Informfully Recommenders Framework

• Informfully Recommenders is an open-source framework that integrates normative targets and diversity-aware methods into every recommendation pipeline stage.
• It extends the Cornac ecosystem by incorporating data augmentation, flexible model training, re-ranking constraints, and tailored evaluation metrics.
• By coupling dynamic intra-session adaptation with interactive, interpretable profiles, the framework advances norm-aware, context-sensitive recommendation research.

Informfully Recommenders is an open-source, end-to-end reproducibility framework that extends the Cornac ecosystem to make diversity-aware, norm-aware recommendation research implementable and repeatable across pre-processing, in-processing, post-processing, and evaluation (Heitz et al., 18 Aug 2025). The surrounding literature also uses the same expression for recommender systems that are information-rich, transparent, interactive, or explicitly grounded in user-facing evidence such as metadata, review text, snippets, graphs, and factual language-model outputs. This suggests a broader research program in which recommendation is treated not only as relevance estimation, but also as the design of recommendation pipelines that expose reasons, support user steering, and operationalize normative targets such as diversity, representation, and balanced exposure (Heitz et al., 18 Aug 2025, Vrijenhoek et al., 2020, Sun, 2024).

1. Conceptual scope and normative foundations

In its narrowest and most literal sense, Informfully Recommenders denotes a normative reproducibility framework for diversity-aware intra-session recommendation in the news domain (Heitz et al., 18 Aug 2025). In that framework, “norm-aware” follows a normative perspective that operationalizes societal norms and values—such as exposure diversity, viewpoint diversity, or balanced representation—directly in recommender design, through data augmentation, model objectives, re-ranking constraints, and evaluation metrics (Heitz et al., 18 Aug 2025). The framework is therefore not only a model library, but an experimental substrate for studying how normative assumptions propagate through the full recommender pipeline.

The normative basis for this agenda is articulated earlier in work on mission-driven news recommendation. That literature argues that recommender systems can either mitigate or exacerbate filter bubbles, selective exposure, and misinformation dynamics by channeling attention, and that common KPIs such as click-through rate and dwell time are short-term proxies rather than direct measures of whether citizens receive diverse and democratically important information (Vrijenhoek et al., 2020). It further grounds diversity in four democratic models—liberal, participatory, deliberative, and critical—each implying different targets for topical diversity, viewpoint diversity, style and complexity, minority voice representation, fragmentation of the public sphere, and activation (Vrijenhoek et al., 2020).

A second conceptual foundation comes from the critique of static collaborative filtering. The paper “Beyond Collaborative Filtering: A Relook at Task Formulation in Recommender Systems” argues that recommender systems are too often conceptualized as predicting missing values in a static user-item interaction matrix, rather than predicting a user’s decision on the next interaction within a dynamic, changing, and application-specific context (Sun, 2024). In that formulation, the relevant object is not merely $R_{ui}$, but a contextual next-decision model such as

$$P(d_{t+1} = i \mid u, \mathcal{H}_t(u), c_t, \mathcal{C}\mathrm{and}_t, \pi_t),$$

where user history, session context, candidate exposure, and position all matter (Sun, 2024). Informfully Recommenders, in the broad sense suggested by the literature, fits this shift toward decision-time context, explicit normative targets, and user-visible rationale.

2. Framework architecture and pipeline design

The Informfully Recommenders framework restructures Cornac’s workflow into four stages, each with two customizable steps, while keeping backward compatibility by making new features optional (Heitz et al., 18 Aug 2025).

Stage	Steps	Main functions
Pre-processing	Data loading/curation; augmentation/validation	userHistory, articlePool, sentiment, named entities, parties, complexity, clusters, categories
In-processing	Flexible splitting; model training/inference	ENMF, LSTUR, NPA, NRMS, VAE, PLD, EPD, 3, RWE-D, D-RDW
Post-processing	Static re-ranking; dynamic re-ranking	G-KL, PM-2, MMR, DAP with POS and ATT simulators
Evaluation	Metrics; visualization	AUC, ILD/EILD, Gini, $\alpha$-nDCG, binomial diversity, RADio, Informfully platform

Inter-stage communication is file-based: Item Pool $\rightarrow$ Candidate List $\rightarrow$ Recommendation List (Heitz et al., 18 Aug 2025). A Save State Manager allows reuse of intermediate artifacts so that only selected stages need to be rerun when experimental conditions change (Heitz et al., 18 Aug 2025). Cornac integration is implemented by extending base_method.py and the static re-ranking evaluator to accept articlePool, and by adding userHistory as an optional parameter for temporal and contextual attributes (Heitz et al., 18 Aug 2025).

The pre-processing stage is especially important because normative constraints are expressed over item attributes. Informfully Recommenders therefore includes multi-language augmentation pipelines for EB-NeRD, MIND, and NeMig, with sentiment analysis by XLM-RoBERTa, named entity recognition via spaCy, political actor and party mapping through Wikidata, text complexity via Textstat, story and event clustering with NetworkX, and category classification by BART zero-shot classification (Heitz et al., 18 Aug 2025). The output of this stage is not yet “norm enforcement”; rather, it is a norm-indexed representation space in which later models and re-rankers can act.

A central device for such control is the Normative Target Distribution, or NTD, described as a JSON-like specification that lists target attributes and desired proportions or quotas, for example over party labels or sentiment bins (Heitz et al., 18 Aug 2025). This target distribution is then consumed by models such as D-RDW and by re-rankers such as G-KL and PM-2.

3. Diversity metrics and normative target distributions

Informfully Recommenders implements both traditional diversity metrics and normative diversity metrics (Heitz et al., 18 Aug 2025). On the traditional side, the framework includes intra-list distance and expected intra-list distance, Gini coefficients, $\alpha$-nDCG, and binomial diversity (Heitz et al., 18 Aug 2025). A standard intra-list distance definition used in the framework is

$$\mathrm{ILD}(L) = \frac{2}{|L|(|L|-1)} \sum_{i<j} \big(1 - \mathrm{sim}(i,j)\big),$$

with similarity computed, for example, over one-hot vectors of attributes such as party, sentiment, or category (Heitz et al., 18 Aug 2025).

On the normative side, the framework adopts the five RADio metrics—calibration, fragmentation, activation, representation, and alternative voices—originally proposed as a bridge between democratic theory and recommender evaluation (Heitz et al., 18 Aug 2025, Vrijenhoek et al., 2020). In the earlier normative formulation, calibration and representation are expressed as Kullback–Leibler divergences:

$$\mathrm{Calibration}_{(r,q)} = \sum_c r(c|u)\log \frac{r(c|u)}{\tilde q(c|u)},$$

$$\mathrm{Representation}_{(p,q)} = \sum_o p(o)\log \frac{p(o)}{\tilde q(o|u)},$$

while fragmentation is defined as one minus Rank-Biased Overlap:

$$\mathrm{Fragmentation}(Q_1,Q_2,s) = 1 - \mathrm{RBO}(Q_1,Q_2,s),$$

and alternative voices are operationalized by a protected-versus-unprotected exposure ratio (Vrijenhoek et al., 2020). Informfully Recommenders implements the same five metrics with Jensen–Shannon divergence and without rank weights in the reported experiments:

$$P(d_{t+1} = i \mid u, \mathcal{H}_t(u), c_t, \mathcal{C}\mathrm{and}_t, \pi_t),$$0

(Heitz et al., 18 Aug 2025).

This difference is not merely notational. The earlier framework ties each target distribution to a model of democracy: reflective targets for participatory systems, equal targets for deliberative systems, and inverse targets for critical systems (Vrijenhoek et al., 2020). Informfully Recommenders operationalizes these commitments through the NTD. A target over parties, sentiment, or other attributes becomes a concrete quota structure that can be enforced by in-processing models or post-processing re-rankers (Heitz et al., 18 Aug 2025). The result is a pipeline in which diversity is not treated as a single scalar objective, but as a set of divergences between observed recommendation distributions and explicitly chosen normative targets.

4. Intra-session adaptation, dynamic optimization, and reinforcement-learning connections

A defining feature of Informfully Recommenders is its explicit treatment of intra-session dynamics (Heitz et al., 18 Aug 2025). Static re-rankers include Greedy-KL, PM-2, and Maximal Marginal Relevance. Greedy-KL greedily selects items to make the recommendation distribution $$P(d_{t+1} = i \mid u, \mathcal{H}_t(u), c_t, \mathcal{C}\mathrm{and}_t, \pi_t),$$1 match a target distribution $$P(d_{t+1} = i \mid u, \mathcal{H}_t(u), c_t, \mathcal{C}\mathrm{and}_t, \pi_t),$$2 by minimizing

$$P(d_{t+1} = i \mid u, \mathcal{H}_t(u), c_t, \mathcal{C}\mathrm{and}_t, \pi_t),$$3

while MMR applies the familiar relevance–redundancy trade-off

$$P(d_{t+1} = i \mid u, \mathcal{H}_t(u), c_t, \mathcal{C}\mathrm{and}_t, \pi_t),$$4

(Heitz et al., 18 Aug 2025). PM-2 instead follows an election-based proportional diversification scheme over intents such as sentiment bins or party labels (Heitz et al., 18 Aug 2025).

Dynamic intra-session re-ranking is implemented through Dynamic Attribute Penalization, or DAP, coupled to a user simulator (Heitz et al., 18 Aug 2025). After each click within a session, future items sharing clicked attributes are penalized:

$$P(d_{t+1} = i \mid u, \mathcal{H}_t(u), c_t, \mathcal{C}\mathrm{and}_t, \pi_t),$$5

with optional position and novelty terms added afterward (Heitz et al., 18 Aug 2025). The default simulator behaviors are POS, which models position-biased clicking, and ATT, which models attribute-biased clicking (Heitz et al., 18 Aug 2025). This makes the framework suitable for offline studies of repeated exposure, evolving session histories, and policy interventions.

The broader recommender literature provides adjacent formalizations for such dynamics. RecoGym formulates recommendation as a Markov Decision Process with alternating organic and bandit events and an objective

$$P(d_{t+1} = i \mid u, \mathcal{H}_t(u), c_t, \mathcal{C}\mathrm{and}_t, \pi_t),$$6

where reward is click-based in the baseline setting (Rohde et al., 2018). Likewise, task-reformulation work emphasizes that recommendation should model next decisions under candidate exposure and context rather than static matrix completion (Sun, 2024). A plausible implication is that Informfully Recommenders can be used not only to compare re-rankers, but also to connect diversity-aware intra-session experimentation with policy-learning and simulation-based evaluation.

A further extension appears in addiction-aware social-media recommendation, where the expression “Informfully Recommenders” is explicitly used for wellbeing-preserving and economically sustainable systems (Bolis et al., 25 Feb 2025). There the recommender objective becomes

$$P(d_{t+1} = i \mid u, \mathcal{H}_t(u), c_t, \mathcal{C}\mathrm{and}_t, \pi_t),$$7

subject to minimum utility constraints, so that heavy use and aftereffects are treated as explicit costs (Bolis et al., 25 Feb 2025). This suggests a widening of the Informfully agenda from diversity control toward broader norm-aware sequential optimization.

5. Explainability, grounding, and informative interaction

A second major strand associated with Informfully Recommenders treats informfulness as a property of user modeling and system interaction. One line of work places user profiles directly in interpretable feature spaces. TEASER learns user profiles in item-metadata space with

$$P(d_{t+1} = i \mid u, \mathcal{H}_t(u), c_t, \mathcal{C}\mathrm{and}_t, \pi_t),$$8

so that recommendations are a linear function of interpretable metadata and users can directly tweak profile weights through

$$P(d_{t+1} = i \mid u, \mathcal{H}_t(u), c_t, \mathcal{C}\mathrm{and}_t, \pi_t),$$9

(Pauw et al., 2022). The model is therefore described as inherently transparent and explainable, and its linear scoring supports interactive recommendation (Pauw et al., 2022). In the Publiq deployment, simulated feedback yielded nDCG@100 gains of $\alpha$0 with one boosted tag and $\alpha$1 with two boosted tags (Pauw et al., 2022).

A related text-based line focuses on sparse users with informative reviews. CUP constructs concise user profiles from review text under a $\alpha$2-token budget and feeds only these into transformer-based recommenders (Torbati et al., 2023). On AM-1K-rich, CUP$\alpha$3 reaches NDCG@5 $\alpha$4 and P@1 $\alpha$5; on GR-1K-rich it reaches NDCG@5 $\alpha$6 and P@1 $\alpha$7 (Torbati et al., 2023). RecCoT instead uses a teacher LLM to generate grounded chain-of-thought for reviews, aligns a BERT student to those CoTs, and caches the resulting embeddings for downstream recommendation; on Amazon Fashion, RecCoT achieves MSE $\alpha$8 versus $\alpha$9 for LoID (Yang et al., 26 Jun 2025). Dynamic Review-based Recommenders extend this idea temporally, coupling ratings and review language in a causal, time-aware model through user and item LSTMs and a LLM conditioned on past state (Cvejoski et al., 2021).

Conversational and language-model-based systems push informfulness further into dialogue and rationale generation. SnipRec decomposes reviews and user utterances into atomic snippets, retrieves evidence through dense retrieval and NLI filtering, and accumulates item scores with a sentiment-aware reciprocal-rank-fusion rule (Sun et al., 2024). After five turns on Yelp restaurants, GPT-4o-mini SnipRec reaches Hits@10 $\rightarrow$0, and SnipRec plus expansion reaches Hits@10 $\rightarrow$1 (Sun et al., 2024). P4LM generates factually grounded, personalized recommendation narratives by conditioning a LLM on collaborative-filtering embeddings and optimizing a joint reward over precision, appeal, personalization, and preference relevance (Jeong et al., 2023). On a held-out test of $\rightarrow$2 user–movie pairs, P4LM reports Precision $\rightarrow$3, Personalization $\rightarrow$4, Appeal $\rightarrow$5, and Preference Relevance $\rightarrow$6 (Jeong et al., 2023).

Graph-grounded personalization provides yet another explicit Informfully blueprint. GraphRAG-IRL combines graph-grounded feature construction, Maximum Entropy IRL pre-ranking, and persona-guided LLM re-ranking (Liang et al., 21 Apr 2026). IRL-MLP with GraphRAG improves NDCG@10 by $\rightarrow$7 on MovieLens and $\rightarrow$8 on KuaiRand over supervised baselines, and persona-guided LLM fusion yields up to $\rightarrow$9 NDCG@10 improvement over the IRL-only baseline on MovieLens ml-1m (Liang et al., 21 Apr 2026). The common pattern across these systems is not a single architecture, but the use of explicitly inspectable intermediate representations—metadata weights, snippets, graph features, or CoTs—that make recommendation more legible and more controllable.

6. Empirical profile, limitations, and research directions

The empirical record associated with Informfully Recommenders is heterogeneous because the term now spans a framework paper and several neighboring system designs. The following reported outcomes illustrate that diversity-aware control, interpretable profiles, grounded dialogue, and graph-grounded reranking are all active components of the area.

System	Setting	Reported outcome
Informfully Recommenders	EB-NeRD, MIND, NeMig news experiments	D-RDW and PLD/EPD are top or near-top on EB-NeRD; dynamic re-ranking atop NPA/NRMS often gives highest activation and representation on MIND (Heitz et al., 18 Aug 2025)
GraphRAG-IRL	MovieLens, KuaiRand	NDCG@10 improves by $\rightarrow$0 and $\rightarrow$1; persona-guided fusion adds up to $\rightarrow$2 on ml-1m (Liang et al., 21 Apr 2026)
SnipRec	Yelp restaurants	Hits@10 $\rightarrow$3 for SnipRec and $\rightarrow$4 with expansion after five turns (Sun et al., 2024)
CUP	Sparse book recommendation	NDCG@5 $\rightarrow$5 on AM-1K-rich and $\rightarrow$6 on GR-1K-rich (Torbati et al., 2023)
RecCoT	Amazon review domains	Fashion MSE $\rightarrow$7, below LoID $\rightarrow$8 (Yang et al., 26 Jun 2025)
TEASER	Interactive recommendation	Publiq nDCG@100 gains of $\rightarrow$9 with one boosted tag and $\alpha$0 with two boosted tags (Pauw et al., 2022)

Several limitations recur. Informfully Recommenders itself reports offline benchmarking only, did not use rank-aware RADio variants in the reported experiments, and did not recompute AUC on re-ranked lists, although future releases are expected to add configurable accuracy evaluation for ranked outputs (Heitz et al., 18 Aug 2025). It also notes augmentation bias risks in NER, sentiment classification, and party mapping, and emphasizes that norm choices are stakeholder-specific rather than universal (Heitz et al., 18 Aug 2025). Related systems expose complementary weaknesses: TEASER is limited by metadata coverage and pure linearity (Pauw et al., 2022); GraphRAG-IRL depends on graph quality, prompt stability, and pre-ranking recall (Liang et al., 21 Apr 2026); SnipRec is evaluated only on Yelp restaurants and remains sensitive to review quality, clustering, and LLM hallucination (Sun et al., 2024); P4LM is single-turn, shows reward trade-offs, and reports a dip in preference relevance relative to other gains (Jeong et al., 2023).

Current directions point toward richer normative control and stronger grounding. In the framework paper, future work includes rank-aware RADio, fair exposure metrics, and NTD extensions beyond parties and sentiment to categories, geography, sources, or temporal recency (Heitz et al., 18 Aug 2025). In adjacent Informfully formulations, balanced usage, aftereffects, and wellbeing preservation become optimization targets alongside engagement (Bolis et al., 25 Feb 2025). This suggests that Informfully Recommenders is evolving from a diversity-aware news evaluation framework into a broader technical vocabulary for recommendation systems that are norm-aware, dynamically adaptive, and explicit about the informational grounds on which they recommend.