The deep separable convolution with DSC NCF model and optimization mechanism of digital economy for intelligent manufacturing under sales order recommendation algorithm

Design of intelligent recommendation platform based on deep learning

In the context of the digital economy, the design of a sales order management recommendation platform for smart manufacturing enterprises must fully consider real-time data processing, dynamic model updates, and personalized user experiences¹⁴. To address this, this study constructs a multi-module collaborative deep learning recommendation platform architecture to meet the complex and dynamic business needs of smart manufacturing enterprises. These requirements are reflected in three aspects: real-time demand adaptation, multi-modal data fusion, and personalized expansion capabilities, specifically as follows: (1) The need to quickly respond to customer order fluctuations and market dynamics, and achieve second-level strategy adjustments through a streaming computing engine; (2) The need to closely integrate various data sources such as real-time production scheduling, multi-level inventory status, and logistics constraints; (3) While maintaining low recommendation response latency, attention should also be paid to differentiated scenarios such as discrete manufacturing and batch customization. These challenges stem from the complexity of dynamically matching production resources with market demands. Based on this, the structure of the intelligent recommendation platform structure based on deep learning is shown in Fig. 1.

Figure 1 shows that the intelligent recommendation platform architecture is built based on a deep learning integration framework. The online computing module uses the deep separable convolution layer of the DSC-NCF algorithm to extract user-item interaction features in real time and dynamically update the recommendation model, while the offline training module embeds historical behavior data through Item2Vec and trains a multi-modal fusion model by combining production plan data such as inventory status and equipment load rates. In the product recommendation process, user behavior data is transmitted in real time to the online module through the log collection module to generate an initial recommendation list, and then the business logic module filters and sorts the results based on enterprise rules such as inventory thresholds and delivery cycles to ensure compatibility with production strategies. The intelligent features of this mechanism are reflected in three aspects: dynamic adaptability enables the system to respond to the supply chain in real time and dynamically adjust recommendation priorities; multi-modal decision-making optimizes order commitment by fusing multi-dimensional data such as user behavior, inventory status, and equipment loads. Rule compatibility embeds production constraints such as safety stock through the business logic module to effectively avoid the risk of over-selling goods.

The web display module is the interface where the recommendation platform directly interacts with users. It is primarily responsible for receiving user input requests and displaying recommendation results¹⁵. Through a user-friendly interface design, it collects user behavior data such as browsing, clicking, searching, and purchasing. These data are then transmitted in real-time to the log collection module¹⁶. The log collection module is the core of data collection for the platform. It receives user behavior data from the web display module. After preliminary processing, the data is stored in a distributed database for subsequent analysis and processing^17,18. The online computation module is the key component for achieving real-time recommendations. This module uses real-time user behavior data and product attribute information to dynamically calculate user profiles and product similarities through deep learning models¹⁹. Based on these calculations, the recommendation model is updated in real-time, providing users with immediate and personalized recommendations²⁰. The offline computation module focuses on processing large-scale historical datasets to optimize and train more in-depth and accurate recommendation models. By batch-processing historical data, it uncovers hidden patterns and trends, thereby optimizing the structure and parameters of the recommendation model²¹. The business logic module is the core decision-making unit of the recommendation platform. It coordinates data flow and task scheduling among all modules. It receives personalized recommendation requests from the web display module, calls real-time recommendation results from the online computation module, and combines optimized models from the offline computation module to generate the final recommendation list^22,23. Through the collaborative work of these modules, the deep learning-based sales order management recommendation platform for smart manufacturing enterprises achieves efficient data processing, dynamic model updates, and personalized recommendation services.

User behavior feature extraction based on Item2Vec word embedding model

In intelligent recommendation systems, user behavior analysis serves as the core to understand user preferences and optimize recommendation results. User behavior refers to the interactive actions of users on the platform, including but not limited to product browsing, clicking, searching, purchasing, rating, and returning. These behavioral data are recorded through fields such as timestamps, operation types, and contextual information, forming user behavior sequences. The purposes of analyzing user behavior are threefold: (1) mining explicit needs and implicit preferences through behavior sequences; (2) capturing the temporal changes in user interests to support real-time recommendation updates; (3) providing initial recommendations for new users based on group behavior patterns.

In the recommendation process of this study, user behavior feature extraction is a key step connecting raw data and recommendation algorithms. Based on this, this section adopts the Item2Vec model as the user behavior feature extraction method. This is because the Item2Vec model can effectively address the deficiency of traditional methods that rely on user-item rating matrices and struggle with data sparsity by modeling behavior sequences. Meanwhile, it can enhance the temporal modeling capability of dynamic sequences while maintaining low computational costs. Additionally, the Item2Vec model can efficiently learn the contextual semantic relationships between items and convert user behavior sequences into low-dimensional semantic vectors, which is suitable for the continuity characteristics of user behavior in intelligent manufacturing scenarios and provides good industrial scalability for multi-modal recommendation optimization.

Principle of Item2Vec model

Item2Vec is an embedding model inspired by Word2Vec. It transforms user behavior sequences into item vectors, capturing similarities between items²⁴. This model is particularly suitable for recommendation systems. It effectively extracts user behavior features for personalized recommendations. The structure of the Item2Vec model is shown in Fig. 2.

The core idea of Item2Vec is to treat user behavior sequences as “sentences,” where each item is similar to a “word” in a sentence. In this way, the model learns the contextual relationships between items, generating low-dimensional vector representations of items^25,26.

Similar to Word2Vec, Item2Vec has two main training methods: Continuous Bag-of-Words (CBOW) and Skip-Gram. Skip-Gram is more commonly used in recommendation systems because it can better capture the local dependencies between items.

In the Skip-Gram model, the goal is to predict other items in its context through the central item^27,28. Specifically, for a given item sequence $\:{w}_{1},{w}_{2},\dots\:,{w}_{K}$, the goal of the model is to maximize the following objective function:

$$\:\frac{1}{K}\sum\:_{i=1}^{K}\:\sum\:_{j\ne\:i}^{K}\:\text{l}\text{o}\text{g}p\left({w}_{j}\right|{w}_{i})$$

(1)

$\:p\left({w}_{j}\right|{w}_{i})$ indicates the probability that the item $\:{w}_{j}$ appears under the condition of the central item $\:{w}_{i}$. This probability can be calculated by the following equation:

$$\:p\left({w}_{j}\right|{w}_{i})=\sigma\:({u}_{i}^{T}{v}_{j}\left)\prod\:_{k=1}^{N}\:\sigma\:\right(-{u}_{i}^{T}{v}_{k}) p$$

(2)

$\:{u}_{i}$ and $\:{v}_{j}$ are vector representations of the central item $\:{w}_{i}$ and the contextual item $\:{w}_{j}$, respectively. $\:\sigma\:\left(x\right)=\frac{1}{1+{e}^{-x}}$ is a Sigmoid function, and N is the number of negative samples.

The purpose of negative sampling is to reduce the computational complexity²⁹. Specifically, for each positive sample, the model randomly samples N negative samples from the thesaurus, so that the distribution of negative samples satisfies the following formula:

$$\:P\left(w\right)=\frac{\text{c}\text{o}\text{u}\text{n}\text{t}\text{e}\text{r}(w{)}^{0.75}}{{\sum\:}_{u\in\:W}\text{c}\text{o}\text{u}\text{n}\text{t}\text{e}\text{r}(u{)}^{0.75}}$$

(3)

Generation of item vector and feature extraction

In Item2Vec, each item $\:{w}_{i}$ has two vectors: the head word vector $\:{u}_{i}$ and the context vector $\:{v}_{i}$. The final item vector can be generated in the following ways:

(1)

Directly use the headword vector: take $\:{u}_{i}$ as the feature vector of article $\:{w}_{i}$.
(2)

Vector addition: add $\:{u}_{i}$ and $\:{v}_{i}$, that is, $\:{u}_{i}$+$\:{v}_{i}$.
(3)

Vector splicing: splice $\:{u}_{i}$ and $\:{v}_{i}$ into a longer vector, namely $\:[{u}_{i}^{T},{v}_{i}^{T}{]}^{T}$.

In the sales order management of smart manufacturing enterprises, user behavior sequences (such as browsing, purchasing, and searching) can be transformed into item sequences and input into the Item2Vec model³⁰. Through training, the model learns the similarities between items and generates low-dimensional vector representations of items. These vectors can serve as features of user behavior for personalized recommendations in the recommendation system.

Based on this, this study chooses to directly use the central word vector as the final feature representation of items for the following reasons: The central word vector $\:{u}_{i}$ has certain advantages in computational efficiency and model simplification. It can directly capture the core semantic features of items in user behavior sequences, avoiding the extra computational costs caused by vector addition or concatenation, making it more suitable for large-scale real-time recommendation scenarios, and particularly suitable for the order management needs of intelligent manufacturing enterprises. In addition, the independence of the central word vector gives it good compatibility with the DSC-NCF algorithm. The deep separable convolution can effectively extract the interaction features between users and items, while avoiding the interference of redundant information on the local feature focus of the convolution layer.

For example, for the item $\:{w}_{i}$ purchased by the user, the model can calculate its similarity with all other items, thus recommending items similar to $\:{w}_{i}$. Similarity is usually calculated by cosine similarity:

$$\:\text{similarity}({w}_{i},{w}_{j})=\frac{{u}_{i}\cdot\:{u}_{j}}{\parallel\:{u}_{i}\parallel\:\parallel\:{u}_{j}\parallel\:}$$

(4)

In this way, Item2Vec not only captures user behavior preferences but also provides rich semantic information for the recommendation system, improving recommendation accuracy and personalization.

In summary, the user behavior features extraction method based on Item2Vec effectively utilizes user behavior sequences to generate semantically meaningful item vectors. This provides strong technical support for sales order management in smart manufacturing enterprises³¹.

NCF intelligent recommendation algorithm based on depth separable convolution

In intelligent recommendation systems, NCF is a deep learning-based recommendation algorithm. It can learn complex nonlinear relationships between users and items³². However, traditional NCF models face high computational complexity when handling large-scale data. To improve model efficiency and performance, this study proposes an NCF intelligent recommendation algorithm combined with depth wise separable convolution, as shown in Fig. 3.

In the NCF model, user and item features are represented in low dimensions through the embedding layer³³. However, this representation may lose some important local features. By introducing depth wise separable convolution, the model can extract interaction features between users and items more finely³⁴. Specifically, depth wise separable convolution can be applied to the embedding feature maps of users and items. It captures local features through channel-wise convolution and then fuses features through point-wise convolution. This operation not only preserves the semantic information of the original features but also uncovers deeper interaction features.

In the multi-layer perception (MLP) part of NCF, traditional fully connected layers require significant computational resources. By introducing depth wise separable convolution, some fully connected layers can be replaced with depth wise separable convolutional layers. This reduces the computational complexity of the model^35,36,37.

For example, assume that the input feature dimension is D and the output dimension is D′, the traditional fully connected layer needs D×D′ parameters. However, the deep separable convolution only needs D + D×D′ parameters, which significantly reduces the parameters of the model.

In addition, to further enhance the semantic understanding capability of the recommendation system, this study introduces Product Reviews (PR) as auxiliary input to the model. Product reviews contain fine-grained user feedback on products (such as functional evaluations, usage experiences, etc.), which can compensate for the sparsity of user behavior data (such as clicks, purchases) and capture users’ implicit preferences. The encoding process of product reviews is as follows:

i.

Text preprocessing: Perform operations such as word segmentation, stop word removal, and standardization on reviews to extract key semantic information.
ii.

Word vector representation: Use a pre-trained BERT model to map each review into a fixed-dimensional semantic vector. BERT’s context-aware capability can effectively capture emotional tendencies and fine-grained features in reviews.
iii.

Feature fusion: Concatenate the review vector generated by BERT with the Item Embedding (IE) to form an enhanced item feature representation:

$$\:{\text{v}}_{\text{i}\text{t}\text{e}\text{m}}^{{\prime\:}}=\text{C}\text{o}\text{n}\text{c}\text{a}\text{t}({\text{v}}_{\text{i}\text{t}\text{e}\text{m}},{\text{v}}_{\text{r}\text{e}\text{v}\text{i}\text{e}\text{w}})$$

(5)

$\:{\text{v}}_{\text{i}\text{t}\text{e}\text{m}}$ represents the original embedding vector of the item, and $\:{\text{v}}_{\text{r}\text{e}\text{v}\text{i}\text{e}\text{w}}$ represents the semantic vector of the comment.

By fusing product review information, DSC-NCF algorithm can more comprehensively model the interaction between users and items, thus improving the accuracy and interpretability of recommendation results.

Source link