Data information recommendation method based on graph neural network

Abstract

The invention discloses a data information recommendation method based on a graph neural network, which captures a high-order relationship between a user and an article layer by layer by using neighborhood aggregation through the graph neural network, effectively relieves the problem caused by data sparsity, and improves the recommendation efficiency by aggregating neighborhood information of the user and the article through the graph neural network. According to the method, richer embedded representations are generated for each user and article, an effective user interest model can still be established through indirect information of high-order neighbors even under the condition that direct interaction data of the users and the articles are rare, in addition, a dynamic attention mechanism is adopted to fuse multi-modal information of user interaction features, text features and image features, and the user interest model can be established through the multi-modal information of the user interaction features, the text features and the image features. And the weight of each mode is adjusted according to the task scene, so that the recommendation individuation and accuracy are improved.

Description

Technical Field

[0001] The present invention relates to the field of data information technology, and in particular to a data information recommendation method based on graph neural network. Background Art

[0002] With the rapid development of the Internet and mobile applications, the problem of information overload has become increasingly prominent. Faced with massive amounts of information, users often find it difficult to quickly and effectively find content of interest to them. In order to help users obtain personalized information, recommendation methods have emerged and have been widely used in e-commerce, social media, music, video and other fields. Common recommendation algorithms include content-based recommendation, collaborative filtering recommendation, hybrid recommendation, etc. However, traditional recommendation methods have the following main problems:

[0003] Most existing recommendation methods rely on explicit interaction data between users and items, such as ratings and purchases. However, in applications, the interaction data between users and items is often very sparse, resulting in many users having no direct interaction records with items, making it difficult for recommendation methods to establish effective user preference models.

[0004] Traditional recommendation methods mostly rely on direct interaction data between users and items, while ignoring more complex high-order relationships between users and items. For example, two users who have not interacted may have an indirect connection through a common item. Similarly, two items may have potential similarities through a common user. Traditional collaborative filtering or content-based recommendation methods have difficulty capturing these high-order user-item relationships, resulting in low accuracy of recommendation results.

[0005] Most existing recommendation methods only use single user behavior data, ignoring other multimodal information generated by users, such as text comments, item descriptions, and pictures. The use of single-modal data makes the recommendation method have great deficiencies in multi-dimensional information fusion and cannot fully explore the user's interest characteristics;

[0006] Based on this, a data information recommendation method based on graph neural network is needed. Summary of the invention

[0007] To achieve the above object, the present invention provides the following solution: a multimodal data information recommendation method based on graph neural network, comprising the following steps:

[0008] Step 1: Obtain the interaction data between users and items, and construct a bipartite graph of users and items, where users and items are nodes in the graph, and the interaction behaviors between users and items are edges in the graph. The initial features of the user and item nodes are updated through neighborhood aggregation through the graph neural network to generate embedded representations of users and items; in the l+1th graph neural network, the embedding representation update formula of user u is:

[0009] Among them, σ is the activation function, W (l) is the training weight matrix, AGG is the neighbor node aggregation function, Nbr(u) is the neighbor item set of user u, is the embedding representation of user u at layer l+1;

[0010] Step 2: Get the user's comment text and the item description text, use the pre-trained language model BERT to encode the comment text and description text, and generate the text feature representation of the user and item. The text feature representation is as follows: in, is the text feature of user u, is the text feature of item v, T u and T v are the comment text of user u and the description text of item v respectively;

[0011] Step 3: Obtain the image data of the object, use the convolutional neural network (CNN) to process the image data of the object, and generate the visual feature representation of the object. The feature representation of the object image is: in, is the image feature of item v, l v is the image data of item v;

[0012] Step 4: Fuse the multimodal features of user and item interaction features, text features, and image features, assign weights to each modality through dynamic attention, and calculate the final feature representation of users and items. The specific steps are as follows:

[0013] Among them, αm and βm are adaptive modal weights, h u and h v is the mth modal feature of users and items;

[0014] Step 5: High-order information propagation is performed through a multi-layer graph neural network. Each layer of the network updates the representation of users and items to capture the high-order relationship between users and items.

[0015] The update formula for user node representation is:

[0016] The update formula for item node representation is:

[0017] in, is the embedding of item v at the l+1th layer, Nbr(u) and Nbr(v) represent the neighbor sets of user u and item v respectively. and is the training weight matrix;

[0018] Step 6: Calculation and generation of recommendation scores: Based on the multimodal feature representation of users and items, the recommendation scores are generated by calculating the inner product of the user and item representations:

[0019]

[0020] in, is the recommendation score of user u for item v, and The final embedding representation of users and items.

[0021] Further preferably, in the step one, the initial features of the user node and the item node are obtained through an embedding layer, and the embedding layer generates their initial vector representations based on the historical behavior data of the user and the item, and maps the user's historical interaction behavior, click records or rating data through an embedding matrix, so that each user and item has a vector representation of a fixed dimension as the input of the graph neural network.

[0022] Further preferably, the aggregation function of the step one adopts a mean aggregation method, that is, the representation of the user or item node is updated by calculating the mean of the representations of its neighboring nodes. For each user node, the features of its direct neighboring item nodes are aggregated, and a new user node representation is generated by weighted averaging or simple mean calculation of the features of these item nodes; similarly, the feature representation of the item node is updated by calculating the feature mean of the neighboring user nodes.

[0023] Further preferably, in the step 2, the language model BERT is a pre-trained bidirectional transformer model, which captures contextual information in comments or description texts and generates high-dimensional semantic embeddings. BERT obtains contextual dependencies in sentences through its bidirectional attention, maps each word in user comments or item descriptions to a vector in a high-dimensional space, and processes the vectors through a multi-layer neural network to generate a high-dimensional vector representation of a fixed length related to the original text.

[0024] Further preferably, in the step three, the convolutional neural network includes multiple convolutional layers and pooling layers, which are used to extract local and global features of the item image and generate a high-dimensional image embedding representation. The image input of the item captures local features through multiple layers of convolution operations. The operation of each layer of convolution kernel can extract low-level features of edges and textures in the item image, and after passing through multiple convolutional layers, the high-level semantic features of the object image are gradually extracted; the pooling layer is used for downsampling, reducing the dimension of the feature map, retaining the main features, and reducing the computational complexity; the convolutional features are mapped to a vector space of fixed dimensionality through a fully connected layer to generate an image feature embedding of the object.

[0025] Further preferably, in the step four, the attention network is composed of a multi-layer perceptron, and weights are assigned according to the feature importance of each modality. For the multimodal feature representation of each user and item, the attention network first initializes the weights of the features of different modalities, and then calculates the importance of each modality to the current recommendation task through the multi-layer perceptron, and outputs the attention score of the modality; the attention score is then normalized by a softmax operation to generate the final weight of each modality; the final user or item feature representation is the weighted sum of the features of each modality.

[0026] Further preferably, in the step five, the high-order information propagation step is implemented by stacking a multi-layer graph neural network, each layer of the network updates the feature representation of the user and item nodes once, and gradually integrates the high-order information from the neighboring nodes. Through the iterative propagation of the multi-layer graph neural network, the user node not only updates the item node features from its direct neighbors, but also obtains the influence of indirect neighbor nodes through multi-layer superposition, that is, gradually aggregates information through second-order, third-order and higher-order neighbor nodes; each layer of the graph neural network performs neighborhood aggregation and activation operations on the node features updated in the previous layer.

[0027] Further preferably, in step 6, the recommendation score is calculated based on the inner product of the final feature representation of the user and the item. The higher the recommendation score, the more the recommended item meets the user's interest preference. The final feature representation of the item Inner product operation to generate the user's recommendation score for the item The higher the score, the more the item matches the user's interests. All items are sorted according to their scores, and the item with the highest score is selected as the recommendation result.

[0028] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0029] 1. The present invention uses graph neural networks to capture high-order relationships between users and items layer by layer through neighborhood aggregation, effectively alleviating the problems caused by data sparsity. Graph neural networks generate richer embedding representations for each user and item by aggregating neighbor information of users and items. Even when the direct interaction data between users and items is scarce, effective user interest models can still be established through indirect information from high-order neighbors. Compared with traditional recommendation algorithms, they can capture potential user preferences in sparse data environments.

[0030] 2. The present invention can transmit and capture the complex high-order relationships between users and items layer by layer through multi-layer propagation of graph neural networks. Traditional recommendation methods can only rely on direct interaction data, while the present invention uses high-order information propagation to achieve in-depth mining of potential connections between users and items, making the recommendation method more expressive when processing user preferences, significantly improving the accuracy and diversity of recommendation results.

[0031] 3. The present invention combines multimodal feature fusion with dynamic attention, and adaptively allocates weights according to the contribution of each modal data to the recommendation task, thereby solving the problem of insufficient utilization of multimodal information in the prior art. By fusing user behavior data, text descriptions of items, images and other information, the present invention understands user interests and item characteristics from multiple perspectives. Compared with the traditional reliance on single behavior data, the present invention can make full use of multimodal information and provide more comprehensive and detailed recommendation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0033] Figure 1 It is a schematic diagram of the present invention. DETAILED DESCRIPTION

[0034] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0035] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0036] Example

[0037] like Figure 1 As shown, a data information recommendation method based on a graph neural network in this embodiment includes:

[0038] Step 1: Construction and encoding of user-item interaction data. In the multimodal data information recommendation method based on graph neural network of the present invention, in order to process the interaction data between users and items, it is necessary to represent the interaction between users and items as a bipartite graph. In the bipartite graph, users and items are respectively used as nodes of the graph, and the click, browse, purchase, and rating interactions between users and items are used as edges in the graph. The user's historical interaction records are obtained through a database or log file, and the data is converted into the form of a bipartite graph, as follows;

[0039] Acquisition of initial features: After constructing the bipartite graph of users and items, it is necessary to generate the initial feature representation for each user and item node. The initial features are obtained through the embedding layer. The embedding layer generates the initial vector representation based on the historical behavior data of users and items. Specifically, the user's click records, browsing history, and rating data are mapped into vectors of fixed dimensions through the embedding matrix. The embedding representations of user u and item v are defined as follows: Among them, Embedding(u) and Embedding(v) represent the embedding matrices of users and items respectively. The initial embedding representations of users and items not only capture the user's preferences and behavior patterns for items, but also provide high-quality input for the subsequent aggregation process of the graph neural network.

[0040] Neighbor aggregation of user and item nodes: This is achieved through neighborhood aggregation. For each user node u, its neighbor nodes are the set of item nodes that interact with it, denoted as Nbr(u). Each layer of the graph neural network updates the user's representation based on the features of the user's neighbor items. In the lth layer of the graph neural network, the embedding representation of user u is updated by the feature aggregation of its neighbor item nodes. The formula is as follows:

[0041]

[0042] The aggregation function AGG uses mean aggregation to perform weighted average or simple mean calculation on the features of neighbor nodes;

[0043] Item node update process: For each item node v, its neighbor nodes are the set of user nodes that interact with it, denoted as Nbr(v). Similar to the update of user nodes, the representation of item nodes is also updated through the features of their neighbor users. In the l-th layer of the graph neural network, the representation of item node v is updated as follows:

[0044]

[0045] Among them, the item node is also averaged and aggregated through the features of its neighboring users to update its embedding representation;

[0046] In the implementation process of the present invention, the mean aggregation is selected as the aggregation function to effectively handle the interference caused by extreme values ​​or abnormal data. The advantage of mean aggregation is that it gives equal weight to all neighbor nodes, ensuring that the contribution of each neighbor node can be reasonably considered, thereby eliminating the adverse effects of extreme behaviors on node representation. Through the multi-layer graph neural network, the node representation of users and items will fuse all information from their neighbor nodes to generate the final embedded representation;

[0047] Step 2: Extraction and encoding of text features. The present invention uses a pre-trained BERT model to perform deep semantic encoding on text data to generate high-quality text feature embedding;

[0048] In the recommendation method, the user's interactive behavior on the item is not only reflected in the click, browse and purchase data, but also accompanied by text comments. The comments usually contain the user's subjective feelings and specific feedback on the item, which can reflect the user's actual needs and preferences. The item itself is usually accompanied by text description information, which details the function, use or other characteristics of the item. These descriptions can provide more information about the item.

[0049] The text data consists of two parts: one is the user's comment text on the item, denoted as T u ; The second is the description text of the item, denoted by T v ;

[0050] The present invention adopts the pre-trained BERT model. BERT is a bidirectional encoder based on transformers, which can capture the contextual information in the text. That is, it can not only predict the subsequent text based on the previous text, but also understand the meaning of the previous text through the subsequent text. The input of BERT is a text sequence after word segmentation. It calculates the relationship between each word in the sentence and other words through the self-attention mechanism, and generates a high-dimensional context-related embedding representation. The comment text T of user u u After being encoded by the BERT model, the corresponding text embedding representation is generated BERT processes the input text sequence through a multi-layer transformer structure and finally outputs a text representation vector of fixed dimension: Among them, T u It is the user's comment text on the item, and BERT converts it into a high-dimensional text embedding vector Since the BERT model can capture bidirectional contextual dependencies, similarly, the description text T of item v v It is also encoded through the BERT model to generate a text embedding representation of the item BERT processes item description text in a similar way to user review text, which can be expressed as follows: Among them, T v is the description text of the item, BERT converts it into a high-dimensional text embedding vector By extracting the semantics of item description text, we can better understand the characteristics of items.

[0051] Step 3: Extraction and encoding of image features. The present invention extracts and encodes the image data features of the object through a convolutional neural network to generate a high-dimensional image embedding representation of the object. The convolutional neural network can extract semantic information from the local and global features of the object image and capture the details and overall structure of the image. It first extracts the object image l v Preprocessing makes the input image adapt to the input layer of the convolutional neural network. Then, after multiple convolutional layers and pooling layers, the features in the item image are gradually extracted. The convolution layer slides the image through the convolution kernel to extract the edge and texture local features in the image, and through layer-by-layer convolution operations, the high-level semantic features in the item image are gradually captured. The pooling layer is responsible for downsampling the feature map after convolution, reducing the dimension of the feature map, while retaining important image information, reducing computational complexity, and avoiding overfitting. In the multi-layer iterative process of convolution and pooling, the visual features of the item are fully extracted and converted into high-dimensional feature vectors. Finally, the extracted image features are mapped to a fixed-dimensional vector space through the fully connected layer to generate the final image embedding representation of the item. This high-dimensional vector representation can comprehensively describe the visual information of the object, provide visual feature data for the present invention, and accurately match the interests and needs of the user.

[0052] Step 4: This step adopts multimodal feature fusion and adaptively adjusts the weights of each modality through the dynamic attention mechanism. The multimodal feature representations of users and items are obtained from the previous steps. For user u, the multimodal features include interaction features Text Features Like features of other modalities, the multimodal features of item v also include interactive features. Text Features And image features Features of different modes have different properties and information carrying capacity, so simply superimposing them linearly is likely to lead to information loss or redundancy. The final feature representation of the present invention and the article is generated by weighted summation of the features of each mode, and the formula is expressed as follows:

[0053] in, and They represent the features of users and items in the mth mode respectively, am and βm are the weights of the mth mode, and the weights are calculated through the dynamic attention mechanism. The attention mechanism learns the features of each mode through the multi-layer perceptron MLP, calculates its importance score for the current recommendation task, and then normalizes the score through the softmax function to generate the weight of each mode. When the user's behavior is mainly reflected in text comments, the text features will obtain a higher weight; in visually oriented scenarios, such as fashion e-commerce or picture recommendations, image features will be given a higher weight. Through this dynamic weight allocation method, the most important features for the recommendation results are screened out in the multimodal information, thereby generating more targeted user and item embedding representations and improving the accuracy of recommendations.

[0054] Step 5: High-order information propagation and node representation update. In the multimodal data information recommendation method based on graph neural network of the present invention, in order to further improve the recommendation accuracy, the present invention transmits high-order neighborhood information layer by layer through the multi-layer graph neural network GNN, so that the feature representation of users and items not only includes the node information of direct interaction, but also captures the interaction relationship from higher order, forming a richer embedded representation. In the constructed user-item bipartite graph, the initial features of each user node u and item node v are updated in the first layer of the graph neural network through the information of their direct neighbor nodes;

[0055] The representation update formula of the user node is:

[0056] The update formula of the item node is:

[0057] By adopting the mean aggregation method, the node representation of users and items will gradually integrate the feature information of their neighbor nodes in each layer of the graph neural network. In the first layer, the representation of users only includes the features of items with which they have direct interactions. However, as the number of graph neural network layers increases, the representation of users can not only include the information of directly interacting items, but also gradually integrate the information of other users connected to these items, thus realizing the propagation of high-order information. Similarly, the embedded representation of items is also continuously transmitted through user nodes to obtain the information of indirectly interacting users, thereby forming a richer feature representation of items. As the number of propagation layers l increases, user and item nodes can gradually capture second-order, third-order and even higher-order neighbor information.

[0058] For example, through a two-layer graph neural network, users can not only capture the features of items with which they directly interact, but also further capture the information of other users associated with them through these item nodes, thereby effectively expanding the expressive power of user embedding representations. Similarly, the features of item nodes will also obtain more user behavior features in high-order propagation, increasing the richness of item features. Through the superposition of multiple layers of graph neural networks, the representation update process of users and items realizes the gradual aggregation of high-order neighbor information, and obtains deeper relationship information layer by layer.

[0059] Step 6: Calculation and generation of recommendation scores. In the previous steps, the features of users and items have been integrated through multimodal feature fusion and high-order information propagation to generate the final embedded representation containing rich information, which are: (the final embedding representation of the user) and (The final embedded representation of the item). This invention uses the vector inner product method to perform inner product operations on the embedded representations of the user and the item to generate a recommendation score. This score indicates the interest intensity of user u in item v. The formula for calculating the recommendation score is:

[0060] The result of the inner product is interpreted as the similarity in the feature space of the user and the item. The larger the inner product value, the higher the match between the user and the item, and the greater the user's interest in the item. This calculation method generates the user's score for each item. Then, the items are sorted according to their recommendation scores, and the items with the highest scores are selected as recommendation results and presented to the user.

[0061] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0062] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A multimodal data information recommendation method based on graph neural network, characterized in that: The following steps are involved: Step 1: Obtain the interaction data between users and items, and construct a bipartite graph of users and items, where users and items are nodes in the graph, and the interaction behaviors between users and items are edges in the graph. The initial features of the user and item nodes are updated through neighborhood aggregation through the graph neural network to generate embedded representations of users and items; in the l+1th graph neural network, the embedding representation update formula of user u is: Among them, σ is the activation function, W (l) is the weight matrix trained on the lth layer, AGG is the aggregation function used to aggregate neighbor node features, Nbr(u) is the set of neighbor items of user u, is the embedding representation of user u at layer l+1; Step 2: Get the user's comment text and the item description text, use the pre-trained language model BERT to encode the comment text and description text, and generate the text feature representation of the user and item. The text feature representation is as follows: in, is the text feature of user u, is the text feature of item v, T u and T v are the comment text of user u and the description text of item v respectively; Step 3: Obtain the image data of the object, use the convolutional neural network (CNN) to process the image data of the object, and generate the visual feature representation of the object. The feature representation of the object image is: in, is the image feature of item v, l v is the image data of item v; Step 4: Fuse the multimodal features of user and item interaction features, text features, and image features, assign weights to each modality through dynamic attention, and calculate the final feature representation of users and items. The specific steps are as follows: Among them, αm and βm are adaptive modal weights, h u and h v is the mth modal feature of users and items; Step 5: High-order information propagation is performed through a multi-layer graph neural network. Each layer of the network updates the representation of users and items to capture the high-order relationship between users and items. The update formula for user node representation is: The update formula for item node representation is: in, is the embedding of item v at the l+1th layer, Nbr(u) and Nbr(v) represent the neighbor sets of user u and item v respectively. and is the training weight matrix; Step 6: Calculation and generation of recommendation scores: Based on the multimodal feature representation of users and items, the recommendation scores are generated by calculating the inner product of the user and item representations: in, is the recommendation score of user u for item v, and The final embedding representation of users and items.

2. According to claim 1, a data information recommendation method based on graph neural network is characterized in that: In step one, the initial features of the user node and the item node are obtained through an embedding layer, and the embedding layer generates their initial vector representations based on the historical behavior data of the user and the item, and maps the user's historical interaction behavior, click records or rating data through an embedding matrix, so that each user and item has a vector representation of a fixed dimension as the input of the graph neural network.

3. According to claim 1, a data information recommendation method based on graph neural network is characterized in that: The aggregation function of step 1 adopts the mean aggregation method, that is, the representation of the user or item node is updated by calculating the mean of the representation of its neighboring nodes. For each user node, the features of its direct neighboring item nodes are aggregated, and a new user node representation is generated by weighted average or simple mean calculation of the features of these item nodes; similarly, the feature representation of the item node is updated by calculating the feature mean of the neighboring user nodes.

4. According to claim 1, a data information recommendation method based on graph neural network is characterized in that: In the step 2, the language model BERT is a pre-trained bidirectional transformer model that captures contextual information in comments or description texts and generates high-dimensional semantic embeddings. BERT obtains contextual dependencies in sentences through its bidirectional attention, maps each word in user comments or item descriptions to a vector in a high-dimensional space, and processes the vectors through a multi-layer neural network to generate a fixed-length high-dimensional vector representation related to the original text.

5. The data information recommendation method based on graph neural network according to claim 1 is characterized in that: In the step three, the convolutional neural network includes multiple convolutional layers and pooling layers, which are used to extract local and global features of the item image and generate a high-dimensional image embedding representation. The image input of the item captures local features through multi-layer convolution operations. The operation of each layer of convolution kernel can extract low-level features of edges and textures in the item image. After passing through multiple convolutional layers, high-level semantic features of the item image are gradually extracted; the pooling layer is used for downsampling, reducing the dimension of the feature map, retaining the main features, and reducing the computational complexity; the convolutional features are mapped to a vector space of fixed dimensionality through a fully connected layer to generate an image feature embedding of the item.

6. The data information recommendation method based on graph neural network according to claim 1 is characterized in that: In the step 4, the attention network is composed of a multi-layer perceptron, and weights are assigned according to the feature importance of each modality. For the multi-modal feature representation of each user and item, the attention network first initializes the weights of the features of different modalities, and then calculates the importance of each modality to the current recommendation task through the multi-layer perceptron, and outputs the attention score of the modality; The attention scores are then normalized using the softmax operation to generate the final weights for each modality; the final user or item features are represented as the weighted sum of the features of each modality.

7. The data information recommendation method based on graph neural network according to claim 1 is characterized in that: In the step five, the high-order information propagation step is implemented by stacking multiple layers of graph neural networks. Each layer of the network updates the feature representation of the user and item nodes once, and gradually integrates the high-order information from the neighboring nodes. Through the iterative propagation of the multi-layer graph neural network, the user node not only updates the item node features from its direct neighbors, but also obtains the influence of indirect neighbor nodes through multi-layer superposition, that is, gradually aggregates information through second-order, third-order and higher-order neighbor nodes; each layer of the graph neural network performs neighborhood aggregation and activation operations on the node features updated in the previous layer.

8. The data information recommendation method based on graph neural network according to claim 1 is characterized in that: In step 6, the recommendation score is calculated based on the inner product of the final feature representation of the user and the item. The higher the recommendation score, the more the recommended item meets the user's interest preference. The final feature representation of the item Inner product operation to generate the user's recommendation score for the item The higher the score, the more the item matches the user's interests. All items are sorted according to their scores, and the item with the highest score is selected as the recommendation result.

Images