-
Notifications
You must be signed in to change notification settings - Fork 0
Predictive Methods
Predictive methods in CRISP-T encompass supervised and unsupervised machine learning algorithms applied to the numeric (and optionally text-derived) portion of the corpus. Their role within triangulation is to quantify patterns that can be cross-validated against qualitative insights (topics, categories, memos). Prediction is thus positioned as confirmatory rather than exploratory.
| Category | Methods | Purpose |
|---|---|---|
| Supervised Classification | Decision Trees (--cart), K-Nearest Neighbors (--knn), Support Vector Machines (--cls), Neural Networks (--nnet), LSTM Text Classification (--lstm) |
Outcome prediction, rule extraction, pattern confirmation |
| Regression | Linear/Logistic (--regression) |
Association strength, directionality |
| Clustering | K-Means (--kmeans), Numeric clustering metadata |
Structure discovery, group comparison |
| Dimensionality Reduction | PCA (--pca) |
Latent structure, variance concentration |
| Association Analysis | Apriori rules (MCP association_rules) |
Co-occurring feature sets |
| Topological Data Analysis | TDABM (--tdabm via crispt/crispviz) |
Shape and connectivity of high-dimensional data |
- Perform textual analysis: topics (
--topics), summarization, sentiment. - Generate predictive models:
crisp --inp corpus_folder --ml --cart --regression --kmeans --pca --out corpus_folder. - Extract metadata (e.g.,
decision_tree_feature_importance,regression_coefficients). - Triangulate: compare high-importance variables with recurring narrative themes.
- Refine grounded theory categories: corroborate or challenge initial qualitative constructs.
In sales enablement data, clustering reveals a segment with high training_hours and low time_to_first_deal. Decision tree rules highlight api_errors_rate as discriminative for churn. Narratives coded with theme "technical friction" overlap this segment, producing multi-modal evidence supporting a theory of tooling reliability driving retention.
Clinical + narrative dataset on chronic fatigue: PCA concentrates variance in components loading on sleep_balance and inflammation_marker. Logistic regression finds both significant for high_fatigue. Topics about "unrestful sleep" triangulate quantitatively and qualitatively, strengthening a code about inflammatory-sleep interaction.
- Feature importance: treat as hypothesis-supporting signals, not causal proof.
- Clusters: examine distribution of narrative keywords within each numeric cluster.
- PCA components: map loadings > |0.4| to textual themes mentioning those constructs.
- Association rules: interpret high-confidence itemsets alongside narratives referencing joint occurrences.
- Small sample sizes undermine statistical stability.
- High-dimensional sparse text-derived features may inflate variance.
- Predictive success without qualitative corroboration risks atheoretical modeling.
- Maintain an audit trail: record CLI commands and outputs as memos.
- Use semantic search (
--semantic) to pull exemplar documents for numeric findings. - Avoid premature theoretical closure; treat conflicting model signals as prompts for further qualitative comparison.
- Mettler et al. (2025) — Guidance on computational text methods in qualitative IS research.
- Rudkin & Dlotko (2024) — Structural exploration via Ball Mapper.