1. Introduction

In the era of information overload, users and commodities in e-commerce platforms are different nodes, and the correlation relationship between these nodes and nodes jointly forms a complex information network, which is a complex network ^[1]. In the massive data, although the recommendation system can be used to recommend suitable products for users, how to combine deep learning and other new technologies such as deep learning to improve the accuracy of personalized product recommendation is a key problem. For example, Amazon’s deployment of a recommendation system, which can greatly increase page views and conversion rates. Through personalized recommendation algorithm, it can not only recommend appropriate products for users, but also enhance the data set for the research of recommendation algorithm ^{[2, 3]}. The e-commerce platform is now able to create more precise user profiles, enhancing the effectiveness of recommendation strategies and facilitating more accurate suggestions ^[4]. Nevertheless, the swift evolution of the Internet and new media, coupled with the generation of vast amounts of information, has led to an explosion of user demands, prompting extensive research and application of recommendation systems ^{[5, 6]}. As these systems play a crucial role in advancing the new media landscape, an increasing number of individuals from diverse fields, each with distinct professional backgrounds and industry expertise, have engaged in comprehensive studies on this topic ^[7]. However, within these investigations, the sheer volume and diversity of data, as well as the complexities of data processing, have led to challenges in finding solutions quickly. Additionally, the recommendation system itself poses a significant concern: how to efficiently manage and organize a large volume of meaningful and relevant information ^{[8, 9]}. Recommendation system faces many challenges, so the recommendation system has attracted high attention from many people. One of the main reasons is that they believe that recommendation systems can provide people with “solutions,” a more efficient tool to help others solve problems ^[10].

There is a lot of research work on the recommendation system, among which the two main technologies are more representative, namely, clustering algorithm and deep learning. The clustering algorithm classifies the users by analyzing their characteristics, so as to achieve better classification results for the recommendation system. Deep learning, on the other hand, uses its powerful computing power to process user data to obtain better recommended results. Deep learning-based methods are also widely used in recommendations ^[11]. The associations between nodes is dynamically changing and heterogeneous, forming a heterogeneous information network. Heterogeneous information network contains more node features and rich semantic features of correlation relations, which is more suitable for application in user-product recommendation. At the same time, some recommended research works are also carried out from the perspective of network embedding, proposing Deepwalk model and node2vec model, which aim to obtain information about the neighborhood of nodes ^{[12, 13]}. To better characterize the graph heterogeneity, a new algorithm based on the meta-path random walk is proposed to better mine the heterogeneity and improve the semantic correlation in the graph. The establishment of intelligent college English writing training system helps to realize the role change of teachers and students. Constructivism theory believes that under certain learning situations, learners use necessary learning materials to acquire knowledge through meaning construction. Therefore, the theory emphasizes the student-centered and requires students to change from the object of knowledge to the active builder of knowledge ^{[14, 15]}.

Teachers should focus not only on delivering writing instruction but also on emphasizing the process of students’ English writing. Consequently, their roles have evolved from traditional knowledge providers to designers and facilitators of the writing process. Their primary objective is to foster creativity among students during writing activities. Simultaneously, students transition from being passive recipients of knowledge to active participants in the writing process ^[16]. They engage in completing writing assignments and actively revise their compositions based on received feedback. Through this iterative process of writing, students address their weaknesses and continually enhance their English writing skills. Furthermore, the English writing training system serves not only as a supplementary tool for instruction but also as a vital component of the overall writing teaching framework. It enriches and extends classroom learning, boosts teachers’ efficiency, and cultivates greater interest in English writing among students. The intelligent college English writing training system enhances traditional teaching methods by creating a multi-dimensional, interactive learning environment. It supplements classroom instruction, addressing its limitations and improving overall writing teaching activities ^[17]. The online mode boosts students’ motivation, writing efficiency, and interest in writing. The interactive teaching design breaks the monotony of traditional learning and fosters collaborative group work. Additionally, the system tracks students’ time spent on tasks, their performance, and writing progress, providing valuable data for language teaching research. The analysis and statistics features enable teachers to quickly assess and compare students’ writing abilities, facilitating more efficient and targeted teaching approaches ^{[18, 19]}.

2. A Multi-Interaction Feature Fusion Approach for Heterogeneous Information Networks with Multi-class Behavioral Feature Fusion

3. Recommended Methods for Time-Aware Feature Fusion Based on Collaborative Filtering

3.1. An Embedded Representation of the Absolute and Relative Time of the CNN Fusion

Above proposed absolute time embedded and relative time is based on the current interaction feature fusion algorithm of time information, can effectively record for the interaction between users and items and modeling of feature fusion, and also can be used to dig between users and items or the interaction between users and other users of effective information, so as to explore more user recommended rule ^{[20, 21]}. At the same time using the sinusoid function of the time information conversion, to the hidden layer of the operation of coding, in the process of coding users and other users or goods interaction will periodically transform, introduced the f parameter to adjust the vector, the last time embedded ER is through the hidden layer vector layer by layer accumulated results, the embedding also represents the time embedding matrix of user interaction ^{[22, 23]}. Time perception and feature fusion refers to the transformation of observing things in temporal order. Usually, the historical value of the same variable is used to predict the future value, or some predictors can be added to predict the future value, so as to solve the problem of sparse data in multi-interaction feature fusion. First, to obtain the complex temporal dependence between items in the fusion of user behavior characteristics ^{[24, 25]}. In heterogeneous information networks, meta-path-based recommendation algorithms are also widely used. The definition of meta-path is that one can link two different objects or the same kind of objects on a complex network. There are two main problems we need to solve at present. First, the traditional network modeling method only extracts some information in the existing interactive system ^[26]. Fig. 1 for the English writing quality assessment algorithm flow chart, at the same time, most of the information can be obtained from the heterogeneous information network recommended method research also mostly only consider how to use the user’s static information modeling, thus ignoring how to use some dynamic changes of the user behavior for interaction features fusion. Secondly, in recommending users, it is crucial to effectively model user behavior preferences by considering various behavioral characteristics derived from extensive data. This involves making recommendations based on the potential relationships between these behaviors, integrating the historical characteristics of users or items. Initially, the study introduces the concept of community detection, which examines the connectivity of object nodes and the density of nodes associated with users or items. The structure of these nodes within the detected communities is primarily geared towards data mining, allowing for clustering based on similar user traits or product characteristics. This approach aids in a deeper analysis of the complex network, highlighting its unique properties and inherent functions.

Fig. 1. Flow chart of the English writing quality assessment algorithm.

Fig. 2. Learner feedback loop optimization algorithm.

Those who have a large number of the same interests or find common friends or common concern accounts in each community may be members of the same community. In view of the shortcomings of the above recommended methods based on homogeneous information networks, this chapter studies the multi-interaction feature fusion method of heterogeneous information networks. Some of the above methods are limited to random walk between specific types of nodes, does not consider the semantic correlation between two objects or multiple objects, and the method proposed in this chapter not only combines the heterogeneity of user behavior in heterogeneous information network, thus further improve the effectiveness and accuracy of recommendation. Fig. 2 for learner feedback loop optimization algorithm diagram, for the above accuracy is low, to address the issues of low efficiency and significant computational challenges, this paper proposes a method for fusing multiple interaction features within heterogeneous information networks. Unlike existing recommendation approaches, this method employs an end-to-end model that captures interactions among multiple users and products, leveraging the heterogeneity of user behavior and the temporal aspects of their interactions. By analyzing the behavioral characteristics of different users, the study introduces a graph convolutional neural network and presents a heterogeneous information network recommendation model called MBFF. This model first assesses user behavior characteristics, then embeds these features through calculations, followed by a fusion process. Ultimately, this approach aims to enhance both the accuracy and efficiency of the results obtained from fusing multiple interaction features.

3.2. Writing Skills Training

This paper presents a recommendation approach for heterogeneous information networks by integrating the node information of both users and products. It encompasses user nodes and product nodes, effectively transforming loosely connected product characteristics into a cohesive user-product network. The process includes iterative updates of embeddings to enhance the model’s effectiveness. The model aims to recommend preferred products based on available user and product information, as well as their associative relationships with other items. By utilizing a fusion method based on graph convolutional neural network embeddings, the model facilitates feature extraction for homogeneous user behavior. Fig. 3 illustrates the architecture of the multi-interaction feature fusion model. Initially, user and product nodes within the network are identified, followed by the embedding of these nodes in the user-product network. Subsequently, graph neural network convolutions are performed to assess preferences within the convolution layer, which are then pooled to mitigate overfitting. Overfitting occurs when machine learning models become overly complex, closely aligning with error values on certain datasets while failing to generalize effectively to real-world or test data, resulting in increased error rates and diminished model performance on data beyond the training set. However, when the model is applied to the real data or model and test data, the error value will become higher, and the generalization ability of the model will become worse, and other data besides training cannot be expressed. The above process is called overfitting phenomenon. In the recommended model for heterogeneous information network.

Over-smoothing typically results in a high degree of similarity among node embeddings. The use of multi-layer Graph Convolutional Networks (GCNs) can cause the output features of each node to become excessively smooth, making it challenging to differentiate between distinct nodes, particularly in cases with small sample sizes. The GCN architecture is notable for its effective Laplacian smoothing performance and strong learning representation capabilities. In essence, GCNs represent a specific form of Laplacian smoothing, which functions like a low-frequency filter. This allows the characteristics of central nodes in the network to diffuse, causing the features of neighboring user nodes to resemble those of the central node, thus benefiting recommendation algorithms. When a user or product node has multiple neighboring nodes, it can establish self-connections. Node degree serves as a critical measurement in this context, where the degree is considered a computational unit. Fig. 4 for real-time error correction function use frequency and accuracy assessment, the performance of the user preferences of goods, that is to say, a user u for a commodity interaction behavior less, often more able to reflect the user of the product, it borrowed the idea of collaborative filtering, can better reflect the user’s preference for goods. After the linear convolution layer is formed, if the nodes calculated by high-order aggregation are directly used, then the phenomenon of overfitting will occur, which will directly affect the training effect of the recommended model. Therefore, we need to pool the nodes for high-order aggregation.

Fig. 3. Architecture diagram of the multi-interaction feature fusion model.

Fig. 4. Frequency and accuracy assessment diagram of real-time error correction function.

Fig. 5. Assessment chart of the impact of multimodal input on writing creative inspiration.

Following convolution, data pooling is used to streamline computations between nodes, enhancing processing speed and robustness of the extracted features. Unlike maximum pooling, average pooling reduces estimation variance from limited neighborhood sizes, minimizing errors in higher-order aggregation calculations. The pooling layer acts as a feature selection and information filtering process, where some data loss is a trade-off for improved computational efficiency, as faster processing tolerates fewer errors. After applying linear convolution and average pooling to node behavior characteristics, the user memory network generates user embeddings. These embeddings are fused with product information through iterative read/write operations to select the most relevant products for recommendations. A graph attention mechanism assigns weights to various behavioral features in heterogeneous networks, allowing for the fusion of homogeneous behavior characteristics. This results in a final node representation after the calculations. The memory network and external storage unit are introduced to store and selectively read user interaction data. This unit updates the user’s behavioral feature fusion, generating an embedded user representation. After processing the embedded data, product embeddings are used to predict the user’s next behavior and recommend products. The dataset is initialized, and target and adjacent nodes are extracted from the heterogeneous network as input to the memory network. After updating the nodes, a model predicts the recommended score between users and product nodes, outputting the results.

4. Research on the Application of NLP Algorithm Based on Multi-interaction Feature Fusion in English Writing Teaching

In traditional recommendation algorithms, user preferences are often derived from static behaviors, neglecting the current or future preference levels for certain items. This limitation leads to less effective recommendations. The feature fusion algorithm addresses these issues by integrating users’ long-term historical interests with their current preferences, capturing more detailed user features. By incorporating temporal information, feature fusion algorithms account for time-sensitive user behaviors and contextual semantics. For example, some algorithms integrate specific temporal data, such as timestamps, directly into the model. However, while these methods enhance recommendations, they often fall short of ideal performance, lacking precision in suggesting preferred items and increasing computational complexity. As a result, this approach was abandoned. This chapter takes an alternative approach by analyzing multi-class interaction behaviors between users and their neighbor nodes, combined with heterogeneous information networks, to offer time-based recommendations. The emphasis is placed on crucial temporal information within user interaction records, taking into account factors like current item preferences and temporal modeling. This strategy highlights the significance of real-time modeling, which captures the dynamic relationship between users and time, ultimately leading to a more accurate and personalized recommendation system. By acknowledging the evolving nature of user behavior, the algorithm enhances its accuracy in suggesting items that resonate with users’ shifting interests.

Fig. 6. Student feedback response time and satisfaction assessment chart.

However, when different users in similar contexts receive almost identical recommendations, it can restrict the individuality of suggestions. This phenomenon can diminish the accuracy of product recommendations, resulting in a lower degree of match between user needs and suggestions. Often, the fusion of user characteristics with temporal interaction data is overlooked, limiting the ability to fully leverage critical time-related behavioral insights. This gap prevents the modeling of long-term preferences based on the relative timing of user interactions and hinders the exploration of relationships between users and items, ultimately impacting the predictive accuracy of the recommendation algorithm regarding users’ future preferences.

The feature fusion algorithm comprises an embedding layer, a self-attention mechanism layer, a feed-forward neural network, and a dynamic fusion network. The model begins by utilizing collaborative filtering techniques to extract temporal information from user interaction records alongside relevant product data for recommendations. Fig. 6 for student’s feedback response time and satisfaction evaluation diagram, and then further in the user interaction record embedded time information and article features and embedded fusion and self-attention mechanism and feed forward neural network for learning, finally the user preferences dynamic fusion, and recommended for the user in the candidate items. The core problem of recommendation algorithms is to understand the user’s preferences and capture their users’ hobbies. According to the change of users’ needs, a large number of recommendation methods according to the information recorded by user interaction are proposed, and such methods can more accurately recommend users. Table 1 has showed the manual annotation function and icon for teachers.

The basic operation of its conventional multi-interaction feature fusion is to convert the user’s historical feature fusion from left to right into a vector in the recommendation model, and to recommend it based on this one-way vector from left to right. Due to the aforementioned issues, multi-interaction feature fusion methods often consider the extraction of user interaction records as an ideal scenario. These methods assume that the interaction records between users and items can fully encapsulate both temporal and item-specific information, overlooking various limitations. For instance, this assumption may result in historical features being inadequately integrated, preventing the complete embedding of users’ specific time-related information into the item data. Additionally, it becomes challenging to capture detailed temporal characteristics of interactions between users and items, as well as to extract valuable insights from the state of feature fusion.

To address these challenges, this paper introduces the concepts of relative time embedding and absolute time embedding within the time embedding layer. The objective is to make the high-order dynamic information in user interaction records more manageable and straightforward, thereby enhancing the accuracy of item recommendations for users. In this context, time information is denoted as T, which encompasses both the timestamps of user interactions and the intervals between these interactions. This paper highlights that the feature fusion representing time and the historical feature fusion associated with the user exhibits a one-to-one correspondence with the timestamps. Thus, time embedding is categorized into two types: absolute time embedding and relative time embedding, providing a more nuanced understanding of temporal data in user interactions. Table 2 for MBFF model based on Jingdong data set comparison experiment, the absolute time coding matrix EA can extract interaction with the user items contains the time information, the matrix received the time information is instant, but also can detect the user on which timestamp specific interact with items. Compared with the relative time embedding, the absolute time embedding can be automatically updated. The accuracy of relative time is usually not as high as that of absolute time. Relative time is a concept of time that can be compressed, which usually represents a general range. The main purpose of relative time embedding is also to solve the difficulties that users are prone to appear in the process of interaction. Relative time embedding ER represents the encoding of different

Fig. 7. Student feedback response time and satisfaction assessment chart.

user interaction pairs with different time intervals, and relative time embedding can observe the last behavioral feature fusion of users, so as to better recommend the next time according to the historical characteristics of their users.

In the time-embedded layer, the fusion of absolute and relative time with user interaction records helps uncover complex behavior patterns. As users interact more with features, their behavior becomes increasingly time-dependent, with shorter intervals between interactions as the relationship with items strengthens. The introduction of an attention mechanism effectively addresses challenges in multi-feature fusion, improving the interpretability and accuracy of user recommendations. By combining relative and absolute time embeddings with attention, the system enhances recommendation precision and provides clearer insights into user behavior.

Fig. 7 shows the evaluation chart of students’ feedback response time and satisfaction. Meanwhile, the attention mechanism has been fully utilized in many fields, such as statistical learning, language recognition technology, image processing, natural language processing and other aspects. The attention mechanism has significantly advanced the effectiveness of various training models, such as the Transformer architecture. The Transformer model leverages the attention mechanism to enhance training speed, although its inherent complexity is quite high. This model is capable of translating content from one language to another, utilizing the attention mechanism to facilitate this process. The application of attention allows for more efficient training and calculations compared to conventional neural networks, demonstrating its superiority in handling complex tasks.

5. Experimental Analysis

The self-attention layer contains three different linear layers, It can convert the input time or item features into a matrix of query, real value and keyword, After performing the transformation of the input matrix, The next step will use related functions such as Soft max functions to calculate the record of user interactions with items, Fig. 8 shows the assessment diagram of the relationship between sentence complexity and text fluency, And to assess how close their users are to the viewed objects, Then the Value matrix in the linear layer will be calculated with the conversion matrix calculated above to weight and sum the results, Thus further obtaining the representation matrix of the user’s interaction record.

Fig. 8. Assessment diagram of the relationship between sentence complexity and text fluency.

Fig. 9. Assessment chart of vocabulary richness and writing score.

This paper utilizes an advanced attention mechanism that outperforms conventional attention mechanisms in quickly and effectively identifying the interaction timestamps between users and products. Fig. 9 illustrates the assessment of vocabulary richness and writing scores. By employing this attention mechanism, the model can dynamically and flexibly assign varying weights to user behaviors, allowing for more sophisticated calculations. Additionally, this mechanism facilitates faster computation speeds compared to traditional attention methods, resulting in superior performance. In contrast, conventional attention mechanisms may be constrained by their own structural limitations, which can hinder their ability to recommend relevant features to users.

Select the self-attention mechanism to represent the interaction record between the user and the object, extract the user and the adjacent time interval, and the extracted time information contains different time intervals and time stamps. Fig. 10 is the evaluation diagram of the impact of different interaction features on the writing quality. When using the self-attention mechanism, it is necessary to prepare hidden features for the attention module, the hidden features provided by the article information V in the embedded layer and the time stamp T in the time information of its interaction.

Fig. 10. Assessment plot of the impact of different interaction characteristics on writing quality.

The time embedding introduced in this paper diverges from previous approaches that solely focus on absolute time embedding, neglecting the relative time embedding aspect. The model developed here calculates the embeddings of both users and objects independently, allowing for the retention of specific temporal information, including the time intervals of user interactions with various items. Fig. 11 presents an evaluation diagram of error types associated with the NLP algorithm and highlights the timestamps for different behaviors. In this framework, the embedding layer integrates self-attention, whereas traditional time embedding methods rely solely on absolute time embedding, which only captures the timestamps of user interactions with objects. The time embedding is multiplied by the query and the keyword matrix, and the results are weighted and summed in a relatively similar way to the previous relative time embedding.

Fig. 11. NLP algorithm identifies error type assessment plot.

Fig. 12. Evaluation chart of students’ writing performance improvement.

Then items embedded combined with the attention mechanism for the item of detailed information features representation learning, the user interaction items embedded is also depends on the query matrix and keyword matrix for data processing, Fig. 12 for students writing performance evaluation, and the results of weighted sum. Then use Soft max function to merge the object embedding and time embedding, the results with the real value matrix in the object embedding, finally will get a complete time information and object interaction with the user embedded representation, the embedding will be a representation of the user’s target object learning, and assign appropriate weight.

6. Conclusion

This paper analyzes user interaction behavior with neighboring nodes and the specific time information in user interaction records for recommendation purposes. Traditional recommendation algorithms typically focus only on the user’s attributes, overlooking the characteristics of neighboring and higher-order nodes in the user’s network. To address this, the study introduces a user behavior fusion approach that integrates heterogeneous information networks, enabling a more comprehensive analysis of user data and embedding of homogeneous relationships. This approach results in a global fusion of user behaviors, providing predictive scores based on the unique features of user actions. By exploring advanced user behavior patterns, the method enhances the effectiveness of interaction feature fusion. The intelligent college English writing training system, designed with advanced educational concepts, redefines the roles of teachers and students. It integrates various resources, including computer networks, natural language processing (NLP), UI processing technologies, and automated feedback mechanisms, to improve students’ independent writing skills. This system overcomes the limitations of traditional methods in process management, personalized guidance, and writing material planning. In comparison, students using the NLP-based system saw a 15% improvement in writing performance, while the traditional group improved by only 7%. Both groups initially had a grammatical error rate of 20%, but after the intervention, the NLP group reduced errors to 8%, while the traditional group saw 14%. Furthermore, the NLP group exhibited a 20% increase in unique vocabulary usage, compared to just 10% in the traditional group. Syntax complexity and text coherence also improved more significantly in the NLP group (18% and 22%, respectively) than in the traditional group (9% and 11%).

Incorporating time information from user interaction records, this study contrasts the traditional approach-focused on short-term user interests-with a method that fully utilizes both relative and absolute time. By combining time and item information characteristics with an attention mechanism, the system provides more personalized and accurate recommendations through collaborative item selection.