FHPLF: Federated Hash Projected Latent Factor

• FHPLF is a federated hash learning model that represents users and items with binary latent factors and uses projected Hamming distance for precise similarity measurement.
• It replaces large real-valued gradient tensors with binary gradient-like matrices, significantly reducing communication costs and enhancing privacy.
• The model incorporates Secure Binary Gradient Reassembly (SBG-PEU) to obscure user interaction data during transmission, achieving high accuracy with robust privacy protection.

Searching arXiv for the specified papers to ground the article in current records. Search query: (Desai et al., 2023) Federated Hash Projected Latent Factor (FHPLF) is a federated hash learning model for decentralized recommendation in which users and items are represented by binary latent factors, predicted ratings are computed through Projected Hamming Distance (PHD), and inter-client communication is based on binary gradient-like matrices rather than large real-valued gradient tensors (He, 24 Jun 2026). A closely related line of work appears in "Heterogeneous federated collaborative filtering using FAIR: Federated Averaging in Random Subspaces," which addresses federated latent-factor training under heterogeneous device memory constraints by compressing embedding tables through hashing-based random projections (Desai et al., 2023).

1. Problem setting and conceptual scope

FHPLF is situated at the intersection of Hash Learning (HL), Federated Learning (FL), and collaborative filtering. The motivating problem is that traditional HL methods typically require users to upload personal data to a central server, while most FL methods rely on transmitting large-scale real-valued gradient information. The former conflicts with increasingly stringent data security regulations, and the latter incurs high communication overhead and potential privacy risks (He, 24 Jun 2026).

The 2026 FHPLF formulation addresses this setting through three stated innovations. First, it replaces real-valued gradient matrices with binary gradient-like matrices, thereby reducing computation, storage, and communication costs while enhancing privacy protection. Second, it uses Projected Hamming Distance for similarity modeling so that individual binary bits can contribute with unequal importance. Third, it introduces Secure Binary Gradient Reassembly and Privacy-Enhanced Upload (SBG-PEU) to reduce the risk of user interaction leakage during transmission (He, 24 Jun 2026).

A parallel but distinct formulation appears in FAIR, which was introduced for heterogeneous federated collaborative filtering. FAIR is motivated by the fact that recommendation models often contain embedding tables that exceed the memory constraints of many user devices. It therefore constructs consistent and collapsible subspaces defined by hashing-based random projections, allowing clients with arbitrary compression levels to participate in training (Desai et al., 2023). This suggests that the phrase “hash projected latent factor” can denote either a specific binary-code federated recommender, as in FHPLF proper, or a broader family of federated latent-factor methods that use hashing to reconcile privacy, memory, and communication constraints.

2. Binary latent factors and the FHPLF objective

FHPLF defines a user-code matrix and an item-code matrix as

$$W=[w_1,\dots,w_{|U|}]\in\{+1,-1\}^{|U|\times D},\qquad Q=[q_1,\dots,q_{|I|}]\in\{+1,-1\}^{|I|\times D},$$

where $w_u\in\{+1,-1\}^D$ and $q_i\in\{+1,-1\}^D$ are the binary latent factors for user $u$ and item $i$ (He, 24 Jun 2026).

Prediction is based on PHD. Using the paper’s Eq. 6, the predicted rating is

$$\hat r_{u,i} =\frac{1}{D}\Bigl(D-\delta_{PH}(w_u,q_i)\Bigr) =\frac{3}{4}+\frac{1}{4D}\Bigl(w_u^\top q_i-\sum_d w_{u,d}+\sum_d q_{i,d}\Bigr).$$

This form departs from plain Hamming modeling by embedding an asymmetric projection term into the similarity calculation (He, 24 Jun 2026).

The optimization target is the squared reconstruction error with balance penalties: $$\min_{W,Q} \sum_{u\in U}\sum_{i\in I_u} \bigl(r_{u,i}-\hat r_{u,i}\bigr)^2 + \alpha\Bigl\|\sum_d w_u\Bigr\|^2 + \beta\Bigl\|\sum_d q_i\Bigr\|^2,$$ subject to $w_u,q_i\in\{+1,-1\}^D$. The role of $\alpha$ and $\beta$ is explicitly to trade off reconstruction error against the balance constraints $w_u\in\{+1,-1\}^D$0 and $w_u\in\{+1,-1\}^D$1 (He, 24 Jun 2026).

The model therefore combines two objectives that are often in tension in federated recommendation: compact binary representation and decentralized optimization. Because the latent factors are discrete from the outset, FHPLF does not treat binarization as a post hoc compression step; it makes binary structure the primary optimization domain.

3. Projected Hamming Distance and binary gradient-like updates

The asymmetric PHD used by FHPLF is defined for binary codes $w_u\in\{+1,-1\}^D$2 as

$w_u\in\{+1,-1\}^D$3

Here $w_u\in\{+1,-1\}^D$4 masks out dimensions where $w_u\in\{+1,-1\}^D$5, so the similarity calculation focuses on user-important bits (He, 24 Jun 2026).

For server-side item updates, FHPLF does not exchange real-valued $w_u\in\{+1,-1\}^D$6. Instead, each client computes a local gradient-like scalar for bit $w_u\in\{+1,-1\}^D$7 of item $w_u\in\{+1,-1\}^D$8: $w_u\in\{+1,-1\}^D$9 with $q_i\in\{+1,-1\}^D$0 defined by the paper’s Eq. 18, and then binarizes it as

$q_i\in\{+1,-1\}^D$1

Collecting these across $q_i\in\{+1,-1\}^D$2 yields the binary gradient-like vector $q_i\in\{+1,-1\}^D$3 (He, 24 Jun 2026).

This mechanism has two explicit consequences. Communication cost drops from $q_i\in\{+1,-1\}^D$4 bits to $q_i\in\{+1,-1\}^D$5 bits, and sign-only transmission curbs inversion attacks on real gradients (He, 24 Jun 2026). The stated interpretation is that binary communication is not merely a compression heuristic; it is part of the privacy model and of the representational design.

4. Federated optimization workflow and SBG-PEU

FHPLF operates in rounds. Initialization begins with the server choosing $q_i\in\{+1,-1\}^D$6 and each client $q_i\in\{+1,-1\}^D$7 choosing $q_i\in\{+1,-1\}^D$8. In each communication round $q_i\in\{+1,-1\}^D$9, the server broadcasts $u$0 to all clients. Each client then performs a local update of $u$1 by Discrete Coordinate Descent using Eqs. 11–15: for each bit $u$2, it computes the local gradient-like term $u$3, forms

$u$4

and updates $u$5 if $u$6. For the global item update, the client builds $u$7, performs SBG-PEU, and uploads $u$8. The server then aggregates $u$9, computes $i$0, and updates $i$1. The process ends when all bits of $i$2 stabilize (He, 24 Jun 2026).

SBG-PEU is designed to obscure which coordinates of $i$3 are nonzero. A client randomly decomposes $i$4 into $i$5 binary fragments satisfying

$i$6

It keeps one fragment locally, sends each of the remaining $i$7 fragments to distinct peers, receives fragments from peers, aggregates them into $i$8, and uploads that quantity. Because

$i$9

the server’s bit-wise aggregation remains valid while no individual fragment reveals the original support of $$\hat r_{u,i} =\frac{1}{D}\Bigl(D-\delta_{PH}(w_u,q_i)\Bigr) =\frac{3}{4}+\frac{1}{4D}\Bigl(w_u^\top q_i-\sum_d w_{u,d}+\sum_d q_{i,d}\Bigr).$$0 (He, 24 Jun 2026).

The privacy objective is therefore structural rather than purely perturbative. Instead of adding noise in the differential-privacy sense, FHPLF reassembles binary fragments so that global bit-wise statistics are preserved but local interaction structure is obscured. A plausible implication is that the method is aimed at protecting user-item incidence information even when the transmitted object is already binarized.

5. Empirical performance, efficiency, and privacy

The reported experiments use two real-world datasets: Amazon (D1), with $$\hat r_{u,i} =\frac{1}{D}\Bigl(D-\delta_{PH}(w_u,q_i)\Bigr) =\frac{3}{4}+\frac{1}{4D}\Bigl(w_u^\top q_i-\sum_d w_{u,d}+\sum_d q_{i,d}\Bigr).$$1, $$\hat r_{u,i} =\frac{1}{D}\Bigl(D-\delta_{PH}(w_u,q_i)\Bigr) =\frac{3}{4}+\frac{1}{4D}\Bigl(w_u^\top q_i-\sum_d w_{u,d}+\sum_d q_{i,d}\Bigr).$$2, density $$\hat r_{u,i} =\frac{1}{D}\Bigl(D-\delta_{PH}(w_u,q_i)\Bigr) =\frac{3}{4}+\frac{1}{4D}\Bigl(w_u^\top q_i-\sum_d w_{u,d}+\sum_d q_{i,d}\Bigr).$$3, and Epinion (D2), with $$\hat r_{u,i} =\frac{1}{D}\Bigl(D-\delta_{PH}(w_u,q_i)\Bigr) =\frac{3}{4}+\frac{1}{4D}\Bigl(w_u^\top q_i-\sum_d w_{u,d}+\sum_d q_{i,d}\Bigr).$$4, $$\hat r_{u,i} =\frac{1}{D}\Bigl(D-\delta_{PH}(w_u,q_i)\Bigr) =\frac{3}{4}+\frac{1}{4D}\Bigl(w_u^\top q_i-\sum_d w_{u,d}+\sum_d q_{i,d}\Bigr).$$5, density $$\hat r_{u,i} =\frac{1}{D}\Bigl(D-\delta_{PH}(w_u,q_i)\Bigr) =\frac{3}{4}+\frac{1}{4D}\Bigl(w_u^\top q_i-\sum_d w_{u,d}+\sum_d q_{i,d}\Bigr).$$6. Hyperparameters include $$\hat r_{u,i} =\frac{1}{D}\Bigl(D-\delta_{PH}(w_u,q_i)\Bigr) =\frac{3}{4}+\frac{1}{4D}\Bigl(w_u^\top q_i-\sum_d w_{u,d}+\sum_d q_{i,d}\Bigr).$$7 bits, a learning rate $$\hat r_{u,i} =\frac{1}{D}\Bigl(D-\delta_{PH}(w_u,q_i)\Bigr) =\frac{3}{4}+\frac{1}{4D}\Bigl(w_u^\top q_i-\sum_d w_{u,d}+\sum_d q_{i,d}\Bigr).$$8 “as in standard FL,” and validation-tuned penalty weights $$\hat r_{u,i} =\frac{1}{D}\Bigl(D-\delta_{PH}(w_u,q_i)\Bigr) =\frac{3}{4}+\frac{1}{4D}\Bigl(w_u^\top q_i-\sum_d w_{u,d}+\sum_d q_{i,d}\Bigr).$$9 (He, 24 Jun 2026).

Dataset	Users / Items	Density
Amazon (D1)	$$\min_{W,Q} \sum_{u\in U}\sum_{i\in I_u} \bigl(r_{u,i}-\hat r_{u,i}\bigr)^2 + \alpha\Bigl\|\sum_d w_u\Bigr\|^2 + \beta\Bigl\|\sum_d q_i\Bigr\|^2,$$0	$$\min_{W,Q} \sum_{u\in U}\sum_{i\in I_u} \bigl(r_{u,i}-\hat r_{u,i}\bigr)^2 + \alpha\Bigl\|\sum_d w_u\Bigr\|^2 + \beta\Bigl\|\sum_d q_i\Bigr\|^2,$$1
Epinion (D2)	$$\min_{W,Q} \sum_{u\in U}\sum_{i\in I_u} \bigl(r_{u,i}-\hat r_{u,i}\bigr)^2 + \alpha\Bigl\|\sum_d w_u\Bigr\|^2 + \beta\Bigl\|\sum_d q_i\Bigr\|^2,$$2	$$\min_{W,Q} \sum_{u\in U}\sum_{i\in I_u} \bigl(r_{u,i}-\hat r_{u,i}\bigr)^2 + \alpha\Bigl\|\sum_d w_u\Bigr\|^2 + \beta\Bigl\|\sum_d q_i\Bigr\|^2,$$3

Evaluation covers rating error (MAE, RMSE), ranking (HR@10, MRR@10, NDCG@10), communication cost, and privacy under gradient inversion attacks. On D1, FHPLF attains MAE $$\min_{W,Q} \sum_{u\in U}\sum_{i\in I_u} \bigl(r_{u,i}-\hat r_{u,i}\bigr)^2 + \alpha\Bigl\|\sum_d w_u\Bigr\|^2 + \beta\Bigl\|\sum_d q_i\Bigr\|^2,$$4 and RMSE $$\min_{W,Q} \sum_{u\in U}\sum_{i\in I_u} \bigl(r_{u,i}-\hat r_{u,i}\bigr)^2 + \alpha\Bigl\|\sum_d w_u\Bigr\|^2 + \beta\Bigl\|\sum_d q_i\Bigr\|^2,$$5; on D2, it attains MAE $$\min_{W,Q} \sum_{u\in U}\sum_{i\in I_u} \bigl(r_{u,i}-\hat r_{u,i}\bigr)^2 + \alpha\Bigl\|\sum_d w_u\Bigr\|^2 + \beta\Bigl\|\sum_d q_i\Bigr\|^2,$$6 and RMSE $$\min_{W,Q} \sum_{u\in U}\sum_{i\in I_u} \bigl(r_{u,i}-\hat r_{u,i}\bigr)^2 + \alpha\Bigl\|\sum_d w_u\Bigr\|^2 + \beta\Bigl\|\sum_d q_i\Bigr\|^2,$$7. In the reported ranking results, FHPLF reaches on D1 HIT@10 $$\min_{W,Q} \sum_{u\in U}\sum_{i\in I_u} \bigl(r_{u,i}-\hat r_{u,i}\bigr)^2 + \alpha\Bigl\|\sum_d w_u\Bigr\|^2 + \beta\Bigl\|\sum_d q_i\Bigr\|^2,$$8, MRR@10 $$\min_{W,Q} \sum_{u\in U}\sum_{i\in I_u} \bigl(r_{u,i}-\hat r_{u,i}\bigr)^2 + \alpha\Bigl\|\sum_d w_u\Bigr\|^2 + \beta\Bigl\|\sum_d q_i\Bigr\|^2,$$9, and NDCG@10 $w_u,q_i\in\{+1,-1\}^D$0; on D2 it reaches HIT@10 $w_u,q_i\in\{+1,-1\}^D$1, MRR@10 $w_u,q_i\in\{+1,-1\}^D$2, and NDCG@10 $w_u,q_i\in\{+1,-1\}^D$3 (He, 24 Jun 2026).

Relative to the communication baselines shown in Table 4, PFedRec/RFRec require $w_u,q_i\in\{+1,-1\}^D$4 communication, whereas LightFR and FHPLF both require $w_u,q_i\in\{+1,-1\}^D$5. The corresponding table entries are $w_u,q_i\in\{+1,-1\}^D$6 and $w_u,q_i\in\{+1,-1\}^D$7 for PFedRec/RFRec on D1 and D2, versus $w_u,q_i\in\{+1,-1\}^D$8 and $w_u,q_i\in\{+1,-1\}^D$9 for FHPLF. The same table labels FHPLF as “High” in privacy, “Fast” in inference, and “High” in accuracy (He, 24 Jun 2026).

Under gradient inversion attacks, the privacy table reports that SBG-PEU yields the highest reconstruction error among the compared upload schemes: on D1, MAE $\alpha$0 and RMSE $\alpha$1; on D2, MAE $\alpha$2 and RMSE $\alpha$3. These values are higher than both LightFR and unSBG-PEU in the reported comparisons (He, 24 Jun 2026).

The paper’s synthesis states three headline conclusions: binary gradients slash communication by $\alpha$4 and thwart gradient inversion; PHD boosts accuracy over plain Hamming FL by endowing bit-level importance; and SBG-PEU fragments gradient-like signals to foil adversaries attempting to infer user-item pairs (He, 24 Jun 2026).

6. Relation to FAIR and hash-projected federated latent factors

FAIR addresses a different bottleneck: the inability of many clients to store full embedding tables during federated training. Let the full server parameter vector be $\alpha$5. A client $\alpha$6 with memory-capacity ratio $\alpha$7 stores only $\alpha$8 parameters, with $\alpha$9 chosen as the largest power of two not exceeding $\beta$0. FAIR defines a sparse projection matrix $\beta$1 from a single universal hash function $\beta$2, and forms client-specific subspaces by collapsing columns via $\beta$3 (Desai et al., 2023).

These subspaces are described as “consistent and collapsible.” Each row of $\beta$4 has exactly one $\beta$5, different columns are orthogonal, and whenever $\beta$6, the column space of $\beta$7 is contained in that of $\beta$8. Reduction from server to client is performed by

$\beta$9

which averages the coordinates of $w_u\in\{+1,-1\}^D$00 hashing to the same bucket, while recovery from client to server is $w_u\in\{+1,-1\}^D$01. Since the client stores $w_u\in\{+1,-1\}^D$02 numbers instead of $w_u\in\{+1,-1\}^D$03, the compression ratio is $w_u\in\{+1,-1\}^D$04 (Desai et al., 2023).

The federated protocol follows FedAvg structure. In each round the server picks a fresh hash seed, samples clients, computes $w_u\in\{+1,-1\}^D$05, defines $w_u\in\{+1,-1\}^D$06, sends $w_u\in\{+1,-1\}^D$07, receives updated $w_u\in\{+1,-1\}^D$08, reconstructs $w_u\in\{+1,-1\}^D$09, and aggregates $w_u\in\{+1,-1\}^D$10. Clients reconstruct $w_u\in\{+1,-1\}^D$11, optimize $w_u\in\{+1,-1\}^D$12 locally for $w_u\in\{+1,-1\}^D$13 epochs, and return $w_u\in\{+1,-1\}^D$14. In the homogeneous case where all clients share the same subspace matrix $w_u\in\{+1,-1\}^D$15, the method reduces exactly to FedAvg on the reparameterized problem $w_u\in\{+1,-1\}^D$16, and under the standard assumptions listed in the paper—$w_u\in\{+1,-1\}^D$17-smoothness, $w_u\in\{+1,-1\}^D$18-strong convexity, uniformly bounded stochastic gradient variance, and bounded second moment—it attains the usual $w_u\in\{+1,-1\}^D$19 convergence rate on non-i.i.d. data (Desai et al., 2023).

The reported FAIR experiments span implicit-feedback ranking on Goodreads-100 and AmazonProduct-100, explicit-feedback rating prediction on Goodreads-100, and general-model tests on MNIST and FEMNIST. On Goodreads-100 ranking, centralized/full-FedAvg obtains $w_u\in\{+1,-1\}^D$20 NDCG@20 and FAIR with heterogeneous $w_u\in\{+1,-1\}^D$21–$w_u\in\{+1,-1\}^D$22 compression attains $w_u\in\{+1,-1\}^D$23, compared with $w_u\in\{+1,-1\}^D$24 for model-loss and $w_u\in\{+1,-1\}^D$25 for data-loss. On Amazon-100, full FedAvg obtains $w_u\in\{+1,-1\}^D$26 NDCG@20 and FAIR obtains $w_u\in\{+1,-1\}^D$27, compared with $w_u\in\{+1,-1\}^D$28 and $w_u\in\{+1,-1\}^D$29 for model-loss and data-loss. On Goodreads-100 rating, FAIR with $w_u\in\{+1,-1\}^D$30–$w_u\in\{+1,-1\}^D$31–$w_u\in\{+1,-1\}^D$32 compression obtains MSE $w_u\in\{+1,-1\}^D$33, versus $w_u\in\{+1,-1\}^D$34 for the LowRank baseline. On MNIST, FAIR reaches up to $w_u\in\{+1,-1\}^D$35 versus full FedAvg $w_u\in\{+1,-1\}^D$36, and at compressions up to $w_u\in\{+1,-1\}^D$37 still attains approximately $w_u\in\{+1,-1\}^D$38. On FEMNIST, FAIR matches or slightly trails full FedAvg under moderate compression and is the only method reported to run under extreme $w_u\in\{+1,-1\}^D$39 compression (Desai et al., 2023).

The relationship between FHPLF and FAIR is therefore one of architectural affinity rather than identity. FHPLF learns binary latent factors and exchanges binary gradient-like signals, whereas FAIR trains large real-valued embedding tables in hash-defined random subspaces. The shared theme is that both methods use hashing or projection to make federated latent-factor learning feasible under decentralized constraints. The divergence is that FHPLF centers discrete representation learning and privacy-enhanced upload, while FAIR centers heterogeneous memory accommodation and subspace consistency.

Both lines also leave open questions. For FHPLF, the paper notes that bit-wise discrete coordinate descent is still NP-hard in the worst case, and that acceleration strategies, continuous relaxations, adaptive bit-importance weights, dynamic fragment sizes in SBG-PEU, and theoretical analysis of convergence rates and privacy budgets under differential privacy extensions remain for future study (He, 24 Jun 2026). FAIR, by contrast, already supplies a homogeneous-case convergence guarantee, but its principal emphasis is on heterogeneous-capacity training rather than binary-code privacy (Desai et al., 2023).