Dynamic Action Masking and Priority Experience Replay in Deep Reinforcement Learning News Recommenders

Abstract

News consumption differs from stable-preference domains: relevance decays within hours and user interest shifts non-sequentially across sessions, motivating deep reinforcement learning (DRL) approaches that plan for sequential engagement rather than static next-click accuracy. A recent proposal, dynamic action-masking deep reinforcement learning (DAMDRL), combines a three-channel GRU state encoder, a TD-error-weighted priority experience cache, and a per-step action mask within a double deep Q-network to address this setting, reporting NDCG@10 and MRR@10 gains over plain DQN and three supervised neural baselines. Independent audits of the wider DRL-recommendation literature, however, have found that a substantial share of such reported gains are not independently reproducible, and in some cases originate from incidental representation-learning effects rather than genuine policy improvement. This paper re-examines the DAMDRL proposal against that critical literature. We restructure the described method into an explicit pipeline and identify eleven specific items — dataset identity, hyperparameters, and training protocol among them — required for replication but absent from the original report. We recompute every comparative and ablation result as explicit percentage deltas, finding NDCG@10 gains of 243–333% over an SVD++ floor and 18–26% losses when priority caching or loop-avoidance exploration is ablated. These patterns are consistent with, but do not rule out, an upstream representation-learning explanation raised by recent reassessment studies of comparable frameworks. We conclude that DAMDRL's central claim is numerically consistent with its own tables but not yet independently verifiable, and provide a reporting checklist to close that gap.