--- title: 'FHPLF: Federated Hash Projected Latent Factor' url: https://www.emergentmind.com/topics/federated-hash-projected-latent-factor-fhplf type: topic --- # FHPLF: Federated Hash Projected Latent Factor Searching arXiv for the specified papers to ground the article in current records. Search query: 2311.01722 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 [2606.26192]. 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 [2311.01722]. ## 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 [2606.26192]. 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 [2606.26192]. 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 [2311.01722]. 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 \(i\) [2606.26192]. 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 [2606.26192]. 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 \(\sum_d w_{u,d}=0\) and \(\sum_d q_{i,d}=0\) [2606.26192]. 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 \(z_u,z_i\in\{-1,+1\}^D\) as \[ \delta_{PH}(z_u,z_i) =\bigl\|z_u-P_{z_u}(z_i)\bigr\|_H =\sum_{d=1}^D\Bigl[z_{u,d}\;\textrm{XOR}\;\bigl(z_{u,d}\land z_{i,d}\bigr)\Bigr] =\sum_{d=1}^D\bigl[z_{u,d}\land(\lnot z_{i,d})\bigr]. \] Here \(P_{z_u}(z_i)=z_u\land z_i\) masks out dimensions where \(z_{u,d}=-1\), so the similarity calculation focuses on user-important bits [2606.26192]. For server-side item updates, FHPLF does not exchange real-valued \(\nabla q_{i,k}\). Instead, each client computes a local gradient-like scalar for bit \(k\) of item \(i\): \[ \Delta q^u_{i,k} =\bigl(r_{u,i}-\tilde r_{u,i}\bigr)(w_{u,k}+1), \] with \(\tilde r_{u,i}\) defined by the paper’s Eq. 18, and then binarizes it as \[ \Delta\hat q^u_{i,k}=\operatorname{sign}(\Delta q^u_{i,k})\in\{+1,-1\}. \] Collecting these across \(i\in I_u\) yields the binary gradient-like vector \(\Delta Q_{u,k}=[\Delta\hat q^u_{i,k}]_{i\in I_u}\) [2606.26192]. This mechanism has two explicit consequences. Communication cost drops from \(O(|I|\times D\times 32\text{–}64)\) bits to \(O(|I|\times D)\) bits, and sign-only transmission curbs inversion attacks on real gradients [2606.26192]. 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 \(Q\in\{\pm1\}^{|I|\times D}\) and each client \(u\) choosing \(w_u\in\{\pm1\}^D\). In each communication round \(t=1,\dots,T\), the server broadcasts \(Q\) to all clients. Each client then performs a local update of \(w_u\) by Discrete Coordinate Descent using Eqs. 11–15: for each bit \(k\), it computes the local gradient-like term \(\Delta w_{u,k}\), forms \[ w^*_{u,k}=\Delta w_{u,k}-\alpha\sum_{d\neq k}w_{u,d}, \] and updates \(w_{u,k}\leftarrow\mathrm{sign}(w^*_{u,k})\) if \(w^*_{u,k}\neq 0\). For the global item update, the client builds \(\Delta Q_{u,k}\), performs SBG-PEU, and uploads \(\Delta Q^{\mathrm{fake}}_{u,k}\). The server then aggregates \(\Delta\hat q^u_{i,k}\), computes \(Q_k^*=\sum_u\Delta\hat q^u_{i,k}-\beta\sum_{d\neq k}Q_{i,d}\), and updates \(q_{i,k}\leftarrow\mathrm{sign}(Q^*_{i,k})\). The process ends when all bits of \(W,Q\) stabilize [2606.26192]. SBG-PEU is designed to obscure which coordinates of \(\Delta Q_{u,k}\) are nonzero. A client randomly decomposes \(\Delta Q_{u,k}\) into \(P\) binary fragments satisfying \[ \sum_{p=1}^P\Delta Q^p_{u,k}=\Delta Q_{u,k}. \] It keeps one fragment locally, sends each of the remaining \(P-1\) fragments to distinct peers, receives fragments from peers, aggregates them into \(\Delta Q^{\mathrm{fake}}_{u,k}\), and uploads that quantity. Because \[ \sum_u \Delta Q^{\mathrm{fake}}_{u,k}=\sum_u \Delta Q_{u,k}, \] the server’s bit-wise aggregation remains valid while no individual fragment reveals the original support of \(\Delta Q_{u,k}\) [2606.26192]. 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 \(|U|=35{,}736\), \(|I|=38{,}121\), density \(0.14\%\), and Epinion (D2), with \(|U|=10{,}706\), \(|I|=8{,}945\), density \(0.31\%\). Hyperparameters include \(D=64\) bits, a learning rate \(\eta\) “as in standard FL,” and validation-tuned penalty weights \(\alpha,\beta\) [2606.26192]. | Dataset | Users / Items | Density | |---|---:|---:| | Amazon (D1) | \(35{,}736 / 38{,}121\) | \(0.14\%\) | | Epinion (D2) | \(10{,}706 / 8{,}945\) | \(0.31\%\) | 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 \(0.6093\) and RMSE \(0.8358\); on D2, it attains MAE \(0.7835\) and RMSE \(1.0973\). In the reported ranking results, FHPLF reaches on D1 HIT@10 \(0.9903\), MRR@10 \(0.6931\), and NDCG@10 \(0.7675\); on D2 it reaches HIT@10 \(0.9981\), MRR@10 \(0.8135\), and NDCG@10 \(0.8606\) [2606.26192]. Relative to the communication baselines shown in Table 4, PFedRec/RFRec require \(O(|I|\,D\,\times 64)\) communication, whereas LightFR and FHPLF both require \(O(|I|\,D)\). The corresponding table entries are \(0.0364\) and \(0.0085\) for PFedRec/RFRec on D1 and D2, versus \(0.0023\) and \(0.0005\) for FHPLF. The same table labels FHPLF as “High” in privacy, “Fast” in inference, and “High” in accuracy [2606.26192]. Under gradient inversion attacks, the privacy table reports that SBG-PEU yields the highest reconstruction error among the compared upload schemes: on D1, MAE \(1.5020\) and RMSE \(1.8449\); on D2, MAE \(1.5403\) and RMSE \(1.8826\). These values are higher than both LightFR and unSBG-PEU in the reported comparisons [2606.26192]. The paper’s synthesis states three headline conclusions: binary gradients slash communication by \(>16\times\) 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 [2606.26192]. ## 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 \(\theta\in\mathbb{R}^n\). A client \(i\) with memory-capacity ratio \(\alpha_i\in(0,1]\) stores only \(m_i\le \lfloor \alpha_i n\rfloor\) parameters, with \(m_i\) chosen as the largest power of two not exceeding \(\alpha_i n\). FAIR defines a sparse projection matrix \(S_i\in\{0,1\}^{n\times m_i}\) from a single universal hash function \(h:\{0,\dots,n-1\}\to\{0,\dots,m_{\max}-1\}\), and forms client-specific subspaces by collapsing columns via \(h(a)\bmod m_i\) [2311.01722]. These subspaces are described as “consistent and collapsible.” Each row of \(S_i\) has exactly one \(1\), different columns are orthogonal, and whenever \(m_i