An industrial app development-oriented crowd-sourcing demand acquisition method

Abstract

The application realizes a crowd-sourcing demand acquisition method for industrial APP development by means of the method in the field of artificial intelligence. Based on the data acquired from the industrial APP platform and the mobile APP market, two models of user demand acquisition method and macro demand prediction method are constructed to realize the crowd-sourcing demand acquisition system. The method provided by the application is directed against the problem that the user demand of the industrial APP and the mobile APP is not consistent and the amount of the industrial APP comment data is small, and combines the technology of transfer learning to pre-train the model through a large amount of mobile APP comment data, and then fine-tune the model using a small amount of industrial APP comment data. In view of the problem that part of the industrial APP has no field classification and the problem that the existing research lacks demand prediction of the market from the macro level, the APP is classified by multi-label through the description information, and the macro demand prediction of the market is carried out according to the field label and the release time of the APP.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method for acquiring collective intelligence requirements for industrial app development. Background Technology

[0002] In recent years, industrial apps have received widespread attention and have been deployed extensively on various industrial internet platforms. However, the development of industrial apps still faces two problems: (1) some industrial apps fail to meet users' actual needs in terms of both functional and non-functional features; and (2) the development of industrial apps is uneven across different application areas, failing to meet the needs of some application areas. How to accurately obtain the real needs for industrial apps has become one of the key issues driving the further development of industrial apps.

[0003] To address the issue of software requirement gathering, a crowdsourcing-based requirement gathering approach has been proposed, aiming to extract requirement information from user feedback and reviews using methods such as crowdsourcing and big data analytics. Existing research on crowdsourcing-based requirement gathering largely focuses on the mobile app field, primarily extracting user needs from mobile app reviews.

[0004] Currently, many research efforts have attempted to address the challenge of intelligent requirements gathering methods for industrial app development. To ensure that apps meet the real needs of the user base, existing research utilizes mobile app review-based requirements gathering methods, including requirements type classification methods, functional requirements gathering methods, and non-functional requirements gathering methods.

[0005] Because industrial app reviews and mobile app reviews have inconsistent data characteristics, and the data volume of industrial app reviews is small, existing research methods are not applicable to obtaining user needs from industrial app reviews. Furthermore, since some industrial apps lack industry categorization and existing research lacks macro-level market demand forecasting, developers cannot understand the market demand in various industrial sectors.

[0006] This invention addresses the issues of poor quality in numerous industrial apps and the uneven development of industrial apps across various sectors. It researches a method for obtaining actual user needs based on user reviews, helping developers update and improve their apps and enhance their quality. Simultaneously, it studies a macro-demand forecasting method for the industrial app market, helping developers predict which industrial sectors will urgently need industrial apps in the future, thereby improving the uneven development of industrial apps across different sectors. Summary of the Invention

[0007] To address this, this invention first proposes a crowdsourced demand acquisition method for industrial app development. Based on user review data obtained from industrial app platforms and mobile app markets, it constructs two models: a user demand acquisition method and a macro-demand prediction method. The user demand acquisition method focuses on users, acquiring demands from their perspective, while the macro-demand acquisition method focuses on industry sectors, acquiring demands from a macro perspective. One model derives key phrases from user reviews, while the other predicts which industrial sectors will urgently need industrial apps in the future based on the number of industrial apps in various industrial sectors over a past period. The latter is used before industrial app development to help developers identify which sectors urgently need industrial apps, while the former is used after industrial app development to help developers iterate and update the industrial app. The resulting crowdsourced demand acquisition system ultimately outputs key phrases representing user demands from the reviews.

[0008] The specific steps of the user demand acquisition method are as follows: First, positive and negative samples are generated. The mobile app reviews are matched with entities from specific online information sources, and matched phrases are selected and labeled as demand phrases. Then, phrases appearing two or more times in each mobile app review are selected and also labeled as demand phrases. These labeled demand phrases are used as positive samples. Next, a portion of continuous word sequences are randomly extracted from the mobile app reviews and used as negative samples. Afterwards, feature extraction is performed. The mobile app review dataset and industrial app reviews obtained through transfer learning are fed into the pre-trained model RoBERTa. RoBERTa encodes the review text. Following BERT, RoBERTa uses Transformer as the feature extractor, enabling… Using the Transformer's multi-head attention mechanism, an attention mechanism map is generated for each positive and negative sample phrase. Then, demand phrase classification is performed. After obtaining the attention mechanism map for each sample, the attention mechanism map is input into a convolutional neural network (CNN) for binary classification training. The CNN classifies phrases into demand phrases or non-demand phrases based on the attention mechanism map of each phrase. After the CNN is trained, industrial APP review data is fed into the pre-trained model for encoding. Then, attention mechanism maps are calculated for each randomly generated word sequence. These attention maps are fed into the CNN classifier for classification to obtain demand phrases. Finally, the obtained demand phrases are grouped according to the APP to which they belong. Demand phrase clustering is performed on the demand phrases of each review under each APP. The core of the clustering is the user's focus.

[0009] The macro demand forecasting method predicts which industrial sectors urgently need industrial apps by using the release time and domain tags of industrial apps. It includes two parts: multi-tag classification and trend forecasting. By selecting a domain classification standard, all industrial apps are classified by domain tags. Industrial apps are classified by multi-tag based on their description information. Then, based on the tags and release time, it predicts which sectors will urgently need industrial apps in the future.

[0010] The positive and negative sample generation method is as follows: unsupervised extraction of requirements from comment data is performed. Starting from the characteristics of the data itself, the requirements phrases are automatically extracted from the text based on their characteristics. Phrases that appear twice or more in a comment are taken as requirements phrases.

[0011] The feature extraction method is as follows: by using a self-attention mechanism, the degree of connection between phrases and other words is mined.

[0012] The proposed demand phrase classification method is as follows: a classifier is trained using positive and negative sample datasets. RoBERTa is used for feature extraction on the dataset. In its default mode, RoBERTa has 12 layers, each with 12 attention heads. For a text with N words, the input data is treated as an image with dimensions N and 144 channels. This transforms the problem of phrase text classification into an image classification problem. Given a multi-channel attention mechanism image, the method determines whether a phrase is a demand phrase. A two-layer convolutional neural network (CNN) model is used to classify the multi-channel attention mechanism image.

[0013] The transfer learning technique involves training the model using mobile app reviews. After the model is fully trained, it is then transferred to an industrial app review dataset for fine-tuning. Given the high similarity between mobile app review data and industrial app review data (in terms of word frequency, text length, text format, and semantics), and the relatively small size of the industrial app review dataset, the original CNN model is not retrained. Instead, the fully connected layers are retrained, and finally, a sigmoid layer is added after the fully connected layers to perform transfer learning on the industrial app review data.

[0014] The clustering method for demand phrases is as follows: unsupervised K-means algorithm is used to cluster demand phrases, and cosine similarity is used to calculate the distance between words.

[0015] The multi-label classification part obtains a large amount of APP description information from crawlers of various industrial internet platforms. The description information includes the industrial fields to which the industrial APPs are applicable. The industrial APPs are classified into different fields based on the APP description information. First, the model is trained on the existing labeled data. Then, it is applied to other unlabeled data. A combination of ALBERT and TextCNN is used for multi-label classification. ALBERT is used as the encoding of the dataset, and TextCNN is used to extract features from the encoded text vectors. Convolutional pooling is performed using different convolution kernels, and then the data is concatenated. To prevent the model from overfitting, some data is discarded during training. Finally, the data is input into a fully connected layer, and the multi-label classification result is obtained through the sigmoid function.

[0016] The trend prediction section performs macro-trend predictions based on the number of domain tags and the release time of industrial apps. Using a monthly time unit, it counts the number of industrial apps in each domain up to the end of each month. Based on the number of industrial apps in each past month, it predicts the number of industrial apps in each domain for the next three months using a multinomial regression equation. The sklearn package in Python generates a multinomial regression equation for each domain. The average monthly growth rate of industrial apps over the past year was 5.8%. After predicting the change in the number of industrial apps for each industrial domain for the next three months, domains with an average growth rate exceeding 10% in the prediction results are selected and considered as industrial domains with strong demand.

[0017] The technical effects to be achieved by this invention are as follows:

[0018] To address the discrepancy between user needs of industrial apps and mobile apps, as well as the limited amount of review data for industrial apps, this paper combines transfer learning techniques. The model is pre-trained using a large amount of mobile app review data, and then fine-tuned using a small amount of industrial app review data. Furthermore, to address the lack of domain classification for some industrial apps and the absence of macro-level market demand forecasting in existing research, this paper performs multi-label classification of apps based on descriptive information and forecasts macro-level market demand based on the app's domain labels and release date. Attached Figure Description

[0019] Figure 1 Overall framework diagram;

[0020] Figure 2 A framework for obtaining requirements based on industrial app reviews;

[0021] Figure 3 Demand forecasting methodology framework;

[0022] Figure 4ALBERT+TextCNN method diagram;

[0023] Figure 5 Trend forecasting methodology framework; Detailed Implementation

[0024] The following are preferred embodiments of the present invention, which are described in conjunction with the accompanying drawings. However, the present invention is not limited to these embodiments.

[0025] This invention proposes a crowdsourcing requirement acquisition method for industrial app development. The overall framework diagram of the crowdsourcing requirement acquisition system is shown below. Figure 1 This invention proposes a demand acquisition method based on comments and a macro-demand forecasting method based on industrial app description information and release time, based on data obtained from mobile app markets and domestic and international industrial app platforms.

[0026] This invention comprises two parts: user demand acquisition and macro-level demand prediction. It targets the application scenario of industrial app development. The user demand acquisition method mines key phrases representing user needs from user comments. The input of this method is user comments in English text format, and the output is the key phrases representing user needs from the comments. In a specific scenario, for an industrial app, the user's comment input might be "If some more example models added, it would be great, or we need to analyze more on this. Really a great implementation." The user demand acquisition method extracts the demand from the user comment, discovering that the user wants more example models added, thus the output is "example models added." Macro-level demand prediction predicts which sectors urgently need industrial apps based on the number of industrial apps developed in various industrial app sectors over a past period. These industrial sectors include "electronic information," "wind power," and "photovoltaic," totaling 17 sectors. Therefore, the output of macro-level demand prediction is a few specific industrial sectors. One part is the key phrases derived from user comments, and the other is the prediction of which industrial sectors will urgently need industrial apps in the future based on the number of industrial apps developed in various industrial sectors over a past period. The latter is used before industrial app development to help developers identify which areas urgently need industrial apps, while the former is used after industrial app development to help developers iterate and update the industrial app.

[0027] 1. Methods for obtaining user requirements

[0028] One current problem with industrial apps is their inability to meet users' actual needs. Existing research mainly focuses on user needs classification, functional needs acquisition, and non-functional needs acquisition, primarily revolving around mobile app reviews. However, due to the limited amount of industrial app review data and the discrepancy between industrial app user needs and mobile app user needs, existing methods are not directly applicable to the industrial app field. Therefore, this invention considers using transfer learning to compensate for the limited amount of industrial app data. By introducing transfer learning techniques, a model is pre-trained using a large number of mobile app reviews, and then fine-tuned using industrial app reviews. Furthermore, since mobile app reviews have a large amount of data and lack labeling, this invention uses an unsupervised learning approach to automatically generate a large number of samples from the review data based on specific rules. Positive samples represent phrases that embody user needs. Through this unsupervised phrase mining method based on transfer learning, the true user needs are obtained from the reviews, thereby addressing the current situation where industrial apps cannot meet users' actual needs. Figure 2 The overall framework of this method is illustrated. It mainly consists of five parts: positive and negative sample generation, feature extraction, demand phrase classification, transfer learning, and demand phrase clustering.

[0029] This method is proposed to address the issue of limited data volume in industrial app reviews. Its innovation lies in training the model on a small sample size using transfer learning, then fine-tuning the model specifically for industrial apps.

[0030] Positive and negative sample generation. This invention uses an unsupervised approach to extract requirements from comment data, without relying on external manual classification. The generation of positive and negative samples is also completely automatic. This invention starts with the characteristics of the data itself, automatically extracting requirement phrases from the text based on their features. According to the characteristics of requirement phrases, words that appear repeatedly in text often indicate requirement phrases for a given paragraph; a paragraph refers to a series of sentences revolving around a core word. Therefore, based on this characteristic, this invention selects phrases that appear two or more times in a comment as requirement phrases.

[0031] Feature extraction. Feature extraction involves extracting features from the text that distinguish between demand phrases and non-demand phrases. These features help this invention determine the role of each sample phrase in the text and its importance. Demand phrases are the central word sequence surrounding a series of sentences. Since demand phrases are the core of a passage, they must have strong connections with other words in the text. In existing work, self-attention mechanisms can effectively reveal the attention distribution in sentences. Through self-attention, the degree of connection between phrases and other words can be explored.

[0032] Demand phrase classification. Demand phrase classification is to build a classifier to determine whether a phrase is a demand phrase. This invention uses positive and negative sample datasets to train the classifier. In feature extraction, RoBERTa is used to extract features from the dataset. RoBERTa[1] has 12 layers in the default mode, and each layer has 12 attention heads. Therefore, for a text with N words, the input data can be regarded as an image with a length and width of N and 144 channels. The problem of phrase text classification is transformed into an image classification problem. This invention determines whether a phrase is a demand phrase by giving a multi-channel attention mechanism image and uses a two-layer convolutional neural network (CNN) model to classify the multi-channel attention mechanism image.

[0033] Transfer learning. Transfer learning refers to training a model using mobile app reviews, and then fine-tuning the model on an industrial app review dataset after the model has been fully trained. Because mobile app review data and industrial app review data are highly similar, including word frequency, text length, text format, and semantics, and because the industrial app review dataset is relatively small, the parameters of the original CNN model are not retrained; instead, the fully connected layers are retrained. Finally, a sigmoid layer is added after the fully connected layers to perform transfer learning on the industrial app review data.

[0034] Demand phrase clustering. Demand phrase clustering refers to clustering keywords from different reviews within the same app. Since each app has multiple reviews, and each review may contain multiple keywords, submitting all of these keywords to the developers would undoubtedly be an additional burden. Therefore, demand phrase clustering is necessary. This invention employs the unsupervised K-means algorithm for demand phrase clustering. K-means is a classic distance-based clustering algorithm; this invention uses cosine similarity to calculate the distance between words.

[0035] The complete process is as follows. First, the present invention matches the entities in the mobile APP reviews with those in the wiki encyclopedia [2], filters out the matched phrases, and marks these filtered phrases as demand phrases, because the entities in the wiki encyclopedia are some non-everyday nouns, and when they appear in the reviews, they are usually aspects that users are concerned about regarding the APP. Then, phrases that appear twice or more in each mobile APP review are filtered out and marked as demand phrases. These marked demand phrases will be used as positive samples. Then, the present invention randomly extracts some continuous word sequences from the mobile APP reviews and uses these randomly obtained word sequences as negative samples. The mobile APP review dataset is put into the pre-trained model RoBERTa. RoBERTa encodes the review text. RoBERTa follows BERT and uses Transformer as the feature extractor. It uses the multi-head attention mechanism of Transformer to generate the corresponding attention mechanism map for each positive and negative sample phrase. After obtaining the attention mechanism map corresponding to each sample, the present invention inputs the attention mechanism map into the convolutional neural network CNN for binary classification training. Since CNNs are commonly used for training image recognition and classification and have good adaptability to image classification, this invention uses CNNs to classify phrases into demand phrases or non-demand phrases based on the attention mechanism graph of each phrase. After the CNN is trained, the entire model is trained. This invention randomly extracts a large number of continuous word sequences from industrial APP review data. For example, the phrase "this app need payment method" is classified into continuous word sequences such as "this app", "app need", "payment method", and "need payment". The industrial APP review data is fed into a pre-trained model for encoding. Then, attention mechanism graphs are calculated for each randomly generated word sequence. These attention graphs are fed into a CNN classifier for classification to obtain demand phrases. Finally, the obtained demand phrases are grouped according to the APP to which they belong. The demand phrases in each review under each APP are clustered, and the core of the clustering is the user's focus.

[0036] 2. Macroeconomic Demand Forecasting Methods

[0037] To address the issue that existing industrial apps cannot meet the needs of certain industrial sectors, this method predicts which industrial sectors urgently require industrial apps by analyzing their release time and sector tags. Since major domestic industrial apps categorize their sectors differently, and some platforms do not categorize industrial apps by sector, this method requires selecting a sector classification standard to classify all industrial apps using sector tags. An industrial app may contain multiple sector tags. Because the description information of an industrial app includes information about its applicable sectors, this method first needs to classify industrial apps using multiple tags based on their description information, and then predict which sectors will urgently need industrial apps in the future based on the tags and release time. This invention designs as follows... Figure 3 Demand forecasting methods.

[0038] This method mainly consists of two parts: multi-label classification and trend prediction.

[0039] Multi-label classification. This method obtains a large amount of APP description information from various industrial Internet platforms. The description information includes the industrial fields applicable to industrial APPs. This method classifies industrial APPs into different fields through the description information of APPs. According to the crawling results, some platforms have assigned field labels to APPs, but a large number of platforms have not classified their APPs with field labels. Demand prediction analysis should include as many publicly available industrial APP datasets in China as possible. Therefore, this method first needs to train the model with existing labeled data and then apply it to other unlabeled data. Unlike multi-class classification, multi-label classification divides a target into multiple categories. For example, in industrial APPs, the same industrial APP can simultaneously include multiple labels such as smart agriculture, water conservancy monitoring, smart villages, etc. These labels involve multiple fields. An industrial APP may belong to multiple labels at the same time, or it may not belong to any label. These are all problems that may be encountered in multi-label classification. In multi-label classification tasks, a piece of data may have multiple labels. This method uses a combination of ALBERT[3] and TextCNN for multi-label classification. ALBERT is used as the encoding for the dataset, and TextCNN extracts features from the encoded text vectors. Convolutional pooling is performed using different kernels before the data is concatenated. To prevent overfitting, this method discards a portion of the training data, preventing it from participating in the training process. Finally, the data is input into a fully connected layer, and the multi-label classification result is obtained through the sigmoid function. Figure 4 The diagram shows a multi-label classification method.

[0040] Trend Prediction. After classifying each industrial app by domain label based on its description information, this method uses the number of domain labels and the release time of the industrial apps to make macro-trend predictions, helping developers understand which industrial sectors will have strong demand in the future. The release time span of the industrial apps used in this method is from August 15, 2017 to March 30, 2022. This method uses a monthly time unit, counting the number of industrial apps in each sector up to the end of each month, and predicting the number of industrial apps in each sector for the next three months based on the number of industrial apps in each past month. Since the model input of this method is relatively simple, only the number of industrial apps in the past is considered as the sole influencing factor when predicting the number of industrial apps in the future. Therefore, this method considers using a multinomial regression equation for prediction, and uses the sklearn package in Python to generate a multinomial regression equation for each sector. The average monthly growth rate of industrial apps over the past year was 5.8%. After predicting the change in the number of industrial apps for the next three months for each industrial sector, this method selects sectors with an average growth rate of more than 10% in the prediction results, considering them as industrial sectors with strong demand. Figure 5 The diagram shown is a framework diagram of the trend prediction method.

[0041] To address the lack of existing research on the market demand for industrial apps, this invention proposes a macro-level demand forecasting method based on the description information and release time of industrial apps. This method helps developers predict which industrial sectors urgently need industrial apps. The invention uses multi-label classification technology to categorize industrial apps by sector based on their description information. Then, based on the sector classification results and release time, it predicts market demand, identifying which sectors will experience high demand in the future.

Claims

1. A method for acquiring collective intelligence requirements for industrial app development, characterized in that: Based on user review data obtained from industrial APP platforms and mobile APP markets, two models are constructed: a user demand acquisition method and a macro demand prediction method. The user demand acquisition method focuses on users and obtains demands from user reviews, while the macro demand acquisition method focuses on industry sectors and obtains demands from a macro perspective. This realizes a collective intelligent demand acquisition system. Key phrases are derived from user reviews, and the distribution of the number of industrial apps in various industrial sectors over a period of time is used to predict the industrial apps that will be urgently needed in the future. Finally, the key phrases representing user demands in the reviews and the industrial sectors that urgently need industrial apps are output. The specific steps of the user demand acquisition method are as follows: First, positive and negative samples are generated by matching the entities in the comments of the mobile APP with the entities of a specific network information source, filtering out the matched phrases, and marking these filtered phrases as demand phrases. Then, phrases that appear twice or more in each mobile APP comment are filtered out and marked as demand phrases as well. These marked demand phrases will be used as positive samples. Then, a portion of continuous word sequences are randomly extracted from the mobile APP comments and these randomly obtained word sequences are used as negative samples. Next, feature extraction is performed. The mobile APP review dataset and the industrial APP reviews obtained through transfer learning are put into the pre-trained model RoBERTa. RoBERTa encodes the review text. Following BERT, RoBERTa uses Transformer as the feature extractor and uses the multi-head attention mechanism of Transformer to generate a corresponding attention mechanism graph for each positive and negative sample phrase. Then, demand phrase classification is performed. After obtaining the attention mechanism map corresponding to each sample, the attention mechanism map is input into the convolutional neural network (CNN) for binary classification training. The CNN is used to classify the phrase as a demand phrase or a non-demand phrase based on the attention mechanism map of each phrase. After the CNN training is completed, the industrial APP review data is fed into the pre-trained model for encoding. Then, attention mechanism graphs are calculated for each randomly generated word sequence. These attention graphs are fed into the CNN classifier for classification to obtain demand phrases. Finally, the demand phrases are grouped according to the APP to which they belong. Demand phrase clustering is performed on the demand phrases of each review under each APP. The core of the clustering is the user's focus. The macro-demand forecasting method predicts which industrial sectors urgently need industrial apps by using the release time and domain tags of the industrial apps. It includes two parts: multi-tag classification and trend forecasting. By selecting a domain classification standard, all industrial apps are classified by domain tags. Industrial apps are classified by multi-tag based on their description information. Then, based on the tags and release time, the sectors that will urgently need industrial apps in the future are predicted and fed back to the users.

2. The method for acquiring collective intelligence requirements for industrial APP development as described in claim 1, characterized in that: The positive and negative sample generation method is as follows: unsupervised extraction of requirements from comment data is performed. Starting from the characteristics of the data itself, the requirements phrases are automatically extracted from the text based on their characteristics. Phrases that appear twice or more in a comment are taken as requirements phrases.

3. The crowd-sourcing requirement acquisition method for industrial APP development according to claim 2, characterized in that: The feature extraction method is as follows: by using a self-attention mechanism, the degree of connection between phrases and other words is mined.

4. The crowd-sourcing requirement acquisition method for industrial APP development according to claim 3, characterized in that: The proposed demand phrase classification method is as follows: The classifier is trained using positive and negative sample datasets. RoBERTa is used to extract features from the dataset during feature extraction. In its default mode, RoBERTa has 12 layers, each with 12 attention heads. For a text with N words, the input data is treated as an image with dimensions N and 144 channels. This transforms the problem of phrase text classification into an image classification problem. Given a multi-channel attention mechanism image, the method determines whether the phrase is a demand phrase. A two-layer convolutional neural network (CNN) model is used to classify the multi-channel attention mechanism image.

5. The crowd-sourcing requirement acquisition method for industrial APP development according to claim 4, characterized in that: The transfer learning technique involves training the model using mobile app reviews. After the model is fully trained, it is then transferred to an industrial app review dataset for fine-tuning. Given the high similarity between mobile app review data and industrial app review data (in terms of word frequency, text length, text format, and semantics), and the relatively small size of the industrial app review dataset, the original CNN model is not retrained. Instead, the fully connected layers are retrained, and finally, a sigmoid layer is added after the fully connected layers to perform transfer learning on the industrial app review data.

6. The crowd-sourcing requirement acquisition method for industrial APP development according to claim 5, characterized in that: The clustering method for demand phrases is as follows: unsupervised K-means algorithm is used to cluster demand phrases, and cosine similarity is used to calculate the distance between words.

7. The method for acquiring collective intelligence requirements for industrial APP development as described in claim 6, characterized in that: The multi-label classification part obtains a large amount of APP description information from crawlers of various industrial internet platforms. The description information includes the industrial fields to which the industrial APPs are applicable. The industrial APPs are classified into different fields based on the APP description information. First, the model is trained on the existing labeled data. Then, it is applied to other unlabeled data. A combination of ALBERT and TextCNN is used for multi-label classification. ALBERT is used as the encoding of the dataset, and TextCNN is used to extract features from the encoded text vectors. Convolutional pooling is performed using different convolution kernels, and then the data is concatenated. To prevent the model from overfitting, some data is discarded during training. Finally, the data is input into a fully connected layer, and the multi-label classification result is obtained through the sigmoid function.

8. The crowd-sourcing requirement acquisition method for industrial APP development according to claim 7, characterized in that: The trend prediction section performs macro-trend predictions based on the number of domain tags and the release time of industrial apps. Using a monthly time unit, it counts the number of industrial apps in each domain up to the end of each month. Based on the number of industrial apps in each past month, it predicts the number of industrial apps in each domain for the next three months using a multinomial regression equation. The sklearn package in Python generates a multinomial regression equation for each domain. The average monthly growth rate of industrial apps over the past year was 5.8%. After predicting the change in the number of industrial apps for each industrial domain for the next three months, domains with an average growth rate exceeding 10% in the prediction results are selected and considered as industrial domains with strong demand.

Images