14 KiB
14 KiB
tags:: /unread, #zotero title:: @TOSEM-2025-0054_Proof_hi item-type:: document original-title:: TOSEM-2025-0054_Proof_hi language:: en links:: Local library, Web library
-
Attachments
- PDF {{zotero-imported-file RNTJP3AD, "TOSEM-2025-0054_Proof_hi.pdf"}}
-
Notes
-
Annotazioni
(1/4/2025, 14:25:59)
- “Software Engineering and Methodology” (“TOSEM-2025-0054_Proof_hi”, p. 1) #ffd400
-
- “Automated Software Engineering” (“TOSEM-2025-0054_Proof_hi”, p. 1) #ffd400
* *
- “Transactions on Software Engineering” (“TOSEM-2025-0054_Proof_hi”, p. 1) #ffd400
* *
- “we did not compare our method with the state-of-the-art TPL recommendation method” (“TOSEM-2025-0054_Proof_hi”, p. 1) #5fb236
* *
- “we did not evaluate the performance of our method in” (“TOSEM-2025-0054_Proof_hi”, p. 1) #5fb236
* *
- “function information of the app itself.” (“TOSEM-2025-0054_Proof_hi”, p. 2) #a28ae5
* *
- “each TPL in the TPL context equally” (“TOSEM-2025-0054_Proof_hi”, p. 2) #a28ae5
* *
- “ining the hidden relationships among candidate TPL” (“TOSEM-2025-0054_Proof_hi”, p. 2) #a28ae5
* *
- “the TPL context,” (“TOSEM-2025-0054_Proof_hi”, p. 2) #a28ae5
* *
- “LLM to mine app functions from description texts.” (“TOSEM-2025-0054_Proof_hi”, p. 2) #a28ae5
* *
- “Finding the synergy between different TPLs is another time-consuming process for developer” (“TOSEM-2025-0054_Proof_hi”, p. 3) #5fb236
* *
- “CF-based recommendation methods are often susceptible to the popularity bias [14, 28], that is, a small fraction of popular TPLs - those that are used by a large number of apps - dominate the prediction results and most other TPLs are ill-served.” (“TOSEM-2025-0054_Proof_hi”, p. 3) #5fb236
* *
- “mogenization of node” (“TOSEM-2025-0054_Proof_hi”, p. 3) #a28ae5
* *
- “ignore the function information of the app” (“TOSEM-2025-0054_Proof_hi”, p. 3) #a28ae5
* *
- “value that reflects the match degree between app product and candidate TPL.” (“TOSEM-2025-0054_Proof_hi”, p. 3) #5fb236
* *
- “Fig. 1. TPL usage for two different apps” (“TOSEM-2025-0054_Proof_hi”, p. 4) #5fb236
* *
- “we mine their description texts.” (“TOSEM-2025-0054_Proof_hi”, p. 4) #a28ae5
* *
- “app product” (“TOSEM-2025-0054_Proof_hi”, p. 4) #5fb236
* *
- “candidate TPL.” (“TOSEM-2025-0054_Proof_hi”, p. 4) #5fb236
* *
- “function information of app products” (“TOSEM-2025-0054_Proof_hi”, p. 4) #5fb236
* *
- “reliability and diversity” (“TOSEM-2025-0054_Proof_hi”, p. 4) #5fb236
* *
- “Meanwhile, since T3 and T5 are as relevant to the development of video game as T2, we should pay more attention to T3 and T5 among T1, T3, and T5 to explore whether T2 can synergize with them to complete some video game-related development tasks.” (“TOSEM-2025-0054_Proof_hi”, p. 5) #a28ae5
* *
- “App. App refers to mobile application.” (“TOSEM-2025-0054_Proof_hi”, p. 6) #ffd400
*Is the proposed approach only for supporting the development of mobile applications? *
- “by app product.” (“TOSEM-2025-0054_Proof_hi”, p. 6) #ff6666
*which one? *
- “TPLs in the historical data.” (“TOSEM-2025-0054_Proof_hi”, p. 6) #ffd400
*what do you mean? *
- “apps in the historical data” (“TOSEM-2025-0054_Proof_hi”, p. 6) #ffd400
*What do you mean? *
- “we first mine the functions from its description text by utilizing LLM” (“TOSEM-2025-0054_Proof_hi”, p. 6) #e56eee
* *
- “app product and each candidate TPL in the TPL pooling, and thereby establishing a TPL recommendation list.” (“TOSEM-2025-0054_Proof_hi”, p. 6) #ffd400
*This is a potential and initial list, isn't it? *
- “pp product and candidate TPL” (“TOSEM-2025-0054_Proof_hi”, p. 7) #ffd400
*only one candidate TPL is given? Or more than one recommended items are given? *
- “provided” (“TOSEM-2025-0054_Proof_hi”, p. 7) #ffd400
*to be provided, isn't it? *
- “To provide better support for TPL recommendation, we employ LLM to mine the functions from the description text of app product.” (“TOSEM-2025-0054_Proof_hi”, p. 7) #ffd400
*it's uncommon to have such a text before the release of the wanted app. *
- “generating” (“TOSEM-2025-0054_Proof_hi”, p. 8) #ffd400
*generating a list of app functions can be a threat to validity. Check if it is stated later in the paper. I'm expecting a research question on this point. *
- “First, Atten-TPL generates the m-dimensional vectors for candidate TPL and TPLs in the TPL context using One-hot Encoding [31], where m is the number of TPLs in the TPL pooling. Second, Atten-TPL multiplies the vector of candidate TPL and the ones of TPLs in the TPL context with two trainable weight matrixes W q ∈ Rm×dk and W k ∈ Rm×dk to obtain Q ∈ R1×dk and K ∈ Rn×dk , respectively (n is the number of TPLs in the TPL context).” (“TOSEM-2025-0054_Proof_hi”, p. 8) #ffd400
*A concrete example is necessary here. An illustrative project with the corresponding TPL and the corresponding represenation (even abstract) to help reader understand how the proposed encoding works. *
- “Task” (“TOSEM-2025-0054_Proof_hi”, p. 8) #2ea8e5
* *
- “Function definition” (“TOSEM-2025-0054_Proof_hi”, p. 8) #2ea8e5
* *
- “Function examples” (“TOSEM-2025-0054_Proof_hi”, p. 8) #2ea8e5
* *
- “we provide five function examples, which are manually extracted from the description texts of existing app products” (“TOSEM-2025-0054_Proof_hi”, p. 8) #5fb236
* *
- “Non-functional content” (“TOSEM-2025-0054_Proof_hi”, p. 8) #2ea8e5
* *
- “Input and output” (“TOSEM-2025-0054_Proof_hi”, p. 8) #2ea8e5
* *
- “Atten-TPL” (“TOSEM-2025-0054_Proof_hi”, p. 8) #e56eee
* *
- “each candidate TPL” (“TOSEM-2025-0054_Proof_hi”, p. 8) #5fb236
* *
- “Generating the vector for TPL context.” (“TOSEM-2025-0054_Proof_hi”, p. 8) #2ea8e5
* *
- “app product” (“TOSEM-2025-0054_Proof_hi”, p. 8) #5fb236
* *
- “not every TPL in the TPL context can make a positive contribution.” (“TOSEM-2025-0054_Proof_hi”, p. 8) #a28ae5
* *
- “truly useful for the task” (“TOSEM-2025-0054_Proof_hi”, p. 8) #e56eee
* *
- “m-dimensional vectors” (“TOSEM-2025-0054_Proof_hi”, p. 8) #5fb236
* *
- “formed randomly from the vectors of all TPLs in the TPL context.” (“TOSEM-2025-0054_Proof_hi”, p. 8) #ffd400
*This is not clear! *
- “After that, the model applies a softmax operation to normalize all attention scores so that they are added up to 1.” (“TOSEM-2025-0054_Proof_hi”, p. 8) #5fb236
* *
- “match degree between app product and candidate TPL.” (“TOSEM-2025-0054_Proof_hi”, p. 9) #a28ae5
* *
- “(3)” (“TOSEM-2025-0054_Proof_hi”, p. 9) #5fb236
* *
- “Evaluating the match degree between app prod” (“TOSEM-2025-0054_Proof_hi”, p. 9) #2ea8e5
* *
- “activations, bias, and model weights” (“TOSEM-2025-0054_Proof_hi”, p. 9) #e56eee
* *
- “sentence-level tasks” (“TOSEM-2025-0054_Proof_hi”, p. 10) #5fb236
* *
- “We create a dataset consisting of positive triples and negative triples based on the hold-one-out strategy [7], and each triple contains three parts, including: candidate TPL, TPL context, and function list.” (“TOSEM-2025-0054_Proof_hi”, p. 10) #ffd400
*The creation of this dataset is also a critical point of the approach, which can represent a threat to validity. *
- “The function list is obtained by utilizing LLM to process the description text of app product.” (“TOSEM-2025-0054_Proof_hi”, p. 10) #ffd400
*As I said the fact that the function lists are obtained by means of LLMs represents a crucial threat to validity because creating a clean and correct dataset is of paramount importance in this case. *
- “Meanwhile, we also establish multiple negative triples for each app, and the number of negative triples is ten times the number of positive triples” (“TOSEM-2025-0054_Proof_hi”, p. 10) #5fb236
* *
- “softmax classifier” (“TOSEM-2025-0054_Proof_hi”, p. 10) #a28ae5
* *
- “Second, Atten-TPL calculates the mean for the vectors of all functions and uses the result as the vector of function list.” (“TOSEM-2025-0054_Proof_hi”, p. 10) #ffd400
*Also here, can you give an illustrative example? *
- “hold-one-out strategy” (“TOSEM-2025-0054_Proof_hi”, p. 10) #a28ae5
* *
- “each triple contains three parts, i” (“TOSEM-2025-0054_Proof_hi”, p. 10) #a28ae5
* *
- “In each operation, we remove a TPL from the TPL list of app product, the removed TPL is treated as candidate TPL, and the remaining TPLs are used as the TPL context. The function list is obtained by utilizing LLM to process the description text of app product” (“TOSEM-2025-0054_Proof_hi”, p. 10) #ffd400
*how many times has this been done? For how many product, how many candidate TPLs? *
- “we remove a TPL from the TPL list of app product” (“TOSEM-2025-0054_Proof_hi”, p. 10) #a28ae5
* *
- “Positive triples and negative triples are labeled 1 and 0, respectively.” (“TOSEM-2025-0054_Proof_hi”, p. 10) #a28ae5
* *
- “How does Atten-TPL perform on the TPL recommendation?” (“TOSEM-2025-0054_Proof_hi”, p. 10) #2ea8e5
* *
- “function information of app product actually improve the performance of Atten-TPL?” (“TOSEM-2025-0054_Proof_hi”, p. 11) #2ea8e5
* *
- “attention mechanism positively affect” (“TOSEM-2025-0054_Proof_hi”, p. 11) #2ea8e5
* *
- “Can Atten-TPL provide help for app developers in software engineering practice?” (“TOSEM-2025-0054_Proof_hi”, p. 11) #ffd400
*This is a qualitative research question. Are used involved?Let's see what the authors have done for this. *
- “MALib dataset [14], a public real-world dataset that contains 61,722 apps, 827 distinct TPLs, and 725,502 app-TPL usage records.” (“TOSEM-2025-0054_Proof_hi”, p. 11) #a28ae5
* *
- “Since the MALib dataset does not provide the description texts needed for our recommendation model, we wrote a web crawler to scrape the description texts for apps in it from app store or third-party website.” (“TOSEM-2025-0054_Proof_hi”, p. 11) #ffd400
*This can be a potential threat to validity. Have you double checked the crawled data and checked that they are linked to the projects correctly? *
- “44,873 apps, and these apps were used as the dataset of our experiment.” (“TOSEM-2025-0054_Proof_hi”, p. 11) #ffd400
*Have you double checked that the crawled information is consistent with the corresponding apps? *
- “TPL author, TPL Dependency” (“TOSEM-2025-0054_Proof_hi”, p. 11) #5fb236
* *
- “0.2,” (“TOSEM-2025-0054_Proof_hi”, p. 11) #5fb236
* *
- “The larger the value, the more random the generated text, while the smaller the value, the less random the generated text, resulting in the more organized and precise response.” (“TOSEM-2025-0054_Proof_hi”, p. 11) #5fb236
* *
- “2rm” (“TOSEM-2025-0054_Proof_hi”, p. 11) #ffd400
*what is it? *
- “Atten-TPL mines app functions from the description text by employing ChatGPT” (“TOSEM-2025-0054_Proof_hi”, p. 15) #ffd400
* *
- “n addition to the TPL context, the input of our recommendation method includes the description text of app product. Considering that the description text must be completed by app developers before the application is” (“TOSEM-2025-0054_Proof_hi”, p. 19) #5fb236
* *
- “In this paper, we propose a novel TPL recommendation model called Atten-TPL. Atten-TPL can analyze the match degree between app product and each candidate TPL in the TPL pooling by utilizing a deep neural network to mine the hidden relationships among candidate TPL, the TPL context and the functions of app product. By employing the attention mechanism, Atten-TPL can pay more attention to TPLs in the TPL context that are truly useful for the evaluation task. Furthermore, to effectively gain the functions provided by apps, we use LLM to mine their description texts. Compared to existing state-of-the-art TPL recommendation methods, Atten-TPL can return more reliable recomemndation results and performs well in terms of diversity. In future work, we plan to propose a method to generate the explanatory texts for TPL recommendation lists, which will assist developers in understanding and utilizing the recommended TPLs. The robust text generation capability of LLM holds promise for achieving this goal.” (“TOSEM-2025-0054_Proof_hi”, p. 20) #a28ae5
* *
- “released, our recommendation model is highly practical in real-world development.” (“TOSEM-2025-0054_Proof_hi”, p. 20) #5fb236
* *
- “Meanwhile, even if the developers of app product do not provide the description text, our model can still rely on the TPL context to recommend the available TPLs for them with acceptable results (The experimental result of RQ2 can demonstrate this). Furthermore, since the TPL ecosystem is dynamic, the app and TPL poolings can be updated and the recommendation model can be retrained periodically, enabling the model to integrate new or updated TPLs.” (“TOSEM-2025-0054_Proof_hi”, p. 20) #ffd400
* *