Forced-Choice Neural Cognitive Diagnostic Model (FCNCD)

• FCNCD is a neural diagnostic model that uses forced‐choice item blocks and Q-matrix constraints to assess latent personality traits.
• It integrates distinct embeddings for participant proficiency, item difficulty, and discrimination through nonlinear mapping while enforcing monotonicity for interpretability.
• The model accommodates PICK, RANK, and MOLE formats using a weighted BPR loss, demonstrating robust performance on both real-world and simulated datasets.

Forced-Choice Neural Cognitive Diagnostic Model (FCNCD) is a deep learning-based cognitive diagnostic model for forced-choice personality testing. It is designed for settings in which participants choose among closely related options, a format used to lower the risk of response distortion in personnel selection, career development, and mental health assessment. FCNCD is presented as a model that overcomes limitations of traditional models, accommodates the three most common item block types in forced-choice tests, creates interpretable participant and item parameters under item unidimensionality, models participant–item interactions through nonlinear mapping and multilayer neural networks, and enforces a monotonicity assumption to improve diagnostic interpretability. Its reported validation on real-world and simulated datasets emphasizes accuracy, interpretability, and robustness (Li et al., 20 Jul 2025).

1. Problem setting and representational assumptions

FCNCD operates on forced-choice item blocks. Its inputs are a participant index $n$, represented as a one-hot vector $\mathbf{x}_s \in \{0,1\}^N$, and an item block $z_\ell$ containing $t$ items $\{e_{i_1}, \dots, e_{i_t}\}$. Each item $e_m$ has a one-hot encoding $\mathbf{x}_e \in \{0,1\}^M$ and is linked by a Q-matrix to exactly one latent dimension among $K$ dimensions. The notation used in the formulation includes $N$ participants, $M$ items, $\mathbf{x}_s \in \{0,1\}^N$0 latent dimensions, and block size $\mathbf{x}_s \in \{0,1\}^N$1. The participant’s latent ability is denoted $\mathbf{x}_s \in \{0,1\}^N$2, but in FCNCD it is learned implicitly through the participant representation $\mathbf{x}_s \in \{0,1\}^N$3 (Li et al., 20 Jul 2025).

A central representational assumption is unidimensionality at the item level. In forced-choice blocks, each item is associated with exactly one latent dimension via a Q-matrix row $\mathbf{x}_s \in \{0,1\}^N$4. FCNCD uses this structure to ensure that only the relevant latent dimension is used when scoring a given item. This makes the model diagnostically structured rather than a generic ranking network.

The model learns three conceptually distinct feature streams: participant proficiency, item difficulty, and item discrimination. This division is important because FCNCD does not treat forced-choice ranking as an unstructured preference problem. Instead, it preserves psychometric roles for participant and item factors while embedding them in a neural architecture. A plausible implication is that the model is intended to bridge cognitive diagnostic modeling and neural representation learning rather than replace psychometric parameterization with a purely black-box ranking system.

2. Neural architecture and feature construction

The participant embedding is parameterized by a learnable tensor

$\mathbf{x}_s \in \{0,1\}^N$5

with participant representation

$\mathbf{x}_s \in \{0,1\}^N$6

For items, FCNCD learns separate embeddings for difficulty and discrimination: $\mathbf{x}_s \in \{0,1\}^N$7 and computes

$\mathbf{x}_s \in \{0,1\}^N$8

Unidimensional masking is then applied through the Q-matrix. In the mathematical formulation, the participant factor is written as

$\mathbf{x}_s \in \{0,1\}^N$9

This masking step is the mechanism that extracts the single relevant latent dimension for each item.

After the initial embeddings are formed, FCNCD applies nonlinear mapping to each of the three streams. For participant proficiency, item difficulty, and item discrimination, the model uses sigmoid-activated fully connected layers: $z_\ell$0

$z_\ell$1

$z_\ell$2

where each $z_\ell$3 and $z_\ell$4.

The interaction vector is defined by

$z_\ell$5

This expression couples discrimination with the proficiency–difficulty contrast. The architecture therefore preserves a psychometric reading of the interaction term: discrimination modulates how participant proficiency relative to item difficulty contributes to the output score.

3. Output layer, ranking mechanism, and block-level prediction

FCNCD produces an item score through a two-layer monotonic multilayer perceptron. The first hidden representation is

$z_\ell$6

and the final output is

$z_\ell$7

with $z_\ell$8 and $z_\ell$9 constrained element-wise to be nonnegative. For each item $t$0 in the block, the model computes a score $t$1, and block-level prediction is obtained by sorting the set $t$2 (Li et al., 20 Jul 2025).

This scoring mechanism places FCNCD in the family of ranking-based forced-choice models rather than categorical choice models. The model does not directly predict only a single selected option; instead, it generates a score for each item in the block and derives the block response pattern from the induced ordering. This is essential for handling multiple block formats within a single framework.

The architecture therefore supports a common ranking backbone across different response protocols. This suggests that the model’s flexibility comes from the supervision and loss construction as much as from the neural parameterization itself.

4. Accommodation of PICK, RANK, and MOLE block types

FCNCD is explicitly described as applicable to the three most common item block types found in forced-choice tests. These are PICK, RANK, and MOLE.

For PICK, the participant selects one “best” item, and the assigned ranks are $t$3. For RANK, the response is a full ordering $t$4. For MOLE, one item is marked most with rank $t$5, one item least with rank $t$6, and the remaining items are assigned middling status (Li et al., 20 Jul 2025).

The treatment of these block types is unified by the scoring-and-sorting procedure. FCNCD computes an item score for each item in the block and then interprets the resulting ordering according to the response format. The distinction between formats enters both through the target ranks and through the training loss, especially for MOLE, where equal-rank middle items require special handling.

A common misconception in forced-choice modeling is that a model specialized for one response format cannot naturally accommodate others. FCNCD is presented specifically as a counterexample to that view: its architecture is shared, while the ranking interpretation and pairwise supervision adapt to PICK, RANK, and MOLE. At the same time, the item-level unidimensionality constraint is maintained across all three formats by masking the participant representation with the Q-matrix row corresponding to the item.

5. Monotonicity and interpretability

FCNCD incorporates a psychometric monotonicity assumption: higher latent ability in the item’s dimension should imply a higher expected score. This is enforced mathematically through nonnegative weights in the final layers,

$t$7

which in turn imply

$t$8

Under this constraint, increasing the participant’s proficiency-related feature cannot decrease the output $t$9, which is the mechanism used to preserve interpretability (Li et al., 20 Jul 2025).

The monotonicity constraint is not a decorative addition to the architecture. It is directly tied to the diagnostic meaning of the participant factor. Because FCNCD is intended as a cognitive diagnostic model rather than merely a ranking predictor, interpretability requires that the latent proficiency signal behave in a psychometrically consistent direction.

The reported interpretability metric is Degree of Agreement (DOA), specifically associated with monotonicity. The experiments report that monotonicity yields $\{e_{i_1}, \dots, e_{i_t}\}$0, far above baselines and random. At the same time, the ablation results indicate that predictive accuracy is only marginally affected by dropping monotonicity. This addresses another common misunderstanding: monotonicity is not presented primarily as a device for maximizing PRA or LRA, but as a device for improving interpretability while preserving competitive predictive performance.

6. Training objective, optimization, and evaluation protocol

FCNCD is trained with a weighted Bayesian Personalized Ranking (BPR) loss over item pairs within each block: $\{e_{i_1}, \dots, e_{i_t}\}$1 where $\{e_{i_1}, \dots, e_{i_t}\}$2 are the ground-truth ranks and $\{e_{i_1}, \dots, e_{i_t}\}$3 is a hyperparameter. For RANK blocks, the loss is averaged over all $\{e_{i_1}, \dots, e_{i_t}\}$4 item pairs. For MOLE blocks, pairs with equal rank among the middle items are ignored through an indicator mask. Regularization is provided by standard weight decay on $\{e_{i_1}, \dots, e_{i_t}\}$5 and $\{e_{i_1}, \dots, e_{i_t}\}$6. Optimization uses AdamW with Xavier initialization and early stopping on validation PRA/LRA (Li et al., 20 Jul 2025).

The evaluation metrics are Pairwise Rank Accuracy (PRA), Listwise Rank Accuracy (LRA), and Degree of Agreement (DOA) for interpretability. The reported datasets are:

• MAP: real, 1 433 participants, 88 blocks of 3 items, 24 dims, RANK
• BFI: real, 372 participants, 20 blocks of 3, 10 dims, RANK
• sim-mole: simulated, 1 000 participants, 120 blocks of 4, 24 dims, MOLE

The key findings are that FCNCD achieves the highest PRA/LRA on MAP and sim-mole and is competitive on BFI. It is also reported as robust to $\{e_{i_1}, \dots, e_{i_t}\}$7, batch size, and training split. The ablation studies confirm the benefit of nonlinear mapping, pairwise BPR weighting, and monotonic constraints for interpretability.

These training choices show that FCNCD is optimized as a ranking model with psychometric structure. The use of weighted BPR aligns the loss with relative ordering inside each block, and the format-specific treatment of item pairs allows the same model to accommodate distinct forced-choice response schemes.

7. Extensions, applications, and research significance

The reported potential extensions include incorporating response times (RT) as an additional feature stream to capture processing speed, embedding item text through BERT or similar models to enrich item difficulty and discrimination beyond the Q-matrix, adapting FCNCD for computerized adaptive testing in forced-choice format using the learned neural embeddings to drive item selection, and applying the framework to other forced-choice settings such as preference elicitation in recommender systems and pairwise comparison in crowdsourcing (Li et al., 20 Jul 2025).

These proposed directions indicate the intended scope of FCNCD beyond the immediate personality-testing setting. In particular, the suggestion to add RT as another feature stream is consistent with the model’s existing multi-stream architecture for proficiency, difficulty, and discrimination. The proposal to use BERT or similar models suggests a path toward hybrid psychometric–language modeling of item content. The computerized adaptive testing extension points to a use of learned embeddings not only for scoring but also for sequential item selection.

Within the paper’s framing, FCNCD occupies an intermediate position between traditional psychometric forced-choice models and modern neural ranking systems. It retains Q-matrix-based dimensional structure, interpretable participant and item parameters, and a monotonicity assumption, while using nonlinear mapping and multilayer neural networks to model participant–item interactions. A plausible implication is that its significance lies less in replacing psychometric theory than in recasting forced-choice cognitive diagnosis in a neural parameterization that remains diagnostically interpretable.