TWI504273B - Multimedia content recommendation system and method - Google Patents

Description

Multimedia content recommendation system and method

The present proposal relates to a multimedia content recommendation system and method, and more particularly to a multimedia content recommendation system and method for recommending a group including a plurality of users.

Today's multimedia technology allows Internet service providers to offer a variety of services, such as users can watch movies or TV shows on the Internet. In these applications, various multimedia content can be simultaneously transmitted to a large number of users. Take Internet Protocol Television as an example, which is a type of Broadband TV. IPTV is a system for transmitting television information using a broadband network as a medium, and transmits a digital television service to a user through an internet protocol on a broadband.

However, due to the proliferation of multimedia content provided, it is difficult for users to find something that suits their preferences in a large amount of multimedia content. Therefore, some Internet TV systems provide built-in referral services to predict user preferences and to pick out multimedia content that users may want to watch. However, current recommendation systems generally only recommend for a single user, such as a content-based or Collaborative-Filtering recommendation system.

Content-oriented is based on the user's past preferences, looking for multimedia content with common points of content. For example, other movies directed by the same director can be recommended for the director that the user likes. However, this method has the problem that the direction of the recommended results remains the same. Concord filtering is for each multimedia content, based on the user's preference for different types of content and the similarity of the multimedia content to these types. However, for new users, due to insufficient existing usage records, it is difficult to achieve effective prediction and recommendation, and there is a cold start problem.

Furthermore, when a group with multiple users views multimedia content together, today's recommendation system cannot effectively integrate multiple user preferences, and then recommend for these users. Existing group recommendation techniques can only be recommended by combining single recommendation results (Merging recommendations) or by combining Merging user profiles. The combined single recommendation result is first recommended for each single user using a single recommended algorithm, and then the single recommendation result is taken as the group recommendation result. The combined personal preference is to average the preferences of all users in the group, and then based on the average preference and the algorithm recommended by the single person. However, these combined methods have the recommendation that the results do not meet the needs of the group members, and cannot satisfy the preferences of multiple users at the same time.

The present invention relates to a multimedia content recommendation system and method for providing at least one recommended content from a plurality of multimedia content of a plurality of content types for a group including a plurality of users.

According to an embodiment, the multimedia content recommendation method may include: reading each user's rating of one of the content types; according to the individual ratings, calculating a member similarity of the group, the member similarity represents the user of the group The degree of similarity between individual ratings; and based on the similarity of the members and a similar threshold, a first group recommendation process or a second group recommendation process is performed to base these individual ratings and The machine learning algorithm or a group viewing process selects at least one multimedia content of at least one content type as at least one recommended content.

According to an embodiment, the multimedia content recommendation system may include: a person evaluation database, a usage history database, a member similarity calculation module, and a recommendation module. The personal rating database is used to store each user's personal rating of these content types; the usage history database is used to store the usage history of the group. The member similarity calculation module is configured to calculate the similarity of the members of the group according to the individual ratings, and the similarity of the members represents the degree of similarity between the individual ratings of the users of the group. The recommendation module includes a first recommendation module and a second recommendation module. The recommendation module starts the first group recommendation module or the second group recommendation module according to the comparison result of the member similarity and the similar threshold value, according to the individual evaluation, and according to the machine learning algorithm or the group viewing. In the process, at least one multimedia content of at least one content type is selected as the at least one recommended content.

The detailed features and advantages of the present invention are described in detail below in the embodiments, which are sufficient to enable any skilled artisan to understand the technical contents of the present invention and to implement the present invention, and to disclose the contents, the scope of the patent, and the drawings according to the present specification. Anyone familiar with the relevant art can easily understand the purpose and advantages of this proposal.

The present invention provides a multimedia content recommendation system and method for a multi-user to provide at least one recommended content from a plurality of multimedia content of a plurality of content types for a group including a plurality of users. . The multimedia content recommendation system and method analyzes the similarity between the group organization users and can provide appropriate recommendation content according to the similarity. The system and method can be applied to a system for providing multimedia content through a server, such as an Internet Protocol Television (IPTV), a network television system, or a network broadcast system, and can also be applied to a local host computer having multimedia content ( Local host).

Please refer to FIG. 1 and FIG. 2, which are respectively a block diagram of a multimedia recommendation system of an embodiment and a flowchart of a multimedia content recommendation method.

The multimedia content recommendation system may include a user database 11, a multimedia content database 12, a person rating database 13, a usage history database 14, a member similarity calculation module 20, and a recommendation module. 40.

The user database 11 is used to store various materials of the user who uses the system to listen to or view the multimedia content, such as the user's identification code, facial image, gender, age, or time zone in which the system is often used. The multimedia content recommendation system may also include a non-active human interface or an active human interface to identify the current system user or for the user to interact with the system.

The non-active human interface can be, for example, a camera. The multimedia content recommendation system can automatically capture the image in front of the camera, capture all the face parts in the image, and compare the face in the image with the face image stored in the user database 11 to know the group. member. In other words, the user can be identified with the inactive human-machine interface and the user database 11, and the group and group members who are currently using the system can be further analyzed.

The active human-machine interface can be, for example, a touch panel or a remote controller, and the user can log in to the system through the active human-machine interface, input user data, or rate the favorite multimedia content that has been listened to or viewed. The scores entered by the single user for the multimedia content of the various content types can be stored in the personal rating database 13 as a user rating one of these content types. In addition to the personal rating entered by the user, for the user or the content type that has not yet entered the personal rating, a preset value may be filled in as the initial value. For example, the personal rating can be a positive number between 0 and 1, and the preset value can be set to 0.5.

The multimedia content recommendation system and method can be recommended for groups of 3 to 6 members, for example, can be implemented in a television set in a general household. And the system and method can record all possible groups and group members in the user database 11. For example, a family of four people with parents, mothers, children, and women in the family can record parent, child, parent, parent, father, mother, parent, father, father, mother, and mother. , mother and daughter and children a total of 11 groups, and their respective member code.

The multimedia content recommendation system and method can also be recommended for groups of non-specific objects, such as a video wall disposed outside a building or an advertising television placed in a store. Multiple template members and template groups can be created in the user member database. For example, a restaurant group with a customer group for a young female group can create a plurality of template members and model groups in the user database 11 in advance, and store these template members in the personal evaluation database 13 according to the preferences of young women. Personal rating. In this way, suitable multimedia content can be recommended for unspecified objects of a particular ethnic group.

The multimedia content library 12 can store multimedia content belonging to different content types. For example, the multimedia content may be a video file, an audio file, an image file, or a web address of a streaming media. The content type can be, for example, sports, drama, news, cartoons, movies, music or finance; and each content type can be subdivided into smaller content types, for example, movies can be divided into action, comedy, love, horror, plot Or content type such as family.

The usage history database 14 is used to store a usage history of the group, and uses the viewing behavior of the history record group. For example, in the usage history, all multimedia content that is viewed (or listened to) when the recording system is used can be recorded as the multimedia that has been viewed, and a viewing time of each multimedia that has been viewed and a group that views the multimedia content are recorded.

The member similarity calculation module 20 may first read each user's personal evaluation of the content type from the personal evaluation database 13 (step S200), and then calculate the member similarity of the group according to the individual evaluation (step S300). The member similarity represents the degree of similarity between the individual ratings of these users of the group. The recommendation module 40 can include a first recommendation module 60 and a second recommendation module 80. After calculating the member similarity, the recommendation module 40 can determine whether the member similarity is greater than a similar threshold (step S400). When the member similarity is less than or equal to the similar threshold, the first recommendation module 60 executes a first group recommendation procedure (step S510); when the member similarity is greater than the similar threshold, the second recommendation module 80 performs A second group recommendation procedure (step S520). In other words, the multimedia content recommendation method can find the appropriate recommended content by different recommendation programs according to the similarity degree of the group members.

Please refer to "FIG. 3", which is a flowchart of S300 of an embodiment. First, the member similarity calculation module 20 takes the individual rating as a plurality of feature vectors (step S310), and individually calculates a cosine function value (cosα) of a vector angle α between the two of the feature vectors (step S320). If the vector angle α between two feature vectors is smaller, the difference between the two feature vectors is smaller; and the smaller the vector angle α is, the larger the cosine function value of the vector angle α is. Therefore, the average value of the cosine function values can be taken as the member similarity (step S330).

Taking a three-family family with a parent, a mother, and a child as an example, assume that the three individual users have different feature vectors for sports, drama, news, cartoons, movies, music, and financial content types (0.6, respectively). 0.5, 0.8, 0.2, 0.4, 0.6, 0.9), (0.7, 0.3, 0.6, 0.3, 0.5, 0.5, 0.8) and (0.8, 0.3, 0.3, 0.5, 0.5, 0.5, 0.2). Substituting these feature vectors into the formula cosα= After that, the cosine function value corresponding to the parent group is 0.977, the cosine function value corresponding to the parent subgroup is 0.877, and the cosine function value corresponding to the parent subgroup is 0.804. The average of these cosine function values can then be calculated to be 0.886, and 0.886 is taken as the member similarity. According to different implementation examples, the maximum cosine function value or the smallest cosine function value may also be used as the member similarity.

According to another embodiment, the vector angle θ can also be calculated as a member dissimilarity (or group complexity). The recommendation module 40 can determine whether the member dissimilarity is greater than a different threshold. When the membership disparity is greater than the dissimilar threshold, the first group recommendation procedure is executed; when the membership disparity is less than or equal to the disparity threshold, the second group recommendation procedure is performed.

The first recommendation module 60 (the first group recommendation program) uses a machine learning algorithm for recommendation, wherein the machine learning algorithm can be, for example, an artificial neural network (ANN). Artificial neural networks) or genetic algorithms. The following description uses a neural network as an example. Please refer to "4th", "5th" and "6th", which are respectively a block diagram of the multimedia content recommendation system of an embodiment, a block diagram of the first recommendation module, and a first group recommendation. Flow chart of the program.

The first recommendation module 60 may include a type of neural network prediction module 62; and the multimedia content recommendation system may further include a group member weight database 15 for storing a group member weight of each user of the group. .

The first recommendation module 60 may first read the group member weight of each user of the group by the group member weight database 15 (step S610); calculate according to the neural network, the individual rating, and the group member weight. A first predicted rating of the group for the content type is obtained (step S620). According to the first prediction rating, the first recommendation module 60 may select at least one multimedia content of the at least one content type as the recommended content (step S630).

Please refer to "Figure 7" and "Figure 8", which are schematic diagrams of a neural network such as an embodiment. The neural network 90 can generally include an input layer 91, a hidden layer 93, and an output layer 95; wherein the input layer 91 can include a plurality of input layer nodes 92, and the hidden layer 93 can include a plurality of hidden layer nodes 94, an output layer 95 can include an output layer node 96. These nodes may also be referred to as neurons in the neural network 90. The neural network 90 will primarily receive two sets of inputs, one being the input values P _u1 to P _{un of the} input layer node 92 and the other being the initial neuron weights. The initial neuron weights can include the neuron weights between the input layer node 92 and the hidden layer node 94, and the neuron weights between the hidden layer node 94 and the output layer node 96. The θ of the neural network 90 is the neuron bias value.

Taking the neural network 90 of Fig. 8 as an example, the product of W ₁ and P _u1 is one of the external inputs of this node (neuron). And for this neuron, its complete external input is equal to all input values multiplied by the corresponding neuron weights, which can be expressed as . After adding the external input value to its own bias value, you can get it. Then, after the activation function is converted, a new nerve pulse is generated and transmitted to the next neuron in the neural network 90. Wherein the activation parameter can be, for example, O ₁ = . However, different bias value calculation methods and different activation parameters may also be used in the neural network 90, and are not limited to the above examples.

The neural network prediction module 62 of the first recommendation module 60 will use the architecture of the neural network 90 described above for multimedia content recommendation. The neural network prediction module 62 sets the individual ratings of the group members to a plurality of input values of the neural network 90; and sets the group member weights to a plurality of initial neuron weights of the neural network 90 to As the neuron weight between the input layer node 92 and the hidden layer node 94, and the neuron weight between the hidden layer node 94 and the output layer node 96. In other words, among the neural network 90, the individual rating is taken as the input value of the input layer node 92, wherein the number of neurons in the input layer 91 is the same as the number of users in the group. As the number of users in the group changes, the number of input layer nodes 92 and hidden layer nodes 94 also changes. Each group member weight represents the influence of the corresponding user in the group, and these initial weight values affect the operation of the entire class-like neural network 90.

Based on the neural network 90, these input values, and these initial neuron weights, the neural network prediction module 62 can calculate the first prediction rating. The first predicted rating represents the degree to which the predicted group has a preference for each content type. The first predicted rating may include a plurality of first predicted rating values, and the first predicted rating values individually correspond to the content type.

For a neural network 90 that includes a group of three users, refer to "Picture 9". The neural network prediction module 62 can calculate the first prediction rating using the following formula: R _G ( i )= ( i ) ‧ W _jn + θ _j ) Wo _jm + θ _m ).

Where i is the identification code of the content type, which is a positive integer; j is the number of hidden layer nodes 94, which is 2 in the "Example 9"embodiment; m is the number of output layer nodes 96, in "9th figure In the example, it is 1; n is the number of input layer nodes 92 (equivalent to the number of group members), which is 3 in the example of "Fig. 9"; R _G (i) is the group for the i The first predicted rating value of the content type; P _un (i) is the personal evaluation (input value) of the nth user for the i-th content type; W _jn is the connection input layer node 92a, 92b, 92c and the hidden layer The neuron weight between nodes 94a, 94b, 0 ≦ W _jn ≦ 1; Wo _jm is the neuron weight between the hidden layer nodes 94a, 94b and the output layer node 96 0 ≦ Wo _j ≦ 1; θ _j is hidden The neuron bias values of the layer nodes 94a and 94b; θ _m are the neuron bias values of the output layer node 96. Further, the simulation data and the neural network 90 shown in "Fig. 9" can be used to obtain data such as Table 2 described later.

For a neural network 90 that includes a group of four users, please refer to "Picture 10". The neural network prediction module 62 can also calculate the first prediction rating using the above formula. However, in the embodiment of Fig. 10, j is 3; m is 1; n is 4; W _jn is the nerve connecting the input layer nodes 92a, 92b, 92c, 92d and the hidden layer nodes 94a, 94b, 94c. Yuan weight; Wo _jm is the weight of the neurons connecting the hidden layer nodes 94a, 94b, 94c and the output layer node 96; θj is the neuron bias value of the hidden layer nodes 94a, 94b and 94c; θ _m is the output layer node 96 neuron bias value.

In addition, although there are only one hidden layer 93 in the implementation examples of "8th", "9th" and "10th", the number of hidden layers 93 can be increased to make the prediction result more precise or reduce. Several items of layer 93 are hidden to speed up the operation.

The first speculative module 60 may sort the first predicted grading values in the first predicted grading from large to small to obtain the content type corresponding to the largest first predicted grading value, and then find the content type from the multimedia content database 12 At least one multimedia content belonging to this content type is taken as the recommended content. The first speculation module 60 can also take out a plurality of content types corresponding to the largest first plurality of first prediction ratings, and then use the multimedia content belonging to the content types as the recommended content.

According to an embodiment, the first recommendation module 60 may further include a weight training module 64 to train and obtain group member weights suitable for such neural network 90. Please refer to "Figure 11" for a flow chart of weight training for an example.

The multimedia content recommendation system can further include a group rating database 16 for storing a group rating of the content types. Both the group rating and the individual rating can be entered by the user through the human interface, but the group rating represents the overall group's preference for multimedia content. For example, after the group consisting of two fathers and sons has finished watching the news, the news can be entered through the human-machine interface as the group rating of the content type of the "news". However, for groups or content types that have not yet entered a group rating, you can first fill in the default value as the initial value. For example, the group rating can be a positive number between 0 and 1, and the preset value can be set to 0.5.

The weight training module 64 may read the group-to-group rating of the content type after providing the recommended content to the group (step S660). The group member weights can then be trained using a learning algorithm, a personal rating, and a group rating (step S660). The learning algorithm can be, for example, a back-propagation algorithm or a steepest descent method. Taking the reverse transfer algorithm as an example, the initial weight training module 64 may first generate each of the partial weights θ and the neuron weights W in the neural network 90 by random numbers, and calculate the output values of the output layer nodes 96 to The error between a target value. If the error is greater than a stop condition, the neuron weight is corrected according to the error, and the modified neuron weight W is transmitted back to the upper layer of the neuron, and the group member weight is corrected, and repeated until the error is less than a stop condition.

In addition, the first recommendation module 60 may further update the group member weights stored in the current group member weight database 15 with the trained group member weights. In other words, the first right group recommendation procedure can flexibly adjust the decision-making importance of the group members in the group according to the user's reaction. It should be noted that the weight of the same user in different groups may vary. For example, the weight of the group members corresponding to the father in the group of parents and the group of father and son may be different.

In summary, when the group similarity is less than or equal to the similar threshold, the first recommendation module 60 may use the neural network 90 to obtain the first rating according to the individual rating and the group member weight, and then according to the first A rating provides recommendations. The first recommendation module 60 can train and update the group member weights according to the user's feedback, so that subsequent predictions are more accurate.

When the member similarity is greater than the similar threshold, the second recommendation module 80 executes the second group recommendation procedure. Since the group members have similar preferences, they can be performed in a manner similar to personal recommendation, or the personal rating database 13 can be updated according to the feedback of the group on the recommendation results. The second group recommendation program can use the personal rating and the usage history to make recommendations. Please refer to "FIG. 4" and "Twelfth Diagram" and "FIG. 13", which are respectively a block diagram of a second recommendation module of an embodiment and a flow chart of the first group recommendation procedure.

The second recommendation module 80 can include a person rating merge module 82, a content-based prediction module 84, a Collaborative-Filtering prediction module 86, and a prediction merge module 88. The content-oriented prediction module 84 calculates a second prediction rating for the content type by using the content-oriented algorithm, and uses the synergistic filtering algorithm to calculate a third prediction rating for the content type. S810).

In more detail, the personal rating merge module 82 first averages all individual ratings of the users in the group to obtain a combined rating. The content-oriented prediction module 84 then uses the content-oriented algorithm, the combined rating, and the existing multimedia content to calculate a second prediction rating for the content type. The synergistic filtering prediction module 86 may first select a user from the group as a group representative, and obtain the individual representative of the group representative from the personal rating database 13, and then use the synergistic filtering algorithm. The personal rating of the group representative and the existing multimedia content calculations obtain a third prediction rating for the content type.

The second predicted rating and the third predicted rating may each include a plurality of second predicted rating values and a third predicted rating value, and the second predicted rating values and the third predicted rating values are individually corresponding to the content type.

Next, the prediction merge module 88 reads the usage history of the group (step S820), wherein the usage history records the group's viewing behavior. The usage history may include a plurality of viewed multimedia viewed by the group during a past unit period and a viewing time of each of the viewed multimedia. When the multimedia content recommendation system is used, the recorded multimedia viewed by the user, the viewing time of the viewed multimedia, and the group viewing the multimedia content can be continuously recorded.

The prediction merge module 88 determines a usage date obtained according to the viewing behavior of the group in the past unit period (step S830), and sets the proportion of the usage days in the unit period as a proportional parameter (step S840). ). The number of days used is the number of days that the group uses the multimedia content recommendation system or method during the most recent unit period in the past. The predictive merge module 88 may calculate the usage days and the proportional parameters for the plurality of groups having different members, and then combine the second predicted rating with the third predicted rating.

Please refer to "Fig. 14" which is a schematic diagram of the unit period of an embodiment. The multimedia content recommendation system can retrieve a history of a unit period 120 from the usage history 110. The unit period 120 can be, for example, 7 days, 3 days, or 1 day. The second recommendation module 80 can retrieve the history of the unit period 120 from the current backtracking from the usage history 110. In other words, the predictive merge module 88 may first count the total number of days in which all the viewed multimedia is viewed, and then set the proportion of the total number of days in which the viewed multimedia is viewed in the unit period 120 as a proportional parameter. For example, the unit period 120 is 7 days, and in the past 7 days, if there is a 5-day group using the multimedia push content recommendation system or method, the usage day number 130 is 5, and the proportional parameter is .

According to another embodiment, the predictive merge module 88 may also count, for each content type, the total number of days in which all the viewed multimedia of the targeted content type is viewed, and then use the total number of days in the unit period 120. The ratio is set to the proportional parameter. For example, in the past 7 days, when the group watched the news for 5 days and the movie for 2 days, the scale parameter for the news was , the scale parameter for the movie is . In this way, the proportional parameter corresponding to each content type can be obtained.

After calculating the proportional parameter, the prediction combining module 88 further obtains a fourth prediction rating according to the proportional parameter weighted average second prediction rating and the third prediction rating (step S850). The fourth predicted rating may include a plurality of fourth predicted rating values, and the fourth predicted rating values are individually corresponding to the content type.

In step S850, the second recommendation module 80 can calculate the fourth prediction rating according to the following formula: R _G ( i )= β × ( i )+(1- β )× ( i ).

Where i is an identification code of the content type, which is a positive integer; R _G (i) is a fourth prediction rating for the i-th content type in the fourth prediction rating; β is a proportional parameter; ( i ) is a second predicted rating for the i-th content type in the second predicted rating; ( i ) is the third predicted rating for the i-th content type in the third predicted rating.

For example, when the second prediction rating is (0.3, 0.2, 0.6, 0.1, 0.7, 0.4, 0.6), the third prediction rating is (0.9, 0.1, 0.6, 0.1, 0.6, 0.1, 0.5), and the scale parameter is 5/7. Then, the fourth prediction rating can be calculated as (0.729, 0.171, 0.6, 0.1, 0.871, 0.314, 0.571).

The second recommendation module 80 then selects at least one multimedia content of the at least one content type as the recommended content according to the fourth predicted rating (step S860). Similar to the first recommendation module 60, the second recommendation module 80 may sort the fourth predicted rating values from large to small to obtain the content type corresponding to the largest fourth predicted rating value, and then from the multimedia content database. At least one multimedia content belonging to the content type is found in 12 as the recommended content. The second recommendation module 80 can also take out a plurality of content types corresponding to the largest previous four prediction ratios, and then use the multimedia content belonging to the content types as the recommended content.

The synergistic filtering prediction module 86 searches for the third prediction rating obtained by the multimedia content of the common point according to the user's recent preference; therefore, the accuracy of the third prediction rating increases as the number of usage days 130 increases. And more close to the preferences of the group of users. Based on this consideration, the prediction merge module 88 gives a third prediction and comparison of large proportional parameters when the number of usage days is high, so that the fourth prediction ratio obtained by the weighted average is more accurate.

According to an embodiment, the second recommendation module 80 may further include a person rating update module 89. Please refer to "figure 15", which is a flow chart of the second group recommendation procedure of an embodiment.

The personal rating update module 89 first sets at least one viewed multimedia whose viewing time is greater than a first viewing time threshold to at least one actual viewing multimedia (step S870), and at least the viewing time is greater than a second viewing time threshold. An actual viewing multimedia is set to at least one favorite multimedia (step S880), wherein the second viewing time threshold is greater than the first viewing time threshold. In order to find multimedia content that you want to watch or listen to, groups of users may switch between different multimedia content in a short time. The personal rating update module 89 can set a first viewing time threshold in order to find multimedia content that the group actually has to watch or listen to. Similarly, the personal rating update module 89 can set a second viewing time threshold in order to find a favorite multimedia that the group continues to view.

Please refer to "16A" to "16E", which are schematic diagrams of the proportional parameters of an embodiment. It can be seen from "Fig. 16A" to "16E" that, in general, the longer the viewing time, the higher the individual rating of the multimedia content. And the personal rating update module 89 can set different first viewing time thresholds and second viewing time thresholds for different content types. For example, the length of a cartoon is usually 30 minutes, and the length of a movie is generally greater than 90 minutes. Therefore, the first viewing time threshold and the second viewing time threshold for the movie should obviously be greater than the first viewing time threshold and the second viewing time threshold for the cartoon.

In addition, the personal rating update module 89 can additionally set a threshold of viewing times for different content types. If the number of actually viewed multimedia is greater than the threshold of the number of viewing times, it means that this is the multimedia content that the group often watches or listens to.

For each content type, the individual rating update module 89 can update the current one according to the number of days of use, the number of days in which the multimedia is actually viewed, the number of multimedia actually viewed, the number of favorite multimedia, and the current personal rating. Personal rating (step S890). In fact, the personal rating update module 89 can sequentially view the multimedia content of each content type according to the number of days of use, the number of days of the actual viewing multimedia corresponding to the content type, and the actual viewing multimedia corresponding to the content type. Calculate the new individual rating by the number, the number of favorite multimedia corresponding to the content type, and the individual rating corresponding to the content type; then update the current personal rating with the new personal rating (that is, the original original before the update) Personal rating). Since the members of the group are highly similar, the individual ratings of all group members can be updated with a new personal rating.

According to an embodiment, the personal rating update module 89 may first calculate the number of corresponding actual viewing multimedia for each content category, and determine whether the number of the actual viewing multimedia is greater than the corresponding viewing threshold. For the content category that has confirmed that the corresponding number of actual viewing multimedia is greater than the viewing threshold, the personal rating update module 89 calculates the personal rating of the heart according to the following formula: .

Where i is the identification code of the content type and is a positive integer; ( i ) for the ith content type value in the new individual rating; ( i ) is the i-th content type value in the current personal appraisal; δ is the number of days in which the actual viewing multimedia corresponding to the i-th content type is viewed; β is the number of days used; t _i is the i-th content The viewing time of the type of the viewed multimedia; th _{1, i} is the first viewing time threshold corresponding to the i-th content type; th _{2, i} is the second viewing time threshold corresponding to the i-th content type; (t _i >th _1,i ) is the number of actual viewing multimedia corresponding to the i-th content type; n(t _i >th _2,i ) is the number of favorite multimedia corresponding to the i-th content type .

Assume that the unit period is 7 days, and the number of days used is 5 days; and for the content type of the news, the first viewing time threshold is 5 minutes, the viewing threshold is 10 times, and the second viewing time threshold is 15 minutes; The group watched the news for 4 days in 7 days, but the number of times the news continued to watch more than 10 minutes (that is, the number of multimedia actually viewed) was 20 times, and the news continued to watch more than 15 minutes (that is, the number of multimedia users) The number is 15 times, and the original personal rating is 0.5 for the news. Since the number of actual viewing multimedia is greater than the threshold of the viewing frequency, the personal rating update module 89 can calculate a new personal rating according to the above formula. ×0.5=0.7. After obtaining a new personal rating, the personal rating update module 89 can update all members of the group to the individual rating of the news to 0.7. In addition, the personal rating update module 89 can be updated after each unit period has elapsed.

In summary, when the group similarity is greater than the similar threshold, the second recommendation module 80 can obtain the fourth rating according to the individual rating, the content-oriented algorithm, the synergistic filtering algorithm, and the usage history, and then according to the The four ratings provide recommendations. The second recommendation module 80 can calculate a new personal rating according to the usage history, and then update the individual rating of each user in the group, so that the subsequent prediction is more accurate.

In addition, the user database 11, the multimedia content database 12, the personal rating database 13, the usage history database 14, the group member weight database 15, the group rating database 16, and the member similarity calculation module 20 are provided. And the recommendation module 40 can be integrated into a set-top box (STB), a television tuner or a universal serial bus television tuner (USB TV tuner).

Performance simulation data for the multimedia content recommendation system and method are provided below.

The simulation data uses the Polylens database (M.O' Connor et al., 2001) and the MovieLens database (http://www.movielens.org/), which provide data for training and testing the multimedia content recommendation system. Set (data set). In the simulation, the Polylens format is used, and the MovieLens database is used as the personal rating database 13. The group data used in the simulation data is shown in Table 1.

Taking the number of group members equal to 3 as an example, the simulation data and the neural network 90 shown in Fig. 9 can be used to obtain data such as Table 2.

The first predictions obtained are shown in Table 3.

Next, the first group recommendation program and the combined single recommendation result of the prior art are compared, wherein the combined personal preference is to average the individual ratings of all users in the group, and then based on the average individual rating and single person. The recommended algorithm is recommended. For performance evaluation, calculate the mean absolute error using the following formula: E = .

Where E is the average absolute error; R _G (i) is the first predicted rating of the group for the ith content type; P _G (i) is the group rating of the group for the ith content type; N is The amount of multimedia content being predicted.

The average absolute error data calculated by the simulation data is as shown in Table 4.

As can be seen from Table 4, for groups with multiple users (group members), regardless of the similarity of the members between the users, the average absolute error of the recommended content obtained by the first group recommendation program is Compared with the average absolute error of the combined single recommendation results of the prior art, it can be seen that the recommendation accuracy of the first group recommendation procedure is higher.

In addition, below, and for a plurality of groups having different member similarities, the first group recommendation program and the second group recommendation program are respectively tested, and the average absolute error data is obtained as shown in Table 5, and the time taken for the recommendation is performed. As shown in Table 6.

As can be seen from Table 5 and Table 6, when the member similarity is very high (for example, higher than 0.7), the average absolute error of the second group recommendation procedure is only slightly higher than the average absolute error of the first group recommendation procedure, but the second The time taken by the group recommendation program is overwhelmingly less than the time spent by the first group recommendation program. Therefore, the multimedia content recommendation method and system can process the group with higher member similarity by the second group recommendation program to speed up the processing.

The above detailed description of the preferred embodiments is intended to provide a clear description of the features and spirit of the present invention, and is not intended to limit the scope of the present invention. On the contrary, the purpose is to cover the various changes and equivalence arrangements within the scope of the patent application to which this proposal is intended.

11. . . User database

12. . . Multimedia content database

13. . . Personal rating database

14. . . Usage history database

15. . . Group member weight database

16. . . Group rating database

20. . . Member similarity calculation module

40. . . Recommended module

60. . . First recommendation module

62. . . Neural network prediction module

64. . . Weight training module

80. . . Second recommendation module

82. . . Personal evaluation merger module

84. . . Content-oriented prediction module

86. . . Concord Filtering Prediction Module

88. . . Predictive merge module

89. . . Personal rating update module

90. . . Neural network

91. . . Input layer

92, 92a, 92b, 92c, 92d. . . Input layer node

93. . . Hidden layer

94, 94a, 94b, 94c. . . Hidden layer node

95. . . Output layer

96. . . Output layer node

110. . . Use history

120. . . Unit period

130. . . Number of days used

Figure 1 is a block diagram of a multimedia recommendation system of an embodiment.

Figure 2 is a flow chart of a multimedia recommendation method of an embodiment.

Figure 3 is a flow chart of an embodiment S300.

Figure 4 is a block diagram of a multimedia recommendation system of an embodiment.

Figure 5 is a block diagram of a first recommended module of an embodiment.

Figure 6 is a flow chart of the first group recommendation procedure of an embodiment.

Figure 7 is a schematic diagram of a neural network such as an embodiment.

Figure 8 is a schematic diagram of a neural network such as an embodiment.

Figure 9 is a schematic diagram of a neural network such as an embodiment.

Figure 10 is a schematic diagram of a neural network such as an embodiment.

Figure 11 is a flow chart of the weight training of an embodiment.

Figure 12 is a block diagram of a second recommended module of an embodiment.

Figure 13 is a flow chart of a second group recommendation procedure of an embodiment.

Figure 14 is a schematic illustration of the unit period of an embodiment.

Figure 15 is a flow chart of a second group recommendation procedure of an embodiment.

16A-E are schematic diagrams of proportional parameters of an embodiment.

Claims (19)

A multimedia content recommendation method for providing at least one recommended content from a plurality of multimedia content of a plurality of content types for a group including a plurality of users, the multimedia content recommendation method comprising: reading each of the uses Comparing one person of the content types; according to the individual ratings, calculating a member similarity of the group, the member similarity represents the degree of similarity between the individual ratings of the users of the group And performing a first group recommendation procedure or a second group recommendation procedure based on the comparison of the membership similarity and a similar threshold value, based on the individual ratings, and based on a machine learning algorithm or A usage history of the group selects at least one of the multimedia content of the content type as the at least one recommended content. The multimedia content recommendation method of claim 1, wherein the step of calculating the similarity of the member of the group according to the individual ratings comprises: using the individual ratings as a plurality of feature vectors; A cosine function value of a vector angle between two of the feature vectors; and an average of the cosine function values as the member similarity. The method for recommending a multimedia content according to claim 1, wherein when the member similarity is less than or equal to the similar threshold, the first group recommendation program is executed, and the first group recommendation program includes: reading a group member weight of each user of the group; calculating a first prediction of the content type of the group according to the machine learning algorithm, the individual ratings, and the group member weights And determining, according to the first predicted rating, at least one of the multimedia content of the content type as the at least one recommended content. The multimedia content recommendation method of claim 3, wherein the group is calculated according to the machine learning algorithm, the individual ratings, and the group member weights of the group. The step of the first predictive rating includes: setting the individual ratings to a plurality of input values of the machine learning algorithm; setting the group member weights to a plurality of initial neuron weights of the machine learning algorithm; The first prediction rating is calculated based on the machine learning algorithm, the input values, and the initial neuron weights. The multimedia content recommendation method of claim 4, wherein the first group recommendation program further comprises a weight training program, the weight training program comprising: reading the group to compare the group of the content types And training the group member weights using a learning algorithm, the individual ratings, and the group rating. The multimedia content recommendation method of claim 5, wherein the weight training program further comprises: updating the current group member weights with the trained group member weights. The method for recommending a multimedia content according to claim 3, wherein when the member similarity is greater than the similar threshold, the second group recommendation process is executed, and the second group recommendation process includes: using a content guide The algorithm calculates a second prediction rating for the content types, and calculates a third prediction rating for the content types by using a synergistic filtering algorithm; reading the usage history of the group, the using The history records the viewing behavior of the group; the number of days of use obtained according to the viewing behavior of the group in the past unit period; the proportion of the usage days in the unit period is set as a proportional parameter And selecting, according to the proportional parameter, the second prediction rating and the third prediction rating to obtain a fourth prediction rating; and selecting, according to the fourth prediction rating, at least one multimedia content of the content type as the at least A recommended content. The multimedia content recommendation method of claim 7, wherein the usage history includes a plurality of viewed multimedia viewed by the group during the past unit period and a viewing time of each of the viewed multimedia. The second group recommendation program further includes: setting the at least one viewed multimedia whose viewing time is greater than a first viewing time threshold to at least one actual viewing multimedia; and the viewing time is greater than a second viewing time threshold At least one of the actual viewing multimedia is set to be at least one favorite multimedia, wherein the second viewing time threshold is greater than the first viewing time threshold; and the number of days in which the actual viewing multimedia is viewed according to the number of usage days, the at least The current personal ratings are updated by the number of actual viewing multimedia, the number of the at least one favorite multimedia, and the current personal ratings. The multimedia content recommendation method of claim 8, wherein the number of days of use, the number of days in which the actual viewing multimedia is viewed, the number of the at least one actual viewing multimedia, and the at least one favorite multimedia The number and the current individual ratings, the steps of updating the individual ratings include: sequentially, for each of the content types, according to the number of usage days, the content of the content type corresponding to the actual viewing multimedia The number, the number of the at least one actual viewing multimedia corresponding to the content type, the number of the at least one favorite multimedia corresponding to the content type, and the individual ratings corresponding to the content type for the content category Calculate the new individual ratings; and update the current individual ratings with the new individual ratings. A multimedia content recommendation system for a plurality of users, configured to provide at least one recommended content from a plurality of multimedia content of a plurality of content types for a group including a plurality of users, the multimedia content recommendation system comprising: a person a rating database for storing each user's rating of the content type; a usage history database for storing a usage history of the group; a member similarity calculation module for Calculating a similarity of the members of the group, the similarity of the members representing the degree of similarity between the individual ratings of the users of the group; and a recommendation module, including a first recommendation a module and a second recommendation module, the recommendation module starts the first recommendation module or the second recommendation module according to the comparison result of the member similarity and a similar threshold value, according to the individual ratings And selecting at least one of the multimedia content of the content type as the at least one push according to a machine learning algorithm or the usage history of the group Content. The multi-user multimedia content recommendation system of claim 10, wherein the member similarity calculation module uses the individual ratings as a plurality of feature vectors; and separately calculates one of the feature vectors A cosine function value of the vector angle; and the average of the cosine function values is taken as the member similarity. The multi-user multimedia content recommendation system of claim 10, further comprising: a group member weight database for storing a group member weight of each user of the group; When the member similarity is less than or equal to the similar threshold value, the recommendation module starts the first recommendation module, and the first recommendation module reads the group member weights from the group member weight database; Determining, according to the machine learning algorithm, the individual ratings and the group member weights, a first predicted rating of the content type of the group; and selecting at least one content according to the first predicted rating At least one of the multimedia content of the type is the at least one recommended content. The multi-user multimedia content recommendation system of claim 12, wherein the first recommendation module sets the individual ratings to a plurality of input values of the machine learning algorithm; The weight is set to a plurality of initial neuron weights of the machine learning algorithm; and the first prediction rating is calculated according to the machine learning algorithm, the input values, and the initial neuron weights. The multimedia content recommendation system for a multi-user according to Item 13 of the claim, further comprising: a group rating database for storing a group rating of the content type of the group; and the first The recommendation module reads the group rating of the group for the content types; and trains the group member weights using a learning algorithm, the individual ratings, and the group rating. The multi-user multimedia content recommendation system of claim 14, wherein the first recommendation module updates the current group member weight database with the trained group member weights. Group member weights. The multi-user multimedia content recommendation system of claim 14, wherein the personal rating database, the usage history database, the group member weight database, the group rating database, and the member are similar The degree calculation module and the recommendation module are integrated in a set-top box, a television card or a universal serial bus bar. The multimedia content recommendation system for a multi-user, as described in claim 12, wherein the usage history records the viewing behavior of the group, and when the member similarity is greater than the similar threshold, the recommendation module starts the first a second recommendation module, wherein the second recommendation module calculates a second prediction rating for the content types by using a content-oriented algorithm, and calculates a content type for the content types by using a synergistic filtering algorithm. a predictive rating; reading the usage history from the usage history database; a usage date obtained according to the viewing behavior of the group in the past one unit period; the usage days are occupied by the unit period The ratio is set to a proportional parameter; the second predicted ranking is weighted and averaged according to the proportional parameter, and a fourth predicted rating is obtained; and at least one of the content types is selected according to the fourth predicted rating. The multimedia content is the at least one recommended content. The multimedia content recommendation system for a multi-user as described in claim 17, wherein the usage history includes a plurality of viewed multimedia that the group has watched during the past period of the unit, and each of the multimedia that has been viewed. a viewing time; the second recommendation module sets the at least one viewed multimedia whose viewing time is greater than a first viewing time threshold to at least one actual viewing multimedia; the viewing time is greater than a second viewing time threshold At least one of the actual viewing multimedia is set to be at least one favorite multimedia, wherein the second viewing time threshold is greater than the first viewing time threshold; and according to the number of usage days, the number of days in which the actual viewing multimedia is viewed, At least one of the actual number of multimedia viewings, the number of the at least one favorite multimedia, and the current personal ratings, and the current personal ratings in the personal rating database are updated. The multimedia content recommendation system for a multi-user, as described in claim 18, wherein the second recommendation module sequentially corresponds to the content type according to the usage date and the corresponding content type. The number of days in which the actual viewing multimedia is viewed, the number of the at least one actual viewing multimedia corresponding to the content type, the number of the at least one favorite multimedia corresponding to the content type, and the content type for the content type Corresponding to the individual ratings, calculating the new individual ratings; and updating the current individual ratings in the personal rating database with the new personal ratings.