A diversified recommendation method for cross-platform data

Abstract

The application discloses a diversified recommendation method for cross-platform data and relates to the technical field of recommendation systems, and comprises the following steps: constructing a local sensitive hash function according to a distance function; constructing a service sub-index by hashing service quality data of each platform by using the local sensitive hash function; connecting the service sub-index of each platform to form a final service index; constructing a service similarity graph according to the service index; constructing an approximate recommendation algorithm by fusing an accuracy function and a diversity function; and generating an accurate and diverse service recommendation list for a target user according to the service similarity graph. The application utilizes the'similarity preserving' characteristic of the local sensitive hash function, shares the service hash index value among platforms, and does not involve the transmission of original data. On the premise of protecting the original data containing user privacy, the application utilizes multi-source data to provide accurate and diverse recommendation results for users and improves user satisfaction.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of recommendation system, and particularly relates to a diversified recommendation method for cross-platform data. BACKGROUND

[0002] With the rapid development of the Internet and the rapid popularization of mobile terminals, the massive resource information brings the problem of "information overload", and a large amount of meaningless information causes great trouble for people to select valuable information. In view of this situation, the recommendation system emerges as the times require. Through the analysis of the user historical behavior record, the recommendation system further grasps the personal hobby of the user, so that the user can be actively provided with the information that he wants. In the recommendation system, the collaborative filtering technology is most widely used, and its biggest feature is "field independence", which only needs to make recommendations to the target user according to the similarity between groups.

[0003] Although the collaborative filtering technology is popular, there are still some deficiencies in the actual application process, which are as follows: (1) The traditional collaborative filtering technology usually needs to obtain a large amount of user historical records. In the real scene, users may call services on different platforms to meet their own needs, and accordingly, the service quality data generated will be distributed in different platforms. Since the service quality data contains the privacy information of the user, in order to prevent privacy leakage, most service providers are unwilling to share the service quality data of their users. Relying on incomplete user historical data for recommendation may affect the accuracy of the recommendation result; (2) Most traditional collaborative filtering methods often consider that the services in the recommendation list are independent of each other, so there may be redundant items or repeated items in the recommendation list, and the recommendation result generated thereby cannot cover the extensive interests of the user, thereby affecting the user satisfaction. Therefore, how to protect the privacy of the user and use multi-source data to provide accurate and diversified recommendation results for the user so as to improve the user satisfaction is a problem to be solved. SUMMARY

[0004] In order to overcome the above problems in the prior art, the present application provides a diversified recommendation method for cross-platform data.

[0005] The technical scheme adopted by the present application to solve its technical problems is: a diversified recommendation method for cross-platform data, comprising the following steps:

[0006] Step one, constructing a local sensitive hash function according to a distance function;

[0007] Step two, according to the quality data of the service under each platform, using the local sensitive hash function to hash the service quality data to construct a service sub-index;

[0008] Step 3: Connect the service sub-indexes under each platform to form the final service index, and build a service similarity graph based on the service index;

[0009] Step four: Construct an approximate recommendation algorithm that integrates accuracy and diversity functions, and generate an accurate and diverse service recommendation list for the target user based on the service similarity graph.

[0010] The aforementioned method for diversified recommendations based on cross-platform data, in step one, employs a cosine similarity-based locality-sensitive hash function. The specific process includes:

[0011] Let S be a set of points in A-dimensional space. Given points s and s' in S, if h satisfies condition (1), then h is called a locality-sensitive hash value.

[0012] (1)

[0013] Where h(s) and h(s') represent the hash values ​​of points s and s', respectively; p1 and p2 represent probability thresholds and p1 > p2; ε1 and ε2 both represent distance thresholds; and Dis(s,s') represents the distance between points s and s'.

[0014] The above-mentioned diversified recommendation method for cross-platform data, specifically includes step two:

[0015] Let the user set be N = {1, 2, …, n}, the service set be M = {1, 2, …, m}, and the platform set be R = {1, 2, …, r}. When user j calls service i on platform r, the generated quality data is represented as d. i,j,r Based on user call records, for service i under platform r, its user data is modeled as a vector. , where N r For users under platform r, calculate the hash index value of service i according to formula (2). ,

[0016] (2)

[0017] Where, vector It is a hyperplane used to divide space.

[0018] The aforementioned method for diversified recommendations based on cross-platform data selects multiple hash function pairs. The hash is used to obtain multiple boolean values, thus obtaining the sub-index of service i under platform r.

[0019] The above-mentioned diversified recommendation method for cross-platform data includes, in step three, the following steps: each platform publishes the service sub-index obtained in step two, with services as points and similarity relationships between services as edges. That is, when the hash index values ​​of two services are the same, the two services can be considered similar, and an edge connection is established between the two services, thereby completing the establishment of the service similarity graph.

[0020] The above-mentioned diversified recommendation method for cross-platform data, specifically includes step four as follows:

[0021] (1) For target users On the platform For service i that has not been called, quality prediction is performed based on the similarity relationship between services using the service similarity graph.

[0022] (2) From the perspective of accuracy, define the Acc(Q) function to ensure that services with high prediction quality are recommended to the target user:

[0023] ,

[0024] Where Q represents the recommendation list, This indicates the predicted quality value;

[0025] (3) Regarding the diversity of recommendations, consider the direct and indirect diversity of the recommendation list. The direct diversity between services in the recommendation list Q is measured by Jaccard distance; the indirect diversity between services in the recommendation list Q is measured by the expansion rate of points in the service similarity graph. The specific Div(Q) function is defined as follows:

[0026] ,

[0027] in, Measure the expansion rate of list Q. Calculate the direct diversity among services in list Q. This represents the size of the extended set of set Q. Indicates service QoS vector, Indicates service The distance between the two Jaccards, This represents the weighting factor, used to weigh... and ;

[0028] (4) Combining (2) and (3), we can obtain the diversified service recommendation problem with a recommendation list length of K, namely the K-DSR problem. The mathematical form of the K-DSR problem is as follows:

[0029] ,

[0030] Where |Q| = K, K represents the weighting factor, used to balance accuracy and diversity; K represents the length of the recommendation list. Indicates target user A collection of services that have not been invoked.

[0031] The aforementioned method for diversified recommendations based on cross-platform data designs a greedy approximation algorithm to solve the K-DSR problem, providing users with an accurate and diverse list of services.

[0032] The greedy approximation algorithm is described as follows: To implement the K-DSR problem, a function is defined.

[0033] ,

[0034] in, The weighting factor represents the ratio between accuracy and diversity; Q represents the recommendation list, which is selected using a greedy algorithm. Make Reach the maximum.

[0035] The aforementioned method for diverse recommendations based on cross-platform data includes a greedy approximation algorithm specifically described below:

[0036] Input: Service similarity graph, target users Target platform Recommended list length Service quality data;

[0037] Output: A recommendation list generated for the target user. ;

[0038] Initialization list It is an empty set; for the target user On the platform Collection of services not yet invoked Select service i to meet This adds service i to the list. Until the recommended list Length reaches stop.

[0039] The beneficial effects of this invention are that it provides a diversified recommendation method for cross-platform data. By utilizing the "similarity preservation" characteristic of locality-sensitive hash functions, service hash index values ​​are shared between platforms without involving the transmission of original data. Under the premise of protecting user privacy in the original data, it uses multi-source data to provide users with accurate and diverse recommendation results, thereby improving user satisfaction. Attached Figure Description

[0040] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0041] Figure 1 This is a flowchart of an embodiment of the present invention;

[0042] Figure 2 These are accuracy effect diagrams for different platforms in embodiments of the present invention;

[0043] Figure 3 The diagrams show the diverse effects of different platforms in the embodiments of the present invention. Detailed Implementation

[0044] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0045] This invention provides a diversified recommendation method for cross-platform data, such as... Figure 1 As shown, it includes the following steps:

[0046] Step 1: Construct a locality-sensitive hash function based on cosine similarity.

[0047] Since different distance functions correspond to different locality-sensitive hash functions, we choose the locality-sensitive hash function corresponding to cosine similarity here.

[0048] Let S be a set of points in A-dimensional space. Given points s and s' in S, if h satisfies condition (1), then h is called a locality-sensitive hash value.

[0049] (1)

[0050] Where h(s) and h(s') represent the hash values ​​of points s and s', respectively; p1 and p2 represent probability thresholds and p1 > p2; ε1 and ε2 both represent distance thresholds; and Dis(s,s') represents the distance between points s and s'.

[0051] Step 2: Based on the service quality data of each platform, use the locality-sensitive hash function to hash the service quality data and construct a service sub-index. The services mentioned may refer to movies, music, etc.

[0052] The method for constructing a hash index is as follows:

[0053] Let the user set be N = {1, 2, …, n}, the service set be M = {1, 2, …, m}, and the platform set be R = {1, 2, …, r}. When user j calls service i on platform r, the generated quality data is represented as d. i,j,rBased on user call records, for service i under platform r, its user data is modeled as a vector. , where N r For users under platform r, calculate the hash index value of service i according to formula (2). ,

[0054] (2)

[0055] Where, vector It is a hyperplane used to divide space.

[0056] Since LSH is a probabilistic search technique, relying on only one hash function cannot adequately reflect the "similarity preservation" characteristic of LSH. Therefore, multiple hash function pairs are selected. Hash the values ​​to obtain multiple boolean values, thus obtaining the sub-index of service i under platform r.

[0057] (2).

[0058] Step three: Following step two, we obtain the hash sub-index of service i under platform r. Each platform publishes its sub-index, and the sub-indexes from the R platforms are connected to obtain the complete index of the service. Using services as nodes and the similarity relationships between services as edges—that is, when two services have the same hash index value, they are considered similar, and an edge is established between them—we complete the service similarity graph. Here, the platform only publishes the index values ​​and does not need to publish the actual quality data of the services, thus protecting user privacy.

[0059] The specific method for step four is as follows:

[0060] (1) For target users On the platform For service i that has not been called, quality prediction is performed based on the similarity relationship between services using the service similarity graph.

[0061] (2) From the perspective of accuracy, define the Acc(Q) function to ensure that services with high prediction quality are recommended to the target user:

[0062] ,

[0063] Where Q represents the recommendation list, This indicates the predicted quality value;

[0064] (3) Regarding the diversity of recommendations, consider the direct and indirect diversity of the recommendation list. The direct diversity between services in the recommendation list Q is measured by Jaccard distance; the indirect diversity between services in the recommendation list Q is measured by the expansion rate of points in the service similarity graph. The specific Div(Q) function is defined as follows:

[0065] ,

[0066] in, Measure the expansion rate of list Q. Calculate the direct diversity among services in list Q. This represents the size of the extended set of set Q. Indicates service QoS vector, Indicates service The distance between the two Jaccards, This represents the weighting factor, used to weigh... and ;

[0067] (4) Combining (2) and (3), we can obtain the diversified service recommendation problem with a recommendation list length of K, namely the K-DSR problem. The mathematical form of the K-DSR problem is as follows:

[0068] ,

[0069] Where |Q| = K, K represents the weighting factor, used to balance accuracy and diversity; K represents the length of the recommendation list. Indicates target user A collection of services that have not been invoked.

[0070] Since the K-DSR problem is NP-hard, a greedy approximation algorithm is designed to solve it, providing users with an accurate and diverse list of services. The greedy approximation algorithm is described below:

[0071] To implement the K-DSR problem, define a function. :

[0072] .

[0073] Use a greedy algorithm to select the service list. Make To reach the maximum, the pseudocode is shown in Algorithm 1:

[0074] Algorithm 1: Greedy algorithm

[0075] Input: Service similarity graph, target users Target platform Recommended list length Service quality data;

[0076] Output: A recommendation list generated for the target user. ;

[0077] First, initialize the list. It is an empty set; secondly, for the target user On the platform Collection of services not yet invoked Select service satisfy Thus, the service Add to list Until the recommended list Length reaches stop.

[0078] The accuracy and diversity effects of the recommendation method proposed in this embodiment are evaluated. To verify that the recommendation results generated by integrating data from multiple platforms are better than those generated by relying solely on data from a single platform, effect diagrams from two platforms are selected. Here, the MovieLens dataset is divided into two parts to construct data for different platforms, thus verifying the effectiveness of the above recommendation method on different platforms. The effect diagrams are as follows. Figure 2 and Figure 3 As shown, PDSR is the method proposed in this invention, while BPR, MPR, EF, MF-based, and LSH-based are the selected comparison methods. BPR, MPR, EF, and MF-based methods implement recommendations considering only single-platform data, while the LSH-based method considers recommendations considering multi-platform data. To verify the effectiveness of the proposed method, MAE and ILD metrics are selected for evaluation. MAE (Mean Absolute Error) is an accuracy evaluation metric, calculated as shown below.

[0079] ,

[0080] Where K represents the length of the recommendation list, and Q is the recommendation list generated for the user. Indicates the predicted value. This represents the original value. The smaller the MAE value, the more accurate the recommendation result. Figure 2 As shown, ILD (Inter List Diversity) is a diversity evaluation indicator, and its calculation method is as follows.

[0081] ,

[0082] in, Indicates service The higher the ILD value, the more diverse the recommendation results.

[0083] according to Figure 2 and Figure 3 The results show that the PDSR method proposed in this invention can integrate data from multiple platforms while protecting privacy, and each platform can achieve accurate and diverse recommendation results by utilizing multi-source data.

[0084] The above embodiments are merely exemplary embodiments of the present invention and are not intended to limit the present invention. The scope of protection of the present invention is defined by the claims. Those skilled in the art can make various modifications or equivalent substitutions to the present invention within its spirit and scope of protection, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of the present invention.

Claims

1. A diversified recommendation method for cross-platform data, characterized in that: Includes the following steps: Step 1: Construct a locality-sensitive hash function based on the distance function; Step 2: Based on the service quality data of each platform, use a locality-sensitive hash function to hash the service quality data and build a service sub-index; Step 3: Connect the service sub-indexes under each platform to form the final service index, and build a service similarity graph based on the service index; Step 4: Construct an approximate recommendation algorithm that integrates accuracy and diversity functions, and generate an accurate and diverse service recommendation list for the target user based on the service similarity graph; Step four specifically includes: (1) For target users On the platform Services that have not been called i Using a service similarity graph, services that have not been called are identified based on their similarity relationships. i Perform quality prediction; (2) From the perspective of accuracy, the definition Acc ( Q The function ensures that services with high prediction quality are recommended to the target user. ; in, Q This represents a list of recommendations. Indicates the predicted quality value; (3) Regarding the diversity of recommendations, consider the direct and indirect diversity of the recommendation list. Q The direct diversity between services in the list is measured using Jaccard distance; for the recommendation list Q The indirect diversity among services is measured by the expansion rate of points in the service similarity graph; specifically Div ( Q The function is defined as follows: ; in, Measure list Q expansion rate, Calculation list Q The direct diversity of services in China Represents a set Q The size of the extended set, Indicates service QoS vector, Indicates service The distance between the two Jaccards, This represents the weighting factor, used to weigh... and ; (4) Combining parts (2) and (3), we can obtain the recommended list with a length of . K The problem of diversified service recommendations, namely K -DSR issue, K The mathematical form of the DSR problem is as follows: ; Among them, | Q | = K， This represents the weighting factor, used to balance accuracy and diversity; K Indicates the length of the recommendation list. Indicates target user A collection of services that have not been invoked.

2. The diversified recommendation method for cross-platform data according to claim 1, characterized in that, In step one, a cosine similarity-based locality-sensitive hash function is selected. The specific process includes: set up A A set of points in 3D space S Given S Points in s and s’ ,like h If condition (1) is satisfied, then it is called h Locality-sensitive hash value, （1）; in, h ( s )and h ( s’ ) represent points respectively s and s’ The hash value; Represents the probability threshold and ; Both represent distance thresholds; Point s and s’ The distance.

3. The diversified recommendation method for cross-platform data according to claim 1, characterized in that, Step two specifically includes: Let the user set be N ={1,2, …, n The service set is M ={1,2, …, m The platform collection is as follows R ={1,2,…, r }, when user j On the platform r Calling services i The generated quality data is represented as follows: Based on user call records, for the platform r The following services i Model its user data as vectors ,in, Indicates platform r For users below, calculate the service according to formula (2). i hash index value , （2）; Where, vector It is a hyperplane used to divide space.

4. The diversified recommendation method for cross-platform data according to claim 3, characterized in that, Select multiple hash function pairs Hash the data to obtain multiple boolean values, thus providing the service. i On the platform r Sub-indexes.

5. The diversified recommendation method for cross-platform data according to claim 1, characterized in that, Step three specifically includes: each platform publishes the service sub-index obtained in step two, with services as points and similarity relationships between services as edges. That is, when the hash index values ​​of two services are the same, the two services can be considered similar, and an edge connection is established between the two services, thereby completing the establishment of the service similarity graph.

6. The diversified recommendation method for cross-platform data according to claim 1, characterized in that, Regarding the above K - A greedy approximation algorithm is designed to solve the DSR problem, providing users with an accurate and diverse list of services; The greedy approximation algorithm is described as follows: To achieve... K -DSR problem, defining a function ; in, This represents the weighting factor, used to balance accuracy and diversity; Q This represents the recommendation list, and a greedy algorithm is used to select recommendations from it. Make Reach the maximum.

7. The diversified recommendation method for cross-platform data according to claim 6, characterized in that, The greedy approximation algorithm is specifically as follows: Input: Service similarity graph, target users Target platform Recommended list length Service quality data; Output: A recommendation list generated for the target user. ; Initialization list It is an empty set; for the target user On the platform Collection of services not yet invoked Select service i to meet This adds service i to the list. Until the recommended list Length reaches stop.

Images