workflow behaviour-pattern-recognition-in-time-series template: define-unit -> characterise -> loop(define-unit (cluster) -> contextualise -> visualise -> abstract (interpret) -> assess -> characterise (refine features)) -> define-unit (label) -> build-model -> loop(characterise (predict) -> visualise -> assess -> build-model (refine)) -> generate-knowledge description: "Human-in-the-loop visual analytics workflow for building classification models that recognise behavioural patterns in movement trajectories through iterative feature engineering and example selection" # ============================================================================ # INPUT # ============================================================================ artifact D_trajectories : entities origin: given internal structure: sequence embedment: {set, time} features: - id: f_movement value structure: vector value type: {numeric, spatial} description: "Geographic positions, speeds, headings" description: "Multivariate movement trajectories of multiple vessels over extended time period" # ============================================================================ # STEP 1: PARTITION INTO TRAJECTORY EPISODES # ============================================================================ transform T_partition : intent: define-unit manner: "time-partitioning with overlapping windows" input: D_trajectories output: D_episodes actor: hybrid description: "Partition continuous trajectories into overlapping episodes of appropriate duration to capture behavioural patterns" artifact D_episodes : entities internal structure: episode embedment: {set, time} features: - id: f_entity_id value structure: atomic value type: reference description: "Entity identifier" - id: f_temporal_extent value structure: atomic value type: temporal description: "Episode time span" description: "Time-bounded trajectory segments potentially containing behavioural patterns of interest" # ============================================================================ # STEP 2: EXTRACT INITIAL SYNOPTIC FEATURES # ============================================================================ transform T_extract_features : intent: characterise manner: "compute interval-based synoptic features" input: D_episodes output: F_synoptic actor: machine description: "Abstract elementary positional data into summary statistics representing behavioural aspects such as speed level, trajectory curvature, and spatial context" artifact F_synoptic : feature(D_episodes) value structure: vector value type: numeric description: "Initial set of synoptic features capturing speed characteristics, movement curvature, and spatial relationships" # ============================================================================ # STEP 3: FEATURE REFINEMENT LOOP # ============================================================================ loop L_feature_refinement: purpose: "Iteratively refine feature space to align with interpretable behavioural patterns" until: "Features adequately distinguish target patterns from other behaviours; episode groups show clear behavioural coherence" body: # ======================================================================== # Step 3a: Cluster episodes by feature similarity # ======================================================================== transform T_cluster : intent: define-unit manner: "cluster-by-similarity" input: D_episodes, F_synoptic output: D_clusters, F_cluster_label actor: machine description: "Group episodes with similar feature profiles to reveal natural behavioural clusters" artifact D_clusters : entities internal structure: group/cluster embedment: set description: "Groups of episodes exhibiting similar behavioural characteristics" artifact F_cluster_label : feature(D_episodes) value structure: atomic value type: categorical description: "Cluster membership identifier for each episode" # ======================================================================== # Step 3b: Project into low-dimensional space # ======================================================================== transform T_project : intent: contextualise manner: "projection by dimensionality reduction" input: D_episodes, F_synoptic output: D_projection_space, A_projection actor: machine description: "Create low-dimensional projection space from feature similarity and arrange episodes within it to expose behavioural groupings" artifact D_projection_space : entities internal structure: elementary features: - id: f_dimensions value structure: atomic value type: numeric description: "Number of spatial dimensions (2)" description: "Computed 2D projection space where proximity reflects feature similarity" artifact A_projection : arrangement(D_episodes) context: D_projection_space principle: "similarity-preserving projection from feature space" description: "Spatial arrangement where proximity indicates behavioural similarity" # ======================================================================== # Step 3c: Visualise clusters # ======================================================================== transform T_vis_clusters : intent: visualise manner: "coordinated multiple views" input: A_projection, F_cluster_label, D_episodes output: V_clusters actor: machine description: "Display episodes in projection space with cluster colouring, linked to geographic trajectory views and feature distribution displays" artifact V_clusters : visualisation(A_projection, F_cluster_label, D_episodes) layout: "coordinated views: projection plot, geographic map, feature distributions" form: "coloured points, trajectory lines, distribution charts" encoding: "projection position from arrangement; colour from cluster membership; trajectory shape and feature values from episode data" description: "Interactive visualisation enabling assessment of cluster quality and behavioural coherence" # ======================================================================== # Step 3d: Interpret patterns # ======================================================================== transform T_interpret : intent: abstract manner: "interpret behavioural patterns from clusters" input: V_clusters, F_synoptic output: P_patterns actor: human description: "Interpret clusters as behavioural patterns; assign preliminary semantic labels to discovered groupings" artifact P_patterns : pattern(D_clusters) representation form: "tentative behavioural labels with descriptions" description: "Conceptual interpretations of discovered behavioural patterns with preliminary semantic labels" # ======================================================================== # Step 3e: Assess feature adequacy # ======================================================================== transform T_assess_features : intent: assess manner: "evaluate feature adequacy and cluster coherence" input: V_clusters, D_clusters, F_synoptic, P_patterns output: K_feature_quality actor: human description: "Assess whether episodes within clusters exhibit coherent behavioural patterns and evaluate cluster separation; determine if feature refinement is needed" artifact K_feature_quality : knowledge(D_clusters, F_synoptic) representation form: "quality judgment" description: "Expert judgment of feature adequacy: whether current features achieve meaningful behavioural discrimination and cluster coherence, and whether refinement is needed" if K_feature_quality indicates refinement needed: then: transform T_refine_features : intent: characterise manner: "feature space engineering" input: D_episodes, F_synoptic, K_feature_quality output: F_synoptic_r actor: hybrid description: "Refine features based on assessment: add new features, apply transformations, adjust feature selection to better capture behavioural distinctions" artifact F_synoptic_r : feature(D_episodes) value structure: vector value type: numeric description: "Updated synoptic features with refined or expanded feature set" assign: F_synoptic := F_synoptic_r else: exit loop L_feature_refinement end loop L_feature_refinement # ============================================================================ # STEP 4: LABEL REPRESENTATIVE EXEMPLARS # ============================================================================ transform T_label_exemplars : intent: define-unit manner: "interactive selection and labelling of representative exemplars" input: D_episodes, D_clusters, A_projection, V_clusters, P_patterns output: D_exemplars, F_pattern_labels actor: human description: "Select cluster cores or high-quality representatives through visual inspection; label episodes as instances of identified behavioural patterns" artifact D_exemplars : entities internal structure: group/cluster embedment: set features: - id: f_cluster_size value structure: atomic value type: numeric description: "Number of labelled episodes per pattern type" description: "Labelled training set of behavioural pattern exemplars representing diverse pattern types discovered in the data" artifact F_pattern_labels : feature(D_exemplars) value structure: atomic value type: categorical description: "Expert-assigned behavioural pattern labels for supervised learning" # ============================================================================ # STEP 5: TRAIN INITIAL CLASSIFICATION MODEL # ============================================================================ transform T_train_model : intent: build-model manner: "supervised classification model training" input: D_exemplars, F_synoptic, F_pattern_labels output: M_classifier actor: hybrid description: "Train machine learning classifier to map synoptic features to behavioural pattern labels using labelled exemplars" artifact M_classifier : model(F_synoptic, F_pattern_labels) model type: "classifier" representation form: "supervised learning model" description: "Computational model for automated recognition of behavioural patterns from trajectory episodes" # ============================================================================ # STEP 6: MODEL REFINEMENT LOOP # ============================================================================ loop L_model_refinement: purpose: "Validate and refine model to ensure reliable pattern recognition" until: "Model errors are understood and minimised; classification accuracy is acceptable for intended application" body: # ======================================================================== # Step 6a: Apply model for prediction # ======================================================================== transform T_predict : intent: characterise manner: "apply model for prediction" input: D_episodes, F_synoptic, M_classifier output: F_predicted_labels actor: machine description: "Apply trained model to classify episodes based on their feature profiles" artifact F_predicted_labels : feature(D_episodes) value structure: atomic value type: categorical description: "Model-predicted behavioural pattern labels" # ======================================================================== # Step 6b: Visualise classification results # ======================================================================== transform T_vis_predictions : intent: visualise manner: "comparative error analysis display" input: D_episodes, D_exemplars, F_pattern_labels, F_predicted_labels, A_projection output: V_predictions actor: machine description: "Visualise classification results showing agreement and disagreement between predictions and human labels; highlight uncertain classifications" artifact V_predictions : visualisation(D_episodes, F_pattern_labels, F_predicted_labels, A_projection) layout: "coordinated views with classification results" form: "colour-coded trajectories and confidence indicators" encoding: "position from arrangement; colour from predicted class; opacity or highlighting indicates classification confidence or error" description: "Visual display of model performance showing accuracy, systematic errors, and uncertain classifications" # ======================================================================== # Step 6c: Assess model performance # ======================================================================== transform T_assess_model : intent: assess manner: "evaluate model performance and diagnose errors" input: V_predictions output: K_model_quality actor: human description: "Analyse classification errors to determine causes: ambiguous features, borderline cases, insufficient training coverage, labelling inconsistencies; decide if refinement is needed" artifact K_model_quality : knowledge(M_classifier, D_episodes) representation form: "quality judgment with diagnostic insights" description: "Understanding of model performance: identification of error types, error patterns, and potential causes; assessment of whether model refinement is needed" if K_model_quality indicates refinement needed: then: transform T_refine_model : intent: build-model manner: "model or training set adjustment" input: D_exemplars, F_synoptic, F_pattern_labels, K_model_quality, M_classifier output: M_classifier_r actor: hybrid description: "Refine model based on error analysis: adjust parameters, modify training examples, or revise feature selection" artifact M_classifier_r : model(F_synoptic, F_pattern_labels) model type: "classifier" representation form: "refined supervised learning model" description: "Updated behavioural pattern recognition model with improved performance" assign: M_classifier := M_classifier_r else: exit loop L_model_refinement end loop L_model_refinement # ============================================================================ # STEP 7: SYNTHESISE FINDINGS # ============================================================================ transform T_synthesise : intent: generate-knowledge manner: "formulate-statements" input: P_patterns, K_model_quality, K_feature_quality output: K_findings actor: human description: "Synthesise insights from pattern discovery and model development process" artifact K_findings : knowledge(M_classifier, P_patterns) representation form: "statements and insights" description: "Discovered behavioural pattern subtypes and their characteristics; understanding of pattern manifestations in data; insights into data-driven pattern recognition compared to specification-based approaches"