Causal Autoencoder-like Generation of Feedback Fuzzy Cognitive Maps with an LLM Agent

TL;DR

A large language model (LLM) can map a feedback causal fuzzy cognitive map (FCM) into text and then reconstruct the FCM from the text, which resembles the operation of an autoencoder (AE).

Abstract

A large language model (LLM) can map a feedback causal fuzzy cognitive map (FCM) into text and then reconstruct the FCM from the text. This explainable AI system approximates an identity map from the FCM to itself and resembles the operation of an autoencoder (AE). Both the encoder and the decoder explain their decisions in contrast to black-box AEs. Humans can read and interpret the encoded text in contrast to the hidden variables and synaptic webs in AEs. The LLM agent approximates the identity map through a sequence of system instructions that does not compare the output to the input. The reconstruction is lossy because it removes weak causal edges or rules while it preserves strong causal edges. The encoder preserves the strong causal edges even when it trades off some details about the FCM to make the text sound more natural.

Figures (4)

• Figure 1: Autoencoding Fuzzy Cognitive Maps (FCMs) with a single LLM agent and multi-prompting: The input FCM is on the top-left. Its edge matrix $E$ is on the bottom-left. The edge matrix colors correspond to the edge weights. Higher edge weights correspond to brighter colors. The LLM agent with the encoding prompt converts the input FCM into the text description latent I. This text description is a detailed description of the input but sounds unnatural. The LLM agent with the content-editing prompt reworks latent I into latent II. The result sounds more natural but sacrifices some detail. The LLM agent with the decoding prompt reconstructs FCMs from their text description in latent I and latent II. The top-right FCM shows that unnatural yet detailed latent I gives a lossy FCM reconstruction. The less detailed yet natural sounding latent II gives a lossier reconstructed FCM in the bottom-right FCM.
• Figure 2: Autoencoding for the clinical depression FCM: (a) Edge matrix $E$ corresponding to a 14-node FCM that models the causes of clinical depression. The nodes $C_1$--$C_{14}$ from the 1st column of Table \ref{['tab:depression_fcm']} are along the rows and columns. The source nodes are along the rows and the target nodes are along the columns. The element on the ${i}^{\text{th}}$ row and ${j}^{\text{th}}$ column gives the weight on the edge from $C_i$ to $C_j$. Brighter color on the edge matrix corresponds to bigger edge weight or strength. (b) Reconstructed edge matrix $E_{1}$ from the encoded latent I summary. The reconstruction is more accurate because the input text was detailed even if it sounded unnatural. (c) Reconstructed edge matrix $E_{2}$ from the encoded latent II summary that refined the encoded text with the content editing prompt or sub-task. The reconstruction is not as accurate because the input text was not as detailed although it sounded natural. Many non-zero edge weights in $E$ changed to zero in ${E}_2$. The edge weights that correspond to the flipped nodes $C_4$, $C_9$, and $C_{10}$ from the $3^{\text{rd}}$ column of Table \ref{['tab:depression_fcm']} are negative. These negative edges are in red. (d) The adjusted reconstructed edge matrix from latent II text with flipped nodes. The reconstruction is lossy but it preserves the stronger causal connections with larger edge weights.
• Figure 3: FCM autoencoding for a strongly-connected depression FCM subset: (a) The edge matrix $E$ corresponding to the subset of nodes $C_2$, $C_3$, $C_7$, $C_{12}$, $C_{13}$, and $C_{14}$ from the depression FCM model described by Table \ref{['tab:depression_fcm']} and Figure \ref{['fig:depression_fcm']}. The concept nodes from Table \ref{['tab:pruned_depression']} index the rows and columns. The rows list the source nodes and the columns list the target nodes. The element on the ${i}^{\text{th}}$ row and the ${j}^{\text{th}}$ column gives the weight on the directed causal edge from the ${i}^{\text{th}}$ source node to the ${j}^{\text{th}}$ target node. The brighter colors correspond to the larger (stronger) causal edge weights. (b) Reconstructed edge matrix from the encoded latent I summary. Many non-zero edge weights in $E$ are here zero but most of the bigger edge weights remain non-zero. (c) The reconstructed edge matrix from latent II summary that refined the encoded text with the content-editing prompt. Many non-zero edge weights in $E$ changed to zero in $\hat{E}_2$ while most of the strongly connected edges remained non-zero.
• Figure 4: FCM autoencoding for celiac disease classifier: (a) The edge matrix $E$ corresponding to the 8-node FCM model that classifies celiac disease. The concept nodes from table \ref{['tab:celiac_disease']} index the rows and columns. The element on the ${i}^{\text{th}}$ row and the ${j}^{\text{th}}$ column gives the weight on the directed causal edge from $C_i$ to $C_j$. The brighter colors correspond to larger (stronger) causal edge weights. (b) Reconstructed edge matrix from the encoded latent I summary. Many non-zero edge weights in $E$ are zero but most of the high-magnitude edge weights remain non-zero. (c) Reconstructed edge matrix from the encoded latent II summary that refined the encoded text with the content editing prompt or sub-task. Many non-zero edge weights in $E$ are here zero but most of the high-magnitude edge weights remain non-zero.