CN114647787A - A user-personalized recommendation method based on multimodal data - Google Patents

Abstract

The invention provides a user personalized recommendation method based on multi-modal data, and relates to the technical field of network recommendation. According to the method, after historical behavior records and mapping logs allowed to be collected by a user are obtained, the characteristics of all objects related to the historical behavior records and the characteristics of each object to be recommended are extracted and mapped to the same multi-dimensional space, multi-modal data are integrated, then a reinforcement learning model is used as a recommendation system agent, the agent is trained through the collected user records, and the trained recommendation agent is used for recommendation, so that user-customized recommendation of the multi-modal data is achieved. The user historical behavior record, the mapping log, the object set to be recommended and the like oriented in the method provided by the invention can all contain objects (such as texts, pictures, videos and the like) in a plurality of fields, and the difference among multiple modes is blurred by a method of integrating after feature extraction, so that the problem that a traditional recommendation system can only be applied to a single field for recommendation is solved.

Description

A user-personalized recommendation method based on multimodal data

technical field

The invention relates to the technical field of network recommendation, in particular to a method for user personalized recommendation based on multimodal data.

Background technique

In recent years, with the rapid development of the Internet, modern society has become an information-based and digital society, with data flooding the entire world, and information explosion has become the norm. However, in the face of a large amount of data, the utilization rate of information by users is reduced, that is, the problem of information overload (Information over load) occurs. In this regard, recommendation system is one of the key technologies to effectively solve the problem of information overload. In fact, with the rapid development of Internet, Internet of Things and cloud computing technologies, personalized recommendation systems have now become the standard configuration of Internet products. Internet users will face Internet information in the fields of e-commerce, video, news and music. It has been closely related to the recommendation system.

The recommender system makes personalized recommendations to users by acquiring the user's historical behavior data, such as web page browsing data, purchase history or item ratings, etc. The solutions to the recommendation problem include recommendation systems based on content filtering, recommendation systems based on collaborative filtering and hybrid recommendation systems; with the rise of deep learning algorithms, various recommendation algorithms using neural networks are also widely used in the implementation of recommendation systems. Such as recurrent neural network, convolutional neural network, generative adversarial network, etc.; in addition, there are also recommendation technology based on reinforcement learning, recommendation technology based on heterogeneous network, etc.

Most of the current recommendation systems only provide a solution to the problem of user-personalized recommendation based on a single modality, that is, they only learn the user's historical data in a single field (such as news, video, music, etc.), so they can only capture users in a single field. , and make recommendations based on it, which greatly limits the diversity of recommendation results.

SUMMARY OF THE INVENTION

The technical problem to be solved by the present invention is to provide a user-personalized recommendation method based on multimodal data in view of the deficiencies of the above-mentioned prior art. , learn user preferences and form user portraits, which can then recommend multimodal data and expand the diversity of recommendation results.

In order to solve the above-mentioned technical problems, the technical scheme adopted by the present invention is:

A user-personalized recommendation method based on multimodal data, comprising the following steps:

Step 1: Obtain the total object set I and the to-be-recommended object set L ₀ , and specify the types of multimodal information that need to be oriented.

Step 2: For each oriented mode, apply a corresponding feature extraction algorithm to extract features from each object in the total object set and map them to the same mathematical multidimensional space S.

Step 3: Obtain and accumulate user historical behavior records and image logs. The historical behavior record is the user's click history before the image, that is, the list of objects that have been clicked; the image log is the list of objects displayed to the user and the user's click behavior on these objects, 1 means click, 0 means no click.

Step 4: Initialize the recommendation system, input the multi-dimensional space S obtained in step 2 and the user historical behavior records and image logs obtained in step 3 into the recommendation system, and set the recommendation system agent and reinforcement learning environment parameters.

Step 5: Execute the training of the recommender system agent.

Step 6: Use the trained recommender system agent to process the object set L ₀ to be recommended, use the recommendation system to simulate the user's interaction behavior facing the object set L ₀ to be recommended, and generate a corresponding image log D.

Step 7: Extract the set of objects interacted by the user in the image log D, that is, the set of personalized interaction objects predicted by the recommender system agent corresponding to the user, as the total recommendation list L.

Step 8: According to different requirements, the total recommendation list L is processed to generate a multimodal recommendation list.

Step 9: Recommend the generated multimodal recommendation list to obtain specific user interaction results, thereby generating an image log. The agent is further trained and updated every time the log accumulates to a given n.

The beneficial effects of adopting the above technical solutions are: in the user-personalized recommendation method based on multi-modal data provided by the present invention, after obtaining the historical behavior records and image logs that the user allows to collect, extract all objects involved therein, and The features of each object to be recommended are mapped into the same multi-dimensional space to integrate multi-modal data, and then use the reinforcement learning model as the recommendation system agent, train the agent through the collected user records, and use the trained recommendation The agent recommends, so as to realize the user-personalized recommendation for multimodal data. Compared with the prior art, the user's historical behavior records, image logs, and object sets to be recommended in the method proposed by the present invention can all include objects in multiple fields (such as text, pictures, videos, etc.), and can be extracted by feature extraction. The post-integration method blurs the distinction between multiple modalities, thereby solving the problem that a traditional recommender system can only be applied to a single domain for recommendation.

Description of drawings

1 is a flowchart of a recommendation method provided by an embodiment of the present invention;

FIG. 2 is a flowchart of a training method for an agent (rainbow model) provided by an embodiment of the present invention.

Detailed ways

The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and embodiments. The following examples are intended to illustrate the present invention, but not to limit the scope of the present invention.

As shown in FIG. 1 , the user personalized recommendation method based on multimodal data in this embodiment is as follows.

Step 1: Obtain the total object set I and the to-be-recommended object set L ₀ , and specify the types of multimodal information that need to be oriented.

Step 2: For each oriented mode, apply a corresponding feature extraction algorithm to extract features from each object in the total object set and map them to the same mathematical multidimensional space S. For example, CNN or RNN convolutional neural network is used to extract features for images, tf-idf algorithm or cnn convolutional neural network is used for text extraction, video features are extracted by P3D (Pseudo-3DResidual Networks) residual network, and audio is extracted by discrete wavelet transform (DWT). ) or perceptual linear prediction (PLP) algorithms to extract features, etc.

Step 3: Obtain and accumulate user historical behavior records and image logs. The historical behavior record is the user's click history before the image, that is, the list of objects that have been clicked; the image log is the list of objects displayed to the user and the user's click behavior on these objects, 1 means click, 0 means no click.

Step 4: Initialize the recommendation system, input the multi-dimensional space S obtained in step 2 and the user historical behavior records and image logs obtained in step 3 into the recommendation system, and set the recommendation system agent and reinforcement learning environment parameters.

Step 5: Execute the training of the recommendation system agent, as shown in Figure 2, the process is as follows:

Step 5.1: The agent requests a user-related record from the user's historical behavior record and image log obtained in step 3.

Step 5.2: The memory returns the record, and through the result of step 2, the corresponding object is replaced with the feature representation.

Step 5.3: The agent forms a user portrait according to the user's historical behavior records, and executes the images in sequence according to the user portrait. If the agent's action is consistent with the real log, the agent will receive a reward, and if the agent's action is inconsistent with the real log, it will not be awarded. Reward or get punished.

Step 5.4: After executing all the entries in the user image log, it is regarded as completing a reinforcement learning frame. Judgment: if the training termination condition in the reinforcement learning environment parameters set in step 4 is satisfied, the agent training is completed, and then go to step 6 , otherwise return to step 5.1.

Step 6: Use the trained recommender system agent to process the object set L ₀ to be recommended, use the recommendation system to simulate the user's interaction behavior facing the object set L ₀ to be recommended, and generate a corresponding image log D.

Step 7: Extract the set of objects interacted by the user in the image log D, that is, the set of personalized interaction objects predicted by the recommender system agent corresponding to the user, as the total recommendation list L.

Step 8: According to different requirements, process the total recommendation list L to generate a multimodal recommendation list;

When it is necessary to mine users' broad interests in multiple fields, or to diversify recommendation result modalities, the total recommendation list L can be simply sorted and output according to specific scene requirements to generate a multimodal recommendation list L ₁ .

When it is necessary to recommend users in other fields according to a certain object (such as adding pictures for news, recommending text hyperlinks for movies, etc.), from the multi-dimensional space S, according to the nearest neighbor principle, select the closest distance to the current object. Other modal objects, generate a multimodal recommendation list L ₂ .

When it is necessary to combine different levels of knowledge expression to make more accurate and effective recommendations to users, the multimodal fusion algorithm is used to fuse the features of different modalities, and the agent is trained based on the fusion features to generate the same The object's recommendation list L3 with multimodal expressions, such as news headlines with news covers.

Step 9: Recommend the generated multimodal recommendation list to obtain specific user interaction results, thereby generating an image log. The agent is further trained and updated every time the log accumulates to a given n.

This embodiment sequentially executes the method of the present invention based on MIcrosoft News Dataset (MIND for short) data set:

Step 1: Since the MIND dataset only contains news text data, which does not meet the needs of multimodal tasks, the spiderFlow crawler tool (https://github.com/chenyuansgit/spiderFlow) crawls according to the news links provided in the MIND dataset The news contains corresponding picture data, forming a picture data set corresponding to the news text. Therefore, the obtained total object set I is all news texts and news pictures, and the object set L ₀ to be recommended is the image log news text and the corresponding picture set of the news in the test set, and it is clear that the types of multimodal information that need to be oriented are text and images.

Step 2: For the two modalities of text and images, the corresponding feature extraction algorithms are respectively applied. The news text features are weighted by the Tf-idf term to extract the basic text features, and then the chi-square test algorithm is used for feature selection. In order to reduce the dimension of the basic text features obtained above, the generalization ability of the model is stronger, overfitting is reduced, and the understanding between features and feature values is enhanced. The news picture features are extracted by the pre-trained VGG16 neural network in the keras library, and finally the extracted features of each object in the total object set are mapped to the mathematical in the same multidimensional space S.

Step 3: Read the user historical behavior records and image logs accumulated in the data set.

Step 4: Initialize the recommendation system, input the multi-dimensional space S obtained in step 2 and the user historical behavior records and image logs obtained in step 3 into the recommendation system, and set the recommendation system agent and reinforcement learning environment parameters (rainbow model as an example): action space It is an integer from 0 to 10, divided by 10 to predict the probability of the object being clicked; the state space is a multi-dimensional matrix, reflecting the user portrait and the characteristics of the current object to be determined; the reward is obtained from the prediction result of the image log output auc indicator decision.

Step 5: Execute the training of the recommendation system agent, as shown in Figure 2, the process is as follows:

Step 5.1: The agent requests a user-related record from the set obtained in step 3;

Step 5.2: The memory returns the record, and through the result of step 2, the corresponding object is replaced with the feature representation;

Step 5.3: According to the user's historical records, the agent uses the Mean-Shift algorithm to cluster and separate the user's preference features to form a user portrait, and predicts the user's click-through rate on the objects in the image log one by one according to the obtained user portrait. The auc index calculated by the real record is given to the agent as a reward;

Step 5.4: After executing all the entries in the user image log, it is regarded as completing a reinforcement learning frame. Judgment: if the training termination conditions in the reinforcement learning environment parameters set in step 4 are met, the agent training is completed, otherwise, return to the step 5.1.

Step 6: Use the trained recommender system agent to process the image log in the test set, predict the user's click probability on each object according to the user's historical behavior record in the test set, calculate the auc index, and record the prediction result D.

Step 7: Use the object with high click probability in the prediction result D as the medium recommendation list L.

Step 8: According to different requirements, the total recommendation list L can be processed to generate a multimodal recommendation list; the following requirements processing process is as follows:

When it is necessary to mine users' broad interests in multiple fields, or to diversify recommendation result modalities, the total recommendation list L can be simply sorted and output according to specific scene requirements to generate a multimodal recommendation list L ₁ .

When it is necessary to recommend users in other fields based on a certain object (such as adding pictures for news, or recommending news hyperlinks for pictures, etc.), from the multi-dimensional space S, according to the nearest neighbor principle, select the closest distance to the current object. other modal objects, generate a multimodal recommendation list L ₂ .

Step 9: Recommend the generated multimodal recommendation list to obtain specific user interaction results, thereby generating an image log. The agent can be further trained and updated whenever a given n logs are accumulated.

Through the above steps, the finally obtained experimental result indicators are shown in Table 1 (step 5 adopts different recommendation models respectively).

Table 1 Experimental results of various recommendation models

The reinforcement learning model of the present invention is used as the recommendation method for the recommendation system agent, and the test is carried out on the MIcrosoft News Dataset (MIND for short) data set. The results show that the performance indicators of the reinforcement learning agent and the traditional recommendation model are similar, that is, it can be Competent for recommended tasks. The data of the MIND data set in the picture mode is expanded by the crawler to make it a data set of text and picture mixed mode, and further experiments are carried out. From the test results, the indicators of the present invention can also meet the requirements for The user's personalized recommendation requirements of multimodal data meet the requirements of the recommendation system.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, but not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that it can still be The technical solutions described in the foregoing embodiments are modified, or some or all of the technical features thereof are equivalently replaced; and these modifications or replacements do not make the essence of the corresponding technical solutions depart from the scope defined by the claims of the present invention.

Claims (7)

1. a user personalized recommendation method based on multimodal data, is characterized in that: the method comprises the following steps: Step 1: Obtain the total object set I and the object set L ₀ to be recommended, and specify the type of multimodal information that needs to be oriented; Step 2: For each mode of orientation, apply the corresponding feature extraction algorithm to extract features from each object in the total object set and map them to the same mathematical multidimensional space S; Step 3: Obtain and accumulate user historical behavior records and image logs; Step 4: Initialize the recommendation system, input the multi-dimensional space S obtained in step 2 and the user historical behavior record and image log obtained in step 3 into the recommendation system, and set the recommendation system agent and reinforcement learning environment parameters; Step 5: Execute the training of the recommender system agent; Step 6: use the trained recommendation system agent to process the object set L ₀ to be recommended, use the recommendation system to simulate the user's interaction behavior towards the object set L ₀ to be recommended, and generate a corresponding image log D; Step 7: Extract the object set that the user interacted with in the image log D, that is, the personalized interactive object set predicted by the recommender system agent corresponding to the user, as the total recommendation list L; Step 8: According to different requirements, process the total recommendation list L to generate a multimodal recommendation list; Step 9: Recommend the generated multimodal recommendation list to obtain specific user interaction results, thereby generating image logs; when the logs accumulate to a given n, further train and update the agent. 2. The user-personalized recommendation method based on multimodal data according to claim 1, wherein the historical behavior record in the step 3 is the click history record of the user before the image, that is, the clicked object list; The image log is the list of objects displayed to the user and the user's click behavior on these objects, 1 means click, 0 means no click. 3. The user-personalized recommendation method based on multimodal data according to claim 1, wherein the specific process of the step 5 is: Step 5.1: The agent requests a user-related record from the user's historical behavior record and image log obtained in step 3; Step 5.2: The memory returns the record, and through the result of step 2, the corresponding object is replaced with the feature representation; Step 5.3: The agent forms a user portrait according to the user's historical behavior records, and executes the images in sequence according to the user portrait. If the agent's action is consistent with the real log, the agent will receive a reward, and if the agent's action is inconsistent with the real log, it will not be awarded. to reward or receive punishment; Step 5.4: After executing all the entries in the user image log, it is regarded as completing a reinforcement learning frame. Judgment: if the training termination condition in the reinforcement learning environment parameters set in step 4 is satisfied, the agent training is completed, and then go to step 6 , otherwise return to step 5.1. 4. The method for user personalized recommendation based on multimodal data according to any one of claims 1-3, characterized in that: in the step 2, the image uses a CNN convolutional neural network to extract features; the text uses tf- Features are extracted by idf algorithm or CNN convolutional neural network; video features are extracted by P3D residual network, and audio features are extracted by discrete wavelet transform or perceptual linear prediction algorithm. 5. The user-personalized recommendation method based on multimodal data according to any one of claims 1-3, characterized in that: in step 8, when it is necessary to mine the extensive interests of users in multiple fields, or When diversified recommendation result modalities are required, the total recommendation list L is simply sorted according to specific scene requirements and then output to generate a multi-modal recommendation list L ₁ . 6. The method for user personalized recommendation based on multimodal data according to any one of claims 1-3, characterized in that: in the step 8, when it is necessary to perform a user recommendation in other fields according to a certain object When recommending, from the multi-dimensional space S, other modal objects closest to the current object are selected according to the nearest neighbor principle, and a multi-modal recommendation list L ₂ is generated. 7. The user-personalized recommendation method based on multimodal data according to any one of claims 1-3, characterized in that: in step 8, when it is necessary to combine different levels of knowledge expression to perform more accurate and effective on users When recommending, a multi-modal fusion algorithm is used to fuse the features of different modalities, and the agent is trained based on the fused features to generate a recommendation list L3 with multi-modal expressions for the same object.

Images