Looking into Black Box Code Language Models

Language Models (LMs) have shown their application for tasks pertinent to code and several code~LMs have been proposed recently. The majority of the studies in this direction only focus on the improvements in performance of the LMs on different benchmarks, whereas LMs are considered black boxes. Besides this, a handful of works attempt to understand the role of attention layers in the code~LMs. Nonetheless, feed-forward layers remain under-explored which consist of two-thirds of a typical transformer model's parameters. In this work, we attempt to gain insights into the inner workings of code language models by examining the feed-forward layers. To conduct our investigations, we use two state-of-the-art code~LMs, Codegen-Mono and Ploycoder, and three widely used programming languages, Java, Go, and Python. We focus on examining the organization of stored concepts, the editability of these concepts, and the roles of different layers and input context size variations for output generation. Our empirical findings demonstrate that lower layers capture syntactic patterns while higher layers encode abstract concepts and semantics. We show concepts of interest can be edited within feed-forward layers without compromising code~LM performance. Additionally, we observe initial layers serve as ``thinking'' layers, while later layers are crucial for predicting subsequent code tokens. Furthermore, we discover earlier layers can accurately predict smaller contexts, but larger contexts need critical later layers' contributions. We anticipate these findings will facilitate better understanding, debugging, and testing of code~LMs.

MobileConvRec: A Conversational Dataset for Mobile Apps Recommendations

Existing recommendation systems have focused on two paradigms: 1- historical user-item interaction-based recommendations and 2- conversational recommendations. Conversational recommendation systems facilitate natural language dialogues between users and the system, allowing the system to solicit users' explicit needs while enabling users to inquire about recommendations and provide feedback. Due to substantial advancements in natural language processing, conversational recommendation systems have gained prominence. Existing conversational recommendation datasets have greatly facilitated research in their respective domains. Despite the exponential growth in mobile users and apps in recent years, research in conversational mobile app recommender systems has faced substantial constraints. This limitation can primarily be attributed to the lack of high-quality benchmark datasets specifically tailored for mobile apps. To facilitate research for conversational mobile app recommendations, we introduce MobileConvRec. MobileConvRec simulates conversations by leveraging real user interactions with mobile apps on the Google Play store, originally captured in large-scale mobile app recommendation dataset MobileRec. The proposed conversational recommendation dataset synergizes sequential user-item interactions, which reflect implicit user preferences, with comprehensive multi-turn conversations to effectively grasp explicit user needs. MobileConvRec consists of over 12K multi-turn recommendation-related conversations spanning 45 app categories. Moreover, MobileConvRec presents rich metadata for each app such as permissions data, security and privacy-related information, and binary executables of apps, among others. We demonstrate that MobileConvRec can serve as an excellent testbed for conversational mobile app recommendation through a comparative study of several pre-trained large language models.

MobileRec: A Large-Scale Dataset for Mobile Apps Recommendation

Recommender systems have become ubiquitous in our digital lives, from recommending products on e-commerce websites to suggesting movies and music on streaming platforms. Existing recommendation datasets, such as Amazon Product Reviews and MovieLens, greatly facilitated the research and development of recommender systems in their respective domains. While the number of mobile users and applications (aka apps) has increased exponentially over the past decade, research in mobile app recommender systems has been significantly constrained, primarily due to the lack of high-quality benchmark datasets, as opposed to recommendations for products, movies, and news. To facilitate research for app recommendation systems, we introduce a large-scale dataset, called MobileRec. We constructed MobileRec from users' activity on the Google play store. MobileRec contains 19.3 million user interactions (i.e., user reviews on apps) with over 10K unique apps across 48 categories. MobileRec records the sequential activity of a total of 0.7 million distinct users. Each of these users has interacted with no fewer than five distinct apps, which stands in contrast to previous datasets on mobile apps that recorded only a single interaction per user. Furthermore, MobileRec presents users' ratings as well as sentiments on installed apps, and each app contains rich metadata such as app name, category, description, and overall rating, among others. We demonstrate that MobileRec can serve as an excellent testbed for app recommendation through a comparative study of several state-of-the-art recommendation approaches. The quantitative results can act as a baseline for other researchers to compare their results against. The MobileRec dataset is available at https://huggingface.co/datasets/recmeapp/mobilerec.

Proactive Prioritization of App Issues via Contrastive Learning

Mobile app stores produce a tremendous amount of data in the form of user reviews, which is a huge source of user requirements and sentiments; such reviews allow app developers to proactively address issues in their apps. However, only a small number of reviews capture common issues and sentiments which creates a need for automatically identifying prominent reviews. Unfortunately, most existing work in text ranking and popularity prediction focuses on social contexts where other signals are available, which renders such works ineffective in the context of app reviews. In this work, we propose a new framework, PPrior, that enables proactive prioritization of app issues through identifying prominent reviews (ones predicted to receive a large number of votes in a given time window). Predicting highly-voted reviews is challenging given that, unlike social posts, social network features of users are not available. Moreover, there is an issue of class imbalance, since a large number of user reviews receive little to no votes. PPrior employs a pre-trained T5 model and works in three phases. Phase one adapts the pre-trained T5 model to the user reviews data in a self-supervised fashion. In phase two, we leverage contrastive training to learn a generic and task-independent representation of user reviews. Phase three uses radius neighbors classifier t o m ake t he final predictions. This phase also uses FAISS index for scalability and efficient search. To conduct extensive experiments, we acquired a large dataset of over 2.1 million user reviews from Google Play. Our experimental results demonstrate the effectiveness of the proposed framework when compared against several state-of-the-art approaches. Moreover, the accuracy of PPrior in predicting prominent reviews is comparable to that of experienced app developers.

AARSynth: App-Aware Response Synthesis for User Reviews

Responding to user reviews promptly and satisfactorily improves application ratings, which is key to application popularity and success. The proliferation of such reviews makes it virtually impossible for developers to keep up with responding manually. To address this challenge, recent work has shown the possibility of automatic response generation. However, because the training review-response pairs are aggregated from many different apps, it remains challenging for such models to generate app-specific responses, which, on the other hand, are often desirable as apps have different features and concerns. Solving the challenge by simply building a model per app (i.e., training with review-response pairs of a single app) may be insufficient because individual apps have limited review-response pairs, and such pairs typically lack the relevant information needed to respond to a new review. To enable app-specific response generation, this work proposes AARSynth: an app-aware response synthesis system. The key idea behind AARSynth is to augment the seq2seq model with information specific to a given app. Given a new user review, it first retrieves the top-K most relevant app reviews and the most relevant snippet from the app description. The retrieved information and the new user review are then fed into a fused machine learning model that integrates the seq2seq model with a machine reading comprehension model. The latter helps digest the retrieved reviews and app description. Finally, the fused model generates a response that is customized to the given app. We evaluated AARSynth using a large corpus of reviews and responses from Google Play. The results show that AARSynth outperforms the state-of-the-art system by 22.2% on BLEU-4 score. Furthermore, our human study shows that AARSynth produces a statistically significant improvement in response quality compared to the state-of-the-art system.

Measuring LDA Topic Stability from Clusters of Replicated Runs

Background: Unstructured and textual data is increasing rapidly and Latent Dirichlet Allocation (LDA) topic modeling is a popular data analysis methods for it. Past work suggests that instability of LDA topics may lead to systematic errors. Aim: We propose a method that relies on replicated LDA runs, clustering, and providing a stability metric for the topics. Method: We generate k LDA topics and replicate this process n times resulting in n*k topics. Then we use K-medioids to cluster the n*k topics to k clusters. The k clusters now represent the original LDA topics and we present them like normal LDA topics showing the ten most probable words. For the clusters, we try multiple stability metrics, out of which we recommend Rank-Biased Overlap, showing the stability of the topics inside the clusters. Results: We provide an initial validation where our method is used for 270,000 Mozilla Firefox commit messages with k=20 and n=20. We show how our topic stability metrics are related to the contents of the topics. Conclusions: Advances in text mining enable us to analyze large masses of text in software engineering but non-deterministic algorithms, such as LDA, may lead to unreplicable conclusions. Our approach makes LDA stability transparent and is also complementary rather than alternative to many prior works that focus on LDA parameter tuning.