Random Forests Can Hash

Hash codes are a very efficient data representation needed to be able to cope with the ever growing amounts of data. We introduce a random forest semantic hashing scheme with information-theoretic code aggregation, showing for the first time how random forest, a technique that together with deep learning have shown spectacular results in classification, can also be extended to large-scale retrieval. Traditional random forest fails to enforce the consistency of hashes generated from each tree for the same class data, i.e., to preserve the underlying similarity, and it also lacks a principled way for code aggregation across trees. We start with a simple hashing scheme, where independently trained random trees in a forest are acting as hashing functions. We the propose a subspace model as the splitting function, and show that it enforces the hash consistency in a tree for data from the same class. We also introduce an information-theoretic approach for aggregating codes of individual trees into a single hash code, producing a near-optimal unique hash for each class. Experiments on large-scale public datasets are presented, showing that the proposed approach significantly outperforms state-of-the-art hashing methods for retrieval tasks.

💡 Research Summary

The paper “Random Forests Can Hash” introduces a novel semantic hashing framework that leverages random forest ensembles to generate compact binary codes suitable for large‑scale image retrieval. Traditional random forests excel at classification but produce inconsistent binary patterns across trees for samples belonging to the same class, making them unsuitable for hashing where similarity preservation is essential. To overcome this limitation, the authors propose two complementary innovations.

First, they replace the conventional decision‑stump split function with a “transformation learner” that operates in a learned sub‑space. At each node a linear transformation (e.g., a PCA‑like projection) is learned from the data; the split is then performed on the projected values. This mechanism forces samples from the same semantic class to follow identical paths within a tree, yielding a sparse binary code where exactly d‑1 bits are set to ‘1’ (for a tree of depth d). Consequently, each tree produces a class‑consistent hash fragment.

Second, they address the aggregation of these fragments across the forest. Let B_i denote the (2^d‑2)‑bit code block generated by the i‑th tree, and let B be the collection of all M blocks. Given a target hash length L, the goal is to select k ≤ L/(2^d‑2) blocks that maximize the mutual information I(B_selected ; B_remaining). This information‑theoretic criterion prefers blocks that provide complementary information, thereby increasing the overall discriminative power of the final concatenated hash. When class labels C are available for a subset of the training data, a semi‑supervised term λ I(B_selected ; C) is added, allowing the method to exploit both labeled and unlabeled samples. The selection is performed once during training; at test time the chosen blocks are simply concatenated, incurring no additional computational cost.

The experimental evaluation covers three widely used benchmarks: MNIST, CIFAR‑10, and the PubFig face dataset. For MNIST, a forest of 64 trees with depth 3 is trained, and the method is tested under three training‑size regimes (6 000, 100, and 30 samples per class). Even with only 30 samples per class, the proposed “ForestHash” achieves 88 % precision and 68 % recall, substantially outperforming state‑of‑the‑art supervised hashing methods such as FastHash, TSH, and HDML. On CIFAR‑10, the baseline “ForestHash‑base” (using simple decision stumps) already surpasses all compared supervised methods at Hamming radius 0, and the full version with transformation learners and information‑theoretic aggregation (denoted “ForestHash”) reaches 32 % precision and 31 % recall at radius 0, with further gains at radius ≤ 2. Finally, on PubFig, using only 30 training faces per subject (5 992 faces total), the method attains 97 % precision and 85 % recall, again beating recent hashing approaches. Across all datasets the encoding time per sample is on the order of a few microseconds, demonstrating the method’s suitability for real‑time large‑scale retrieval.

In conclusion, the paper makes three key contributions: (1) it is the first to repurpose random forests for semantic hashing; (2) it introduces a sub‑space transformation learner that enforces intra‑tree hash consistency; and (3) it proposes an information‑theoretic code‑block selection scheme that yields near‑optimal, class‑specific binary codes. The approach combines feature learning, ensemble classification, and similarity‑preserving hashing in a unified pipeline that can be extended to multimodal data and resource‑constrained platforms. Limitations include the fixed shallow depth (d = 3) and modest number of trees (M = 64) used in experiments; scaling to deeper forests or longer hash lengths remains an open question, as does the computational overhead of learning the transformation matrices compared to traditional CART splits. Nonetheless, the results demonstrate that random forests, when equipped with appropriate learning and aggregation mechanisms, can serve as a powerful and efficient alternative to deep‑learning‑based hashing for large‑scale image retrieval.