Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

1. In the Introduction, it is stated that *"But the nutritional needs of dogs can be very different depending on their size, age, activity level, and whether they have any diseases"* in the lines preceding lines 35–40. However, the subsequent example—*"Owners of large breeds, on the other hand, often prefer dry food because it costs less"*—appears to have a minor logical coherence issue. 2. The paper mentions that *"part of the data was collected using web scraping"* but fails to specify the exact data sources for the web crawler (e.g., pet food e-commerce platforms, veterinary literature databases). Although the official website of Hill's Pet Nutrition is cited later, it is recommended to refine the description of data sources in the early data collection section. 3. The study utilized 330 types of feed (Table 5) for training and classification. For high-dimensional vector spaces such as embeddings from large language models like E5 or TF-IDF, is the sample size too small? It is recommended to discuss whether the insufficient sample size has led to the risk of overfitting. 4. In Table 7, the cosine similarity for the *"Musculoskeletal and Joint Care"* category is -0.69 (a negative value). The paper does not explain the physical meaning of *"negative similarity"* and only states that *"there are significant differences in ingredient and nutrient composition"*, resulting in a somewhat vague interpretation. 5. The paper uses the Jaccard coefficient to measure the similarity between the generated recipes and commercial recipes, and employs this as an evaluation criterion. Commercial pet foods usually contain a large number of fillers (e.g., starch-based binders) that are unnecessary in homemade diets. Therefore, the *"dissimilarity"* between the generated recipes and commercial recipes does not necessarily indicate poor quality. It is recommended that the authors revise the interpretation of the Jaccard coefficient in the discussion section, acknowledging that it is only a similarity metric rather than an absolute quality indicator. 6. It is recommended to supplement the axis labels for Figure 9 and Figure 10. In addition, Figure 9 mentions that *"dry feeds are mainly concentrated on the left side and wet feeds on the right side"*, but does not provide quantitative indicators such as the quantity ratio of dry/wet feeds or the clustering separation degree. 7. The paper mentions that *"lexical variants were normalized and names were standardized for ingredient descriptions"* but does not specify the exact rules (e.g., the basis for unifying *"meat and animal by-products"* into *"chicken meat"*, and the rationality explanation for mapping *"vegetables"* to *"carrots"*). 8. Although the paper does not require a standard ablation experiment, as a system involving linear programming, is it lacking in sensitivity analysis experiments? How can the stability of this system be proven? If the user-input weight of the dog changes slightly by 0.5 kg, will the recommended recipe undergo drastic changes? 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This research presents a hybrid approach for generating personalized canine diets, integrating techniques from natural language processing, machine learning, and mathematical optimization. The article is highly suitable for the journal Informatics, as it applies artificial intelligence, machine learning, and transformer models for semantic analysis to create a specialized system within the field of veterinary informatics.

The study evaluates two primary methods for the automated annotation of missing data regarding specialized foods: TF-IDF and the E5 (Enhanced Embedding Encoder) model. The E5 model demonstrated a higher average precision of 88.2% compared to 85.4% for TF-IDF, owing to its superior ability to capture semantic contexts.

For ingredient extraction, both machine learning and statistical approaches were utilized. The machine learning method (Ridge Classifier) generated more stable and semantically consistent recipes, achieving an average Jaccard coefficient of 0.43. This method successfully transfers contextual information between functional categories. In contrast, the statistical method (using χ² criteria and PMI) provided greater variability (coefficient of 0.19) and identified ingredients specifically distinguishable for given categories, though it proved more sensitive to data sparsity in small samples.

Linear programming (LP) was employed to finalize ingredient ratios. The system balances macronutrients (protein, fat, carbohydrates) while maximizing specific indicators critical to the food's functional purpose, such as moisture or protein content.

The literature sources are appropriate and comprehensive, covering fundamental nutrient databases (USDA FoodData Central), specialized veterinary resources (Hill’s Pet Nutrition), international nutritional standards (FEDIAF), and scientific developments in algorithmic optimization and machine learning. The sources are exceptionally current, with many published in 2023, 2024, and 2025.

The research could be further improved by incorporating micronutrients (vitamins and minerals) and technological additives, which are essential for the complete nutritional balance of industrial products. Currently, the system faces difficulties in categorizing general health needs such as "Aging Care" or "Skin Health" due to a lack of unique, specific key terms, unlike more specialized areas like "Heart Care" or "Kidney Care". Finally, the authors acknowledge a limitation regarding the lack of precise information on ingredient proportions in commercial diets, which may lead simulated recipes to deviate from actual production standards.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf