Home / Curated / Algorithms / Weighted Multiview Convex Mixture Clustering

Weighted Multiview Convex Mixture Clustering

Algorithms AIMech:WeightedMultiviewConvexMixtureClustering DRAFT

Aliases: Weighted Multiview CMM Clustering

A multiview exemplar-clustering procedure that learns shared exemplar weights and view weights through outer view-responsibility updates and an inner shared-weight fit, then assigns observations to a specified set of high-weight exemplars.

Identity Notes & Concept Scope This is the Tzortzis-Likas 2010 weighted multiview CMM procedure. Despite the inherited CMM name, its joint objective is not convex. It uses the log of a view-weighted sum of exemplar scores, whereas the 2009 algorithm sums view log likelihoods. This record distinguishes the general component-value formulation from the distance-only experimental variant that omits density base factors, and does not assert that marginal densities on different view spaces automatically define a joint generative model. No normalized identity collision across all 442 pinned AIO IDs, labels/scoped synonyms and 487 maintained IDs, names and aliases. Full-checkout searches include ignored and hidden files: 8 matched lines in the worktree and 9 in main. All six distinct matched texts were read; they concern the prior exemplar-clustering frontier, its source/query URLs and an unrelated FreeFem lexer token. Seven annotated AIO nodes and seven full native neighbors were inspected. The same pinned bytes retain the previously read 442 class labels and 138 semantic label/synonym rows. Absence is bounded to this snapshot and corpus.

Claims & Grounded Evidence

5 Claims

Mechanism #outer-view-and-inner-exemplar-updates

Given fixed positive component values f_vij, let Q_vi(q)=sum_j q_j f_vij and maximize L(q,pi)=sum_i log(sum_v pi_v Q_vi(q)) with q and pi on their respective simplices. From the current parameters compute gamma_iv=pi_v Q_vi/sum_w pi_w Q_wi. Set pi_v_new=(1/N)sum_i gamma_iv. Hold these outer responsibilities fixed while iterating q_j <- q_j/N*sum_iv gamma_iv f_vij/Q_vi(q), using the current inner q in the denominator. Repeat inner updates to their stated tolerance, then recompute outer responsibilities. The paper starts both weight vectors uniformly and uses inner L1-change and outer objective-change stopping conditions.

supports Source: Multiple View Clustering Using a Weighted Combination of Exemplar-Based Mixture Models · Locator: Section IV-B, printed p. 1929 / PDF p. 5, Equations 11-13 and Algorithm 1; Appendix A, printed pp. 1936-1937 / PDF pp. 12-13, Equations 22-28

Scope #objective-and-convergence-scope

The objective is separately concave in either weight vector with the other fixed, but not jointly concave in general. The nested surrogate updates support objective ascent for admissible exact arithmetic; a finite inner tolerance is an approximate M-step. Uniform starts, small objective changes or successful source examples do not prove global optimality, a unique limit, or that every stationary outcome is a local maximum. Uniform pi also leaves a log-of-sum objective, not the unweighted sum-of-logs objective.

supports Source: Multiple View Clustering Using a Weighted Combination of Exemplar-Based Mixture Models · Locator: Section IV-B, printed pp. 1928-1929 / PDF pp. 4-5; Appendix A, printed pp. 1936-1937 / PDF pp. 12-13; curator two-view objective/update comparison

Limitation #view-domain-and-density-factors

The source writes a sum of marginal densities Q_v(x_v) as a density of the full view tuple. That sum is not automatically normalized on the product observation space without an appropriate common observation model or background factors. For two- and three-state view alphabets, its total mass is 3*pi_1+2*pi_2. The finite positive-score objective still has meaningful normalized responsibilities and surrogate updates, but does not by itself establish the asserted generative interpretation. View-specific density scales affect the learned pi; they are not pure measures of semantic view quality.

supports Source: Multiple View Clustering Using a Weighted Combination of Exemplar-Based Mixture Models · Locator: Section IV-A, printed p. 1928 / PDF p. 4, Equation 9; Appendix A, printed p. 1936 / PDF p. 12, Equations 17-21; curator product-mass and coordinate-rescaling examples

Limitation #distance-only-experimental-variant

The general outer gamma calculation retains C_v(x_vi) from f_vij=C_v(x_vi)exp(-beta_v*d_vij). These factors cancel within the inner q update but not generally across views in gamma. The experiments explicitly omit them from the outer update to work from distances alone; this optimizes a modified affinity-score objective unless the factors cancel across views. With exp(-beta*||x-m||^2), the normalized Gaussian factor is (beta/pi)^(d/2), rather than the printed (beta/(2*pi))^(d/2). Source benchmark results for omitted factors do not certify the normalized-density formulation.

supports Source: Multiple View Clustering Using a Weighted Combination of Exemplar-Based Mixture Models · Locator: Section IV-B, printed p. 1929 / PDF p. 5, third implementation remark; Section V-A, printed p. 1930 / PDF p. 6, Gaussian factor and explicit removal; curator Gaussian-integral check

Mechanism #hard-assignment-and-cost

After fitting, select a specified M largest-q exemplars and anchor each to its own cluster. Assign other observations by maximizing q_j sum_v pi_v f_vij among the selected exemplars. The full-component denominator is shared and cancels for this comparison; the selected-subset scores need not sum to one as M-cluster probabilities without renormalization. Ties need an explicit policy. Dense training costs O(N^2 V tau tau_inner), excluding initial distances. Kernel squared distances can supply a feature-space realization, but a generic dissimilarity matrix does not automatically establish a normalized feature-space density.

supports Source: Multiple View Clustering Using a Weighted Combination of Exemplar-Based Mixture Models · Locator: Section IV-B, printed pp. 1929-1930 / PDF pp. 5-6, Equations 14-15; Section IV-C, PDF p. 6; Appendix B, printed p. 1937 / PDF p. 13, Equations 29-34

⚠️ Curated Limitations & Negative Results

All views must be aligned to the same observation and exemplar indices. The source explicitly does not model general dependence between views and notes failures with systematic inter-view bias.
The number of final exemplars and each component scale remain supplied choices. Hard assignment is separate from soft fitting and has no global clustering guarantee. Zero weights remain excluded; positivity and nonzero denominators matter.
The empirical comparisons use complete-view filtering, specified cluster counts, label-based scale selection and optional kernel-k-means fine-tuning. They do not establish general noise immunity, semantic calibration of view weights, or universal superiority. No author software or historical benchmark was reproduced.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:UnsupervisedClustering` Unsupervised Clustering	related_match	The procedure fits exemplar weights and constructs a grouping from unlabeled observations. The inspected AIO target is a clustering task; this algorithm-to-task association asserts neither equivalence nor an upstream subclass.

Evidence Sources & Primary Literature

1 Source

Source ID	Publication / Document	Type / Access	Reference & Checksum
`weighted-source`	Multiple View Clustering Using a Weighted Combination of Exemplar-Based Mixture Models Grigorios F. Tzortzis, Aristidis C. Likas	primary paper full_text	ID: doi:10.1109/TNN.2010.2081999 sha256:6b6c13d90ee9…

Curation Provenance

Curator: Codex

Method: ai_assisted

Model: GPT-6

Created: 2026-10-12T05:43:14Z

Updated: 2026-10-12T05:43:14Z

Source File: data/curated/WeightedMultiviewConvexMixtureClustering.yaml