A personalized recommendation technology suitable for platform software cold start users

Abstract

The application relates to a personalized recommendation technology suitable for cold start users of platform software, which comprises the following steps: dividing users into new users and old users according to the sparsity of an interaction matrix; learning embedding representation space by using a hidden model for the old users, representing the users and items as d-dimensional embedding vectors, and calculating user-item similarity by vector cosine similarity; segmenting the clustering sequence of the old users by using a k-means clustering algorithm according to the user embedding vectors; and integrating new users into the existing user embedding space and predicting the item preference of the new users.The application has the beneficial effect that the most popular item preferences calculated in advance by the old users in the cluster are recommended to the corresponding new users, and the recommendation effect of the platform is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of information recommendation, and more particularly to a personalized recommendation technology suitable for cold start users of platform software. BACKGROUND

[0002] Many platform software relies heavily on recommendation system to improve user experience. However, it is a great challenge to profile the cold start users (i.e. new users who have not generated enough interaction behaviors or usage records with the platform software) and recommend the items (such as songs or short videos) that the users are interested in.

[0003] Existing recommendation technologies for cold start users are mostly based on collaborative filtering or deep learning models. Among them, the collaborative filtering technology first constructs the interaction matrix of users and items and the similarity matrix between users, and calculates the recommendation list of users based on the interaction matrix and the similarity matrix; while the deep learning model aims to directly calculate the similarity of the representation vectors of users and items, and combines the rating matrix of users and items to obtain the recommendation list of users. However, neither collaborative filtering nor deep learning model recommendation technology can make good improvements to the personalized recommendation of cold start users, because new users have not interacted with the software or have only performed a small amount of interaction behaviors in the initial stage of using the platform software. On the other hand, new users often rely on the recommended content of the platform software in the initial stage, and are likely to give up using the platform software quickly and decisively because they are not interested in the recommended content of the platform software. SUMMARY

[0004] The present application aims to overcome the deficiencies in the prior art and provides a personalized recommendation technology suitable for cold start users of platform software.

[0005] In a first aspect, a personalized recommendation technology suitable for cold start users of platform software is provided, comprising:

[0006] S1, obtaining the interaction information of users with the platform, and establishing an interaction matrix between users and items according to the interaction information; dividing the users whose interaction matrix sparsity is greater than a preset threshold into new users, and dividing the users whose interaction matrix sparsity is less than or equal to the preset threshold into old users;

[0007] S2, learning an embedding representation space by using a latent model for the old users, representing the users and items as d-dimensional embedding vectors, and calculating the user-item similarity by vector cosine similarity;

[0008] S3, dividing the clustering sequence of old users by using a k-means clustering algorithm according to the user embedding vectors, and representing each user cluster by the average amount of user embedding vectors;

[0009] S4, integrating the new user into the existing user embedding space and predicting the new user's item preferences.

[0010] As preferred, S2 comprises:

[0011] S201, generating the user-item low-dimensional embedding vector representation of the old users using the Alternating Least Squares method as the implicit model according to the interaction matrix between the users and the items;

[0012] S202, establishing a PMI matrix, decomposing the item-item matrix using a distributed implementation of the Singular Value Decomposition algorithm based on the co-occurrence of objects in different item sets, and obtaining the item-item low-dimensional embedding vector representation of the old users by averaging the item vectors in the interaction history of the old users;

[0013] S203, calculating the user-item similarity by inner product or cosine similarity.

[0014] As preferred, S4 comprises:

[0015] S401, integrating input features from different sources, including demographic information and interaction information;

[0016] S402, adding the embedding vectors of the demographic information and the embedding vectors of the interaction information into a fully connected layer, and inputting the latent feature representation after full connection into a feedforward neural network for model training;

[0017] S403, inputting the input features of the new user into the trained model to obtain the predicted embedding vector of the new user;

[0018] S404, according to the predicted embedding vector of the new user, including the new user into the corresponding old user cluster, which is the cluster center closest to the predicted embedding vector of the new user, and then recommending the most popular items in the corresponding old user cluster according to the old user-item interaction history.

[0019] As preferred, in S201, an interaction matrix M between users and items is constructed, M(i,j) is equal to 1 if the ith user and the jth item have interaction behavior, otherwise it is 0, and the Alternating Least Squares method is used as the implicit model to obtain the user-item low-dimensional embedding representation vector of the old users by weighted matrix decomposition M.

[0020] As preferred, in S202, the PMI matrix is represented as:

[0021]

[0022] Wherein, p(x), p(y) are the probabilities of items x, y appearing in different item sets of the platform, and p(x, y) represents the probability of items x and y appearing in an item set at the same time.

[0023] As preferred, in S3, in the process of dividing the clustering sequence, the most popular item preference list recommended for each old user is obtained.

[0024] As preferred, in S401, when the demographic information is missing, the information is completed in a user-guided manner.

[0025] In a second aspect, a personalized recommendation device suitable for platform software cold start users is provided, which is used to execute the personalized recommendation technology suitable for platform software cold start users according to any one of the first aspect, and comprises:

[0026] The dividing module is configured to obtain the interaction information of the user with the platform, establish an interaction matrix between the user and the item according to the interaction information, divide the user whose interaction matrix sparsity is greater than a preset threshold into a new user, and divide the user whose interaction matrix sparsity is less than or equal to the preset threshold into an old user.

[0027] The computing module is configured to learn an embedding representation space by using an implicit model for the old user, represent the user and the item as a d-dimensional embedding vector, and calculate the user-item similarity by using vector cosine similarity.

[0028] The dividing module is configured to divide the clustering sequence of the old user by using a k-means clustering algorithm according to the user embedding vector, and represent each user cluster by using the average amount of the user embedding vector.

[0029] The predicting module is configured to integrate the new user into the existing user embedding space, and predict the item preference of the new user.

[0030] In a third aspect, a computer storage medium is provided, and the computer storage medium stores a computer program; when the computer program runs on a computer, the computer program makes the computer execute the personalized recommendation technology suitable for platform software cold start users according to any one of the first aspect.

[0031] In a fourth aspect, a computer program product is provided, and when the computer program product runs on a computer, the computer program product makes the computer execute the personalized recommendation technology suitable for platform software cold start users according to any one of the first aspect.

[0032] The beneficial effects of the present application are: the present application integrates new users into the existing embedding space, so that the most similar items can be retrieved for each user through accurate or approximate nearest neighbor search and some similarity measures, and each cold user is provided with completely personalized item recommendations. In addition to the neural network prediction provided by the present application, each new user is also assigned to the nearest old user cluster cluster of the predicted embedding vector of the new user, and the most popular item preferences of the old users in the cluster are recommended to the corresponding new users. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 The model architecture diagram of the personalized recommendation system for cold start users of platform software provided by the present application;

[0034] Figure 2 The overall flowchart of the personalized recommendation technology for cold start users of platform software provided by the present application;

[0035] Figure 3 The flowchart of learning embedding representation space of old users provided by the present application;

[0036] Figure 4 The flowchart of predicting item preferences of new users provided by the present application;

[0037] The reference signs are explained: interactive information module 1, population information module 2, item stream quantity 3, user age 4, and retrieval stream quantity 5. DETAILED DESCRIPTION

[0038] The present application will be further described below in conjunction with the embodiments. The following description of the embodiments is only for the purpose of helping to understand the present application. It should be noted that for ordinary people in the technical field, some modifications can be made to the present application without departing from the principles of the present application, and these improvements and modifications also fall within the protection scope of the claims of the present application.

[0039] Example 1:

[0040] In order to solve the problem that it is difficult to provide personalized recommendations for cold start users, the present application provides a personalized recommendation system for cold start users of platform software, as shown in Figure 1 The system includes: an interactive information module 1 and a population information module 2. The interactive information of the interactive information module 1 mainly includes music stream embedding representation, composer embedding representation and music retrieval embedding representation. The population information of the population information module 2 mainly includes same city old user embedding representation and same age old user embedding representation. The additional information input of the model includes item stream quantity 3, user age 4 and retrieval stream quantity 5. The final model output is the low-dimensional embedding representation of the new user to be predicted.

[0041] The personalized recommendation system for the cold start user of the platform software is adopted, the old user and the new user are distinguished according to the data information sparsity of the user-item interaction matrix, then the two kinds of collaborative filtering implicit models are used for the old user, a user embedding space is learned, the similarity between them is calculated through the inner product or cosine similarity, then the user embedding based on the above is used, the k-means clustering algorithm is used to calculate the division of the old user, each user cluster is represented by the average of the user embedding, the new user is integrated into the existing embedding space, and finally the item preference of the new user is predicted, so that the overall design of the cold start problem of the recommendation system is realized.

[0042] Embodiment 2:

[0043] As shown in Figure 2 The application provides a personalized recommendation technology suitable for the cold start user of the platform software, which comprises:

[0044] S1, obtaining the interaction information of the user and the platform, establishing the interaction matrix between the user and the item according to the interaction information; dividing the user whose interaction matrix sparsity is greater than a preset threshold into a new user, and dividing the user whose interaction matrix sparsity is less than or equal to the preset threshold into an old user.

[0045] S2, for the old user, two kinds of collaborative filtering implicit models are used to learn an embedding representation space, the user and the item are represented as d-dimensional embedding vectors, and the user-item similarity is calculated through vector cosine similarity.

[0046] Specifically, S2 comprises:

[0047] S201, according to the interaction matrix between the user and the item, using the alternating least square method as the implicit model, generating the user-item low-dimensional embedding vector representation of the old user.

[0048] Specifically, in S201, the elements of the interaction matrix are defined by heuristic rules, and the low-dimensional embedding vector representation of the user-item of the old user is obtained through weighted matrix decomposition. For example, the user-item interaction matrix M is constructed, if the ith user and the jth item have interaction behavior, M(i,j) is equal to 1, otherwise 0, the alternating least square method is used as the implicit model to decompose M through weighted matrix, and the user-item low-dimensional embedding representation vector of the old user is obtained.

[0049] S202. Construct the PMI matrix, inspired by the co-occurrence context of word2vec, based on the co-occurrence of objects in different item sets, such as music playlists in music software. Then, use a distributed implementation of the singular value decomposition algorithm to decompose the item-item matrix, thereby generating a 128-dimensional embedding vector representation for each item. Finally, obtain the low-dimensional item-item embedding vector representation of old users through the item vectors in the average old user interaction history.

[0050] In S202, the PMI matrix is ​​represented as follows:

[0051]

[0052] Where p(x) and p(y) are the probabilities of items x and y appearing in different sets of items on the platform, and p(x,y) represents the probability that items x and y appear simultaneously in the set of items.

[0053] Then, a distributed implementation of the singular value decomposition (SVD) algorithm is used to decompose the project-project matrix. SVD is a decomposition method applicable to any matrix, given a matrix P. d×d There always exists a singular value decomposition:

[0054] P d×d =U d×d ∑ d×k V k×d T

[0055] Define the embedding vector representation as having dimension k, meaning each item is represented by a k-dimensional vector. Using a singular value decomposition model, we can obtain a d×k dimensional matrix ∑ d×k We call this the item embedding vector matrix, which generates an embedding vector representation of dimension k (=128) for each item.

[0056] Finally, by averaging the item vectors of the interaction history of long-term users on the platform software, a low-dimensional embedding vector representation is derived for long-term users.

[0057] S203. In the above embedding space, both users and items are represented as d-dimensional vectors. The similarity between users and items is calculated by the inner product and cosine similarity, which facilitates the calculation of the list of most popular item preferences recommended to each old user in the later steps.

[0058] Specifically, in the aforementioned embedded space, the user and projects All are expressed as k-dimensional vectors, and the similarity between users and items is calculated by inner product and cosine similarity, which facilitates the calculation of the most popular item preference list for each old user in the following steps.

[0059]

[0060] S3, according to the user embedding vector, the k-means clustering algorithm is used to segment the clustering sequence of old users, and each user cluster is represented by the average value of the user embedding vector.

[0061] Specifically, in S3, the clustering sequence of old users is segmented by using the k-means clustering algorithm in the embedding space described above, using k=1000 clusters / segments, and each user segment is represented by the average value of its user embedding vector. During the segmentation of the clustering sequence, the application also pre-calculates the most popular item preference list for each old user, so that in the following steps, the new user is semi-personalized to recommend the top N items in the popular item preference recommendation list of the user cluster to which it belongs (N is defined by the platform software development status and development needs).

[0062] S4, automatically integrate new users into the existing user embedding space and predict the item preferences of new users.

[0063] S4 includes:

[0064] S401, integrate input features from different sources, including demographic information and interaction information; when demographic information is missing, information completion is performed by user guidance.

[0065] Specifically, S401 includes: (1) Collect user demographic information: during registration, all users specify their demographic information, including age and nationality, which is included in the input characteristics, and the corresponding embedding is obtained by the average value of the embedding of old users of the same age and the same country. (2) Collect and process user interaction information. (3) Information completion by using on-boarding, i.e. user guidance. Two steps are added to the new user guidance process after registration: measurement indicators and insights. The application measures the activation rate and engagement indicators respectively, and measures the influence of different signal collection time points on user experience, so as to find a suitable signal collection time to guide new registered users to judge their preferences and submit them to the online recommendation system. Such interaction information is used as a small but crucial amount of interaction information of new users during the registration period.

[0066] S402, add the embedding vectors of demographic information and interaction information to the fully connected layer, and input the fully connected latent feature representation to the feedforward neural network for model training.

[0067] For example, the embedding vectors of the population information of fixed size representation and the embedding vectors of the interaction information are connected to form an input layer of the feedforward neural network with a dimension of 5139 or 2579 (considering the low-dimensional embedding vector representation of user-item or the low-dimensional embedding vector representation of item-item, respectively), while the neural network has three hidden layers with dimensions of 400, 300 and 200, respectively, and an output layer with a dimension of 256 or 128 (considering the low-dimensional embedding vector representation of user-item or the low-dimensional embedding vector representation of item-item, respectively). In addition to the output, the network uses the RELU function as the activation function at each layer, and performs batch normalization processing after the activation function processing. The model of the old user is trained by the present application, and the mean square error between the predicted user embedding and the actual value in the user-item alternating least squares hidden model vector space or the item-item singular value decomposition hidden model vector space is minimized by iteration, while the learning rate is set to 0.0001, the batch size is 512, and the training round is 130 or 100 (considering the low-dimensional embedding vector representation of user-item or the low-dimensional embedding vector representation of item-item, respectively).

[0068] S403, input the input features of the new user into the trained model using the model trained on the old user, obtain the predicted embedding vector of the new user, and integrate the new user.

[0069] S404, according to the predicted embedding vector of the new user, include the new user into the corresponding old user cluster, the corresponding old user cluster is the cluster center closest to the predicted embedding vector of the new user, and then recommend the most popular items in the corresponding old user cluster according to the old user-item interaction history.

[0070] It should be noted that the present application integrates the new user into the existing embedding space. Therefore, the most similar items can be retrieved for each user through accurate or approximate nearest neighbor search and some similarity measures, and completely personalized item recommendations are provided for each cold user. However, as shown by the results in Embodiment 3 described below, such a strategy still leads to noisy results for users who rarely or have no usage data. Therefore, the present application instead adopts a semi-personalized recommendation strategy. In addition to the neural network prediction provided by the present application, the present application also includes the new user into the pre-computed old user sequence, specifically, each new user is assigned to the old user cluster closest to the predicted embedding vector of the new user. Then, the most popular item preferences pre-computed by the old users in the cluster are recommended to the corresponding new user.

[0071] Embodiment 3:

[0072] As shown in Table 1, the present application provides comparative experimental results of different methods, wherein UT-ALS Embeddings represents a low-dimensional embedding vector feature representation of user-item, and TT-SVD Embeddings represents a low-dimensional embedding vector feature representation of item-item.

[0073] In the experiment, the semi-personalized cold start method used in the present application is compared with the popularity algorithm, the registration flow algorithm, the input feature clustering algorithm, the semi-personalized DropoutNet model and the semi-personalized MeLU model, respectively, and the comparison index is the accuracy, the recall rate and the normalized cumulative benefit based on the recommendation system.

[0074] According to the conclusion in Table 1, the cold start user recommendation method provided by the present application achieves the best effect, which is helpful to improve the benefit of the platform software.

[0075] Table 1

[0076]

Claims

1. A personalized recommendation technique for cold start users of platform software, characterized in that, Comprising: S1, obtaining the interaction information between the user and the platform, and establishing the interaction matrix between the user and the project according to the interaction information; Dividing the user with the interaction matrix sparsity greater than the preset threshold into a new user, and dividing the user with the interaction matrix sparsity less than or equal to the preset threshold into an old user; S2, learning embedding representation space by using implicit model for the old user, representing the user and the project as d-dimensional embedding vector, and calculating the user-project similarity by vector cosine similarity; S3, dividing the clustering sequence of the old user by using k-means clustering algorithm according to the user embedding vector, and representing each user cluster by the average amount of user embedding vector; S4, integrating the new user into the existing user embedding space, and predicting the project preference of the new user; S4 comprises: S401, integrating input features from different sources, including demographic information and interaction information; S402, adding the embedding vector of demographic information and the embedding vector of interaction information into the full connection layer, and inputting the latent feature representation after full connection into the feedforward neural network for model training; S403, inputting the input features of the new user into the trained model to obtain the predicted embedding vector of the new user; S404, according to the predicted embedding vector of the new user, integrating the new user into the corresponding old user cluster, the corresponding old user cluster being the cluster center closest to the predicted embedding vector of the new user, and then recommending the most popular project in the corresponding old user cluster according to the old user-project interaction history.

2. The personalized recommendation technique for platform software cold start users according to claim 1, wherein S2 Comprising: S201, generating the user-project low-dimensional embedding vector representation of the old user by using the alternating least squares method as the implicit model according to the interaction matrix between the user and the project; S202, establishing the PMI matrix, decomposing the project-project matrix by using the distributed implementation of the singular value decomposition algorithm based on the co-occurrence of objects in different project sets, and obtaining the project-project low-dimensional embedding vector representation of the old user by averaging the project vectors in the interaction history of the old user; S203, calculating the user-project similarity by inner product or cosine similarity.

3. The personalized recommendation technique for platform software cold start users according to claim 2, characterized in that, In S201, the user-project interaction matrix M is constructed, M(i,j) is equal to 1 if the ith user and the jth project have interaction behavior, otherwise it is 0, and the user-project low-dimensional embedding representation vector of the old user is obtained by using the alternating least squares method as the implicit model through weighted matrix decomposition M.

4. The personalized recommendation technique for platform software cold start users according to claim 3, characterized in that, In S202, the PMI matrix is represented as: Wherein, p(x), p(y) are the probabilities of the occurrence of projects x, y in different project sets of the platform, and p(x,y) represents the probability of the simultaneous occurrence of projects x and y in the project set.

5. The personalized recommendation technique for platform software cold start users according to claim 4, characterized in that, In S3, in the process of dividing the clustering sequence, the most popular project preference list recommended for each old user is obtained.

6. The personalized recommendation technique for platform software cold start users according to claim 5, characterized in that, In S401, when the demographic information is missing, the information is completed by user guidance.

7. A device for personalized recommendation for cold start users of platform software, characterized in that, The personalized recommendation technology suitable for platform software cold start user comprises: The division module is configured to acquire interaction information of a user with a platform, establish an interaction matrix between the user and a project according to the interaction information, divide a user whose interaction matrix sparsity is greater than a preset threshold into a new user, and divide a user whose interaction matrix sparsity is less than or equal to the preset threshold into an old user. The calculation module is configured to learn an embedding representation space by using a latent model for the old user, represent the user and the project as d-dimensional embedding vectors, and calculate user-project similarity by using vector cosine similarity. The segmentation module is configured to segment a clustering sequence of the old user by using a k-means clustering algorithm according to the user embedding vectors, and represent each user cluster by using an average amount of the user embedding vectors. The prediction module is configured to integrate the new user into an existing user embedding space, and predict a project preference of the new user.

8. A computer storage medium, characterized in that, The computer storage medium stores a computer program; when the computer program runs on a computer, the computer program causes the computer to execute the personalized recommendation technology for a platform software cold start user according to any one of claims 1 to 6.

9. A computer program product, characterised in that, When the computer program product runs on the computer, the computer program product causes the computer to execute the personalized recommendation technology for a platform software cold start user according to any one of claims 1 to 6.

Images