Description
Decomposes a users-by-features usage matrix into additive, non-negative archetypes so each user is described as a weighted mix of interpretable usage styles.
NMF factorises a non-negative data matrix into two non-negative matrices, producing parts-based, additive representations. Unlike PCA, factors have no negative loadings, so each is interpretable as a coherent pattern that adds to others. The treatise's transfer: apply to clickstream matrices where rows are users or sessions and columns are features, actions or screens. NMF yields K archetypes, each a distinct usage style, and users are described as weighted combinations rather than forced into exclusive clusters, which matches reality where one person is part 'reporter', part 'collaborator'. Choosing K uses reconstruction error, stability across runs and interpretability. Archetypes ground personas in behavioural data. Rated MEDIUM priority.
Attributes
- Discipline
- Computational & Machine Learning
- Loss type
- Unmigrated
- Goal
- Segment users
- Stage
- Discovery, Post-launch
- Data
- Quantitative
- Effort
- Weeks
Steps
- Build a users-by-features matrix of counts or durations over a defined window.
- Normalise (for example TF-IDF style or per-user scaling) to reduce dominance of heavy users.
- Fit NMF for a range of K (for example 3-12) with multiple initialisations.
- Choose K using reconstruction error, stability and interpretability.
- Label archetypes from top-loading features and validate with qualitative interviews of high-weight users.
- Track archetype weights over time and use them for segmentation and personas.
Sources
- Lost Methods in UX — A Complete Treatise (Gopinath)
- Lee DD & Seung HS (1999) Learning the parts of objects by non-negative matrix factorization. Nature 401:788-791
- Paatero P & Tapper U (1994) Positive matrix factorization. Environmetrics
- Lee & Seung, Learning the parts of objects by non-negative matrix factorization (1999), Nature
- Brunet et al., Metagenes and molecular pattern discovery using matrix factorization (2004), PNAS
Also known as
- NMF
- Behavioural archetype decomposition
- Behavioural archetypes via NMF
- Parts-based decomposition
- Non-negative Matrix Factorisation (NMF) for Usage Archetypes
Categories
- Analytics & Behaviour
- Computational & Machine Learning