Algorithmic advances in smart TV content recommendation: a structured evidence-mapping review

Recommendation systems have become an invisible layer of modern life, quietly shaping what billions of people watch on their living room screens. Unlike mobile apps, where users actively tap and swipe, television viewing is often passive, shared, and spread across several household members. A new review published in Frontiers in Big Data on 14 September 2026 argues that this difference makes television one of the hardest domains for recommendation technology, and it maps the field in unprecedented detail.

The review, led by Zhen Chen and colleagues, systematically analyzed 120 studies published between 2010 and 2026, covering everything from linear models to large language models. The authors found that television recommendation systems differ from mobile apps in three key ways: viewing is passive, content is shared, and usage is spread across household members. These differences, they argue, make television one of the most challenging domains for recommendation technology.

The review found that television viewing is passive, content is shared, and usage is spread across household members. The authors argue that these differences make television one of the most difficult domains for recommendation systems.

The review concluded that television viewing is passive, content is shared, and usage is spread across household members. The authors argue that these differences make television one of the hardest domains for recommendation technology.

The methodological backbone of the review deserves attention because it illustrates how modern evidence mapping differs from a traditional narrative survey. The authors began with 2,526 database-level results drawn from six sources, then applied a combined deduplication and title/abstract screening stage that removed 2,414 records. That left 112 unique studies for full-text assessment, of which 87 were excluded and 25 met the eligibility criteria. These 25 newly eligible studies were merged with an initial corpus of 95 studies to produce the final 120-study evidence map. Every record in the corpus carries a DOI, with Crossref verification covering 116 entries and six stable arXiv records retained separately, a detail that signals an unusually rigorous approach to source integrity in a literature where citation rot is common.

Reliability of the screening decisions was tested through an independent dual-reviewer audit. Two reviewers assessed 126 records, comprising all 120 included studies plus six representative boundary exclusions, after completing a randomized 20-record calibration exercise. Their agreement reached 97.6 percent, or 123 of 126 judgments, with a Cohen’s kappa of 0.788, a value generally interpreted as substantial agreement. Six records were excluded by both reviewers, and the three discordant judgments were resolved through joint full-record review. The authors are careful to note that this coefficient describes only the eligibility audit and not the earlier title and abstract stage or the taxonomy coding, an honest boundary on what the statistic can claim.

The coding scheme itself reflects the specific character of television as a recommendation surface. Each study received one primary technical category and one of six mutually exclusive primary functional goals, with secondary technical labels reserved for hybrid methods so that overlap between paradigms could be analyzed without distorting the primary counts. This design choice matters because much contemporary recommender research is deliberately hybrid, combining sequence models with graph structures or blending collaborative signals with content features, and single-label taxonomies often obscure that reality. Quality was appraised through separate domains rather than a composite score, covering design, data transparency, reproducibility, evaluation rigor, external validation, and deployment, which allows readers to see precisely where the evidence is strong and where it is thin.

The scope definition is equally deliberate. TV-specific studies directly examined smart TV, IPTV, over-the-top, video-on-demand, live-TV, channel, or programme recommendation. Enabling or contextual studies qualified only when they addressed a prespecified mechanism that transfers directly to television, such as shared accounts, long-form content representation, temporal availability, privacy, or on-device inference. Crucially, enabling studies were not treated as direct evidence of effectiveness on television platforms, a distinction that protects the review from one of the most common weaknesses in cross-domain literature synthesis, namely the silent importation of results from adjacent domains as if they had been validated on the target platform.

The search strategy was frozen before execution. Six domain phrases were crossed with six prespecified challenge blocks to form 36 conceptual variants, and all 216 database-native search strings were recorded alongside the per-database result counts. The formal sources were Google Scholar, ACM Digital Library, Web of Science, IEEE Xplore, Scopus, and Engineering Village, searched in April and May 2026. Primary-discovery attribution among the 120 included records was distributed unevenly across these databases, with IEEE Xplore contributing 39 records, Google Scholar 27, ACM Digital Library 20, Web of Science 18, Engineering Village 10, and Scopus 6. This attribution pattern is a useful diagnostic for other reviewers, since it shows how much the choice of database shapes the composition of a final corpus even when the underlying literature is the same.

The technical landscape that emerges from the map spans deep learning, sequence analysis, graph theory, shared-account modeling, and multimodal approaches. Each of these families addresses a distinct television constraint. Sequence models are naturally suited to viewing histories that unfold over days and weeks, where the order of consumption carries information that set-based models discard. Graph methods can represent relationships among programmes, channels, genres, and viewers, which is valuable when explicit feedback is scarce and structure must substitute for ratings. Multimodal approaches exploit the fact that television content is intrinsically rich in audiovisual signals, metadata, and text, allowing models to characterize a programme even when few users have interacted with it.

Shared accounts remain perhaps the most distinctive challenge. When a single profile aggregates the tastes of several household members, standard personalization assumptions break down, and naive averaging can produce recommendations that satisfy no one. Research on this problem, much of it developed in adjacent domains, offers techniques for separating latent users within an account or for generating recommendations that serve a group rather than an individual. The review’s decision to include such work only under an explicit transfer rationale, while flagging it as contextual rather than direct evidence, gives practitioners a clearer picture of how mature each proposed solution actually is for the living room setting.

Passive viewing compounds the feedback problem. Remote controls offer little interaction, explicit ratings are unusual, and much of what a television observes is co-viewing rather than deliberate choice. Signals that mobile platforms take for granted, such as taps, scrolls, and dwell time on a screen, have weaker or different analogues on television, where a programme may run for an hour in the background of household life. This scarcity of direct feedback pushes the field toward implicit signals, temporal patterns, and content-side features, and it raises the stakes for evaluation design, because a model tuned to noisy implicit feedback can look strong offline and fail in deployment.

Long-form and live content introduce further complications. A film or a multi-season series is semantically rich in a way that a short video clip is not, and live schedules impose temporal availability constraints that on-demand catalogs do not. A recommendation that arrives after a programme has aired is worthless in the linear context, so models must reason about what is available now, not merely what is similar to past preferences. The review’s inclusion of temporal availability as a qualifying transfer mechanism acknowledges that this scheduling dimension is not incidental but central to the television recommendation problem.

The evidence-reporting profile across data transparency, reproducibility, evaluation design, external validation, and deployment validation is likely to be the review’s most consequential contribution for practitioners. In many areas of recommender research, the gap between offline benchmark performance and deployed-system outcomes is well documented, and the review’s separation of external validation and deployment as distinct reporting domains makes that gap visible in the television literature specifically. Studies validated only on public datasets, for example, can be distinguished from those tested on proprietary platform data or in production, allowing readers to weigh claims accordingly.

The review also positions itself against prior surveys of recommender methods, cold-start strategies, graph models, privacy, and multimodal learning, arguing that their conclusions do not transfer automatically to television. This claim is defensible given the domain’s constraints, and the structured evidence map provides a mechanism for testing it, since cross-classifying technical paradigms against functional goals reveals which combinations have accumulated direct television evidence and which rest on extrapolation from other domains.

For the research community, the review identifies several open territories. Industrial systems are noted as a promising but weakly validated direction, reflecting the perennial difficulty that production platforms publish less about their algorithms than academia would like. Emerging uses of large language models fall into the same category, attracting substantial attention but lacking the deployment evidence that would establish practical value on television platforms. The authors’ framework, by separating established findings from weakly validated directions, gives both researchers and platform engineers a shared vocabulary for discussing maturity rather than novelty.

The time window of the corpus, spanning 2015 through 2025, captures a decade in which the field moved from matrix-factorization-era methods through deep sequence models to graph neural networks and, at the frontier, language-model-based approaches. Mapping that decade within a single consistent taxonomy makes it possible to see paradigm shifts rather than simply enumerate papers. The retention of two older foundational references as background, excluded from the evidence map proper, further sharpens the temporal boundary of the analysis.

Ultimately, the review’s value lies less in any single finding than in its infrastructure. The frozen queries, the recorded per-database counts, the explicit inclusion and exclusion rules, the dual-reviewer audit, and the domain-separated quality appraisal together form a template that other domain-specific evidence maps could adopt. For a field where claims often rest on heterogeneous benchmarks and unpublished industrial experience, that kind of methodological transparency is itself a form of scientific progress, and it sets a standard against which future television recommendation research can be measured.

Subject of Research: Algorithmic advances in smart TV content recommendation: a structured evidence-mapping review

Article Title: Algorithmic advances in smart TV content recommendation: a structured evidence-mapping review

Article References: Chen, Z., He, J., Gong, Y., Liang, J., & Li, J. (2026). Algorithmic advances in smart TV content recommendation: a structured evidence-mapping review. Frontiers in Big Data, 9, Article 1927033. https://doi.org/10.3389/fdata.2026.1927033

Image Credits: AI Generated

DOI: 10.3389/fdata.2026.1927033

Keywords: recommendation systems, television, machine learning, Algorithmic, advances, smart, content, recommendation, structured, evidence-mapping, review, scientific research