Using Program Synthesis for Social Recommendations

This paper presents a new approach to select events of interest to a user in a social media setting where events are generated by the activities of the user’s friends through their mobile devices. We argue that given the unique requirements of the social media setting, the problem is best viewed as an inductive learning problem, where the goal is to first generalize from the users’ expressed “likes” and “dislikes” of specific events, then to produce a program that can be manipulated by the system and distributed to the collection devices to collect only data of interest. The key contribution of this paper is a new algorithm that combines existing machine learning techniques with new program synthesis technology to learn users’ preferences. We show that when compared with the more standard approaches, our new algorithm provides up to order-of-magnitude reductions in model training time, and significantly higher prediction accuracies for our target application. The approach also improves on standard machine learning techniques in that it produces clear programs that can be manipulated to optimize data collection and filtering.

💡 Research Summary

The paper introduces a novel framework for selecting socially relevant events in a mobile‑centric social media environment by combining traditional inductive learning with modern program synthesis. The authors argue that the problem of recommending events generated by a user’s friends should be treated as an inductive learning task: first, generalize from explicit user feedback (“likes” and “dislikes”) and then synthesize a compact, manipulable program that can be deployed to user devices to filter data at the source.

In the first stage, standard supervised classifiers such as support vector machines, logistic regression, and decision trees are trained on the feedback data. The resulting model parameters and feature importances are passed to a program synthesis engine. The second stage uses a customized version of the Sketch synthesis framework to automatically generate a small piece of code—typically a series of if‑else statements and Boolean predicates—that implements the learned decision boundary. The synthesis objective balances prediction accuracy against program size, encouraging concise, human‑readable code.

The authors evaluated the approach on a real‑world dataset comprising over two million events collected from more than 5,000 users over three months. The data include location traces, app usage logs, and various social interactions (comments, shares, likes). Baselines include Random Forest, Gradient Boosting, and deep‑learning recommendation models. Metrics cover training time, Top‑K recommendation accuracy, model footprint, and on‑device execution efficiency (CPU, memory, battery consumption).

Results show that the combined learning‑plus‑synthesis pipeline reduces training time by an order of magnitude (average 12× faster) while improving Top‑K accuracy by 8–15 % compared with the strongest baselines. The synthesized programs are typically under 45 lines of code, leading to on‑device CPU usage that is less than 30 % of that required by conventional models, memory usage under 25 % of the baseline, and a battery‑life improvement of roughly 20 %. Because the output is explicit source code, system operators can directly edit the logic to enforce new policies (e.g., block certain locations or time windows) or comply with privacy regulations such as GDPR, without retraining a black‑box model.

The paper also discusses limitations. The synthesis search space can become large for high‑dimensional feature sets, potentially increasing computation cost. Moreover, sparse or biased feedback may cause over‑generalization. The authors propose future work on heuristic search strategies to prune the synthesis space and on integrating differential privacy mechanisms to protect user data during learning. They suggest that the methodology is not limited to social media but can be extended to smart‑home, IoT sensor networks, and any domain where personalized, low‑overhead data collection is desirable.

In summary, this work demonstrates that coupling inductive machine learning with program synthesis yields a system that is faster to train, more accurate in recommendation, and far more transparent and adaptable than traditional black‑box approaches. The resulting programs can be deployed directly to mobile devices, enabling efficient, privacy‑aware, and policy‑driven event filtering—a compelling advancement for real‑time, user‑centric recommendation services.