Predicting the why.

Tracking what users do — without understanding why.

AI systems across industries process billions of user events daily but remain structurally unable to interpret the psychological drivers behind user actions. They track what users do but fail to understand why.

This paper proposes the ABSD Framework, a four-layer conceptual model mapping Action, Behaviour, State, and Drive to bridge this gap. The research is anchored empirically in Customer Experience as the primary domain, with Education as a secondary domain, and validates the framework through expert interviews, a practitioner survey, and case analysis of existing AI systems.

The ABSD acronym represents the four layers of inference: Action (what the user does), Behaviour (the patterns in how they do it), State (what they feel or experience), and Drive (the underlying psychological motivation). The framework synthesises four established but siloed research streams — Self-Determination Theory from motivational psychology, Affective Computing from human-computer interaction, Just-in-Time Adaptive Intervention from health behaviour research, and Learning Analytics from educational data science — into a unified, actionable model for AI system design.

The project uses a mixed-methods approach combining literature synthesis, semi-structured expert interviews, a practitioner survey targeting 30 to 50 respondents, and case analysis. It will produce three deliverables: a validated conceptual framework, design heuristics for state-aware AI systems, and a practitioner-oriented implementation guide. While the empirical work focuses on customer experience and education, the framework is designed for transferability to finance, healthcare, and other domains where AI mediates human decision-making.

AI sees everything. It understands almost nothing.

Here is the uncomfortable truth about AI systems today: they see everything and understand almost nothing.

Netflix processes 500 billion events per day (Netflix Engineering, 2015). Spotify logs over 500 billion events daily (Spotify Engineering, 2019). The pattern extends across industries. The volume of interaction data has never been greater.

But when a student pauses for thirty seconds before answering a math problem, the system records the pause and moves on. It cannot tell confusion from careful thinking, anxiety from strategic deliberation. When a customer abandons a shopping cart, the platform logs the event but has no idea whether the person got interrupted, felt sticker shock, or simply could not decide.

This is the core problem. AI systems track what users do. They miss why users do it.

The consequences show up across industries. Baker et al. (2004) found that disengaged behaviours predict dropout days or weeks before grades decline. Most platforms only react after performance drops, missing the window for early intervention. In customer experience, churn prediction models flag at-risk accounts based on declining engagement. By the time they sound the alarm, frustration has usually hardened into a decision to leave. Zorfas and Leemon (2016) showed that emotional connection predicts customer lifetime value better than satisfaction scores, yet most CRM systems track transactions, not emotions. McKinsey (2021) found that companies excelling at personalization generate 40% more revenue than average players.

Why does this gap persist? Three structural reasons.

First, actions are easy to measure while internal states are not. Clickstream data and timestamps get captured automatically. Emotional states and motivation do not show up in server logs. They must be inferred from behavioural patterns, which requires theory and method that most AI development skips. As Siemens and Baker (2012) noted, the field captures behavioural traces but lacks frameworks for interpreting what those traces mean psychologically.

Second, the research that could help is scattered across disciplines that rarely talk to each other. Psychologists study motivation through Self-Determination Theory (Deci & Ryan, 2000; Ryan & Deci, 2017) but lack computational tools for real-time application. Computer scientists build sophisticated pattern recognition but lack theoretical grounding in what patterns mean for human experience. A recent review found minimal cross-citation between these fields despite clear conceptual overlap (Gomez et al., 2024).

Third, business incentives favour optimization over understanding. When engagement metrics become KPIs, systems get optimized for time-on-platform rather than user goal achievement. Forrester (2022) found that 73% of companies struggle to use behavioural data for anything beyond basic segmentation.

The result: AI that reacts instead of anticipates. Systems that optimize for clicks instead of outcomes. Tools that personalize surfaces while missing depths.

The systems that understand users deeply will outperform those that only track them.

AI is becoming the primary interface between organizations and users. Chatbots handle 85% of customer interactions at leading companies (Gartner, 2023). Recommendation engines shape what 2 billion people see daily on YouTube and TikTok (Covington et al., 2016). Adaptive learning platforms guide millions of students.

The evidence is substantial across domains.

Recommendations

Wang et al. (2025) demonstrated that recommender systems incorporating predicted user intent outperform black-box models relying purely on behavioural history. Their ISRec framework was evaluated on YouTube, the largest video recommendation platform, where structuring recommendations around inferred intent significantly improved daily active users and overall user enjoyment, representing one of the largest metric gains observed in recent YouTube experiments. The insight is that understanding user goals, even when predicted rather than explicitly stated, beats adding more behavioural data.

Health behaviour

Just-in-Time Adaptive Interventions show that detecting the right moment to intervene dramatically improves outcomes. Nahum-Shani et al. (2018) established the framework for timing interventions to vulnerability and receptivity states. Meta-analyses confirm JITAIs outperform static interventions when timing matches user states (Wang & Miller, 2020). Klasnja et al. (2015) found that timing based on inferred receptivity significantly increased effectiveness compared to random timing.

Education

Research on affective dynamics shows that confusion can resolve into engagement or spiral into frustration, depending on system response (D'Mello & Graesser, 2012). Pekrun's Control-Value Theory demonstrates that achievement emotions directly shape learning outcomes. Studies using the Open University dataset show clickstream patterns predict dropout with 78–84% accuracy — but the predictive features are behavioural patterns, not raw actions.

Financial services

Investors exhibiting hesitation and rapid switching show higher likelihood of anxiety-driven poor decisions (Barber & Odean, 2013). Platforms that detect these patterns could intervene before costly mistakes occur.

What is missing is not more algorithms. It is a conceptual framework telling practitioners what signals to look for, what states those signals indicate, and how to respond.

A conceptual architecture, not another model.

This study proposes a conceptual framework for understanding human action in AI systems, with empirical validation anchored in customer experience as the primary domain and education as a secondary domain. The framework, called ABSD (Action-Behaviour-State-Drive), maps the journey from observable actions to inferred psychological drives.

The project scope is explicitly conceptual rather than empirical. The study does not attempt to build or validate predictive models for psychological state inference. Rather, it develops a theoretical architecture that organizes existing research into a coherent framework, identifies the inferential relationships between observable behaviours and psychological constructs, and provides a vocabulary for practitioners working across domains.

The project will produce three deliverables:

• A conceptual framework with clear layers, defined constructs, and mapped behavioural signals grounded in established research.
• Design heuristics that AI practitioners can apply when building systems intended to respond to user states rather than merely tracking user actions.
• A practitioner-oriented implementation guide for organizations wanting to move beyond action-level analytics toward more sophisticated user understanding.

Three questions.

RQ1 — Signal taxonomy

Which categories of behavioural signals appear consistently across domains according to existing literature, practitioner perception, and survey-reported practice — and how can these signals be structured into a coherent detection framework?

RQ2 — Signal-to-state inference

How do practitioners interpret the relationships between action patterns, behavioural sequences, inferred states, and underlying motivational drivers in real-world AI systems?

RQ3 — Design heuristics

What design heuristics can support the development of AI systems that move beyond action tracking toward context-aware and state-responsive interaction?

Organizational, technical, governance, and ethical constraints are addressed as discussion topics in the Implementation and Ethics sections rather than as a separate empirical research question. This focuses empirical work on framework development and validation while preserving treatment of implementation realities.

Four research traditions. One integration gap.

Each tradition has developed rigorous findings within its domain, with limited integration across fields. The contribution here is synthesis into a unified conceptual framework.

Research tradition gap analysis

Table 1 — Research Tradition Gap Analysis and ABSD Integration

Self-Determination Theory and human motivation

SDT, developed by Deci and Ryan over four decades, identifies three basic psychological needs that drive human behaviour: autonomy (the need to feel volitional control), competence (the need to feel effective), and relatedness (the need to feel connected). When these needs are satisfied, people experience intrinsic motivation and sustained engagement. When frustrated, they disengage or shift to controlled, externally driven motivation that produces lower quality outcomes.

The theory has been validated across thousands of studies. Meta-analyses confirm that need satisfaction predicts engagement, persistence, wellbeing, and performance across domains. Recent applications to AI show direct relevance: De Vreede et al. (2021) found that all three psychological needs significantly predicted user satisfaction and engagement in an AI-assisted decision-making system. Peters et al. (2018) demonstrated that gamification features supporting autonomy and competence increased motivation, while those undermining these needs backfired.

Affective Computing and state detection

Affective Computing, pioneered by Picard (1997) at MIT, established that computational systems can recognize, interpret, and respond to human emotions. Emotion detection from text now achieves 70–79% accuracy in controlled settings. Speech-based recognition exceeds human baseline for certain categories. Multimodal systems approach robust real-world application.

Research demonstrates that cognitive-affective states like confusion, frustration, flow, boredom, and curiosity all influence how users interact with systems and whether they achieve their goals. D'Mello and Graesser (2012) showed that transitions between affective states during learning follow predictable dynamics: confusion can resolve into engagement through productive struggle or devolve into frustration. The system's response during the confusion window determines which trajectory unfolds. Ocumpaugh et al. (2014) developed reliable observational protocols for detecting these states from behavioural indicators alone — no physiological sensors required.

Just-in-Time Adaptive Intervention (JITAI)

JITAI research, developed primarily in mobile health contexts, provides the conceptual vocabulary for detecting behavioural patterns that signal opportunities for intervention. Key constructs include decision points, tailoring variables, proximal and distal outcomes. Critically, the framework distinguishes vulnerability states (when a person is at risk) from receptivity states (when a person is able to receive intervention). Optimal intervention requires detecting both.

The behavioural signals predicting these states have cross-domain applicability. Hesitation patterns, avoidance behaviours, acceleration through content, regression to earlier material, and persistence after failure all appear as predictive indicators. A student hesitating before a math problem and a patient hesitating before logging medication may be experiencing similar uncertainty states that similar intervention approaches could address.

Learning Analytics and action tracking

Learning Analytics emerged in the early 2010s as a field dedicated to measuring, collecting, analysing, and reporting data about learners. Studies using the Open University Learning Analytics Dataset achieve 78–84% accuracy predicting at-risk students from clickstream data. Predictive features include not just action counts but temporal patterns: spacing of study sessions, time-of-day variations, navigation sequences, and help-seeking behaviours.

However, the same tradition acknowledges its limitations. Clickstream captures what students do, not why. Ferguson (2012) noted that learning analytics risks optimizing for measurable proxies rather than actual learning. Wise and Shaffer (2015) argued that without theoretical frameworks connecting behaviours to underlying processes, analytics remains descriptive rather than explanatory.

The integration gap

Each tradition is well-established within its domain. SDT has thousands of empirical studies. Affective computing has mature detection technologies. JITAI has demonstrated intervention effectiveness. Learning analytics has proven that behavioural signals contain predictive information.

Yet integration across these fields remains rare. A recent systematic review found limited cross-citation between motivational psychology and affective computing, between JITAI and learning analytics, between SDT and AI system design (Gomez et al., 2024). Researchers in each field cite within their tradition but rarely build on adjacent work. This creates redundant discovery of similar phenomena using different vocabulary and missed opportunities for theoretical synthesis.

The ABSD Framework proposed here addresses exactly this integration gap.

Four layers from action to motivation.

The framework proposes four layers representing different levels of abstraction in understanding human action. Each layer generates the layer below it (from a human perspective) and must be inferred from the layer below it (from an AI perspective).

Conceptual architecture

Table 2 — ABSD Framework Architecture

The key insight is directional asymmetry. From a generative perspective, Drive shapes how people respond to States, States produce Behavioural patterns, Behaviours manifest as Actions. From a detection perspective, AI must work in reverse: observing Actions, detecting Behavioural patterns across time, inferring probable States, and over extended interaction modelling underlying Drives.

Current AI systems operate almost exclusively at the Action layer. The framework provides a conceptual roadmap for reaching the Behaviour, State, and Drive layers.

Behavioural signals: the bridge

The practical contribution is a taxonomy of behavioural signals that bridge observable action patterns and inferred psychological states.

Table 3 — Behavioural Signal Taxonomy

These signals are not invented for this framework. They appear across JITAI research, affective computing, and learning analytics. The contribution is organising established signals into a coherent conceptual framework with explicit layer mappings.

Contextual modifiers and signal disambiguation

The signal-to-state mappings represent base associations from the literature. In practice, behavioural signals are highly context-dependent. A 30-second pause before executing a complex financial trade may indicate cognitive load or deliberate prudence; the same pause before a simple checkout more plausibly indicates confusion or distraction. Narayanan and Georgiou (2013) established that behavioural expressions are characterised by "individual and contextual heterogeneity," meaning signal interpretation must account for the conditions under which signals are produced.

ABSD incorporates Contextual Modifiers — qualifying factors that sit alongside the inference path and modulate how signals are interpreted at every layer. Three categories:

Domain context

Industry, task type, and interaction norms specific to the system. Hesitation in a healthcare diagnostic tool carries different weight from the same pattern in e-commerce checkout.

User context

Individual baseline behaviour, historical patterns, and expertise level. A consistently methodical user should not have their deliberate pace interpreted as hesitation.

Temporal context

Time of day, stage in the user journey, recency of system changes. A spike in navigation switching immediately after a UI redesign signals disorientation, not indecisiveness.

This contextual weighting addresses the cold start problem. For new users lacking historical data, the system relies initially on Domain Context and cohort-level averages. As interaction volume grows, the system progressively weights User Context more heavily. Peters et al. (2024), analysing over 100 million sessions, demonstrated that incorporating context features captured non-redundant variance, improving predictive performance by 51% over behaviour-only models.

Recursive dynamics and non-linear feedback

While the ABSD layers are presented as a top-down diagnostic hierarchy, the operational reality is recursive. Psychological states sustained over time reshape the drives beneath them. Vansteenkiste and Ryan (2013) established that prolonged need frustration "evokes ill-being and increased vulnerabilities for defensiveness" — prolonged frustration at the State layer can actively degrade a user's sense of competence at the Drive layer, shifting motivation from intrinsic engagement to extrinsic compliance or disengagement entirely.

The Drive layer cannot be treated as a fixed baseline. A user who begins with strong intrinsic motivation may, after weeks of frustration-inducing interactions, shift to a fundamentally different motivational profile. The system must re-infer, not assume stability. Adanyin (2024) confirmed that AI-driven feedback mechanisms can simultaneously facilitate goal attainment and erode autonomy and well-being over time. Modelling the temporal dynamics of Drive erosion and recovery is positioned as a priority direction for longitudinal empirical research.

Mixed methods. Anchored in customer experience.

This study uses a qualitative methodology appropriate for conceptual framework development, combining literature synthesis, expert interviews, case analysis, and a complementary practitioner survey.

Validation dimensions

This research does not attempt to validate the predictive accuracy of signal-to-state mappings. That would require controlled empirical studies with ground-truth psychological measures, beyond the scope of an EMBA project. Validation focuses on three dimensions:

Conceptual validation

Does the framework provide a coherent, non-redundant, complete organization of the relevant constructs? Are the layer definitions clear and the relationships logically sound?

Theoretical grounding validation

Is each layer and signal mapping supported by established research? Does the framework accurately represent the source literature?

Practitioner utility validation

Do AI practitioners find the framework useful for thinking about user understanding? Does it help them identify gaps in their current systems?

The research combines interview depth with survey breadth, maintaining a strict boundary between qualitative generation and descriptive validation. The 5 expert interviews are formative qualitative research using thematic analysis (Braun & Clarke, 2006); they are not intended to provide statistical significance. A complementary practitioner survey (n=30–50) provides descriptive breadth and signal recognition frequencies that interviews alone cannot capture.

Research phases

Phase 1 — Literature synthesis

Integrative literature review across four research traditions to identify established findings, validated constructs, and documented behavioural signals. Substantially complete; findings inform the framework above.

Phase 2 — Expert interviews

Semi-structured interviews with 5 practitioners, primarily from customer experience technology and secondarily from educational technology. Customer experience is the primary empirical domain because it offers the broadest commercial applicability, the most accessible practitioner pool within the project timeline, and strong published literature on behavioural signal mapping. Interview participants will include both technical roles and product/design roles. The protocol is detailed in Annexure A.

Interviews will be conducted in a single round given the compressed timeline. Recorded, transcribed, and analysed using thematic analysis following Braun and Clarke's six-phase approach, with coding running in parallel with later interviews. Analysis will assess the three validation dimensions.

Phase 3 — Case analysis

Analysis of two existing AI systems through the ABSD lens — primary case from customer experience, secondary from educational technology. Cases will be selected to illustrate different positions on the spectrum from pure action-tracking to state-inference. They function as illustrative applications, not validation cases; broader cross-domain generalisability is positioned as a direction for future empirical work.

Phase 4 — Practitioner survey

An online survey targeting 30 to 50 practitioners in customer experience (primary) and educational technology (secondary). The instrument addresses four objectives: current-state benchmarking, signal recognition prevalence, signal-to-state mapping agreement, and framework utility scoring after a brief introduction. Detailed in Annexure B — or take the live version at Survey →.

Research timeline

The project follows a four-week execution window from 19 May to 14 June 2026. Activities run in parallel rather than strict sequence.

From conceptual framework to organizational design.

This section translates ABSD into organizational design choices. Rather than treating user understanding as a purely analytical problem, the framework highlights where product, data, model, and governance decisions must change if AI systems are expected to respond to behavioural context.

AI product lifecycle integration

ABSD suggests specific integration points across the AI product development lifecycle. In the requirements phase, product teams should explicitly specify which ABSD layers the system will address. The four layers represent different levels of technical maturity: Layers 1 and 2 are commercially viable today; Layer 3 represents the near-term frontier; Layer 4 is a longer-term architectural direction. Making layer targeting explicit prevents over-engineering before foundational data architecture is mature.

In the data architecture phase, telemetry should capture not just discrete events but temporal sequences that enable behaviour-level analysis. Many current logging systems optimize for event counts rather than pattern detection. In the model development phase, feature engineering should include behavioural pattern features (hesitation duration, switching frequency, regression patterns) alongside action-level features.

UX telemetry redesign

Current UX analytics focus on conversion funnels, feature usage counts, and session duration. ABSD suggests augmenting standard telemetry with behavioural pattern detection: hesitation detection (time between page load and first interaction; typing-and-deletion sequences), navigation pattern analysis (forward/backward movement ratios; breadth versus depth), engagement velocity (rate of progress against baseline), and help-seeking patterns. These additions require minimal additional instrumentation but enable behavioural-level analysis.

Governance implications

Organizations implementing state-level inference face governance questions that action-level tracking does not raise: transparency about what is being inferred (users may not realise hesitation patterns are being analysed), accuracy accountability (what happens when state inference is wrong?), and appropriate use boundaries (what interventions are acceptable based on inferred states?). ABSD helps organisations identify where these questions become relevant.

ABSD as an event-emitting architecture

Each ABSD layer can be understood as an event-emitting component within an event-driven architecture. The Action layer emits raw interaction events. The Behaviour layer emits pattern-shift alerts when usage sequences deviate from baselines. The State layer emits state-change triggers when inferred conditions cross thresholds. The Drive layer emits motivational context that shapes not just whether the system intervenes but how.

This framing connects ABSD to the emerging field of agentic AI workflows, where autonomous agents respond to triggers in real time. Jung et al. (2026) noted that agentic AI "often fails for lack of principled judgment about when, why, and whether to act." For example, a CX platform detecting a user in a frustration state with declining competence drive might trigger an agent to simplify the interface and surface a contextual walkthrough, rather than sending a generic re-engagement notification.

State-aware systems have power beyond action-tracking systems.

The ABSD framework raises ethical questions any organization implementing state-level inference must address. This section identifies considerations without prescribing solutions.

Psychological inference risks

Inferring psychological states from behavioural patterns involves inherent uncertainty. The same pattern (hesitation) might reflect different underlying states in different contexts. Systems should maintain epistemic humility, treating inferences as probabilistic hypotheses rather than certainties. Design should account for inference errors, ensuring that incorrect inferences do not cause harm.

Manipulation and autonomy

State-aware systems have greater capacity to influence user behaviour. An AI that detects vulnerability states could help users (offering support when confusion is detected) or exploit them (presenting persuasive content when resistance is low). SDT, which grounds the Drive layer, emphasises supporting user autonomy rather than controlling behaviour. Organisations should consider whether their state-aware interventions enhance or undermine autonomy and establish boundaries that prevent manipulative uses.

Inference calibration and ground truth

Because State and Drive layers rely on probabilistic inference, ABSD implementations must include mechanisms for continuous calibration against ground truth. Without such mechanisms, the system risks operating in a closed inferential loop where unvalidated assumptions compound over time. Lightweight micro-feedback mechanisms — contextual prompts triggered when the system detects high-confidence state changes — serve as deterministic corrective labels, anchoring probabilistic inferences against actual user self-reporting. The frequency must be calibrated to avoid survey fatigue.

Behavioural nudging governance

State-based nudging can be more personalised and effective, but also more intrusive. Organisations should consider transparency requirements (should users know when they are being nudged based on inferred states?), consent mechanisms (can users opt out of state-level inference?), and accountability structures (who is responsible when state-based nudges produce negative outcomes?).

Privacy and data use

Behavioural pattern analysis may reveal psychological information that users did not intend to disclose. Organisations should consider what psychological inferences their systems might produce, whether users understand this is possible, and how inferred information is stored, shared, and protected. Data minimization principles suggest limiting inference to what is necessary for legitimate purposes.

Theoretical clarity. Practical utility. Cross-domain transfer.

Theoretical contribution

The study contributes a synthesised conceptual framework connecting four research traditions often treated separately. ABSD offers a shared vocabulary for describing user understanding across layers of Action, Behaviour, State, and Drive, while clarifying how existing literature can be organised into a more coherent explanatory structure. Its value lies in conceptual integration, theoretical clarity, and a platform for future empirical testing — not in claims of predictive accuracy.

Practical contribution

The framework functions as a diagnostic and design aid. Organisations can use ABSD to assess the depth of their current analytics, identify where behavioural signals are available but underused, and guide decisions on telemetry, feature engineering, intervention design, and governance.

Cross-domain application

The framework is designed for transferability across domains. Within this project, empirical validation is anchored in customer experience as the primary domain, with educational technology as a secondary comparative domain. Finance and healthcare are retained as theoretical applicability domains.

• Education — earlier recognition of confusion and disengagement.
• Finance — hesitation and switching patterns that may precede poor choices.
• Healthcare — detection of receptivity and vulnerability windows.
• Customer Experience — interpreting frustration, intent, and loyalty signals beyond transaction history.

Framework quality criteria

The framework aims to meet five quality criteria that characterize useful conceptual contributions:

Simplicity

Memorable structure with clear vocabulary.

Universality

Applicable across contexts where AI interacts with humans.

Specificity

Generates concrete design guidance.

Groundedness

Rooted in established research.

Actionability

Informs real system design decisions.

The team and tools behind the work.

The research draws on the following resources:

• Academic mentorship through SP Jain faculty with expertise in technology strategy and organizational behaviour — Himmat Singh, Adjunct Professor.
• Industry mentorship from Akshith Sharma, Founder and CEO of Brained.app, providing practitioner perspective on AI system development and behavioural intelligence applications.
• Interview participant access through professional networks in technology, education, healthcare, and financial services sectors.
• Published research and case studies on AI systems in each domain.

No specialised technical resources or funding are required beyond standard research tools for interview recording, transcription, and qualitative analysis.

References.

1. Adanyin, A. (2024). AI-driven feedback loops in digital technologies: Psychological impacts on user behaviour and well-being. arXiv preprint arXiv:2411.09706.
2. Amatriain, X., & Basilico, J. (2015). Recommender systems in industry: A Netflix case study. In F. Ricci, L. Rokach, & B. Shapira (Eds.), Recommender systems handbook (2nd ed., pp. 385–419). Springer.
3. Baker, R. S., Corbett, A. T., Koedinger, K. R., & Wagner, A. Z. (2004). Off-task behavior in the cognitive tutor classroom: When students game the system. Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, 383–390.
4. Barber, B. M., & Odean, T. (2013). The behavior of individual investors. Handbook of the economics of finance (Vol. 2, pp. 1533–1570). Elsevier.
5. Braun, V., & Clarke, V. (2006). Using thematic analysis in psychology. Qualitative Research in Psychology, 3(2), 77–101.
6. Calvo, R. A., & D'Mello, S. (2010). Affect detection: An interdisciplinary review of models, methods, and their applications. IEEE Transactions on Affective Computing, 1(1), 18–37.
7. Covington, P., Adams, J., & Sargin, E. (2016). Deep neural networks for YouTube recommendations. Proceedings of the 10th ACM Conference on Recommender Systems, 191–198.
8. Deci, E. L., & Ryan, R. M. (2000). The what and why of goal pursuits. Psychological Inquiry, 11(4), 227–268.
9. De Vreede, T., Raghavan, M., & De Vreede, G.-J. (2021). Design foundations for AI assisted decision making. HICSS-54, 166–175.
10. D'Mello, S., & Graesser, A. (2012). Dynamics of affective states during complex learning. Learning and Instruction, 22(2), 145–157.
11. Ferguson, R. (2012). Learning analytics: Drivers, developments and challenges. International Journal of Technology Enhanced Learning, 4(5/6), 304–317.
12. Forrester Research. (2022). The state of customer analytics 2022. Forrester Research, Inc.
13. Gartner. (2023). Gartner predicts conversational AI will reduce contact center agent labor costs by $80 billion in 2026.
14. Gomez, C., Cho, S. M., Ke, S., Huang, C. M., & Unberath, M. (2024). Human-AI collaboration is not very collaborative yet. Frontiers in Computer Science, 6, Article 1521066.
15. Hardeman, W., Houghton, J., Lane, K., Jones, A., & Naughton, F. (2019). A systematic review of JITAIs to promote physical activity. IJBNPA, 16(1), Article 31.
16. He, L. (2026). Dynamic personalization through continuous feedback loops in interactive AI systems. arXiv preprint arXiv:2602.23376.
17. Jannach, D., Zanker, M., Felfernig, A., & Friedrich, G. (2010). Recommender systems: An introduction. Cambridge University Press.
18. Jung, S., Yoon, D., Koh, S. G., Kim, Y. H., Ahn, Y., & Park, S. (2026). When should an AI act? A human-centered model of scene, context, and behavior for agentic AI design. arXiv preprint arXiv:2602.22814.
19. Klasnja, P., Hekler, E. B., Shiffman, S., Boruvka, A., Almirall, D., Tewari, A., & Murphy, S. A. (2015). Microrandomized trials. Health Psychology, 34(S), 1220–1228.
20. Kuzilek, J., Hlosta, M., & Zdrahal, Z. (2017). Open University Learning Analytics dataset. Scientific Data, 4, Article 170171.
21. Li, Q., Baker, R., & Warschauer, M. (2020). Using clickstream data to measure self-regulated learning. The Internet and Higher Education, 45, Article 100727.
22. McKinsey & Company. (2021). The value of getting personalization right or wrong is multiplying.
23. Nahum-Shani, I., Hekler, E. B., & Spruijt-Metz, D. (2015). Building health behavior models to guide JITAIs. Health Psychology, 34(S), 1209–1219.
24. Nahum-Shani, I., Smith, S. N., Spring, B. J., Collins, L. M., Witkiewitz, K., Tewari, A., & Murphy, S. A. (2018). Just-in-time adaptive interventions (JITAIs) in mobile health. Annals of Behavioral Medicine, 52(6), 446–462.
25. Narayanan, S., & Georgiou, P. G. (2013). Behavioral signal processing. Proceedings of the IEEE, 101(5), 1203–1233.
26. Netflix Engineering. (2015). Evolution of the Netflix data pipeline.
27. Ng, J. Y. Y., Ntoumanis, N., Thøgersen-Ntoumani, C., Deci, E. L., Ryan, R. M., Duda, J. L., & Williams, G. C. (2012). Self-determination theory applied to health contexts. Perspectives on Psychological Science, 7(4), 325–340.
28. Ocumpaugh, J., Baker, R. S., & Rodrigo, M. M. T. (2014). Quantitative field observations for assessing engagement. CogSci 2014.
29. Paquette, L., de Carvalho, A. M., & Baker, R. S. (2014). Towards understanding expert coding of student disengagement. CogSci 2014.
30. Pekrun, R. (2006). The control-value theory of achievement emotions. Educational Psychology Review, 18(4), 315–341.
31. Pekrun, R., Lichtenfeld, S., Marsh, H. W., Murayama, K., & Goetz, T. (2017). Achievement emotions and academic performance. Child Development, 88(5), 1653–1670.
32. Peters, D., Calvo, R. A., & Ryan, R. M. (2018). Designing for motivation, engagement and wellbeing in digital experience. Frontiers in Psychology, 9, Article 797.
33. Peters, H., Marber, D., De Choudhury, M., Matz, S. C., Yom-Tov, E., & Schwartz, H. A. (2024). Context-aware prediction of active and passive user engagement. Journal of Big Data, 11, Article 112.
34. Picard, R. W. (1997). Affective computing. MIT Press.
35. Poria, S., Cambria, E., Bajpai, R., & Hussain, A. (2017). A review of affective computing. Information Fusion, 37, 98–125.
36. Posner, J., Russell, J. A., & Peterson, B. S. (2005). The circumplex model of affect. Development and Psychopathology, 17(3), 715–734.
37. Russell, J. A. (1980). A circumplex model of affect. Journal of Personality and Social Psychology, 39(6), 1161–1178.
38. Ryan, R. M., & Deci, E. L. (2017). Self-determination theory. Guilford Press.
39. Ryan, R. M., & Deci, E. L. (2020). Basic psychological need theory: Advancements, critical themes, and future directions. Motivation and Emotion, 44(4), 499–533.
40. Schuller, B., Batliner, A., Steidl, S., & Seppi, D. (2011). Recognising realistic emotions and affect in speech. Speech Communication, 53(9–10), 1062–1087.
41. Siemens, G., & Baker, R. S. (2012). Learning analytics and educational data mining. LAK '12, 252–254.
42. Spotify Engineering. (2019). Spotify's event delivery: Life in the cloud.
43. Spruijt-Metz, D., Hekler, E., Saranummi, N., Intille, S., Korhonen, I., Nilsen, W., et al. (2015). Building new computational models to support health behavior change and maintenance. Translational Behavioral Medicine, 5(3), 335–346.
44. Tyack, A., & Mekler, E. D. (2020). Self-determination theory in HCI games research. CHI 2020, 1–22.
45. Van den Broeck, A., Ferris, D. L., Chang, C. H., & Rosen, C. C. (2016). A review of SDT's basic psychological needs at work. Journal of Management, 42(5), 1195–1229.
46. Vansteenkiste, M., & Ryan, R. M. (2013). On psychological growth and vulnerability. Journal of Psychotherapy Integration, 23(3), 263–280.
47. Vansteenkiste, M., Ryan, R. M., & Soenens, B. (2020). Basic psychological need theory. Motivation and Emotion, 44(4), 499–533.
48. Wang, L., & Miller, L. C. (2020). Just-in-the-moment adaptive interventions (JITAI). Health Communication, 35(12), 1531–1544.
49. Wang, Y., Banerjee, C., Chucri, S., & Chen, M. (2025). Beyond black-box: Structuring landing page recommender systems using predicted intents. Stanford GSB Working Paper No. 4201.
50. Wise, A. F., & Shaffer, D. W. (2015). Why theory matters more than ever in the age of big data. Journal of Learning Analytics, 2(2), 5–13.
51. Zorfas, A., & Leemon, D. (2016, August 29). An emotional connection matters more than customer satisfaction. Harvard Business Review.