Basic direction and realization path of PE teaching innovation in PSS based on deep learning model

Huiming Ke, Yan Wang3C 2023 · DOI

TL;DR

This paper explores how deep learning, specifically Long Short-Term Memory (LSTM) networks, can be used to steer the future of Physical Education (PE) in primary and secondary schools. By treating educational trends as time-series data, the authors developed a model that predicts the most effective paths for teaching innovation, significantly outperforming traditional machine learning baselines.

What Problem Does This Paper Address?

Physical Education in primary and secondary schools (PSS) often suffers from a "demonstration-imitation" rut. Teachers explain a movement, students mimic it, and the cycle repeats. This traditional model frequently leads to poor student engagement and fails to foster a lifelong interest in health.

The authors argue that while many researchers call for "innovation," there is a lack of data-driven guidance on what specifically should change. The challenge lies in the fact that educational trends are not random; they are sequential and context-dependent. To solve this, the paper proposes using predictive modeling to identify which innovative directions—such as shifting teaching philosophies or diversifying content—are most likely to succeed based on historical educational data.

Methods

The researchers propose a CLNN model (a hybrid of Convolutional Neural Networks and LSTM).

  • Why LSTM? Standard neural networks treat inputs independently. However, innovation in a curriculum is a "time-series" problem where previous methods influence future ones. LSTMs are designed to handle these long-term dependencies without suffering from the "gradient disappearance" issues common in simpler recurrent networks.
  • Architecture: The model uses a three-tier structure (70, 50, and 25 intermediate units). It employs a CNN-like approach to extract local features from "user behavior" (educational data points) and an LSTM layer to capture temporal relationships.
  • Optimization: The team used the "CNN" algorithm for gradient training with a learning rate of 0.001 and a 70/30 split for training and testing.
  • Output: A softmax layer maps the results into a probability region (0,1) to classify the most viable innovation directions.

Datasets

The paper utilizes a dataset focused on "innovation direction" and "user behavior" within the context of sports innovation methods. While the specific origin of the raw data (e.g., a specific national database or school district) is not explicitly named, the authors describe extracting features related to teaching thoughts, objectives, and content systems. They used this data to generate "education vectors" for the model.

Key Insights and Ideas

  • Innovation as a Path, Not a Point: The most significant insight is the application of time-series logic to pedagogy. By recognizing that educational reform is a sequence of related events, the authors justify why LSTM—a tool usually reserved for speech recognition or stock market prediction—is actually the perfect fit for curriculum design.
  • The "Health First" Pivot: The model identifies a critical shift from "PE as fitness" to "PE as health education." This suggests that innovation shouldn't just focus on making kids faster or stronger, but on the "combination of PE and health education" to promote lifelong wellness.
  • The CLNN Advantage: The authors found that while standard CNNs are good at accuracy, they struggle with decision-making logic. By combining CNNs with LSTMs, the model captures both the "what" (local features) and the "when/why" (temporal relationships), leading to more robust predictions.

Key Results

The proposed CLNN/LSTM model was tested against a battery of traditional classifiers, including Support Vector Machines (SVM), Random Forest (RF), and Naive Bayes (NB).

  • Quantitative Superiority: The LSTM-based model outperformed the baseline models across four key metrics: Accuracy, Precision, F1-score, and AUC. Specifically, it scored between 0.0532 and 0.2323 higher than the baselines.
  • Comparison to Deep Learning: Even when compared to a standard CNN-RNN hybrid, the CLNN model was 0.0422 higher on average.
  • Qualitative Directions: The model successfully identified three primary pillars for PE innovation:
  • Teaching Ideas: Moving toward "seeking knowledge and innovation."
    1. Process/Objectives: Shifting focus toward "exercise independence" and "peaceful coexistence."
    2. Content System: Moving toward a diversified curriculum that integrates "thought education" with physical activity.

Limitations and Future Work

The authors acknowledge that the PE system is "completely dynamic," meaning a model trained today might become obsolete as societal values shift.

  • Generalizability: The paper does not deeply discuss how the model performs across different cultural or socioeconomic school environments, which could drastically change what "innovation" looks like.
  • Implementation: While the model predicts what to change, the "realization path" remains high-level. Future work could involve integrating these AI predictions into actual teacher-training platforms to see if the predicted "innovations" translate to better student outcomes in the real world.

Summary updated

Abstract

At present, the traditional model of PE in PSS (PSS)has seriously affected the quality of PE teaching in PSS and the perception of PE among primary and secondary school students. Because of the urgent need for innovation in PE in PSS, this study proposes the LSTM model to achieve an accurate prediction of the innovation direction of PE in PSS. Based on the LSTM model, the user behavior is classified by extracting the important features of the innovation direction. Expression to achieve accurate prediction of the future development direction of PE. Using the data confusion matrix to estimate the prediction accuracy of the LSTM model, the four evaluation indicators of Accuracy, Precision, F1, and AUC are 0.0532~0.2323 higher than the baseline model. The prediction results of PE teaching innovation in PSS from three aspects of teaching thought, teaching content, teaching objectives and essence are output, which has obvious guiding significance for the overall optimization of PE classrooms in PSS. This result shows that the LSTM prediction model has important practical value.
BibTeX
@inproceedings{ke2023basic,
  author = {Huiming Ke and Yan Wang},
  title = {{Basic direction and realization path of PE teaching innovation in PSS based on deep learning model}},
  booktitle = {3C},
  year = {2023},
  doi = {10.17993/3ctecno.2023.v12n1e43.70-85},
}
Figure