig. 3a: An exemplary schematic diagram of the item label frequency matrix of the music recommendation method according to an embodiment of the invention.

Fig. 3b: An exemplary schematic diagram of a project user rating matrix of the music recommendation method according to an embodiment of the invention.

Fig. 3c: An exemplary schematic diagram of the item user tag matrix of the music recommendation method according to an embodiment of the invention.

Fig. 4a: An exemplary schematic diagram of the IF matrix of the music recommendation method according to an embodiment of the invention.

Fig. 4b: An exemplary schematic diagram of the UTF matrix of the music recommendation method according to an embodiment of the invention.

Fig. 5a: An exemplary schematic diagram of the original item label frequency matrix of the music recommendation method according to an embodiment of the invention.

Fig. 5b: An exemplary schematic diagram of a frequency matrix of transformed artist tags in the music recommendation method according to an embodiment of the present invention.

Figure 6: A schematic diagram of the experimental results of the OSCP, ATP, and MMR embodiments of the present invention in terms of RMSE.

Figure 7: The present invention compares the top k% experimental results of the MMR embodiment and the conventional recommendation system in terms of RMSE.

Figure 8: The present invention compares the overall experimental results of the MMR embodiment and the conventional recommendation system in terms of RMSE.

Figure 9: The present invention compares the experimental results of the top N test items on the NDCG ranking list of the MMR embodiment and the conventional recommendation system.

Figure 10: A schematic diagram of the experimental results of the present invention for evaluating the effect of different proportions of training sets on RMSE using data set 1.

Figure 11: The present invention compares the experimental results of the MMR embodiment and the conventional IB method in terms of RMSE with the data set 2.

Fig. 12: The present invention compares the experimental results of the MMR embodiment and the conventional RWR method in terms of NDCG with the data set 2.

Figure 13: A schematic diagram of the experimental results of the present invention comparing the NMFN embodiment and the conventional IB method in terms of RMSE.

In order to make the above and other objects, features, and advantages of the present invention more comprehensible, the preferred embodiments of the present invention will be specifically described below in conjunction with the accompanying drawings, which are described in detail below. Furthermore, the terms of direction mentioned in the present invention, such as up, down, top, bottom, front, back, left, right, inner, outer, side, surrounding, center, horizontal, horizontal, vertical, longitudinal, axial, Radial, most The upper layer or the lowermost layer, etc., are for reference only to the directions of the attached drawings. Therefore, the directional terminology is used to illustrate and understand the present invention, not to limit the present invention.

Referring to FIG. 1, the music recommendation method embodiment of the present invention can be executed by a computer that recommends music. The computer can be an electronic device with data processing functions, such as a cloud computing platform (Cloud Computing Platform), various A computer (Computer) or a portable data processing device (Portable Computing Apparatus), etc. In this embodiment, the computer system may include a processor, a memory, and a display, wherein The storage is, for example, various hard drives, etc., which is electrically connected to the processor, and its architecture is understandable to those with ordinary knowledge in the technical field to which the present invention belongs, and will not be described here again. The storage can be used to store the execution of the method The database required for the process, such as: contains pre-stored data or temporary data, etc .; the processor is, for example, a central processing unit (CPU), etc., which can execute an operating software to implement the execution process of the method, such as : Data format conversion and data feature extraction, etc., but not limited to this; the display is, for example, an LCD liquid crystal screen, etc., which is electrically connected to the processor, and its architecture belongs to the present invention Those with ordinary knowledge in the technical field can understand and will not repeat them here. The display can be used to output the result data processed by the processor, for example, to recommend several music items in the ranking list.

Please refer to FIG. 1 again. The above method embodiment of the present invention may include an offline preprocessing step S1 and an online prediction step S2. Among them, the offline pre-processing step S1 can meet the needs of using ratings to represent user preferences, and can speed up the online prediction process. For the needs of using ratings to represent user preferences, a project user can be broadcast by statistical calculation. The item-user playcount matrix is transformed into an item-user rating matrix for use in accelerating the online prediction step S2. In one embodiment, the offline preprocessing step S1 may first merge the converted project user rating matrix and the project label frequency matrix into a hybrid matrix. Then, the mixed matrix is optimized into a non-negative matrix-based matrix (NMF-based matrix) by performing a conventional non-negative matrix factorization (NMF) algorithm. Finally, the non-negative decomposition-based matrix and an artist-tag frequency matrix can be used. The two similarity matrices can be based on calculating the similarity of items, and a similarity matrix (Optimized based on social content optimization) can be generated. Social-Content-based similarity matrix) and an artist-tag-driven similarity matrix.

Please continue to refer to FIG. 1. In an embodiment, the input of the execution stage of the online prediction step S2 may include an un-purchased item and the similarity matrix optimized based on social content (Optimized Social -Content-based similarity matrix), the artist-tag-driven similarity matrix (Artist-Tag-Driven similarity matrix) and the item user rating matrix (Item-User rating matrix), and the output is a ranking list (ranking list) . The whole process starts with the user's visit (active user ’s visit), for example, the user is using a website or running a computer program to purchase music. Then, each unpurchased item of the user is regarded as a target item. For each target item, the similarity matrix optimized based on social content and the similarity matrix driven by the artist tag can be used to determine The top k% of the most relevant items, where k is a positive number, such as 20, 21.5, 25, 30, 32.5, 35 ... etc. Based on this related item, the rating of each unpurchased item can then be predicted. Finally, all unpurchased items can be sorted according to the predicted ratings, thereby generating the ranking list, which can be used as a reference for related applications, such as: a basis for recommending users to purchase items. The following examples illustrate the implementation of the above embodiments of the present invention, but not limited thereto. The following examples illustrate the implementation of the offline pre-processing phase and the online prediction phase of the above method embodiments, but not limited thereto.

For example, the main tasks in the offline pre-processing stage are: (1) rating conversion and (2) constructing the similarity matrix optimized based on social content and the similarity matrix driven by artist tags, as described in detail later.

In terms of the conversion task of rating, usually for a music recommendation system, the user's preference can be expressed by two types, namely explicit and implicit. However, from a usability point of view, it is more convenient for users to use implicit expressions than explicit expressions, because users do not need to give ratings as explicit expressions in non-rating recommendation systems, but users are in the process of using Relevant information can be left as an implicit expression. As a result, for a recommendation system based on social information, how to describe user preferences using implicit ideology is an important issue. For example, the number of user plays can be regarded as implicit ideology. From this point of view, the embodiments of the present invention propose a formula for expressing the number of playbacks as a rating space to meet the requirement of using ratings to represent user preferences.

Figure 2 shows an example of converting the number of plays into a rating. The whole process includes three In the first sub-step, for each user, the number of plays can be divided into two ranges by a threshold T, which is defined by the following formula (1): T = μ- τ * σ, (1) where T represents the threshold value, μ represents a mean value of the number of playback times, σ represents a standard deviation of the number of playback times, and τ is a weight parameter. The above formula (1) is called traditional votes, where the more times the item is played, the higher the user ’s interest in the item. Specifically, for the user, the average number of her / his play times provides His / her average preference. Therefore, those with a lower standard deviation than the number of play times are considered to indicate negative preferences, while those with a higher standard deviation than the number of play times are considered to indicate positive preferences.

In the second substep, the range below T is further divided into two equivalent subranges, relative to the range number set {1,2}, and the range above T is divided into the range number set {3, 4,5}. In the third sub-step, if a certain number of plays is within a certain range, it can be converted into the quoted range number. In this process, the determination of τ is based on the actual rating data collected by the rating system disclosed by the present invention. In one embodiment, by referring to the actual data distribution, the experimental data of τ can be set to 0.5. Finally, the project user rating matrix can be constructed as one of the input sources for this online prediction stage.

The following example illustrates the construction of the similarity matrix based on social content optimization. In one embodiment, a similarity matrix based on social content optimization can be constructed by integrating the number of plays and tag information, with the intention of using tag information as a rating (Playing times) information supplement. Therefore, the first operation process at this stage is to combine the number of plays and the tag information. During this operation, the tag information can be represented by the item tag frequency matrix, and the play number information can be represented by the item user broadcast matrix. After converting the project user broadcast matrix to the project user rating matrix, the project tag frequency matrix and the project user rating matrix can be merged into a matrix named Item-User-Tag (IUT) matrix In the mixed information matrix, the definition can be as defined below. It should be noted that the element value in the item label frequency matrix and the item user rating matrix are the normalized item label frequency value and the normalized item User rating.

Definition 1:

Suppose there are a certain number of unique users U = {u ₁ , u ₂ , ..., u _{| U |} } in the database to represent {user ₁ , user ₂ , ,,, user _{| U |} } Collection; and there are a certain number of unique items I = {itm ₁ , itm ₂ , ..., itm _i , ..., itm _{| I |} }, used to represent {item ₁ , item ₂ , ..., item _i , ..., items _{| I |} }, i is a numbered value ranging from 1 to a positive integer I; and there are a certain number of unique tags T = {tag ₁ , tag ₂ , ..., tag _j , ..., tag _{| T |},} {tag to indicate _1, tag _2, ..., _J tags, ..., tags _{| T |}} is set, j is a positive integer ranging from 1 to the number of T value. If an item user rating matrix IUR _{I → U} and an item label frequency matrix ITF _{I → U are given} , I → U is used to represent the permutation vector of each item in the I set to all users in the U set. The user tag matrix can be defined as: IUT _{I → UT} [v _{i, m} ] where UT is the set of the combination of U set and T set, and I → UT represents all items in the I set to all elements in the UT set Permutation and combination, IUT _{I → UT} represents the project user label (IUT) matrix, v _{i, m is} used to represent any element in the project user label matrix IUT _{I → UT} (such as v _1,1, etc.), v is the range Normalized values from 0 to 1, | UT | = | U | + | T | and 0 <m ≦ | UT |, for example: Figures 3a, 3b, and 3c reveal an item-tag frequency matrix (item-tag frequency matrix), an item-user rating matrix, and an example of an item-user tag matrix.

After the user tag matrix of the project is generated, the second operation process at this stage can execute the NMF algorithm to approximate the preferred information matrix based on the NMF matrix. In fact, NMF is a conventional matrix factorization method, which is understood by those with ordinary knowledge in the technical field to which it belongs. This method is mainly used for factor analysis of non-negative element matrix. Therefore, this operation process can reduce the data structure Dimension to discover potential factors and make separate predictions more effective. Based on the definition 1, during this operation, the project user label matrix can be approximated as the product of two sub-matrices, as follows: IUT _{I → UT} [ v _{i , m} ] IF _{I → F} [ iv _{i , f} ]. UTF _{UT → F} [ utv _{m, f} ] ^T where IF and UTF represent the two-factor matrix, each item i and user m are modeled by a set of one-factor vector F, 0 <f ≦ | F |, the two The elements of the matrix are all non-negative elements. As mentioned above, v _{i, m is} used to represent any element in the item user label matrix IUT _{I → UT} , and so on, and iv _{i, f is} used to represent the factor matrix IF _{I →} any element in _F , utv _{m, f is} used to represent any element in the factor matrix UTF _{UT → F} , and the remaining parameters are defined as above.

In this process, the objective function can be shown as the following formula (2): The definition of the parameters in the above formula (2) is the same as above, and will not be repeated here. The above formula (2) can be minimized using two iterative update algorithms, which can be understood by those with ordinary knowledge in the technical field to which it belongs, and will not be repeated here. The result can be as follows: Among them, the parameter definition in the above formula is the same as above, and will not be repeated here.

According to the NMF algorithm, the IUT can be decomposed into two factor matrices including IF and UTF. The IF is a matrix based on NMF, which is used in the next operation process. In an embodiment, | F | may be set to 1000. Figures 4a and 4b show examples based on factor matrices illustrated in figures 3a, 3b and 3c.

Based on the NMF matrix, the final operation of this sub-stage is to use the similarity calculation between items to generate the similarity matrix based on social content optimization. The similarity matrix based on social content optimization can be used to reduce online For projected costs, the similarity of the project can be defined in Definition 2 described below.

Definition 2:

According to the unique item | I | defined above and the | F | based on the NMF matrix, the feature vectors of the item x and y can be expressed as {iv _{x, 1} , iv _{x, 2} , ..., iv _{x, | F |} } and {iv _{y, 1} , iv _{y, 2} , ..., iv _{y, | F |} }, therefore, the similarity of optimization based on social content between itm _x and itm _y can be The definition is shown in the following formula (3): Among them, OSCsim (itm _x , itm _y ) represents the value of Optimized Social-Content-based similarity between projects x (ie itm _x ) and y (ie itm _y ), iv _{x, f} It is used to indicate that any element in the factor matrix IF _{I → F} is substituted by the item x (ie, iv _{i, f} , i = x), and iv _{y, f is} used to indicate that the factor matrix IF _{I → F} is substituted by the item y Any element (ie ivi _{, f} , i = y), the remaining parameters are defined as above.

According to definition 2, the similarity between the items can be calculated as elements of the similarity matrix optimized based on social content, and the similarity matrix can be defined as in definition 3.

Definition 3:

According to the unique items | I | of Definition 1 and Definition 2, the similarity matrix based on social content optimization can be defined as follows: OSC _{I → I} [ OSCsim _{x, y} ] where OSCsim _{x , y} represents the above item x , The value of the optimized similarity based on social content between y, which is defined as formula (3) above, represents any element of the similarity matrix OSC _{I → I} optimized based on social content. Table 2 is an example of the similarity matrix optimized based on social content by the above embodiment, where itm ₁ , itm ₂ , itm ₃ , and itm ₄ are different item numbers.

The following example illustrates the construction method of the artist tag-driven similarity matrix. In this example, another value of the item tag information is that it can improve music recommendation as information of artists. It can be seen from observation that most users 'music preferences are quite stable for certain artists, that is, it is effective to mine users' music preferences from preferred artists. Therefore, the artist's label can be used as useful information to suggest the user's preferences. The first operation in this sub-stage is to generate an artist tag frequency matrix, which can be defined as defined in the following definition 4.

Definition 4:

Based on definition 1, it is assumed that there are n unique artists in the database AT = {art ₁ , art ₂ , ..., art _a , ..., art _n }, which is used to represent {artist ₁ , artist ₂ , ..., artist _{| U |} }, a is a numbered value ranging from 1 to a positive integer n, and a single artist art _a contains _a set of executable items itm _z , where art _a = ∪ itm _z , itm _z I. Given the item-tag frequency (ITF) matrix ITF _{I → T} [ itf _{i, j} ], the artist tag frequency matrix can be defined as follows: ATX _{AT → T} [ atf _{a, j} ]

Among them, AT → T means the arrangement and combination of all the artists in the AT set to all the tags in the T set, atf _{a, j} means any element in the artist ’s tag frequency matrix ATX _{AT → T} (such as atf _1,1, etc.) , It means the value of the element itf _i in the label frequency matrix of the item, _{i in j} is substituted by itm _z , and the remaining parameters are defined as above.

Figures 5a and 5b are examples for explaining how to generate the artist label frequency matrix. Figure 5a is a schematic diagram of the list of artists and labels of the original project label frequency matrix. . In the examples shown in Figures 3a, 3b, and 3c, suppose there are two unique artists and four unique tags in the database, where art ₁ = {itm ₁ , itm ₂ } and art ₂ = {itm ₃ , itm ₄ }. According to definition 4, the label feature vector of art ₁ is {(1 + 1) / 2, (0.8 + 1) / 2, (0 + 0.2) / 2, (0.6 + 1) / 2} = {1,0.9, 0.1,0.8}. Similarly, the label feature vector of art ₂ is {1,0.5,0.3,0}.

After the artist tag frequency matrix is generated, the artist tag driving similarity matrix can be derived by calculating the similarity between items, which is defined in Definition 5 as described later.

Definition 5:

Assuming that art _a and art _b each contain itm _x and itm _y , based on the artist tag frequency matrix, the project similarity between itm _x and itm _y using artist tag-driven similarity can be defined as follows (4): Among them, ATsim (itm _x , itm _y ) represents the value of artist-tag-driven similarity (Artist-Tag-Driven similarity) between items x (ie itm _x ) and y (ie itm _y ), ATsim (art _a , art _b ) represents the value of Artist-Tag-Driven similarity between artists _{a a} and art _a , p represents a positive number ranging from 0 to the threshold T, and atf _{a, p} represents the artist ’s label frequency matrix ATX _{AT → T} , any element atf _{a, j} , j is substituted by p, atf _{b, p} represents the artist label frequency matrix ATX _{AT → T} , any element atf _{a, j} , a, j Substitute the values entered in b and p, respectively, and define the remaining parameters as above.

The meaning behind Definition 5 is that the similarity of the project can be expressed by the similarity of the artist, That is, if two users like the same artist, they may have the same music preferences. Therefore, the artist tag-driven similarity matrix can be derived from the similarity between items, which is defined in Definition 6 as described later.

Definition 6:

Based on definitions 1, 4 and 5, assuming that there are n unique artists in the database AT == {art ₁ , art ₂ , ..., art _a , ..., art _n }, the artist tag-driven similarity matrix can be The definition is as follows: ATS _{I → I} [ ATsim _{x, y} ] where ATsim _{x , y} represents the artist tag-driven similarity between the above items x , y, and the artist tag-driven similarity matrix ATS _{I → I} For the element in, the remaining parameters are defined as above. Table 3 is an example of the artist tag-driven similarity matrix based on Figures 5a and 5b and Definition 6, where itm ₁ , itm ₂ , itm ₃ , and itm ₄ are different project numbers.

The following example illustrates the construction method of the fusion similarity matrix. In this example, based on social content optimization and artist tag-driven similarity, these two similarities can be fused as shown in Definition 7.

Definition 7:

Assuming that art _a and art _b each contain itm _x and itm _y , given the relative similarity based on social content and artist tag drive are OSCsim (itm _x , itm _y ) and ATsim (itm _x , itm _y ) , respectively. The fusion similarity (FU) between itm _x and itm _y can be defined as described below: FUsim (itm _x , itm _y ) = OSCsim (itm _x , itm _y ) * ATsim (itm _x , itm _y ) (5) Among them, FUsim (itm _x , itm _y ) represents the value of Fusion Similarity between items x (ie itm _x ) and y (ie itm _y ), and the remaining parameters are defined as above.

Considering Table 2 and Table 3, the example of the fusion similarity matrix can be shown in Table 4, where itm ₁ , itm ₂ , itm ₃ , and itm ₄ are different project numbers.

In addition, the following example illustrates the above online prediction method. In this example, since the item similarity has been generated in the offline pre-processing stage, the main purpose of the online prediction is to use the generated item similarity to rate unpurchased items , The entire process of online prediction will be triggered by an active user visit. Then, the unknown ratings are predicted one by one. After all unknown ratings are predicted, the mechanism will sort the unpurchased items to generate a ranking list. In detail, the forecast first uses the project similarity to determine the top k% of projects that are most relevant to the target project. Secondly, for an unknown rating, it can be calculated by definition 8.

Definition 8:

According to the above definition, given a user project rating matrix UIR _{U → I} [r _{s, i} ], which is the transpose matrix of the aforementioned project user rating matrix, r _{s, i} represents the matrix UIR _{U → I} Any element. Suppose the set of items with the highest correlation for a target item itm _x for a user u _s is RI _s = ∪ itm _c , where itm _x I and itm _c I. In this way, the rating of the unpurchased project itm _x can be defined as described below: among them, S indicates that the user is not available to comment item itm estimate _{value x,} r _{s, c s} U indicates that the user is not available for assessment items itm _x equivalent of the other parameters are as defined above.

It should be noted that according to three similarity matrices, the prediction model can be defined as described below:

Following the above embodiment, assume that a user-to-item matrix is shown in Table 1: If the target item for a user u ₃ (user 3) is itm ₁ (item 1), The target rating It can be expressed as follows:

The following examples illustrate related experiments to verify the effectiveness of the above method embodiments.

In terms of experimental data, it can be collected from a collection of user listening behaviors provided by Last.fm. Last.fm is a popular social music website, providing users with online listening (online listening) and tagging services (tagging services). Through this platform, 30 million users can mark the music they have listened to to describe their musical tastes. Therefore, the data from Last.fm is widely used as experimental data. Our experimental data contains two groups, namely Data Set 1 and Data Set 2 (as shown in Table 5). Data Set 1 includes 912 users, 27303 items and 3937 tags, while Data Set 2 has 319 users. 40992 items and 235921 tags. The main difference is that data set 1 contains the record log in 2010 and data set 2 contains the record log in 2013. In our experiment, for each user, about 20% of the items are randomly selected As test data, and other ratio items are used as training data. Table 5 reveals the detailed description of the experimental data. It should be noted that the dimension IF of the optimization matrix can be reduced to 1000 during this work process. As mentioned above, the ratio conversion, τ is set to 0.5.

The following example illustrates the ranking method. As mentioned earlier, the recommendation process can be broken down into two stages, namely, rating prediction and project selection. The two stages of practice use two types of measures, namely, rating error and precision. In general, the rating error is adopted as the rating measurement factor in the rating prediction phase, which is used to indicate the relationship between the predicted rating and the underlying facts The accuracy is used as the rating index of the selected item of the project, and it is used to indicate the ratio of the test item to the top N results in the ranking list (the highest N suggestions).

However, it is not easy to rank the recommendation system based on Top-N Recommendation (TNR) using conventional accuracy, because the ranking list contains unpurchased items and test (rating) items. For the known accuracy, the test items are regarded as basic facts based on the TNR recommendation system. Therefore, for this type of recommendation system, the prediction of its success lies in one aspect, that is, the obtained item should be one of the test items. That is, items that are not test items will be considered as incorrect results. That is, this measurement method seems inappropriate, because there is no evidence that the unpurchased items are negative for users.

In this example, a simple example can be used to explain this. Suppose there are 6 items {item 1, item 2, item 3, item 4, item 5, item 6} in the database. For a user, indicate The set of test items for the specified item rated by the user is {Item 1, Item 2}. Then, suppose that the ranking list of the recommendation system based on TNR derives {item 3, item 2, item 4, item 1, item 6, item 5}. For a TNR-based recommendation system, the accuracy is defined as (correct value / N), where the correct value represents the number of forward items, and N is the resulting item value. Therefore, if the value of N is 2, the item {item 3, item 2} obtained in this example. Obviously, only item 2 is correct, so the accuracy is 1/2. In general, in this example, the accuracy is 0/1, 1/2, 1/3, 2/4, 2/5 and 2/6, where the values of N are 1, 2, 3, 4, 5 and 6. This example shows that it is not feasible to identify the unpurchased item as a false prediction, because item 3, item 4 and item 5 may be the user's positive items.

Therefore, in this example, a rough measurement called NDCG (Normalized Discounted Cumulative Gain) is used as the rating index. In this rating, all test items of each user are first sorted by the number of normalized play times, where the play times are considered to be related to relevance judgment or scoring. It should be noted that the rated items in the test are regarded as unrated items, so the unrated items are composed of inspection items and unpurchased items. After all unrated items are rated and predicted, a ranking list consisting of test items and unpurchased items is generated to select the top N test items from the ranked list. The NDCG is defined as follows (7): Among them, IDCG (ideal normalized discounted cumulative gain) is a normalization factor, e represents the ranking position, and rel represents the correlation judgment or score using the normalized play times. From the above definition of NDCG, the difference between conventional accuracy and NDCG is that using NDCG rating only considers test items, but using conventional accuracy considers both test items and unpurchased items.

Tables 6 and 7 are examples of how to calculate NDCG. Suppose there are 6 unrated items in the rating process. The rating process includes a test item combination {itm ₁ , itm ₂ , itm ₃ , itm ₄ , itm ₅ } and an unpurchased item {itm ₆ }. In Table 6, the ranking set is {itm ₄ , itm ₆ , itm ₁ , itm ₃ , itm ₅ , itm ₂ }, where the forecast rating set is {4.7,4.2,3.8,3.5,3.3,3.0} and related The correlation score set is {0.9, NULL, 0.4, 0.7, 0.1, 0.2}. In this example, because itm _{6 is} not assigned a relevance score, it cannot be used to consider ratings. Therefore, the DCG without considering the unpurchased project itm ₆ is 1.468. Similarly, the IDCG derived from Table 3 is 1.511, so the NDCG is 1.468 / 1.511 = 0.97.

In addition to NDCG, the other indicators used for evaluation are RMSE (root mean square error), which is defined as: Where r represents the basic fact, Represents the predicted rating value and is used to represent the test of the test data set (test). Overall, NDCG shows the predicted ranking ratio relative to the ideal ranking, and RMSE shows the error variance between the predicted rating and the underlying facts. In other words, the lower the rms error, the lower the error, the higher the accuracy, and the better the recommendation. In contrast, the higher the NDCG, the better the recommendation. In order to ensure sound experimental evaluation results, both NDCG and RMSE are used to verify the above method.

The following examples illustrate the experimental results. In the experiment, the evaluation is mainly based on several concepts: (1) the performance evaluation of the individual models proposed by the present invention including OSCP, ATP, and MMR; (2) the fusion model proposed by the present invention and the existing The effectiveness comparison between RMSE and NDCG among well-known recommendation systems; (3) The effectiveness evaluation of all comparison methods.

The following description evaluates the effectiveness of the individual models proposed by the present invention in Data Set 1. Before evaluating the multimodal recommendation system proposed by the present invention, it must be noted that the effectiveness of the individual models proposed by the present invention. In this evaluation process, the relevant items of the top k percentages mentioned above are used as the basis for prediction. Figure 6 is the experimental results and shows some important points. First, the increase in RMSE increases with the related items. From this, it can be seen that the increase in related items and the increase in noise make the performance decrease. Secondly, the performance of OSCP is better than that of ATP, and MMR is the best. It can be seen that without OSCP or ATP, this fusion mode will not achieve high-quality music recommendation results. In short, OSCP and ATP play a key role in the actual fusion model, and because MMR is the best predictive model, it is used in this example as the main method proposed by the present invention and compared with other well-known recommendation systems.

The following describes the effectiveness comparison between the fusion model proposed by the present invention and the existing well-known recommendation system. In terms of comparison methods, in order to make the experiment solid, we choose 15 better conventional recommendation systems for comparison, including memory-based), content-based and model-based recommendation systems. Table 8 describes the information on the experimental comparison method. Most recommendation systems are widely known, and since some methods to be compared only produce ranking lists, such methods are evaluated by NDCG.

Table 8 Comparison method list

The following describes the experimental analysis of the sparse ratings of data set 1, as shown in Table 9, in this experiment, we completely selected 40866 item ratings as unknown values to predict. However, since the conventional IB method suffers from the problem of sparse ratings, only 34,870 item ratings can be successfully predicted via the IB method. That is, there will be 5,996 item ratings that cannot be predicted by the conventional IB method. In contrast, the MMR disclosed in the present invention has only 66 item ratings that cannot be predicted. Obviously, the rating prediction range of the MMR disclosed by the present invention is much better than that of the conventional IB method, indicating that the MMR method disclosed by the present invention can truly significantly alleviate the problem of sparse ratings. Based on this analysis result, In the comparative experimental method described later, the number of ratings of this test item is set to 34870.

The following describes the content of using RMSE evaluation in data set 1. Due to the sparse ratings and the diversity of ratings may bring unsatisfactory prediction results, it needs to be investigated through detailed experimental evaluation. Therefore, in subsequent experiments, the method of comparison is to use RMSE and NDCG to display and rate the sparseness and the ability to assess the diversity of rating issues. First, this paragraph describes the RMSE comparison results of different methods. It should be noted that the top k% of related project / user predictions are used as the basis for this experiment.

Figure 7 depicts some important experimental results provided by the RMSE. First, the MMR method of the present invention is the best, and the conventional UB method is the worst; second, the conventional user-based method is compared to the conventional based The effectiveness of the project method is worse; third, the results of the conventional project-based method and the conventional similarity fusion (SF) method are quite close; fourth, although the conventional UPTR method is a content-based method, it does not It is not better than the conventional project-based method and the conventional SF method. The possible reason is that the recommendation results of the conventional UPTR method are based on the user's ideas; fifth, the RMSE of the conventional UB method and the conventional UPTR follow As related users increase and decrease. This result shows that the conventional IB and SF methods are very different from the MMR method of the present invention. It can be seen that more users can bring better results to the user-based method, but more projects cannot promote the prediction based on the project method. The value may be that the user ’s behavior in similar listening and marking is very relevant, but the project does not have this feature.

Inherited, this is because the conventional model-based methods such as SVM, DT, Bayes, SVD ++, MF, NMF, and NIMF do not consider the top k percentage rate correlation items when predicting the rating, which is based on the model. Method, memory-based method and the MMR method of the present invention compare the best overall evaluation method of RMSE. It can be seen from Figure 8 that, first, the Bayesian The efficiency of the method is the worst; second, the performance of the conventional SVD ++, IB, MF, NMF and SF methods are quite close, slightly better than the conventional UPTR method; third, the overall knowledge is based on the project-based method and the conventional SF The method is superior to the conventional user-based method and the conventional model-based method; fourth, although the conventional SF method combines the conventional user-based and project-based CF methods, it still cannot bring out the conventional CF-based method Significantly better results. Overall, the experimental results show that the method of the present invention is better than other conventional methods in terms of RMSE, that is, from the perspective of RMSE, the method proposed by the present invention can effectively integrate social content optimization and artist tag-driven Similarity effectively predicts users' rating results.

The following describes the evaluation contents of the data set 1 using NDCG, including the comparison between the MMR of the present invention and the conventional methods of RWR, PureSVD, NIMF, SVD ++, IB, TAGA, and UB. The final evaluation check uses the NDCG method. Figure 9 depicts the experimental results in terms of NDCG, which is explained below. First, the performance of the conventional UB method is the worst, and the performance of the MMR method of the present invention is the best. In this experiment, while considering 10% of related items, the NDCG of the MMR of the present invention can reach about 0.8; second, The results of Xizhi's content-based and model-based methods are very close, which are better than Xizhi's memory-based methods, that is, the use of RMSE results cannot reveal the true strength of the recommendation system in predicting user preferences. Overall, the method proposed by the present invention can bring better results in music recommendation than many comparison methods.

The following illustrates the scalability evaluation content of data set 1 using RMSE, the purpose is to clarify the effectiveness of this experiment in the case of different training set sizes. The experimental results shown in Figure 10 are that the more training data, the higher the effectiveness. However, the effectiveness of the training size between 80% and 90% is quite close, but 70% and 80% are slightly different, which further shows that the method proposed by the present invention can produce stable results at different training scales .

The following describes the evaluation content of using RMSE and NDCG in data set 2. Based on the above evaluation results of data set 1, it can be seen that the music recommendation performance of the method proposed by the present invention in terms of RMSE, NDCG, and sparseness is far superior to all current methods. . In order to make the evaluation results more testable, the data set 2 was further used for evaluation. As shown in Table 5, the two data sets have different listening and marking behaviors. Therefore, our purpose is to investigate whether the effectiveness of listening varies and the marking behavior. Since the best comparison methods using Dataset 1 are the conventional IB and RWR methods for RMSE and NDCG, respectively, we still choose the conventional knowledge IB and RWR do For the comparison method using dataset 2. Regardless of RMSE or NDCG, Figures 11 and 12 show that the MMR method proposed by the present invention, even using Dataset 2, is still superior to the conventional IB and RWR methods, in particular, even in the case of changes in listening and marking behavior The method of the present invention can still truly achieve high-quality music recommendation performance.

The following describes the results of the efficiency evaluation. In addition to the effectiveness evaluation, the efficiency of each user is another problem in practical applications. In order to solve this problem, a comparative experiment was conducted to show the time predicted by a method of comparison for a user. Table 10 shows the experimental results. Overall, almost all of these methods can achieve user predictions within 1 second, and only the time required for the conventional SVM and RWR methods is greater than 1 second. From a usage point of view, it is feasible if the result can be returned in about a second. Therefore, only the conventional RWR method cannot meet the time requirement. In addition, although the conventional RWR method can truly achieve high benefits, the execution time is so high that users cannot accept it.

The following describes the relevant discussion of the above experimental results: (1) In the experiment, OSCP comprehensive rating and label information, and the conventional IB method uses rating information. Using Figures 5 to 7 to compare the conventional OSCP method with the IB method, the conventional OSCP method performs better than the conventional IB method. This result shows that the joint information is more robust than the rating information in dealing with the problems of rating diversity and sparse rating, which also shows that the tag information does help the recommendation system to predict user preferences. (2) As can be seen from Figures 6 to 8, the method proposed by the present invention is the best method in terms of RMSE and NDCG. This is an important finding. From these results, it is not easy to perform well in both RMSE and NDCG, for example: the conventional IB method is good in RMSE, but poor in NDCG; According to observation, although the prediction of the conventional IB method can be close to the actual truth of the rating, the relevant ranking cannot be close to the truth; in addition, both the RMSE and NDCG of the conventional IB and SVD ++ methods perform well and stable. Therefore, not a good RMSE method is a poor NDCG method. On the contrary, the method that performs well for both RMSE and NDCG is really reliable for music recommendation. (3) In fact, the above comparison method is a fusion-based method that uses several information sources and combines different CFs to form a fusion mechanism, such as MMR, SF, NIMF, SVD ++, TagA, RWR, UPTR, etc. However, experimental results show that not all fusion-based methods are beneficial to both RMSE and NDCG. On average, the method proposed by the present invention can bring stable results in RMSE and NDCG. (4) In addition to the above experimental analysis, it is necessary to use an optimization method based on NMF. For this reason, the present invention has completed another method, namely, NMF-based neighborhood prediction (NMFN) using an optimized rating matrix (user item rating matrix). Figure 13 shows the relevant experimental results. The NMFN method using the optimized rating matrix is superior to the IB method with no optimization of the rating matrix. The experimental results show that the optimization operation can really promote the rating prediction. In addition, although NMFN uses an optimized rating matrix, its performance is still not as good as the MMR method of integrating and optimizing information in the present invention. It can be seen that the MMR method of the present invention is indeed used for music compared to other methods. Recommend a truly reliable method.

In addition, according to other embodiments of the present invention, the above-mentioned music recommendation method embodiment can be implemented as a computer program product, which can be written in a programming language (such as C or Java, etc.), and the computer program product can be stored in a computer. The read storage device, such as an optical disc, etc., when a computer loads the computer program product in the storage device, the music recommendation method embodiment can be executed according to the computer program product.

Although the present invention has been disclosed in preferred embodiments, it is not intended to limit the present invention. Anyone who is familiar with this skill can make various changes and modifications without departing from the spirit and scope of the present invention, so the present invention The scope of protection shall be defined as the scope of the attached patent application Whichever prevails.

Claims (5)

A music recommendation method for a computer that recommends music. The method includes the steps of: storing an item user broadcast matrix, an item label frequency matrix, and an artist label frequency matrix in a memory of the computer; using the One of the computer's processors performs operations, converts the project user broadcast matrix into a project user rating matrix, merges the project user rating matrix and the project label frequency matrix into a project user label matrix, and then projects the user label The matrix is calculated based on a non-negative matrix factorization method to generate a non-negative decomposition matrix, which is calculated based on a similarity algorithm to generate a similarity matrix based on social content optimization, and the artist label frequency matrix is based on The similarity algorithm is calculated to generate an artist tag-driven similarity matrix; the processor responds to a query from a user, and generates several target items based on the user ’s unpurchased items. The social content-based optimization The similarity matrix and the artist-driven similarity matrix determine several correlations Project, based on the related items, a rating prediction algorithm is used to predict a rating of each unpurchased item, and each unpurchased item is sorted according to the rating to generate a ranking list; and the ranking list is output by a monitor of the computer To recommend several music items in the ranking list; where the processor converts the elements of the project user ’s broadcast matrix to the elements of the project user ’s rating matrix according to a threshold, the threshold is defined as follows: T = μ-τ * σ, where T represents the threshold, μ represents an average value of the user's play times, σ represents a standard deviation of the user's play times, and τ is a weight parameter. The music recommendation method as described in item 1 of the patent application scope, wherein the related items are determined by the top k% related items of the similarity matrix optimized based on social content and the similarity matrix driven by the artist ’s tags, where k is a positive number. The music recommendation method as described in item 1 of the patent application scope, wherein the threshold range is divided into two equivalent sub-ranges, and the two equivalent sub-ranges respectively include a rating value set, each The number of rating values is different. The music recommendation method as described in item 1 of the patent application scope, wherein the weight parameter is 0.5. A computer program product. After the program is loaded and executed through a computer, the computer can perform the music recommendation method as described in any one of items 1 to 4 of the patent application.